On an Implementation of the

Hilber-Hughes-Taylor Method
Dan Negrut1
Department of Mechanical Engineering,
in the Context of Index 3
The University of Wisconsin,
Madison, WI 53706
e-mail: [email protected]
Differential-Algebraic Equations
of Multibody Dynamics
Rajiv Rampalli
e-mail: [email protected] „DETC2005-85096…
Gisli Ottarsson The paper presents theoretical and implementation aspects related to a numerical inte-
e-mail: [email protected]
grator used for the simulation of large mechanical systems with flexible bodies and
contact/impact. The proposed algorithm is based on the Hilber-Hughes-Taylor (HHT)
Anthony Sajdak implicit method and is tailored to answer the challenges posed by the numerical solution
e-mail: [email protected]
of index 3 differential-algebraic equations that govern the time evolution of a multibody
MSC.Software,
system. One of the salient attributes of the algorithm is the good conditioning of the
Ann Arbor, MI 48105
Jacobian matrix associated with the implicit integrator. Error estimation, integration
step-size control, and nonlinear system stopping criteria are discussed in detail. Simula-
tions using the proposed algorithm of an engine model, a model with contacts, and a
model with flexible bodies indicate a 2 to 3 speedup factor when compared against
benchmark MSC.ADAMS runs. The proposed HHT-based algorithm has been released in
the 2005 version of the MSC.ADAMS/Solver. 关DOI: 10.1115/1.2389231兴

Introduction Mq̈n+1 + Cq̇n+1 + Kqn+1 = Fn+1 共2c兲

The Hilber-Hughes-Taylor Method: General Considera- Based on Eqs. 共2a兲 and 共2b兲, qn+1 and q̇n+1 are functions of the
tions. The Hilber-Hughes-Taylor 共HHT兲 method 共also known as acceleration q̈n+1, which in Eq. 共2c兲 remains the sole unknown
the alpha-method兲 关1兴 is widely used in the structural dynamics quantity that is obtained as the solution of a linear system. This
community for the numerical integration of a linear set of second method is implicit and A-stable 共stable in the whole left-hand
ordinary differential equations 共ODEs兲. This problem is obtained plane兲 关3兴 provided 关1兴
at the end of a finite element discretization. Provided the finite
1 共␥ + 1/2兲2
element approach is linear, the equations of motion assume the ␥艌 ␤艌 共3兲
form 2 4
Mq̈ + Cq̇ + Kq = F共t兲 共1兲 The only combination of ␤ and ␥ that leads to a second-order
integration formula is ␥ = 1 / 2 and ␤ = 1 / 4. This choice of param-
The p ⫻ p mass, damping, and stiffness matrices, M, C, and K, eters produces the trapezoidal method, which is both A-stable and
respectively, are constant, the force F 苸 R p depends on time t, and second order. The drawback of the trapezoidal formula is that it
q 苸 R p is the set of generalized coordinates used to represent the does not induce any numerical damping in the solution, which
configuration of the system. makes it impractical for problems that have high-frequency oscil-
A precursor of the HHT method is the Newmark method 关2兴, in lations that are of no interest or parasitic high-frequency oscilla-
which a family of integration formulas that depend on two param- tions that are a byproduct of the finite element discretization pro-
eters ␤ and ␥ is defined: cess 关4,5兴. Thus, the major drawback of the Newmark family of
integrators was that it could not provide a formula that was
A-stable and second order and displayed a desirable level of nu-
h2 merical damping. The HHT method came as an improvement be-
qn+1 = qn + hq̇n + 关共1 − 2␤兲q̈n + 2␤q̈n+1兴 共2a兲
2 cause it preserved the A-stability and numerical damping proper-
ties, while achieving second-order accuracy when used in
conjunction with the secondorder linear ODE problem of Eq. 共1兲.
q̇n+1 = q̇n + h关共1 − ␥兲q̈n + ␥q̈n+1兴 共2b兲 The idea proposed in 关1兴 actually does not pertain the expression
of the Newmark integration formulas, but rather the form of the
These formulas are used to discretize at time tn+1 the equations of discretized equations of motion in 共2c兲. The new equation in
motion 共1兲 using an integration step size h: which the integration formulas of Eqs. 共2a兲 and 共2b兲 are substi-
tuted is
1
Corresponding author. Mq̈n+1 + 共1 + ␣兲Cq̇n+1 − ␣Cq̇n + 共1 + ␣兲Kqn+1 − ␣Kqn = F共t̃n+1兲
Contributed by the Design Engineering Division of ASME for publication in the
JOURNAL OF COMPUTATIONAL AND NONLINEAR DYNAMICS. Manuscript received April 1, 共4兲
2006; final manuscript received July 5, 2006. Review conducted by Hamid M. Lan-
karani. Paper presented at the ASME 2005 Design Engineering Technical Confer- where
ences and Computers and Information in Engineering Conference 共DETC2005兲, Sep-
tember 24–28, 2005, Long Beach, California, USA. t̃n+1 = tn + 共1 + ␣兲h 共5兲

Journal of Computational and Nonlinear Dynamics JANUARY 2007, Vol. 2 / 73

Copyright © 2007 by ASME

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 As indicated in 关4兴, the HHT method will possess the advertised is X, its time derivative is Ẋ, and the assumption is that, in
stability and order properties provided ␣ 苸 关−1 / 3 , 0兴 and
its implicit form, Eq. 共8a兲 properly and uniquely defines Ẋ
1 − 2␣ 共1 − ␣兲2 as a function of the state of the system through the user-
␥= ␤= 共6兲 specified function d. If nd represents the number of differ-
2 4
ential states, then X , d 苸 Rnd.
The smaller the value of ␣, the more damping is induced in the 2. User-defined variables, which can technically be regarded as
numerical solution. Note that, the choice ␣ = 0 leads to the trap- aliases or definition equations. A set of nv user-defined vari-
ezoidal method with no numerical damping. ables V 苸 Rnv is typically specified through an equation of
The Multibody Dynamics Problem. In the case of a multi- the form
body system, without any loss of generality, the set of generalized V − v共q,q̇,q̈,X,␭,V,F,t兲 = 0nv 共8b兲
coordinates considered henceforth is as follows: for each body i
its position is described by the vector ri = 关xi , y i , zi兴T, while its which, during the solution sequence, are solved 共or rather
orientation is given by the array of local 3-1-3 Euler angles 共see, evaluated兲 simultaneously with the equations of motion and
for instance, 关6兴兲, ei = 关␺i , ␪i , ␾i兴T. Consequently, for a mechanical the kinematic constraint equations. Here v 苸 Rnv is a user-
system containing nb bodies, q = 关rT1 eT1 ¯ rTn eTn 兴T 苸 R p, p = 6nb. defined function that depends on other system states as in-
b b
Note that this set of position generalized coordinates is augmented dicated in Eq. 共8b兲.
with deformation modes when flexible bodies are present in the 3. External force definition, F, which allows the user to define
model. the set of n f applied forces F 苸 Rn f that act on the system.
In any constrained mechanical system, joints connecting bodies This is the mechanism through which a complex tire model
restrict their relative motion and impose constraints on the gener- can be interfaced to a vehicle model, that supports user-
alized coordinates. Kinematic constraints are then formulated as defined bushing elements, custom nonlinear damper and
algebraic expressions involving generalized coordinates friction models, and the like.

⌽共q,t兲 = 关⌽1共q,t兲 ¯ ⌽m共q,t兲兴T = 0 共7a兲 F − f共q,q̇,q̈,X,Ẋ,␭,V,F,t兲 = 0n f 共8c兲

where m is the total number of independent constraint equations
that must be satisfied by the generalized coordinates throughout Equations 共7a兲–共7d兲 comprise a system of index 3 DAE 关9兴,
the simulation. Although in its current implementation the algo- where the notion of DAE index, as far as its definition and mean-
rithm handles nonholonomic constraints in Pfaffian form 关7兴, for ing are concerned, continues to be a research topic and the inter-
brevity, the case of holonomic constraints is assumed henceforth. ested reader is referred to 关10兴 and the recent work 关11兴 for a more
Differentiating Eq. 共7a兲 with respect to time leads to the veloc- in-depth analysis of the concept. It is known that differential-
ity kinematic constraint equation algebraic equations are not ordinary differential equations 关12兴.
Analytical solutions of Eqs. 共7a兲 and 共7d兲 automatically satisfy
⌽q共q,t兲q̇ + ⌽t共q,t兲 = 0 共7b兲 Eqs. 共7b兲 and 共7c兲, but this is no longer true for numerical solu-
where the overdot denotes differentiation with respect to time and tions. In general, the task of obtaining a numerical solution of the
the subscript denotes partial differentiation, ⌽q = 关⳵⌽i / ⳵q j兴, for DAE of Eqs. 共7a兲–共7d兲 is substantially more difficult and prone to
1 艋 i 艋 m, 1 艋 j 艋 p. The acceleration kinematic constraint equa- intense numerical computation than that of solving ordinary dif-
tion is obtained by differentiating Eq. 共7b兲 with respect to time: ferential equations. For an account of relevant work in the area of
numerical integration methods for the DAE of multibody dynam-
⌽q共q,t兲q̈ + „⌽q共q,t兲q̇…qq̇ + 2⌽qt共q,t兲q̇ + ⌽tt共q,t兲 = 0 共7c兲 ics the reader is referred to 关3,9,13–18兴 and references therein.
The state of the mechanical system changes in time under the The theory and attractive features associated with the HHT
effect of applied forces such that Eqs. 共7a兲–共7c兲 are satisfied at all method have been derived in conjunction with a linear second-
times. The time evolution of the system is governed by the order ODE. The only similarity between Eqs. 共1兲 and 共7d兲 is that
Lagrange multiplier form of the constrained equations of motion they are both second order and qualitatively obtained from New-
共see, for instance 关8,6兴兲, ton’s second law. In 关19兴, for the purpose of stability and conver-
gence analysis, the constrained equations of motion are tackled in
M共q兲q̈ + ⌽qT共q兲␭ = Q共q̇,q,t兲 共7d兲 a stabilized index 2 DAE framework. The HHT method is also
where M共q兲 苸 R p⫻p
is the generalized mass, and Q共q̇ , q , t兲 苸 R p is discussed in 关20兴 and more recently in 关21兴, where the proposed
implementation is based on a technique that accounts for viola-
the action 共as opposed to the reaction ⌽qT共q兲␭兲 force acting on the
tions in the position and velocity constraints in a stabilization
generalized coordinates q 苸 R p. These equations are neither linear framework similar to the one proposed in 关22兴. There are also
nor ordinary differential as is the case in Eq. 共1兲, first and fore- several Runge-Kutta-based approaches for highly oscillatory me-
most because the solution q共t兲 must also satisfy the kinematic chanical system simulation that, like the HHT method, display the
constraint equations in Eq. 共7a兲. These constraint equations lead in attractive attribute of selectively damping frequency at the high
Eq. 共7d兲 to the presence of the reaction force ⌽qT共q兲␭, where ␭ end of the spectrum. In 关23兴, a singly diagonal Runge-Kutta
苸 Rm is the Lagrange multiplier associated with the kinematic 共SDIRK兲-based method allows the user to choose, within certain
constraints. bounds, the diagonal value in the formula and thus control the
In addition to the equations of motion and kinematic constraint, amount of numerical damping associated with the algorithm. The
several classes of equations need to be considered in a general- role of the diagonal element in the formula becomes similar to
purpose mechanical simulation package: that of the ␣ parameter in the HHT method. An approach based on
additive Runge-Kutta methods that has the potential to accurately
1. Ordinary differential equations, which in the most general
handle highly oscillatory multibody dynamics simulation was in-
case are provided in implicit form
troduced in 关24兴 and further discussed in 关25兴. These novel Runge-
d共Ẋ,X,q,q̇,q̈,␭,V,F,t兲 = 0nd 共8a兲 Kutta-based algorithms are mathematically sound, but they re-
quire more time to achieve, vis-a-vis industrial-strength
This type of differential equation is encountered, for in- applications, the level of acceptance currently associated with the
stance, when controls are active in the system, such as is the well-established HHT method.
case in cars with an anti-lock braking system 共ABS兲, active The method used in this paper was also considered in 关26兴, and
suspension control, and so forth. The state of the controller then revisited in 关5兴. Thus, this work does not propose a new

74 / Vol. 2, JANUARY 2007 Transactions of the ASME

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 integration method 共the HHT method兲, but rather an algorithm 共Mq̈兲n+1 + 共1 + ␣兲共⌽qT␭ − Q兲n+1 − ␣共⌽qT␭ − Q兲n = 0 共10兲
based on this method, which 共a兲 proposes a step-size control strat-
egy at the position level that works in conjunction with both the For notational simplicity, when obvious, the dependency of some
second-order differential equations of motion and the first-order quantities on q and/or q̇ and/or time t will be omitted, as was done
equations that typically come from hydraulic or control systems; in Eq. 共10兲. From an implementation standpoint it is more advan-
共b兲 proposes a choice of unknowns 共accelerations and Lagrange tageous to scale the previous equation by 1 / 共1 + ␣兲 to obtain the
multipliers, respectively兲 in the Newton-type algorithm associated equivalent form
with the implicit discretization scheme that results in optimal con- 1 ␣
dition number with extremely positive impact on the robustness of 共Mq̈兲n+1 + 共⌽qT␭ − Q兲n+1 − 共⌽T␭ − Q兲n = 0
the algorithm; 共c兲 defines a stopping criteria for the nonlinear 1+␣ 1+␣ q
system that is tied to the level of user-specified accuracy in an 共11a兲
analytically sound way; and 共d兲 withstood the test of industrial
strength applications as it was implemented and released in a Discretization of the rest of the problem equations leads to
commercial simulation software that used it successfully in con- ⌽共qn+1,tn+1兲 = 0 共11b兲
junction with models with more than 18,000 states 共engine
model兲.
d共Ẋn+1,Xn+1,qn+1,q̇n+1,q̈n+1,␭n+1,Vn+1,Fn+1,tn+1兲 = 0 共11c兲

Vn+1 − v共qn+1,q̇n+1,q̈n+1,Xn+1,␭n+1,Vn+1,Fn+1,tn+1兲 = 0
The Proposed Algorithm
共11d兲
In addition to the Newmark formulas of Eq. 共2兲 an integration
formula is required for the solution of the first-order ODE in Eq.
共8a兲. The general formula considered is Fn+1 − f共qn+1,q̇n+1,q̈n+1,Xn+1,Ẋn+1,␭n+1,Vn+1,Fn+1,tn+1兲 = 0
共11e兲
Xn+1 = Xn + hẊn + h␳共Ẋn+1 − Ẋn兲 共9a兲
Everywhere in Eq. 共11兲, in an index 3 DAE direct approach, the
in which case Newmark integration formulas of Eq. 共2兲 are used to express q
⳵Xn+1 and q̇ as a function of q̈, while Eq. 共9a兲 is used to discretize the
= ␳hInd 共9b兲 set of ordinary differential equations 共express X as a function of
⳵Ẋn+1
Ẋ兲. A Newton-type algorithm 关30兴 is used to solve the resulting
where throughout this paper Is stands for the identity matrix of system of nonlinear equations for the set of unknowns 共in this
dimension s. The discussion about the choice of the parameter ␳ is
order兲 q̈, ␭, Ẋ, V, and F. Note that the generalized force Q of Eq.
deferred to a later section.
For the multibody dynamics problem stated, the unknowns of 共7d兲 is obtained by projecting the force states F along the gener-
interest are the generalized positions, velocities, and accelerations alized coordinates q; that is, Q共q , F兲 = ⌸F, where the projection
q, q̇, and q̈, respectively, the Lagrange multipliers ␭, the applied- operator ⌸ = ⌸共q兲 depends on the choice of generalized coordi-
force states F, the user-defined variables 共aliases兲 V, and the states nates. The iterative algorithm requires at each iteration 共k兲 the
associated with the user-defined ordinary differential equations, solutions of the linear system

冤 冥
that is, X and Ẋ. The index 3 DAE multibody dynamics problem M̂ ⌽qT 0 0 −⌸
that must be solved to compute these quantities is neither linear
nor ordinary differential, and the HHT method is thus applied for ⌽q 0 0 0 0
a different class of problems from that originally designed for. 共dq̈ + ␥hdq̇ + ␤h2dq兲 d␭ dX˙ + ␳hdX dV dF
Rather than approaching the solution within an index 2 framework
关19兴 or using a stabilization approach 关20,21兴, the proposed algo- − 共vq̈ + ␥hvq̇ + ␤h2vq兲 − v␭ − ␳hvX I − vV − vF
rithm uses the implicit Newmark formulas to discretize the equa- − 共fq̈ + ␥hfq̇ + ␤h fq兲2
− f␭ − fX˙ − ␳hfX − fV I − fF
tions of motion and requires that the position-level kinematic con-
straint equations be satisfied at the end of each time step. This is a

冤冥冤冥
共k兲
direct index 3 approach, and it requires at each integration time ⌬q̈ − e1 共k兲
step the solution of a nonlinear system of equations. The theoret- ⌬␭ − e2
ical foundation of this algorithm is provided by the stability and
convergence results and observations in 关27,28,26,5兴. However, ⫻ ⌬Ẋ = − e3 共12兲
we are not aware of a global convergence proof for the HHT ⌬V − e4
method when used in conjunction with the fully nonlinear index 3
problem of multibody dynamics. Current work is underway for a ⌬F − e5
rigorous global convergence proof for a stabilized index 2 formu- where ei are the residuals in satisfying the set of discretized equa-
lation 关29兴. The global convergence for the direct HHT-based in- tions of motion, constraint equations, discretized first-order differ-
dex 3 formulation remains very much an open question, but the ential equations, variable definition equations, and applied force
results obtained with the proposed algorithm provide an early definition equations, respectively, and, unless otherwise specified,
validation of the good properties associated with the HHT all the quantities in e1 – e5 are evaluated at time tn+1:
method.
At the cornerstone of the proposed algorithm lies the simple 1 ␣
e1 = 共Mq̈兲n+1 + 共⌽qT␭ − Q兲n+1 − 共⌽T␭ − Q兲n
idea on which the HHT method is built: recycle the Newmark 1+␣ 1+␣ q
integration formulas, but slightly change the equations of motion
to account for the set of forces acting on the system at two con- 1
secutive integration points. The algorithm is modified to a small e2 = ⌽共q,t兲
extent to accommodate the set of differential and algebraic equa- ␤h2
tions 共8a兲–共8c兲 that a general-purpose simulation package would
have to handle. Thus, in the spirit of the original HHT formula- e3 = d共Ẋ,X,q,q̇,q̈,␭,V,F,t兲 共13兲
tion, the discretization of the multibody dynamics equations of
motion yields e4 = V − v共q,q̇,q̈,X,␭,V,F,t兲

Journal of Computational and Nonlinear Dynamics JANUARY 2007, Vol. 2 / 75

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 e5 = F − f共q,q̇,q̈,X,Ẋ,␭,V,F,t兲 error in advancing the simulation from step n to n + 1. Once the
local integration error is available, an algorithm is implemented to
The matrix M̂ in Eq. 共12兲 is defined as ensure that this error stays smaller than a user-prescribed toler-

冋 册
ance. Based on a linearization of the equations of motion in Eq.
⳵e1 1 1 共7d兲 along with an asymptotic expansion of the solution q, a strat-
M̂ = = M + ␤h2 共Mq̈兲q + 共⌽qT␭兲q − 共II F兲q
⳵q̈ 1 + ␣ 1+␣ egy for estimating the local integration error in positions is pre-
共14兲 sented. Following the same approach, namely, linearization and
asymptotic expansion, an estimate of the local integration error is
Note that the nonlinear equations associated with the position also provided for the state X associated with the ordinary differ-
kinematic constraints are scaled by 1 / ␤h2 in order to improve the ential equations defined by Eq. 共8a兲.
conditioning of the coefficient matrix in Eq. 共12兲. This is a com-
promise reached after considering the following options: 共a兲 have Local Integration Error in Positions Coordinates. The approxi-
the level-zero positions, q, and differential states, X, be the un- mation of the local integration error is similar to the approach
proposed in 关34兴 for the Newmark method. The discussion is go-
knowns 共replacing q̈ and Ẋ兲, in which case some entries in the
ing to focus on Eq. 共4兲, because locally a linearization of Eq. 共7d兲
Jacobian matrix in Eq. 共12兲 end up divided by ␤h2; 共b兲 have q̈ and leads to the previous form. Thus, Eq. 共4兲 is rewritten as
Ẋ be the unknowns, in which case the second row in the Jacobian
matrix comes multiplied by ␤h2; and 共c兲 same as in 共b兲, except Mq̈n+1 + 共1 + ␣兲共Cq̇n+1 + Kqn+1兲 − ␣共Cq̇n + Kqn兲 = F共t̃n+1兲
that the set of positions kinematic constraint equations is scaled by
共15兲
1 / ␤h2. Option 共a兲 is implemented by the default integrator used in
the MSC.ADAMS simulation package 关31兴 共here entries get di- with t̃n+1 defined as in Eq. 共5兲.
vided by a factor ␤0h rather than ␤h2, as the second-order equa- For the purpose of computing the local integration error, the
tions of motion are reduced to an equivalent first-order system of usual assumption is that the configuration at time tn, 共qn , q̇n , q̈n兲,
differential equations that is then solved with a BDF-type integra- is perfectly consistent. That is, it satisfies the equations of motion,
tor 关32兴兲. On numerous occasions this has been observed to be the along with the time derivatives of the equations of motion:
cause of numerical problems once the step-size becomes very
small and consequently some entries in the Jacobian become ex- Mq̈n + Cq̇n + Kqn = Fn 共16a兲
tremely large. A bad Jacobian condition number ensues, and the
quality of the Newton corrections becomes poor. The option 共b兲
was not embraced because the problem at 共a兲 plagues this ap- តn + Cq̈n + Kq̇n = Ḟn
Mq 共16b兲
proach as well, though in a more subtle way. If h becomes very The Newmark integration formula of Eq. 共2兲 is rewritten in the
small, the second row of the Jacobian matrix is scaled by ␤h2, equivalent form
which practically makes all the entries in this row very small and
thus leads to ill conditioning. Option 共c兲 proved a good solution h2
qn+1 = qn + hq̇n + q̈n + ␤h2x 共17a兲
because typically the type of error that one sees in satisfying the 2
position kinematic constraint equations is very small. It is never
that these constraint equations are problematic in a simulation, but
q̇n+1 = q̇n + hq̈n + h␥x 共17b兲
rather some discontinuity in the model that causes the step-size h
to assume small values. But if h is small, when advancing the
simulation the position constraint violation stays very small, and q̈n+1 = q̈n + x 共17c兲
the norm of e2 always remains bounded. A formal proof of this where the unknown x represents the change in the value of accel-
result is provided in 关33兴, which also discusses the nonsingular
eration from time tn to tn+1. The goal is to compute an estimate of
character of the coefficient matrix in Eq. 共12兲 when h → 0, and the the error at the end of one integration step 共the local integration
convergence of the iterative Newton scheme. Thus, a salient fea- error兲
ture of the approach is that it eliminates the ill conditioning typi-
cally associated with the index 3 integration of the DAE of multi- ␦n+1
q
= qn+1 − q̃n+1 共18兲
body dynamics. Two factors are responsible for this: 共i兲 the
position kinematic constraint equations are appropriately scaled, where q̃n+1 is the exact solution of the initial value problem
and 共ii兲 the set of unknowns q̈ and ␭ are consistent in the sense
that they are qualitatively of the same kinematic level, that is, two Mq̈ + Cq̇ + Kq = F 共19兲
共as opposed to mixing q, which is level zero, with ␭, which is that starts in the configuration 共qn , q̇n , q̈n兲 at t = tn.
level two兲. Using Taylor’s theorem, one obtains q̃n+1 as
With the corrections computed as the solution of the linear sys-
tem of Eq. 共12兲, the numerical approximation of the solution is h2 h3
improved at each iteration as q̈共k+1兲 = q̈共k兲 + ⌬q̈共k兲, ␭共k+1兲 = ␭共k兲 q̃n+1 = qn + hq̇n + q̈n + q តn + O共h4兲 共20兲
2 6
+ ⌬␭共k兲, Ẋ共k+1兲 = Ẋ共k兲 + ⌬Ẋ共k兲, V共k+1兲 = V共k兲 + ⌬V共k兲, F共k+1兲 = F共k兲
+ ⌬F共k兲. The following sections present in detail the answer to The local integration error ␦n+1
q
becomes available as soon as the
three key questions: 共a兲 When is the computed solution accurate acceleration correction x is available. In order to obtain an esti-
enough? 共b兲 How does one select the integration step-size h? 共c兲 mate for x, based on Eqs. 共15兲 and 共17兲

冋
When should one stop the Newton-type iterative process that
computes at each integration step the unknowns q̈, ␭, Ẋ, V, and M共q̈n + x兲 + 共1 + ␣兲 C共q̇n + hq̈n + h␥x兲

冊册
F? Recall that once q̈ and Ẋ are available, Eqs. 共2a兲, 共2b兲, and
共9a兲 are used to evaluate q, q̇, and X, respectively.
冉
+ K qn + hq̇n +
h2
2
q̈n + ␤h2x − ␣共Cq̇n + Kqn兲
Implementation Details for Proposed Algorithm
= Fn + 共1 + ␣兲Ḟnh + O共h2兲 共21兲
Estimating the Local Integration Error. Since an approxima-
tion of the global error at time tn+1 cannot be obtained in general, where Taylor’s theorem was used to expand F共tn + 共1 + ␣兲h兲. Using
the goal is to produce an approximation of the local integration Eqs. 共16a兲 and 共16b兲,

76 / Vol. 2, JANUARY 2007 Transactions of the ASME

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 តnh + O共h2兲
关M + 共1 + ␣兲h␥C + 共1 + ␣兲␤h2K兴x = 共1 + ␣兲Mq ␦n+1 ⬇ 共␳ − 2 兲hxd
1
d
共34兲
共22兲 This is an effective way of computing the local integration error
Denoting D = M + 共1 + ␣兲h␥C + 共1 + ␣兲␤h2K, since D−1 = M−1 because all the quantities on the right side of the previous equa-
+ O共h兲 · I p, the equation tion are available at the end of the corrector phase.
Note that Eq. 共33兲 is relevant for ␳ ⫽ 0.5. The choice ␳ = 0.5
តnh + O共h2兲
Dx = 共1 + ␣兲Mq 共23兲 corresponds to the trapezoidal formula, for which one additional
term in the Taylor expansion would need to be considered
leads to throughout the derivation. This is qualitatively similar to the pre-
តnh + O共h2兲
x = 共1 + ␣兲q 共24兲 sentation herein and is not detailed further. Other choices of ␳
苸 共1 / 2 , 1兴 are viable, and it is insightful to compare Eq. 共9兲 with
and therefore the Newmark formula of Eq. 共2b兲. This idea can be taken one step
␦n+1 = qn+1 − q̃n+1 = h3共␤共1 + ␣兲 − 6 兲q
តn + O共h4兲 further and combined with the introduction of a fictitious variable
1
q
共25兲
Z, defined as Ż = X. In this case Eq. 共8a兲 leads to a second-order
តn from Eq. 共24兲 and ignoring the higher-order
Substituting for q equation in Z, in which case straight Newmark can be applied to
terms leads to find the solution X. This approached is followed in 关35兴.

␦n+1
q
冉
⬇ ␤−
1
6共1 + ␣兲
冊
h 2x 共26兲 The Accuracy Test. With the local truncation error in positions
q and differential states X obtained as indicated in Eqs. 共26兲 and
共34兲, the numerical integrator has to certify at time tn+1 the accu-
which provides an effective way of computing the local integra-
racy of the newly computed solution. Two tests performed to this
tion error, since the required quantities are available at the end of
end are used to accept or reject the integration step. The tests are
the corrector stage.
based on the value of the position and differential states composite
Local Integration Error in X States. A necessary condition for errors, eq and ed, respectively:

冑 兺冉 冊 冑 兺冉 冊
the differential equations of Eq. 共8a兲 to be locally well defined is p nd
that 关9兴 det共⳵d / ⳵Ẋ兲 ⫽ 0 holds in a neighborhood of the current 1 ␦i,n+1
q 2
1 ␦i,n+1
d 2
eq = ed = 共35兲
system configuration. Assuming that the user-defined form for d p i=1 Y iq nd i=1 Y iX
satisfies this requirement, by using the implicit function theorem
where Y qi = max共1 , max j=1,. . .,n兩qi,j兩兲 and ␦i,n+1
q
, 1 艋 i 艋 p, is the ith
and Taylor’s theorem, Ẋ can be locally expressed explicitly as a
component of ␦n+1.
q
function of X and time t:
The composite error is compared with the user-prescribed error
Ẋ = AX + b共t兲 共27兲 ⑀. Introducing the notation
where A is a constant matrix that depends on the configuration of p⑀2
␺q ⬅ 共36a兲
the system at the time when the linearization is carried out, and b 关␤ − 1/6共1 + ␣兲兴2
is a function of time. One additional time derivative leads to
the error test eq 艋 ⑀ is equivalently expressed as
Ẍ = AẊ + ḃ共t兲 共28兲 ␺q
储x储2q 艋 共36b兲
The integration formula used to integrate the differential equations h4
in Eq. 共8a兲 is equivalently expressed as
where 储 · 储q represents a weighted norm 关36兴 defined as 储x储q
Xn+1 = Xn + hẊn + ␳hxd 共29兲 ⬅ 关兺i=1
p
共xi / Y qi 兲2兴1/2.
For the local integration error in the differential states, introduc-
where xd = Ẋn+1 − Ẋn. The goal is to produce an approximation of ing the notation
the local integration error when advancing the simulation from tn
nd · ⑀2
to tn+1. To this end, suppose that X̃n+1 is the exact solution at tn+1 ␺d = 共37a兲
共␳ − 1/2兲2
starting from 共tn , Xn兲, while Xn+1 is the approximate solution as
computed by the proposed algorithm. By using Taylor’s theorem, the accuracy test ed 艋 ⑀ leads to the requirement
1
X̃n+1 = Xn + hẊn + 2 h2Ẍn + O共h3兲 共30兲 ␺d
储xd储X2 艋 共37b兲
h2
Considering the definition of the local truncation error ␦n+1
d
where the corresponding weighted norm is defined as 储x储X
⬅ Xn+1 − X̃n+1, based on Eqs. 共29兲 and 共30兲,
⬅ 关兺i=1
nd
共xi / Y Xi 兲2兴1/2. Note that all the quantities that enter the ac-
␦n+1
d 1
= ␳hxd − 2 h2Ẍn + O共h3兲 curacy tests in Eqs. 共36b兲 and 共37b兲 are available after the nonlin-
ear discretization system of Eq. 共11兲 is solved, and a decision is
Thus ␦n+1
d
is available as soon as xd becomes available. Note that made at that point whether the newly computed solution is ac-
xd should satisfy cepted or rejected.

Ẋn + xd = AXn + bn + A共hẊn + ␳hxd兲 + hḃn + O共h2兲 共31兲 The Step-Size Selection. Step-size selection plays a central
role in the numerical integration algorithm. If eq Ⰶ ⑀ and ed Ⰶ ⑀,
Substituting for Ẋn from Eq. 共27兲 into Eq. 共31兲 and performing CPU time is wasted in computing a solution that unnecessarily
simple manipulations yields exceeds the user-demanded accuracy. At the other end of the spec-
trum, a step-size selection mechanism that is too aggressive leads
xd = hẌn + O共h2兲 共32兲 to a large number of integration steps at the end of which the user
accuracy requirements are not met. The effort to perform such an
Therefore,
integration step is wasted, as the integration step is discarded for a
␦n+1 = 共␳ − 2 兲h2Ẍn + O共h3兲 new attempt with a more conservative step-size h. To strike the
1
d
共33兲
right note, the integration step-size is always chosen such that the
which leads to error at the end of the next integration step is precisely equal to

Journal of Computational and Nonlinear Dynamics JANUARY 2007, Vol. 2 / 77

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 the one deemed acceptable by the user and quantitatively defined
by ⑀. By ignoring the terms of order h4 and higher, and denoting
តi,n, Eq. 共25兲 suggests that the position com-
ci = 关␤共1 + ␣兲 − 1 / 6兴 q
posite error is proportional to the cube of the step-size h. Ideally,
the new step-size hnew is selected such that

⑀ = hnew
3
冋 兺冉 冊 册
1
p
ci
p i=1 Y i
2
1/2

Therefore, e / ⑀ = h3 / hnew
3
, from where

hnew = h⑀1/3 冋冑 冉
1
p
␤−
1
6共1 + ␣兲
冊
h2 · 储x储q 册 −1/3

By defining
储x储2q · h4
⍜q = 共38a兲
␺q
the position-based criterion for selecting the step-size becomes Fig. 1 Poly-V accessory belt.
sh
q
hnew = 共38b兲
⍜1/6
q

As recommended in 关16兴, a safety factor s = 0.9 was used to scale words, no room is left for errors in finding the numerical solution
the value of the new step-size. The superscript q was added to at the end of one integration step. Finding an approximate solution
indicate that this value of the new step-size is based on position translates into solving a different initial value problem, which can
considerations. be close to or far from the original problem based on how accurate
The approach for computing hnew d
follows step by step the the nonlinear system of Eq. 共11兲 is solved and the nature of the
q original initial value problem itself. In summary, based on these
position-based selection of hnew. Defining

冋兺 冉 冊 册
nd 1/2
two remarks, the corrector stopping criterion adopted here is that
2
␳ − 1/2 Ẍi,n the relative difference between e and e共k兲 should stay smaller than
sd =
冑nd Yi a threshold value denoted by c. A typical value recommended in
i=1
the literature is c = 0.001 关16兴. The local integration error at the
it can be concluded that the error depends quadratically on the end of one time step is

冉 冊
step-size h, like ed = sdh2. Therefore, hnew should be selected such
that ⑀ = sdhnew
2
, which leads to hnew = h⑀1/2e−1/2. Hence, hnew 1 h2
eq = ␤ − 储x储q
= h共␺d / h · 储xd储2X兲1/4. By defining
2 6共1 + ␣兲 冑p

储xd储X2 · h2 After iteration k the approximation obtained is

⍜d = 共39a兲
␺d
the differential-based criterion for selecting the step-size becomes
冉
eq,共k兲 = ␤ −
1
冊
h 2 共k兲
6共1 + ␣兲 冑p
储x 储q

sh The question is what k should be such that eq,共k兲 is close to eq

d
hnew = 共39b兲
⍜1/4
d
within 0.1% 共c = 0.001兲; that is, 兩eq − eq,共k兲兩 艋 c兩eq兩. Since eq is not
available, the test is replaced by
where s = 0.9 is a safety factor 关16兴.
The Correction Stage. The last issue that needs to be ad-
dressed is how accurate the quantities x and xd of Eqs. 共17兲 and
冏 ⑀
冏
eq − eq,共k兲
艋c 共40兲
共29兲, respectively, should be computed. These quantities are ob- where ⑀ is the user-prescribed error. Note that the goal of the
tained as the solution of an iterative Newton-type algorithm that step-size control is to keep eq as close as possible to ⑀; therefore,
recycles the coefficient matrix in Eq. 共12兲 and its LU factorization, substituting the original condition with Eq. 共40兲 is acceptable.
and which requires at least one evaluation of the residuals in Eq. Then,
共13兲, followed by a forward/backward substitution to retrieve the
corrections in the unknowns. However, one corrector iteration
might be as expensive as doing all of the above but preceded by
an evaluation and LU factorization of the coefficient matrix of the
冏
兩eq − eq,共k兲兩 艋 ␤ −
1 h2
6共1 + ␣兲 冑p
冏
储x − x共k兲储q 共41兲

linear system of Eq. 共12兲, two operations that are expensive and and an approximation for 储x − x共k兲储q is needed. Since for the
should be kept to a minimum. Newton-type algorithm employed the convergence is linear, there
Suppose that x is approximated by x共k兲, the value obtained after is a constant ␰ that for convergence must satisfy 0 艋 ␰ ⬍ 1 such
k corrector iterations. Therefore, according to Eqs. 共25兲 and 共35兲, that 关30兴
the composite error eq is actually computed based not on the value
x, but rather on x共k兲, which will lead to a value eq,共k兲. It is therefore 储⌬x共k+1兲储q 艋 ␰储⌬x共k兲储q 共42兲
important to have a good approximation x共k兲 for x if the algorithm 共k兲
where ⌬x represents the correction at iteration k, x 共k+1兲
= x共k兲
is to produce a reliable measure of the local integration error 共a + ⌬x共k兲. Immediately,
similar argument holds for the differential error ed兲. Another rea-
son for having an accurate approximation is that the stability and ␰
convergence results associated with a numerical integrator are de- 储x − x共k+1兲储q 艋 储⌬x共k兲储q 共43兲
1−␰
rived under the assumption that the numerical solution is com-
puted to the specifications of the integration formula; in other The value ␰ is going to be approximated by

78 / Vol. 2, JANUARY 2007 Transactions of the ASME

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Fig. 2 X-component of reaction force Fig. 3 HHT and GSTIFF differences

储⌬x共k兲储q erations prediction is obtained by taking q̈n+1 = q̈n and Ẋn+1 = Ẋn,
␰ ⬇ ␰k = 共44兲
储⌬x共k−1兲储q which is equivalent to setting x = 0 and xd = 0. A new strategy is
Based on Eq. 共41兲, proposed based on polynomial extrapolation, in which a polyno-

冉 冊
mial of order up to three is used to produce an initial guess for the
1 h2 ␰ unknowns. The approach used in 关4兴 is thus obtained by setting
兩eq − eq,共k兲兩 艋 ␤ − 储⌬x共k兲储q 共45兲
冑
6共1 + ␣兲 p 1−␰ the degree of the interpolation polynomial to zero. The polyno-
mial extrapolation is based on Newton divided differences and
The condition of Eq. 共40兲 is then satisfied as soon as uses Horner’s scheme for evaluation of the interpolant at time tn+1

冉 冊
␰
1−␰
2
储⌬x共k兲储2q 艋 c2
␺q
h4
共46兲
关36兴. The heuristics for choosing the degree of the interpolant
would be based on the size of the numerical derivatives of the
solution 共evaluated by divided differences兲, and on the “smooth-
Note that at the right of the inequality sign are quantities that ness” of the numerical solution as reflected in the number of re-
remain constant during the corrector iterative process, while at the jected integration steps.
left are quantities that change at each iteration. Likewise, note that
the stopping criterion of Eq. 共46兲 can be used only at the end of Summary of Key Formulas. Summarized below are the an-
the second iteration because only then can an approximation of swers to the questions 共a兲 What is the stopping criteria for the
the convergence rate ␰ be produced. In other words, the proposed nonlinear discretization algebraic system? 共b兲 How is the integra-
approach will not be able to stop the iterative process after the first tion error computed? and 共c兲 How is the step-size controlled?
iteration. This is not a matter of great concern, however, because Summary of Key Formulas for Handling of the Generalized
models as simple as a one-body pendulum are already nonlinear Coordinates. Notation.
and require more than one iteration.
Qualitatively, the same approach used for the positions-based p⑀2 储x储2qh4
␺q ⬅ ⍜q = 共48a兲
stopping criterion is used for the differential states. Without get- 关␤ − 1/6共1 + ␣兲兴2 ␺q
ting into details, this will lead to the following stopping criterion:
Prediction. Performed based on divided differences 共Newton in-
冉 冊␰d 2
1 − ␰d
␺d
储⌬xd共k兲储X2 艋 c2 2
h
共47兲
terpolation and Horner’s scheme for extrapolation at tn+1兲.
Correction. Linear convergence rate allows for computation of
␰ 共Eq. 共44兲兲. Stopping criterion:
The Prediction Stage. For the Newton-type algorithm used to
find the solution of Eq. 共13兲, a good starting point is essential both
for convergence and for reducing the effort in finding the approxi-
冉 冊␰
1−␰
2
储⌬x共k兲储2q 艋 c2
␺q
h4
共c = 0.001兲 共48b兲

mation of the solution at time tn+1. In 关4兴, the generalized accel- Accuracy check. Performed after corrector converged,

Journal of Computational and Nonlinear Dynamics JANUARY 2007, Vol. 2 / 79

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Fig. 4 Alternator angular velocity
Fig. 5 Alternator force difference

⍜q 艋 1 共48c兲 ried out until the conditions of Eqs. 共48b兲 and 共49b兲 are simulta-
Step-size selection. With a safety factor s = 0.9, neously satisfied. The new step size is chosen as hnew = min 共hnew
q
,
hnew兲. An integration step is not accepted unless both accuracy
d
sh
q
hnew = 共48d兲 conditions of Eqs. 共48c兲 and 共49c兲 are satisfied.
⍜1/6
q

Summary of Key Formulas for Handling of the Differential Numerical Experiments

States. Notation.
The proposed algorithm has been implemented in the commer-
n d⑀ 2 储xd储X2 h2 cial simulation package MSC.ADAMS and released in its
␺d = ⍜d = 共49a兲 2005 version. The algorithm has been extensively tested with
共␳ − 1/2兲2 ␺d
more than 1600 mechanical systems of various complexity. Three
Correction. Linear convergence rate allows for computation of ␰d. representative numerical experiments aimed at comparing the
Stopping criterion: HHT-based algorithm and GSTIFF, the default integrator in the

冉 冊␰d 2
1 − ␰d
␺d
储⌬xd共k兲储X2 艋 c2 2
h
共c = 0.001兲 共49b兲

Accuracy check. Performed after corrector converged,

⍜d 艋 1 共49c兲
Step-size selection. With a safety factor s = 0.9,
sh
d
hnew = 共49d兲
⍜1/4
d
In multibody dynamics simulations, the number of differential
states is orders of magnitude smaller than the number of states
associated with the generalized coordinates; that is, nd Ⰶ p. Nev-
ertheless, the stopping criteria as well as the selection of the new
step-size hnew take into account both the position and differential
states. For stopping the Newton-type algorithm, iterations are car- Fig. 6 Track subsystem model

80 / Vol. 2, JANUARY 2007 Transactions of the ASME

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Fig. 7 Acceleration and velocity of track 8

MSC.ADAMS 关31兴 simulation package, are presented herein. The tensioner 共deviation pulley兲. The poly-V belt provides drive by
comparison primarily focuses on efficiency issues, although the adhesion and, compared with conventional belts of the same
accuracy of the results is touched upon. width, it augments the contact area, increasing power transfer. The
larger of the pulleys 共the lowest one in the picture兲 is the engine
A Poly-V Belt Model. The model in Fig. 1 is an accessory crankshaft pulley. Right above it is the water-pump pulley, and at
drive for a car engine with a poly-V belt 共V-ribbed belt兲 wrapped the left is the alternator pulley. There is a tensioner in between the
around three pulleys 共crank, water pump, and alternator兲 and one crankshaft and alternator pulleys; its pivot point is shown in the
figure as the little ring just outside the belt. The tensioner uses a
rotational spring element that includes damping and stiffness ef-
fects. The units used for this model are Newton, kilograms, mil-
liseconds, and millimeters.
One driving torque is on the crankshaft, and two resisting
torques act on the alternator and water pulleys. The belt is mod-
eled by using 100 segments connected by a set of 400 VFORCE
elements 关31兴. The length of the simulation was 200 ms, which
results in more than a full revolution of the belt. The GSTIFF
integrator was run with ERROR= 1 . E-4, while HHT was run with
ERROR= 1 . E-5 共this is the ⑀ variable of Eqs. 共48a兲 and 共49a兲兲.
Because of the different error control strategies employed by the
two integrators, it has been noticed that GSTIFF typically can be
run with a more lax ERROR for results that are qualitatively simi-
lar to HHT. The simulation time for GSTIFF was 1286 s, while
HHT completed the simulation in 485 s. The simulation was run
on a Windows 2000 machine, with Pentium III CPU, and 512 MB
RAM.
Figure 2 shows the X-component of the reaction force in the
revolute joint that connects the alternator with the rest of the sys-
Fig. 8 All-terrain vehicle „ATV… tem. The agreement of the results is very good: the peak differ-

Journal of Computational and Nonlinear Dynamics JANUARY 2007, Vol. 2 / 81

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Table 1 Top ten von Mises hot spots for flexible frame

GSTIFF HHT
Stress
difference
Stress Time Stress Time GSTIFF
No. Point 共lbf/ inch2兲 共s兲 No. Point 共lbf/ inch2兲 共s兲 versus HHT

1 23,956 300,738 1.104 1 23,956 300,768 1.104 0.01%

2 26,654 204,503 0.746 2 26,654 204,274 0.746 −0.11%
3 33,918 204,075 0.206 3 33,918 204,191 0.206 0.06%
4 46,577 200,688 0.746 4 46,577 200,354 0.746 −0.17%
5 34,560 198,106 0.211 5 34,560 197,887 0.211 −0.11%
6 46,580 193,729 0.326 6 46,580 192,929 0.326 −0.41%
7 30,060 190,386 0.692 7 30,060 189,911 0.692 −0.25%
8 24,704 173,882 0.281 8 24,704 173,355 0.281 −0.30%
9 36,156 171,990 1.446 9 36,156 172,746 1.446 0.44%
10 23,007 170,191 1.100 10 23,007 170,375 1.100 0.11%

ence between the two sets of results is less than 3%; see Fig. 3. tances, frequent place changes, and demanded 90% machine
Figure 4 shows the time variation of the angular velocity of the availabilities are standard requirements in the industry. The type
alternator. The plot displays good correlation between the results of simulation reported here is typical in the virtual prototyping
obtained with the GSTIFF and the HHT integrators. Figure 5 con- cycle when trying to meet these requirements and extend life
firms that the difference between the angular velocity computed cycles of the track and drive systems.
with GSTIFF and HHT integrators is less than 1%. This value is The model in Fig. 6 contains a set of 61 moving parts. It has
smaller than the 3% noticed for the difference in force in the one planar joint, 54 revolute joints, one translational joint, four
alternator joint. This is an expected trend all across the simulation fixed joints 共a joint that removes all six degrees of freedom兲, one
results, where the quality of the velocity level variables is better inline joint primitive, and one motion. This results in a model with
than the quality of the force/acceleration variables. Although not 61 degrees of freedom. The relatively large number of degrees of
shown here, the position-level variables for the two integrators are freedom indicates that the motion of this system is not controlled
practically identical, and in general they are qualitatively better as much through constraints 共joints兲 as through the geometry of
than the velocity-level results obtained with the two integrators. the components and 690 three-dimensional contact elements
A Track Model. The track presented in Fig. 6 is a detailed 共sprocket-track, roller-track, track-track, and track-ground con-
model of a subsystem of a low-mobility hydraulic mining exca- tacts兲. The HHT integrator resulted in a set of 4675 equations for
vator. Weight and extreme operating conditions cause high me- the dynamic analysis of this model.
chanical stresses on crawler tracks, especially in the case of big A 14 s simulation was run on a Windows XP Dell Precision
hydraulic excavators of 1000 tons and higher. Long haulage dis- M530 machine, with 2 GB ECC RAM and 2.8 GHz HyperThread

Fig. 9 Comparison of vertical reaction force

82 / Vol. 2, JANUARY 2007 Transactions of the ASME

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Fig. 10 Engine pitch angular velocity Fig. 11 Angular velocity difference

Xeon CPUs. A first set of results was obtained with the default
version of the ADAMS/Solver. The key integrator settings were as A Model With Flexible Bodies. The model in Fig. 8 is an
follows: GSTIFF integrator, stabilized index 2 共SI2兲 DAE formu- all-terrain vehicle 共ATV兲 with a flexible frame. Beneath the ATV
lation 关37兴, ERROR= 1E − 2, KMAX= 1 共to reflect the 3D contact- is a four-post shaker device, which is used to simulate a durability
induced discontinuous nature of the simulation兲, MAXIT= 10. The event represented by a simple-harmonic translational displace-
HHT integrator was run in the beta version of the 2005 release, ment constraint at each of the four posts 共pitch-type excitation, of
with the following key settings: ERROR= 1E − 5, MAXIT= 10, approximately 2 Hz兲.
DAE formulation was index 3 关27,28兴. Note the difference be- The frame component in Fig. 8 is an MSC.ADAMS flexible-
tween the ERROR settings in the two cases. This is explained by body modal representation 关38兴 that was created with the
the different ways in which the error is quantified by the two MSC.Patran finite element package; it contains 134 modes 共from
integrators. In this context, the SI2-GSTIFF integrator has a much 56.7 Hz to 13.1 kHz兲. The purpose of this simulation is to recover
stricter interpretation of the user-set ERROR. In has been noticed the von Mises stresses in the flexible body and identify critical
that in order to obtain qualitatively comparable results, the HHT stress locations in the frame.
ERROR setting should be two orders of magnitude more stringent The steering system has a motion constraint that applies a ro-
than that of the SI2-GSTIFF integrator. To be on the safe side, for tational displacement function, causing the front wheels to turn
this model the HHT integrator was run with an even tighter error left to right in a sinusoidal fashion. The tires of the vehicle interact
setting, ERROR= 1.E-5, which actually is the HHT default setting with the shaker by means of three-dimensional solid-to-solid con-
for this attribute. tact forces. A rider weight of 230 lbs has been approximated with
The speedup obtained when using the HHT integrator was more lumped masses and distributed between the steering column as-
than fivefold: it took 1713 s for the simulation to complete when sembly and frame as 30 and 200 lbs, respectively. Remaining
using the HHT integrator, while it took 8988 s for the GSTIFF parts in the model that are attached to the frame, such as the
integrator to finish the same simulation. The quality of the new engine, are modeled as lumped masses and are connected by using
results is very good. The accelerations are always the most likely fixed joints.
to show differences. Differences can only rarely be noticed in Altogether there are 28 moving parts, 43 joints, 5 motions, and
velocity results, while the quality of the position level results is 1 flexible body, leading to a total number of 532 equations for the
almost always very good in both integrators. For comparison, in HHT integrator and 927 equations for the Index-3 GSTIFF 共I3兲
Fig. 7 the acceleration and velocity results are displayed for track formulation. The MSC.ADAMS modeling units used for this ex-
no. 8. The results match overall very well everywhere with the periment were poundគforce, poundគmass, inch, and second.
exception of some spikes that are explained by the sensitivity of The duration of the simulation was 2.0 s, which allows the ATV to
the simulation with respect to the large number of contact forces move up and down several times. The GSTIFF integrator was run
present in the model. with ERROR= 1e − 4, while HHT was run with ERROR= 1e − 5.

Journal of Computational and Nonlinear Dynamics JANUARY 2007, Vol. 2 / 83

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Considering the typical use of a simulation like this, namely order 5 or 6, the HHT method cannot produce a solution of similar
in durability 共fatigue兲 analysis, a maximum time step 共HMAX quality in comparable CPU time because of the low-order integra-
= 5e − 4兲 was specified for accurate post-processing hot-spot tion formulas employed. However, this scenario is not very com-
stress-recovery. mon, because most real-life large models contain discontinuities
The simulation CPU time for GSTIFF was 367.08 s, while or stiff mechanical components that typically limit the BDF inte-
HHT completed the simulation in 122.47 s. The simulation was gration order to 1 or 2. As seen in the numerical experiments
run by using MSC.ADAMS 2005 on a Windows 2000 laptop presented, in these cases the HHT-based algorithm has proved to
computer with a single 2.20 GHz Pentium 4 CPU and 1 GB be very competitive.
RAM. The accuracy of the results is good, occasional spikes in accel-
There is excellent agreement in location of critical stresses as erations and reaction forces being explained by the use of a vari-
presented in Table 1; the difference in peak stress between the two able step integration algorithm for the solution of an index 3 DAE
sets of results is less than 0.5%. The Z-component of the reaction problem, an operation that is conjectured in 关9兴 to further reduce
force in the spherical joint that connects the right half of the rear the order of an already low-order method. Quantitatively, the
suspension component to the frame is presented in Fig. 9. The plot simulation results can be improved by decreasing the user-
shows good correlation between the force results obtained with specified integration error; qualitatively, the results could be im-
the GSTIFF and HHT integrators. Figure 10 presents the time proved by using a Runge-Kutta method as proposed in 关24兴, using
variation of the angular velocity of the engine assembly and is an the generalization of the HHT method as proposed in 关39兴, or
indicator of the severity of the ATV pitching behavior. Figure 11 reducing the index of the problem in an approach similar to the
confirms that the difference between the angular velocity com- one proposed in 关37兴, an alternative that is currently under inves-
puted with GSTIFF and HHT integrators is less than 1.6%. tigation 关29兴.

Conclusions Acknowledgment
The HHT method used in structural dynamics and later intro- The authors thank Dipl.-Ing. Holger Haut of the University of
duced in the context of multibody dynamics simulation in 关26兴 Aachen, Germany, for providing images, results plots, and timing
serves as the starting point of an algorithm implemented and vali- results for the track subsystem simulation, and Andrei Schaffer of
dated in an industrial strength mechanical system simulation pack- MSC.Software for his suggestions. The authors thank an anony-
age. Strategies for corrector stopping criteria, error estimation, mous reviewer for pointing out additional reference material for
and step-size control were presented in detail. A set of real-life the concept of DAE index.
numerical experiments indicates that simulations are at least two
to three times faster when compared with the default BDF-based References
integrator used in ADAMS 关31,32兴. An explanation for the im- 关1兴 Hilber, H. M., Hughes, T. J. R., and Taylor, R. L., 1977, “Improved Numerical
proved performance is based on three key observations. 共1兲 The Dissipation for Time Integration Algorithms in Structural Dynamics,” Earth-
most time consuming part of simulation is the computation of the quake Eng. Struct. Dyn., 5, pp. 283–292.
Jacobian associated with the nonlinear discretization system. The 关2兴 Newmark, N. M., 1959, “A method of computation for structural dynamics,” J.
Engrg. Mech. Div., 85, pp. 67–94.
proposed algorithm contains heuristics to reduce as much as pos- 关3兴 Ascher, U. M., and Petzold, L. R., 1998, Computer Methods for Ordinary
sible the number of Jacobian evaluations. Unlike the BDF integra- Differential Equations and Differential-Algebraic Equations, SIAM, Philadel-
tor employed by GSTIFF, in which terms of the integration Jaco- phia, PA.
bian can become disproportionately large as a result of a scaling 关4兴 Hughes, T. J. R., 1987, Finite Element Method—Linear Static and Dynamic
Finite Element Analysis, Prentice-Hall, Englewood Cliffs, NJ.
by the inverse of the step-size, the proposed integrator employs a 关5兴 Cardona, A., and Geradin, M., 1994, “Numerical Integration of Second Order
different approach where certain values are multiplied 共never di- Differential-Algebraic Systems in Flexible Mechanics Dynamics,” in
vided兲 by the step-size prior to populating the Jacobian. As long as Computer-Aided Analysis of Rigid and Flexible Mechanical Systems, M. F. O.
the step-size does not significantly change over several consecu- S. Pereira and J. A. C. Ambrosio, eds., Kluwer Academic, Dordrecht.
关6兴 Shabana, A. A., 1994, Computational Dynamics, Wiley, New York.
tive time steps, this approach better supports the recycling of the 关7兴 Pars, L. A., 1965, A Treatise on Analytical Dynamics, Wiley, New York.
Jacobian. 共2兲 When compared with the BDF Jacobian, the HHT 关8兴 Haug, E. J. 1989, Computer-Aided Kinematics and Dynamics of Mechanical
Jacobian is numerically better conditioned, thereby leading to Systems, Prentice-Hall, Englewood Cliffs, NJ.
more reliable corrections in the Newton-type iterative approach 关9兴 Brenan, K. E., Campbell, S. L., and Petzold, L. R., 1989, Numerical Solution
of Initial-Value Problems in Differential-Algebraic Equations, North-Holland,
for large problems. Typically, this results in a smaller number of New York.
corrector iterations. This desirable attribute is further enhanced by 关10兴 Le Vey, G., 1994, “Differential Algebraic Equations—A New Look at the
the fact that since certain partial derivatives are scaled by the Index,” Technical report, Institut de Recherche en Informatique et Systemes
step-size h, or by h2 prior to populating the Jacobian, small errors Aleatoires.
关11兴 Kunkel, P., and Mehrmann, V., 2006, Differential Algebraic Equations. Analy-
in these partial derivatives are going to have a less negative effect sis and Numerical Solution, European Math. Soc. Textbooks in Mathematics,
on the overall quality of the Jacobian. 共3兲 On one hand, the BDF Zurich.
formulas of order higher than one contain regions of instability in 关12兴 Petzold, L. R., 1982, “Differential-Algebraic Equations are not ODE’s,” SIAM
the left plane. The higher the order, the smaller the region of 共Soc. Ind. Appl. Math.兲 J. Sci. Stat. Comput., 3共3兲, pp. 367–384.
关13兴 Bauchau, O. A., Bottasso, C. L., and Trainelli, L., 2003, “Robust Integration
stability. On the other hand, BDF intrinsically is designed to maxi- Schemes for Flexible Multibody Systems,” Comput. Methods Appl. Mech.
mize the integration order/step-size. Because of these two con- Eng., 192, pp. 395–420.
flicting attributes, particularly for models that are mechanically 关14兴 Eich-Sollner, E., and Fuhrer, C., 1998, Numerical Methods in Multibody Dy-
stiff 共models with stiff springs, flexible bodies, etc., that lead to namics, Teubner-Verlag, Stuttgart.
关15兴 Anderson, K. S., and Oghbaei, M., 2005, “A Dynamics Simulation of Multi-
systems with large eigenvalues close to the imaginary axis兲, an body Systems Using a New State-Time Methodology,” Multibody Syst. Dyn.,
order/step-size choice often lands the BDF integrator outside the 14共1兲, pp. 61–80.
stability region 关16兴. These integration time-steps typically end up 关16兴 Hairer, E., and Wanner, G., 1991, Solving Ordinary Differential Equations,
being rejected, and smaller step-sizes are required to advance the Vol. II of Computational Mathematics, Springer-Verlag, New York.
关17兴 Lubich, C., Nowak, U., Pohle, U., and Engstler, C., 1995, “MEXX—
simulation. This is a nonissue with the HHT method, which is a Numerical Software for the Integration of Constrained Mechanical Multibody
fixed low-order method with good stability properties in the whole Systems,” Mech. Struct. Mach., 23, pp. 473–495.
left plane. 关18兴 Potra, F. A., 1993, “Implementation of Linear Multistep Methods for Solving
It should be pointed out that there are situations when BDF- Constrained Equations of Motion,” SIAM 共Soc. Ind. Appl. Math.兲 J. Numer.
Anal., 30共3兲, pp. 774–789.
type formulas are going to work significantly faster. These are the 关19兴 Yen, J., Petzold, L., and Raha, S., 1996, “A Time Integration Algorithm for
cases where BDF can sustain a high integration order throughout Flexible Mechanism Dynamics: The DAE ␣-Method,” Technical report TR96-
the simulation. If the model simulated allows BDF to work at 024, Dept. of Comp. Sci., University of Minnesota.

84 / Vol. 2, JANUARY 2007 Transactions of the ASME

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 关20兴 Garcia de Jalon, J., and Bayo, E., 1994, Kinematic and Dynamic Simulation of Optimization and Nonlinear Equations, Prentice-Hall, Englewood Cliffs, NJ.
Multibody Systems. The Real-Time Challenge, Springer-Verlag, Berlin. 关31兴 MSCsoftware, 2005, ADAMS User Manual. Also available online at http://
关21兴 Cuadrado, J., Dopico, D., Naya, M. N., and Gonzalez, M., 2004, “Penalty, www.mscsoftware.com
Semi-recursive and Hybrid Methods for MBS Real-Time Dynamics in the 关32兴 Gear, C. W., 1971, Numerical Initial Value Problems of Ordinary Differential
Context of Structural Integrators,” Multibody Syst. Dyn., 12, pp. 117–132. Equations, Prentice-Hall, Englewood Cliffs, NJ.
关22兴 Baumgarte, J., 1972, “Stabilization of Constraints and Integrals of Motion in 关33兴 Gavrea, B., Negrut, D., and Potra, F., 2005, “The Newmark Integration
Dynamical Systems,” Comput. Methods Appl. Mech. Eng., 1, pp. 1–16. Method for Simulation of Multibody Systems: Analytical Considerations
关23兴 Owren, B., and Simonsen, H., 1995, “Alternative Integration Methods for 共IMECE 2005-81770兲,” in Proceedings of the International Mechanical Engi-
Problems in Structural Dynamics,” Comput. Methods Appl. Mech. Eng., 122, neering Congress and Exposition, Orlando, FL, ASME, New York.
pp. 1–10. 关34兴 Zienkiewicz, O. C., and Xie, Y. M., 1991, “A Simple Error Estimator and
关24兴 Jay, L. O., 1998, “Structure Preservation for Constrained Dynamics With Su- Adaptive Time-Stepping Procedure for Dynamic Analysis,” Earthquake Eng.
per Partitioned Additive Runge-Kutta Methods,” SIAM J. Sci. Comput. Struct. Dyn., 20, pp. 871–887.
共USA兲, 20, pp. 416–446.
关35兴 Bruls, O., Duysinx, P., and Golinval, J., 2005, “A Unified Finite Element
关25兴 Schaub, M., and Simeon, B., 2002, “Automatic h-Scaling for the Efficient
Framework for the Dynamic Analysis of Controlled Flexible Mechanisms,” in
Time Integration of Stiff Mechanical Systems,” Multibody Syst. Dyn., 8, pp.
Proceeding of Multibody Dynamics ECCOMAS Thematic Conference, June
329–345.
关26兴 Cardona, A., and Geradin, M., 1989, “Time Integration of the Equation of 21-24, Madrid, Spain.
Motion in Mechanical Analysis,” Comput. Struct., 33, pp. 881–820. 关36兴 Atkinson, K. E., 1989, An Introduction to Numerical Analysis, 2nd ed., Wiley,
关27兴 Lötstedt, C., and Petzold, L., 1986, “Numerical Solution of Nonlinear Differ- New York.
ential Equations With Algebraic Constraints I: Convergence Results for Back- 关37兴 Gear, C. W., Gupta, G., and Leimkuhler, B., 1985, “Automatic Integration of
ward Differentiation Formulas,” Math. Comput., 174, pp. 491–516. the Euler-Lagrange Equations With Constraints,” J. Comput. Appl. Math., 12,
关28兴 Brenan, K., and Engquist, B. E., 1988, “Backward Differentiation Approxima- pp. 77–90.
tions of Nonlinear Differential/Algebraic Systems,” Math. Comput., 51共184兲, 关38兴 Craig, R. R., and Bampton, M. C. C., 1968, “Coupling of Substructures for
pp. 659–676. Dynamics Analyses,” AIAA J., 6共7兲, pp. 1313–1319.
关29兴 Jay, L. O., and Negrut, D., “On an Extension of the HHT Method for Index 3 关39兴 Chung, J., and Hulbert, G. M., 1993, “A Time Integration Algorithm for Struc-
Differential Algebraic Equations,” Electron. Trans. Numer. Anal., submitted. tural Dynamics With Improved Numerical Dissipation: The Generalized-␣
关30兴 Dennis, J., and Schnabel, R., 1983, Numerical Methods for Unconstrained Method,” ASME J. Appl. Mech., 60共2兲, pp. 371–375.

Journal of Computational and Nonlinear Dynamics JANUARY 2007, Vol. 2 / 85

Downloaded From: http://computationalnonlinear.asmedigitalcollection.asme.org/ on 01/29/2016 Terms of Use: http://www.asme.org/about-asme/terms-of-use