Sei sulla pagina 1di 9

ACCURATE INTEGRATION OF NONLINEAR SYSTEMS USING

NEWMARK EXPLICIT METHOD

S.-Y. Chang *
Department of Civil Engineering
National Taipei University of Technology
Taipei, Taiwan 10608, R.O.C.

ABSTRACT
In the step-by-step solution of a linear elastic system, an appropriate time step can be selected based on
analytical evaluation results. However, there is no way to select an appropriate time step for accurate
integration of a nonlinear system. In this study, numerical properties of the Newmark explicit method
are analytically evaluated after introducing the instantaneous degree of nonlinearity. It is found that the
upper stability limit is equal to 2 only for a linear elastic system. In general, it reduces for instantaneous
stiffness hardening and it is enlarged for instantaneous stiffness softening. Furthermore, the absolute
relative period error increases with the increase of instantaneous degree of nonlinearity for a given product
of the natural frequency and the time step. The rough guidelines for accurate integration of a nonlinear
system are also proposed in this paper based on the analytical evaluation results. Analytical evaluation
results and the feasibility of the rough guidelines proposed for accurate integration of a nonlinear system
are confirmed with numerical examples.

Keywords : Instantaneous degree of nonlinearity, Stability, Accuracy.

1. INTRODUCTION tem through the introduction of the step degree of


nonlinearity to monitor the stiffness change within a
Numerical properties of step-by-step integration time step [8,9].
methods [4-7,11-13,15,18-20] have been well explored Although the introduction of step degree of nonlin-
through many evaluation techniques [1-3,16,17,21]. earity enables the analytical evaluations of step-by-step
However, these works are based exclusively on linear integration methods in the solution of nonlinear systems,
elastic systems so that the numerical characteristics of it is still difficult to construct a guideline as that used
stability, accuracy, convergence, amplitude distortion, for a linear elastic system to obtain an accurate solution,
period distortion and overshooting can be analytically where the ratio of integration time step over period less
evaluated. It is inherently assumed that these numeri- than 1/10 or 1/20 is a good rule of thumb for obtaining
cal properties can be extrapolated to nonlinear systems, reliable results [1,14]. This is because the step degree
but this must be considered as conjecture. In fact, the of nonlinearity is determined by the stiffness change
characteristics of the step-by-step solution of a nonlin- within a time step. Hence, this step degree of nonlin-
ear system might not be reliably predicted by the ana- earity depends upon the starting stiffness of the time
lytical evaluation results for linear elastic systems. For step and step size. Therefore, it is very hard to esti-
example, an unconditionally stable integration method mate step the degree of nonlinearity before the step-by-
for linear elastic systems might lose its stability for step integration. This might result in the difficulty in
nonlinear systems [16]. The numerical characteristic selecting an appropriate time step to yield accurate so-
of the response depends upon how the step-by-step in- lutions since the relative period error is closely related
tegration method interacts with the overall solution to the step degree of nonlinearity for a nonlinear system.
procedure, whether the dynamic equilibrium of forces is In this study, a parameter named the instantaneous de-
met approximately or exactly at each time step, and gree of nonlinearity is introduced instead of the step
whether the solution is direct or iterative. Because of degree of nonlinearity in evaluating the numerical
the complexity of the step-by-step solution of nonlinear properties of a step-by-step integration method for
systems, numerical properties of step-by-step integra- nonlinear systems. Using these properties, rough
tion methods are often examined by numerical means. guidelines to select an appropriate time step for accurate
Recently, a technique was proposed to evaluate a pseu- integration of a linear or nonlinear system are proposed
dodynamic algorithm in the solution of a nonlinear sys- herein and confirmed with numerical examples. For
*
Professor

Journal of Mechanics, Vol. 25, No. 3, September 2009 289


illustration, the Newmark explicit method is chosen for 3. STEP-BY-STEP INTEGRATION
this investigation. PROCEDURE

In the analysis of a step-by-step integration method, it


2. NEWMARK EXPLICIT METHOD
is needed to rewrite this method in a recursive matrix
form [1,3,17,21]. Thus, the computing sequence of the
The general formulation of the Newmark explicit
(i + 1)-th time step for the Newmark explicit method is
method for a single degree of freedom system can be
described. Because the acceleration, velocity and dis-
expressed as
placement are available at the end of the (i)-th time step
mai +1 + cvi +1 + ki +1 di +1 = fi +1 the displacement di+1 can be directly computed from the
second line of Eq. (1). Hence, the restoring force can
1
di +1 = di + (Δt )vi + (Δt ) 2 ai (1) be determined from an assumed mathematical force-
2 displacement relationship. After substituting the ac-
1 celeration ai+1, which is a function of vi+1, into the first
vi +1 = vi + (Δt )(ai + ai +1 )
2 line of Eq. (1) the velocity can be calculated by
where m, c and k are the mass, viscous damping coeffi- −1
⎡ 1 ⎤ ⎡ 1 1 ⎤
cient and stiffness, respectively; di+1, vi+1 and ai+1 are vi +1 = ⎢ m + (Δt )c ⎥ ⎢ mvi + (Δt )mai + (Δt )( fi +1 − ri +1 ) ⎥
approximations to the displacement, velocity and accel- ⎣ 2 ⎦ ⎣ 2 2 ⎦
eration, respectively, at the (i + 1)-th time step. The (4)
symbol ki+1 is the secant stiffness at the end of (i + 1)-th Finally, the first line of Eq. (1) is used to compute the
time step. acceleration ai+1. This procedure can be repeated until
For a nonlinear system, the stiffness may change the desired response is achieved.
from the initial stiffness k0 to ki+1, which is the stiffness The application of the Newmark explicit method to
at the end of (i + 1)-th time step. To measure the compute the free vibration response for a damped,
stiffness change from the beginning of motion to the nonlinear single degree of freedom system can be
end of the (i + 1)-th time step, the parameter of δi+1, rewritten in a recursive matrix form and it is found to be
which is referred to as the instantaneous degree of
nonlinearity at the end of the (i + 1)-th time step, is de- Xi +1 = A i +1 Xi (5)
fined as
where Xi+1 = [di+1, (Δt) vi+1, (Δt)2 ai+1]T is defined, and
k Ai+1 represents the amplification matrix for the (i + 1)-th
δi +1 = i +1 (2)
k0 time step. The explicit expression of Ai+1 is found to
be
Noticed that δi+1 is only the ratio of the instantaneous
stiffness at the end of the (i + 1)-th time step over the ⎡ ⎤
initial stiffness and is nothing to do with the history of ⎢ 1 ⎥
1 1
stiffness. Apparently, δi+1 = 1 implies that there is no ⎢ 2 ⎥
stiffness change for the instantaneous stiffness at the ⎢ 1 1 1 1 ⎥
⎢ − δi +1 Ω 02 1− δi +1 Ω02 − δi +1 Ω0 2

end of the (i + 1)-th time step when compared to the A i +1 =⎢ 2 2 2 4 ⎥
initial stiffness. Whereas, the instantaneous stiffness is ⎢ 1 + ξΩ0 1 + ξΩ0 1 + ξΩ0 ⎥
hardening at the end of the (i + 1)-th time step for δi+1 > ⎢ ⎥
⎢ ⎛ 1 2 ⎞⎥
1 and it is softening for δi+1 < 1. In addition, for δi+1 > − ⎜ δi +1 ξΩi + δi +1 Ω0 ⎟
⎢ −δi +1 Ω 02 −(2ξΩ0 + δi +1 Ω02 ) ⎝ 2 ⎠⎥
1 the larger the δi+1, the higher the instantaneous stiff- ⎢ ⎥
ness hardening. Similarly, in the range of 0 < δi+1 < 1 ⎣ 1 + ξΩ0 1 + ξΩ0 1 + ξΩ0 ⎦
the smaller the δi+1, the higher the instantaneous stiff- (6)
ness softening.
In order to study the viscous damping effect in the where Ω0 = ω0(Δt) and ω0 = k0 / m is the natural
nonlinear time history analysis the viscous damping frequency of the system at the beginning of the motion.
coefficient, which is determined from the initial stiff-
ness [1,5] is adopted in the following analysis. In fact,
the viscous damping coefficient may be expressed as
4. ANALYSIS
ci +1 = 2ξ k0 m (3)
Numerical properties of the Newmark explicit
for the (i + 1)-th time step. In this expression, the method in the solution of a nonlinear system can be
symbol ξ is a viscous damping ratio, which is usually evaluated through the spectral analysis of Eq. (5). In
specified before the step-by-step integration. Thus, the fact, after the eigen-decomposition of the amplification
viscous damping coefficient is a constant during the matrix Ai+1 [1], the characteristic equation for the (i +
whole step-by-step integration procedure. 1)-th time step for this integration method is found to be

290 Journal of Mechanics, Vol. 25, No. 3, September 2009


⎡ 2 − δi +1 Ω02 1 − ξΩ0 ⎤ where j = −1 , ΩiD+1 = 1 − ξ2 Ωi +1 and ξi+1 is the
λ ⎢λ 2 − λ+ ⎥=0 (7) numerical damping ratio at the (i + 1)-th time step. To
⎣ 1 + ξΩ0 1 + ξΩ 0 ⎦
estimate the period distortion in a step-by-step integra-
This equation reveals that there are three eigenvalues tion procedure, it is necessary to obtain the computed
and one of them is zero. This equation can be used to phase angle Ωi+D
1. In fact, it can be computed from Eq.
determine the numerical properties in the solution of (9) and is
nonlinear systems.
Stability ⎛ 1 − 1 δ Ω2 ⎞
ΩiD+1 = cos −1 ⎜⎜ 2 i +1 0 ⎟⎟ (10)
In Eq. (7), if there are two complex conjugate eigen- ⎝ 1 + ξΩ0 ⎠
values and |λ1,2| ≤ 1, a bounded oscillatory response for
the (i + 1)-th time step can be achieved. In fact, based After finding the computed phase angle Ωi+ D
1 , the rela-
on these stability conditions the stability limits for the
tive period error, which is usually considered as a
Newmark explicit method are found to be
measure of period distortion in the step-by-step integra-
⎛ 2 ⎞ tion, can be computed by
0 < Ω0 ≤ Ωu0 , Ωu0 = ⎜ ⎟ δi +1 − ξ
2
(8)
⎝ δi +1 ⎠ Ti +1 − Ti +1
Pi +1 = (11)
Ti +1
The upper stability limit exists only if δi + 1 ≥ ξ . Thus, 2

the stability limits are closely related to the instantane-


where Ti+1 = 2π/ωi+1 is the true period of the system at
ous degree of nonlinearity and viscous damping ratio.
the end of the (i + 1)-th time step and Ti +1 = 2π / ωi +1 is
Variations of the upper stability limit Ωu0 versus the
the computed period corresponding to the computed
instantaneous degree of nonlinearity δi+1 for different
natural frequency ωi +1 . In this definition, a positive
values of ξ are plotted in Fig. 1. The upper stability
relative period error implies that the period is elongated
limit of Ωui decreases with the increasing of δi+1 for
while it is shrunk for a negative value.
the curve of ξ = 0. For example, the upper stability
Variations of the relative period error (Ti +1 −
limit Ωi extends to 200 as δi+1 = 0.0001 while it reduces
Ti +1 ) / Ti +1 versus Ω0 for δi+1 = 0.5, 1 and 2 and ξ = 0,
to 0.2 as δi+1 = 100. A similar phenomenon is also
0.1 and 0.2 are shown in Fig. 2. In this figure, the ab-
found for the other curves except each curve starts from
solute relative period error increases with the increasing
δi + 1 ≥ ξ2 for ξ ≠ 0 and it increases gradually at first and
value of Ω0 in each curve. This property is the same as
then decreases with the increasing of δi+1. This reveals
that for a linear elastic system. For an instantaneous
that the upper stability limit will be shrunk for instanta-
stiffness softening system, such as δi+1 = 0.2, an increase
neous stiffness hardening while it is enlarged for in-
of viscous damping will lead to the increase of period
stantaneous stiffness softening. It is also found that
elongation as Ω0 is small while for linear elastic and
the increase of the viscous damping ratio will reduce the
instantaneous stiffness hardening systems, such as δi+1 =
upper stability limit.
1 and 2, an increase of viscous damping results in the
Accuracy increase of period shrinkage as Ω0 is small. It is also
If there are two complex conjugate eigenvalues in Eq.
found that numerical damping accompanies with period
(7), they can be expressed as
distortion although the value of Ω0 tends to zero.
λ1,2 = e − ξ i+1 Ω i+1 (cos ΩiD+1 ± j sin ΩiD+1 ) (9)

Fig. 1 Variation of upper stability limit with δi+1 Fig. 2 Variation of relative period error with Ω0

Journal of Mechanics, Vol. 25, No. 3, September 2009 291


Viscous Damping
Since the amplitude decay is closely related to the
numerical damping ratio, it is interesting to compute the
numerical damping ratio ξi+1 for the (i + 1)-th time
step for a nonlinear system. For this purpose, it can be
computed from Eq. (9) and is

−1 ⎛ 1 − ξΩ0 ⎞
ξi +1 = ln ⎜ ⎟⎟ (12)
Ωi +1 ⎜⎝ 1 + ξΩ0 ⎠
Figure 3 shows the variation of the numerical damp-
ing ratio with Ω0 for δi+1 = 0.5, 1 and 2. It seems that
the numerical damping ratio becomes constant within Fig. 3 Variation of numerical damping ratio with Ω0
the upper stability limit for any curve. It is worth not-
ing that the curves for ξ = 0 as δi+1 = 0.5, 1 and 2 are
overlapped together. Thus, the Newmark explicit
method possesses no numerical dissipation for zero
viscous damping. Numerical damping is generally
generated for nonzero viscous damping. In addition, it
increases with the increasing viscous damping ratio for
a given δi+1. Meanwhile, for a given nonzero viscous
damping ratio, it decreases with the increasing value of
δi+1. It is also interesting to find that the numerical
damping ratio ξ is almost equal to the given viscous
damping ratio ξ as δi+1 = 1.
Spectral Radius Fig. 4 Variation of spectral radius with Ω0
The spectral radius is defined as max| λj| for j = 1, 2,
and 3, where λj is an eigenvalue of the characteristic an appropriate time step for accurate integration of a
equation. The variation of the spectral radius with Ω0 linear or nonlinear system either based on period or ac-
is shown in Fig. 4 for δi+1 = 0.5, 1 and 2. In this figure, curacy consideration. However, a conservative way to
the point with the abrupt change of slope is the bifurca- choose an appropriate time step is proposed next. This
tion point where complex conjugate eigenvalues bifur- relies upon the estimation of the maximum instantane-
cate into real and distinct eigenvalues. The spectral ous degree of nonlinearity from the actual instantaneous
radius is equal to 1 for zero viscous damping for any degree of nonlinearity. For this purpose, the following
δi+1. For each curve, the spectral radius is decreased rough guidelines are proposed to estimate the maximum
with the increase of Ω0, and then turns upward steeply instantaneous degree of nonlinearity
after the bifurcation point. It is also found that the
δmax = 1 for instantaneous stiffness softening or
spectral radius decreases with the increase of the speci-
linear elastic systems
fied viscous damping ratio as δi+1 is fixed.
δmax = 2 for instantaneous stiffness hardening or
unknown systems
5. ROUGH GUIDELINES FOR ACCURATE Apparently, δmax = 1 is taken for a linear elastic sys-
INTEGRATION OF NONLINEAR SYSTEMS tem. It is conservative to take δmax = 1 for any instan-
taneous stiffness softening system since its instantane-
Since the numerical properties of Newmark explicit ous stiffness is always smaller than the initial stiffness.
method can be expressed as a function of Ω0 and the Meanwhile, it is very rare for a system its instantaneous
instantaneous degree of nonlinearity, it may become stiffness will become twice of the initial stiffness.
feasible to choose an appropriate time step for accurate Hence, it seems appropriate to take δmax = 2 for an in-
integration of a nonlinear system. This is because that stantaneous stiffness hardening system. It is natural to
the value of Ω0 can be obtained after determining the assume the most stringent constraint of δmax = 2 for an
initial natural frequency and the actual instantaneous unknown structural system.
degree of nonlinearity might be replaced by an esti- After finding the initial natural frequency and maxi-
mated maximum instantaneous degree of nonlinearity. mum instantaneous degree of nonlinearity, an appropri-
The proposed details are described next. ate time step may be selected for accurate integration of
It is clear that the actual time history of instantaneous a nonlinear system. A stable computation can be
degree of nonlinearity cannot be known before the step- achieved if Eq. (8) is met. After substituting δmax = 1
by-step integration although it is inevitable for selecting for linear elastic and instantaneous stiffness softening

292 Journal of Mechanics, Vol. 25, No. 3, September 2009


systems as well as δmax = 2 for instantaneous stiffness the instantaneous degree of nonlinearity at the (i + 1)-th
hardening or unknown systems into Eq. (8), the maxi- time step for this mode is found to be
mum time step to satisfy the upper stability limit is
2
found to be ⎡ ω( j ) ⎤
δi(+j1) = ⎢ i(+j1) ⎥ (14)
⎣ ω0 ⎦
2 1− ξ 2
(Δt ) max ≤ for linear elastic or instantaneous
ω0 where ω(0 j ) and ωi(+j1) are the natural frequencies of
stiffness softening systems the j-th mode based on the initial stiffness and the stiff-
ness at the end of the (i + 1)-th time step. Conse-
2−ξ 2
(Δt ) max ≤ for instantaneous stiffness hardening quently, the relative period error for this mode at the (i +
ω0
1)-th time step can be obtained from Fig. 2 based on
or unknown systems δi+ 1 , ωi +1 and the size of integration time step.
( j) ( j)
The
(13) assumptions made to estimate the maximum instanta-
In addition to stability consideration, an accuracy con- neous degree of nonlinearity for a single degree of
sideration must be also satisfied for accurate integration freedom system are also taken for each mode of a mul-
of a linear or nonlinear system. The ratio of time step tiple degree of freedom system. Consequently, reliable
over period less than 1/10 or 1/20 has been considered solutions can be achieved if the dominant modes are
as a guideline for accurate integration of a linear elastic accurately integrated.
system [1,14]. It is manifested from Fig. 2 for ξ = 0
and δi+1 = 1 that the relative period error is about 0.41%
as Δt/Ti = 1/20 and 1.69% as Δt/Ti = 1/10. Since the 7. NUMERICAL EXAMPLES
relative period error for a nonlinear system depends not
only upon Δt/Ti but also δi+1 for each time step, it seems
inappropriate to propose a rough guideline by con- Numerical properties of the Newmark explicit
straining Δt/Ti within a certain limit, such as 1/10 or method for a nonlinear system can be evaluated through
1/20 for a linear elastic system, without considering the the introduction of instantaneous degree of nonlinearity.
value of δi+1. Alternatively, it is appropriate to con- To confirm these analytical results and the rough guide-
sider a relative period error less than 1% for each time lines proposed for accurate integration of linear and
step as a rough guideline for accurate integration of a nonlinear systems, some numerical examples are exam-
linear or nonlinear system. Hence, after finding the ined.
initial natural frequency and maximum instantaneous In this study, a two-story shear-beam type structure is
degree of nonlinearity, an appropriate time step can be considered. The stiffness of each story is assumed to
chosen by using Fig. 2. It should be mentioned that be in the form of
the selection of a time step is also affected by other
factors, such as the rapid changes of dynamic loading k = k0 ⎡⎣1 + α(Δu ) 2 ⎤⎦ (15)
[10]. Only the structural nonlinearity is emphasized in
this study and thus other factors are neglected. where k0 is the initial stiffness and Δu is a story drift.
The stiffness is made up of two parts. One is the linear
part of k0 and the other is the nonlinear part of αk0(Δu)2.
6. INSTANTANEOUS DEGREE OF It is clear that α = 0 mimics a linear elastic system while
NONLINEARITY FOR MULTIPLE α ≠ 0 it simulates a nonlinear system. Two nonlinear
DEGREE OF SYSTEM systems are examined and their structural properties are
defined as follows

It is recognized that a nonlinear multiple degree of System (1)


freedom system cannot reduce to uncoupled single de- m1 = m2 = 103 kg k1 = 105 ⎡⎣1 + 2(Δu )2 ⎤⎦
gree of freedom systems for a whole step-by-step inte-
k2 = 105 ⎡⎣1 + 2(Δu )2 ⎤⎦
gration procedure. However, it seems this can be done
for each time step, and hence the numerical properties
for the step-by-step solution of a nonlinear single degree System (2)
of freedom system can be applied to these uncoupled m1 = m2 = 103 kg k1 = 107 ⎡⎣1 − 102 (Δu )2 ⎤⎦
single degree of freedom systems. This highly relies k2 = 104 ⎡⎣1 − 10−3 (Δu )2 ⎤⎦
upon the determination of the natural frequencies of the
modes of interest in each time step since the instanta- It is clear that System (1) is an instantaneous stiffness
neous degrees of nonlinearity for these modes can be hardening system due to a positive nonlinear part in
further computed based on these frequencies. This each story while System (2) is an instantaneous stiffness
procedure is described next. After finding the natural softening system due to a negative nonlinear part in
frequency of the j-th mode for the (i + 1)-th time step, both stories. The natural frequencies of System (1) are

Journal of Mechanics, Vol. 25, No. 3, September 2009 293


found to be 6.18 and 16.18 rad/sec based on the initial
stiffness matrix while for System (2) they are found to
be 3.16 and 100.05 rad/sec. Both systems are excited
by a ground acceleration of 50sin(t) at the base. The
small applied frequency of 1 rad/sec is intentionally
chosen so that the dynamic loading can be easily cap-
tured [8]. The displacement response obtained from
the Newmark explicit method with Δt = 0.001 sec is
considered as an “exact” solution for comparison for
both systems. The rough guidelines proposed for se-
lecting an appropriate time step for accurate integration
of a nonlinear system are employed in this investigation.
In addition, numerical properties of relative period error,
instantaneous degree of nonlinearity and upper stability
limit are also discussed in these numerical examples. Fig. 5 Displacement responses for System (1)

System (1)
An appropriate time step must be chosen based on at
least stability and accuracy considerations. Since Sys-
tem (1) is an instantaneous stiffness hardening system,
the maximum instantaneous degree of nonlinearity of
δmax = 2 must be taken according to the proposed rough
guidelines. Thus, the maximum time step to have a
stable computation requires that (Δt ) max ≤ 2 − ξ2 / ω(2)
0
= 0.087 sec , where ω(2)0 = 16.18 rad/sec and ξ = 0.
Since the natural frequency of the second mode is much
larger than that of the first mode for System (1), it is
anticipated that the response contribution from the sec-
ond mode to the total response is insignificant. Hence, Fig. 6 Time history of relative period error for
a reliable solution can be obtained if the first mode is System (1)
accurately integrated. As a result, the maximum time
step of (Δt)max = 0.055 sec is found from Fig. 2 for ξ = 0
and δi+1 = 2 as indicated by a solid circle. This time
step will lead to a relative period error less than about
1% for the first mode.
Numerical results are plotted in Fig. 5 for System (1).
In order to discuss these results based on analytical
evaluation results, response time histories of relative
period error, instantaneous degree of nonlinearity and
upper stability limit are also plotted in Figs. 6 to 8, re-
spectively. In Fig. 6, the relative period errors for the
first and second modes are only about 2.13 × 10−6 and
1.35 × 10−5, respectively, for Δt = 0.001 sec. Hence,
both modes are very accurately integrated and thus the
numerical solution can be considered as an “exact” so- Fig. 7 Time history of instantaneous degree of
lution for comparison. On the other hand, they be- nonlinearity for System (1)
come about 0.0065 and 0.045 for the first and second
mode, respectively, for Δt = 0.055 sec. Apparently, the
first mode is accurately integrated while considerable to satisfy the upper stability limit. This implies that
period distortion is found in the second mode. Since the analytical evaluation results obtained from the in-
the second mode response is negligible a reliable solu- stantaneous degree of nonlinearity can reliably predict
tion can still be achieved as shown in Fig. 5. The nu- the behaviors of the numerical solutions.
merical results obtained from Δt = 0.088 sec blow up In Fig. 7, the instantaneous degree of nonlinearity of
early due to the violation of the upper stability limit in δi(1) varies between 1 and 1.8 for the first mode while
the second mode. In fact, this time step is slightly lar- for the second mode δi(2) varies between 1 and 1.6.
ger than the maximum time step of (Δt)max = 0.087 sec Thus the instantaneous degree of nonlinearity is always

294 Journal of Mechanics, Vol. 25, No. 3, September 2009


able solution is dominated by stability consideration and
thus the time step of Δt = 0.019 sec is used. On the
other hand, the time step of Δt = 0.021 sec is also ap-
plied to perform the step-by-step integration and is in-
tended to confirm numerical instability.
Figure 9 displays the displacement responses of Sys-
tem (2). In addition, the time histories of relative pe-
riod error, instantaneous degree of nonlinearity and up-
per stability limit are also plotted in Figs. 10 to 12, re-
spectively. In Fig. 9, the numerical solution obtained
from Δt = 0.019 sec is almost coincided with that ob-
tained from Δt = 0.001 sec while that obtained from Δt =
0.021 sec results in numerical instability very early.
Fig. 8 Time history of upper stability limit for System This can be thoroughly explained by Figs. 10 to 12.
(1) Figure 10 reveals that the relative period errors are
about 1.48 × 10−4 and 0.24 for the first and second
modes, respectively. Since the first mode is very ac-
greater than 1 for both modes for System (1). This is curately integrated and the second mode is drastically
consistent with the assumption of 0 < δi ≤ δmax = 2 for larger than the first mode, the numerical solution is very
an instantaneous stiffness hardening system. Appar- accurately achieved although the second mode is sig-
ently, the numerical stability problem is dominated by nificantly distorted. In Fig. 11, the instantaneous de-
the second mode. The values of Ω(2) 0 are found to be gree of nonlinearity for the first mode of δi(1) for Sys-
0.88 and 1.42 in correspondence to Δt = 0.055 and 0.088 tem (2) varies between 0.94 and 1 while for the second
sec, respectively. In Fig. 8(b), it is found that mode δi(2) varies between 0.98 and 1. Thus, the in-
0 = 1.42 for Δt = 0.088 sec, which is shown by a
Ω(2) stantaneous degree of nonlinearity is always less than 1
solid thick line, is greater than the upper stability limit for both modes. The values of Ω(2) 0 are found to be
of Ω(2~
0
u)
at the early time of t = 1.67 sec and thus in- 1.90 and 2.10 corresponding to Δt = 0.019 and 0.021 sec,
stability occurs. This time is in a good agreement with respectively. In Fig. 12(b), it is found that
the time of the numerical explosion occurred in the dis- Ω 0 = 2.10 for Δt = 0.021 sec, which is shown by a
(2)

placement response time history as shown in Fig. 5. solid thick line, is greater than the upper stability limit
Thus, the reliability of the stability analysis is evident. of Ω(2~0
u)
at the very early time of t = 0.21 sec and thus
instability occurs after this time instant. It seems that
System (2)
the rough guidelines proposed are feasible at least for
In System (1), the maximum time step to have an
the numerical examples studied herein. However, a
accurate solution is dominated by accuracy considera-
further study is still needed.
tion while for System (2) it is proposed to be dominated
by stability consideration. For this purpose, System (2)
is designed to have a relatively very high second mode. 8. CONCLUSIONS
Since the story stiffness reduces after the structure de-
forms System (2) is an instantaneous stiffness softening
system. Thus, δmax = 1 is taken according to the pro- Numerical characteristics of the Newmark explicit
posed rough guidelines. As a result, the maximum method in solving a nonlinear system are evaluated
through the introduction of the instantaneous degree of
time step to have a stable computation,
nonlinearity for each time step. It is found that the
(Δt ) max ≤ 2 1 − ξ2 / ω(2)
0 = 0.020 must be satisfied, upper stability limit generally varies with the instanta-
where ω(2) 0 = 100.05 rad/sec and ξ = 0. Again, the neous degree of nonlinearity and a viscous damping
natural frequency of the second mode is much larger ratio. In fact, this limit will be enlarged for instanta-
than that of the first mode. Therefore, the response neous stiffness softening and it reduces for instantane-
contribution from the second mode to the total response ous stiffness hardening when compared to the case of
is negligible and an accurate integration of the first stiffness invariant, whose upper stability limit is 2. In
addition to that the period distortion increases with the
mode will provide a reliable solution. Hence, to meet
product of natural frequency and step size, it is also
accuracy consideration, the maximum time step of found that its relative period error also increases with
(Δt)max = 0.15 sec is found from Fig. 2 for ξ = 0 and δi+1 the increasing step degree of nonlinearity for a given
= 1 as indicated by a hollow circle. This time step will product of natural frequency and time step in the solu-
lead to a relative period error less than about 1% for the tion of a nonlinear system. Based on these numerical
first mode. It is apparent that the time step to a reli- properties, rough guidelines for accurate integration of

Journal of Mechanics, Vol. 25, No. 3, September 2009 295


Fig. 9 Displacement responses for System (2) Fig. 12 Time history of upper stability limit for System
(2)

ACKNOWLEDGEMENTS

The author is grateful to acknowledge that this study


is financially supported by the National Science Council,
Taiwan, R.O.C., under Grant No. NSC-95-2221-E-027-
099.

REFERENCES

1. Bathe, K. J., Finite Element Procedures in Engineering


Fig. 10 Time history of relative period error for Analysis, Prentice-Hall, Inc. Englewood Cliffs, New
System (2) Jersey (1996).
2. Belytschko, T. and Schoeberle, D. F., “On the Uncondi-
tional Stability of an Implicit Algorithm for Nonlinear
Structural Dynamics,” Journal of Applied Mechanics,
17, pp. 865−869 (1975).
3. Belytschko, T. and Hughes, T. J. R., Computational
Methods for Transient Analysis, Elsevier Science Pub-
lishers B.V., North-Holland (1983).
4. Chang, S. Y., “A Series of Energy Conserving Algo-
rithms for Structural Dynamics,” Journal of the Chinese
Institute of Engineers, 19, pp. 219−230 (1996).
5. Chang, S. Y., “Improved Numerical Dissipation for
Explicit Methods in Pseudodynamic Tests,” Earthquake
Engineering and Structural Dynamics, 26, pp. 917−929
(1997).
6. Chang, S. Y., “The γ-Function Pseudodynamic Algo-
Fig. 11 Time history of instantaneous degree of rithm,” Journal of Earthquake Engineering, 4, pp.
nonlinearity for System (2) 303−320 (2000).
7. Chang, S. Y., “Explicit Pseudodynamic Algorithm with
Unconditional Stability,” Journal of Engineering Me-
linear elastic and nonlinear systems are proposed and chanics, ASCE, 128, pp. 935−947 (2002).
are confirmed with numerical examples. It seems that 8. Chang, S. Y., “Nonlinear Error Propagation Analysis
the rough guidelines proposed herein may also apply to for Explicit Pseudodynamic Algorithm,” Journal of En-
other step-by-step integration methods. However, it is gineering Mechanics, ASCE, 129, pp. 841−850 (2003).
of need to further study the appropriateness of the as- 9. Chang, S. Y., “Error Propagation in Implicit Pseudody-
sumptions made for estimating the maximum instanta- namic Testing of Nonlinear Systems,” Journal of Engi-
neous degree of nonlinearity. neering Mechanics, ASCE, 131, pp. 1257−1269 (2005).

296 Journal of Mechanics, Vol. 25, No. 3, September 2009


10. Chang, S. Y., “Accurate Representation of External 17. Hughes, T. J. R., The Finite Element Method, Prentice-
Force in Time History Analysis,” Journal of Hall, Inc., Englewood Cliffs, N.J. (1987).
Engineering Mechanics, ASCE, 132, pp. 34−45 (2006). 18. Newmark, N. M., “A Method of Computation for
11. Chang, S. Y., “Enhanced, Unconditionally Stable Ex- Structural Dynamics,” Journal of Engineering
plicit Pseudodynamic Algorithm,” Journal of Engi- Mechanics Division, ASCE, 85, pp. 67−94 (1959).
neering Mechanics, ASCE, 133, pp. 541−554 (2007). 19. Wilson, E. L., Farhoomand, I. and Bathe, K. J.,
12. Chang, S. Y., “Improved Explicit Method for Structural “Nonlinear Dynamic Analysis of Complex Structures,”
Dynamics,” Journal of Engineering Mechanics, ASCE, Earthquake Engineering and Structural Dynamics, 1,
133, pp. 748−760 (2007). pp. 241−252 (1973).
13. Chang, S. Y., Huang, Y. C. and Wang, C. H., “Analysis 20. Wood, W. L., Bossak, M. and Zienkiewicz, O. C., “An
of Newmark Explicit Integration Method for Nonlinear Alpha Modification of Newmark’s Method,” Interna-
Systems,” Journal of Mechanics, 22, pp. 321−329 tional Journal for Numerical Methods in Engineering,
(2006). 15, pp. 1562−1566 (1981).
14. Clough, R. W. and Penzien, J., Dynamics of Structures, 21. Zienkiewicz, O. C., The Finite Element Method,
McGraw-Hill, Inc, International Editions (1993). McGraw-Hill Book Co (UK) Ltd. Third Edition (1977).
15. Hilber, H. M., Hughes, T. J. R. and Taylor, R. L., “Im-
proved Numerical Dissipation for Time Integration Al-
gorithms in Structural Dynamics,” Earthquake Engi-
neering and Structural Dynamics, 5, pp. 283−292
(1977).
16. Hughes, T. J. R., “Stability, Convergence and Growth
and Decay of Energy of the Average Acceleration
Method in Nonlinear Structural Dynamics,” Computer (Manuscript received July 17, 2007,
and Structures, 6, pp. 313−324 (1976). accepted for publication September 9, 2008.)

Journal of Mechanics, Vol. 25, No. 3, September 2009 297

Potrebbero piacerti anche