Sei sulla pagina 1di 16

Computational Mechanics (1992) 9, 359-374

Computational
Mechanics
9 Springer-Verlag 1992

Mixed finite element formulations in the time domain


for solution of dynamic problems
D. Aharoni and P. Bar-Yoseph
Computational Mechanics Laboratory, Faculty of Mechanical Engineering, Technion--Israel Institute of Technology, Haifa
32000, Israel

Abstract. This paper presents a two-field mixed finite element formulation in the time domain. Several strategies are presented
by which the variational formulations are produced. One strategy is based on Hamilton's law, and is suitable for problems
where the mechanical energy of the system is given.
Another strategy uses the Galerkin method, which is suitable when the equations of motion are given. This strategy can
either be based on a continuous or discontinuous approximation of the physical variables in time. All formulations were
analyzed, and their characteristics in terms of truncation error (phase and amplitude) and stability are presented. This analysis
was verified by numerical results. Comparison of many formulations leads to the conclusion that discontinuous mixed finite
element method in the time domain is superior to all others in terms of stability and accuracy.

1 Introduction

The numerical solution of dynamic problems traditionally involves two stages: Discretizing the
continuum into finite elements, and solving the generated ordinary differential equations by classical
time integration schemes (e.g. Runge-Kutta, Newmark, finite difference, etc.). An alternative
strategy is the space-time finite element method first proposed by Fried (1969), Oden (1969), Argyris
and Scharpf (1969). This method has many potential advantages: It offers a unified solution strategy,
which can improve computational efficiency and simplify the application to cases where the energy,
rather than the equations of motion are known. It also makes the task of adaptive mesh refinement
relatively simple (e.g. Johnson 1987; Bajer 1989). Moreover, high numerical efficiency can be
achieved by using spectral elements in time, i.e. elements with high order polynomials. An example
of the comparison of finite element method (FEM) in time with other methods can be found in
Peters and Izadpanah (1988). An efficient sapce-time spectral finite element method was recently
developed by Zrahia and Bar-Yoseph (1991). Another implication of the spectral method is that
higher order accurate algorithms can be developed by choosing higher-order elements. This makes
it unnecessary to choose different schemes for different error levels, thus enabling a versatile
program. All of these advantages lead to the solution of numerous problems in mechanics by using
time finite element method.
The variational formulation is the basis of the development of FEM. It can be divided into
three different types of formulations:
(1) Extremum variational principles.
(2) Physically based on-extremum, variational formulations.
(3) Weighted residuals, such as the Galerkin method with continuous and discontinuous approximations in time.
An example of(l) is the Gurtin principle (1964). Some FEM work was done using this principle
of which recent example is Jing (1991). The main drawback of this principle is that it is not
applicable to nonlinear problems.
As an example of (2), Hamilton's principle and law have been thoroughly discussed in the
literature, beginning with the remarks of Leitmann (1963) and Smith and Smith (1974) on the
extremum of Hamilton's principle. Oden (1969) and Fried (1969) also based their developments on

360

Computational Mechanics 9 (1992)


q

k
m

Fig. 1. Mass-spring harmonic system

the time FEM of Hamilton's principle. Bailey (1975, 1976, i980) offered the use of the more general
Hamilton's law of varying action. Simkins (1978, 1981) used Bailey's procedure with Hermite cubic
polynomials as base functions, that require continuity of the variables and their derivatives. This
requirement makes it unsuitable for problems such as shock waves. Hermite base functions were
also used by Sorek and Blech (1982), and by Baruch and Rift (1982, 1984).
Smith and Smith (1977), and Smith (1977, 1979, 1981, 1984) questioned all the above interpretations of Hamilton's principle. They proposed a weighted residuals Galerkin method, which
they showed to be numerically equivalent to Hamilton's law, without the semantics associated
with the word "variational". This leads to formulation type (3), in which the differential equations
are applied directly, rather than via a variational principle. Borri et al. (1985a, b) presented
Hamilton's weak principle, where the trailing terms in Hamilton's law were expressed as a function
of the known and unknown momenta Po and Ps, rather than as a function of -~?L
- , as was done

0O

before. This eliminates the need to use Hermite polynomials of C 1 continuity, while in the
functional the variables need C o continuity only, thus enabling the use of linear base functions.
Borri also stated that ~q are actually weighting functions, leading to the equivalent Galerkin
method. Borri's formulation is analogous to the hybrid FEM, where one variable is defined in the
domain, and the other on the boundaries only. A similar, more general bilinear formulation was
presented by Peters and Izadpanah (1988).
Mello et al. (1990) and Borri et al. (1991a, b), presented a general three-field approach. Primal
and mixed formulations (similar to the one developed independently in this paper) are developed
from the general formulation. These are discussed in the following sections.
Another approach assumes the unknown fields to be discontinuous in time. Lesaint and
Raviart (1974) were the first to present this approach for scalar linear hyperbolic equations, and
showed the equivalence to the Runge-Kutta method. Bar-Yoseph (1989) extended this idea
through flux splitting with an alternating sweep in the forward and backward space directions,
for multi-dimensional nonlinear and quasi-linear hyperbolic equations with shock fronts. The
drawback of this scheme is that it is computationally inefficient, and its generalization seems
possible only in special cases. Further development was made by Bar-Yoseph and Elata (1990),
that gave an answer to the efficiency problem by moving the nodes to Gauss points, thus cutting
down the number of operations needed. They were also able to reconstruct the exact solution of
some problems by using titled elements.
The aim of this paper is to present simple mixed FEM formulations in time of types (2) and
(3) for solving dynamic problems. All formulations are analyzed for their error (phase and
amplitude), convergence rate and stability.
Since any n degrees of freedom coupled system can be decomposed into n uncoupled scalar
equations, it can be established that the entire coupled system reduces to consideration of the
individual modal equation. Thus the analysis was done on the scalar single degree of freedom
harmonic mass-spring oscillator with no friction or viscosity (Fig. 1).
2 Mixed variational formulation based on Hamilton's law

A mixed formulation allows independent fields in the domain for the generalized displacement q,
and momentum p. Two strategies for developing the mixed functional are used:

D. Aharoni and P. Bar-Yoseph: Mixed finite element formulations in the time domain

361

(1) A Lagrange multiplier is utilized to enforce the compatibility relations between p and 0 in a
similar way to that of Washizu (1980).
(2) Using Hamilton's canonical integral. This strategy was also developed independently by Borri
et al. (1991a, b).

2.1 Strategy (1)


It can be shown that every Lagrange multiplier 2, represents the generalized momentum p.
Therefore the mixed functional is
tf

I = ~ [ L + p ' ( q - v)]dt

(2.1)

to

where v is the generalized velocity, t o is the initial time, t: is the final time and L is the Lagrangian
and is defined by
L(q, q, t):= T(q, q, t) = V(q, t)

(2.2)

where T is the kinetic energy and V is the potential energy. The definition of the generalized
momentum is given by
0T

9L
= --.

P:= 90

(2.3)

90

For the mass-spring harmonic oscillator model example, the mixed functional is given by
ty

i=f[ 89

1 2 +P'(gl- v)]dt
gkq

(2.4)

to

and the generalized momentum is


OL
p

- -

(2.5)

my.

g0
Substituting Eq. (2.5) into Eq. (2.4) gives

I= l P'O-~mP to

kq2 dt

(2.6)

which is the mixed functional for this example.

2.2 Strategy (2)


This strategy uses the definition of the Lagrangian after employing Legendre's transformation
to include the Hamiltonian H.
L(q, q, t):= p ' q -- H(p, q, t).

(2.7)

The mixed functional through this definition of the Lagrangian is


tf

tf

to

to

I= S Ldt= ~ [ p ' q - U ] d t

(2.8)

and is called Hamilton's canonical integral.


For the mass-spring model example the functional is

I=![p'(l-(~mpZ+~kq2)]dt.
The functionals in Eq. (2.9) and Eq. (2.6) are equivalent.

(2.9)

362

Computational Mechanics 9 (1992)

It should be noted that the Hamiltonian is the total mechanical energy only when it is not
directly dependent on time.

2.3 The variational formulation


Obtaining the mixed functionals is the first step towards the numerical solution. The second
step is to take the variation of these functionals and substitute the result in Hamilton's law
(e.g. Baruch and Rift 1982), which is
6I -- 0-~T-6q t: = O.
(2.10)
v0
Ito
Replacing the boundary trailing term with the definition of the generalized momentum leads to
6I - p . 8qltts -- 0.

(2.11)

Substituting Eq. (2.9) or Eq. (2.6) into Eq. (2.11) yields the following variational formulation

t:~ O -

P .6p+p.5(l-kq.Sq dt=p.6q

to

(2.12)

I tO

Integrating the term q.@ by parts leads to the following alternative formulation

"[

I _ q . 6 p _ l p . 6 p +p'6(I- kq.6q dt=p.c~qlt/o-q.fpl;~o.

to

(2.13)

Other formulations may be produced in the same way. For example, another integration by
parts yields

I 4"' P-l-P'6P-P6q-kq' q

to

dt=0.

(2.14)

However, the two formulations in Eqs. (2.12) and (2.13) are sufficient.
In the formulation of Eq. (2.12) the variables p and q appear as independent fields, p and 6p
have no continuity requirements at the time boundaries (C-t continuity), while q and 5q have
to be continuous at the boundary (C o continuity). The formulation in Eq. (2.13) is the same as in
Borri et al. (1991a, b). The variables p and q are also independent fields, but they have no
continuity requirements (C- ~ continuity), while the test functions 6p and 6q have to be continuous
at the boundary, and differentiable in the domain (C o continuity).
In the next section the discontinuous approach is presented, where the continuity requirement
on the test functions is eliminated, and the boundary trailing terms are shown to have the property
of satisfying the initial conditions on average.

3 Mixed formulation based on the Galerkin method

3.1 Continuous Galerkin approach


The Galerkin method is an example of the more general weighted residuals method. This method
is based on weighting all of the motion equations and other conditions by the same base functions
used to approximate the variables. For the dynamical case, the Euler-Lagrange equations of
motion are (Goldstein 1950)
d-~

-~q=

where Q is the generalized force.

(3.1)

D. Aharoni and P. Bar-Yoseph: Mixed finite element formulations in the time domain

363

The compatibility equation gives the relation between the variables


0L
-- =p.

(3.2)

04

Substituting Eq. (3.2) in Eq. (3.1) yields two-field (p, q) equations of motion
OL
--=POq

Q.

(3.3)

A different way of representing the differential equations is through Legendre's transformation


to give Hamilton's canonical equations (Lanczos 1991)
~H
~3q + Q'

/~-

~H
0 - c3p"

(3.4)

In both cases the initial conditions are


t = to:

q=

q(to), v = p ( t o ) .

(3.5)

Weighting the canonical equations using the Galerkin method yields the following weak formulation

!E(gTl~p)'6P+([~+OoHq - Q ) ' 6 q ] d t = O

(3.6)

where the initial conditions are directly imposed.


It should be noted that weighting the equations in the Euler-Lagrange form (e.g. Eq. (3.3)),
rather than the canonical form, yields similar results.
For the mass-spring model example, the canonical equations are

= - kq,

1
m

0 = --P.

(3.7, 3.8)

After weighting these equations, the following weak formulation is obtained

O--

P "6p+(p+kq).6q

to

dt=0.

(3.9)

Equation (3.9) is equivalent to Eq. (2.14) (except for the sign of the test function 6q which will not
change the solution). Integration by parts yields Eq. (2.12) and Eq. (2.13), as pointed out by Smith
(1984).
The choice of approach depends on the representation of the problem. If it is represented
through a set of differential equations then it is simpler to use the Galerkin method. On the other
hand, if the problem is represented by an integral form of mechanical energies, it is simpler to use
the approach based on Hamilton's law. (Landau and Lifshitz 1976)

3.2 Discontinuous Galerkin approach


In this formulation the canonical equations are weighted by the Galerkin method, and the
continuity of the variables is weakly enforced by a jump operator, thus satisfying the initial
conditions in an average sense (see Fig. 3.1).
The general variational formulation is
to

O-

-6p + [~ + - - -

0q

Q "6q dt + [Po -p(to)]'6q(t0) + [q0 - q(to)]'6p(to) = 0

(3.10)

where q(to) , p(to) are the initial conditions from the previous time step, and Po, qo are the initial

364

Computational Mechanics 9 (1992)


qf Pf

At

qo Po
q(t0) P(to)
Fig. 3.1. Discontinuous element

variables of the new time step. As a result of the discontinuous scheme, qo and q(to) as well as Po
and p(to), are not necessarily equal.
This formulation is a direct extension of Lesaint and Raviart (1974) in which the second-order
equation is transformed into two first-order equations. For the mass-spring example, the formulation is represented by

0-

P "6p + ([J + kq)'6q dt + [Po -p(t0)]'6q(to) + [qo - q(to)]'6p(t0) = 0.

(3.11)

to

Here p, q are independent fields, thus p, q, 6p, 6q have no continuity requirements at the time
boundaries (C- 1 continuity). It should be noted that this formulation is less demanding than the
formulation in Eq. (2.13) where the test functions are required to be continuous on the boundary
(C o continuity). This advantage is especially significant when space dimension is introduced. It
eliminates the need to use transition elements between meshes when adaptation is used. Moreover,
it is possible to improve efficiency by moving the nodal points to Gauss points (e.g. Bar-Yoseph
and Elata 1990).
Integration by parts as demonstrated in Appendix A shows the numerical equivalence of the
discontinuous formulation (Eq. (3.11)) and the continuous one (Eq. (2.13)), assuming 6p, 6q are
continuous and nodal points are on the boundaries.
4 The finite element formulation

In the numerical formulation of the problem we assume that q} and ~ are base functions which
are used to interpolate variables q and p, respectively. As stated, the approximation is independent.
Thus, for example, the degrees of the polynomials do not have to be equal. Discussion on this
subject is given in Sect. 6. The test functions 6p, 6q are also approximated by the same base
functions according to the Bubnov-Galerkin method
q,~t~Tu, O ~ t ~ T u , 6q~@T, 6 i 7 ~ T, p,,~lJTv, ] ~ V T v ,
6 p ~ V T, ( ~ i l ~v/T
(4.1)
where
uT={qt,q2,'",qN},

vT={p*,P2,''',PM}"

Here, p~, q~ are the approximated values of the variables at the nodal points, and N, M are the
number of nodal points in the element depending on the degree of the polynomial of the base
function. Substituting these approximations (4.1) in the variational formulations yields an algebraic
system of equations. For the formulation in Eq. (2.12) the system of equations is

I!
--

~Td t

1 tI ~ ~ r d t
m to

tf

tf

to

to

~ ~@Tdt

~ t~Tdt

I:t
=

(4.2)

w h e r e f T = {0,0,...,0) a n d f T= --p(to),O,...,O,p(tf)}.
The dimension of vector fx(M) is equal to the dimension of vector r, and the dimension of
vectorf2(N) is equal to the dimension of vector u. The contribution of the nodals in the domain
is zero, because this simple case has no internal or external forces.

D. Aharoni and P. Bar-Yoseph: Mixed finite element formulations in the time domain

365

For the formulation in Eq. (2.13) the following system of equations are obtained

~Tdt

f'f

-- S ~ T d t
W/ to

to

tf

tf

{;t={;:t

j" ~ X d t

- - k S (l[~]~Tdt
to

,43,

tO

where the right hand vector has both boundary trailing terms
fiT= {-q(to),0,...,0, q(t:)}

and fT=2 {--P(t0),0,"',0,p(tz)}"

For the discontinuous approach a different system of equations is obtained in the general case.
However, for the approximation used (e.g. Sect. 6) results were the same as those in Eq. (4.3)

,o

m to

tf

tf

k~Tdt

~ q}~Tdt

fl

(4.4)

f2

where the right hand vector is


f v = { _ [ - q 0 _ q ( t o ) ] , 0 ,...,0,0},

fT=2 { - - [ P 0 - - p ( t 0 ) ] ' 0 ' ' ' ' ' 0 ' 0 ) "

In order to demonstrate a solution procedure we substitute a linear Lagrangian base function for
both variables p and q in formulation (2.13) to obtain the algebraic system of equations (given
generally in Eq. (4.3))
1
1
At
At - q(to)
Uol
2
2
3m 6m
1
2

1
2

kAt
3

kAt
6

At
6m

At
3m

U:l

1
2

1
2

Vol

Uf

(4.5)

-p(to)

kAt
kAt
1
1
V:l
Vf
6
3
2
2
We now have several options for imposing the initial conditions. One option is to impose them
on the average, i.e. the initial conditions from the last time step p(to) and q(to) are left at the right
hand vector, then the unknown generalized momentum and displacement v: and u: are moved to
the coefficient matrix. The system of equations for this option is
m

At

At

3m

6m

At
6m

At
3m

1
2

1
2

1
2

--

1
2

----

kAt
3

kAt
6

kAt
6

kAt
3

1
2

1
2

Uo

-- q(to)

Uf

0
(4.6)

Uo

-p(to)

vf

Another option is to insert the initial conditions directly, as done with boundary conditions,

366

Computational Mechanics 9 (1992)

thus enforcing u o to be exactly equal to q(to) and v0 to be equal to p(to). After moving vy and us
to the coefficient matrix, the following system of equations is obtained
m

At

At

6m

3m

kAt
6

kAt
3

1
2

0
1
2

uo

uf

I q(to)

I~

.= )
I

Vo

I p(to)

Vf

(4.7)

In both cases different schemes were generated. A comparison and analysis of many other options
are given in Sect. 6.

5 Stability and error analysis


Since any n degrees of freedom coupled system can be decomposed into n uncoupled scalar
equations, it can be established that the entire coupled system reduces to considerations of the
individual model equation (e.g. Hughes 1987). Thus the analysis was done on the scalar single
degree of freedom harmonic oscillator equation with no damping or friction
+ co2X = 0

(5.1)

where co2 =--k and co is the natural frequency of the system, and X is the magnitude of the
m
displacement.
The analytical solution of Eq. (5.1) is

X(t) = Clei~'t + C2 e-i~

(5.2)

where C1 and C2 are constants that depend on the initial conditions.


Transforming each of the numerical formulations to the discrete form
x,+ 1 = Ax"

(5.3)

where x" and x "+1 are the known and unknown vectors, respectively, and A is the numerical
amplification matrix.
The solution to the discrete form of Eq. (5.3) is
X = O~2~ + 0221

(5.4)

where D 1 and D2 are constants, X is the discrete displacement and 21,22, are the complex
eigenvalues of the amplification matrix A, and are represented as
21 = e "la',

22 = e "2a'

(5.5)

where/~1 and ]-/2 are conjugate complex numbers, i.e.,

#1 = a + ib,

t~2 = a - lb.

(5.6)

Substituting Eq. (5.5) into Eq. (5.4) yields another form of the discrete solution
X = D 1 eanAt eibnat q- D 2 eanAte- ibnAt.
By comparing this solution to the analytical one, the physical meaning of a and b is identified:
a is proportional to the numerical damping, and b to the numerical frequency.

D. A h a r o n i a n d P. Bar-Yoseph: M i x e d finite element f o r m u l a t i o n s in the time d o m a i n

367

When solving dynamical problems it is common procedure to separate the error in numerical
schemes, into dissipation, namely, amplitude error, and dispersion, namely, phase error.
The amplitude error ex is found from the modulus of the eigenvalues of the amplification matrix
A, as it is proportional to the damping by (Lapidus and Pinder 1982)
I)~i[ = ]Di e""ate -+ibnAt[ = IDi ea"a'(cos bnAt +_ i sin bnAt)[

D i e a"at.

(5.7)

Giving the amplitude error definition


e~:= 1 -I;~[.

(5.8)

It should be noted that numerical dissipation is sometimes an advantage, especially when solving
shock wave problems, where it is needed to filter the high frequency noise made by dispersion.
Dispersion in waves is generated when the phase velocity is a function of the frequency. The
physical significance of dispersion in one dimension (time), is the difference between the analytical
natural frequency and the numerical frequency multiplied by the time step At.
The phase in the analytical expression is mAt, and in the numerical expression it is given by
a r c t a n ( I m ( 2 ) ~ = a r c t a n ( s i n ( b A t ) ) = bAt
\ Re (2) ]
\ cos (bAt)

(5.9)

where Im0~) represents the imaginary part of the eigenvalue, and Re(2) the real part. The phase
error % is then defined by
% : = c o a t - arctan (Im(2) ~.
\Re(2)/

(5.10)

In order to verify the results of the phase error, we can use the amplitude modulation (AM) caused
when one wave is subtracted from another one of similar frequency. This is the case when the error
expression is defined as the difference between the analytical and numerical displacement. For
simplicity the case without dissipation is chosen, causing the error to be affected only by the phase
error.
E R R O R = a cos(cot - (p) - a cos(bt - q)) = - 2a sin(coast - ~0)sin(iAcot)

(5.11)

where a and (p are the amplitude and phase respectively, and


b+co

coav -

Aco = co -- b.

(5.12)

A typical behavior of the error as a function of time is given in Fig. 5.1.


Measuring the wavelength of the envelope of the error plotted as a function of time, gives
Aco, which is the phase error, divided by At. An example for the reduced integration case (details
given in Sect. 6) with a linear approximation is given in Fig. 5.2. Half the wavelength was measured
from the graph to give 0.5z ~ 7500s for the case where At=0.1 and co= 1. From the equality

0.75
0.5
0.25

P~

0
-0.25
-0.5
-0.75
-1

500

1000 1~00 sooo 2500


t (s)

000 3500 4000

Fig. 5.1. A m p l i t u d e m o d u l a t i o n of the error as a function of time

368

Computational Mechanics9 (1992)


2

1.5
1
0.5
0

-0.5
-1
-1.5
-2
2500

5000

7500

1.000e+04

t (s)

sin

~-t

= sin

Fig. 5.2. Error envelopeas a function of time for reduced integration

Acot ,Am is extracted to give Aco = 8.32-10 . 4 making the phase error to be

e,o = AcoAt = 8.32.10-5. This result is identical to the result from Eq. (5.10) which is for this case
1
e~ ~ ~(coAt) 3 = 8.32"10- 5.
Another way of finding the truncation error is by subtracting the analytical amplification
matrix from the numerical one. The subtraction follows expansion of every c o m p o n e n t of the
matrices into a Taylor series. The local truncation error is the largest c o m p o n e n t in the resultant
matrix. Because of error accumulation in every step, the global truncation error is an order of
magnitude greater than the local one. For the mass-spring model the analytical amplification
matrix is

icos ., 1
me) sin(c0At)

(5.13)

cos(coAt) _J

The truncation error for this example is found from the resultant matrix
AA = A, - A.

(5.14)

To verify the results of the global truncation error, found by the above procedure, a posteriori
convergence rate in [I'll 1 n o r m is evaluated, where the error is defined by
blU - u [ I , : =

T/At

n=l

~ abs(U-u)~At

a b s ( U ~ - u ")

(5.15)

where U and u are the exact and approximate solutions respectively at discrete nodal points, At
is the time step and T is the end of the time interval taken. As shown in Sect. 6 results from the
a posteriori and a priori evaluations are identical.
Convergence of a numerical formulation requires consistency and stability (Lax equivalence
theorem). Consistency can be determined from the truncation error. Stability is determined by the
eigenvalue expression of the numerical amplification matrix. The stability condition is that the
eigenvalue modulus is less than or equal to unity, i.e.
[).il < 1

i= 1,2,...,N

(5.16)

where N is the dimension of the amplification matrix.


An example of the stability and error analysis procedure is given for the formulation in Eq.
(2.13). Results for other formulations are stated in Sect. 6.

D. Aharoni and P. Bar-Yoseph: Mixed finite element formulations in the time domain

369

Convergence rate
-2
-3

M-5
-6
-7
-I

-1'.2

-1'.4

-1.6

-1.8

-2

Fig. 5.3. Convergence rate

Log~t

The amplification matrix for this formulation is taken from Eq. (4.3)
1
--

(_

36(coAt))

((oAt) 4 d- 4(o~At) 2 + 36

mco(2(coAt) 3 - 36(coAt))

(coAt)4 + 4(~At) 2 + 36
where co =

2(o~At)3 +

mo)

(coAt)4 + 4(teAt) 2 + 36

Uf} =

v:

14(coAt) 2 + 36

(5.17)

14(e)At) 2 + 36

(coAt)4 + 4(coAt) 2 + 36

is the natural frequency.

With the aid of a symbolic program, the two conjugate complex eigenvalues were found
21,2 = (-T-) (2(oAt)3 - 36(coAt))i(_+) 14(coAt)2( -T-)36
(coAt)4 + 4(oAt) 2 + 36

(5.18)

Table 6.1. Comparison of a number of formulations with linear base functions for p and q. p:= coat
Formulation

Stability
condition

ez

Mixed, type (2.13)


I.C. for q andp on the average

none

7~p

Mixed, type (2.12) and (2.13)


I.C. for p on the average
and q in line Uo

1 2

<x/i5

Local
truncation
5

2~0 p
17 a

Convergence
ratea

72P

1 2

U p

none

Mixed, type (2.12) and (2.13)


I.C. forp in line v:, q in line Uo

p<6

~5p 3

1 2
~p

Mixed with reduced integration


I.C. forp in line v0, q in line uo

none

1
12--Pa

1 a
i2 p

Borri's hybrid formulation

p<,~

~p~

Mixed, type (2.12) and (2.13)


I.C. for p in line Vo, q in line uo

~p

a Rounded off (with 1~ accuracy) a posteriori numerical evaluation for both variables p and q

370

Computational Mechanics 9 (1992)

In order to compute the amplitude error and stability, the modulus is needed
4((coAt)

13(coAt)4 q- 72(coAt)2 + 324)

~o-At-)2~
~Xt~ + 3-@

]21 = k /

(5.19)

It can be seen that the modulus expression is less than unity for every coat i.e.
121 < 1

(5.20)

namely, this formulation is unconditionally stable.


The amplitude error ex is evaluated from the same expression by a Taylor series expansion
ez = 1 -121

(5.21)

~ (coAt)4

72
The phase error e~ is found from Eq. (5.10), also by a Taylor series expansion
9 / 2 ( c o A t ) 3 - 36(coAt)'~ (coAt)5
% = coAt - arctant . . . . . .
9~
\ - 14(coAt)2 + 36 J
270

(5.22)

The truncation error is found from the matrix AA, which is the difference between the analytic
and numerical amplification matrices, after expanding every component to a Taylor series. The
following matrix was obtained for this example
(coAt)
AA~

72
19(coAt)5

19(coAt)5]

S0

(5.33)

(coAt)~/ "

L 10~

72

_1

The dominant component gives the local truncation eror, which for this case is O((coAt)4), making
the global truncation error to be O((coAt)3).
The a posteriori convergence rate is calculated through the norm in Eq. (5.15). The result for
this example compared with Borri's hybrid formulation, is shown in Fig. 5.3. As mentioned above

Table 2. Comparison of a number of formulations with parabolic base functions for p and q
Formulation

Stability
condition

Mixed, type (2.13)


I.C. for q a n d p on the average

none

Mixed, type (2.12) and (2.13)


I.C. forp in line Vo and q in line Uo

none

Mixed, type (2.12)


I.C. for p in lines vs and ui,
for q in lines v0 and u o

e~,

Local
truncation

Convergence
rate"

1 6
72~P

1
7
42~P

1 6
7~6P

1 3

10oP

60P

p < x/~

~p3

Mixed, type (2.12) and (2.13)


I.C. forp in line v~, q in line u o

p < x/41 - 1
2

2S1
uu p3

Mixed with reduced integration


I.C. forp in line v0, q in line u o

none

720
pl s

1
~/zuP5

Borri's hybrid formulation

p < x/~

1 5
14~ p

1
7 ~ p5

66P
1

~p

p3

" Rounded off (with 1~ accuracy) a posteriori numerical evaluation for both variables p and q

2
4

D. Aharoni and P. Bar-Yoseph: Mixed finite element formulations in the time domain

371

there is a correlation between the a posteriori evaluation and the computed global truncation
error. It should also be noted that as a result of the coupling between the variables p and q, both
have the same convergence rate.
The same analysis was used for other possible formulations, and the results are summarized
in Tables 6.1 and 6.2.
6 Numerical results and conclusions
A number of mixed formulations in FEM for the solution of dynamical problems have been
presented. Tables 6.1 and 6.2 summarize results for some of the possible formulations with equal
approximations for q and p. The analysis was carried out with the same procedures as shown in
the example of Sect. 5. All formulations were characterized by the stability condition phase and
amplitude errors, global truncation error and a posteriori convergence rate. It is obvious from
these results that the formulation in Eq. (2.13), where initial conditions for p and q are satisfied in
the average sense, is superior all the others presented: It is unconditionally stable, and has the
highest convergence rate. The error is usually presented by the formula
(6.1)

/[ e II1 "~ Ch=

where C is a constant independent of h, h is the element size and c~is proportional to the degree
of interpolation polynomial. For the case of formulation in Eq. (2.13), the error takes the form
]1e [11 ~ CP 2k+ 1

(6.2)

where k is the polynomial degree of the base function, and p = coat. For the case of Eq. (2.12) it is
of the form
1[e ]]1 ~ CP k

(6.3)

and in the hybrid case or in the reduced integration option it is of the form
IIe [11 ~ Cp 2k.

(6.4)

In the hybrid formulation, p is approximated only on the boundaries. This solves the problem of
imposing initial conditions with linear base functions, but the formulation has many disadvantages
compared with the mixed formulation. Convergence rate is an order of magnitude less then that
of the discontinuous scheme (e.g. Eq. (6.4) as compared with Eq. (6.2)). It has a growing instability
condition, as the degree of the polynomial rises (see also Peters and Izadpanah 1988). The mixed
formulation with reduced integration has similar characteristics to the hybrid formulation except
for the fact that it is unconditionally stable. Reduced integration is obtained by using less Gauss

II

III

IV
Fig. 6.1. Setting of reduced integration matrix

Computational Mechanics 9 (1992)

372
Table 6.3. Comparison of different combinations for base functions ofp and q
Global
truncation
error

Minimum
for
approximation

Maximum
for
approximation

parabolic

O(pz)

parabolic

linear

O(p2)

linear

parabolic

O(p)

linear

parabolic

O(p z)

parabolic

linear

O(p)

linear-linear
o(p ~)
linear-linear
O(p3)
linear-linear
o(p)
linear-linear
O(p)
linear-linear

parabolic-parabolic
O(p ~)
parabolic-parabolic
O&)
parabolic-parabolic
o&)
parabolic-parabolic
0(02)
parabolic-parabolic

cubic

parabolic O ( p )

parabolic-parabolic cubic-cubic
O(p2)
O(p3)

cubic

parabolic

O(pz)

parabolic-parabolic cubic-cubic
O(p2)
O(p3)

cubic

parabolic

O(p2)

parabolic-parabolic cubic-cubic
O(p2)
O(p3)

Formulation

q~

Mixed, type (2.13)


I.C. for q and p on the average
Mixed, type (2.13)
I.C. for q and p on the average
Mixed, type (2.12)
I.C. for p on the average, q in line Uo
Mixed, type (2.12)
I.C. for p in line vo and q in line Uo
Mixed, type (2.12)
I.C. for p on the average, q in line v~
Mixed, type (2.12)
I.C. for p in line v0 and q in line u0
Mixed, type (2.12)
I,C. for p in line Voand q in line vI
Mixed, type (2.12)
I.C. for p on the average q in line vI

linear

o(p)

O(p~)

points than needed for the exact integration. The motivation was to produce a symmetric scheme
in terms of the weight given to every time level. F o r the linear base functions, integrating with one
Gauss point yields exact integration for quadrants II and IV (see Fig. 6.1) because linear polynomials
are integrated, while in quadrant I and III reduced integration is obtained because the polynomials
are parabolic. Integrating the parabolic base function with two Gauss points yields similar results.
The main disadvantage in reduced integration is that its application to more complicated problems
will not necessarily produce the same effect. It should be noted that the hybrid formulation with
parabolic base functions has exactly the same characteristics as the formulation offered by Baruch
and Rift (1984). They used Hermite base functions which enforce continuity of the velocity. It is
also more difficult to generate higher order C 1 continuity Hermite base functions, making the
hybrid formulation more attractive.
In addition to accuracy and unconditional stability, the mixed formulation offers another
advantage, specially valuable when space-time wave problems are solved, which is small dissipation.
This is needed in order to filter the high frequency noise made by dispersion.
The mixed formulation makes it possible to approximate the variables p and q with different
degrees of polynomials, since they are independent. The expected advantage is the efficiency gained
from getting the same accuracy with a smaller matrix to solve.
Examination of many different combinations for the degree of polynomials o f p and q unfortunately give for most cases a convergence rate lower or equal to the minimum expected (Table 6.3).
The expected Min. and Max. are obtained from the degree of the polynomials of the base function,
based on the results for the equal degree of polynomials. It should be noted that no lamitations
were found on the combination of degree of polynomials, in constrast to mixed formulation in
elastic problems (e.g. Pian et al. 1983).
In conclusion, it is obvious from the results that the mixed formulation with initial conditions
taken on the average (discontinuous), with equal base functions for the approximation of the
variables p and q, is superior to all others presented. This scheme has been extended to space-time
wave equations. Results will be published in the future.

D. Aharoni and P. Bar-Yoseph: Mixed finite element formulations in the time domain

373

Acknowledgements
The research was sponsored by the Technion V.P,R. fund, Bernstein Fund for promotion of research. The generous help of
the Gutwirth Fellow is gratefully acknowledged.

Appendix A
Integration by part of Eq. (3.11) yields
t/

to

Ddt + [qo - q(to)] "3p(to) + q'3Plt/o + [Po -p(to)] "3q(to) + P't~qlttfo = 0

(A.1)

where
D:=i-q'6[j-lP'c~P-P'C~(l+kq'~q;m

(A.2)

which leads to
t:

~ D dt + qz'3p(t:) - q(to)'6p(to) + p:'6q(tz) - p(to)'g~q(to) = O.

(A.3)

to

Equations (A.3) and (2.13) are numerically equivalent.


References
Argyris, J. H.; Scharpf, D. W. (1969): Finite elements in time and space. Nucl. Eng. Des. 10, 456-464
Bajer, C. I. (1989): Adaptive mesh in dynamic problems by the space-time approach. Comput. Struct. 33, 319-325
Bar-Yoseph, P. (1989): Space-time discontinuous finite element approximations for multi-dimensional nonlinear hyperpolic
systems. Comput. Mech. 5, 149-160
Bar-Yoseph, P.; Elata, D. (1990): An efficient L2 Galerkin finite element method for multi-dimensional nonlinear hyperbolic
system. Int. J. Numer. Meth. Eng. 29, 1229-1245
Baruch, M.; Rift, R. (1982): Hamilton's principle, Hamilton's law - 6" correct formulations. AIAA J. 20, 687-692
Baruch, M.; Rift, R. (1984): Time finite element discretization of Hamilton's law of varying action. AIAA J. 22, 1310-1318
Bailey, C. D. (1975): Application of Hamilton's law of varying action. AIAA J. 13, 1154-1157
Bailey, C. D. (1976): Hamilton, Ritz and elastodynamics. J. Appl. Mech. 43, Trans. ASME 98/E, 684-688
Bailey, C. D. (1980): Hamilton's law and the stability of noncons~vative continuous systems. AIAA J. 18, 347-349
Borri, M.; Lanz, M.; Mantegazza, P. (1985a): Comment on time finite element discretization of Hamilton's law of varying
action. AIAA J. 23, 1457-1458
Borri, M.; Ghiringhelli, G. L.; Lanz, M.; Mantegazza, P.; Merlini, T. (1985b): Dynamic response of mechanical systems by a
weak Hamiltonian formulation. Comput. Struct. 20, 495-508
Borri, M.; Mello, F. J.; Atluri, S. N. (1991): Primal and mixed forms of Hamilton's principle for constrained rigid body systems:
numerical studies. Comput. Mech. 7, 205-220
Borri, M.; Mello, F. J.; Atluri, S. N. (1991): Variational approaches for dynamics and time-finite-elements: numerical studies.
(in press)
Fried, I. (1969): Finite element analysis of time-dependent phenomena. AIAA J. 7, 1170-1173
Goldstein, H. (1950): Classical Mechanics. Reading: Addision-Wesley
Gurtin, M. E. (1964): Variational principles for linear initial problems. Quaterly Appl. Math. 22, 252-256
Hughes, T. J. R. (1987): The Finite Element Method. Linear Static and Dynamic Finite Element Analysis. New Jersey: Prentice
Hall, Inc.
Jing, H. S. (1991): Space-time partial hybrid stress element for linear elastodynamic analysis of anisotropic materials. Commun.
Appl. Numer. Methods 7, 39-45
Johnson, C. (1987): Numerical Solutions of the Partial Differential Equations by the Finite Element Method. Cambridge:
Cambridge University Press
Lanczos, C. (1991): The Variational Principles of Mechanics. 4th ed. Toronto: The University of Toronto Press
Landau, L. D.; Lifshitz, E. M. (1976): Mechanics, 3rd ed. Oxford: Pergamon Press
Lesaint, P.; Raviart, P. A. (1974): On a finite element method for solving the neutron transport equation. In: De Boor, C. (ed.).
Mathematical Aspects of Finite Elements in Partial Differential Equation. New York: Academic Press
Lapidus, L.; Pinder, G. F. (1982): Numerical Solution of Partial Differential Equations in Science and Engineering. New York:
John Wiley
Leitmann, G. (1963): "Some remarks on Hamilton's principle". J. Appl. Mech. 2, Trans. ASME, 623-625

374

Computational Mechanics 9(1992)

Mello, F. J.; Borri, M.; Atluri, S. N. (1990): The finite element methods for large rotational dynamics of multiboldy systems.
Comput. Struct. 37, 231-240
Oden, J. T. (1969): A general theory of finite elements II. Applications. Int. J. Numer. Meth. Eng. 1,247-259
Peters, D. A.; Izadpanah, A. P. (1988): hp-Version finite elements for the space-time domain. Comput. Mech. 3, 73-88
Pian, T. H. H.; Chen, D. P.; Kang, D. (1983): A new formulation of hybrid/mixed finite element. Comput. Struct. 16, 81-87
Simkins, T. E. (1978): Unconstrained variational statements for initial and boundary-value problems. AIAA J. 16, 559-563
Simkins, T. E. (1981): Finite elements for initial-value problems in dynamics. AIAA J. 9, 1357-1362
Smith, D. R.; Smith, C. V. Jr. (1974): When is Hamilton's principle an extremum principle? AIAA J. 12, 1573-1576
Smith, C. V. Jr.; Smith, D. R. (1977): Comment on: Application of Hamilton's law of varying action. AIAA J. 15, 284-286
Smith, C. V. Jr. (1977): Discussion on: Hamilton, Ritz and elastodynamics. J. Appl. Mech., Trans. ASME 44/E, 796-797
Smith, C. V. Jr. (1979): Comment on: Unconstrained variational statements for initial and boundary-value problems. AIAA J.
17, 126-127
Smith, C. V. Jr. (1981): Comment on: Hamilton's law and the stability of nonconservative systems. AIAA J. 19, 415 416
Smith, C. V. Jr. (1984): Comment on: Hamilton's principle, Hamilton's l a w - - 6 n correct formulations. AIAA J. 22, 1181-1182
Sorek, S.; Blech, J. J. (1982): Finite-element technique for solving problems formulated by Hamilton's principle. Comput. Struct.
15, 533-541
Washizu, K. (1980): Variational Methods in Elasticity and Plasticity, 3rd ed. Oxford: Pergamon Press
Zrahia, U.; Bar-Yoseph, P. (1991): Space-time spectral element methods for solution of second order hyperbolic equations.
(submitted for publication)
Communicated by S, N. Atluri, October 14, 1991

Potrebbero piacerti anche