Vous êtes sur la page 1sur 20

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING, VOL.

36, 3937-3956 (1993)

INCREMENTAL KINEMATICS FOR FINITE ELEMENT APPLICATIONS


M. M. RASHID

Engineering Mechanics and Material Modeling Department, Sandia National Laboratorre?, Alhicqueryue. NeN Mexico, 87/85, U S A

SUMMARY
A kinematical algorithm is proposed in which the input is the incremental deformation gradient corresponding to an increment in motion for a time step, and the output is a constant rate-of-stretching tensor and a rotation tensor. The concept of algorithmic objectivity is discussed, with a stronger statcmcnt of objectivity than those advanced in previous works being set forth. The importance of strong objecticity is illustrated for applications in which moderately large rotation increments are to be expected. Accuracy of the proposed algorithm is discussed and illustrated with example calculations: and it is shown that the accuracy of the algorithm may be extended to arbitrary order. Also, the computational effort associated with the proposed algorithm is seen to compare favourably to that of other algorithms.

1. INTRODUCTION
This paper presents a means for processing the incremental motions generated by finite element solution procedures for use in constitutive integration algorithms. The basic premise of the paper is that, while the global solution procedures found in all updated-Lagrangian finite element formulations generate only discrete deformation increments, the evolution equations defining the material state generally depend on the complete history of the deformation. It is therefore necessary to assume a deformation path over each time step for the purpose of updating the material state. To be useful, the assumed deformation path over each time step must satisfy certain requirements. Among these requirements is the need for the assumed path to be of such a form that it is convenient for use in the constitutive integration algorithm. Although the precise implications of this requirement will depend on the particular constitutive relations of interest, the algorithm presented here produces a constant stretching motion such that D = constant and W = 0 (where D is the stretching rate, W is the vorticity, and D W is the spatial velocity gradient), followed by a stepwise rotation. This representation is appropriate for a wide variety of inelastic constitutive models. The deformation path is constructed for each time step given the incremental deformation gradient for the step, which in turn is trivial to compute based on quantities normally available in the finite element environment. Another requirement of any incremental algorithm is that it be objective, in some sense. Here the notions of weak and strong incremental objectivity are introduced. Essentially, a weakly objective kinematical algorithm is objective for special input motions, namely, pure rotation and pure stretch. The notion of strong objectivity is related to the algorithmic behaviour in the presence of both stretching and rotation. Tt is shown that the algorithm of Hughes and Winget, as well as other, related algorithms (e.g. that of Flanagan and Taylor), are only weakly objective. This characteristic can cause unintended coupling between the rotational part of the motion and

0029-598 1/93/233937-20$15.00 0 1993 by John Wiley & Sons, Ltd.

Received 21 July 1992 Revised 13 April 1993

3938

M . M . RASHID

the stress update whenever moderately large rotation increments are present, as illustrated in Section 4. However, lack of strong objectivity is of little consequence when both the rotation increment and the stretching increment are small, as is often the case in finite element applications. Section 2 of the paper dispenses with some preliminary kinematics, and provides a statement of the problem to be addressed. The algorithm itself is presented in Section 3, including a step-by-step summary for purposes of implementation. A synopsis of the computational effort required by the algorithm is also given in Section 3. Example calculations are presented in Section 4, wherein the effects of the strong objectivity of the proposed algorithm are illustrated for motions involving moderately large rotation increments. The relationship of the present algorithm to previously proposcd ones is also discussed in Section 4. Finally, some kinematical aspects of the computation of tangent moduli are addressed in Section 5. and a geometrical interpretation of the rotation part of the algorithm is given in the appendix. Notation is standard throughout, with Cartesian tensors bcing employed. The switch between co-ordinate-free notation and indicia1 notation is made freely; where tensorial components appear, they should be understood to represent Cartesian components. An interposed dot indicates double contraction, and the summation convention is in force. Finally, the expression x = O(c) is taken to have the standard meaning lin~,-,ox/t: # 0 and constant.
2. KINEMATICAL PRELIMINARIES

In finite element applications involving inelastic material behaviour, it is generally necessary to integrate rate-type constitutive equations over each time increment to obtain the updated material state. In the constitutive rate equations, quantities relating to the rate of deformation and rotation appear as forcing functions. The global solution procedures used in all finite element codes require that the constitutive update be performed given only the deformation increment, but not the deformation history, over each time step. However, except in thc case of elasticity, the updated material state clearly depends on the deformation history over the step, since the kinematical forcing functions that appear in the constitutive rate equations are pointwise functions of the deformation and spin rates. It is therefore clear that, for the purposes of integrating the constitutive rate equations, an assumption must be made regarding the variation of the imposed deformation over each time increment, given the total deformation increment. In formulating this assumption, the following three considerations should be borne in mind:
1. The assumed deformation history over each time step should reproduce the given total deformation increment when evaluated at the end of the step, as accurately as possible if not exactly. 2. The assumed deformation history should be reasonable in the sense that it should not involve any wild excursions enroute to the total deformation increment. Such excursions could produce unrealistic results when the constitutive equations are integrated. 3. The assumed deformation history should be such that the Constitutive rate equations are rendered as easy to integrate as possible.

The issue of incrementaE ohjectiviry as defined by Hughes and Winget falls within the domain of Item 1 above. Essentially, Hughes and Winget define an incrementally objective kinematical algorithm as one in which the incremental motion that is returned by the algorithm is identical to the input incremental motion, whenever the input motion is a pure rotation. An algorithm will be called strongly objective if, in addition to the requirements for weak objectivity, it computes a stretching part that is independent of the input rotation, when the input incremental motion

INCREMENTAL KINEMATICS

3939

involves both stretch and rotation. These definitions will be made more precise in the next section, after the necessary notation has been introduced. The implications of Item 3 will depend, to some extent, on the form of the rate constitutive equations. For the purposes of this development, it will be assumed that the constitutive rate equations are most easily integrated when the imposed deformation history is one of constant stretching rate (D = constant) and zero spin (W = 0), where

= $(L

+ LT),

W = $(L - LT)

(1)

where L is the spatial velocity gradient. The implications of this specification are discussed below. A few definitions are now introduced. The generic time increment under consideration is assumed t o be such that t E It,, t,+ 1. The configurations of the material element under consideration at I= t, and t = t,+ are denoted by K, and K, + 1 , respectively. The incremental motion over the time increment is assumed to be given in the form of the inverse of the deformation gradient p of ti,,+ with respect to K , , which may be written as

Here ;(x) is the incremental displacement field for the time step, and x is the position vector of material points in t i , + l . It is remarked that E- is readily obtained within a finite element environment, since for a typical element e, dG/ax is given by

c;Bj
-=

OXj

u:,,,

B;,,, (no sum on e )

(3)

In (3), u:,, (i = 1,2,3; A = 1,2,. . . , number of nodes per element) are the components of incremental nodal displacements for element e and for the current time step, whereas BTA are the components of the shape function gradient for node A and element r , evaluated at time t = t,+ and at the quadrature point of interest. It is emphasized that (3) was written assuming the B;,,, that are available in the finite element code are the components of the shape function gradient with respect to K,+ 1 . This is the case in many large-deformation finite element formulations. If instead the BT,, are the components of thc shape function gradient with respect to K, (i.e. evaluated at t = t,), then 6 is given by

where X is the position vector of material points in

ti,,,

and where now

- ufAB;,, (no sum on e) (5) OX, within element e. With regard to the kinematical algorithm described in the next section, the procedure will be developed using 6 as input. while it is noted that 6 may be used in place of @ - with only minor modification. Also, it bears repeating that fi is not the deformation gradient with respect to some global reference configuration, but rather the incremental deformation gradient of ti,+ with respect to K , . Hence, fi = F,+ I F,-l, where F, is the total deformation gradient at time t,. As mentioned previously, a suitable time-varying deformation F(t), t E [t,, t,+ must be developed subject to
~

-_

dBi

F(t,) = 1, F(t,+l) = E (6) In line with the comments made above equation (l), a deformation history is sought such that

3940
FF-I
=D

M. M. RASHID

(constant and symmetric), t,, < t < t,+,

(74 (7b) ( 74

F(t;+ ) = 0 (symmetric positive definite)


F(t,,+ l ) = Rfi
=

(R proper orthogonal)

Equation (7b) is a consequence of (7a). According to (7a)47c), the deformation is assumed to vary over the time step in such a way that stretching occurs along fixed principal directions and without rotation in 1 E (t,,, t,+, ), with all of the rotation occurring impulsively at t = r,,, with no additional stretch.

For future use, it is noted that any rotation tensor R may be expressed in the form
Rij = d j j

+ (1 - cos @ ( p i p j

dij) - sin Osijkpk

(8)

in which p is the unit vector along the rotation axis, 0 is the rotation angle, and E i j k = f 1 or 0 is the alternator symbol. In deriving an approximation to R in the next section, use will be made of the distance function d(R1, R2) between two rotations R 1and R 2 . The metric on the manifold SO (3) of three-dimensional rotations is usually defined so that d(R, , R,) is equal to the minimum angle of rotation required to take R 1 into R,, i.e. d (R,, R,) is the angle of rotation of RzRr around a fixed axis. Using (8) and letting pl, p, and 0, ,8, be the rotation axes and rotation angles of R , and R 2 , respectively, it can be shown that
4 sin2 [d(R,, R2)]
=

2[S?(l

+ C,) + S$(1 + C,)]- SfS?


-

+ 4aS, S2(1 - c1 cz - c1C,) + a2[4(1 C1 - C2 + C1C , ) - 6 S : S $ ] - 4a3S1 s2(1 - c1- c2 + c1c,) a4(1 c1- c, + c, c2)z
-

(9)

In (9), a = p1 - p 2 ,S, = sin 0, and C, = cos O,, where a = 1,2. Equation (9) serves to define the distance function on SO(3) when d(R', R 2 ) < n/2, which is not restrictive for the purposes considered herein. Finally, it is noted that (9) reduces to the expected expression

[d(R', R2)]'

0:

+ Q$

2nfflB2 + c4

(10)

when 1O,1 d E < 1. It is here remarked that the deformation history given by (7aH7c) is convenient for the integration of the usual rate form of isotropic plasticity, i.e.

. F = C * ( D - DP)

(1 1)

In (1l), C is the rank-four isotropic elastic modulus tensor, DPis a function of Cauchy stress T and any hardening parameters that may exist, and f = T + TW - WT represents the Jaumann stress rate. With the deformation history (7aH7c), ( 1 1) becomes

T = C .(D - DP), D constant

(12)

where now only the material rate of Cauchy stress, and not a corotational rate, appears on the left-hand side. The solution to (12) is evaluated at t = t n + l to obtain Ti;, from which the updated Cauchy stress is obtained as R'fR'. Some authors have replaced the Jaumann rate of Cauchy stress in (11) with other co-rotational stress rates, e.g. that involving C2 = RRT instead of W (see discussions by Johnson and B a m m a n ~A ~ ,t ~ l ~ r iLee , ~ et al.,' Dafalias6 and Dienes'). (Here, R arises from the polar decomposi-

INCREMENTAL K I N E M A H C S

3941

tion F = RU of the total deformation gradient.) The justification for this replacement is often based on the aphysical oscillations in stress components observed when (1 1) is used to determine the stress response in large shearing motions. However, use of the so-called Green-McInnis rate (i.e. that based on Q instead of W) introduces an awkward dependence of the stress on the chosen global reference configuration that may be no more physical than the stress oscillations in simple shear predictcd by the Jaumann-rate form of plasticity and hypoelasticity. In the context of incremental kinematics, this dependence manifests itself as the need to carry descriptors of the total deformation from step to step. In this paper, the view is taken that any history dependence of the constitutive response, including dependence on a particular, favoured reference configuration, should be explicitly accounted for through suitable state variables.' The criteria (6) and (7aH7c) were formulated consistent with this point of view, and in particular such that the desired description of the deformation history over each time step [i.e. (6) and (7aH7c)l could be determined based on the incremental motion only (as represented by k or @),independent of any global reference configuration.

'

3. AN ALGORITHM FOR INCREMENTAL KINEMATICS Calculation


(f

the stretching rate

Equations (7aH7c) indicate that the rotation tensor R (from the polar decomposition F = R6) must be extracted from F-'. An approximate method for obtaining R that is both accurate and efficient will be presented shortly. First, however, an approximate D is obtained from p- To this end, it is noted that

'.

(fj-l)2

=F-l@-T

= e-1

(13)

where = @'@. From (7a), (7b) and the assumed constancy of D in (r,, r,+ ,), common (fixed) principal directions; one is therefore justified in writing
*

0, c and D share
(14)

= exp(At D),

At

t, + 1 - t,

Equation (14) may be inverted to yield D = (l/At) log

0, or, using (13),

1 D = - log { At

[(e

1)112]

'}

The indicated operations in (1 5) are justified by the co-axiality of D and the argument of the log function, and by the positive-definiteness of [In principal, the functions appearing in (15) are evaluated by writing in spectral form and then applying the indicated operations to the principal values.] Next, it is noted that 2. differs from 1 by only a small amount, since the stretches associated with the increment in motion over a single time step are expected to be small. Accordingly, the function of appearing in (15) is expanded in a Taylor series around 1:

c.
'

c-'

Retaining only a finite number of terms in the series expansion (16) results in some truncation error. Specifically, integrating U = DU exactly subject to the initial condition U(l,) = 1, and where D is obtained from the first few terms of (16), results in a value of U(t,,+,) that differs slightly from 6. In the next section, it is demonstrated that the first two terms provide a highly accurate approximation to D for finite element applications. Before proceeding to the rotation part of the algorithm, a few observations are made regarding

3942
(1 6). First, U is written as

M.M.RASHID

ij=l+Y
where Y is symmetric and small in magnitude compared to unity. Using (17j, expressed as
c-1=

(17)

e-' may be
(18)

1 - 2 y + 3 y 2 - 4y3 + . . .

from which it is evident that the kth term in the expansion (16) is of O(Yk). Accordingly, (16) is rewritten as

) .One possible scheme by which to compute D" is to simply where D"is an approximation to I calculate F - ' as given by (2), then form -' using (13), and finally use in (19). However, if the computations are carried out in single precision, this procedure can lead to undesirable round-off error since differs from 1 by terms that are small in magnitude. To remedy this problem, it is noted that

e-'

c-1-

1= AAT-

A-

AT

1-@-I=-

a;
dX

(20)

By using (20) in (l9), problems caused by round-off error are avoided since all components of A (and not just the off-diagonal ones) are typically small compared to unity. For applications in which the bulk response is much stiffer than the deviatoric response, it may be desirable to represent the dilatational part of the incremental motion exactly, while allowing for some error in the deviatoric part. This is easily accomplished by computing the deviatoric part of D" according to (19), but replacing the spherical part by

1 trD" = ~ - 1 o g J, J At
Culculation o f the rotution tensor

[det

@'I

Turning now to the determination of R, it is emphasized that the approximation to the rotation tensor R returned by the algorithm must be strictly proper orthogonal. Using (17) and expanding 0 - I in a Taylor series around 1, @ - I becomes
@ - I =

' i f

(y - y

+ y3 -

. . .)R'.

(22)

Next, the vector a is defined by

and Q by

INCREMENTAL KINEMATICS

3943

Using (8)?(22) and (23), it may be shown that

& = sin 0 [ 1 + p.Yp - tr Y + O(Y2)]'12

and

a
7=

+ O(Y)

2JQ

where 0 and p are the rotation angle and rotation axis of R, respectively. In writing (25),it has been assumed that 0 E [0, n),since rotation angles in the range 0 E [n, 2 n ) are accommodated by reversing the direction of p. Next, the trace of F - ' is taken to produce

from which one obtains

COSO

= f(trFpl -

1) + O(Y)

(27)

Approximate values of sin2 0 and cos2 0, here named sin2 0" and cos2 Q", are now defined by cosz 0"
=

3P2[1 ( P + Q ) ] +~ _ _ _ ( p + Q)'
-

2P3[l - (f' + Q)1 ( p + QI3


(29)

sin2 8" = 1 - cos2 8" in which P is defined by

P = + [ t r P ' - 112 (30) A rotation tensor 8" which is an approximation to R, but which is nonetheless exactly proper orthogonal: is now defined using the form (8) together with (28) and (29):

The reason for defining cos2 0" and sin2 0 "as in (28) and (29, rather than simply setting sin2 0" = Q, cos2 8"
=

1- Q

(321

is related to the conditioning of (32) for B near n/2. Specifically, it is evident from (25) that Q is close to unity when 0 = n/2. However, when 0 = nj2 exactly, Q differs from unity slightly due to the terms of O(Y) appearing in (25). This error in Q results in a relatively large error in cos 0" if cos 0" is obtained by extracting the square root of (32)2,due to the behaviour of the square root function near zero. Consequently, the function of P and Q on the right-hand side of (28) was chosen so that cos2 B" = P when 0 is near n/2, whereas cos' 0" ;z 1 - Q when 0 is near zero, with a smooth transition from one case to the other. More precisely, using (28) limp~.o (cos2Od/P)= 11 for any Q, whereas limQ,o(sin2 8 " / Q ) = 1 for any P. Also, it is noted that (28) results in 0 d cos20" d 1 always, since P , Q 3 0. In extracting cos 0" from (28), the choice of sign is based on the range of the rotation angle 0"; i.e. cos 0" r 0 if 8" E [O, n/2)and < 0 if 0 E ixj2, n).(It should be noted that B" can always be taken to lie in the range [0, n), since the sensc of the rotation is automatically determined by the direction of a,) Accordingly, the sign of cos 8" is set by examining the sign of (tr F - - 1). Finally, two observations are made regarding the evaluation of (31). First, from (28) and (29), the term siniY/2JQ may be written as
7

(33)

3944

M. M. RASHID

(I

which is easily evaluated when Q is close to zero. Secondly, the evaluation of the term - cosOa)/4Q in (31) is problematic when Q is close to zero; in this case, the truncated Taylor series 1 - costr 1 P2 - 1 2 ( P - 1) (P - 2 j ( P - 1OP 32) =+Q 32p2 + Q2 64P3 4Q 8

Q3

_____.__

1104 - 992P + 376P2 - 72P3 + 5P4 - + 0(Q4) 512 P4

(34)

should be used. The approximation given in (34) is within 10 of the exact functional value when Q is 0.01 or smaller. It can be shown that d(R", 8 ) is of O(Y).As will be demonstrated in Section 4, this level of accuracy is virtually always adequate for large-deformation finite element computations. However, it i s possible to increase the order of accuracy of R a arbitrarily by preconditioning @ before using it in (23) and (30). Specifically, 6 = 0 R' can be premultiplied by an approximation to U, so that the result differs from fi' by terms of O(Ykj. k > 1. Again expanding in a Taylor series around C = 1,
~

'

~'

'

u = 1 -+(c-'- 1) + &c-l - I)* -&(e-'


A
~

113

+ ...

(35)

where -- 1 is obtained from (20).It can be shown that premultiplying F-' by N terms of (35) and then using the reFult in (23) and (30) to compute R a results in d(R", R ) = O(YN)for arbitrary rotation angles. However, it is emphasized that the additional accuracy realized by preconditioning is rarely worth the associated computational expense (see below for an estimate of the computational expense involved in this and other kinematical algorithms). Finally, it is mentioned that the direction of a departs significantly from that of p as the rotation angle 8 approaches 180". This behaviour is ultimately related to the fact that a must change sign at 0 = 180". For this reason, the algorithm should be restricted to rotations of not more than, say, 178" - in - practice. This restriction can be removed by replacing the direction of a with that of E , , ~ F ~ , , , F z; ccjkF;k ~ F;! when 0 is near 180" (see the discussion on geometrical interpretation in the appendix). Alternatively, in the event that the rotation angle is found to be near 180", the input F - ' may be postmultiplied by a rotation of 180" around the direction of a before entering the kinematical algorithm.* In this way the rotation that is returned by the algorithm is rendered small, and the incremental rotation for the step is found by composing this small rotation with the transpose of the one used to postmultiply @ - I . In any case, these procedures are invoked rarely if at all, since a 180" rotation in a single increment is so large that it is unlikely to be encountered in practice.
A -

'

Outline of the algorithm The computational steps in the proposed algorithm are outlined below. The number of floating point operations involved in each step are indicated. The input to the algorithm is A = 1 - 6 as defined in (20), and is computed according to (3).
Step I. (42 operations):

'

Compute

C-'

- 1 = AAT - A - A'

*This idea was advanced by Dr. Sam Key.

LNCREMENTAL KINEMATICS

3945

Step 2. (50 operations):

1 Approximate D by D" = -- - { At Step 3. (6 operations):


Compute a, = c l l k F l i l = Step 4 . (11 operations): ComputeQ=ix,a,,tr@-'-l

$(e ' - 1) + i(c '


-

1)2)

--

EllkAlk

and P = $ ( t r E - ' -

1j2

Step 5. (24 operations, plus two evaluations of square root):

Compute cos2 0" from (28) and take the square root, and compute sin Oa/2& from (33).

Step 6. (21 operations using first three terms of equation (34)):


Compute (1 - cos P)/4Q using the sign of (tr 6 close t o zero.
Step 7. (24 operations):
~

- 1) for the sign

of cos 8"; use (34) when Q is

The total computational effort involved in the algorithm is 178 floating point operations, plus two evaluations of square root. This level of effort compares favourably with that associated with other kinematical algorithms, as discussed in the next section. Preconditioning E with the first two terms of (35) requires 54 additional operations, whereas preconditioning with the first three terms, yielding d(R", R ) = O(Y3), requires 66 additional operations. After the stress integration has been performed using D as given in Step 2 and with W = 0, the final updated stress is obtained by rotating the stress forward by R ' . This rotation requires 75 floating point operations. It bears mention that all of the operations involved in the forgoing algorithm vectorize fully on vector processors.
~

'

Objectivity of' the algorithm In order to discuss objectivity of kinematical algorithms, it is convenient to introduce the following definition: the output ofa kinematical algorithm is the deformation gradient Pd= fi.6. that results from exact integration over [t,, rT+ '1 of the description of the motion produced by the algorithm, subject to the initial condition F"(t,,) = 1. For example, in the algorithm described above, 6"= where = exp [Ar D"] and R" is as given in (31). An algorithm will be called weakly objectiue if

Raca, ca

1. 6"= @ whenever fi is proper orthogonal and 2. R a = 1 whenever 6 is positive definite symmetric.


Here @ is the input incremental deformation gradient, as before. 'Incremental objectivity' as defined by Hughes and Winget' corresponds to Item 1 above, whereas Pinsky et al.9 include both Criteria 1 and 2 in their definition of incremental objectivity. A kinematical algorithm will be called strongly objective if Criteria 1 and 2 above for weak objectivity are satisfied, and in addition if 3.

6" remains unchanged when

is replaced by Q@,where Q is an arbitrary rotation tensor.

3946

M. M. RASHID

Comparing the two definitions, it is clear that the essential difference lies in the behaviour of the stretch in the presence of rotation: weak objectivity requires that there be no output rotation for a pure stretch input, and no output stretch for a pure rotation input, whereas strong objectivity requires in addition that the output stretch be unaffected by the input rotation for non-vanishing input stretch. This stronger statement is perhaps closer to the general concept of objectivity encountered in mechanics; namely, that two motions that differ only by a rotation should induce responses that differ only by the same rotation. Strong objectivity obviously implies weak objectivity, but strong objectivity does n o t require that the algorithm in question be exact (i.e. Pa = F always). Indeed, an algorithm may return approximate values of R a and fixand still be strongly objective, so long as 0 .is independent of the input rotation. That the proposed algorithm is strongly objective (and therefore weakly objective) follows from the facts that u = 0 and cos 6' = 1 when fi = 1, implying that fia = 1 in this case, and that D"is independent of R trivially (see equations (23), (24), (28) (31) and (19)). 4. PERFORMANCE AND COMPAKlSQN WITH OTHER ALGORITHMS
The algorithm of Hughes and Winget

The proposed algorithm may be directly compared to the algorithm of Hughes and Winget' (see also Reference lo), since the input and output of the two algorithms arc essentially the same. In the procedure of Hughes and Winget (hereafter referred to as HW), the spatial gradient G of the displacement field evaluated at the mid-poinr of the time step is used to form both R " and D" for the step. To obtain G, it is necessary to make some assumption regarding the manner in which the nodal velocities vary over the step; in Reference 1, they are taken to be constant. Making this assumption, G may be obtained by evaluating the shape function gradient B:,, at the time-step mid-point. Alternatively, G may be expressed in terms of the deformation gradient for the step as

G = 2(@- I)(@ + I )

'

(36)

An expression similar to (36) is also given by Pinsky et Using notation consistent with that of Section 3, Hughes and Winget approximate D (constant over the step) by

and R by

R a = (1 *o)(l -)w)~', o )(G - GT) (38) The origin of the HW algorithm is as follows. First, the assumption is made that the muterial velocity field is constant in time over the step, i.e. the velocity of each material point is assumed to be independent of time. Then, the spatial velocity gradient corresponding to this velocity field is evaluated at the mid-point of the time step, yielding 1/At G. The approximate stretching rate tensor D"is set equal to the symmetric part of this mid-point velocity gradient, and o is set equal to the antisymmetric part. The approximate rotation tensor R" is obtained by integrating R = WR using the modified mid-point rule with a single step. This approximate integration obviously requires knowledge of the mid-point spin o only. Rubinstein and Atluri" have proposed a similar procedure for the determination of ria from the mid-point spin o, except that they employ the Cayley-Hamilton theorem to obtain an exact expression for exp(At o) instead of using the modified mid-point rule. Pinsky et al.,' o n the other hand, prove that, if R = WR is

IKCREMENTAL KINEMATICS

3947

integrated approximately using a generalized linear single-step scheme, then the result is strictly proper orthogonal only when the modified mid-point rule is used. Hughes and Winget show that haas given above is strictly proper orthogonal for arbitrary F, and that D = 0 and R a = fr when P = , where k is a rotation of less than 180'-. In the terminology of this paper, then, the HW algorithm is at least weakly objective. However, it turns out that the H W algorithm is not strongly objective as defined in the previous section. Indeed, using (37) and (16), one can show that
At

D" = Y - +Yz + by3 + . . .

+ fI($(UZ - ZY) + $(ZY' + 82(+(Z2Y + YZ2) + tzvz -+- . . . 1. +

- Y'Z)

+ ...}
(39)

for small Y aizd sinall rotation angle 8, where Z , = 6 i j k p k and p is the rotation axis of R. NOW, if the deformation K, -+ x,+ is a pure stretch (is. if 8 = O), then substitution of (18) into (16) and comparison of the result with (39) reveals that D" = D + O(Y ,) In this case. This is the same order of accuracy as that obtained by using (19) to calculate Da in the algorithm of Section 3. However, if 8 # 0, and in particular if the rotation angle is fairly large. then (39) indicates that D" = I) + O(Y), i.e. the error in D" is first order in Y. The consequences of this characteristic can be illustrated by considering the following problem. imagine a block of isotropic, linearly elastic material subjected to a motion whose deformation gradient is given by

[ F i j ] = [Rik] [.!Ikj] =

where v is the Poisson's ratio of the material. The motion given by (40) gives rise to a uniaxial stress state along the material direction that is aligned with the s,-axis at t = 0, along with a simultaneously occurring rotation around the x,-axis. Let (1) = n and p = IO4/E ( E the Young's modulus), so that at t = 1.0, the tensile axis will again be aligned with the x,-axis and the stress will have reached a value of lo4. Now, imagine partitioning the interval 0 d t < 1.0 into 6 equal time steps, so that Ar = and t,* = n At, n = 0, 1, . . . , 6. The incremental deformation gradient @ , of FC, with respect to K , , is given by

cos cut sin o t

- sin wt

Itbt
0
1

0
1 - l$C

cos tot
0

0 O

, t 30

(40)

1Ov p t I

P, = F,F;-ll
,

(41)

where F,, rz = 0. 1, . . , 6, is computed according to (49) with 1 replaced by 1,. The P, obtained from (41) is used in both the algorithm of Section 3 (with no preconditioning) as well as in the HW algorithm as represented by (361438). In both cases, the increment in stress is computed according to

AT = Al(A1 tr D" + 2pD")

(42)

which. is exact within the approximation of small total stretch. In (42),?, and p are the usual Lami: constants corresponding to E = 10' and v = 0.3. After incrementing the stress, the final updated stress tensor is obtained by rotating forward using k". The computed solutions using both the algorithm of Section 3 and the HW algorithm are compared to the exact solution in Figure 1. It is emphasized that. although the rotation increment in each step is 30": the stretches are very small, with a total stretch on the order of The significant error exhibited by the H W algorithm i s related to the f-i-term in (39), which in turn is ultimately due to the assumption of constant nodal velocities within each time step. However, it is

3948
10000.0

M M RASHID

7500.0

exact solution and proposed algorithin

5000.0

2500.0

0.0

-2500.0

-5000.0 0.0

0.2

0.4

0.6

0.8

1.o

TIME (SEC)
Figure 1. Comparison of the exact solution to those generated by the proposed algorithm (circles) and thc Hughes-Winget algorithm (squares) for a linearly elastic material element simultaneously undergoing stretching and rotation. Total rotation is 180', whereas total stretch in the direction of nonzero stress is 0.001

further emphasked that the error illustrated in Figure 1 is a direct consequence of the fairly large rotation increments, which are only occasionally encountcred in finite element practice. In writing (39), it has been assumed that the HW algorithm is used in the same mode as the algorithm proposed in Section 3, in the sense that the stress update is first performed subject to D" and W = 0, after which the incremented stress is rotated by R". In Reference I it is suggested that these steps be performed in the reverse order [see their equations (1 1) and (12)]. If the rotation is performed first, followed by a pure stretch along fixed directions, then the expression for the corresponding stretching rate D is (43) where = 1 + y is the left stretch tensor in the polar decomposition @ = qk,and y is symmetric and small. Using y instead of Y, the expression for I)" [as defined by (37)] that replaces (39) is

AtD" = y - i y 2 + b y 3 + ' . . + 8{:(yZ - Zy) + $(Zy2 - y2%)+ . . .$

+ UZ{i(ZZy + y z q + 2ZyZ +

. .} + . . .

(44)

Comparison of (39) with (44) reveals that D"as defined by (37)can be used equally well to update the stress (and other state variables) either before or after the rotation has been applied. The error incurred by using the same D in both cases is of the same order as the #-terms in (39) and (44), which are present in any case. Turning now to the rotation part of the HW algorithm, it can be shown that the error in R a given by (38) is second order in Y, i.e. that d(R", k)= O(Yz). To achieve this order of accuracy with the first two terms of using the algorithm of Section 3, it is necessary to precondition

@-'

INCKkMkNTAL KINbMATlCS

3949

(35). A comparison between the rotation part of the proposed scheme and that of the HW algorithm may be illustrated by considering the following example problem. Let gl, = i l k u k , , where [fi,,] = diagjl - ~ L Y1, * a , 1 4x1 (45)

Furthcr, let R be specified by the rotation axis and rotation angle


[/4]=--(1,1,1), J 3 3

B = - 17 4

(44)

Using (45),the magnitude of Y is I Y 1 = (Y Y) = a. Both the proposed algorithm and the HW algorithm were used to compute the approximate rotation R dcorresponding to the incremental motion given by (45) and (46). The results are plotted in Figure 2 for 0 < CI < 0.04, with the distance d(k,6) being plotted versus a for both algorithms. From Figure 2, it is clear that the rotation part of the H W algorithm is extremely accurate. In the context of the proposed algorithm, the accuracy gains to be realized by preconditioning are evident from Figure 2. However, CI = 0.04 represents a fairly large incremental stretch in finite element practice; the error in fiawithout preconditioning still seem5 to be acceptable for most applications. In any case, any degree of accuracy can be obtained by preconditioning without incurring undue computational expense (see the comments at the end of Section 3). Finally, it is observed that the HW algorithm as expressed by (36j ( 3 8 ) requires approximately 216 floating point operations to return R a and D. This may be compared with the 232 floating point operations plus two evaluations of square root required by the algorithm of Section 3 using first-order preconditioning (1 78 plus two square roots without preconditioning). The 216 figure is based on the assumption that (or k) serves as the input to the algorithm. In most finite element formulations, it is more computationally efficient to form G using (36) than to compute G directly by recomputing the shape function gradients at the time step mid-point.
Cornparison to other procedures

Flanagan and Taylor have presented a kinematical algorithm that is similar in motivation to both the algorithm of Section 3 and the HW method, in that the constitutive update procedure is provided with a deformation path which simplifies the objective stress rate to simply a material rate. In their algorrthm (hereafter referred to as FT), Flanagan and Taylor consider the Green-McInnis objective stress rate rather than the Jaumann rate. Accordingly, the FT algorithm generates an approximate deformation path in which D = constant and R = constant over t E ( t n ,t n f l ) , after which R changes stepwise to its updated value at t = t,+ I . Here R is the rotation tensor associated with the gradient of the total deformation. This deformation path may be contrasted with that given by (7aH7c), in which the rotation associated with the incremental deformation is held at R = 1 over the step. In the FT algorithm, the mid-point values of D and W are obtained as in the HW algorithm. Then, a result of Dienes, in which 0 = RRT is related to D, W and V (the total left stretch tensor), is applied. The updated total rotation tensor is then obtained by applying the modified mid-point rule as in Reference 1, but using i2 instead of W. Since the total R is available in the FT scheme, the constitutive integration is performed in the so-called unrotated configuration. That is, D is back-rotated to RZD R, for the purposes of performing the stress update, after which the updated stress is forward-rotated to R, + T R;f for use in the statement of momentum balance. The left stretch tensor V is updated using a single forward Euler step, and is stored for use in Dienes relation for 0.

3950
1.500

M. M.RASHID

0.000 L j

_STRAIN MAGNITUDE (ALPHA)

0.050

' i

0.040

z w

PROPOSED ALGORITHM, SECOND-ORDER PRECONDITIONING

0
[r

\
0.030

b a 3 a:
w

HUGHES-WINGET ALGORITHM
0.020

fl8
,
\

g
F

2
0.010

PROPOSED ALGORITHM, FIRST-ORDER

0.000

STRAIN MAGNITUDE (ALPHA)

Figure 2. Error in the computed rotation for a dcfurmation increment that corresponds to a 45' rotation-about the (1, I , 1) oxis, with a stretch as given by equation 145). I n this figure, thr: rotation error i s defined as d(R" R)/(z/4)

The behaviour of the F T algorithm is similar to that o f the WW algorithm. In particular, the coupling between the stretching and ro:ation as exhibited by equalions (39) and (44)is present for large rotation increments, due to tht: LIFC of the mid-point velocity gradient and assumed constant material velocity field. The I T algorithm requires approximately 500 floating point operations,

INCREMENTAL KINEMATICS

395 1

including the determination of the mid-point velocity gradient from [using ( 3 6 ) ] and the back-rotation of D, but not including the forward rotation of the updated stress. It bears mention that the F T algorithm has been in successful use in the PRONTO and JACI3 family of finite element codes for several years. Attention is now focused on the procedure described by Pinsky et ~ 1 Although . ~ this algorithm is not truly a kinematical algorithm in that it does not provide a deformation path for the constitutive integration, it has some features relevant to the present discussion. The essential aspect in Pinsky et a/. is the observation that any generalized rate of stress (such as the Jaumann co-rotational rate or the Truesdell rate) may be written as a malerial rate of an appropriately transformed stress followed by the reverse transformation. This conclusion arises naturally in their formulation, as they define all tensor fields over the manifold of body points, rather than using Cartesian tensors defined over a manifold of spatial points as is done here. Within this context, all generalized rates may be thought of as special forms of the Lie derivative, in which the tensor quantity in question is pulled back to some other configuration, the material rate is taken, and the result is pushedjimvard to the current configuration. Pinsky et aL9 consider rate forms of hyperelasticity and hypoelasticity, in which the stress rate is approximated by a weighted average of the end-point stress tensors pulled back to a common configuration. An equation for the updated stress results from equating this weighted average to the pull-back of the stress rate evaluated at an intermediate time point. This latter quantity generally involves the rate-ofstretching D and the stress evaluated at the intermediate point. Pinsky et nL9 assume that the material velocity field is constant over the step, which leads to the restriction that the intermediate point must be chosen to be the time mid-point if the overall procedure is to be (weakly) objective. However, due to the constant-material-velocity-field assumption, the stress update procedure of Pinsky et aL9 is not strongly objective for the same reasons given in connection with the HW algorithm. They give a numerical example problem (their Section 6.3) that is similar to the problem of Figure 1 above, but which involves a total stretch of 2.0 and a total rotation of 360. A departure of the numerical solution from the exact one is not seen in their calculation, since both the stretch and the rotation increments are of the same order of magnitude, rendering the error in the stress small [see equation (39)l. Rerunning their example problem using their algorithm, but with a total stretch ratio of 1.01 instead of 2, resuits in a comparison between the computed and exact solutions similar to that shown in Figure 1. A completely different class of constitutive update procedures has recently emerged in which kinematical algorithms of the type considered here play no role. Examples of such procedures, here referred to as non-rate-form update methods, have been advanced by, e.g., Moran et aI.I4 and Weber and Anand; and, in the context of the so-called operator-split methodology, by The essential idea is to first integrate a rate equation for the Simo and Ortiz and Ortiz er plastic part of the deformation, after which the elastic deformation and the stress are obtained by function evaluation. Perhaps chief among the attractions of the non-rate-form approach is that strong objectivity is trivially satisfied. However, this approach is limited by two essential aspects: i t only applies to constitutive formulations for which there exists some form of kinematical decomposition of the motion into elastic and plastic parts (e.g. the multiplicative decomposition F = FTPj,and it is useful only when the plastic evolution equation can be adequately integrated with a singlc-step integration rule, so that the imposed deformation is sampled only at the time-step end-points. These limitations imply that there are many situations of practical interest for which it is necessary or desirable to appeal to a treatment of the incremental kinematics of the type described herein.

3952

M.M. RASHID

5. THE TANGENT MODULUS TENSOR


In quasi-static finite element codes, and in dynamic codes in which the equations of motion are integrated using an implicit operator, it is generally necessary to provide the global solution algorithm with a tangent modulus tensor. The tangent modulus appears in the expression for rate of change of internal nodal forces with respect to incremental nodal displacements, which in turn is used in the global Newton-Raphson (or other suitable) iterative solution procedure. Since the derivative of the incremental deformation gradient F with respect to incremental nodal displacements is element-formulation dependent, the focus in this section will be on the derivative of updated stress with respect to 6-l. As with the kinematical algorithm itself, the corresponding expressions taking F as given are obtained by slightly modifying those given below. Let the initial and updated Cauchy stress for time step n be T, and T,+ respectively. Also, let the Cauchy stress obtained by integrating the constitutive equations over t E ( t n , t,+ subject to D = D" [as given by (19)] and W = 0, be ?. ?'hen,

T,+ = RaTRdT
where Ra is given by (31). The objective is to compute the rank-four tensor

(47)

It will be assumed that the rank-four tensor

39 C=(49) dU is available from the constitutive update routine. In connection with (49), it is noted that

At ~-O(Y)

dD"

a? +

For economy of presentation, R' and F - ' will hereafter be replaced with R and A, respectively. Combining (47)-(49), it is seen that (in component form)

The terms i?R/BA and dU/BA are not trivial to calculate, and are the essential results of this section. The approach taken here is to look for linear relations between A and R,and between A and U. To this end, consider the polar decomposition F = RU, from which one obtains

A
Taking the rate of the expression (U
&j

A(RU

+ RU)A

(52)

AA', using the result in (52), and rearranging leads to

- FikAImRmj4- UikAmlRjm -k + & R j k )

(53)

Recalling that U differs from I by terms of O(Y), one arrives at

A similar procedure leads to

INCREMENTAL KINEMATICS

3953

The O(1) terms given in (54) and (55) are expected to be entirely adequate for use in computing the tangent modulus tensor C. Combining (55), (54) and (51), the final expression IS
Cij~ =
fmn {Rjn[$dkm6,1

- dn,lRtk
f

!iRIkRml]

+ Rcrn[:6kndjl

- 6nlRjk

$R)kRnll}

RlrnRwRjvCnrrnk

(56)

6. CONCLUSION

The purpose of the kinematical algorithm proposed herein is to express the deformation increment for a time step in a form that is convenient for use in constitutive integration of the incremental algorithms. The kinematical algorithm takes as input the inverse 8 deformation gradient for the step. which is easily obtained from the incremental displacements and the shape-function gradients. The output is a rate-of-stretching tensor D" and a rotation tensor R", such that the gradient of the deformation corresponding to D = Dd(constant), W = 0 for t E [ t n ,t n + followed by stepwise application of the rotation Ra, is very nearly $ at t = t n +1 . A feature of the algorithm is that it can be made arbitrarily accurate simply by incurring the additional expense of retaining more terms. That is, D"can be made arbitrarily close to the exact value (l/At) l o g o , and R a to the exact value R, where 6 = Rc.For the basic form of the algorithm presented in Section 3, it is shown that D" = D O(Y3), and that d(R", R ) = O(Y), where Y = 6 - 1 and where d( ,.) represents the distance function on the three-dimensional rotation manifold SO(3).This level of accuracy is generally more than adequate for finite element applications. The important issue of objectivity is discussed in relation to both the present algorithm as well as to previously-proposed algorithms. Previous treatments have interpreted algorithmic objectivity in terms of pure input rotation and pure input stretch. Here, the notion of strong objectivity is introduced, which, in addition to the criteria for weak objectivity, requires that the output stretch (i.e. f l a = exp [At D"] in the present case) remains unchanged when an arhitrury input deformation is altered by a rotation. While this stronger statement of algorithmic objectivity is perhaps closer to the usual concept of objectivity encountered in mechanics, it is not the strongest statement that might be made. Indeed, the most severe requirement would be to simply demand that the output and input deformations be exactly equal for arbitrary input deformations. One of the basic differences between the present algorithm and those proposed by Hughes and Winget,' Flanagan and Taylor,' and Pinsky et aL9 is that the latter three proposals all compute the stretching rate based on the assumption that the material velocity field is constant in time over the time step. This assumption leads to a lack of strong objectivity, i.c. to a coupling between the stretching and rotation parts of the incremental motion. This coupling can lead to significant errors in the stress update whenever the rotation increment is moderately large, even when the incremental stretch is infinitesimal. In the present approach, care is taken to avoid such a coupling between the rotational and stretching parts of the incremental deformation. The result is a strongly objective kinematical algorithm for which the computational effort involved (about 200 floating point operations) is comparable to that of other kinematical algorithms that have been proposed.

ACKNOWLEDGEMENTS

The author would like to express appreciation to Dr. Sam Key for fruitful discussions in connection with this work. The work reported herein was performed at Sandia National Laboratories in Albuquerque, New Mexico, and was supported by the United States Department of Energy under contract DE-AC04-76DP00789.

3954

M . M. RASHID

APPENDIX
The rotation part of the kinematical algorithm presented in Section 3 admits an interesting geometrical interpretation, as follows. Consider two material points with position vectors X 1 and X2 in configuration ti,,. Suppose these two material points move to positions x1 and x 2 in the homogeneous deformation K, -+ K , , + ~ defined by E. Let the unit vector M that lies along X2 - XI be given by the Cartesian components ( M I ,M 2 , M 3 ) = (sin 4 cos i,b, sin $ sin $, cos $), where {$, ( 1 ) are polar angles. Now, the relative displacement (x2 - x , ) - (X2 - X Ij of the two points lies along the vector FM - M. This vector is pictured in Figure 3 for M in the { x l , c 2 } plane and for 6 corresponding to a pure stretch, a pure rotation, and a more general motion involving both stretch and rotation. Figure 3 suggests that the integral of the cross product (FM - M) x M ( = f?M x Mj over all orientations of M vanishes when F is a pure stretch, i.e.

On the other hand, it turns out that this integral lies along the rotation axis of R when

@ = R:

jo2n J.

( k M - M) x M sin 6 d 4 d$

= ~sin O p

-8n 3

where again 8 and p are the rotation angle and rotation axis of

6. More generally,
fl p t- Q(Y)

:1

M)x M s i n 4 d+ d$

= __-sin

871 .

(59)

Carrying out this integral leads to an expression similar to (25),. Finally, and with reference to the comments made in Section 3 regarding the direction of

x2

Figure 3(a)

INCREMENTAL. KfNEMATICS

3955

/
I

,///'-/
I

Figure 3. Plot of the displacement vector (EM - M) when is (a) a pure stretch, (b) a pure totatton and (c) a general deformation

a when 0 z 180", it can be shown that

Jo J o
Evaluation of the integral in (60) yields a vector that is proportional to C,,~F~,,,F;~, whose direction, it turns out, differs from that ofp by terms of O(Y), and is a good approximation to the direction of p when f? is near 180".
REFERENCES
1. T.J. R. Hughes and J. Wingct, 'Finite rotation effects in numerical intcgration of rate constitutive equations arising in large-deformation analyses', lnt. j . numer. methods eng., 15, 1862-1867 (1980).

3956

M. M. RASHID

2. D. P. Flanagan and L. M. Taylor. An accurate numerical algorithm for stress integration with finite rotations, Coniput. Methods Appl. Mech. Eng., 62, 305-320 (1987). 3. G. C. Johnson and D. J. Bammann, A discussion of stress rates in finite deformation problems, Int. J. Solids Strucf., 20, 725- 737 (1984). 4. S . N . Atluri, On constitutive relations at finite strain: hypo-elasticity and elasto-plasticity with isotropic or kinematic hardening, Cornput. Mrthods Appl. Mech. Eng., 43, 137~-171 (1984). 5 . E. H. Lee, R. L. Mallett and T. B. Wertheimer, Stress analysis for anisotropic hardening in finite-deformation plasticity, J. Appl. Mech., 50, 554-560 (1983). 6 . Y. F. Dahlias, Torotational rates for kinematic hardening at large plastic deformations, J . Rppl. Mech., 50, 56 1-565 (1983). 7. J. K. Dienes; On the analysis of rotation and stress ratc in deforming bodies, Acta hfrch., 32, 217-232 (1979). 8. M. M. Rashid, A class of constitutive models for rate-dependent inelasticity in metals, in S.-C. Chou, (ed.), Proc. 12th Army Symposium o n Solid Mechanics, Plymouth, MA, November 1991 pp. 521-537. 9. P. M. Pinsky, M . Ortiz and K. S . Pister, Numerical integration of rate constitutive equations in finite deformation analysis, Comput. Methods 4ppI. Mech. Eng., 40, 137-1 58 (1 983). 10. T. J. R. Hughes, Numerical implementation of constitutive models: rate-independent deviatoric plasticity. in S. Neniat-Nasser, R. J. Asaro and G . A. Hegemier, (eds.), Theoretical Foundation f o r Large-Scak Computations .for Nonlinear Material Rehatlior, Martinus Nijhoff, Dordrecht, 1984 pp. 29 57. 11. R. Rubinstein and S. N. Atluri, Objectivity of incremental constitutive relations over finite time steps in computational finite deformation analyses, Compuf. Methods Appl. Mech. Eny., 36, 277-290 (1983). 12. L. M. Taylor and D. P. Flanagan, PRONT03D a three-dimensional transient solid dynamics program. Sandia National Laborntorirs Report SANDR6-0594, Albuquerque, New Mexico, 1989. 13. J. H. Biflle, JAC3D--a three-dimensional finite element computer program for the nonlinear quasi-static response of solids with the conjugate gradient method, Sundia Nalionul Laboratories Report SANDX7-I.305, Albuquerque, New Mexico, 1993. 14. B. Moran; M. Ortiz and C. F. Shih, Formulation of implicit finite element methods for multiplicative finite deformation plasticity, Int. j . numrr. mehods eng., 29, 483-514 (1990). 15. G. Weber and L. Anand, Finite deformation constitutive equations and a time integration procedure for isotropic, hypcrelastic-viscoplastic solids, Comput. Methods Appl. Mech. Eng., 79, 173-202 (1990). 16. J. C. Simo and M. Ortiz, A unified approach to finite deformation elastoplastic analysis based on the use of hyperelastic constitutive equations, Comput. Methods Appl. Mech. Eng., 49, 221L245 (1985). 17. M. Ortiz, P. M. Pinsky and R. L. Taylor, Operator split methods for the numerical solution of the elastoplastic dynamic problem. Comput. Methods A p p l . Mech. Eng., 39. 137-1 S7 (1983).

Vous aimerez peut-être aussi