You are on page 1of 11

FINITE ELEMENTS

IN ANALYSIS
A N D DESIGN
ELSEVIER Finite Elements in Analysis and Design 24 (1997) 271-281

A new component mode synthesis method:


Quasi-static mode compensation
Wen-Hwa Shyu, Zheng-Dong Ma, Gregory M. Hulbert*
Department of Mechanical Engineering and Applied Mechanics, 2250 G.G. Brown, The University of Michigan, Ann Arbor, MI
48109-2125, USA

Abstract

A new component mode synthesis method is presented in this paper that combines the computational efficiency of the
well-known constraint mode approach with the dynamic compensation accuracy obtained by higher-order expansion
methods. Instead of employing static constraint modes, quasi-static modes are used to capture inertial effects of the
truncated modes. The method is ideally suited for mid-band frequency analysis in which both high-frequency and
low-frequency modes may be omitted. A tuning parameter, designated as the centering frequency, controls the dynamic
range of the quasi-static modes. Numerical examples are provided which demonstrate the improved accuracy of the
proposed method.

Keywords: Structural dynamics; Component mode synthesis; Modal analysis

1. Introduction

Component modes synthesis (CMS) is a well established method for efficiently constructing
models to analyze the dynamics of large, complex structures that are often described by separate
substructure (or component) models. Typically, each substructure is approximated by a set of basis
vectors (Ritz vectors), where the number of vectors is substantially smaller than the number of
physical degrees of freedom (dof). The substructure approximations are then assembled to provide
a global approximation of the structure. CMS is often used to greatly reduce the computational
size (and cost) of large, complex structures for a given level of accuracy. CMS is also employed in
design environments in which the components of structural systems are designed and analyzed
separately.
Numerous CMS methods have been developed since its genesis in the early 1960s. In the first era
of CMS development, of principal concern was the approach taken to connect components, that is,

* Corresponding author. Tel.: (313) 763-4456; fax: (313)647-3170.

0168-874X/97/$17.00 1997 Elsevier Science B.V. All rights reserved


PII S0 1 6 8 - 8 7 4 X ( 9 6 ) 0 0 0 6 6 - 2
272 W.-H. Shyu et al./ Finite Elements in Analysis and Design 24 (1997) 271 281

how to synthesize the equations of the combined system. In addition to rigid-body modes and
normal modes (using fixed boundary conditions) for the components, Hurty first proposed an
interface mode, using the constraint mode, to satisfy displacement compatibility [1]. Subsequent
methods were developed and classified according to the type of boundary conditions used to
calculate the component normal modes and the type of interface mode employed, see, e.g., [1-11].
The work of Goldman [3] and Hou [4] marked a change in the CMS development community.
They proposed a free interface method, in which only the component normal modes (free boundary
conditions) are included. In this way, the components could be combined without the need for
interface modes. Subsequently, many improved free interface methods were developed, see, e.g.,
[5-10]. However, from the viewpoint of mode superposition theory, the free interface method of
Goldman and Hou is a regression as their method usually produces less accurate results than the
methods of Hurty and Craig.
Two important milestones in CMS methodology are marked by the papers of MacNeal [6] and
Rubin [7] and by Hintz [8]. MacNeal noted that a static approximation for the truncated
high-frequency modes can be used to improve accuracy. Rubin related this improvement to the
mode-acceleration method. Hintz showed that static completeness of the interface modes is critical
and defined a general category of statically complete interface modes. Lu and Ma [11] extended
this concept to static completeness of the Ritz basis for any type of component approximation and
proposed a general methodology for obtaining such a basis from a set of given initial Ritz vectors.
Lu and Ma showed that the commonly used CMS methods can be recovered as special cases of
their method. At this state of CMS development, it became clear that any type of boundary
condition may be used to define the component normal modes provided that these modes are
concatenated with static modes that satisfy the static completeness requirement. Provided that the
Ritz basis vectors span the same configuration space, the results obtained using different static
modes are identical. Thus, CMS methods can be characterized as methods that do or do not satisfy
the static completeness condition. Choosing an appropriate reduced set of basis vectors continues
to be an area of research, see, e.g., [12].
While the original focus of CMS methods was to solve problems restricted to low-frequency
ranges, much research has been conducted to develop CMS methods for mid-frequency or high-
frequency problems. Kuhar and Stahle [13] introduced a dynamic transformation (in contrast to
the static transformation method of Guyan) to further reduce the size of the assembled system
problem. Kubomura [14, 15] developed first-order and second-order approximation methods to
treat both low-frequency and high-frequency problems. Yee and Tsuei [16, 17] presented a series of
papers introducing a model force equation in which the system equation can be reduced to include
only boundary dof. Suarez et al. [18, 19] developed alternative higher-order expansion methods.
Min et al. [20] developed a frequency-window technique in which analytic expressions for
the impedance and mobility matrices of coupled substructures are expanded about a central
frequency value. However, these methods have one or more of the following drawbacks: (a) a
nonlinear equation in frequency must be solved by iterative methods; (2) different formulations are
employed for the low-frequency and high-frequency problems; (3) the number of interface dof is
increased.
CMS methods may be considered to have originated from mode-superposition theory, e.g,
Goldman and Hou simply employed the mode-displacement method to develop the first free
interface method. The mode acceleration method has been used extensively by many CMS
W.-H, Shyu et al./ Finite Elements in Analysis and Desion 24 (1997) 271-281 273

developers. Indeed, it can be shown that a necessary and sufficient condition for obtaining
a statically complete Ritz basis is to use a mode acceleration method, i.e., static completeness is
equivalent to the mode acceleration method. However, as noted by Ma and Hagiwara [21], while
the mode acceleration method corrects for the inaccuracy caused by truncating high-frequency
modes, accuracy can decrease, compared to the mode displacement method, if low-frequency
modes are truncated. Thus, no viable CMS method exists for analyzing mid-frequency or high-
frequency problems.
In this paper, a new CMS method is proposed that is based upon the recently developed mode
superposition technique of Ma and Hagiwara [21]. The new method extends the static compensa-
tion technique used in the mode acceleration method via the use of quasi-static compensation. The
new class of "quasi-static modes" (QSM) are suitable for dynamic analysis associated with any
bounded frequency range. The new method possesses the following attributes: (1) Substantially
more accurate results can be obtained with the same number of dof of the synthesized system,
compared to standard CMS methods; (2) Conversely, for a given error tolerance, computational
cost is reduced; (3) For banded frequency analysis, we can omit both high-frequency modes and
low-frequency modes outside the frequency range of interest; (4) Standard constraint mode
formulations are obtained directly; (5) No additional implementation costs are incurred, compared
to the constraint mode formulation.

2. Formulation

In the following, subscripts define particular vectors and matrices and simultaneously define the
dimensionality of the vectors and matrices. For ease of exposition, we will consider the eigenvalue
analysis of an undamped structure. For each component of the structure, the displacement vector,
U, is partitioned into a vector (of length a) of the so-called active dof Ua and Uo, which is the vector
(of length o) of omitted dof (a + o = component's total number of unconstrained dof, neq).
Associated with each component of the structure is the eigenproblem

Moa
(1)

in which K and M denote the respective stiffness and mass matrices, (09, ~) denotes the eigen-
value-eigenvector pair, and ~ is the vector of constraint forces, which depends upon the type of
boundary conditions applied to the active dof. For convenience, in the examples, we have chosen to
use fixed boundary conditions for the active dof. Let ~ , denote the set of n retained normal modes
from (1), where, for efficiency n ,~ neq. For commonly used CMS methods, accuracy requirements
typically dictate that ~ , consists of the first n normal modes. As will be shown in the numerical
examples, this restriction can be relaxed when using QSM compensation.
To enhance accuracy, computational modes, denoted by ~, are appended to the set of normal
modes. Using the classical constraint modes (CM) approach,

,--{ - K1K"
},
(2)
274 W.-H. Shyu et al. / Finite Elements in Analysis and Design 24 (1997) 271-281

where l,a is the identity matrix. These constraint modes emanate from expressing the Xo partition of
the static response vector, X, in terms of)Ca by solving the static reduction problem:


(31

The quasi-static modes are defined by


, e = ( K o{ - ~ o2 c M o o ) (Ko,,-~,2Moa, t
L, ' (4)

where cot is a frequency tuning parameter, denoted as the centering frequency. Analogous to the
derivation of constraint modes, the quasi-static modes are derived by expressing the Xo partition of
the quasti-static response vector X, in terms of X,, by solving the quasi-static reduction problem:
MO~ Xo

Comparing (3) and (5), it is clear that static CM are a special case of QSM, obtained by setting
oc = 0. Without incurring additional expense over the CM technique, the QSM approach is able
to include inertial effects of the omitted dof.
Concatenating the QSM and the normal modes results in the transformation:

U= U, = ~,, @,, ( Q , J /,,, q~,, , (6)

where Qa and Q. are vectors of generalized coordinates associated with the QSM and the normal
modes, respectively. To facilitate coupling the components, (6) is transformed into
U = TZ (7)
in which

T= 1~ 0.. = 1,,. 0,,. (8)

and

In the above, 0a, denotes the a x n null matrix. Now, in the usual manner, the component reduced
mass and stiffness matrices are computed using
IU m~ = TTKT, A1~m~ = TTMT. (10)
where the superposed T indicates matrix transpose.
After assembling all the component reduced mass and stiffness matrices, we obtain the un-
coupled eigenproblem:
(/ - cozM)Z =/~, (11)
W.-H. Shyu et al./ Finite Elements in Analysis and Design 24 (1997) 271-281 275

where 2fir, K and P are the assembled, uncoupled mass matrix, stiffness matrix and force vector,
respectively; 2 is the vector comprising all active and generalized dof in which the interface
boundary dof are repeated. Following [22], the coupled, reduced system may be obtained through
a suitable coupling matrix, S, which incorporates the compatibility conditions, such that
= S ~ red (12)

where ~red is an eigenvector of the assembled, reduced system. Thus,


(~red __ ( Z ) 2 ~ e d ) ~ r e d = 0 (13)

in which
~1 r*d = SX 2f-lS, g tea = SXffS. (14)

3. Numerical examples

Two examples are given to demonstrate the performance of the new CMS method presented in
the previous section. In both examples, the reference solution was computed directly from the
assembled components. Comparison is made between the use of QSM and CM. For all calcu-
lations, MSC/NASTRAN [23] was employed; a DMAP was written to implement QSM. All
component normal modes included were computed using fixed boundary conditions for the active
dof. Comparison of eigenvectors is achieved using Modal Assurance Criterion (MAC) values, in
which

MAC(i, j) = q i)2 (15)

3.1. Rectangular plate

As shown in Fig. 1, a flat rectangular, clamped--clamped plate was discretized using 576 shell
elements. The plate was divided vertically into two equal-size components consisting of 288
elements. The active dof were chosen as those on the vertical symmetry boundary.
For purposes of exposition, two frequency bands were investigated. The first frequency band is
the low-pass band at 400 Hz. In this band, each component has two normal modes that were
included in the transformation matrix. Fig. 2 shows the eigenvalue errors as a function of the
choice of centering frequency (recall that setting the centering frequency to zero recovers the CM
formulation). Each vertical grouping of markers denotes a specific system reference eigenvalue; the
lines are added for the sake of clarity. In the remainder, QSM i denotes that i Hz was used as the
centering frequency. From Fig. 2, it is clear that the eigenvalue accuracy is dependent on the choice
of centering frequency; a more accurate eigenvalue is obtained when the centering frequency is
close to the eigenvalue. As the centering frequency is increased, errors in the eigenvalues below the
centering frequency generally increase. Conversely, errors in eigenvalues above the centering
frequency are always reduced when the centering frequency is increased. If the low-frequency
276 W.-H. Shyu et al./ Finite Elements in Analysis and Design 24 (1997) 271 281

I0001 '

100 t ~", tcnstraint

i ' " , ..........~ ' " , S T+ QSM 200 -~


?Ill IIIIIIIII III IIIIIII
qlllllll IIIIIIIII Ill IIIllll IIIll 17/+
?If[Jill| IIIl[]lll Ill Illllll IIlll
!llll llll IIIIIIIII III lllllll lllll
:IIIIIIII lllllllll Ill IIII111 lllll If/ I / * ' ....... k:-~ ...... t
'III11111 111111111 llr III1111 Iflll
:llllI[ll IIIIIIIII III lllllll IIIII 17/
so l o o 15o 200 2so aoo aso 400
xllll Illll[l[I Ill Illll[I
4111llll llllfllll III IIIIIII II[[I
~equency (Hz)

Fig. 1. Finite element mesh of rectangular plate I-ig. 2. Rectangular plate problem. Eigenvalue errors, in
problem. percent, for CMS-derived system modes in 0-400 Hz
band.

Table 1
Rectangular plate problem. MAC values of CMS-derived system modes using QSM approach with
different centering frequencies

Method Mode

1 2 3 4 5 6

CM 1.0 1.0 1.0 1.0 0.996 0.997


QSM 50 1.0 1.0 1.0 1.0 0.996 0.997
QSM 1O0 1.0 1.0 1.0 1.0 0.996 0.997
QSM 200 0.999 1.0 I.() 1.0 0.996 0.997
QSM 300 0.997 0.999 0.999 1.0 1.0 1.0
QSM 400 0.981 0.996 0.998 1.0 0.991 1.0

eigenvalues are of principal concern, the CM method can always be improved by choosing
a centering frequency just below the first system frequency.
Table 1 lists the MAC values for the system modes obtained using different centering frequencies
with respect to the first six system reference normal modes. Note that the errors (deviation from
unity) are substantially smaller than the eigenvalue errors. Similar to the trend in eigenvalue error
with centering frequency, the high-frequency MAC values improve with increased centering
frequency, while the low-frequency MAC values deteriorate.
The second frequency band was chosen as a mid-frequency band from 600 to 800 Hz. In this
band, each component possesses three normal modes. For the QSM approach, these modes were
included in the transformation matrix. Standard practice using CM employs (at least) all compon-
ent normal modes from 0 Hz to the high-frequency cutoff. For this example, the standard CM
formulation (CM standard) used five normal modes per component. To provide a fair comparison
of accuracy for a fixed cost, we also examined using only the three normal modes in the frequency
W.-H. Shyu et al./ Finite Elements in Analysis and Design 24 (1997) 271-281 277

Table 2
Rectangular plate problem. Eigenvalue errors, in percent, for CMS-derived system modes in the
600-800 Hz band (reference values in parentheses)

Method Mode 10 Mode 11 Mode 12 Mode 13


(651.49 Hz) (693.75 Hz) (759.88 Hz) (783.20 Hz)

QSM 700 0.01 0 0.15 4.87


CM banded 27.9 0.1 8.75 x
CM standard 0.3 0.08 4.21

Table 3
Rectangular plate problem. MAC values for CMS-derived system modes in the 600-800 Hz band

Method Mode 10 Mode 11 Mode 12 Mode 13

QSM 700 1.0 1.0 0.999 0.993


CM banded 0.862 0.997 0.831 x
CM standard 0.999 1.0 0.989

band of interest in the CM transformation matrix (CM banded). In this example, the centering
frequency was chosen at the center of the frequency band, i.e., 700 Hz (QSM 700). Tables 2 and
3 list the eigenvalue errors and the MAC values for the system normal modes between 600 and
800 Hz. Improvements in both the eigenvalues and eigenvectors are clear when using the QSM
method. For both CM methods, the 13th mode was not found in the extended frequency search
band from 600-950 Hz (denoted by x in the table). As has been noted by others, including all
component normal modes below the high-frequency cutoff is necessary. Note, however, the
accuracy of the QSM formulation is still greater despite its reduced cost, in terms of the number of
generalized coordinates (as well as the cost of extracting the eigenpairs).

3.2. Automotive lower control arm (LCA)

Fig. 3 shows the top view of a finite element model of an automotive lower control arm (LCA).
The model consists of 1429 elements (shells, beams and rigid elements) and 1338 nodes. For the
sake of analysis the dof at the two bushing centers and at the ball joint are fixed. The LCA was
divided into four components, as shown in Fig. 4. The 282 dof of the 50 component boundary
nodes were chosen as the active dof, representing a substantial reduction from the 8010 dof of the
complete model. As in the previous example, two frequency ranges were studied: 0-1000 Hz and
6500-7500 Hz. In the low-frequency band, there are no normal modes for any of the components;
in the mid-frequency band, there are 1, 2, 1, and 2 component normal modes for components 1-4,
respectively. Due to the complexity of the LCA normal modes, particularly, in the high-frequency
band, the MAC values were used to distinguish the different modes.
278 W.-H. Shyu et al./ Finite Elements in Analysis and Design 24 (1997) 271 281

Ball Joint

~ - Front Bushing

RearBushing

Fig. 3. Finite element mesh of LCA problem.

Component 1

Component 2

C m p ~ ~;

Component 3

Fig. 4. Decomposition of LCA finite element model.

Tables 4 and 5 list the eigenvalue errors and the MAC values obtained for the two system
normal modes in the 0-1000 Hz range, for the QSM formulation, using a centering frequency of
500 Hz and for the CM formulation. While a modest improvement in eigenvalue is observed, there
is a substantial improvement in the mode 2 MAC value.
Table 6 lists the eigenvalue errors for the five system normal modes between 6500-7500 Hz
obtained using a centering frequency of 7000 Hz, the standard CM formulation that includes all
component normal modes below 7500 Hz (2, 8, 1, 5, for the four respective components), and the
CM formulation including only the component normal modes between 6500-7500 Hz. As ex-
pected, this last approach is entirely inadequate since it is unable to provide any of the system
W.-H. Shyu et al./ Finite Elements in Analysis and Design 24 (1997) 271-281 279

Table 4
LCA problem. Eigenvalue errors, in present, for CMS-derived system modes in 0-1000 Hz
band (reference values in parantheses)

Method Mode 1 Mode 2


(474.90 Hz) (931.57 Hz)

QSM 500 0.02 11.1


CM 1.39 13.8

Table 5
LCA problem. MAC values for CMS-derived system modes in 0-1000 Hz band

Method Mode 1 Mode 2

QSM 500 1.0 0.652


CM 0.999 0.450

Table 6
LCA problem. Eigenvalue errors, in percent, for CMS-derived system modes in the 6500-7500 Hz
band (reference values in parantheses)

Method Mode 25 Mode 26 Mode 27 Mode 28 Mode 29


(6536.3 Hz) (6921.4Hz) (6985.0Hz) (7227.1Hz) (7281.2Hz)

QSM 7000 0.72 0.01 0 0.19 0.32


CM standard 8.95 7.78 12,01 10.29 x
CM banded x x x

Table 7
LCA problem, MAC values for CMS-derived system modes in 6500-7500 Hz band

Method Mode 25 Mode 26 Mode 27 Mode 28 Mode 29

QSM 7000 0.979 1.0 1.00 0.995 0.994


CM standard 0.859 0.907 0.816 0.716 x

n o r m a l m o d e s in the 6 5 0 0 - 7 5 0 0 H z b a n d . I n c l u d i n g all c o m p o n e n t n o r m a l m o d e s p r o v i d e s
a s u b s t a n t i a l i m p r o v m e n t ; h o w e v e r , the last s y s t e m n o r m a l m o d e was n o t f o u n d b e l o w a eigenvalue
s e a r c h f r e q u e n c y o f 9000 Hz. Finally, the excellent a c c u r a c y of the Q S M a p p r o a c h is n o t e d . T a b l e 7
lists the M A C values for the 6 5 0 0 - 7 5 0 0 H z system m o d e s c o m p a r e d to the s y s t e m m o d e s o b t a i n e d
using the Q S M a p p r o a c h a n d the s t a n d a r d C M f o r m u l a t i o n ; the increased a c c u r a c y of the
e i g e n v e c t o r s using Q S M m e t h o d is clear.
280 W.-H. Shyu et al. / Finite Elements in Analysis and Design 24 (1997) 271 281

4. Conclusions

A new component mode synthesis method has been proposed that combines the computational
efficiency of the standard constraint mode (CM) formulation with the high accuracy typically seen
in higher-order (and more expensive) expansion methods. Instead of the static constraint mode,
a quasi-static compensation, or quasi-static mode (QSM), is employed which depends on the value
of a centering frequency parameter. Efficiency is gained by requiring that only the component
normal modes in the frequency band of interest be included in the transformation matrix.
Numerical examples were given which demonstrate the performance improvements of the QSM
method over the C M approach in terms of computing eigenvalues and eigenvectors. In addition,
the results from the automotive lower control arm example suggest that stress computations
performed using the QSM approach should improve greatly upon the values obtained using the
C M method.
Questions unanswered in this paper include how to choose the centering frequency and whether
more than one centering frequency should be employed. Results were presented which showed that
an improper choice of centering frequency, particularly for low frequency response, can actually
decrease accuracy when compared to the C M approach. Current efforts are directed at providing
guidelines for selecting the centering frequency and will be the subject of a subsequent paper. In
addition, the new method suggests an efficient, iterative approach towards obtaining particular
system eigenvalues of interest. Thus, it may provide a numerically efficient alternative to the
eigenvalue iteration method of Min, Igusa and Achenbach.

Acknowledgements

The authors gratefully acknowledge the support of Ford M o t o r C o m p a n y for this work through
the Center for Automotive Structural Durability Simulation at the University of Michigan.

References

[1] W.C. Hurty, "Dynamic analysis of structural systems using component modes", A I A A J. 3, pp. 678-685, 1965.
[2] R.R. Craig, Jr. and M.C.C. Bampton, "Coupling of substructures for dynamic analysis",A I A A J. 6, pp. 1313-1319,
1968.
[3] Ri. Goldman, ~'Vibration analysis by dynamic partitioning", A I A A J. 7, pp. 1152 1154, 1969.
[4] S.N. Hou, "Reviewof modal synthesistechniques and a new approach", Shock Vib. Bull. Part 4, 40, pp. 25-39, 1969.
[5] E.H. Dowell, "Free vibrations of an arbitrary structure in terms of component modes", J. Appl, Mech. 39, pp,
727-732, 1972.
[6] R.H. MacNeal, "'A hybrid of component mode synthesis", Comput. Struct. I, pp. 581-601, 1971.
[7] S. Rubin, "Improved component mode representation", A I A A J. 13, pp. 995-1006, 1975.
[8] R.M. Hintz, "Analytical methods in component modes synthesis", A I A A J. 13, pp. 1007-1016, 1975.
[9] R.R. Craig, Jr. and C.J. Chang, "On the use of attachment modes in substructure coupling for dynamic analysis",
A I A A / A S M E 18th Structures, Dynamics and Materials Conf., Paper 77-405, 1977.
[10] R.R. Craig, Jr. and A.L. Hale, "Block-Krylov component synthesis method for structural model reduction", J.
Guidance, Control Dyn. 11, pp. 562-570, 1988.
W.-H. Shyu et al./ Finite Elements in Analysis and Design 24 (1997) 271-281 281

[11] Y.F. Lu and Z.D. Ma, 'A new component modal synthesis method- with higher accuracy, computational efficiency,
synthesis flexibility and adaptability", Proc. 3rd Internat. Modal Analysis Conf. Orlando, FL, USA, 291-298, 1985.
[12] L. Meirovitch and M.K. Kwak, "Rayleigh-Ritz substructure synthesis for flexible multibody systems", A I A A J. 29,
pp. 1709-1719, 1991.
[13] E.J. Kuhar and C.V. Stahle, "Dynamic transformation method for modal synthesis", A I A A J. 12, pp. 672-678,
1974.
[14] K. Kubomura, "A theory of substructure modal synthesis", J. Appl. Mech. 49, pp. 903-909, 1982.
[15] K. Kubomura, "Component mode synthesis for damped structures", A I A A J. 25, pp. 740-745, 1987.
[16] E.K,L. Yee and Y.G. Tsuei, "Direct component modal synthesis technique for system dynamic analysis", A I A A J.
2% pp. 1083-1088, 1989.
[17] E.K.L. Yee and Y.G. Tsuei ``Imprved methd fr the atera vibratin anaysis f a rbt system J. Prpulsion 6
pp. 165-170, 1990.
1-18] L.E. Suarez and M.P. Singh, "Improved fixed interface method for modal synthesis", A I A A J. 30, pp. 2952-2958,
1992.
[19] L.E. SuarezandE.E Matheu, "A modal synthesis technique based on the force derivative method", J. Vib. Acoustics
114, pp. 209-216, 1992.
[20] K.W. Min, T. Igusa and J.D. Achenbach, "Frequency window method for strongly coupled and multiply connected
structural systems", A S M E J. Appl. Mech. 59, pp. 236-252, 1992.
[21] Z.D. Ma and I. Hagiwara, "Improved mode-superposition technique for modal frequency response analysis of
coupled acoustic-structural systems", A I A A J. 29, 1720-1726, 1991.
[22] R.R. Craig, Jr., Structural Dynamics - An Introduction to Computer Methods, Wiley, New York, 1981.
[23] MSC/NASTRAN User's Manual, Vols. 1 and 2, The MacNeal/Schwendler Corporation, Los Angeles, CA, 1991.

You might also like