Computers and Structures 206 (2018) 42–53

Contents lists available at ScienceDirect

Computers and Structures

journal homepage: www.elsevier.com/locate/compstruc

An improved explicit time integration method for linear and nonlinear

structural dynamics
Wooram Kim ⇑, Jin Ho Lee
Department of Mechanical Engineering, Korea Army Academy at Yeongcheon, Yeongcheon-si, Gyeongsangbuk-do 38900, Republic of Korea

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

Article history: In this article, a new explicit time integration method is developed to analyze linear and nonlinear prob-
Received 12 February 2018 lems of structural dynamics. Like recently developed explicit time integration methods, the new explicit
Accepted 6 June 2018 method can also control the amount of numerical dissipation in the high frequency range. The method is
Available online 23 June 2018
explicit in the presence of the damping matrix, if the mass matrix is diagonal. Due to the unconventional
approximations of the displacement vector, the new method does not require evaluation of the initial
Keywords: acceleration vector and other acceleration vectors. Linear and nonlinear problems of structural dynamics
Linear and nonlinear structural dynamics
can be tackled in a consistent manner, and iterative solution finding procedures are not required. Various
Explicit time integration method
Hermite approximations in time
illustrative problems are used to investigate improved performance of the new explicit method.
Wave propagation Ó 2018 Elsevier Ltd. All rights reserved.
Time finite element method
Collocation method

1. Introduction
Recently, numerous implicit and explicit time integration meth-
ods were proposed for effective analyses of structural dynamics.
Many of the recently developed time integration methods possess
controllable numerical dissipations which are useful for eliminat-
ing the spurious high frequency mode in numerical solutions
[1,2]. In general, implicit methods are unconditionally stable when
they are applied to linear problems, while explicit methods are
only conditionally stable. Due to this fact, dissipative implicit
methods can be used for the high frequency filtering by adopting
considerably large time steps, and numerical dissipations in expli-
cit methods are usually used to improve quality of numerical solu-
tions in wave propagation and impact problems where small time
steps are required.
Other than stability conditions, the biggest difference between
implicit and explicit methods can be found in equation solving pro-
cedures. In implicit methods, the displacement and velocity vec-
tors of current time step are expressed in terms of both
unknown properties of current time step and known properties
of previous time steps. Naturally, implicit methods require factor-
izations of the effective stiffness matrices which are not diagonal to
solve the fully discrete equations.
In linear analyses, factorization of the effective stiffness matrix
is required only once, if the factorized effective stiffness matrix is
⇑ Corresponding author.
E-mail address: kim.wooram@yahoo.com (W. Kim).
stored in additionally allocated memories and reused for next time
steps. In nonlinear analyses, however, the internal force vectors
and the stiffness matrices are often functions of unknown displace-
ment and velocity vectors of current time. Due to this reason, con-
struction and factorization of the effective stiffness matrix are
inevitable in each time step for implicit methods, and each time
step accompanies several times of iterations to obtain converged
nonlinear solutions. In large and complex nonlinear systems, this
may seriously limit solution refinements which can be done by
decreasing sizes of time steps, because factorization of a big effec-
tive stiffness matrix requires huge computational resources.
Details regarding recent development of implicit methods and
their computational aspects can be found in Refs. [2–5].
On the other hand, factorization of any matrices is not required
in explicit methods if the mass matrix is diagonal. Due to this fact,
explicit methods require much less computational effort to
advance a time step compared with implicit methods. In nonlinear
problems, the entries of the mass matrix are usually given as con-
stants, while the damping matrix and the internal force vector are
functions of the current displacement and velocity vectors. Even in
these situations, well designed explicit methods does not require
any factorizations of matrices, if the mass matrix is diagonal. Even
when the mass matrix is not diagonal, factorization of the mass
matrix is required only once, if the factorized mass matrix is stored
in additionally allocated memories and reused for next time steps.
Thus, explicit methods may be more efficient for analyses of large
nonlinear systems that require very small sizes of time steps for a
long duration of time.

https://doi.org/10.1016/j.compstruc.2018.06.005
0045-7949/Ó 2018 Elsevier Ltd. All rights reserved.
 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53
The Newmark method is one of the most broadly used non-
dissipative second-order accurate implicit method [6] for struc-
tural dynamic problems. After the introduction of the Newmark
method, numerous improved implicit methods were developed
based on it. One of the most famous methods developed based
on the Newmark method is the generalized-a method [7]. The
generalized-a method of Chung and Hulbert can control numerical
dissipations of the high frequency limit in a simple and practical
manner. Recently, some implicit methods were developed based
on the time finite element approach. The collocation composite
time integration method of Kim and Reddy [5] and the generalized
composite method of Kim and Choi [4] are recently proposed
second-order accurate implicit methods based on the time finite
element method.
Good explicit methods can also provide very accurate numerical
solutions for very complicated nonlinear problems with much less
computational effort. The 4th-order Runge-Kutta method (the RK4
method) and the central difference method (the CD method) are
standard explicit methods which can be used for structural dynam-
ics. The CL method of Chung and Lee provided acceptably accurate
solutions for the elastic spring-mass nonlinear pendulum problem
[8]. The HC method of Hulbert and Chung [9], the TW method of
Tchamwa and Wielgosz [10,11], and the NB method of Noh and
Bathe [12] was developed for the analyses of wave propagations
and impact problems. Recently, the Soares method [13] was devel-
oped based on the weighted residual approach.
The explicit methods mentioned above have their own advan-
tages and disadvantages. The RK4 method was originally consid-
ered for general first-order ordinary differential equations. By
rearranging equations of structural dynamics as proper first-
order forms, structural dynamics can also be analyzed with the
RK4 method, but this method requires more than four times of
computational cost compared with the CD method. The RK4
method provides a considerably large amount of numerical
dissipation (the minimum spectral radius is about 0.5) in the high
frequency range when applied to the second-order linear single-
degree-of-freedom problem. The CD method is the simplest
non-dissipative explicit method. The CD method is provided as a
standard time integration method in many software packages,
but the CD method becomes implicit in the presence of the viscous
damping terms.
The CL method can maintain explicitness in the presence of
viscous damping terms. However, the amplification matrix of the
CL method has the spurious eigenvalue that may seriously
influence quality of solutions when large time steps are used.
The minimum spectral radius of the most dissipative case of the
CL method is about 0.52 when b ¼ 28=27 is used, and only less
dissipative cases are included in the CL method. Unlike the CL
method, the HC method can include a full range of dissipative
cases, however, the non-dissipative case of the HC method
becomes unconditionally unstable for any choices of time steps in
the presence of viscous damping terms. The amplification matrix
of the HC method also has the spurious root.
The TW method does not require computation of the initial
acceleration vector, and the method is very effective for wave
propagation problems. On the other hand, all dissipative cases of
the TW method are only first-order accurate, thus this method is
not suitable for long term analyses. The Soares method is the only
self starting method among the explicit methods mentioned above.
It has dissipation control capability and provides improved accu-
racy and extended stability limit for linear problems. However
the Soares method cannot be used for nonlinear analyses because
it directly manipulated the linear structural dynamics equations
in a weighted residual sense. The Soares method becomes only first
order accurate in the presence of physical damping terms, and
requires integral evaluation of the external force vector.
43
The NB method is probably the best performing second-order
accurate explicit method among the explicit methods mentioned
above. The NB method is developed based on the strategy similar
to the strategy used in the implicit Bathe method where two
sub-steps were combined to form one complete method. The NB
and Soares method share almost identical spectral characteristics
for the linear undamped single-degree-of-freedom problem. How-
ever, the NB method can be applied to nonlinear analyses in a con-
sistent manner, while the Soares method cannot. Unlike the Soares
method, the NB method requires computation of the initial accel-
eration vector, and the amplification matrix of the NB method
always has spurious root when viscous damping terms are
included.
In this work, we propose a new second-order accurate explicit
time integration method to tackle variety of linear and nonlinear
problems of structural dynamics in a consistent way. In designing
the new explicit method, we wish to accommodate the preferable
attributes of the existing explicit methods and exclude undesirable
attributes by manipulating proper numerical techniques and
methods. To this end, we consider unconventional interpolating
techniques, effective residual minimizing procedures, and effective
computational structures of recently developed explicit methods.
Discussions of this paper will mainly focus on the development
and analysis of the proposed algorithms. Through the interpolating
techniques used herein, we wish to exclude the spurious root of
the amplification matrix of the proposed explicit method. By
adopting unique computational structures of the Noh and Bathe
method, we wish to achieve extended stability limit and improved
spectral characteristics in the proposed method. By using the collo-
cation approach for the time discretization, we also expect that the
proposed method will be applicable to both linear and nonlinear
problems in a consistent manner. In addition to these improve-
ments, we also wish to eliminate computation of the acceleration
vectors (including the initial acceleration vector) in our explicit
method, which is already realized in the Soares method. Simple
and illustrative linear and nonlinear single- and multi-degree-of-
freedom benchmark problems will be used to investigate the linear
and nonlinear performances of the proposed explicit method.
2. An explicit time integration method
The influence of the spurious root of time integration methods
was studied in Ref. [14]. Even though the influence of the spurious
root is not that huge in some of recent time integration methods as
explained in Refs. [8,14], its presence is not completely acceptable
in a mathematical view point. At least, it should be minimized to
achieve good accuracy for the important low frequency modes.
Here, the time finite element method [15–18] based unconven-
tional Hermite type interpolating techniques are employed as a
remedy for this problem. Many of improved methods, such as
the HC, CL and NB methods [8,9,12], have the spurious roots, and
initial and other acceleration vectors should always be computed
and stored in each time step.
A practical way of designing explicit time integration methods
without a spurious root is to exclude time nodal acceleration vec-
tors from the time approximations of the displacement vector. This
can be done by using proper Hermite type interpolation functions
which are associated with time nodal displacement and velocity
vectors for the approximation of the displacement vector [2,19].
Then, the approximated displacement vector, and the first and sec-
ond time derivatives of the approximated displacement vector can
be substituted into the structural dynamics equations to obtain
approximated structural dynamics equations. The time discretiza-
tion can be completed by evaluating the approximated structural
dynamics equations at a certain point of time in a collocation
44 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53

sense. Weighted residual approaches can also be employed, but the the time nodal displacement and velocity vectors at t ¼ Dt, respec-
resulting discrete relations will be applicable to only linear cases as tively. By using u0 and v 0 , the unknown displacement vector u1 is
shown in the case of the Soares method. determined according to
This unique approach considered in this study gives two serious  
M as1 þ n us1 ; v s1 ¼ f s1 ð9Þ
computational advantages to the new explicit method. First, the
spurious root of the amplification matrices can be excluded in where us1 ; v s1 ; as1 , and f s1 are given as
the new explicit method. Second, computation of the initial accel-
eration vector can be omitted, and computation of any acceleration us1 ¼ 1 uðs1 DtÞ ¼ u0 þ s1 Dt v 0 ð10aÞ
vectors is not required throughout entire procedure. Acceleration
vectors can be computed by using the structural dynamics equa- vs 1
¼ 1 u_ ðs1 Dt Þ ¼ v 0 ð10bÞ
tions if it is necessary. Here we present specific developing proce-
2 2
dures of the new explicit method. € ð s 1 Dt Þ ¼
as1 ¼ 2 u ð u1  u0 Þ  v0 ð10cÞ
The standard form of the structural dynamics equations is ð a Dt Þ 2 a Dt
expressed as
f s 1 ¼ f ð t s þ s 1 Dt Þ ð10dÞ
€ ðtÞ þ nðuðtÞ; uðtÞÞ
Mu _ ¼ fðtÞ ð1Þ
After u1 is determined, the unknown displacement vector u2 can
and the initial conditions are given as
also be determined from
uð0Þ ¼ u0 ; _
uð0Þ ¼ v0 ð2Þ  
M as2 þ n us2 ; v s2 ¼ f s2 ð11Þ
where M is the mass matrix, n is the internal force vector, f is the
where us2 ; v s2 ; as2 , and f s2 are
external force vector, u is the displacement vector, u0 is the vector
of initial displacement, and v 0 is the vector of initial velocity. If us2 ¼ 2 uðs2 DtÞ
Eq. (1) is obtained by spatially discretizing original governing s2 2 ða  s2 Þða þ s2 Þ s 2 ð a  s 2 Þ Dt
_
PDEs, the internal force vector is often expressed as nðuðtÞ; uðtÞÞ ¼ ¼ u þ u0 þ v0 ð12aÞ
a2 1 a2 a
_
CuðtÞ þ KuðtÞ where K and C are the stiffness and damping matrices
respectively. 2 s2 a  2 s2
vs ¼ 2 u_ ðs2 Dt Þ ¼ ðu  u0 Þ þ v0 ð12bÞ
For simple handling of formulations, we use the local time 2
a2 Dt 1 a
t ¼ t  t s , where t s is the beginning of the ðs þ 1Þ-th time step or
the end of the s-th time step. By using u0 and v 0 ; uðtÞ can be 2 ð a  3 s2 Þ 2 ð 3 s2  1Þ
€ ð s 2 Dt Þ ¼
as2 ¼ 3 u u2 þ 2 u1
approximated as ða  1Þ Dt 2 a ða  1Þ Dt2
 
2 a2  a  1 þ 3 s2 a þ 3 s2
uðt Þ  1 uðtÞ ¼ 1 /1 ðtÞu0 þ 1 /2 ðtÞv 0 ð3Þ þ u0
a 2 Dt 2
2 ða þ 3 s2  1Þ
where þ v0 ð12cÞ
a Dt
1
/1 ðtÞ ¼ 1; 1
/2 ðtÞ ¼ t ð4Þ
f s 2 ¼ f ð t s þ s 2 Dt Þ ð12dÞ
By including u1 in addition to u0 and v0 ; uðtÞ can be approximated We note that Eqs. (9) and (11) are obtained by substituting
as approximated displacements (given in Eqs. (3), (5), and (7)) and
their time derivatives into Eq. (1) and evaluating the approximated
 ðtÞ ¼ 2 /1 ðtÞu0 þ 2 /2 ðtÞv 0 þ 2 /3 ðtÞu1
uðt Þ  2 u ð5Þ
structural dynamics equations at t ¼ s1 Dt and t ¼ s2 Dt. This is a
typical approach used in the collocation methods, thus numerical
where u1 is the displacement vector at t ¼ a Dt; Dt is the size of a
characteristics of the resulting discrete relation depend on the
time step, and 2 /i ðtÞ are given as
choices of s1 ; s2 and a. Optimal sets of s1 ; s2 and a will be provided
ðt  a Dt Þðt þ a Dt Þ in the next section.
2
/1 ðtÞ ¼  ; In Eq. (9), the only unknown property is u1 . Simply, u1 is com-
ða Dt Þ2

2 puted as
tðt  a Dt Þ t
2
/2 ðtÞ ¼  ; 2
/3 ðt Þ ¼ ð6Þ
a Dt a Dt ða DtÞ2 1   
u1 ¼ M f s1  n us1 ; v s1 þ u0 þ a Dt v 0 ð13Þ
2
Here, a is a parameter which adjusts specific location of u1 in  
0 < t < Dt. In a similar sense, uðtÞ can also be approximated as In linear case n us1 ; v s1 is expressed by using us1 and v s1 given in
Eqs. (10a) and (10b) as
 ðtÞ ¼ 3 /1 ðtÞu0 þ 3 /2 ðtÞv 0 þ 3 /3 ðtÞu1 þ 3 /4 ðtÞu2
uðt Þ  3 u ð7Þ  
n u s 1 ; v s 1 ¼ C v 0 þ K ð u 0 þ a s 1 Dt v 0 Þ ð14Þ
where u2 is the displacement vector at t ¼ Dt, and
ðt  DtÞðt þ a t þ a DtÞðt  a DtÞ t ðt  Dt Þðt  a DtÞ After computing u1 form Eq. (13), u2 is computed as
3
/1 ðt Þ ¼  ; 3 /2 ðtÞ ¼
a2 Dt3 a Dt2 ða  1Þ Dt 2 1    3 s2  1
t 2 
ð t  Dt Þ 
t 2 
ð t  a D t Þ u2 ¼ M f s2  n us2 ; v s2  u1
3
/3 ðt Þ ¼ 3 ; 3 /4 ðtÞ ¼  ð8Þ 2 ð a  3 s2 Þ ða  3 s2 Þa2
ða  a2 Þ Dt3 ða  1Þ Dt3  2 
ða  1Þ a þ a þ 1  3 s2 a  3 s2
We note that Eq. (8) is not the standard set of Hermite cubic þ u0
ða  3 s2 Þa2
interpolation functions. In the set of the standard cubic Hermite
ða  1Þða þ 1  3 s2 ÞDt
interpolation functions, four interpolation functions are associated þ v0 ð15Þ
with the time nodal displacement and velocity vectors at t ¼ 0, and a ð a  3 s2 Þ
 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53 45

 
In linear cases, n us2 ; v s2 is expressed by using us2 and v s2 given in and the initial conditions are
Eqs. (12a) and (12b) as uð0Þ ¼ u0 ; _
uð0Þ ¼ v0 ð19Þ


  2 s2 a  2 s2 where uðtÞ is the displacement, u0 is the initial displacement, v 0 is
n us 2 ; v s 2 ¼ C ð u  u Þ þ v
a2 Dt 1 0
a 0
the initial velocity, n is the damping coefficient, x is the natural

2 
s2 ð a  s2 Þ ð a þ s2 Þ s2 ða  s2 Þ Dt frequency, and f ðtÞ is the external force. By setting M ¼ 1;
þK u þ u þ v C ¼ 2 n x; K ¼ x2 ; u ¼ u, and f ¼ f , and applying Eqs. (13), (15),
2 a 1 2 a 0 0
a
ð16Þ and (17) to Eqs. (18) and (19), the following equations can be
obtained.
Finally, v2 is computed by using u1 and u2 as       ( )
u2 A11 A12 u0 L11 L12 f s1
¼ þ
v2 ¼
2a  3
u2 þ 2
1
u 
ð2 a þ 1Þða  1Þ
u0 
a1
v v2 A21 A22
|fflfflfflfflfflfflfflffl
ffl{zfflfflfflfflfflfflfflfflffl}
v0 L21 L22
|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}
f s2 ð20Þ
ð a  1 Þ Dt a ð a  1 Þ Dt 1 a2 Dt a 0 A L
ð17Þ
where A is the amplification matrix, L is the load operator matrix, u2
Specific procedures of the new explicit method are completely sum- and v 2 are the displacement and velocity at t ¼ Dt, respectively,
marized in Table 1. f s1 ¼ f ðs1 Dt Þ, and f s2 ¼ f ðs1 Dt Þ. The entries of A and L are given as
As shown in Table 1, the new explicit method does not require 1
computation of the initial acceleration vector a0 . Since u0 and v 0 A11 ¼ a X4  X4 þ 8 a n X3
8 ð2 a  3Þ
are given properties at t ¼ 0, the new explicit method can be
started without any additional preparations. This attribute can also 8 n X3  8 a X2 þ 12 X2 þ 16 a  24 ;
be found in the Soares method. Many of other methods require Dt
computation of a0 at t ¼ 0 to start procedures. A12 ¼ a s1 X4  s1 X4 þ 8 a s1 n X3 þ 2 a n X3
8 ð2 a  3Þ
8 s1 n X3  2 n X3 þ 16 a n2 X2  16 n2 X2  4 a X2
3. Analysis of the new explicit method
16 a n X þ 4 s1 X2 þ 4 X2 þ 24 n X þ 16 a  24 ;
Numerical characteristics of the new explicit method are quite
similar to those of the NB and Soares methods when they are X2 X2 þ 8 n X  8
A21 ¼ ;
applied to the undamped linear single-degree-of-freedom prob-

8 Dt 
lem. To investigate stability and accuracy of the new explicit 1 1 1
method, the single-degree-of-freedom problem [20,21] is consid-
A22 ¼ s1 X4 þ s1 n X3 þ n X3 þ 2 n2 X2  X2  2 n X þ 1
8 4 2
ered. The single-degree-of-freedom problems is expressed as ð21Þ

€ ðtÞ þ 2 n xuðtÞ
u _ þ x uðtÞ ¼ f ðtÞ
2
ð18Þ
Dt2 a X2 þ 8 a n X  X2  8 n X þ 4
L11 ¼ 
8 ð2 a  3Þ
Dt2 ða  1Þ
Table 1 L12 ¼ ð22Þ
Summary of the new explicit method. 2a  3
Dt ðX þ 8 nÞ X
Before starting the procedure, set s1 ; s2 , and a as one of following settings: L21 ¼
– s1 ¼ 0:2500; s2 ¼ 0:5, and a ¼ 0:5000 (no dissipation)
8
– s1 ¼ 0:2684; s2 ¼ 0:5, and a ¼ 0:4219 (moderate dissipation) L22 ¼ Dt
– s1 ¼ 0:2831; s2 ¼ 0:5, and a ¼ 0:3370 (maximum dissipation)
Dt 6 0:1 T is recommended, where T is the shortest period of a given system. where X ¼ xDt. Now, numerical characteristics, such as stability
Repeat following procedures n times to obtain solutions at and accuracy, can be investigated by using the amplification matrix
t ¼ Dt; 2 Dt; 3 Dt; . . . ; n Dt. A given in Eqs. (20) and (21).
1. First sub-step (0 6 t 6 aDt)
Accuracy of the new explicit method can be defined by using
– Guess us1 and v s1 , and compute f s1 at t ¼ s1 Dt as the characteristic polynomial of A. The characteristic polynomial
us1 ¼ u0 þ s1 Dt v 0 of A is expressed as
v s1 ¼ v 0
f s1 ¼ fðts þ s1 DtÞ pðkÞ ¼ k2  2 A1 k þ A2 ð23Þ
where v 0 and v 0 are known properties, t ¼ t  t s ,
and ts is the beginning of the s-th time step. where
– Compute u1 as
   1
u1 ¼ ða D2t Þ M1 f s1  n us1 ; v s1 þ u0 þ a Dt v 0 A1 ¼ 2 a s1 X4  3 s1 X4 þ a X4  X4 þ 16 n a s1 X3
2

16 ð2 a  3Þ
2. Second sub-step (0 6 t 6 Dt)
– Guess us2 and v s2 , and compute f s2 at t ¼ s2 Dt as
þ12 n a X3  24 n s1 X3  14 n X3 þ 32 n2 a X2  48 n2 X2
us2 ¼ sa22 u1 þ ðas2aÞð2aþs2 Þ u0 þ s2 ðaas2 Þ Dt v0 16 a X2 þ 24 X2  32 n a X þ 48 n X þ 32 a  48
2

v s2 ¼ a2 Dt ðu1  u0 Þ þ a2a s2 v 0
2 s2

1
f s2 ¼ fðt s þ s2 DtÞ A2 ¼ 2 a s1 X4 þ s1 X4  a X4 þ 16 n a s1 X3
– Compute u2 and v2as 8 ð2 a  3Þ
  
u2 ¼ 2ðað1Þ Dt 2
a3 s2 Þ M1
f s2  n us2 ; v s2  ða33s2s1Þa2 u1 4 n a X3  24 n s1 X3 þ 10 n X3 þ 32 n2 a X2  48 n2 X2
2

ða1Þða2 þaþ13 s2 a3 s2 Þ

þ ða3 s2 Þa2
u0 þ ða1aÞððaaþ13 s2 ÞDt
3 s2 Þ v 0 32 n a X þ 48 n X þ 16 a  24Þ
ð2 aþ1Þða1Þ
v2 ¼ 2 a3
ða1Þ Dt u2 þ a2 ða1Þ Dt u1 
1
a2 Dt u0  a1 a v0 ð24Þ
3. Update u0 and v 0 as u2 and v 2 , respectively,
With the invariants given in Eq. (24), the local truncation error
and repeat from the first sub-step to advance another time step.
[4,7,20] is defined as
46 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53
1 completely extinguished by adjusting the parameters in dissipative
se ð t s Þ ¼ ðuðt s þ DtÞ  2 A1 uðt s Þ þ A2 uðt s  DtÞÞ
ðDt Þ2
where uðtÞ is the exact solution of the single-degree-of-freedom
problem for the unforced case (i.e., f ðtÞ ¼ 0). According to Eq.
(25), the explicit method is k-th order accurate if
 
s e ð t s Þ ¼ O Dt k
The choice of s2 ¼ 12 gives
x3 
se ðts Þ ¼ 96 n2 þ 64 a n2 þ 48 a s1  12 a  72 s1 þ 30
24 ð2 a  3Þ
    
 n u0 þ 32 n2 a þ 3  48 n2 þ 2 a  12 s1 x v 0 Dt2 þ O Dt3
As shown in Eq. (27), the new explicit methods becomes second-
order accurate for any choices of a and s1 if s2 ¼ 12 is used. Since
the viscous damping term is included in Eq. (27), the new algo-
rithms is always second-order accurate even in the presence of
the damping matrix C. We note that the choice s ¼ 12 ensures
second-order accuracy for the new explicit method. Controllable
numerical dissipation capability can be achieved through s1 and a.
The spectral radius [4,12] is a very useful for the investigation of
stability and numerical dissipation in time integration methods.
With the undamped and unforced case of the single-degree-of-
freedom problem (i.e., n ¼ 0 and f ðtÞ ¼ 0), the spectral radius is
defined as
qðAÞ ¼ max ðjk1 j; jk2 jÞ
where k1 and k2 are the eigenvalues of A. A method is stable if
0 < qðAÞ 6 1 is provided. Fig. 1 shows that the explicit method is
conditionally stable. Through different setting of s1 and a, amounts
of numerical dissipations in the new explicit method are effectively
controlled as presented in Fig. 1. We note that proposed values
of a and s1 are numerically determined. For example, a ¼ 0:5;
s1 ¼ 0:25, and s2 ¼ 0:5 give non-dissipative case, and a ¼ 0:337;
s1 ¼ 0:2831, and s2 ¼ 0:5 give the most dissipative case.
The stability limit of the non-dissipative cases of the new expli-
cit method, the NB method and the Soares method for the
undamped problem is identical. We also note that the stability
limit of the non-dissipative cases of these explicit methods is twice
of the central different method.
There are perturbation points (bubble shapes) around X ¼ 2:8
in the spectral radii of the new explicit method as shown in
Fig. 1. We note that these perturbation points cannot be
Fig. 1. Spectral radii of proposed explicit method. X ¼ x Dt is used.
ð25Þ
cases, only the non-dissipative case (s1 ¼ 0:25; s2 ¼ 0:5, and
a ¼ 0:5) does not have a perturbation point. However, these per-
turbation points do not affect stability characteristics of the algo-
rithm, because their peak points are well under the line of
qðAÞ ¼ 1:0 for all recommended settings of parameters.
In fact, the newly developed explicit method shares many of lin-
ð26Þ
ear characteristics with the NB and the Soares methods. These
methods have almost identical numerical characteristics when
they are applied to the linear undamped single-degree-of-
 freedom problems with similar level of numerical dissipations.
Noh and Bathe emphasized that their explicit method with
p ¼ 0:54 could share almost identical spectral radius with the
ð27Þ implicit Bathe method for Dt=T 6 0:3 [12], where the period of
the single-degree-of-freedom problem is 2p=x. The new explicit
method with a ¼ 0:4219; s1 ¼ 0:2684, and s2 ¼ 0:5 shares almost
identical numerical characteristics with the NB method (the case
of p ¼ 0:54). Thus, we recommend a ¼ 0:4219; s1 ¼ 0:2684, and
s2 ¼ 0:5 as the standard setting of parameters for the new explicit
method.
The period elongation and the damping ratio are practical mea-
surements of dispersion and dissipation errors, respectively. They
directly describe important characteristics of time integration
 
algorithms. Here, the relative period error is defined as T  T =T
where, the exact period is T ¼ 2p=x, and the numerically obtained
period is T ¼ 2p=x  . The algorithmic damping error is defined as

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  Þ. Here x
n ¼  lnðA2 Þ=ð2 X  ¼ acrtan A2 =A21  1 = Dt 1  n2
ð28Þ
 ¼x
and X  Dt. Please see Ref. [21] regarding details of the accuracy
measurements.
The period elongation and the damping ratio of the new explicit
method is presented in Figs. 2 and 3. Fig. 2 shows that the period
error of the new explicit method decreases as the amount of
numerical dissipation increases, while Fig. 2 shows that the damp-
ing error of the new explicit method increases as the amount of
numerical dissipation increases.
We note that the accuracy characteristics of the new explicit
method shown in Figs. 2 and 3 are almost identical to those of
the NB method and the Soares method in linear cases, if similar
levels of numerical dissipations are used for the NB method and
the Soares method. In other word, the period and damping errors
obtained from the NB method and the Soares method will super-
pose the result given in Figs. 2 and 3, if the algorithmic parameters
of the methods are set to yield similar levels of numerical dissipa-
tions. However, the spectral radii of the NB method and the Soares
Fig. 2. Period errors of proposed explicit method.
 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53 47

methods used Dt ¼ 0:08, and the TW and CL method used

Dt ¼ 0:04 to equalize overall computational efforts of methods.
For the second case, the new explicit, NB, and Soares methods used
Dt ¼ 0:16, and the TW and CL method used Dt ¼ 0:08 to equalize
overall computational efforts of methods.
As shown in Figs. 4 and 5, the new explicit, NB, and CL methods
presented better performances compared with other explicit
methods for 0:0 6 t 6 3:0. Unlike the undamped case, the Soares
and TW methods presented large period errors, as shown in Figs. 4
and 5. When even larger time step is used, the solutions obtained
from the TW and Soares methods presented noticeably large per-
iod errors, while the solutions obtained from the new explicit,
NB, and CL methods were very accurate. This is due to the lower
order of accuracy of the Soares and TW methods. We note that
the Soares method becomes only first order accurate in the pres-
ence of the viscous damping terms and the TW method is always
first order accurate for all dissipative cases.
Fig. 3. Damping ratios of proposed explicit method. Since general linear performances of time integration methods
can be measured by using the single-degree-of-freedom problem
[17,20], we may conclude that the new explicit, NB, and CL

method do not present any perturbation points (bubble shapes),

which is contrary to the case of the new explicit method as shown
in Fig. 1.
We also note that the computational efforts per time step of the
new explicit method and the NB method are almost identical after
the first time step. However, unlike the new explicit method, the
NB method requires the initial acceleration vector at the first step
which is computed as u € 0 ¼ M1 ðf  Cu_ 0  Ku0 Þ. In linear analyses,
both methods include two sub-steps which accompany two evalu-
ations of matrix and vector operations (i.e., M1 ðf  Cu_  KuÞ,
where u and u_ are known properties from previous steps or
approximated properties at certain time points.). In nonlinear anal-
yses, both methods include two sub-steps which accompany two
evaluations of nonlinear vectors and matrix operations (i.e.,
M1 ðf  nðu; u_ ÞÞ). Other pure vector operations do not include
any matrix operations, thus they do not affect overall computation
times significantly even though the new explicit method includes
less vector operations than the NB method. For this reason, the
computation times of the two methods are almost identical in Fig. 4. Comparison of displacements of damped linear single-degree-of-freedom
practical analyses. However, it should be noted that the new problem with n ¼ 0:1. The new explicit, NB, and Soares methods used Dt ¼ 0:08,
and the TW and CL method used Dt ¼ 0:04 to equalize overall computational efforts
explicit method do not require memory allocations for acceleration of methods.
vectors.

4. Illustrative examples

To verify improved performance of the new explicit method,

four simple and illustrative test examples are solved by using the
new and recently developed explicit methods, and their numerical
solutions are compare with each other. To equalize overall compu-
tational efforts of the explicit methods, the TW and CL methods use
one half of time steps used in the new explicit method, and the NB
and Soares methods used the same size of the time step used in the
new explicit method. The Soares method is not used to solve non-
linear examples, because it is not applicable to nonlinear analyses
in its present form. We also note that 16 significant digits of preci-
sion is use in all computations, and increased precision level of the
computations did not changed results of the analyses conducted
herein.

4.1. Damped linear single-degree-of-freedom problem

Fig. 5. Comparison of displacements of damped linear single-degree-of-freedom

As a damped case, Eq. (18) is solved by various explicit meth-
problem with n ¼ 0:1. The new explicit, NB, and Soares methods used Dt ¼ 0:16,
ods. u0 ¼ 1:0; v 0 ¼ 0:0; n ¼ 0:1; f ðtÞ ¼ 0, and x ¼ 2 p are used for and the TW and CL method used Dt ¼ 0:08 to equalize overall computational efforts
the analysis. For the first case, the new explicit, NB, and Soares of methods.
48 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53
methods are suitable for both damped and undamped cases, while
the Soares method is only suitable for undamped cases.
4.2. Simple nonlinear pendulum problem
Since the new explicit and NB methods share almost identical
linear spectral characteristics, they will give almost identical
numerical solution when they are applied to linear problems. How-
ever, different time integration methods may perform quite differ-
ently in nonlinear cases, even though they have identical linear
spectral characteristic as discussed in [3]. Due to this reason, the
oscillating nonlinear single pendulum [3,4,22] is used to compare
nonlinear performances of the new explicit method and other
explicit methods.
The oscillation of the single pendulum is described in Fig. 6.
Motion of a single pendulum is described by
€h þ x2 sinðhÞ ¼ 0
with initial conditions
hð0Þ ¼ h0
_ solutions of the NB and TW methods presented very large period
hð0Þ ¼ h_ 0
pffiffiffiffiffiffiffiffi
where x ¼ g=L, and hðtÞ is the angle between the rigid rod and the
vertical line at time t. g is the gravitational constant, and L is the
length of the massless rigid rod. Here, we assumed a dimensionless
case of g ¼ 1:0 and L ¼ 1:0.
Important information pertaining to the oscillating pendulum
problem (such as nonlinear period and maximum angle) can be
exactly obtained. For details, please see Refs. [4,22,23]. Here, we
use h0 ¼ 0:0 and h_ 0 ¼ 1:999999238456499 to synthesize a highly
nonlinear case. For the chosen initial conditions, maximum angle
hmax becomes 3:139847324337799, and the period T becomes
33:72102056485366.
For the first case, the new explicit and NB methods used
Dt ¼ T=100; 000 and the TW and CL method used
Dt ¼ T=200; 000 to equalize overall computational efforts of meth-
ods. As shown in Fig. 7, all methods superposed the exact solution
except the TW method. Dt ¼ T=200; 000 is used for the TW
method, but the TW method presented noticeable period error.
For the second case, the new explicit and NB methods used
Dt ¼ T=10; 000 and the TW and CL method used Dt ¼ T=20; 000
to equalize overall computational efforts of methods. With these
larger time steps, the solutions of the new explicit method and
the CL method were the closest to the exact solution, while the
Fig. 6. Description of nonlinear single pendulum [4].
ð29Þ
Fig. 7. Comparison of angles. The new explicit and NB methods used
Dt ¼ T=100; 000 and the TW and CL method used Dt ¼ T=200; 000 to equalize
overall computational efforts of the methods.
ð30aÞ
ð30bÞ
errors.
For the third case, the new explicit and NB methods used
Dt ¼ T=1000 and the TW and CL method used Dt ¼ T=2000 to
equalize overall computational efforts of methods. With these even
larger time steps, the NB and CL methods presented very interest-
ing results. The solution of the NB and CL methods indicated com-
pletely rotating pendulum about the pinned point instead of
oscillating pendulum. The solutions of the new explicit method
and the TW method presented increased period errors compared
with the previous cases of smaller time steps, but their results
showed oscillating pendulum.
In addition to the results of the pendulum problems reported
herein, we also investigated numerical results of the explicit meth-
ods obtained by using other settings of algorithmic parameters. To
be specific, the non-dissipative (p ¼ 0:5) and the most dissipative
pffiffiffi
(p ¼ 2  2) cases are additionally considered for the NB methods,
and the non-dissipative (a ¼ 0:5; s1 ¼ 0:25, and s2 ¼ 0:5) and the
most dissipative (a ¼ 0:337; s1 ¼ 0:2831, and s2 ¼ 0:5) cases are
additionally considered for the new explicit method. For all addi-
tional settings of algorithmic parameters, different accuracy set-
tings were also tested. When 16 significant digits of precision
was use, all additional cases gave numerical predictions similar
to those presented in Figs. 8 and 9. When 8 significant digits of pre-
cision was use, however, the non-dissipative case of the new expli-
cit method (i.e., a ¼ 0:5; s1 ¼ 0:25; s2 ¼ 0:5) presented result of
rotating pendulum, while other cases of the new explicit method
presented results of oscillating pendulum. When slightly increased
accuracy level (10 significant digits of precision) was used, the pre-
dicted result (oscillating pendulum) was also obtained with the
case a ¼ 0:5; s1 ¼ 0:25; s2 ¼ 0:5 of the new explicit method. Unlike
the new explicit method, the NB method could not give predicted
results by using the different settings of parameters and accuracy
levels stated above.
For the completeness of the comparison, an additional case is
also tested. If the initial conditions are chosen as h0 ¼ 0:0 and
h_ 0 ¼ 2:000003, the pendulum should make complete rotation
instead of oscillation. The 10th-order accurate implicit method of
Kim and Reddy [3] and the RK4 methods are used to numerically
obtain the reference solution. For this case, numerically obtained
nonlinear period is T ¼ 15:4895.
When large time steps (i.e., Dt ¼ T=1; 000 for the new and NB
methods and Dt ¼ T=2; 000 for the CL and TW methods) are used,
the CL and NB methods gave correct results (i.e., completely
 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53 49

Fig. 8. Comparison of angles. The new explicit and NB methods used Dt ¼ T=10; 000 Fig. 10. Comparison of angles. A completely rotating case is tested. The new explicit
and the TW and CL method used Dt ¼ T=20; 000 to equalize overall computational and NB methods used Dt ¼ T=1000 and the TW and CL method used Dt ¼ T=2000 to
efforts of the methods. equalize overall computational efforts of the methods.

Fig. 9. Comparison of angles. The new explicit and NB methods used Dt ¼ T=1000 Fig. 11. Comparison of angles. A completely rotating case is tested. The new explicit
and the TW and CL method used Dt ¼ T=2000 to equalize overall computational and NB methods used Dt ¼ T=10; 000 and the TW and CL method used
efforts of the methods. Dt ¼ T=20; 000 to equalize overall computational efforts of the methods.

rotating pendulum), while the newly proposed method and the TW

where s > 0:0. Here we use s ¼ 100:0; u0 ¼ 4:0, and v 0 ¼ 4:0 as
method gave completely opposite results as shown in Fig. 10.
given in Ref. [24]. The new explicit and NB methods used
Among the four methods tested, the CL method presented the
Dt ¼ T=40 and the TW and CL method used Dt ¼ T=80 to equalize
result which is the closest to the reference solution.
overall computational efforts of the methods, where the period T
When smaller time steps (i.e., Dt ¼ T=10; 000 for the new and
is 1:14187. The exact solution was obtained by using the 4th-
NB methods and Dt ¼ T=20; 000 for the CL and TW methods) are
order accurate Runge-Kutta method with very small time step
used, however, the new, NB, and CL methods gave accurate results,
(i.e., Dt ¼ T=10; 000).
while the TW method presented incorrect result as shown in
The elastic spring problem given in Eq. (31) is a conservative
Fig. 11.
system [4]. In conservative systems, the exact displacement-
velocity portrait should be a closed cycle shown as dotted lines
4.3. Elastic spring problem
in Fig. 12. We noticed that displacement-velocity portraits of all
methods presented closed cycles as presented in Fig. 12. However,
The elastic spring equation [24–26] is solved by using various
the displacement-velocity portrait of the TW method noticeably
explicit methods. The motion of the elastic spring is given as
deviated from the exact portrait as shown in Fig. 12(c).
€ þ s tanh ðuÞ ¼ 0
u ð31aÞ On contrary to the displacement-velocity portraits presented in
Fig. 12, the explicit methods presented quite different aspects in
uð0Þ ¼ u0 ð31bÞ displacement solutions when they are tested for a long duration
of time. All four methods gave very correct displacement solutions
_
uð0Þ ¼ v0 ð31cÞ for 0 6 t 6 2 T as presented in Fig. 13(a). As presented in Fig. 13(b),
50 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53
€ þ s tanh ðuÞ ¼ 0; s ¼ 100:0; uð0Þ ¼ 4:0, and uð0Þ
Fig. 12. Phase portrait of the softening spring problem u
method; (d) CL method.
however, the TW method presented the largest period error, and
the CL and NB methods presented the second largest period errors
for 25 T 6 t 6 27 T. We note that the displacement solutions of the
CL and NB methods superposed each other, and they were shown
as the single blue line in Fig. 13(b). The new explicit method gave
the most accurate displacement solution as shown in Fig. 13(b).
4.4. Elastic spring-pendulum problem
The configuration of the two-degree-of-freedom elastic spring-
pendulum problem [4,7] is described in Fig. 14. The governing
equations are given as
m €r  m ðL0 þ r Þ h_ 2  m g cos h þ k r ¼ 0
m 2 r_ h_ þ g sin h
m €h þ ¼0
L0 þ r
with initial conditions
_ cit and NB methods almost superposed each other in the first case,
rð0Þ ¼ r 0 ; hð0Þ ¼ h0 ; r_ ð0Þ ¼ r_ 0 ; hð0Þ ¼ h_ 0
where r and h are the displacements in the radial and circumferen-
tial directions, respectively, m is the mass of the pendulum, g is the
gravity, L0 is the length of the undeformed spring, and k is the spring
constant. In this numerical test, we use m ¼ 1:0; g ¼ 9:81;
L0 ¼ 0:5; k ¼ 98:1; r 0 ¼ 0:25; h0 ¼ p=2 ¼ 1:570796326794896; r_ 0 ¼ 0; 0
_ ¼ 4:0 for 100  T. (a) New explicit method; (b) NB method; (c) TW
and h_ 0 ¼ 0:0. The reference solution was obtained by using the 4th-
order Runge-Kutta method with very small time step (i.e.,
Dt ¼ 0:0001).
In fact, this example was considered by Chung and Lee [8]
to demonstrate improved performance of the CL method. In
Ref. [8] h0 was chosen as 0:6981317007977317. With
h0 ¼ 0:6981317007977317, the CL method was able to give
very accurate solutions. However, we modified h0 as
1:570796326794896 to synthesize a highly nonlinear case for the
investigation of nonlinear performances of the explicit methods.
For the first case, the new explicit and NB methods used
Dt ¼ 0:01 and the TW and CL method used Dt ¼ 0:005 to equalize
overall computational efforts of the methods. In the first case, all
ð32aÞ methods performed well except the CL method. The new explicit
and NB methods gave the most accurate solutions as presented
in Figs. 15 and 16. The solutions of the TW method slightly devi-
ð32bÞ ated from the reference solutions, while the solutions of the CL
method noticeably deviated from the reference solutions as pre-
sented in Figs. 15 and 16. We note that the results of the new expli-
ð33Þ
and appeared as the pink line in the figures.
For the second case, the new explicit and NB methods used
Dt ¼ 0:05 and the TW and CL method used Dt ¼ 0:025. In this case,
only the new explicit and NB methods performed well. The new
explicit and NB methods gave very accurate solutions as presented
in Figs. 17 and 18. The displacements in the radial direction of the
 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53 51

€ þ s tanh ðuÞ ¼ 0; s ¼ 100:0; uð0Þ ¼ 4:0, and uð0Þ

Fig. 13. Displacement solution of the softening spring problem u _ ¼ 4:0. (a) Short term solutions (0 6 t 6 2 T); (b) long term
solutions (25 T 6 t 6 27 T). T ¼ 1:14187.

Fig. 15. Comparison of displacements in the radial direction. The new explicit and
Fig. 14. Description of elastic spring pendulum [4,7]. NB methods used Dt ¼ 0:01 and the TW and CL method used Dt ¼ 0:005 to equalize
overall computational efforts of the methods.

TW method became ‘‘unstable ” after some steps, and the angles in

developed in this article can control algorithmic dissipation through
the circumferential direction of the TW method aso became unsta-
the parameters. Remarkably, the new explicit method could provide
ble as presented in Figs. 17 and 18. We also note that the results of
equivalent or even better nonlinear solutions with less computa-
the new explicit and NB methods superposed each other in the sec-
tional effort due to the exclusion of acceleration vectors.
ond case, and appeared as the pink line in figures.
The advantages of the new explicit method can be summarized
as follows:
5. Concluding remarks
(a) The amplification matrix does not have a spurious root.
The new explicit method was developed based on the time collo- (b) Computations of the acceleration vectors are not required
cation finite element approach and the unconventional Hermite throughout the entire procedure, but only the displacement
type interpolation functions in time. The new explicit method and velocity vectors are involved in.
52 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53
Fig. 16. Comparison of angles in the circumferential direction. The new explicit and
NB methods used Dt ¼ 0:01 and the TW and CL method used Dt ¼ 0:005 to equalize
overall computational efforts of the methods.
Fig. 17. Comparison of displacements in the radial direction. The new explicit and
NB methods used Dt ¼ 0:05 and the TW and CL method used Dt ¼ 0:025 to equalize
overall computational efforts of the methods.
Fig. 18. Comparison of angles in the circumferential direction. The new explicit and
NB methods used Dt ¼ 0:05 and the TW and CL method used Dt ¼ 0:025 to equalize
overall computational efforts of the methods.
(c) Second-order accuracy and control of numerical dissipation
are achieved through the optimized algorithmic parameters.
(d) It can provide equivalent or improved solutions compared
with existing explicit methods, when they are applied to
various challenging nonlinear dynamic problems.
(e) It does not require any iterative nonlinear solution finding
procedures, such as the Newton-Raphson and Picard meth-
ods within a certain time step, but just requires evaluations
of nonlinear vectors which can be easily done by using
known properties of previous steps.
(f) The computer implementation of the new explicit method
can be done easily, if the Noh and Bathe method is already
implemented on computers due to similar computational
structures of the two methods.
As shown in the numerical tests herein, the new explicit
method could provide more accurate solutions compared with
many of well-known explicit methods. Our future study will
include applications of the new explicit method to classical impact
and wave propagation problems to provide more complete infor-
mation regarding the performance of the method to users.
Acknowledgments
The research was carried out with the support of the Republic of
Korea (ROK) Army. The first author truly appreciates supports of
Dalsoo Yoon and Seung-Im Paik.
References
[1] Kim W, Reddy JN. An improved time integration algorithm: a collocation time
finite element approach. Int J Struct Stab Dyn 2017;17(02):1750024.
[2] Kim W, Reddy JN. Effective higher-order time integration algorithms for the
analysis of linear structural dynamics. J Appl Mech 2017;84(7):071009.
[3] Kim W, Reddy JN. A new family of higher-order time integration algorithms for
the analysis of structural dynamics. J Appl Mech 2017;84(7):071008.
[4] Kim W, Choi SY. An improved implicit time integration algorithm: the
generalized composite time integration algorithm. Comput Struct 2018;196:
341–54.
[5] Kim W. Improved time integration algorithms for the analysis of structural
dynamics [Ph.D. Thesis]. Texas A & M University; 2016.
[6] Newmark NM. A method of computation for structural dynamics. J Eng Mech
Divis 1959;85(3):67–94.
[7] Chung J, Hulbert GM. A time integration algorithm for structural dynamics
with improved numerical dissipation: the generalized-alpha method. J Appl
Mech 1993;60:271–5.
[8] Chung J, Lee JM. A new family of explicit time integration methods for linear
and non-linear structural dynamics. Int J Numer Meth Eng 1994;37
(23):3961–76.
[9] Hulbert GM, Chung J. Explicit time integration algorithms for structural
dynamics with optimal numerical dissipation. Comput Methods Appl Mech
Eng 1996;137(2):175–88.
[10] Maheo L, Grolleau V, Rio G. Numerical damping of spurious oscillations: a
comparison between the bulk viscosity method and the explicit dissipative
Tchamwa–Wielgosz scheme. Comput Mech 2013;51(1):109–28.
[11] Tchamwa B, Conway T, Wielgosz C. An accurate explicit direct time integration
method for computational structural dynamics. ASME-PUBLICATIONS-PVP
1999;398:77–84.
[12] Noh G, Bathe KJ. An explicit time integration scheme for the analysis of wave
propagations. Comput Struct 2013;129:178–93.
[13] Soares D. A novel family of explicit time marching techniques for structural
dynamics and wave propagation models. Comput Methods Appl Mech Eng
2016;311:838–55.
[14] Hulbert GM, Chung J. The unimportance of the spurious root of time
integration algorithms for structural dynamics. Commun Numer Methods
Eng 1994;10(8):591–7.
[15] Reddy JN. An introduction to nonlinear finite element analysis: with
applications to heat transfer, fluid mechanics, and solid mechanics. Oxford;
2014.
[16] Hughes TJR. The finite element method: linear static and dynamic finite
element analysis. Courier Corporation; 2012.
[17] Zienkiewicz OC, Taylor RL. The finite element method for solid and structural
mechanics. Heinemann: Butterworth; 2005.
[18] Bathe Klaus-Jürgen. Finite element procedures. Bathe K.J.; 2006.
 W. Kim, J.H. Lee / Computers and Structures 206 (2018) 42–53
[19] Fung TC. Unconditionally stable higher-order accurate hermitian time finite
elements. Int J Numer Meth Eng 1996;39(20):3475–95.
[20] Hilber HM. Analysis and design of numerical integration methods in structural
dynamics [Ph.D. Thesis]. University of California Berkely; 1976.
[21] Hughes TJR. Analysis of transient algorithms with particular reference to
stability behavior. In: Computational methods for transient analysis (A 84-
29160 12-64). Amsterdam, North-Holland; 1983. p. 67–155.
[22] Fung TC. On the equivalence of the time domain differential quadrature
method and the dissipative Runge–Kutta collocation method. Int J Numer
Meth Eng 2002;53(2):409–31.
53
[23] Fung TC. Solving initial value problems by differential quadrature method-part
2: second-and higher-order equations. Int J Numer Meth Eng 2001;50
(6):1429–54.
[24] Xie YM. An assessment of time integration schemes for non-linear dynamic
equations. J Sound Vib 1996;192(1):321–31.
[25] Liu J, Wang X. An assessment of the differential quadrature time integration
scheme for nonlinear dynamic equations. J Sound Vib 2008;314(1):246–53.
[26] Wood WL, Oduor ME. Stability properties of some algorithms for the solution
of nonlinear dynamic vibration equations. Int J Numer Methods Biomed Eng
1988;4(2):205–12.