© 2001 Elsevier Science Ltd. All rights reserved.

Computational Mechanics - New Frontiers for New Millennium

S. Valliappan and N. Khalili, editors. 35

EXPLICIT/IMPLICIT PARTITIONING OF THE

GENERALIZED ALPHA METHOD

W. J. T. Daniel

Department of Mechanical Engineering, University of Queensland,

St Lucia Campus, 4072, Brisbane, Queensland, Australia.
E-mail: [email protected]

ABSTRACT
Most finite element packages use the Newmark algorithm for time integration of structural dynamics.
Various algorithms have been proposed to better optimize the high frequency dissipation of this
algorithm. Hulbert and Chung proposed both implicit and explicit forms of the generalized alpha
method. However, their algorithms use incompatible parameter sets and cannot be used together in an
implicit/explicit spatial partition, whereas this can be done for the Newmark algorithm, as Hughes and
Liu demonstrated, and for the HHT-a algorithm developed from it. The present paper shows that the
explicit generalized alpha method can be rewritten so that it becomes compatible with the implicit
form. All four algorithmic parameters can be matched between the explicit and implicit forms. An
element interface between implicit and explicit partitions can be used, analogous to that devised by
Hughes and Liu to extend the Newmark method. The stability of the explicit/implicit algorithm is
examined and found to exceed that of the explicit partition. The element partition is significantly less
dissipative of intermediate frequencies than one using the HHT-a method.

KEYWORDS
Structural dynamics, finite element analysis, transient dynamics, partitioned methods, time integration.

INTRODUCTION
The generalized alpha method is a generalization of the Newmark method of time integration, widely
used for structural dynamics problems (Chung and Hulbert, 1993). The method uses two a
parameters, evaluating forces at one fraction c^of a cycle, and inertia at a different fraction am. It gives
an effectively optimized way of adding high frequency dissipation to a finite element transient
dynamic simulation of structural dynamics, at least for small deformation analyses. It has also been
adapted to problems with finite deformation and finite rotation updates, eg Kuhl and Crisfield (1999).
In addition to the implicit version, an explicit version is presented in Hulbert and Chung (1996), in
which they evaluate internal forces at the start of each cycle. This choice causes their explicit version
to use different Newmark parameters ß and γ to the implicit version, preventing its use in an
implicit/explicit mesh partition. Such spatial partitioning of integration is developed in Hughes and
Liu (1978), extending the Newmark method and the stability rigorously analysed. Hilber-Hughes-
Taylor (1977) reports on the HHT-a method, which can also be used with an implicit/explicit mesh
36

partition: see Hughes (1987). Partitioning in this way is efficient if a small region of a model is much
suffer than the rest.

In the present paper, it is shown that the explicit generalized alpha method can be rewritten in a way
that is fully compatible with the implicit version. The implicit and explicit generalized alpha methods
can then be coupled in an element partition. The parameter sets used (a/, am, ß and γ) can be matched
in each partition, corresponding to matching the spectral radius at bifurcation in the explicit partition
Pb to the limiting spectral radius poo of the implicit partition.

A CONSISTENT EXPLICIT FORM OF THE GENERALIZED ALPHA METHOD

Consider a predictor of new displacements u from velocities v and accelerations a

un+jP = un + v„ At +η* an At2 where η = 1/2 - β/(1 - α^. (1)

The implicit form of the generalized alpha method can be written using this new predictor *#„+/ to
obtain un+j.aP where un+j.ap = (1 - aß un+Jp + OfU„. The discrete equation of motion can either
(a) predict α^ = (l-am) an+¡ + aman using effective mass Mr or
(b) predict a correction Aun+i = u„+i - un+p using effective stiffness Κγ.
Using mass matrix M, a viscous damping matrix C, and stiffness matrix K, KT is the usual tangent
stiffness of the algorithm:

KT = {(1 - aJ/fßAt2)} M +{(1- af)Y/(ßAt)}C+ (1 - af)K (2)

MT = {ßAt2/(l -αηι)}Κτ = Μ+ {(1 - af)Y/(l - a^JCAt +{(J - af)ß/(l - a^jKAt2 (3)

The resulting discrete equation of motion, where FINT are elastic forces (eg -Ku) is

KTAun+J =MTaam= FEXT + FIN7{un+I.ap) + FDAMp(vn+1.aP) (4)

FEXT are external forces and FDAMP are viscous damping forces. Note the new predictor leads to a form
with no inertia term on the right hand side. This suggests that a rational explicit form can be obtained
simply by ignoring the contribution of K to Μτ, thus in the absence of damping, making the effective
mass the actual mass. It remains to determine the predictor of velocity vn+j.ap suitable for use in Eqn.
4. It may be shown that

Vaf = (l-af) vn+i + af vn = vn + (l-af)fl-//(l - aj} anAt + (1 - αβγ^Ι-α^α^Αί (5)

Hence the appropriate predictor, omitting the term in α^ , is

Vn+i-/ = vn + (1-α/){1-γ/(1 - a„)} anAt or vn+1p = vn + {1 - γ/(1 - aj) anAt (6)

Using the predictors in Equations 1 and 6, which divide the Newmark parameters by (1 - a^), making
them consistent with the terms of Mr, a new explicit version of the generalized alpha method is
obtained. The variation of spectral radius p with Courant number Ω = ωηΑί of the algorithm is shown
in Figure 1. This is identical to that of the algorithm by Hulbert and Chung, but the Newmark
parameters γ and β are now those of the implicit algorithm with its limiting spectral radius poo set so
that pb = poo. The effect on spectral radius of adding an explicit treatment of physical damping, using
the predictor in Eqn. 6, and maintaining different values of damping ratio, is shown in Figure 2.
 37

Figure 1: Spectral radius of the modified explicit generalized-oc method for various parameter settings.
A β=0.25, γ=0.5, af=0.5; B β=0.3543, γ=0.6905, af =0.4048; C β=0.3906, γ=0.75, ccf= 0.375;
D β=0.5102, γ=0.9286, a/=0.2857;E β=0.6944, γ=1.1667, af=0.1667;¥ β=1, γ=1.5, α/=0.0.

1.05

1
Ρ
^^Α
0.95

^ \ Β
0.9

\ \C
\ρ
0.85

0.8 0 0.5 1 1.5 2 25 3

Figure 2: Spectral radius of a single degree of freedom model problem with explicit treatment of
viscous damping added, for various damping ratios ζ. pb = 0.9.
A ζ= 0.025, B ζ=0.05, C ζ= 0.075, D ζ=0.1.

An explicit predictor, multi-corrector approach can be adopted by continuing to correct the predictors
of velocity and displacement using the new values calculated. This has the effect of making the
numerical dissipation in the explicit algorithm tend towards that of the implicit algorithm, although
rather slowly, as the stability limit is approached. As described later, this algorithm retains the stability
limit of the explicit algorithm.

RELATION TO THE ALGORITHM OF HULBERT AND CHUNG

The new algorithm can be regarded as the result of a similarity transformation on the amplification
matrix of the algorithm of Hulbert and Chung, their displacement and velocity values (u, v, a) being
reinterpreted as those at fraction c^of a cycle

Let dn = T un or explicitly
1 \-af
\-a
1
Γ νU nΔί = ν„Δί
η
"„
(7)
1 L a » A ' 2 . _a At\ n
38

The amplification problem for Hulbert and Chung's algorithm can be written in the form Adn+i = Bdn.
If for some mode of the amplification problem, Αα=λΒα where λ is an eigenvalue, then ΑΤιι=λΒΤιι.
Hence the revised algorithm retains the same eigenvalues of its amplification matrix, and hence
identical dissipative properties. The value of η needed to obtain the implicit β parameter in the
resulting algorithm, is consistent with η in equation 1:
1/ ß Λ_ Pb<\-Ph)
/ l
\-am 2(l + pb)2(2-pb)

Another possibility in Eqn. 7 is to set η = 0 and a/ = *Λ, leading to an explicit algorithm which
dissipates high frequencies, while maintaining evaluation offerees at the predicted midpoint of a
cycle. In terms of Hulbert and Chung's explicit Newmark parameters ^ a n d βΕ, the values required
are γ = γΕ~ 1/2 and β = ΡΕ- γ/2. am is unchanged. This algorithm suits analysis of finite deformation.

AN ELEMENT PARTITION
Consider a model partitioned into explicit (E) elements and implicit (I) elements. At interface nodes
sharing E and I elements, we add internal forces on the interface. Let M include the total mass of
interface nodes and FINT and FEXT be the internal and external forces on one side of the interface. M
must be found from lumped element mass matrices on the explicit side. On the explicit side, to
evaluate the internal forces, displacements and velocities at fraction a/ of a cycle are predicted at
explicit and interface nodes using the expressions for un+p and v„+/P in Eqns. 1 and 6 above. This
enables new accelerations to be found at explicit nodes (not including interface nodes) using

an+1E = 1/(1 - aj {ME1 (FINTE(un+1J, vn+1.ap) + FExf) - am aEj (9)

Let M again include the total mass of interface nodes. At interface nodes

MT dam = FEXT + FINT(un+1.ap) + FDAMp(vn+i-aP) + FINTE(un+i.aP) (10)

Nodes fully in the implicit partition are updated the same way, but without the Fm/ loading from
explicit elements. The Newmark formulae can then be used to recover velocities and accelerations
from the displacements, if these are found first, or to find velocities and displacements from
accelerations.

ANALYSIS OF THE ELEMENT INTERFACE

The element interface has been studied using model problems. As, for the partitioned algorithm, modal
analysis cannot be used to relate a model problem to the general case, stability is also studied by
analysis of energy changes. The simplest model problem is one representing a single interface degree
of freedom. The equations can be written using Courant number parameters Q¿ and Ω/2 for each
partition, that is: (KE/M) At2 and (K/M) At2. Let ίί = ΩΕ2 + Ω2.

The spectral radius of the coupled algorithm, evaluated from the single degree of freedom interface, is
plotted versus the Courant number of the actual problem solved (ΩΕ, Ω/ or Ω) in Figure 3 for pb = p*»
= 0.9. A comparison is made to that of the explicit and implicit partitions, with Ω/ΩΕ = 2 held
constant as ΩΕ is varied. In general, the implicit/explicit algorithm shows a bifurcation at a timestep
size very slightly larger than that of the explicit partition. The explicit bifurcation occurs at

ΩΕ = (1+ρ*)ν(2-ρ0) (11)
 39

. ΝΛ ~c

Β
A -
D\

ΩΕΟΓ Ωι or Ω

Figure 3: Spectral radius of the single interface degree of freedom model problem.
Ωι = 2ΩΕ. pb = 0.9. A - gen-ot element partition. B - gen-ot explicit element only.
C - gen-oc implicit element only. D - HHT-a element partition

The dissipation of the partitioned problem can be seen to be intermediate between that of the explicit
and implicit partitions. Also plotted on Figures 3 is the spectral radius an implicit/explicit element
partition of the HHT-a method, using the same poo setting of 0.9. It is seen to cause much more
dissipation of intermediate frequencies.

The setting pb= p00 = 0, is appropriate where the highest frequencies are excited. An example of using
this setting on the mesh of Figure 4 is shown in Figure 4(a). Figure 4(a) shows the displacement in
response to an impulse on the left end. The timestep of 0.5 s is close to the explicit stability limit. The
elements all have the same material properties (Young's modulus of 1 and density of 2), and a
consistent mass matrix was used in the implicit partition. The initial condition chosen maximizes the
high frequency response. The partitioned algorithm is quite stable, and the highest frequencies can be
seen to be rapidly filtered from the solution. For contrast, the response to the pb= pao= 0.99 setting is
given as well in Figure 4(b), showing the high frequency noise on the solution.
EXPLICIT | IMPLICIT | EXPLICIT
10 ELEMENTS 10 ELEMENTS

ίφ
10 ELEMENTS
E = 1,A=1,DENSITY = 2

η
μr w
Time (s) Time
(s)

Figure 4: Axial displacement at the left end of the rod mesh shown, (a) pb=0. (b) pb=0.99.
40

Stability analysis from an energy perspective

A stability analysis can be preformed from an energy perspective to compare the partitioned algorithm
to a case where the stiffness terms of the implicit partition are omitted. The stability of this case is
controlled by the highest natural frequency of the explicit partition. Introducing the stiffness terms K¡
of the implicit partition adds a term -(1/2 - aß [u]TK¡[u] to the change in energy over a cycle,
representing a loss of energy or gain in stability, where [u] is the change in displacement occurring.

CONCLUSIONS
The explicit generalized alpha method can be rewritten in a form that is fully compatible with the
implicit version, having all parameters equal when pb = poo- This permits development of a spatially
partitioned explicit/implicit integrator. An element interface is the simplest possibility, and exceeds
the stability of the explicit partition. The partitioned algorithm is significantly less dissipative than an
explicit/implicit partition of the alternative HHT-a method. The consistent parameters also enable a
predictor/multi-corrector implementation of the explicit algorithm, in which the predictions of un+j.a
and vn+1-a are replaced by new estimates of the weighted averages, although this becomes inefficient,
as the stability limit is approached. An inconsistent explicit form, achieving high frequency
dissipation, while maintaining midpoint evaluation of loading, is also possible. This is advantageous
in finite deformation analyses where midpoint evaluation of loading is crucial to accuracy.

References
Chung, J. and Hulbert, G.M. (1993). A time integration method for structural dynamics with improved
numerical dissipation: the Generalized-oc Method, Journal of Applied Mechanics 60, 371-375.

Hilber, H.M., Hughes, T.J.R. and Taylor (1977), R.L. Improved numerical dissipation for time
integration algorithms in structural dynamics, Earthquake Engineering and Structural Dynamics 5,
283-292.

Hughes, T.J.R. and Liu, W.K. (1978), Implicit-explicit finite elements in transient analysis: stability
theory, ASME Journal of Applied Mechanics 45, 371-374.

Hughes, T.J.R. and Liu, W.K. (1978), Implicit-explicit finite elements in transient analysis:
implementation and numerical examples, ASME Journal of Applied Mechanics 45, 375-378.

Hughes, T.J.R. (1987), The finite element method: linear static and dynamic finite element analysis
Prentice-Hall, Englewood Cliffs, New Jersey.

Hulbert, G.M. and Chung, J. (1996), Explicit time integration algorithms for structural dynamics with
optimal numerical dissipation, Comput. Methods Appl Mech. Engrg. 137, 175-188.

D. Kuhl, D. and Crisfield, M.A. (1999),Energy-conserving and decaying algorithms in non-linear

structural dynamics, Int. Jnl for Num. Meth. Engng.,45, 569-599.