Time integration methods still questions

&]HVDZ%DMHU
Institute of Fundamental Technological Research, Polish Academy of Sciences,
ZL
WRNU]\VND-049 Warsaw, Poland

Streszczenie. : SUDF\ RPyZLRQH ]RVWD\ QLHVWRVRZDQH Z SUDNW\FH PHWRG\ FDNRZDQLD

UyZQD Uy*QLF]NRZ\FK UXFKX 6 RQH QLH]Z\NOH VNXWHF]QH QXPHU\F]QLH RUD] PDM
NRU]\VWQLHMV]H ZDVQRFL RG SRZV]HFKQLH VWRVRZDQHM PHWRG\ 1HZPDUND F]\ PHWRG\
Uy*QLFFHQWUDOQ\FK1LHVWHW\SR]DPLHMVFHPZSXEOLNDFMDFKVDP\FKWZyUFyZQLHZHV]\
dR V]HUV]\FK ]DVWRVRZD 0HWRG\ %RVVDND 3DUND-Housnera, Hilbera-Hughesa-Taylora,
HOHPHQWyZ F]DVRSU]HVWU]HQQ\FK SU]HGVWDZLRQH Z QLQLHMV]HM SUDF\ PRJ E\
EH]SRUHGQLR ]DVWRVRZDQH Z NRPHUF\MQ\FK SDNLHWDFK REOLF]HQLRZ\FK EH] LVWRWQ\FK
modyfikacji.

1. Introduction
Time integration methods for several of years were intensively investigated. In hundreds of
publications both new methods and their properties were broadly described. Unfortunately
nowadays in the engineering practice only few of them are in regular use. Usually the
selection is done taking into account the accessibility of procedures rather than numerical
quality. Newmark method and the central difference method are employed in almost totality
of structural dynamic analysis problems. The simplicity is a great advantage of these methods.
However, alternative computational schemes are not more complicated. They enable the user
a wide range of useful properties instead.
The decision: implicit or explicit methods, depends on the problem to be solved.
Refined spatial mesh decreases the approximation error and strongly increases the
computational time, because of both the total number of degrees of freedom increase and the
time step decrease involved by the stability criterion. However, experiences in the practical
use of time integration methods are low. Some properties of the methods are described in the
academic literature. The Wilson method is too dissipative in lower modes. It requires a time
step smaller than needed for required accuracy. The Houbolt method is even more dissipative
than the Wilson method. It has no parameter to control this property. The damping is
controlled in practice by the time step value.

In the opinion of the author, the best time integration method should has the following
features:
S
should be unconditionally stable,
S
should have the numerical dissipation controlled by a parameter (in a particular case
should have no dissipation),
S
the numerical dissipation should affect higher modes; lover modes should not be affected,
S
numerical effort should be low enough, comparing with explicit methods,
S
should permit computations of non-inertia structures with the motion forced
kinematically.
The last point of the above list concerns for example crashworthiness problems.
In the paper we recall some efficient schemes, rarely used, with interesting nonclassical features. The algorithms which perform computations in practice are as simple as in
commonly applied methods and can be alternatively introduced to computational codes.
The following features are important in practical use:
S
computational cost,
S
accuracy (phase error),
S
stability,
S
damping of high and low frequencies,
S
scheme of the propagation of information (important in wave problems),
S
type of inertia matrices, fundamental for finite displacements and rotations.
Below we discuss the following methods:
S
implicit methods: Newmark, Bossak, Hilber-Hughes-Taylor, space-time element method,
S
semi-implicit methods: Park-Housner, Trujillo.
The scheme of numerical methods applied to structural dynamics is depicted in Fig. 1.

Fig. 1. Numerical methods employed in structural dynamics.

Although discrete methods are broadly applied to structural dynamics, the qualitative progress
has not been made for several recent years. The following questions are still open.
S
Parabolic type of the solution of the hyperbolic differential equation; numerical velocity of
the information flow is higher than the physical wave speed.
S
Inertia matrix does not result in accurate period of vibration, especially if applied to finite
rotations.
S
Spurious oscillations in fine meshes are hardly eliminated.
That is why so many methods have been elaborated, with the hope to improve at least one of
the mentioned features.

2. The Newmark method

The Newmark method [1], well known and commonly applied in computations, is presented
here since it is a particular case of the methods described in successive paragraphs. The force
equilibrium is determined in ti+1. Three equations (Tab. 1) allow to determine the
displacement vector by solving the system of equations. Complementary computations are
carried out to determine velocity and acceleration vector.
Table 1. The Newmark scheme.
S un+1=un+hvn+h2(1/2-)an+h2an+1
S vn+1=vn+h(1-)an+han+1
S Man+1+Cvn+1 +Kun+1 =Fn+1
In the case of =1/2 and  we have unconditionally stable scheme. The maximum
numerical damping of higher frequencies is achieved for =1/4(+1/2)2, with >1/2.

Fig. 2. Spectral radius of the transfer matrix for Newmark method.

3. The Bossak method

The Bossak method [2] is the extension of the Newmark method. The acceleration a is taken
prior to ti+1. The method can successfully replace the Newmark method in all cases.

Table 2. The Bossak scheme.

Sun+1=un+hvn+h2(1/2-) an+h2 an+1
Svn+1=vn+h(1-)an+h an+1
SM(1- B) an+1+ MBan+ Cvn+1 +Kun+1 =Fn+1
In the case of B=0 we have the Newmark method.
Stability conditions are fulfil for
B 

B B/2 

B + B 

The spectral radius, which determine the numerical dissipation, is depicted in Fig. 3. The
following values of parameters are assumed:
B1: B= 0.1, B=0.3025, B=0.6,
B2: B= 0.1, B=0.5000, B=0.6,
B3: B= +0.1, B=0.3025, B=0.6.
The set of Hilber-Hughes-Taylor parameters are as follows:
HHT1: H= 0.1, =0.3025, =0.6,
HHT2: H= 0.3, =0.3025, =0.6.

Fig. 3. Spectral radius for selected time integration methods [2].

The Bossak method is characteristic of a good damping in high frequencies range and less
sensible to the wrong choice than Hilber-Hughes-Taylor method.
4. Hilber-Hughes-Taylor method
The elastic forces are taken here between tn+1 and tn+1 (H is negative in the original
publication [3]), i.e. in tn+1+ H h. In the case of H=0 we have the Newmark method. The
effect of the artificial damping for this method is depicted in Fig. 3. The authors of this
method do not give the range of application, mutual relation between parameters H, and
and their influence on the stability condition. Numerical tests performed by the author of the
present paper proved that the change of the parameters should be done with attention.
The method can be considered as the alternative to the Bossak method. However, since
it contributes potential forces not clearly definite, applications to nonlinear problems should
be investigated.

Table 4. The Hilber-Hughes-Taylor scheme.

S un+1=un+h vn+h2 (1/2-) an+h2 an+1
S vn+1=vn+h(1-)an+h an+1
S Man+1+ (1+H) Kun+1 H K un=Fn+1
5. The Park-Housner method
The Park-Housner method [4] is the example of semi-implicit methods. It employs the best
features of both the implicit and explicit methods: low numerical cost and memory storage
requirements with unconditional stability. The mass matrix M is diagonal. The stiffness
matrix K is split into a sum of triangular matrices. Two systems of equations with triangular
matrices are to be solved. Finally resulting displacement and velocity vectors are computed
(Tab. 5).
Table 5. Park-Housner algorithm.
SForm K and diagonal M
SSplit K into KL and KU (K= KL+KU , KL=KUT)
SBuild the matrices of systems of equations:
L = M(I+h2M-1KL) , U = I+h2M-1KU,
gn+1= h2 [ fn+1+(1-)fn ] + M(un+hvn)
SSolve the systems of equations (triangular matrices):
Lyn+1= gn+1 , Uu*n+1 = yn+1
SSolve:
un+1=1/[u*n+1-(1-)un] , vn+1=1/(h) (un+1-un)-(1-)/ vn
The stability analysis exhibits for which values and the stability is ensured (Fig. 4).

Fig. 4. Spectral radius

of the transfer matrix
for Park-Housner
method.
In only one point (=1/2, =1/2) we have the unconditional stability of the method without
artificial amplitude decrease. The damping for ,>0.5 is significant and affects low
frequencies considerably.
Fig. 4. Stability domain for
Park-Housner method.

The incoincidence with the theoretical solution (Fig. 7)

concerns the oscillation decay. In means that the system
response is trapezoidal rather than rectangular one.

6. Trujillo method
In the Trujillo semi-implicit method the inertia matrix M is diagonal. Matrices K and C are
symmetric and positive definite. The restriction of a diagonal matrix M may not be severe. K
and C are split into lower and upper triangular matrices: KL+KU=K, CL+CU=C. The
symmetric splitting was investigated in the original paper [5].

Table 4. Trujillo scheme.

h
h
h
h2
h
h2
M + CL + KL vj +1/2 = M CU KL vj K uj + (fj + fj +1)

2
4
2
8
2
8

h
(vj + vj +1)
4

h
h
h
h2
h
h2

KU vj +1 /2 K uj +1 /2 + (fj + fj +1/2 )
M + CU + KU
vj +1 = M CL

2
4
2
8
2
8

uj +1/2 = uj +

h
(vj +1/2 + vj +1)
4
K = KL + KU, K = KL + KU, h = 2t ,
uj +1 = uj +1/2 +

t = jh = 2 jt

The test example depicted in Fig. 7 shows that results better coincide with theoretical
response, then those obtained by Park-Housner method.

7. Space-time element method

The space-time finite element method has two formulations. The early one is expressed in
terms of displacements [6,7]. The time space could be split into rectangular, triangular or even
less regular meshes. The displacement formulation results in the three-level scheme
Aqn-1+Bqn+Cqn+1=Fi.
The second type of formulation is expressed in terms of velocities [8,9]. The step-by-step
scheme is the following
Avn+Bvn+1+sn=Fn.
A, B, C are the square matrices, q and v are the displacement and velocity vectors,
respectively, F is the vector of external forces and s is the vector of nodal potential forces,
computed at the end of the previous time step.
Table 6. The algorithm of the space-time element method (velocity formulation).
SForm the matrices K and M

1
SRight-hand-side vector:
F = fn+1 - [K(1 )h - M] vn
2
h
2
[K h + 1 M] vn+1 = F
SSolve the equation:
h
2
SDisplacements:
xn+1=xn+h[vn+(1-)vn+1]
SNodal forces:
sn+1=K xn+1

The numerical dissipation is performed by modifying the formula for displacements

xn+1=xn+h[(1-)vn+vn+1], =1/(1+) .
The important advantage of this method is that it can be directly employed both to dynamic
and quasistatic analysis. For =1.0 the same procedure can be used even if the mass density is
equal to zero. In such a case the kinematic boundary conditions should force the motion. In
the case of positive mass density the unconditional stability is obtained for 2/2.

M>0

M=0

Fig. 5. Spectral radiuses with non zero inertia and without inertia.
The time integration methods (many of them) can be described in the following form:
qi+1 = A qi qi-1
The comparison given in Tab. 7 presents the group of methods, that can be considered as a
particular case of the space-time finite element formulation.
Table 7. Operators for time integration methods.
method
operator A
central difference method
2-
Newmark =0, =1/2
space-time elem. =0
Newmark =1/6, =1/2
4(3-)/(6+)
trapezoidal rule
2(4-)/(4+)
Crank-Nicolson
Newmark =1/4, =1/2
space-time elem. =0.707
space-time elem. =1
4/(2+)

8. Information flow.
The information flow between nodes is important for every wave propagation problem.
Especially strongly non-linear problems are sensitive. We can say that all the discrete
methods exhibit parabolic-type propagation of disturbances. Implicit methods give infinite
speed of the information flow while in explicit methods the speed is limited to the diagonal of
the mesh. Practically in both types the wave propagation exceeds the physical wave speed.

The wave front in the same time is not sufficiently sharp. In certain problems it can be
essential. Fig. 6 shows the flow of information between joints in one time layer and between
successive time layers. The arrows show how the external impulses flow from joint to joint
and how they perturb the mesh.

explicit methods

space-time simplex elements

implicit methods

Trujillo method

Fig. 6. The information flow in selected time integration schemes.

9. Inertia matrix.
Dynamic response strongly depends on the form of inertia matrix. Several forms of inertia
matrices are described in the literature. The diagonal mass matrix is the simplest one. It is
efficient in numerical calculations carried on by central difference method or semi-implicit
methods. The consistent matrix, derived directly from the shape functions results in more
accurate simulation of wave propagation ore vibrations with both the transverse and rotatory
degrees of freedom. Another way [10] of the lumping scheme also does not take into account
complete rotational degrees of freedom.
We can notice that for most of purposes all the methods give sufficient accuracy.
However, the best approach to the theoretical line is obtained with the consistent mass matrix.
Park-Housner and Trujillo methods have the feature observed for lumped mass matrix (since
they use such a matrix). The split of stiffness matrix in the case of semi-implicit methods does
not worsen results. However, the Trujillo method exhibits better quality.
We should emphasize that even if the artificial damping is applied, the spurious
vibrations locally dominate. The decay of oscillations, when consistent and lumped mass
matrices are applied, are similar. However, in the first case the divergence is according to
time, in the second case it is opposed.

Trujillo

STEM, constitutive masses

Park-Housner

STEM, diagonal masses

Fig. 7. Vibrating bar solved by different methods (no damping).

consistent masses

lumped masses

Fig. 8. Artificial damping effect with constitutive and lumped mass matrices.

11. Conclusions.
The comparison of time integration method given in this paper exhibits non classical methods,
elaborated and described in the literature. Although the question of the efficient numerical
tool was discussed in many papers (for example [11-15]), supplementary tests proved
essential advantages of the methods presented in the present paper:
S
semi-implicit methods are efficient and give sufficiently good results both in dynamic and
wave analysis,
S
the Bossak and Hilber-Hughes-Taylor methods are the alternative to the Newmark method,
S
the space-time element method enables both dynamic and quasi-static analysis.

References
1. N.M. Newmark. A method of computation for structural dynamics. Proc. A.S.C.E., 8:6794, 1959.
2. W.L. Wood, M. Bossak and O.C. Zienkiewicz. An alpha modification of Newmarks
method. Int. J. Num. Meth. Eng., 15:1562-1566, 1981.
3. H.M. Hilber, T.J.R. Hughes and R.L. Taylor. Improved numerical dissipation for time
integration algorithms in structural dynamics. Earthquake Eng. And Struc. Dyn., 5:283292, 1977.
4. K. Park and J.M. Housner. Semi-implicit transient analysis procedures for structural
dynamics analysis. Int. J. Num. Meth. Eng., 18:609-622, 1982.
5. D.M. Truhillo. An unconditionally stable explicit algorithm for structural dynamics. Int. J.
Num. Meth. Eng., 11:1579-1592, 1977.
6. =.F]NRZVNL2QWKHDSSOLFDWLRQRIQRQ-rectangular space-time elements. Mech. Teoret.
i Stosow., 21(4):531-542, 1883.
7. C. Bajer. Trangular and tetrahedral space-time finite elements in vibration analysis. Int. J.
Numer. Meth. Engng., 23:2031-2048, 1986.
8. C. Bajer. Space-time finite element formulation for the dynamical evolutionary process.
Appl. Math. and Comp. Sci., 3(2)251-268, 1993.
9. C. Bajer and C. Bohatier. The soft way method and the velocy formulation. Comput. and
Struct., 55(6):1015-1025, 1995.
10. E. Hinton, T. Rock and O.C. Zienkiewicz. A note on mass lumping and related processes
in the finite element method. Earthquake Engng. Struct. Dyn., 4(3):245-249, 1976.
11. R. Mullen and T. Belytschko. An analysis of an unconditionally stable explicit method.
Comp. and Struct., 16:691-696, 1983.
12. T.J.R. Hughes and T. Belytschko. A prcis of developments in computational methods
for transient analysis. J. Appl. Mech., 50:1033-1041 (1983).
13. J. Cushman. Difference schemes or element schemes? Int. J. Num. Meth Engng., 14:16431651, 1979.
14. R. Mullen. An analysis of an unconditionally stable explicit method. Comp. and Struct.,
16(6):691-696, 1983.
15. K.C. Park. Practical aspects of numerical time integration. Comp. and Struct., 7:343-353,
1977.

Supported by KBN: 8 T12C 046 21