Sunteți pe pagina 1din 20

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/265853858

A review of automatic time-stepping strategies


on numerical time integration for structural
dynamics analysis

Article in Engineering Structures December 2014


DOI: 10.1016/j.engstruct.2014.08.016

CITATIONS READS

5 124

4 authors, including:

Diogo Folador Rossi W.G. Ferreira


Universidade Federal do Esprito Santo Universidade Federal do Esprito Santo
1 PUBLICATION 5 CITATIONS 20 PUBLICATIONS 32 CITATIONS

SEE PROFILE SEE PROFILE

Adenilcia Fernanda Grobrio Calenzani


6 PUBLICATIONS 6 CITATIONS

SEE PROFILE

All content following this page was uploaded by W.G. Ferreira on 21 July 2015.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
Engineering Structures 80 (2014) 118136

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Review article

A review of automatic time-stepping strategies on numerical time


integration for structural dynamics analysis
Diogo Folador Rossi a, Walnrio Graa Ferreira a,, Webe Joo Mansur b,
Adenilcia Fernanda Grobrio Calenzani a
a
Department of Civil Engineering, Universidade Federal do Esprito Santo, Vitria, ES, Brazil
b
Program of Civil Engineering, Instituto Alberto Luiz Coimbra de Ps Graduao e Pesquisa de Engenharia, Universidade Federal do Rio de Janeiro, Rio de Janeiro, RJ, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: This paper discusses some of the algorithms available for the automatic adaptive selection of time step
Received 10 November 2013 size, applied to the step-by-step direct time integration methods of structural dynamics problems. Three
Revised 7 August 2014 adaptive strategies based on different concepts are explored and compared: the algorithm of Bergan and
Accepted 8 August 2014
Mollestad (1985), which is based on the current characteristic frequency; the strategy of Hulbert and
Jang (1995), which uses a local error estimator; and the method of Lages et al. (2013), which is based
on the geometric indicator of displacements history curvature. The reviewed strategies are applied to
Keywords:
the Newmark integration scheme to solve various numerical examples of linear dynamic systems, which
Dynamic structural analysis
Automatic time stepping
are presented to compare the performance between the three algorithms that are tested. To conclude, a
Direct time integration methods brief analysis about the considerations of the computational cost is made.
2014 Elsevier Ltd. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
2. Basic structural dynamics concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
2.1. Dynamic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
2.2. Newmark formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3. Automatic time-stepping strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.1. Strategy of Bergan and Mollestad . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.2. Strategy of Hulbert and Jang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.2.1. Normalised error and error tolerance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.2.2. Algorithm for automatic time step size control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.2.3. Application to problems with quiescent start conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.3. Strategy of Lages et al. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.3.1. Indicator of curvature and time step strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.3.2. Curvature regularisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4. Numerical examples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.1. Single-degree-of-freedom systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.2. Shear building with three floors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.3. Simply supported beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.4. A three-bar plane frame structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5. Computational cost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6. Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

Corresponding author. Address: Av. Fernando Ferrari, 514, Goiabeiras, 29075-910 Vitria, ES, Brazil. Tel.: +55 27 40092666, cell: +55 27 99699645.
E-mail addresses: rossi.diogo@gmail.com (D.F. Rossi), walnorio@gmail.com (W.G. Ferreira), webe@coc.ufrj.br (W.J. Mansur), afcalenzani@gmail.com (A.F.G. Calenzani).

http://dx.doi.org/10.1016/j.engstruct.2014.08.016
0141-0296/ 2014 Elsevier Ltd. All rights reserved.
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 119

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

1. Introduction frequency, Bergan and Mollestad [6] proposed an automatic


time-stepping algorithm based on a current characteristic
It is well known that there are two general approaches that can frequency that is estimated using an expression similar to the
be used to analyse the dynamic responses of structural systems. Rayleigh quotient. Their scheme adjusted the time step according
The rst approach employs mode-superposition methods in which to a period measure that is tied to the current stiffness and
the total response is expressed as the sum of individual responses response. To prevent the step size from changing too frequently,
in the various normal modes of vibration. This approach could be they also introduced a tuning function that determines the
useful for nding analytical solutions for the dynamic problem, sensitivity of the time-stepping algorithm.
but it cannot be applied in non-linear systems because of the Zienkiewicz et al. [7] proposed a different approach to time
superposition scheme involved, which typically requires that the step length selection based on a simple expression as an indicator
system remains linear during the process. The second general of local error, which was obtained at little computational cost.
approach that is applicable in the analysis of an arbitrary set of Observing that this indicator does not provide an accurate mea-
nonlinear dynamic equations, which is also effective for addressing sure of the real local error, Zienkiewicz and Xie [8] proposed
coupled linear modal equations, is the numerical step-by-step another method that has a more accurate error estimate. Zeng
direct time integration. These schemes are often used in practice et al. [9] presented an identical error indicator, which has a
because they can be easily applied to solve the cases of both non- slightly simple and intuitive derivation scheme. Li et al. [10] devel-
linear and linear systems numerically and because the analytical oped a new error estimator that was derived from the difference in
solutions for most of the real dynamic problems do not exist or the solutions between an ordinary integration method and
have an excessive computational cost. To compute the numerical another higher-order method, which assumes that the derivatives
solution at specic time ti (where i is an integer that denotes the of the accelerations vary linearly within each time step. The last
number of discrete time steps taken to reach ti), most direct time three methods cited here were applied on the ordinary Newmark
integration methods require the solution to be specied in the [11] time integration process using a posteriori error estimators
previous instant, ti1. Also required is the specied time step size, and strategies that control the time step size by decreasing, main-
Dti = ti  ti1, among other parameter specications for each taining, or increasing it in such a way that the local error fell inside
specic method. or outside the lower and upper limits of a specied tolerance
It is also well known that it is difcult to choose the appropri- interval.
ate time step size for the direct time integration process. Selecting Based on the same approach of local error estimates, Hulbert
a small time step increases the number of evaluations in the and Jang [12] proposed an automatic time-stepping algorithm that
dynamic analysis and provides a more detailed description of was applied to the generalised-a method integration process,
the response; however, selecting a small time step increases the developed by Chung and Hulbert [13], which includes the
computational cost and the time spent on the integration process. Newmark [11], HHT-a [14], and WBZ-a [15] integration methods
The correct choice of the time step size must consider the conict as special cases. Later, Chung et al. [16] developed an a priori error
between the demands of solution accuracy and computational estimator for the generalised-a method, which is computed using
cost (the stability is also important in some cases). The ideal time only information in the previous and current time steps (without
step size maximises the accuracy while minimising the computa- a feedback process, which is required in most conventional a
tional effort. However, the optimal time step may change during posteriori error estimators).
the computation due to changes in the forcing function and/or Recently, Lages et al. [17] introduced a new automatic time
system nonlinearities. Thus, although a xed time step size is typ- step strategy that employs an estimator based on the geometric
ically used for time domain analyses and its choice is frequently indicator of displacement history curvature. The developed esti-
based on intuition and/or experience, the best approach to mator is obtained at little computational effort and is easily
address this parameter is employing an algorithm for automatic applicable to the various existing direct time integration
time step selection. The algorithm should typically seek the larg- methods.
est possible time step while maintaining a prescribed accuracy. It In this paper, three specic automatic time-stepping strategies
should also be efcient and economical and satisfy some specic are explored in more detail and compared with each other. Specif-
criteria. ically, the algorithms of Bergan and Mollestad [6], Hulbert and Jang
Several automatic time-stepping algorithms for dynamic prob- [12], and Lages et al. [17] are discussed when applied to the time
lems have been studied in past research and have been described integration process of the Newmark scheme. These strategies were
in the literature. Some of the early methods were proposed by chosen because they are based on three different concepts: cur-
Hibbitt and Karlsson [1], who developed a strategy based on a rent characteristic frequency, local error estimator, and indicator
measure of the residual force vector computed at the midpoint of curvature, respectively. The reviewed strategies are imple-
of the time interval, and by Oughourlian and Powell [2], who mented with some necessary modications, and the compared
proposed a cheaper method to obtain the mid-step force residual results are presented in numerical examples of linear structures,
through the use of the product between the stiffness matrix and including single-degree-of-freedom and multi-degree-of-freedom
the incremental velocity vector. However, both methods appeared systems.
to have too signicant additional computational cost.
Felippa and Park [3], Park and Underwood [4], and Underwood
and Park [5] proposed an algorithm for step size control that was 2. Basic structural dynamics concepts
applied to the central difference method based on the apparent
highest frequency, which is computed from the highest ratio This section describes the basic concepts of structural dynamics
between incremental accelerations and displacements for all adapted for step-by-step procedures and the equations of the
degrees of freedom. Following the concept of apparent highest Newmark integration method.
120 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

2.1. Dynamic equations Tj


Dt 6 Dtcr p 7
p 1  4b
After spatial discretisation, a structural dynamic system can be
represented by the following matrix system of second-order linear where Tj is the period of natural vibration for the jth normal mode.
ordinary differential equations: This stability condition must be satised for each mode in the
system; thus, the minimum natural period, Tmin, is the critical value.
CtDt
MDt _ KtDt Pt 1 When the incremental dynamic equilibrium Eq. (2) is considered
with Eqs. (5) and (6), the following system of simultaneous linear
where M is the mass matrix, C(t) and K(t) are the viscous damping
equations (which has the form of an incremental pseudostatic
and stiffness matrices at time t (which might be constants in a lin-
_ equilibrium equation) can be obtained:
ear system), Dt, Dt, and Dt are the displacement, velocity, and
acceleration vectors at time t, respectively (superposed dots indi- KDDi DP 8
cate differentiation with respect to time), and P(t) is the load vector
at time t. The initial value problem consists of nding a function where K and DP are the effective stiffness matrix and incremental
that satises (1) for all t 2 [0, tN], tN > 0, and the known initial effective load vector, respectively, given by
conditions: D(0) = D0, D0_ D _ 0.
1 c
The purpose of all numerical direct time integration algorithms K M Ci Ki 9
bDt2i bDt i
is to nd Di, D _ i , and D
i , which are given approximations of D(ti),
_ i , and Dt
Dt i , respectively, in discrete instants ti. The most gen-  
1 _ 1
eral procedure to reach this aim is to consider the following DP DP M Di1 Di1
bDt i 2b
incremental equation in a generic step i on the step-by-step time   
integration [18,19]: c i1 c D
_ i1
Ci  1 Dt i D 10
2b b
i Ci DD_ i Ki DDi DPi
M DD 2
The solution of system (8) provides the displacement
_ i , and DD
i are the displacement, velocity, and accel- increment, and with this result, the velocity increment may be
where DDi, DD
found using the following equation (which does not require the
eration vector increments, respectively, that occur from the begin- i ), which is obtained by combining (5) and (6):
value of D
ning to the end of the time step, DPi = P(ti)  P(ti1), and Ci and Ki
 
are the current incremental (initial tangent) damping and stiffness c c i1  c D_ i1
matrices, respectively, which are assumed to be constants for the
DD_ i DDi 1  Dt i D 11
bDt i 2b b
short increment of time or deformation within the time step for
nonlinear systems analysis. Linear systems can also be repre-
sented by this equation, which becomes simplied when the 3. Automatic time-stepping strategies
matrices Ci and Ki remain constant over the entire analysis. After
the incremental quantities have been determined with some Three adaptive time-stepping procedures will be discussed in
direct integration procedure, the displacement and velocity vec- the following sections, which were chosen because they are based
tors at the end of the time increment can be evaluated by using on three different concepts. First, it is important to dene some
the following: characteristics that make an automatic time-stepping algorithm
efcient, economic, and applicable in real structural dynamic anal-
Di Di1 DDi _ i D_ i1 DD_ i
D 3 yses. Bergan and Mollestad [6] suggested a set of criteria to evalu-
The acceleration at the end of the time step could also be found ate step size control strategies. Briey, these criteria are:
using the increment DD i , but to avoid an accumulation of errors in
a nonlinear analysis, the initial acceleration vector is calculated 1. The optimised time step size should not be inuenced by the
directly from the equilibrium condition at the beginning of the fol- initial step size estimate.
lowing step: 2. The time step should remain constant during a linear stationary
response.
i1 M1 Pi1  Ci D
D _ i1  Ki Di1 4 3. The time step should not be inuenced by the selection of
physical units or by the number of degrees of freedom in the
where the inverse of the mass matrix, M1, is calculated once
dynamic equations.
because it is typically constant.
4. All of the input parameters should be easy to describe.
5. The additional step size selection computational cost should be
2.2. Newmark formulae
as small as possible.
6. The step size should react immediately to sudden changes in a
The integration method of Newmark [11] is one of the most
dynamic response.
popular and widely used time integration algorithms in dynamic
7. The time step should not change without necessity.
analysis. This method uses the following implicit expressions as
assumptions over the displacement and velocity increments:
The last point above has a special importance in the integration
DD_ i 1  cDti D
i1 cDti D
i 5 process. Indeed, a change in the time step of a linear system anal-
ysis means that a new effective stiffness matrix K (given by Eq. (9))
  must be formed and factorised to solve Eq. (8) and to nd DDi,
1
DDi Dti D_ i1 i1 bDt2 D
 b Dt2i D i
i 6 which is an expensive operation. Thus, the time step should be
2
changed only when necessary, and only major changes in the
where c and b are parameters chosen by the analyst to control the response should lead to adjustments in the time step size for linear
stability and accuracy. The method is unstable for c < 1/2 [20]. problems. Too frequent changes in the time step might not intro-
When c = 1/2, the method is unconditionally stable for b P 1/4 duce additional computational cost only in nonlinear analysis, in
[21]. For b < 1/4 and c = 1/2, it is only conditionally stable, and the which the tangent damping and stiffness matrix (Ci and Ki) are
condition for stability is given by the critical time step size [22]: recomputed for almost every step because of nonlinear effects,
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 121

thus changing the matrix K. In this paper, only linear cases are step length does not change unless a major change in the response
compared in numerical examples. The three strategies described has taken place. Outside this plateau, the function f ni is simply 
ni ,
in the following sections have procedures to avoid frequent which makes the setting of the time step in accordance with Eq.
changes in the time step size when it is not necessary. (13) except that a maximum increase has been set at  nm . Bergan
and Mollestad reported that  np should be specied as an input
3.1. Strategy of Bergan and Mollestad parameter (as well as the step-length parameter, k, and the rst
time step size, Dt1) and that typical values of  np would be 1.3 or
The strategy of Bergan and Mollestad [6] proposes the use of an 1.4. The maximum increase in the time step,  nm , can be kept xed
expression that is similar to the Rayleigh quotient, which is dened (not an input parameter), and its value would be 2.0.
as a current characteristic angular frequency, given by the rst of Finally, it is worth noting that, in some steps, the values of DDi
the following equations for step i: could become very small or vanish, such as when passing through
maximum amplitudes. In these situations, it was proposed that x  2i
DDTi Ki DDi 2p could be assigned to its value in the previous step:
 2i
x T i  1=2 12
DDTi MDDi x
 2  2i x
 2i1 when kDDi k < ekDDi1 k
i x 15
Based on this frequency, a corresponding measure of the cur- where e is a xed small number, which is specied to be 0.1. Other
rent characteristic period is calculated by the second of the previ- comments were made in the original work about various aspects of
ous equations. The absolute value of x  2i was taken because it might
the algorithm, but the description given here is necessary for
be negative and T i must be real. The dened frequency x  2i is not
implementation.
any specic natural eigenfrequency, but it reects the incremental
response for all eigenmodes, being an estimate for the dominant
3.2. Strategy of Hulbert and Jang
frequency of the response. The period T i expresses a dominant
increment response period for the total system, and it is tied to
The adaptive strategy of Hulbert and Jang [12] proposes to use
the current stiffness and response, but it is not generally a measure
an algorithm based on the local error estimate approach. The local
of the time elapsed between zero amplitudes. Using this character-
error ei is a measure of the error in the numerical solution for each
istic period, a preliminary estimate of the time increment for the
time step. This error is obtained by the difference: ei Di  Dti ,
next time step is calculated by
where Dt is the solution of the local problem, i.e., it is an analyt-
Dt i1 kT i 13 ical function that satises the following equations:
Ci Dt
MDt _ Ki Dt Pt 16
where the asterisk indicates that this equation is only an estimate
and not the actual new time step size. Here, k is an input constant
that is denoted as the step-length parameter, which is selected Dti1 Di1 _ i1 D_ i1
Dt i1 D
Dt i1 17
according to the desired accuracy and could lie between 0.001 (high Using Taylor series expansions, Hulbert and Jang developed a
accuracy) and 0.1 (low accuracy). A typical value of k would be 0.05, simple expression to dene the local error, which is applied to
which corresponds to 20 time steps for one characteristic period. the generalised-a method integration process. This time integra-
This number is the xed value used in the computing software tion algorithm uses the same Eqs. (5) and (6) of the Newmark
ANSYS 11.0 [23], in which this method is applied. Using the esti- method combined with a modied balance equation, in which
mate Dti1 , the current time-step ratio, 
ni , is dened by the rst two additional parameters are introduced: am and af. When
of the following Eq. (14), and the actual new time increment of am = af = 0, the algorithm reduces to the original Newmark method.
the next time step is calculated by the second of these equations. Therefore, by making these two parameters equal to zero, the local
  error expression developed by Hulbert and Jang, applied to the
ni Dt i1 Dt i1 f ni Dti 14
Dt i Newmark integration process, is given by
  
where f  ni is a tuning function of 
ni that determines the sensitiv-
1 i
ei Dt 2i b  DD 18
ity of the time-stepping algorithm. Bergan and Mollestad discussed 6
some examples of tuning functions in their work and concluded which is identical to the expression presented by Zienkiewicz and
that the best type of such functions has the form of the function Xie [8] and Zeng et al. [9]. This equation is invalid only if b = 1/6
shown in Fig. 1. (ei is equal to zero). However, for this case, an error estimate can
This type of tuning function has a plateau around  ni 1, which be found by using high-order terms in the Taylor series expansion
is dened by a value  np and serves as a protection against making [10]. In this paper, only Eq. (18) will be used to dene the local error
unnecessary adjustments of the time step, what means that the of this strategy.

3.2.1. Normalised error and error tolerance


To specify an appropriate tolerance on the local error for choos-
ing the time step sizes, Hulbert and Jang proposed to relate it to an
input quantity, (Dt/T)target, for which the relationship with the error
tolerance is obtained by considering an unforced, undamped
single-degree-of-freedom model, where T denotes the period of
vibration. To obtain such a relation, the norm of the local error
was scaled using the following equation:
 
kei k 9
RLi sclfaci max kDDi k; sclfaci1 19
sclfaci 10
where sclfac0 = 0. The resulting discussion of this strategy applied to
Fig. 1. Tuning function for the time-stepping algorithm of Bergan and Mollestad. the Newmark scheme demonstrates that given the user specied
122 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

parameter (Dt/T)target, the corresponding tolerance for this norma- Finally, it is important to expose that the user-specied value
lised error can be computed by (Dt/T)target can be used to determine the number lcount. Hulbert
 2   and Jang applied a criterion where this number is dened as the
Dt  1
tol C d C d 2p2 b   20 reciprocal of (Dt/T)target, which can be stated as follows:
T 6 !
1
Common practice dictates that there should be a minimum of lcount int 24
10 time steps per period of the maximum frequency of interest Dt=Ttarget
in the response [12], which suggests that typical values for
(Dt/T)target would be 0.1. 3.2.3. Application to problems with quiescent start conditions
The original work of Hulbert and Jang also suggests another The time-step control strategy described previously cannot be
scheme for the normalised error RLi. However, the expression of used to compute the step size for the rst time increment when
the tolerance in this case was derived numerically by curve tting, the initial conditions and load are zero (quiescent conditions).
and the resulting equation is a function of the high-frequency The normalised error for the rst time step would be independent
dissipation parameter of the generalised-a method, q1, which of Dt for a quiescent start. It can be shown that under such
has no relationship to the Newmark method. For this reason, only conditions, the normalised error is given by
the normalisation scheme of Eq. (19) will be used in this paper.
b  16
RL1 25
3.2.2. Algorithm for automatic time step size control b
After calculating the error tolerance by Eq. (20) and the norma- Since b is an algorithm parameter, the time-step control strategy
lised error RLi for the current interval, the rst procedure of the may not be utilised. Thus, the rst step increment, Dt1, is not
time step control algorithm is to request, at each step, that changed in a quiescent start problem, and this time step size should
lb  tol 6 RLi 6 tol 21 be given as an input parameter.
For the second time step, Dt1 should be used as initial trial step
where lb is a multiplication factor for the lower-bound tolerance, size. Then, if RL2 > tol, the step size is decreased. However, due to
for which a typical xed value would be 0.75. Three situations quiescent start, the value ||DD2||, obtained with the decreased time
can occur from this point: (1) if the condition (21) is satised, then step, may not be large enough to be used as a normalisation factor.
the solution is accepted and the integration process proceeds to the For this reason, the normalisation factor is kept constant in the sec-
next time increment without changing the step size; (2) if ond time step with the value kD02  D1 k, where D02 is the displace-
RLi < lb  tol, then the time step size can increase because the error ment computed for time t2 using the initial step size estimate Dt1.
is considered to be too small. However, this increase in step size
shall occur only for the next time increment, and the solution for 3.3. Strategy of Lages et al.
the current step is taken to be acceptable. Due to the periodic
nature of the local error, the time step length would also be peri- The automatic adaptive strategy proposed by Lages et al. [17] is
odic, and because it is not desirable to change the time step too fre- based on a geometric indicator dened as the curvature of the dis-
quently, a counter cont was introduced to register the number of placement history. Using the denitions of elementary vector cal-
steps in which the situation RLi < lb  tol occurs consecutively. If culus, the parametric representation of a curve in space,
this counter exceeds a specied value, lcount, only then the step described by a given parameter t, can be expressed as a vector
size is increased for the next time increment, in accordance with function rt (Fig. 2). The vector v t a dr=dtjtta 0 is the geo-
the following equations: metric velocity vector of the curve r at time ta, and it is tangent
 1=pinc to the curve at the point r(ta). The corresponding unit tangent vec-
tol
Dti1 f inc  Dt i f inc 22 tor is dened by Tt v t=kv tk, and the arc length function
RLmax
i
s(t) is the integral of the norm ||v(t)|| taken from a specied instant
where pinc is an input parameter that is related to the rate of conver- point t0 to a generic instant t. Based on these concepts, the curva-
gence of the local error and RLmaxi is the maximum value of RL that ture of a curve at time t, which is denoted by j(t), is dened by
has occurred since the counter was reset to zero. The counter is


dT

reset every time that a step size change occurs (an increase or jt

t
26
ds

decrease) or when the condition (21) is satised.


The remaining situation that could occur is: (3) RLi > tol. When Using denition (26), it is possible to show that the curvature
this situation occurs, where the error is computed using a current can be expressed by the following equation:
step size, Dt old
i , then the time increment can be reduced: the current
q
solution is discarded, and a new solution is computed using a smal- v  v a  a  v  a2
ler step size, Dtnew . Two cases can occur when this step size reduc- jt 27
i v  v 3=2
tion is required. In the rst case, if Dt was increased in the previous
step, then the new trial Dt is taken as its value before the step size
increase has occurred. In the other case, i.e., if the step size did not
increase in the previous step (e.g., if it did not change or decreased),
then the new step size is computed using the following equations:
 1=pdec
tol
Dtnew
i f dec  Dtold
i f dec 23
RLi
where pdec has the same denition as pinc and the xed value used
for these parameters was pinc = pdec = 2. As cited before, the solution
at time t i t i1 Dt old
i is discarded, and a new solution is obtained
by using Dt new
i , thereby dening a new value of RLi and restarting
the adaptive strategy for that step. Fig. 2. Parametric representation of a curve in space.
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 123

where a(t) = dv/dt = d2r/dt2 is the geometric acceleration vector


and the symbol  indicates the dot product of two vectors. The
benet of using Eq. (27) is related to the computational economy,
which is the reason why it is recommended. The function jt gives
a positive measure about the geometry of rt that indicates how
much it is curved. The curvature will be close to zero at points
where rt is similar to a straight line. At the other points, jt will
Fig. 3. Intervals of curvature regularisation.
increase or decrease in accordance with the deviation of the func-
tion rt from a straight line. This information is simple, but it can
be used in an efcient time adaptive strategy. 3.3.2. Curvature regularisation
The value of the time increment obtained by Eq. (30) or Eq. (31)
should not be used in the next step solutions without some neces-
3.3.1. Indicator of curvature and time step strategy sary adjustments. In many cases, the values of ji could present a
Let r(t) describe the displacement history of a structure by the highly oscillatory behaviour, which causes many changes in the
rst of the following equations. Thereby, r(t) will be a vector of time step size, and as discussed earlier, this scenario is not typi-
n + 1 elements, where n is the number of degrees of freedom in cally desirable. To prevent this situation from occurring, Lages
the structure. Thus, v(t) and a(t) will be given by the corresponding et al. introduced a method of curvature regularisation, which
temporal derivatives: makes the curvature values change in regular plateaus during the
      step-by-step analysis.
t 1 0 Some techniques of curvature regularisation were suggested,
rt vt _ at 28
Dt Dt Dt but the most appropriate technique was the strategy based on
the maximum interval value: during numerical integration, for
The indicator of displacement history curvature should then be
each time instant, the curvature value is investigated and com-
obtained by an expression that is similar to Eq. (27). Using the vec-
pared with the maximum value that occurs in the interval that cor-
tor functions (28), the curvature should be given by (for a discrete
responds to that instant; the highest value among them is chosen
time ti)
to be used in time step computation with Eq. (31). It was suggested
q that the regularisation interval (Dtreg) be a function of the critical
1 D _iD _ i D i  D i  D _iD i 2 time step size (Dtcr), which is a constant quantity.
ji 3=2
29 The next gures present the scope of the regularisation tech-
1 D_ i  D_ i
nique. Fig. 3 illustrates the step iterations, and Fig. 4 illustrates
This curvature could be used to control the time step size. As the programming process used in this paper.
cited before, ji will be small when the graphs of D are similar to
a straight line and larger when the graphs are more curved. It is
well known that plotting a response history in locations that are
similar to a straight line requires few points to describe the graph,
whereas more points are required in parts that have greater curva-
ture. This fact suggests that a smaller step size should be used in
parts of the response history that have a greater curvature, where
a more detailed solution is needed to guarantee the accuracy. Sim-
ilarly, a larger step size can be used in parts that have a smaller
curvature.
Because the curvature is a positive measure, some relationships
between it and the time step size can be established. Lages et al.
[17] suggested a simple exponential expression to nd the time
step for the next increment, which is given by

Dt i1 Dtmax expcp ji 30

where cp is a positive constant and Dtmax is an upper value to the


time step size, which can be set equal to the critical value in condi-
tionally stable integration methods. This paper proposes another
equation to relate the step size and curvature, which has the form
Fig. 4. Algorithm for regularisation in a maximum interval value, where: Dt reg )
Dt max Regularisation interval, ct ) Positive constant, max jt 0j ; ti ) Maximum value of
Dt i1 31 ji between times t0j and ti , a ) Weighting constant.
1 cp ji
In Eqs. (30) and (31), the time step varies from an upper speci-
ed value (Dtmax) to zero in the limit case when the curvature is
innity. Eq. (31) appeared to be more appropriate than Eq. (30)
because the exponential function varies more than this ratio form
for the same variation in curvature. Moreover, the exponential
form yields a zero time step for a curvature value that is not overly
high, i.e., the number exp(ji) is close to innity in numerical calcu-
lations when ji is close to 700, which causes a zero time step size
in Eq. (30). For this reason, Eq. (31) was then chosen for use in the
numerical examples of this paper. Fig. 5. Single-degree-of-freedom system.
124 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

Table 1 displacements; a simply supported beam with a harmonic load;


Set of parameters. and a plane frame structure composed of three bars with different
(1) (2) applied loads.
m 1.0 1.0
k 1.0 p2/4 4.1. Single-degree-of-freedom systems
c 0.6 0
D0 0 1.0
The spring-mass system shown in Fig. 5 is analysed with two
D_ 0 0 0
p 1.0 0
different sets of parameters, which are given in Table 1. The rst
set of parameters is a forced vibration system, whereas the second
is an undamped free vibration system. The analytical solutions for
both cases are well known from the elementary theory of mechan-
The presented algorithm refers to the maximum curvature ical vibration.
value that has occurred in the previous time interval (jreg i1 on the For the Newmark integration process, c = 1/2 was taken in both
eighth line of Fig. 4). This process is used to minimise time step cases. The value of b was chosen as 1/8 and 1/5 for the rst and sec-
changes in the rst instants of the regularisation interval, Dtreg, ond sets of parameters, respectively. For the damped system, the
where the maximum value between the times t0j and ti, rst time step was chosen as Dt1 = 0.5 s, the value of (Dt/T)target
max jt0j ; t i , does not accurately represent the behaviour of in the strategy of Hulbert and Jang was taken as 1/10, and the
function j(t) when ti is close to t 0j , which could cause undesired parameters cp, ct, a for the method of Lages et al. were set equal
oscillating changes in the time step size. to 40, 1.0, and 0.2, respectively. For the undamped system, the rst
This technique also guarantees that the regularised curvature time step was chosen as Dt1 = 0.01 s, the value of (Dt/T)target was
values are always greater than the non-regularised values, which chosen as 1/100, and the parameters cp, ct, and a were set to 40,
is convenient. The values of the parameters a, ct, and cp will be cho- 1.0, and 1.0, respectively. The algorithm of Bergan and Mollestad
sen for each case separately because the values suggested in the is not compared here because it does not apply to single-degree-
original work that proposes this strategy did not yield reasonable of-freedom systems. The value Dtmax used in the method of Lages
results when applied in the dynamic problems of frame structures, et al. was taken to be equalpto the critical time step
which is treated in this paper. p (Dtcr) in both
cases: this increment is 2 2 2:8284 s and 4 5=p 2:8470 s
The rst time step for problems with quiescent start conditions, for the rst and second cases, respectively.
Dt1, could also not be changed because the initial curvature, j0, is To compare the accuracy of the numerical solutions, the abso-
zero in such cases, which would make the rst time step be set lute error (er) was calculated from the difference between the
equal to the maximum value, Dtmax. Normally, for transient numerical response ui and the exact solution uex(ti) found by ana-
problems, the most important dynamic events occur at the very lytical equations. The absolute error history is shown in Fig. 6(a),
beginning; as a result, it is not convenient to choose an excessively and the time step size histories for the different adaptive strategies
long rst time step. Therefore, this rst step size should be given as are shown in Fig. 6(b), using the rst case for the parameters.
an input parameter in quiescent start problems in the same way as The numerical error is greater when a constant time step is used
in the strategy of Hulbert and Jang. compared to when the adaptive methods are applied on the inte-
gration process. The highest values of this error occurred in the
4. Numerical examples rst stages of the analysis, which are exactly when the most impor-
tant dynamic events occur, before the solution is dissipated by
This section presents a comparison of the results that are given damping.
by the three previous related strategies when applied to simple Both strategies of Hulbert and Jang and Lages et al. start by
numerical analyses of dynamic structural systems. The same inte- adjusting the rst time increment, Dt1, in accordance with the local
gration process (Newmark) was used with the three techniques. error tolerance and initial curvature, respectively.
The rst example studies the model of a single-degree-of-freedom The strategy of Lages et al. yields the smallest error at the begin-
system with two sets of parameters. Then, three different ning, which grows as the solution advances, becoming greater than
multi-degree-of-freedom problems are presented: a shear building the other errors in the nal instants. The strategy of Hulbert and
structure that assumes a lumped mass system with only lateral Jang presents a greater error at the beginning. This behaviour is

0.012 2.5
Constant t Constant t
Hulbert & Jang Hulbert & Jang
0.010 Lages et al. Lages et al.
2
Time step size history (t)

0.008
1.5
Error (er)

0.006

1
0.004

0.5
0.002

0.000 0
0 4 8 12 16 20 0 4 8 12 16 20
Time (t) Time (t)
(a) (b)
Fig. 6. Absolute error and time step size history for the rst set of parameters.
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 125

0.0012 0.018 1.000

0.900
Norm of the local error history (|| e ||)

Non-Regularized Curvature

Normalized local error history (RL)


|| e ||
0.001 RL 0.015 Regularized Curvature
0.800

0.700
0.0008 0.012

Curvature ()
0.600

0.0006 0.009 0.500

0.400
0.0004 0.006
0.300

0.200
0.0002 0.003
0.100

0 0 0.000
0 4 8 12 16 20 0 4 8 12 16 20
Time (t) Time (t)
(a) (b)
Fig. 7. Normalised local error history and curvature history for the rst set of parameters.

0.06 3.0
Non-Regularized Curvature
Constant t Regularized Curvature
Hulbert & Jang
0.05 2.5
Lages et al.
Time step size history ( t)

0.04 2.0
Curvature ()

0.03 1.5

0.02 1.0

0.01 0.5

0 0.0
0 5 10 15 20 0 5 10 15 20
Time (t) Time (t)
(a) (b)
Fig. 8. Time step size history and curvature history for the second set of parameters.

-4 0.007
-5
1x10 1.8x10 Constant t
|| e ||
Normalized local error history (RL)

Hulbert & Jang


Norm of the local error history (|| e ||)

0.006 Lages et al.


-5 RL -4
1x10 1.5x10
0.005
-6 -4
8x10 1.2x10
Error (er)

0.004
-6 -5
6x10 9.0x10
0.003
-6 -5
4x10 6.0x10
0.002

-6 -5
2x10 3.0x10 0.001

0 0
0x10 0.0x10 0
0 5 10 15 20 0 5 10 15 20
Time (t) Time (t)
(a) (b)
Fig. 9. Normalised local error and absolute error for the second set of parameters.

in accordance with the variation in the time steps. As shown in before it is increased again, searching for stabilisation at a target
Fig. 6(b), for the strategy of Hulbert and Jang, the time steps start value given by the input (Dt/T)target, which is approximately
at close to half of the initial value (Dt1) and tend to increase as Dt = 0.628 s. In many steps, the normalised error RL lies below
the analysis advances, which is in accordance with the behaviour the lower tolerance limit lb  tol = 0.0123 in condition (21), as
of the normalised local error. For a short interval of time, the value shown in Fig. 7(a). This observation explains why the time step
of Dt has been set to be equal to the initial estimate (Dt = 0.5 s) increases with time. Fig. 7(a) also illustrates that the error RL never
126 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

Table 2 Table 3
Summary of results of single-degree-of-freedom systems. Set of parameters.

Set T Dtcrit Hulbert and Jang Lages et al. (1) (2)


Dt min Dtmax NTOT Dtmin Dtmax NTOT m1 150 t 180 t
m2 150 t 270 t
(1) 6.283 2.828 0.2782 0.8062 47 0.0690 2.2898 85
m3 150 t 360 t
(2) 4.000 2.847 0.0100 0.0400 576 0.0286 0.0286 701
k1 100,000 kN/m 105,000 kN/m
k2 200,000 kN/m 210,000 kN/m
k3 300,000 kN/m 315,000 kN/m

exceeds the tolerance tol = 0.0164 that is predicted by the algo-


rithm. This graph still shows the history of the local error norm,
e, which is used in the strategy of Hulbert and Jang (the left vertical
axis shows the values of e and the right axis shows the values of
RL).
Still observing Fig. 6(b), the time steps of the strategy of Lages
et al. are initially considerably lower than the others, becoming
greater when the solution advances over time. This pattern occurs
because the curvature has a high initial value, which reduces sub-
stantially as the response is dissipated by the damping. Fig. 7(b)
presents the curvature history, with and without regularisation,
Fig. 11. Applied loads on the shear building systems.
applied in the strategy of Lages et al. The regularised curvature his-
tory in a constant plateau is very sensitive to changes in the param-
eter a, which was taken as 0.2 in this case. The absolute errors for this set of parameters are shown in
The analysis of the second case also illustrates some disadvan- Fig. 9(b). The strategy of Hulbert and Jang provided the largest
tages of the strategy of Lages et al. in this type of system. Fig. 8(a) error, whereas the method of Lages et al. assumed smaller error
presents the time step size history for the second set of parameters. values. These results are in accordance with the time step size
The time steps in the strategy of Lages et al. remained constant behaviour, where the strategy that used larger values of Dt reached
and were equal to the step obtained with the maximum curvature, larger errors.
i.e., Dt = Dtcr/(1 + cpjmax). This pattern occurred because of the Table 2 summarizes the results of analyzes of the two single-
choice a = 1.0 in the curvature regularisation, which makes the reg- degree-of-freedom systems with respect to maximum and mini-
ularised curvature assume a constant value equal to the maximum mum time increments found with adaptive strategies, and the total
obtained from the displacement history, which is reached at the number of points (NTOT) plotted with the different algorithms. The
beginning, as shown in Fig. 8(b). This choice was made because higher the number NTOT, the more expensive is the strategy and the
the actual curvature assumes a permanent oscillatory behaviour, more detailed is the dynamic response. It is concluded that the
which results from the system being in an undamped free vibra- strategy of Lages et al. provided more points in both analyzes.
tion (c = 0). Unless the regularisation interval was taken to be equal
to half of the natural period, no value of a between zero and one
would yield a reasonable behaviour of the regularised curvature 4.2. Shear building with three oors
that would transfer this permanent oscillation to the time step
sizes. To illustrate the solution of a dynamic analysis in a simple
In the strategy of Hulbert and Jang, the time step was adjusted multi-degree-of-freedom system, the structure will be considered,
rapidly to reach the constant value given by the target ratio in this example, to be a shear building with three oors. In this
(Dt/T)target. On the rst steps, the increment is relatively small, type of structure, the supporting columns provide only horizontal
and the computed error RL is less than the lower bound of the displacements in each oor. Thus, the shear building has only three
tolerance; as a result, the step size remains small until the counter degrees of freedom, as shown in Fig. 10.
cont reaches the value lcount = 1/(Dt/T)target = 100, when the time The matrix M of this structure is obtained by the lumped mass
step increases to the permanent value Dt = 0.04 s. This behaviour model, where the masses of the oors are disposed in the rst diag-
can also be seen in Fig. 9(a), which provides the error history of the onal of the matrix. Two cases of shear buildings were tested in this
algorithm. After reaching the permanent time step size, the norm example using the parameters provided in Table 3.
of the local error (||e||) becomes oscillatory, and the normalised In the rst case, there is a transient horizontal load that acts on
error (RL) is controlled within the tolerance interval, assuming the lower oor with a constant value for 0.38 s, becoming zero
small values periodically for a few steps and becoming equal to after that instant, as shown in Fig. 11(a). In the second case, there
the upper bound value tol = 1.3  104 in almost every step. is a horizontal load acting on the upper oor, with a triangular
impulsive load history that has a duration of 0.05 s, which is illus-
trated in Fig. 11(b). With this information, the load vectors for the
two cases can be given, respectively, by

pt 0 0 p3 t T and pt p1 t 0 0 T 32

When solving the linear analysis of these systems, we consid-


ered Rayleighs proportional damping with damping ratios of 15%
in the 1st and 3rd modes of vibration for the rst case and of 5%
for the second case, also in the 1st and 3rd modes.
For the rst case, the mass matrix, stiffness matrix, and natural
eigenfrequencies are formed and found, giving the following
Fig. 10. Shear building model with three oors. results (in tons, kN/m and rad/s, respectively):
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 127

0.015 0.025
First Floor Bergan & Mollestad
Second Floor Hulbert & Jang
Lages et al.

Time step size history - t (s)


Third Floor 0.02
0.01
Displacement (m)

0.015

0.005

0.01

0
0.005

-0.005 0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Time (s) Time (s)
(a) (b)
Fig. 12. Dynamic response history and variation of Dt for the rst case of the shear building.

Table 4
procedures discussed earlier, for which the rst time step was cho-
Parameters of the adaptive strategies used in the shear building examples.
sen as Dt1 = 0.01 s in the three strategies. All of the algorithms pre-
Case Bergan and Mollestad Hulbert and Jang Lages et al. dicted an increase in the time step size value during the rst 0.35 s,
k 
np (Dt/T)target cp ct a sharply reducing it near time t = 0.4 s and increasing it again in the
(1) 0.05 1.4 1/10 2 1.0 0.6
subsequent moments. Because the value of RL1 is greater than the
(2) 0.05 1.3 1/20 25 2.0 0.5 bound tol, the strategy of Hulbert and Jang initiates the integration
process by adjusting the initial time step, and the same is done in
2 3 2 3 2 3 the strategy of Lages et al. The algorithm of Bergan and Mollestad
150 0 0 1 1 0 16:65 uses the given value Dt1 in the rst step integration and reduces it
6 7 6 7 6 7
m4 0 150 0 5 k 1  105 4 1 3 2 5 x 4 39:11 5 immediately afterward. We observed a slight tendency of the time
0 0 150 0 2 5 64:76 step to reach the value Dt = 0.015 s at the end of all of the analyses.
The strategy of Lages et al. provided the lowest values of Dt,
33
whereas the other strategies provided values that were closer to
The damping matrix c can then be formed with Rayleighs each other. Fig. 13(a) presents the curvature sequences used in
damping using the damping ratios, eigenfrequencies, and matrices the strategy of Lages et al.
m and k. The displacement responses for the three oors, obtained Fig. 13(b) presents the variations of the characteristic period T
from the dynamic analysis using a constant time step Dt = 0.001 s, calculated with the strategy of Bergan and Mollestad, as well as
are shown in Fig. 12(a). The parameters of the numerical integra- the ratio Dt=k, which has plateaus of constant values. The three
tion procedure were set equal to c = 1/2 and b = 1/8, and the horizontal traced lines represent the values of the natural modal
parameters for the different adaptive strategies are provided in periods of vibration of the structure, which are equal to
Table 4. T1 = 0.377 s, T2 = 0.161 s, and T3 = 0.097 s. The values of T vary
The value Dtmax of the method of Lages et al. was set equal to between the values of the fundamental period T1 and the mode
the critical time step, Dtcr, in both of the p cases period of the highest frequency, T3, passing through the value T2.
on the shear build-
ing. In the rst case, this value is Dtcr T J 2=p. Fig. 12(b) presents To conrm that the algorithms are independent of the initial
the variation in the time step sizes with all of the adaptive step size Dt1, the same analysis is conducted with a different value

15.0 0.4
Non-Regularized Curvature
Regularized Curvature
(s) T

12.0
Current characteristic period -

0.3
(1/m)

9.0

0.2
Curvature -

6.0

0.1
3.0
T
t /
0.0 0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Time (s) Time (s)
(a) (b)
Fig. 13. Curvatures and characteristic periods for the rst case of the shear building.
128 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

Hulbert & Jang


Bergan & Mollestad 0.025
0.025 t1 = 0.01
t1 = 0.01
t1 = 0.001
t1 = 0.001

Time step size history - t (s)


0.02
Time step size history - t (s)

0.02

0.015
0.015

0.01 0.01

0.005 0.005

0 0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Time (s) Time (s)
(a) (b)
Fig. 14. Variation in the time step size using different values of Dt1.

0.035
Bergan & Mollestad
Hulbert & Jang 8x10
-7
4.00x10-3
0.03
Lages et al. || e ||
Time step size history - t (s)

Norm of the local error history (|| e ||)

RL
0.025
6x10
-7
3.00x10-3

0.02

-7 -3
0.015 4x10 2.00x10

0.01
-7 -3
2x10 10

0.005

0 0x10
0
0.00x100
0.0 0.5 1.0 1.5 2. 0.0 0.5 1.0 1.5 2.0
Time (s) Time (s)
(a) (b)
Fig. 15. Time step size history and normalised error for the second case of the shear building.

1.8 0.5
Non-Regularized Curvature
1.6
Regularized Curvature
Current characteristic period - T (s)

0.4
1.4
Curvature - (1/m)

1.2
0.3
1.0

0.8
0.2
0.6

0.4
0.1
T
0.2
t /
0.0 0
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
Time (s) Time (s)
(a) (b)
Fig. 16. Curvatures and characteristic periods for the second case of the shear building.

Dt1 = 0.001 s using the strategies of Bergan and Mollestad and Hul- Lages et al. was not compared because it does not employ the value
bert and Jang. The compared results are shown in Fig. 14. The Dt1 unless the initial start conditions are quiescent.
behaviour of the values of Dt are generally insensitive toward Fig. 14 indicates that the initial optimised time step would be
selection of the initial time step in both methods. The strategy of between 0.005 s and 0.01 s in both strategies. This optimal time
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 129

oscillatory until approximately the same value was reached. This


number is 1/20 of the average value T = 0.3 s, and the characteristic
period of Bergan and Mollestad oscillates about this value, as
shown in Fig. 16(b) (this result is compatible with the choice
k = (Dt/T)target = 1/20). Fig. 16(b) also indicates that the values of
T vary in the region dened by the fundamental period
T1 = 0.439 s, and the period of the second mode T2 = 0.205 s, oscil-
lating around the average value, which is close to 0.3 s. In the
method of Lages et al., the time step starts smaller and increases
in most parts of the analysis, following the inverse behaviour of
the regularised curvature shown in Fig. 16(a).

Fig. 17. Model of a simply supported beam under sinusoidal load.


4.3. Simply supported beam

A simply supported beam is excited in transverse motion with a


increment is reached in the method of Bergan and Mollestad harmonic point load acting at the midpoint of its length. Fig. 17
immediately after the rst step with the two choices of Dt1. In provides the problem parameters and geometry. The beam is mod-
the strategy of Hulbert and Jang, the initial value Dt1 = 0.01 s is elled by 12 equal beam elements over its length. A Rayleigh pro-
adjusted before the analysis starts, as was noted before, but the portional damping is introduced with damping ratios in the rst
step Dt1 = 0.001 s remains with its small value in the rst steps and third modes of vibration set equal to 10% of the critical.
until the value of RL stays below the limit lb  tol during lcount Two cases were studied using different forcing frequencies and
steps. Despite this different initial behaviour, the adjusted time amplitudes. In the rst case, the forcing frequency was equal to the
step sizes stayed closer in the algorithm of Hulbert and Jang than fundamental frequency x1 = 2p rad/s and the loading amplitude
in the strategy of Bergan and Mollestad. was taken as p0 = 450 N; in the second case, the forcing frequency
The procedure for dening the structural matrices and vectors was equal to the third mode frequency x3 = 18p rad/s and the
are identical when solving the second case of the Shear Building, amplitude was chosen as p0 = 4,500 N.
in which are applied the load p1(t) on the upper oor. The param- The beam was analysed with the Newmark procedure, choosing
eters for the integration method were chosen as c = 1/2 and b = 1/5. c = 1/2 and b = 1/4 and the initial time step Dt1 equal to 0.001 s.
The results are shown in Figs. 15 and 16. The parameters for the The parameters k and  np were set equal to 0.05 and 1.3,
different adaptive strategies are provided in Table 4. respectively, for the strategy of Bergan and Mollestad. The value
This case is a problem that has quiescent start conditions, and (Dt/T)target was taken as 0.05 for the algorithm of Hulbert and
for this reason, all of the algorithms employ the initial increment Jang. The parameters cp, ct, and a for the method of Lages et al.
Dt1 = 0.001 s in the integration of the rst step. The error RL1 in were set equal to 1.0, 1.0, and 0.9, respectively, and the
the strategy of Hulbert and Jang is not used on the adaptive proce- maximum value Dtmax = 0.1 s was employed, which corresponds
dure and has a value that exceeds the limits of Fig. 15(b), which are to 1/10 of the fundamental period T1 = 1 s. The regularisation
obtained directly from Eq. (25). The initial curvature used in the interval was calculated as Dt reg ct Dtmax because Dtcr = + 1 in
method of Lages et al. is zero, and for this reason, the maximum this case. Along with the above-mentioned adaptive strategies,
time increment (Dtmax) is avoided in the rst step by using the the same transient analysis was executed with the computer
user-specied value Dt1. However, this procedure did not avoid program ANSYS 11.0 [23], applying the automatic time-stepping
the small value of the curvature on the second step, which made procedure of the program. Fig. 18 presents the obtained results
the increment Dt2 assume a relatively high value. for the rst mode-forcing frequency case.
In the strategy of Bergan and Mollestad, the time step size oscil- The responses are almost identical for all of the analysis. In the
lated close to the value Dt = 0.015 s. In the algorithm of Hulbert strategy of Bergan and Mollestad, the parameter k corresponds to
and Jang, the increment increased in stages that were not very an ideal time step of 5% of the fundamental period. Fig. 18(b)

Displacement history of the central node Time step size history


0.08 0.10
Constant t
0.09 Ansys
0.06
Bergan & Mollestad
Time step size history - t (s)

0.08 Hulbert & Jang


0.04 Lages et al.
Displacements (m)

0.07
0.02
0.06

0.00 0.05

0.04
-0.02
0.03
-0.04 Constant t
Ansys 0.02
Bergan & Mollestad
-0.06
Hulbert & Jang 0.01
Lages et al.
-0.08 0.00
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)
(a) (b)
Fig. 18. Results of the beam analysis for the rst case of loading.
130 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

Curvatures for Curvatures for


7.00 7.00
Non-Regularized Curvature Non-Regularized Curvature
Regularized Curvature Regularized Curvature
6.00 6.00

5.00 5.00

Curvature - (1/m)
Curvature - (1/m)

4.00 4.00

3.00 3.00

2.00 2.00

1.00 1.00

0.00 0.00
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)
(a) (b)
Fig. 19. Variation in the curvature for different values of a.

Lages et al. Bergan and Mollestad


0.10 0.10
= 0.9 t1 = 0.001
0.09
= 1.0 t1 = 0.1
Time step size history - t (s)

0.08 0.08
Time step size history - t (s)

0.07

0.06 0.06

0.05

0.04 0.04

0.03

0.02 0.02

0.01

0.00 0.00
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)
(a) (b)
Fig. 20. Time step size history for different values of the input data.

Hulbert and Jang Ansys 11.0


0.06 0.14
t1 = 0.001 t1 = 0.001
0.13
t1 = 0.01
t1 = 0.05 0.12
0.05 t1 = 0.1
Time step size history - t (s)

Time step size history - t (s)

0.11
0.10
0.04
0.09
0.08
0.03 0.07
0.06
0.05
0.02
0.04
0.03
0.01
0.02
0.01
0.00 0.00
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)
(a) (b)
Fig. 21. Time step size history for different values of the input data.
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 131

0.049
Constant t 0.040
0.042 Ansys Constant t
Bergan & Mollestad Ansys
0.035 0.035
Hulbert & Jang
Bergan & Mollestad

Time step size history - t (s)


Lages et al.
0.028 Hulbert & Jang
0.030
Displacements (m)

Lages et al.
0.021

0.014 0.025

0.007 0.020
0.000
0.015
-0.007

-0.014 0.010
-0.021
0.005
-0.028

-0.035 0.000
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)
(a) (b)
Fig. 22. Results of the beam analysis for the second loading case.

illustrates that for this strategy, the time step increases rapidly step is less oscillatory and decreases in constant plateaus. In gen-
from the initial low value of 0.001 s up to nearly 0.05 s and has a eral, the value a = 1.0 should be chosen in problems where the
brief fall that is close to the time t = 0.5 s before a new increase peak values of the curvature grow over time, such as in the last
to an unchanged value close to the ideal value. In the strategy of example. An alternative choice to this procedure would be to take
Hulbert and Jang, the target value given by (Dt/T)target is the same a value of ct that makes the regularisation interval Dtreg close to the
as the method of Bergan and Mollestad, which was almost reached oscillation period of the curvature, which is difcult to predict.
after some intermediary plateaus and remained stable. The time The same analysis of this loading case was conducted with the
step history obtained with ANSYS 11.0 is close the history given strategies of Bergan and Mollestad and Hulbert and Jang, taking
by the method of Bergan and Mollestad. The same behaviour different values for the initial value Dt1 equal to 0.1 s and
shown in the previous example with the strategy of Lages et al. 0.05 s, respectively. The comparison results for variations in the
in quiescent start conditions can be observed in this case: the size time step are shown in Figs. 20(b) and 21(a). In the strategy of
of the second time step (Dt2) is close to the maximum value Bergan and Mollestad, the time step decreases rapidly from the
Dtmax = 0.1 s due to an initial curvature that is close to zero; how- high initial value to the same value as in Fig. 18(b), assuming a
ever, the nal sizes of the time increment oscillated in values that similar behaviour. In the strategy of Hulbert and Jang, the initial
were much smaller than those obtained with the other strategies. step size decreased after the rst step due to a quiescent start,
Fig. 19(a) illustrates the variation in the curvature for this analysis, and the time steps reached the target value after some interme-
with and without regularisation, illustrating that the result of jreg diary plateaus, remaining stable. This result proves that both
is very oscillatory between local peak values. This behaviour leads strategies are not dependent on the choice of the initial time step.
to an undesirable oscillation in the time step sizes, which is The variation in the time step was also analysed in the program
eliminated by taking a = 1.0 in the adaptive procedure, leading to ANSYS 11.0, with different initial values Dt1 = 0.01 s and
the curvature history of Fig. 19(b). Dt1 = 0.1 s, for which the results are shown in Fig. 21(b). How-
Fig. 20(a) presents a comparison between the time step size ever, in this case, the time step behaviour changed with the
histories obtained with this adaptive strategy while choosing the choice of the initial step size, and the software made no changes
different values a = 0.9 and a = 1.0. In the second case, the time in the value Dt1 = 0.1 s.

Curvatures for Curvatures for


90.00 90.00
Non-Regularized Curvature Non-Regularized Curvature
80.00 Regularized Curvature 80.00 Regularized Curvature

70.00 70.00
Curvature - (1/m)

Curvature - (1/m)

60.00 60.00

50.00 50.00

40.00 40.00

30.00 30.00

20.00 20.00

10.00 10.00

0.00 0.00
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time (s) Time (s)
(a) (b)
Fig. 23. Variations in the curvature for different values of a.
132 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

Fig. 24. Plane frame steel structure.

Fig. 25. Loading and models of the plane frame structure.

The results for the third mode forcing frequency case are shown nearly constant value of 0.0015 s. This behaviour can be xed by
in Fig. 22. Fig. 22(a) shows the response history at the midpoint of taking a = 1.0, which leads to the variation shown in Fig. 23(b).
the beams length. The transient part of the vibration is gradually In the same manner, the last example illustrates that the value
damped out and the particular part of the solution dominates the a = 1.0 should be chosen when the average behaviour of the peak
response. The characteristic period calculated by Bergan and Mol- curvatures are increasing. As noted before, this procedure is useful
lestad would become nearly equal to the period of the loading in linear analyses, where an oscillating time step can increase the
function, which is 1/9 = 0.111. . .s. In both strategies of Bergan computational cost.
and Mollestad and Hulbert and Jang the ideal time step size would
assume nearly 5% of this period, what is successfully obtained in 4.4. A three-bar plane frame structure
the second strategy, as shown in Fig. 22(b). The result using algo-
rithm of Bergan and Mollestad is very oscillatory, but it remained The following example illustrates the application of the
close to the ideal plateau in most part of the analysis. The time reviewed time adaptive techniques to a problem of a plane frame
steps obtained with ANSYS 11.0 stayed nearly close to 0.01 s, but steel structure, with the geometric and physical properties shown
assumed so high values in some moments, that may have caused in Fig. 24.
a loss of precision in the response seen in Fig. 22(a), compared with The objective is to nd the lateral displacement history Dx(t) of
the others methods. the top-right beamcolumn joint node, as illustrated in the last
In the strategy of Lages et al., the curvature increases and oscil- gure. Some cases of loading were considered. In the rst case, a
lates considerably, as shown in Fig. 23(a). The regularised curva- lateral transient load F1(t) is applied at the same target node, with
ture with a = 0.9 yields an oscillating time step that is close to a the loading function illustrated in Fig. 25(a). To compare the results

Non-subdivided model Subdivided model


0.011 0.022
Bergan & Mollestad Bergan & Mollestad
0.021 Hulbert & Jang
Hulbert & Jang
Lages et al. Lages et al.
0.010 0.020
Time step size history - t (s)

Time step size history - t (s)

Ansys Ansys

0.004 0.004

0.003 0.003

0.002 0.002

0.001 0.001

0.000 0.000
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Time (s) Time (s)
(a) (b)
Fig. 26. Time step size histories.
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 133

Non-subdivided model Subdivided model


4000 400000
Non-Regularized Curvature Non-Regularized Curvature
Regularized Curvature Regularized Curvature

3000 300000
Curvature - (1/m)

Curvature - (1/m)
2000 200000

1000 100000

0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Time (s) Time (s)
(a) (b)
Fig. 27. Curvature history for both models of the frame structure.

3000
Non subdivided Model
Subdivided Model
Curvature - (1/m)

2000

Fig. 29. Loading and model of the plane frame structure.

1000

0
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Time (s)

Fig. 28. Curvatures for the frame structures from time t = 0.05 s.
Fig. 30. First multiple-loading case.
obtained with different nite-element beam meshes, the analysis
is conducted on two different models, which are illustrated in
Fig. 25(b) and (c). In the rst model, there is no internal element
division over the bars, which are described only by the initial
and nal nodes, yielding only four nodes and three bars in the
entire structure. In the second model, the nite element mesh
was rened in 24 equal beam elements of length equal to 0.5 m,
yielding a total of 25 nodes.
Because the support nodes of the structure are completely
restrained, the model of Fig. 25(b) has six degrees of freedom
Fig. 31. Second multiple-loading case.
and the model of Fig. 25(c) has 69 degrees of freedom. The Rayleigh
proportional damping is introduced with damping ratios that are
set to 5% of the critical amount in the rst and third natural modes displacement Dx appear to be identical in all of the analyses and are
of vibration in both models. not shown.
The rst initial step in the dynamic analysis was chosen as The dominant period T calculated by Bergan and Mollestad
Dt1 = 0.001 s. np and k were set equal to 1.4 and 0.05, respectively. stayed close to the fundamental period T1, yielding an ideal time
(Dt/T)target was set equal to 0.1. The parameters ct and a were set step size for this strategy that is equal to 1/20 of T1, which is close
equal to 1.0 and 0.7, respectively, and the value Dtmax in the to 0.00175 s. The target time step of Hulbert and Jang was close to
algorithm of Lages et al. was chosen as 0.03 s, which is close to twice that value because (Dt/T)target is twice k. ANSYS 11.0 yielded
the fundamental period T1  0.0346 s. The parameter cp was set time step size values that are between the two plateaus.
equal to 0.1 and 0.001 in the rst and second models, respectively. When applying the strategy of Lages et al., a characteristic
The Newmark integration process was applied with c = 1/2 and situation of multi-degree-of-freedom systems has occurred, as
b = 1/4, and in the method of Lages et al., the expression cited in Section 3.3.1: the curvature takes such high values that
Dtreg = ctDtmax was used again because there is no Dtcr in this case. the time step would be very small if the parameter cp did not
The step size history results obtained for the divided and non- reduce the value of j. The curvature value at the initial time step
divided models are shown in Fig. 26. The responses for the lateral increases with an increasing number of degrees of freedom, which
134 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

justies the choice of cp = 0.1 in the non-divided bars case and considered, which are given by F2 and F3, as illustrated in Figs. 30
cp = 0.001 in the divided case because the magnitude of the initial and 31, respectively.
curvature in the second model is approximately 100 times greater The variations in the time steps for both loading combinations
than in the rst model, as shown in Fig. 27. are shown in Fig. 32. The same parameter values applied in the
However, this fact occurs only in the initial time steps because previous example were used, changing only the values of a and
in the next instants, the curvatures in both models are approxi- Dtmax from the strategy of Lages et al., to 0.8 and 0.05 s, respec-
mately the same magnitude, as shown in the interval plotted in tively. The previous values of cp, 0.1 and 0.001 were used on the
Fig. 28, which could explain why the time step in the method of cases of F2 and F3, respectively.
Lages et al. starts with close to the same value in both models In the same way as in the last example, the time step sizes
and reaches much higher values at the end of the second model, calculated by Bergan and Mollestad remained close to 1/20 of the
which results from the reduced cp. However, in the rst model, fundamental period in both cases of loading. ANSYS yielded results
the time step size ends with values that are much greater than in with similar behaviour, but with values that were slightly greater,
the other strategies (Fig. 26(a)) because the initial curvature staying closer in the second case. From using the strategy of
already has a value that is much greater than in the subsequent Hulbert and Jang, the time step was rather oscillatory in the nal
instants, which leads to a small initial assumed value for cp in instants in both cases because the normalised error RL stays lower
the nal step calculations. than and greater than the tolerance limits, sequentially.
This phenomenon occurs because when the initial velocity vec- From using the algorithm of Lages et al., the time step was
tor (D_ 0 ) is equal to 0, the curvature at time t = 0 is given by the increased due to the continuous curvature reduction, staying in
norm of the acceleration vector (kD 0 k), which will have a greater the same range of values that was obtained with the other strate-
value when there are more degrees of freedom in the structure. gies. The results were almost the same in both loading cases, even
Even if this behaviour does not occur in the subsequent moments, with very different values of cp, because the rst case has quiescent
the curvature regularisation algorithm continues to use a high per- start conditions, which lead to zero initial curvature, and the sec-
centage of the values that were originally calculated. This phenom- ond case is similar to the previous example, with a very large initial
enon does not occur in quiescent start condition problems because curvature.
the initial acceleration is zero in this case.
To explore the results of the automatic time-stepping proce-
dures for multiple transient loadings, the same plane frame steel 5. Computational cost
structure was analysed with two acting loads, Fa(t) and Fb(t), which
are applied on the nodes shown in Fig. 29. In this case, only the Similarly to Table 2, the following Table 5 summarizes the
subdivided model was considered. results of analyzes of the previous multi-degree-of-freedom sys-
The proportional damping is introduced with a damping ratio of tems, in an attempt to compare the computational cost of the algo-
10% for the rst and third modes. Two groups of loading were rithms. The maximum and minimum time increments found with

First loading case Second loading case


0.0035 0.0030
Bergan & Mollestad Bergan & Mollestad
Hulbert & Jang Hulbert & Jang
0.0030 Lages et al. Lages et al.
0.0025
Time step size history - t (s)
Time step size history - t (s)

Ansys Ansys

0.0025
0.0020

0.0020
0.0015
0.0015

0.0010
0.0010

0.0005
0.0005

0.0000 0.0000
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Time (s) Time (s)
(a) (b)
Fig. 32. Time step size history results for the two cases of applied loads in the frame structure.

Table 5
Summary of results of multi-degree-of-freedom systems.

Example Bergan and Mollestad Hulbert and Jang Lages et al. Dt multip. factors
Dtmin Dt max NTOT Dtmin Dt max NTOT Dtmin Dtmax NTOT
Shear build. (1) 5.4994 18.3063 103 4.4521 18.4874 93 1.5787 18.2792 247 103 seg.
Shear build. (2) 1.0000 21.8786 131 0.3445 14.0845 258 1.0000 33.3247 306
Beam (1) 1.0000 48.5011 120 0.3546 44.8030 227 1.0000 98.0206 249
Beam (2) 1.0000 19.8014 801 0.3549 5.5313 1035 1.0000 35.5056 3234
Frame, load F1 1.9931 16.2555 376 0.0017 33.3489 284 0.9327 211.381 1093 104 seg.
Frame, load F2 2.8048 17.2158 403 1.8334 30.7454 948 2.9520 26.6504 812
Frame, load F3 0.3907 17.2371 451 0.0009 13.3440 1712 3.0195 19.2714 831
D.F. Rossi et al. / Engineering Structures 80 (2014) 118136 135

adaptive strategies are presented, as well as the total number of The global analysis that would be made to compare the compu-
points (NTOT) plotted with the different algorithms. The higher tational cost of the automatic time-stepping strategies, and to give
the number NTOT, the more expensive is the strategy and the more a nal conclusion about the less expensive technique, must con-
detailed is the dynamic response. sider these three criteria together: the total number of steps taken
For the simply supported beam systems, only results of the (NTOT, or the number of plotted points), the cost of individual time
analysis with the same data input for all strategies are shown, step size selection routine (which could be given not only by the
and for the three-bar plane frame systems, the results obtained asymptotic behaviour of efciency, but by the explicit complexity
only with the subdivided models are presented. The last column function f(n), to be more specic) and the frequency of chances
of Table 5 brings the multiplication factors of the values of Dt pre- in the value of Dt (which leads to changes in the matrix K). The
sented in the table. global comparison depends on these aspects, which are affected
Observing the summary results, it can be noted that the strat- by various parameters and become very complex.
egy of Bergan and Mollestad lead to the smallest values of NTOT The accuracy obtained with the numerical analyses by different
in almost all analyses (with higher values of Dt), except in two automatic time-stepping strategies should also be considered for
cases, where the strategy of Hulbert and Jang reached smaller val- completeness, what could be made only by extensive simulation
ues. The strategy of Lages et al., on the other hand, gave bigger val- of various systems and comparison with analytical or more accu-
ues of NTOT in many of the numerical examples, or a value near to rate responses, in the same way it was done in some numerical
the higher one, except for the last case, where the strategy of examples in the previous section. The strategy of Bergan and
Hulbert and Jang gave a value of NTOT greater than twice the value Mollestad is very similar to that used by the software ANSYS,
obtained with Lages et al. By only this comparison, one would con- which has been used and which reached accurate results in
clude that the strategy of Bergan and Mollestad leads to a less practical analyses. The strategy of Hulbert and Jang has itself an
expensive analysis, and that an analysis using the strategy of Lages error estimate that intends to control the accuracy results. The
et al. has a higher computational cost. However, the cost of opera- strategy of Lages et al. should be more applied in future simulations
tion in each time step size selection should be also considered to evaluate accuracy, because it was developed recently.
(which should be small, as it was noted in the statement 5 in the
beginning of Section 3), as well as the accuracy obtained with 6. Conclusions
the numerical analyses.
The computational cost of the strategies in the time step size Observing the formulation and numerical example results, one
selection can be compared by counting the number of numeric can draw some conclusions about the automatic adaptive time-
operations performed by each algorithm. A common technique stepping procedures.
for estimating the efciency of an algorithm is to analyse it accord- The strategy of Bergan and Mollestad [6], which is based on the
ing to the size of the problem, given by the number of processed current characteristic frequency, can easily be implemented in any
elements n. By calculating the number of operations performed numerical integration procedure, facilitating programming. This
on the n elements, a complexity function f(n) can be found for approach is the oldest method among the three studied approaches
the algorithm, and the asymptotic behaviour of the efciency can and has been successfully applied in commercial nite element
be estimate by taking big values of n. For a structural analysis prob- software. However, it was observed that the oscillations of the time
lem, n would be the number of degrees of freedom of the system. increments were too large in some examples when using this
To compare the three automatic time-stepping strategies with strategy (Fig. 22b) despite remaining close to a constant plateau
this approach of complexity function, it should be considered value. This pattern occurs because the frequency x  is close to zero
only the routine of time step size selection in each problem, when the response passes through maximum amplitudes, as the
because the numerical integration process is the same for all norm ||DD|| becomes relatively small in these situations (a fact
algorithms. that has been observed in other research). Even if verifying the
The most expensive operation executed in the routine of Bergan condition (15) for discarding some frequency x  i , the technique
and Mollestad is a matrixvector multiplication that appears on was not very effective, even when using values for e that were dif-
the terms of Eq. (12), which has complexity of n2. The second most ferent from 0.1. In some cases, this fact leads to a very frequent
expensive operation of this strategy is the norm-vector calculation oscillation in the time step sizes. This method also yields no addi-
of Eq. (15) (which has the cost of n) followed by minor xed simple tional gain in the analysis of single-degree-of-freedom systems,
arithmetic operations in each step. Thus, it is found that the strat- which have only one characteristic period. The suggested values
egy of Bergan and Mollestad has complexity of n2 , because its com- of the parameters k,  np ; 
nm , and e were used, but how to dene
plexity function has the form f(n) = An2 + Bn + C. Both strategies of these values remains unclear.
Hulbert and Jang and Lages et al., in the other hand, has complexity The strategy of Hulbert and Jang [12] is very intuitive when
of n, since their complexity function has the form f(n) = Bn + C, using integration errors, which is a concept that has been sug-
because the most expensive operations in the routines of these gested in various other studies (e.g., [7,10]) and appears to be a
algorithms are the norm-vector calculations of Eqs. (19) and (29), more suitable technique for future research, perhaps using more
followed by minor xed operations. recent error estimators [16]. The algorithm was very effective in
It is concluded that the strategy of Bergan and Mollestad is the the solution of single-degree-of-freedom systems and also led to
most expensive when only the additional step size selection compu- time step sizes that were more regular when solving multi-
tational cost is considered. However, it should be noted that the degree-of-freedom systems. However, the last example of a frame
expensive operation of factorising and solving the matrix Eq. (8) is structure in Section 4.4 displayed values of Dt that were very
always performed in the numerical analysis, because it is part of oscillatory and that were calculated by this strategy. Moreover,
the numerical integration process. This operation may have a com- the formulation of the local error presented by Hulbert and Jang
putational cost of complexity greater than n2 in some cases, which is attached to the integration algorithm of the generalised-a
leads the whole algorithm analysis to this order of asymptotic com- method and must be adapted to any other numerical procedure,
plexity anyway. Moreover, the gain in not make unnecessary which is not necessary in the other strategies. Even with the intu-
changes the value of Dt could be better than decrease the cost of itive use of Eq. (21) for tolerance intervals, the time step changing
the individual time step size selection, because it avoid a new factor- Eqs. (22) and (23) are not very intuitive and should be studied
ization of the effective stiffness matrix in Eq. (9) to solve the system. in more depth. Again, the suggested values of the parameters
136 D.F. Rossi et al. / Engineering Structures 80 (2014) 118136

(Dt/T)target, lb, pinc, and pdec were used, but how they are dened [3] Felippa CA, Park KC. Direct time integration methods in nonlinear structural
dynamics. Comput Methods Appl Mech Eng 1979;1718(2):277313.
remains rather unclear.
[4] Park KC, Underwood PG. A variable-step central difference method for
The strategy of Lages et al. [17] is the latest strategy that was structural dynamics analysis part 1. Theoretical aspects. Comput Methods
developed, and it has an interesting and intuitive derivation, which Appl Mech Eng 1980;22(2):24158.
uses the geometric indicator of the displacement history curvature. [5] Underwood PG, Park KC. A variable-step central difference method for
structural dynamics analysis part 2. Implementation and performance
The approach is not attached to the specic choice for the numer- evaluation. Comput Methods Appl Mech Eng 1980;23(3):25979.
ical integration procedure and can be applied to any of them. How- [6] Bergan PG, Mollestad E. An automatic time-stepping algorithm for dynamic
ever, the implementation had some difculties in the relation problems. Comput Methods Appl Mech Eng 1985;49(3):299318.
[7] Zienkiewicz OC et al. A unied set of single step algorithms. Part 1: general
between the curvature indicator and time step size. Despite the formulation and applications. Int J Numer Methods Eng 1984;20(8):152952.
regularisation procedure used in processing the curvature, the cur- [8] Zienkiewicz OC, Xie YM. A simple error estimator and adaptive time stepping
vature is not stabilised to any constant value, which causes a con- procedure for dynamic analysis. Earthq Eng Struct Dynam 1991;20(9):87187.
[9] Zeng LF et al. A posteriori local error estimation and adaptive time-stepping for
tinuous change in the time increment even when the system has a Newmark integration in dynamic analysis. Earthq Eng Struct Dynam
stationary response, which violates the second criterion shown in 1992;21(7):55571.
Section 3. Thus, the strategy needs a better development on an [10] Li XD, Zeng LF, Wiberg N-E. A simple local error estimator and an adaptive
time-stepping procedure for direct integration method in dynamic analysis.
application to structural vibration, particularly to solve the prob- Commun Numer Methods Eng 1993;9(4):27392.
lems observed when analysing the quiescent start condition sys- [11] Newmark NM. A method of computation for structural dynamics. ASCE J Eng
tems (which could have initial curvatures that are close to zero, Mech Div 1959;85(3):6794.
[12] Hulbert GM, Jang I. Automatic time step control algorithms for structural
which leads to large time step sizes) and systems with many
dynamics. Comput Methods Appl Mech Eng 1995;126(12):15578.
degrees of freedom (which could have greater initial curvatures, [13] Chung J, Hulbert GM. A time integration algorithm for structural dynamics
leading to unworkable time step sizes that are close to zero). with improved numerical dissipation: the generalized-a method. J Appl Mech
The computational cost analysis of the three strategies should 1993;60(2):3715.
[14] Hilber HM, Hughes TJR, Taylor RL. Improved numerical dissipation for time
consider various aspects when choosing the one with better perfor- integration algorithms in structural dynamics. Earthq Eng Struct Dynam
mance. For accuracy optimisation only, for example, it could be 1977;5(3):28392.
adopted the procedure of calculate the time step size with all strat- [15] Wood WL, Bossak M, Zienkiewicz OC. An alpha modication of Newmarks
method. Int J Numer Methods Eng 1980;15(10):15626.
egies and take the smaller value of Dt to be used in the step solution. [16] Chung J, Cho E-H, Choi KA. A priori error estimator of the generalized-a
method for structural dynamics. Int J Numer Methods Eng 2003;57(4):53754.
Acknowledgments [17] Lages EN et al. An adaptive time integration strategy based on displacement
history curvature. Int J Numer Methods Eng 2013;93(12):123554.
[18] Clough RW, Penzien J. Dynamics of structures. 3rd ed. Berkeley: Computers &
The authors would like to thank the institutions CNPq, CAPES, Structures Inc.; 1995.
and FAPES, which provided support for this work. [19] Chopra AK. Dynamics of structures: theory and applications to earthquake
engineering. Englewood Cliffs: Prentice Hall; 1995.
[20] Cook RD, Malkus DS, Plesha ME. Concepts and applications of nite element
References analysis. 3rd ed. Madison: John Wiley & Sons; 1989.
[21] Hughes TJR. A note on the stability of Newmarks algorithm in nonlinear
[1] Hibbitt HD, Karlsson BI. Analysis of Pipe Whip. Electric Power Research structural dynamics. Int J Numer Methods Eng 1977;11(2):3836.
Institute rept. no. EPRI NP-1208, Palo Alto; November 1979. [22] Hughes TJR. The nite element method: linear static and dynamic nite
[2] Oughourlian CV, Powell GH. ANSR-III: general purpose computer program for element analysis. Englewood Cliffs: Prentice Hall; 1987.
nonlinear structural analysis. Earthquake Engineering Research Center report [23] ANSYS, INC. Theory reference for ANSYS and ANSYS workbench; 2007.
no. UCB/EERC-82/21, Berkeley; November 1982.

View publication stats

S-ar putea să vă placă și