Computer methods

in applied
mechanics and
englneerlng
Comput. Methods Appl. Mech. Engrg. 137 (19%) 175-188

Explicit time integration algorithms for structural dynamics

with optimal numerical dissipation
Gregory M. Hulbert”,*, Jintai Chungb
“Department of Mechanical Engineering and Applied Mechanics, 2250 G.G. Brown, The University of Michigan, Ann Arbor,
MI 48109-2125, USA
bHanyang University, Ansan, South Korea

Received 25 September 1995

Abstract

A new predictor-corrector explicit time integration algorithm is presented for solving structural dynamics problems. The basis
of the algorithm is the implicit generalized-a method developed by the authors. Like its implicit parent, the explicit generalized-cl!
method is a one-parameter family of algorithms in which the parameter defines the high-frequency numerical dissipation. For a
given value of high-frequency dissipation, the explicit generalized-a method minimizes low-frequency dissipation. The algorithm
can be utilized effectively for structural dynamics calculations in which numerical dissipation is needed to reduce spurious
oscillations. Parameter values are given that enable physical damping to be treated explicitly while maintaining second-order
accuracy without the need for an expensive second corrector pass.

1. Introduction

Numerous efforts during the past several decades have focused on developing implicit time
integration algorithms for structural dynamics that include controllable numerical dissipation in the
high-frequency response domain. The purpose of this numerical dissipation is to reduce the spurious,
non-physical oscillations that sometimes occur due to excitation of spatially unresolved modes. One
basic difficulty in (designing such algorithms is to add high-frequency dissipation without introducing
excessive algorithmic damping in the important low-frequency modes. Numerous dissipative algorithms
have been developed that attain high-frequency dissipation with little low-frequency dissipation while
maintaining second-order accuracy; e.g. the Wilson-8 method [l], the HHT-(r method of Hilber et al.
[2], the WBZ-a method by Wood et al. [3], the p method of Bazzi and Anderheggen [4] and the
8,-method of Hoff and Pahl [5-71. Recently, the authors developed a second-order accurate, implicit
time integration allgorithm that is optimal in the sense that for a given value of high-frequency
dissipation, low-frequency dissipation is minimized [8]. The new algorithm, which we call the
generalized-a method, is a one-parameter family of algorithms; the parameter directly specifies the
amount of high-frequency dissipation of the algorithm. The algorithm can be directly implemented into
standard programs with little additional effort beyond that required to implement the HHT-(r or
WBZ-a methods.
Numerical dissipation is also important (and perhaps more important) when solving structural

* Corresponding author.

00457825/96/$15.00 Q> 1996 Elsevier Science S.A. All rights reserved

PII SOO45-7825(96)01036-5
176 G.M. Hulbert, J. Chung I Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188

dynamics problems using explicit methods. The principal use of explicit time integration methods is for
problems in which the time step size needed for accuracy is of the same order as the step size limit
dictated by the stability limit of an explicit method, e.g. wave propagation and impact problems. The
response of these problems usually is characterized by large gradients or discontinuities in the solution
due to propagating wave fronts. Amongst explicit time integration methods for structural dynamics, the
nearly universal choice is the central difference (CD) method which possesses no numerical dissipation.
If a mesh is constructed so that the critical time step limit for each element is the same, results from the
CD method are nearly optimal. This behavior is best illustrated by the impact of a one-dimensional,
uniform elastic bar; the CD method computes the exact nodal solution when a uniform spatial mesh is
employed. However, in practical computations, it is not feasible to construct a mesh so that the critical
time step limit is the same for all elements (one approach to circumvent this difficulty is to employ
subcycling or multi-time step integration methods). The consequence is that oscillations occur in the
solutions computed using the CD method. An explicit time integration algorithm for structural
dynamics that possesses numerical dissipation was developed by Miranda et al. [9] based upon the
implicit HHT-a method. The explicit and implicit HHT-(I! methods were then combined into a unified
implicit-explicit scheme that was shown to provide improved solutions compared to the implicit-
explicit method developed by Hughes and Liu [lo, 111 based upon the Newmark method.
A second disadvantage with the CD method is the difficulty incurred when a non-diagonal physical
damping matrix is required. Straightforward application of the CD method requires factorization of the
damping matrix, which negates the computational advantages of the CD method. An alternative
approach is to treat the physical damping explicitly by employing a predictor-corrector algorithm and
performing two corrector passes [12].
In this paper we present a predictor-corrector explicit time integration algorithm that has as its
implicit parent the generalized-a method. Similar to its parent, the explicit generalized-a method was
designed so that for a given value of high-frequency dissipation, the low-frequency dissipation is
minimized. In the presence of physical damping, parameter values are given that enable physical
damping to be treated explicitly while requiring only one corrector pass to maintain second-order
accuracy. The algorithm is first developed for the undamped case. Two sets of algorithmic parameters
are given for treating damping terms, which are delineated by application needs. Numerical results are
presented that demonstrate the performance of the explicit generalized-a method.

2. Implicit generalized-a method

We employ the standard, second-order form of the structural dynamics equations of motion, i.e.

MX + F i”t(x, Xi) = F =*(x, X, t) (1)

where M is the mass matrix, F int and F ext are the vectors of internal and external forces, respectively, X
is the (nodal) displacement vector and a superposed dot denotes differentiation with respect to time, 1.
The initial-value problem is to find X = X(t) satisfying (1) for t > 0 and the given initial conditions,
X(O)=d, X(0) = u (2)
The implicit generalized-cl! method to solve (l)-(2) has the form

Ma,+,-, m +Fi”‘(4+,-af, %+,-,) =Frxt(4++ r&-u,, t,+l-,$ (3)

LJn+l = u, + At,,+, ((1 - rk + vn+J (5)

d,=d, u. = v (6)

a, = M -‘(F ‘“‘(d, v, 0) - F i”t(d, v)) (7)

 G. M. Hulbert, J. Chung Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188 177

where

4+,-, = (1 -- +)4+1+ olpn (8)

1. = (I__ a&j+1 + apn (9)

“n+l-Yy

a n+l-a, = (I- s)a,+l + sa, (10)

t n+l-a, = Cl-. +,+I + q4l (11)

in which 12E (0,1, . . . , N - l} , N is the number of time steps and d,, u, and a,, are approximations to
X(t,), %(r,) and %(t,), respectively. In terms of the desired high-frequency dissipation, pm, the
algorithmic parameters are given by

2P, - 1 PCC
cu, =-
Pm+1 ’ af==
(12)

af ’ p=;(;+r)2

3. Predictor-corrector explicit form of the implicit generalized-a method

Following the strategy taken to develop predictor-corrector explicit (PCE) methods for the
Newmark and HIIT-& methods, a PCE implementation of the implicit generalized-o method has the
form

Predictors:

&+I = (13)
K+, = u, +(1-7)Ml+1a, (14)
Balance equation:

Ma,+,-, m +~pinf(&+l-rrfT 6+l-a,) =Fext(k+l-,lr L-qrY t,+v-u,) (15)

Correctors:

d n+1= &+, + P At~+,a,+, (16)

v n+l = -
u,+1 +-Y 4+la,+l (17)
where

L-a, = (I-- +)&+1+ olpn (18)

%+,-a I = (1 -- c+)u”,+i + +vn (19)

and cu,,,, ‘yf, /3 and y are given by (12).

In contrast to the implicit generalized-a method, pm is no longer the value of the spectral radius in the
high-frequency limit for the above PCE method. This may be seen clearly in Fig. 1 which shows the
variation in spectral radius as a function of pm (no physical damping). When pm = 0, the explicit scheme
has no numerical dissipation when the implicit generalized-a method is asymptotically annihilating in
accordance with the zero value of pm. When pm = 1, the PCE algorithm is unconditionally unstable. For
the undamped case, an expression for the stability limit of the algorithm is given by
178 G.M. Hulbert, J. Chung I Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188

iii. , . , ( , ) y

0 0.2 0.4 0.6 0.8 1.0

Fig. 1. Spectral radius of the PCE form of the implicit generalized-.a method.
Fig. 2. Stability limit of the PCE form of the implicit generalized-a method.

(20)

where the s subscript is used to denote the critical limit due to stability and a= w At in which w denotes
the maximum system eigenvalue. As shown in Fig. 2, the stability limit of the PCE implementation
decreases as px increases. Thus, unlike the PCE forms developed for the Newmark and HHT-(U
-1-
aigorithms, the generalized-a method does not admit a PCE form in which the same parametric
relations as the implicit parent can be used.

4. The explicit generalized-w method for undamped systems

To obtain a useful explicit generalized-a (EG-(u) method, we relax the restrictions (12) and derive
new relationships for the algorithmic parameters. To do so, we recall how the implicit generalized-a
m.ethod was deveioned,
The generalized-a method is spectrally defined as a three-step linear multistep method; thus, its
characteristic equation possesses three roots. Two of these roots are the principal roots while the third
is the so-called spurious root. The implicit generalized-a method was constructed by requiring the three
roots to be real and equal in the high-frequency limit, a+= cc).To maximize high-frequency dissipation,
the two principal roots remain complex conjugate except in the high-frequency limit. That is, the
principal roots bifurcate only in the high-frequency limit.
For an explicit method, the notion of high-frequency limit is replaced by the critical limit, 0,. Root
bifurcation results in a decrease in high-frequency dissipation; thus, the two limits of concern for an
,._.-I:_:. __LL__l
e;xpnc~i --- A
mtxnou art: JL, Xd tk bifi_iKZkHi hiit, U~IIUL~U
2---r-A Le.
uy JLb.
A
When designing time integration algorithms from general balance and update equations we have
found that it is best to address accuracy before stability. For second-order accuracy, the EG-cu method
requires

Y’&_X~+ af (21)
which is the same restriction on y as that of the implicit parent algorithm.
The characteristic equation of the EG-(U method for the undamped case may be written as
3

lz (iGi + y;a2)A3-’= 0 (22)

in which Gi and ri are functions of CY,,,

, af, /3 and y.
At the bifurcation limit, all three roots have real values. Let pP and p, denote the values of the
 G.M. Hulbert, J. Chung I Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188 179

principal and spurious roots, respectively, of the algorithm’s characteristic equation at the bifurcation
limit, J& (actually lthe negative values of the roots). The characteristic equation may then be written as

(A + PJ2(A + PS) = 0 (23)

Letting R = &, in (22) and equating (22) and (23), the algorithmic parameters, Q, and OS can be
expressed in terms of p, and p,

2P,P,+ Pp - 1
am
= (P,tl)(ps + 1) (24)

1 -Pp - 2PpPs (1- P,)(l - P,PJ2

p = 9 + (1+ p,)(1+ p,) + (yf(1+ P&*(1 + PJ(2 + P, - P,P,) (25)

f&l = lm5,w + (1 -P&s> (27)

4(I + PJ(2 - P,PS + P,)(3 - Pp + P, - 3PpP,)

a,= - (28)
i- 2(5 -&) + (5 - ‘3PP -Pi + P;)PS - (I - PJ3PZ

The stability limit, (28), holds provided pP # 1 or p, # 1; when pP = p, = 1, then fl, = 2, i.e. the method
has the same stability limit as the CD method. Note that (Yehas no influence on the bifurcation limit or
,... lc11e
“‘1 ,,,.&:,:c.. ,:,:c.
L, 3Ld”IULy U‘UIC,&l..- ^. Ib
Ulllb Ur :,. a^ Ilet;
c--- yalaU,cUZl.
-,...,.-^C^- IITCuf
-. =
- u,
fl CL--
LII~~I__.^
WC____- :-- - = 1; if, in &gioil,
requrrr. q
p, ~ 0,

then we recover the CD method as expected.

The characteristics of the EG-(r method can be studied in terms of pi, and pS. To maximize
high-frequency numerical dissipation, we require p, s p,. One result of this constraint is that Q s Q. If
we restrict p, s pP, then, for a given value of pP, fib is maximized when p, = pP ; this may be seen from
(27) and Fig. 3 which shows the variation in R, as a function of pP and pS. A second consequence of
enforcing pS s p, is, that the value of the spectral radius at the bifurcation limit is pp. This may be seen in
Fig. 4 which shows the variation in the spectral radius curve as a function of p,, for p,, = 0.6. Again note
that the maximum value of Q, occurs when pS = pp. Fig. 4 also shows that the EG-(I! method is similar
to the implicit generalized-o method in that for a given value of high-frequency dissipation, low-
frequency dissipation is minimized when p, = pp. Denoting the value of the spectral radius at 0,, by pb,
we have that p, = clP= pb achieves the optimal balance of low-frequency and high-frequency dissipation.
What results is a two-parameter (elf, P,,) family of EG-cu methods. Recall that in terms of the spectral

1.0r-
I.
0.9 -

0.8 -
pc’i’-
P. =p,-
r- 9 0.7 -
P.= O.Sp,---.
p.= OS& ---.- p.= 0.6-
p.=O.lp,----- 0.6 - p.= 0.4---
p.IO.2p,____.____. p.= 0.2-.-.
0.5 - p.=o.2----

Fig. 3. Bifurcation limit variation of the EG-a! method,

Fig. 4. Spectral radius dependence on p, for the EG-a method (p, = 0.6).
180 G.M. Hulbert, J. Chung i Comput. Methods Appl. Mech. Engrg. 137 (1996) 17.5-188

behavior, ‘Yeis a free parameter. Thus, a one-parameter family of EG-(Y methods may be obtained by
setting CY~
= 1. This algorithm is summarized in the box below. It is unusual in the sense that in the
balance equation, the external and internal forces are evaluated at t, while the acceleration term is
evaluated between t,, and t,, , .

Explicit generalized-a algorithm-undamped case

Ma,+,_,m = Ftxt - F’“‘(d,)

d .+,=dn+Atv,+At2(($P)n,+&+I)

V n+l =v,+At((l-y)a,+ya,+,)

= (I - amja,+l -i lqz,
%+1-a,

2Pb - 1
am= (l+p,)

5 - 3Pb
@ = (1+ Pb)2(2 - Pb)

y=;-am

al, = (I + pb)q2 - Pb

12(1 + Pbj3c2 - Pb)

n, =
10+15p,-p;+p;-p:
pb is the spectral radius value at the bifurcation limit Ob

The variation in spectral radius as a function of pb is shown in Fig. 5. Note that when pb = 0, the
EG-a method possesses an explicit time integration algorithm’s version of asymptotic annihilation. That
is, provided that the critical time step is chosen based upon 4, any high-frequency response of the
structure will be nearly annihilated in one time step. Asymptotic annihilation is an attribute of the Gear
backward differentiation implicit time integration methods that is exploited when solving constrained
dynamical systems. The high-frequency dissipation of the EG-a method may be used similarly for
systems in which the constraint equations are replaced by penalized constraints.
Measures of numerical dissipation and dispersion are provided by the algorithmic damping ratio and
relative period error. Provided the principal roots are complex, they may be expressed as

0.8 -

0.6 - pI = LO-
9 p, = 0.8 ---
p, z 0.6 -----
0.4 - p& = 0.4 ___-
. ,* t 0.2 ------
0.2 - p, = 0.0 ------

0 I. I ..I i .
0 0.5 1.0 1.5 2.0

Fig. 5. Spectral radius dependence of pb for the EG-a method.

 G.M. Hulbert, J. Chung / Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188 181

0.02 0.06 0.09 0.12 0.15 0.18

At/T At/T

Fig. 6. Numerical damping ratio of the EG-a method.

Fig. 7. Relative period error of the EG-a method.

A l-2 = exp( W rlt( - S k i))

where ,$ and 0 are the algorithmic damping ratio and algorithmic frequency, respectively. The relative
period error is given by (T - T) / T, where T = 21r/w and T = 2n/W. The numerical damping ratio of
Al__
L~,t:rn
cU_(y method is shown in Fig. 6 for different va;ues of & ml__
I ne -.--_-Z--l
numttrtcal
~_-_:__ t________
uamp111~ tncreases
L__-
lrurrt
zero when P,, = 1 tlo a maximum value when pi, = 0. The relative period error is shown in Fig. 7. It is
interesting to note that both period elongation and shortening can occur in the EG-a method,
depending on pb. Period error is minimized in the low-frequency domain when pb = 0.3665; more
precisely, period error is minimized when

lip;-36&-39p,-lO=O (30)

The spectral radii of the explicit HHT-a and EG-a methods are compared in Fig. 8. Both algorithms
have the same spectrai radius vaiue at their bifurcation iimits. (This was accompiished by choosing
a = -0.1 in the explicit HHT-a method and then numerically finding the value of spectral radius at the
bifurcation limit. This value was used directly in the EG-a method, i.e. pb = 0.8182.) Note that 4 is
much closer to the CD stability limit of 2 for the EG-a! method than the explicit HHT-a method. Also,
it is noted that the EG-a method spectral radius remains closer to 1 for a larger frequency bandwidth
than for the HHT-a method. This influence on the numerical dissipation can be seen more clearly in
Fig. 9 in which the numerical dissipation of the two algorithms is compared.

1.05 2.0 . I . I , , . I . I
1.8 - 1
Ml0 ---__ PC HHT-a ------ #‘.
--__ 1.6 - #’
--._ EG-o -
O.% 1.4 - I’ _
*. ,
*. a’
0.90 ‘\ 1.2 - a’
P ‘. iz 1.0 - a’
‘, : l’
0.85 .I
lI ly 0.8 - ,’
*’
0.80 EC&o- 0.6 -
.*’
pc HBT* ____-__. 0.4 - r’
O.l6 .*
rI .-
0.2 - __*-
_ _-- H
0.70 r.,.,.,.l 0-H’
0 0.5 1.0 1.5 2.0 0 0.02 0.06 0.00 0.12 0.15 0.18
n At/T

Fig. 8. Spectral radii of the Eg-a (pb = 0.8182) and the PCE HHT-a (a = 0.1) methods.
Fig. 9. Numerical damping ratios of the EG-a (p,, = 0.8182) and the PCE HHT-a (a = 0.1) methods.
182 G.M. Hulbert, J. Chung I Comput.MethodsAppl. Mech. Engrg. 137 (19%) 175-188

5. The explicit generalized-cY method for damped systems

Physical damping may be incorporated into an explicit time integration algorithm by treating the
damping term either implicitly or explicitly. To maintain second-order accuracy, the standard form of
the CD algorithm treats damping terms implicitly. An advantage from treating damping implicitly is
that the stability limit is not decreased, compared to the undamped case. However, to maintain the
computational speed advantage of the CD algorithm requires a diagonal damping matrix (the matrix is
obtained by differentiating Fint with respect to v). Explicit treatment of damping renders many
algorithms first-order accurate; second-order accuracy can be recovered by a second corrector pass
through the balance equation (see e.g. [12, p. 5631). Implicit and explicit treatments of damping using
the EG-a method are given. We are able to maintain second-order accuracy for both approaches
without requiring a second corrector pass.
In the contrast to the undamped case, useful analytic expressions for the bifurcation limit, 4, were
not obtained for the damped case. As a consequence, pp and p, no longer represent the values of the
algorithmic roots at L&. In the following, we will use the same definitions of algorithmic parameters as
for the undamped case, i.e. (24)-(26). Based upon the studies of the undamped case, two limits of p,
are considered: p, = 0 and p, = pp.

5.1. Explicit treatment of physical damping

Explicit treatment of physical damping for the EG-(Y method is attained by using the balance
equation, (15). To maintain second-order accuracy,
czf= 1 (31)

The stability limit for explicit treatment of damping is given by

n = 4( 1 - a,)& - J/16( 1 - cz,J*~* - 4( 1 - 2cu,)(3 - 2a,,, - 4p)

s (32)
3-2a,,,-4p

where 6 is the physical damping ratio. Typical of explicit damping treatments, an estimate of the
maximum damping ratio is required.
Figs. 10 and 11 shows the influence of pp and 6 on the critical limit, Q, for p, = 0 and p, = pp,
respectively. Here, fit denotes the minimum of L&,, 0, and the frequency at which 1A,1 > Ihl,zl first
occurs, i.e. when the spurious root magnitude exceeds that of the principal roots. As with the
undamped case, it may be seen that p, = 0 engenders a more stable algorithm. As the physical damping
increases, the critical limit decreases. With p, = pp = 1, the EG-(Y method is unstable for all non-zero
values of damping. This is not surprising since the algorithmic roots are being forced to have unity value
even though the spectral radius should be less than one due to the physical damping.

p, r 0, p, = 0 -.-.-----
@* = 0.2, p, =o __-----_ -_. -
p,=o.4,p.=o------
p,=a6,p.=o--------- - c' p,=p,=o.4------ :
.Z
E p,=,,,=0.6----.-.- :
p,=O.S, p,=O----
s
a
.o
e
v
0.2 - - -----___
0.2
0 I I . I . I .
0
0 0.2 0.4 0.6. 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
t c

Fig. 10. Influence of 5 and pp on the critical limit of the EG-a method (p, = 0).
Fig. 11. Influence of .$ and pp on the critical limit of the EG-a method (p, = p,).
 G.A4. Hulbert. J. Chung I Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188 183

1.0
0.9
0.8
0.7
0.6
0.5
0.4
p,= 0.6,p.=O-------
0.3 p,=O.l,p.=O-----
0.2
0.1
0

n R

Fig. 12. Spectral radius dependence on p, for the EC&a method (p, = 0, .$ = 0.1).
Fig. 13. Spectral radius dependence on p, for the EG-a method (p, = p,, 5 = 0.1).

Figs. 12 and 13 show the influence of pi, on the spectral radii for a lightly damped case (5 = 0. l), for
p, = 0 and p, = pi,, respectively. With p, = 0, the spectral radius follows similar trends observed in the
undamped case in that as p,, is decreased, the high-frequency dissipation is increased and the critical
limit decreases. H:owever, note that p, = 0 is a poorer choice than p,, = 0.2 since its maximum
high-frequency dissipation is smaller while its low-frequency dissipation is larger. The performance of
the matched root case (p, = p,) appears more erratic. Low-frequency dissipation behaves normally in
that decreasing pp increases low-frequency dissipation. However, in the high-frequency band, for
1.0 > pp > 0.4, decreasing p, increases the critical limit. Values of pp less than 0.4 result in the typical
trend of decreasing critical limit with increasing high-frequency dissipation. This ‘cross-over’ behavior is
actually an attribute of both p, limit cases as can be seen in the critical limit plots of Figs. 10 and 11.
For completeness, the numerical damping and period error curves are given in Figs. 14-17. The
algorithms are well behaved in the low-frequency band; the increase in numerical damping with
decrease in pp is evident as well as the period elongation and compression characteristics observed for
the undamped case.

5.2. Implicit treatment of physical damping

Provided that C+# 1, implicit treatment of damping is obtained by using a balance equation of the
form

0.50 0.50
0.45 0.45
= = 0 ----______
0.40 p, 0.2,p.
pr=0.4,p.dJ__--_ 0.40
0.35 p, = 0.6,~. = '0s-___ II.35
p, = 0.8,~. = O--__
'*2 0.30 p, = l.O,p, = 0 _ ‘U 0.30
0.25 0.25
0.20 0.26
0.15 0.15
0.10 0.10
0 0.04 0.08 0.12 0.16 0 0.04 0.08 0.12 0.16
At/T At/T

Fig. 14. Numerical damping ratio of the EG-a method (p, = 0, .$ = 0.1).
Fig. 15. Numerical damping ratio of the EG-a method (p, = pP, [ = 0.1).
184 G.M. Hulbert, J. Chung / Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188

0.06

0.04

0.02

h h 0

iz R -0.02
p, = p, = 0.2 __---_____
-0.04

p,
=p,
=
p,=o.4,p,=o----- pp=p.=o.4-_--_
-o,,6_ ,+=0.6,p,=O------- -0.06 0.6 __-_-
. p,= 0.8,~.= O-----
-0.08- p, = l.O,p,
= o- -0.08 P, = p.= 0.8---_
t
-0.10
r 1 -0.10
' 1 1
0 0.64 0.08 0.12 0.16 0 0.04 0.08 0.12 0.16
At/T At/T

Fig. 16. Relative period error of the EG-(Y method (p, = 0, 5 = 0.1)
Fig. 17. Relative period error of the EG-a method (p, = p,, 5 = 0.1).

(33)

where u~+~._ is given by (9). With implicit damping treatment, second-order accuracy requirements
impose only tfhe standard constraint on y. However, unlike the undamped case, af has an influence on
the spectral properties of the damped case. The critical limit is shown in Fig. 18 for case p, = 0. With

I . I . I . I 7

2.0 ,_--_---_-____
__----------______
----------SW_______

cl 1.5.-----------.______________
=.-.-.-s-.-w
.Z
2E p, = 1.0,
p,= o-
a 1.0- = = 0 ----
y
6 p, 08, p, -
. ,+= 0.6,
p.=O-------
,0,=0.6,,~.=0-------
0.5 - p,=o.rl,p.=o----- .'i 0.6
u _ p,=o.4,p.=o-----
p,= 0.2.p. = o--------- 0.4 - p,= 0.2,p.=o---------
. p,= 0.o.p. = o---.--- 0.2 L p,=o.o,p,=o-------
o-. I". ” ”
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
Qf c
limitdependence on a, and p, for implicit damping treatment
Fig. 18, Critical (P, = 0, 4 = O.l).
Fig, 19. Critical limit dependence on 6 and pp for implicit damping treatment (Ps = O* ar = 1’2).

,,,=,,,=0.8----
. P,= O.S,p,=O---- 1.0 P,= P,=0.6------
0.4 -
. p,= 0.6.~.=O-------
0.3 -
. p,=o.4+=0-----
0.2 - p, = 6.2.p.
= O___-_-
0.1L p,=6.O,p,~O_--._

0 0.5 1.0 1.5 2.0 “0 0.2 0.4 0.6 0.8 1.0

R “I

Fig. 20. Spectral radius dependence on pp for implicit damping treatment (p, = 0, 5 = 0.1, LY,
= 112).
Fig. 21. Critical limit dependence on a, and pp for implicit damping treatment (p, = p,).
 GM. Hulbert, J. Chung I Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188 185

pP = 1 and CX~ G l/2, the implicit treatment of damping enables L$5 2. With af = l/2, Fig. 19 shows the
stability limit of the algorithm as a function of damping and pp. It is interesting to observe that the
stability limit is 1es:s sensitive to damping values as pP decreases. Treating the damping term with a
midpoint evaluation (CY~ = l/2), Fig. 20 shows the spectral radius variation as a function of p,, for
5 = 0.1. Two differences between the damped and undamped cases are observable. First, the value of
the spectral radius at bifurcation corresponds to neither p, nor 4~. Second, similar to the explicitly
treated damping parameter set, values of pP < 0.4 are not effective since low-frequency dissipation
increases but high-frequency dissipation decreases.
With p, = pP, Fig. 21 shows that the variation in the critical limit is more pronounced with variation in
(yf than for the case p, = 0. As with the explicit damping treatment, pP = ps = 1 is an unstable choice of
parameters for the damped system, at least with the current definitions of algorithmic parameters. The
increased sensitivity also occurs for damping changes; compare Figs. 19 and 22 (cyr= l/2). Midpoint
damping evaluation, CY,= l/2, is recommended since the method has a larger stability limit in this
vicinity. An advantage of the equal root case is that the usable range of pP values increases (to
pP = 0.2); compare the spectral radius plots in Figs. 20 and 23.
Numerical damping and period error curves are shown in Figs. 24-27 for ,$ = 0.1. With implicit
damping treatment, it is possible for the numerical damping to be Eess than the physical damping value.
Such cases suggest that limits should be put on the maximum value of p, as a function of 5.

1.0
0.9
0.8
0.7
0.6
P 0.5
0.4 _ p, = p, = 0.6 ---_-

0.3 _ p,=p,=o.4__-___
0.2 _ p, = p. = 0.2 ---_______ !
p, =p, =O.O_.-.-.-
0.1 I
I I . I .
0 1
0 0.2 0.4 0.6 0.8 1.0 0 0.5 1.0 1.5 2.0
E R

Fig. 22. Critical limit dependence on 4 and pP for implicit damping treatment (p. = pP, nr = 112).
Fig. 23. Spectral radius dependence on p, for implicit damping treatment (p, = p,, 6 = 0.1, CI,= l/2).

10.8 10.8 ‘, . . I I I
10.6 - !’ ,,’
10.6
10.4 - .I ,.‘,’
10.4
./ /’ / ,
10.2 10.2 - ./ / I’ .M
.’ -4 _c-/
1o.o 1o.o : _.0_...-----_ _-c __c-H
‘-=ZZ<_
E 9.8 is 9l?- p, = p, = 0.0 ___._._ ----____ -
1” 9.6 p, = 0.2,p, = 0 __________ ‘ha/ ._
9.6 - p, = p. = 0.2 __________
9.4 p,=o.4,p,=o-----
9.4 - p,=p.=O.4_____
9.2 p, = 0.6,p, = 0 -----
p, = 0.8,p, = 0 --a- 9.2 - ,,, = p, = 0.6 -_-_-
9.0 p, = 1.0.p, = 0 - 9.0 - p, = p. = 0.8----
8.8 8.0
0 0.04 0.08 0.12 0.16 0 0.04 0.08 0.12 0.16
At/T At/T

Fig. 24. Numerical damping ratio for implicit damping treatment (pS = 0, 6 = 0.1).
Fig. 25. Numerical damping ratio for implicit damping treatment (p, = p,, 6 = 0.1).
186 G.M. Hulbert, J. Chung I Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188

0.06
___._.-.------_._.
0.04 0.04

0.02 - ______________------------------__________- 0.02

-0.02
. pr = 0.2, p, = 0 __________
-mu- pp=o.4,p.=o_____ -0.04

-0.m _ p, = 0.6, p* = 0 ----- -0.08

. p, = 0.8,~.
= O---- t p, = p. = 0.6 em__ -
1
-0.08- p,=l.o,p.=o- -0.08 p; = p, = 0.8___-.

-0.10- I I I -0.10 t I 1
0 0.04 0.08 0.12 0.16 0 0.04 0.08 0.12 0.16
At(T At/T

Fig. 26. Relative period error for implicit damping treatment (p, = 0, 5 = 0.1).
Fig. 27. Relative period error for implicit damping treatment (p, = pP, 5 = 0.1).

6. Numerical example

To demonstrate the effectiveness of the numerical dissipation inherent in the EG-(r method, we
consider the impact of a one-dimensional, homogeneous elastic rod with linearly varying cross-sectional
area; see Fig. 28. The bar is moving with a initial speed, uO, when the left end impacts the rigid wall.
The length of the bar (L) is 4; density, Young’s modulus and u0 have unit values; the cross-sectional
area is given by

A(x)=A,+
AL L-A0 x

where A,, = 1 and A, = 0.01 are the cross-sectional areas of the left and right ends, respectively and x is
measured from the left end. Employing linear rod elements, the maximum element eigenvalue is given
bY

(35)

where c is the wave speed; c = m. Using the maximum eigenvalue of all elements as an upper
bound to the maximum eigenvalue of the bar discretized by 400 elements, the critical time step, At,, for
the CD method is At, = 9.939 x 10e3. For the EG-(r method, we let pi, = 0.6; then At, = 9.408 X lo-‘.
Figs. 29 and 30 show the stress distribution in the rod for t,,, = 3, computed using the CD and EG-(w
methods, respectively. It is clear from the figures that the spurious oscillations inherent in the CD
method that result from the non-uniform mesh are reduced substantially by the EG-(U method. For
general problems, to completely eliminate the oscillations most likely requires additional discontinuity-
capturing operators (e.g. [13,14]).

Fig. 28. Configuration of the rod impact problem.

 GM. Hulbert, J. Chung I Comput. Methods Appl. Mech. Engrg. 137 (19%) 175-188 187

-2.5

Fig. 29. Stress distribution in tapered rod at t, = 3 computed using central difference method.
Fig. 30 Stress distribution in tapered rod at t, = 3 computed using EG-a method (ps = 0.6).

7. Conclusions

A predictor-corrector explicit version of the generalized-a (EG-(Y) method has been presented for
structural dynamics. The design and performance of the algorithm were described; it was shown that an
optimal combination of low-frequency and high-frequency dissipation can be obtained within a
one-parameter family of algorithms. Unlike the predictor-corrector explicit methods developed for the
Newmark and HHT-a methods, the algorithmic parameter definitions differ between the implicit and
explicit generalized-a methods. For the physically damped case, an algorithm was presented in which
the damping term may be treated explicitly while maintaining second-order accuracy with one corrector
pass. The importance of the numerical dissipation inherent in the EG-(U method was demonstrated by
comparing its computed response with that of the central difference method for a simple model
problem.
Given the new explicit algorithms and their implicit parent algorithm, it is natural to combine both
schemes into an implicit-explicit family of numerically dissipative algorithms. One obvious application
for such a mixed time integration algorithm in the treatment of constraints in dynamic problems, e.g.
contact/impact problems. Constraint equations often produce ‘stiff’ systems of equations. The
constraint equatiotrs may be treated using the highly dissipative implicit algorithm while the remaining
equations may employ dissipative or non-dissipative implicit or explicit integration. Such an approach
shows promise to accurately capture the dynamic response of a structure without incurring loss of
accuracy or severe time step size restrictions in order to handle the stiffness difficulties introduced by
the constraints.
For the damped case, the parameter set based upon the undamped case engenders effective but
sub-optimal algorithms. Studies are underway to characterize the algorithmic parameter set, in terms of
physical damping, to directly relate the algorithmic parameters to the desired high-frequency dissipa-
tion.

References

[l] E.L. Wilson, A computer program for the dynamic stress analysis of underground structures, SESM Report No. 68-1,
Division of Structural Engineering and Structural Mechanics, University of California, Berkeley, CA, 1968.
[2] H.M. Hilber, T.J.R.. Hughes and R.L. Taylor, Improved numerical dissipation for time integration algorithms in structural
dynamics, Earthquake Engrg. Struct. Dyn. 5 (1977) 283-292.
[3] W.L. Wood, M. Bossak and O.C. Zienkiewicz, An alpha modification of Newmark’s method, Int. J. Numer. Methods
Engrg. 15 (1981) 1.562-1566.
[4] G. Bazzi and E. Anderheggen, The p-family of algorithms for time-step integration with improved numerical dissipation,
Earthquake Engrg. Struct. Dyn. 10 (1982) 537-550.
188 G.M. Hulbert, J. Chung I Comput. Methods Appl. Mech. Engrg. 137 (1996) 175-188

[5] C. Hoff and P.J. Pahl, Development of an implicit method with numerical dissipation from a generalized single-step
algorithm for structural dynamics, Comput. Methods Appl. Mech. Engrg. 67 (1988) 367-385.
[6] C. Hoff and P.J. Pahl, Practical performance of the 0, method and comparison with other dissipative algorithms in structural
dynamics, Comput. Methods Appl. Mech. Engrg. 67 (1988) 87-110.
[7] C. Hoff, T.J.R. Hughes, G. Hulbert and P.J. Pahl, Extended comparison of the Hilber-Hughes-Taylor a-method and the
0,-method, Comput. Methods Appl. Mech. Engrg. 76 (1989) 87-93.
[8] J. Chung and G.M. Hulbert, A time integration algorithm for structural dynamics with improved numerical dissipation: The
generalized-a method, ASME J. Appl. Mech. 60 (1993) 371-375.
[9] I. Miranda, R.M. Ferencz and T.J.R. Hughes, An improved implicit-explicit time integration method for structural
dynamics, Earthquake Engrg. Struct. Dyn. 18 (1989) 643-653.
[lo] T.J.R. Hughes and W.K. Liu, Implicit-explicit finite elements in transient analysis: Stability theory, ASME J. Appl. Mech.
45 (1978) 371-374.
[ll] T.J.R. Hughes and W.K. Liu, Implicit-explicit finite elements in transient analysis: Implementation and numerical
examples, ASME J. Appl. Mech. 45 (1978) 375-378.
[12] T.J.R. Hughes, The Finite Element Method: Linear Static and Dynamic Finite Element Analysis (Prentice-Hall, Englewood
Cliffs, NJ, 1987).
1131 G.M. Hulbert, Discontinuity-capturing operators for elastodynamics, Comput. Methods Appl. Mech. Engrg. 96 (1992)
409-426.
[14] D.J. Benson, A new two-dimensional flux-limited shock viscosity for impact calculations, Comput. Methods Appl. Mech.
Engrg. 93 (1991) 39-95.