You are on page 1of 25

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2001; 50:1931–1955

Conservation properties of a time FE method—part II:


Time-stepping schemes for non-linear elastodynamics

P. Betsch∗; † and P. Steinmann


Chair for Applied Mechanics; Department of Mechanical Engineering; University of Kaiserslautern; Germany

SUMMARY

In the present paper one-step implicit integration algorithms for non-linear elastodynamics are developed. The
discretization process rests on Galerkin methods in space and time. In particular, the continuous Galerkin
method applied to the Hamiltonian formulation of semidiscrete non-linear elastodynamics lies at the heart
of the time-stepping schemes. Algorithmic conservation of energy and angular momentum are shown to
be closely related to quadrature formulas that are required for the calculation of time integrals. We newly
introduce the ‘assumed strain method in time’ which enables the design of energy–momentum conserving
schemes and which can be interpreted as temporal counterpart of the well-established assumed strain method
for nite elements in space. The numerical examples deal with quasi-rigid motion as well as large-strain
motion. Copyright ? 2001 John Wiley & Sons, Ltd.

KEY WORDS: nite element method; integration schemes; energy conservation

1. INTRODUCTION

The objective of this paper is to develop conserving time-stepping schemes for non-linear
elastodynamics. To this end we extend the time nite element approach proposed in Part I [1]
for N -body problems to non-linear elastodynamics.
It is natural to aim at conserving numerical schemes that inherit the conservation properties of
the underlying in nite-dimensional dynamical system. However, customary implicit time-stepping
schemes for elastodynamics such as the Newmark family of algorithms fail to conserve major
invariants of non-linear elastodynamics, namely total energy and total angular momentum, see e.g.
Reference [2]. For example the widely used average acceleration method (or trapezoidal rule) is
known to conserve energy only for the linear theory (see e.g. Reference [3]) thus con ning uncon-
ditional stabiltiy to the linear case. In the non-linear regime the energy method is often employed
for the investigation of the stability properties of time-stepping schemes, see Reference [4] and
references therein. Accordingly, the design of numerical schemes that enable conservation of the
total energy (for autonomous systems) yields automatically unconditional stability. For example

∗ Correspondence to: P. Betsch, Chair for Applied Mechanics, Department of Mechanical Engineering, University of
Kaiserslautern, Erwin-Schrodinger Str., D67663 Kaiserslautern, Germany
† E-mail: pbetsch@rhrk.uni-kl.de

Received 18 October 1999


Copyright ? 2001 John Wiley & Sons, Ltd. Revised 14 March 2000
1932 P. BETSCH AND P. STEINMANN

energy conservation may be enforced by means of a Lagrange multiplier as suggested in Refer-


ence [5], see also Reference [6]. On the other hand, algorithmic energy conservation alone does
not guarantee a satisfactory numerical performance, see Reference [7]. This fact has motivated
the design of the so-called energy–momentum schemes which in addition to energy also conserve
angular momentum, see References [8; 9]. In fact, energy–momentum schemes have been shown
to perform especially well in non-linear structural dynamics, see Reference [10] for an overview
of recent developments.
Another possible approach towards the development of time-stepping schemes for non-linear
elastodynamics is to view the semidiscrete equations resulting from a spatial discretization as a
nite-dimensional Hamiltonian system. It is well known that Hamiltonian systems preserve the
symplectic structure in phase space (see e.g. Reference [11]) and a lot of e ort has gone into
the design of the so-called symplectic integrators, see Reference [12]. However, for non-linear
elastodynamics it has been shown in Reference [8] that energy–momentum conserving meth-
ods have better stability properties than symplectic momentum conserving methods (like e.g. the
mid-point rule).
Yet another approach is the design of energy-decaying algorithms that are sometimes considered
important for practical applications, see e.g. References [13; 14]. In this connection it has been
observed by Armero and Romero [15] that time-stepping schemes that are known to be dissipative
in the linear regime do not necessarily retain this property when applied to non-linear systems.
As outlined above, in the present paper we restrict our attention to the design of mechanical
integrators in the sense of Marsden [16]. A distinguished feature of mechanical integrators is that
they obey fundamental conservation laws of the underlying dynamical system. For non-linear elas-
todynamics we are especially concerned with the algorithmic conservation of angular momentum
and energy.
Most, if not all investigations dealing with conservation properties of time-stepping schemes for
non-linear elastodynamics rely on nite di erence methods. Alternatively, in the present paper we
elaborate on the conservation properties of a time nite element formulation emanating from the
continuous Galerkin (cG) method. The cG method has been introduced by Hulme [17] for the
discretization of systems of non-linear rst-order ordinary di erential equations. Already Hulme
envisaged the possibility of ‘fully discrete’ Galerkin methods resulting from a combination of
nowadays standard nite elements in space with Galerkin methods in time.
Referring to Eriksson et al. [18] there exist essentially two types of Galerkin methods for the
discretization of ordinary di erential equations, namely the aforementioned cG method and the
discontinuous (dG) method which goes back to Lasaint and Raviart [19]. The dG method has
been applied to linear elastodynamics by Hughes and Hulbert [20], see also Reference [21]. As
has been shown in References [22; 23], the dG method can be used for the design of dissipative
time-stepping schemes.
It is interesting to note that time-stepping schemes that have been developed from variational
formulations associated with classical mechanics such as Hamiltons principle or Hamiltons law
(see References [24; 25] for an overview of developments in this direction) often t into the
framework of the cG or dG method.
On the other hand, discrete time-stepping schemes emanating from time nite element formula-
tions can often be related to more customary nite di erence algorithms. For example, Bottasso [26]
has shown that (discontinuous) time nite elements which fall into the framework described by
Borri and Bottasso [27] can be related to certain Runge–Kutta methods. Similarly, the cG method
can be connected with Gauss–Runge–Kutta methods, see Reference [28]. In this connection we

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1933

emphasize that time nite elements relying on the cG method do not t into the framework
provided by Borri and Bottasso [27].
Similar to the N -body problem treated in Part I [1] we show that the conservation properties of
particular time-stepping schemes furnished by the general time nite element formulation (based
on the cG method) are closely connected to quadrature formulas required for the evaluation of
time integrals. In particular, we devise quadrature rules that lead to energy–momentum schemes.
In this connection we introduce the ‘assumed strain method in time’ which can be considered as
temporal counterpart of the well-established assumed strain method for nite elements in space,
see e.g. Reference [29].
This work is organized as follows. In Section 2 we aim at the Hamiltonian formulation of
semidiscrete non-linear elastodynamics resulting from a standard nite element discretization in
space. Section 3 is devoted to the nite element discretization in time based on the continuous
Galerkin method. Section 4 contains a detailed investigation of linear time elements. In Section 5
we present representative numerical examples. Conclusions are drawn in Section 6.

2. HAMILTONIAN FORMULATION OF SEMIDISCRETE ELASTODYNAMICS

The time nite element method developed in Part I [1] relies on the Hamiltonian function of the
nite-dimensional dynamical system under consideration. Therefore, in our treatment of non-linear
elastodynamics, we rst perform a spatial discretization by means of a standard nite element
method. A detailed account of the nite element method for solid mechanics can be found in
References [30; 31]. Our main goal is to cast the semidiscrete equations in Hamiltonian form.
Let B be a regular region in ndim -dimensional Euclidean space (ndim 6 3) occupied by the
reference con guration of a solid continuum body. Furthermore, let I = [0; T ] denote the time
interval of interest. From a kinematic point of view the standard displacement-based nite element
approach employs an approximation of the deformation eld D : B × I → Rndim of the form
nP
node
D(X; t) = NA (X)q A (t) (1)
A=1

Within the material (or Lagrangian) description of motion D(X; t) describes the position of material
point X of body B at time t. Moreover, NA : B → R are the global shape functions associated
with the nodes A = 1; : : : ; nnode and q A : I → Rndim denotes the position vector at time t ∈ I of the
nodal point A. The material velocity of the motion is de ned by v = @D=@t = Ḋ such that the nite
element approximation implies
nP
node
v(X; t) = NA (X) q̇ A (t) (2)
A=1

In general the motion of body B is subject to Dirichlet boundary conditions on @B’ ⊂ @B which
can be easily incorporated into (1), see the above cited references for details. In addition to
that, initial conditions at time t0 , prescribed as initial position D0 : B → Rndim and initial velocity
v0 : B → Rndim , can be taken into account by the nodal quantities q A and q̇ A . The spatial nite
element discretization (1) gives rise to the discrete deformation gradient
nP
node
F = @D=@X = q A ⊗ ∇NA (3)
A=1

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1934 P. BETSCH AND P. STEINMANN

where ∇ denotes the gradient operator with respect to the material co-ordinates. Then the discrete
version of the deformation tensor (or right Cauchy–Green tensor) C = FT F can be written as
nP
node
C= q A · qB ∇NA ⊗ ∇NB (4)
A; B=1

Hyperelastic material is modelled by means of a scalar-valued strain energy density function W(C)
such that the second Piola–Kirchho stress tensor can be calculated via
S = 2DW (C) (5)
R
where DW (C) = @W=@C. Then the strain energy function is given by Vint = B W (C) dV . For
simplicity we assume that the external forces acting on the body can be derived from a potential
function
Z Z
Vext = − %R b · D dV − t · D dA (6)
B @B

where %R : B → R + denotes the reference mass density, b : B × I → Rndim is the applied body force
and t is the prescribed traction boundary condition on @B × I. In view of (1) one obtains
nP
node
Z Z
Vext = − q A · Fext
A A
with Fext = NA %R b dV + NA t dA (7)
A=1 B @B

A ndim
where Fext:I → R are the prescribed external nodal forces. The kinetic energy of the body at
time t is given by
Z
1
T= %R v · v dV (8)
2 B

such that substitution from (2) into (8) leads to


Z
1 nP node
T= MAB q̇ A · q̇B with MAB = %R NA NB dV (9)
2 A; B=1 B

Here MAB are the coecients of the consistent mass matrix M consisting of diagonal sub-matrices
MAB = MAB Indim with A; B = 1; : : : ; nnode . Next, the Lagrangian function L = T − (Vint + Vext ) can be
employed for the calculation of the nodal momenta p A associated with q A according to
nP
node
p A = @L=@ q̇ A = MAB q̇B for A = 1; : : : ; nnode (10)
A=1

Then the Hamiltonian


Pnnode A function follows from the Legendre transformation of L with respect to the
q̇ A ’s as H = A=1 p · q̇ A − L, with the nodal velocities q̇ A being replaced by the nodal momenta
in (10). Accordingly, the Hamiltonian function of semidiscrete non-linear elastodynamics can be
written in the form
1 nP node
H= M −1 p A · pB + Vint + Vext (11)
2 A; B=1 AB
−1
The coecients MAB in (11) are determined by the inverse of the mass matrix M which is
−1 −1
composed of diagonal sub-matrices of the form MAB = MAB Indim with A; B = 1; : : : ; nnode .

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1935

Eventually, Hamilton’s canonical equations of motion can be written in the form


nP
node
−1 B
q̇ A = @H=@p A = MAB p
B=1 for A = 1; : : : ; nnode (12)
ṗ A = −@H=@q A = Fext
A A
− Fint
A A
R
Here, the internal nodal forces Fint follow from Fint = @Vint =@q A = B DW (C) : @C=@q A dV . With
regard to (4), (5) and the symmetry of S, one obtains
nP
node
Z
A
Fint = SAB dV qB with SAB = ∇NA · S ∇NB (13)
B=1 B

Next, we verify that the spatial nite element discretization of non-linear elastodynamics outlined
above inherits fundamental conservation properties of the underlying in nite-dimensional system.

2.1. Conservation of total energy


For the natural dynamical system at hand the Hamiltonian function coincides with the total energy.
Time di erentiation of (11) yields
nP
node node 
nP  A
−1 B
Ḣ = ṗ A · MAB p + A
Fint A
− Fext · q̇ (14)
A; B=1 A

−1
where the symmetry of MAB has been taken into account. With regard to the equations of motion
in (12) one obtains Ḣ = 0 such that the total energy of the semidiscrete dynamical system is
conserved.

2.2. Conservation of total angular momentum


The total angular momentum of the semidiscrete dynamical system about the origin of the inertial
reference frame is given by
Z nP
node
L= %R D × v dV = qA × pA (15)
B A

where (1), (2) and (10) have been employed. Referring to (12)2 , time di erentiation of (15)
leads to
node 
nP  A 
L̇ = q̇ A × p A + q A × Fext A
− Fint (16)
A
Pn
Introduction of the resultant external torque Text = Anode q A × Fext
A
together with (12)1 and (13)
yields
nP
node nP
node
Z
−1 B
L̇ = MAB p × pA + SAB dV q B × q A + Text (17)
A; B=1 A; B=1 B

−1
In view of the symmetry of MAB and SAB , and the skew-symmetry of p B × p A and q B × q A , one
obtains the result
L̇ = Text (18)
such that angular momentum is conserved in the case of vanishing resultant external torque.

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1936 P. BETSCH AND P. STEINMANN

2.3. Conservation of total linear momentum


The total linear momentum of the semidiscrete dynamical system is given by
Z nP
node
I= %R v dV = pA (19)
B A
R Pnnode
where use has been made of (2) and (10) together
Pnnode with the identity B %R NA dV = B=1 MBA which
follows from the completeness condition B=1 NB = 1 for the nodal shape functions. In view of
(12)2 , the rate of change of total linear momentum can be written as
nP
node nP
node  A 
İ = ṗ A = A
Fext − Fint (20)
A=1 A=1

We further obtain the relationship


nP nP
Z n 
node
A
node Pnode
Fint = ∇ NA · S∇NB dV q B = 0 (21)
A=1 B=1 B A=1
Pnnode
where again the property A=1 NA = 1 of the nodal shape functions has been taken intoP account.
nnode A
De ning further the resultant external force of the prescribed external loading as Next = A=1 Fext
one nally obtains
İ = Next (22)
Accordingly, if the resultant external force is zero the total linear momentum of the semidiscrete
dynamical system is conserved.
Our main goal is to perform the remaining time discretization in such a way that the conservation
properties of the semidiscrete system outlined above are inherited. To this end we further develop
the time nite element method outlined in Part I of this work.

3. TIME FINITE ELEMENT DISCRETIZATION

The goal of this section is to perform the time discretization of semidiscrete elastodynamics treated
above by using the general time nite element formulation developed in Part I [1] of this work.
To this end we rst give a brief summary of the advocated time nite element method. For a
compact description it proves convenient to introduce the following vector notation:

q = q1 ; : : : ; qnnode ∈ Rndof
 and z = (q; p) ∈ R2ndof (23)
p = p1 ; : : : ; pnnode ∈ Rndof

Accordingly, the nodal quantities q A ∈ Rndim and p A ∈ Rndim , A = 1; : : : ; nnode , corresponding to the
spatial discretization described in the last section are summarized in the nodal displacement vector
q ∈ Rndof and the vector of nodal linear momenta p ∈ Rndof , respectively. Together they constitute
the nodal phase space with co-ordinates collected in z ∈ R2ndof . The canonical equations in (12)
are now equivalent to the following initial value problem. Find z : I → R2ndof such that
ż = JDH (z) for 0¡t 6 T
(24)
z(0) = z0

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1937

with the skew-symmetric 2ndof × 2ndof (symplectic) matrix


 
0 I
J= (25)
−I 0
Here 0 and I denote the ndof × ndof zero and identity matrix, respectively. We further infer from
(12) that
 
Fint − Fext 1 nnode

DH (z) = with Fint = Fint ; : : : ; Fint ∈ Rndof (26)
M−1 p
A
where the internal nodal forces Fint : I → Rndim , A = 1; : : : ; nnode , are given by (13). Similarly, the
A
external nodal forces Fext follow from (7).
We next perform the time discretization of (24). Let 0 = t0 ¡t1 ¡ · · · ¡te = T be a partition of
the time interval I = [0; T ] of interest. In the following we focus on a typical time nite element
on the sub-interval In = [tn−1 ; tn ] corresponding to the time step n, n = 1; : : : ; e. It proves convenient
to introduce a transformation to a master element with local co-ordinate ∈ ˙I = [0; 1], de ned by

(t) = [t − tn−1 ]=hn with hn = tn − tn−1 (27)

Let Pk (0; 1) denote the set of polynomials of degree k on the interval (0; 1). Then the proposed
time nite element method reads: Find zh ∈ Pk (0; 1)2ndof that satis es the continuity condition
zh (0) = zn−1 and
Z 1
 
Jzh · dzh =d − hn JDH zh d = 0 (28)
0

h k−1 2ndof
for all z ∈ P (0; 1) . Here zn−1 denotes the algorithmic approximation to z at time tn−1 ,
resulting from the previous time step n − 1. For the initial time step, i.e. when n = 1, zn−1 is
prescribed by the initial conditions, i.e. z0 = (q0 ; Mq̇0 ). The trial functions zh ∈ Pk (0; 1)2ndof are
given by
P
k+1
zh ( ) = MI ( )zI (29)
I =1

where the nodal shape functions MI ( ), I = 1; : : : ; k + 1, coincide with Lagrange polynomials of


degree k such that MI ( J ) = IJ and zI = zh ( I ) are the nodal values of zh . Accordingly, the
global approximation to z : I → Rndim remains continuous. On the other hand, the test functions
zh ∈ Pk−1 (0; 1)2ndof are given by
P
k
zh ( ) = M̃I ( )zI (30)
I =1

with reduced shape functions M̃I ( ) determined by the relation


P
k+1 P
k
dzh =d = MI0 ( )zI = M̃I ( )z̃I (31)
I =1 I =1

The quantities z̃I , I = 1; : : : ; k, in (31) consist of linear combinations of the nodal values zI ,
I = 1; : : : ; k + 1, see Table I. Note that the nite element approximation (30) implies interelement

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1938 P. BETSCH AND P. STEINMANN

Table I. Nodal shape functions MI ( ), reduced shape functions M̃I ( ) and


associated quantities z̃I for polynomial approximations of degree k = 1; 2.

MI ( ) M̃I ( ) z̃I

k =1 M1 = 1− M̃1 = 1 z̃1 = z2 −z1


M2 =
k =2 M1 = [2 −1][ −1] M̃1 = 1− z̃1 = − 3z1 + 4z2 − z3
M2 = − 4[ 2 − ] M̃2 = z̃2 = z1 − 4z2 + 3z3
M3 = [2 − 1]

discontinuities of the test funtions. It is precisely this fact that leads to the recursive character of
the resulting time nite element formulation. Introduction of (29) and (30) into (28) eventually
yields the set of algebraic equations
Z 1 Z 1
P
k 
M̃I M̃J d z̃J − hn J M̃I DH zh d = 0 (32)
J =1 0 0

for I = 1; : : : ; k. The equations in (32) can be used to obtain a family of implicit one-step schemes,
a speci c member of which can be obtained according to the following two steps:
Step 1: The selection of the polynomial degree k speci es the nite element formulation. Con-
cerning the integrals in (32) involving only the shape functions MI and M̃I , exact
integration can be readily performed.
Step 2: The quadrature formula used for the evaluation of the remaining integrals in (32),
R1
namely 0 M̃I DH (zh ) d , nally determines a particular time-stepping scheme.
Step 2 plays a central role concerning the conservation properties and the accuracy of the
resulting algorithm. The implications of step 2 will be further investigated in the ensuing treatment
of particular time nite element formulations corresponding to k = 1.

3.1. Conservation of energy


3.1.1. Inherent energy conservation property. We now show that the time nite element formu-
lation (32) obeys the fundamental law of conservation of energy for any polynomial degree k
in a natural way. To this end consider (32) with autonomous Hamiltonian H (z). After scalar
multiplication with Jz̃I and subsequent summation over I = 1; : : : ; k, one obtains
Z 1 Z 1
P
k P
k 
M̃I M̃J d z̃J · Jz̃I − hn M̃I DH zh d · JT Jz̃I = 0 (33)
I; J =1 0 I =1 0

Employing (31) and the property JT = J−1 leads to


Z 1 Z 1
d h d  d h
z · J zh d − hn DH zh · z d = 0 (34)
0 d d 0 d

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1939

Due to the skew-symmetry of J the relation a·Ja = 0 holds for any a ∈ R2ndim . Now the fundamental
theorem of calculus can be used in (34) to obtain
Z 1 Z 1
 d h d   
DH zh · z d = H zh ( ) d = H zh ( (tn )) − H zh ( (tn−1 )) = 0 (35)
0 d 0 d

Accordingly, the total energy is conserved in the sense that Hn = Hn−1 , where Hn denotes the total
energy at time tn and Hn−1 denotes the total energy at time tn−1 .

3.1.2. Condition for algorithmic energy conservation. In many practical applications of the time
R1
nite element method at hand the exact evaluation of the time integrals 0 M̃I DH (zh ) d corre-
sponding to step 2 is not feasible. On the other hand, in view of (34) the time nite element
formulation implies
Z 1
P
k 
M̃I DH zh d · z̃I = 0 (36)
I =1 0

which is identical with Hn − Hn−1 = 0 in the case of exact time integration. Thus application
of numerical quadrature to the evaluation of the time integrals in question in general destroys
energy conservation. In the following we provide a condition which has to be ful lled by speci c
quadrature rules if algorithmic energy conservation is to be retained. To this end we further consider
(36) which in the case of vanishing external loading (Fext = 0) is equivalent to
"Z Z 1 #
1
Pk
−1 h
M̃I Fint d · q̃ I + M M̃I p d · p̃I = 0 (37)
I =1 0 0

where use has been made of (26). On the other hand, in view of (29) one has
Z 1 Z
P 1
k+1
M̃I ph d = M̃I MJ d pJ (38)
0 J =1 0

such that the exact evaluation of the time integrals associated with the inertial forces can be easily
performed. Thus the corresponding terms in (37) yield
Z 1 Z
P k
−1 h 1 1 d h 
M M̃I p d · p̃I = p · M−1 ph d = Tn − Tn−1 (39)
I =1 0 2 0 d
where (31) has been employed. In (39) Tn and Tn−1 denote the kinetic energy at time tn and tn−1 ,
respectively. Consequently, for algorithmic energy conservation the quadrature rule applied to the
calculation of the remaining time integrals in (37) has to ful ll the condition

k R
P 1
0
M̃I Fint d · q̃ I = Vintn − Vintn−1 (40)
I =1

Here Vintn and Vintn−1 denote the total strain energy at time tn and tn−1 , respectively. In the subsequent
construction of energy–momentum schemes we will use (40) as design condition for nonstandard
quadrature rules.

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1940 P. BETSCH AND P. STEINMANN

Figure 1. Linear time element with two nodes.

4. LINEAR TIME FINITE ELEMENTS

We now focus on linear time elements associated with linear trial functions, i.e. k = 1. According to
(29) one gets zh ( ) = M1 ( )z1 + M2 ( )z2 with nodal shape functions M1 ( ) = 1 − and M2 ( ) = .
The time nodes are located, respectively, at 1 = (tn−1 ) = 0 and 2 = (tn ) = 1 on the master
element, see Figure 1. Due to the continuity of the trial functions z1 = z( (tn−1 )) = (q1 ; p1 ) is
prescribed. Thus the task consists of nding z2 = z( (tn )) = (q2 ; p 2 ), that is, the nodal position
vectors and the linear momenta corresponding to time tn . This can be accomplished by solving
the system of non-linear equations furnished by the general time nite element method in (32).
Accordingly, setting k = 1 in (32) yields
Z 1

z2 − z1 − hn J DH zh d = 0 (41)
0

A more explicit form of (41) can now be written in the form


hn nP
node  
q2A − q1A − M −1 pB + p2B = 0
2 B=1 AB 1
Z 1 for A = 1; : : : ; nnode (42)
A A
 A A

p2 − p1 + hn Fint − Fext d = 0
0

Eventually, the quadrature rule employed for the evaluation of the time integral in (42) determines
a speci c time-stepping scheme. For instance, employing the mid-point rule leads to the mid-point
di erence method whereas application of the trapezoidal rule leads to the average acceleration
method, see Part I [1] for more details.
Our next objective is the design of non-standard quadrature rules which furnish energy–
momentum methods, that is, time-stepping schemes that conserve both angular momentum and
energy. To this end we next provide a condition for the design of quadrature rules which has to
be ful lled if conservation of angular momentum is to hold.

4.1. Algorithmic conservation of angular momentum


R1 A
We now elaborate on possible restrictions on the evaluation of 0 Fint d imposed by algorithmic
conservation of angular momentum. To this end let us consider the total angular momentum about
the origin at time tn , denoted as L2 = L(z2 ) = L(z( (tn ))), given by
nP
node
L2 = q2A × p2A (43)
A=1

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1941

Incorporating (42)1 and (42)2 yields


  Z 1
nP
node hn nP
node
−1 B
 A 
L2 = L1 − hn q1A + MAB p1 × A
Fint − Fext d (44)
A=1 2 B=1 0

On the other hand (42)1 and (42)2 imply the relationship


Z 1 Z 1
hn nP
node
−1 B
 A A
 1   A 
MAB p1 × Fint − Fext d = q2A − q1A × A
Fint − Fext d (45)
2 B=1 0 2 0

such that (44) leads to


Z 1
nP
node  
L2 = L1 − hn qAh 1
2 × A
Fint A
− Fext d (46)
A=1 0
 
Here qAh ( 12 ) = q1A + q2A =2 which conforms with (30). Consequently, in the case of vanishing
external loading, algorithmic conservation of angular momentum holds in the sense that L1 = L2
provided that the condition

nP
node R1 A 1

0
Fint d × qAh 2 =0 (47)
A

is ful lled. Accordingly, conservation of angular momentum severely restricts the number of
R1 A A
applicable quadrature rules for the calculation of 0 Fint d , where Fint is given in (13.)

4.2. Quadrature rules


We next elaborate on possible quadrature rules which satisfy condition (47) for conservation of
angular momentum. In a rst step we consider
Z 1 nP
node
Z 1Z nP
node
Z 1Z
A
Fint d = SAB dV qBh d ≈ SAB dV d qBh ( 12 ) (48)
0 B=1 0 B B=1 0 B
R1
where use has been made of (13). Note that qAh ( 12 ) = 0 qAh ( ) d such that qAh ( 12 ) can be inter-
preted as the average of qAh in In . The time integration on the right-hand side of (48) is nally
completed by the speci cation of the quadrature rule applied to the quantities SAB . For example,
R1
application of the mid-point rule, i.e. 0 SAB d ≈ SAB | = 1=2 , coincides with the consistent applica-
tion of the mid-point rule to the left-hand side of (48). Then the resulting time-stepping scheme
is equivalent to the mid-point di erence scheme (cf. Section 4:4:1 in Part I [1]) which is well
known to conserve angular momentum.
Note, however, that any reasonable quadrature rule can be employed for the calculation of
R1 R1 A
S d . In the second step, which eventually completes the evaluation of 0 Fint
0 AB
d in (48), we
focus on this point.

4.3. Energy–momentum schemes


We next complete the design of quadrature rules such that in addition to conservation of angular
momentum total energy is preserved as well. To this end we start from (48) where, with regard

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1942 P. BETSCH AND P. STEINMANN

to (13) and (5), the evaluation of the time integral


Z 1 Z 1
SAB d = ∇NA · S d ∇NB (49)
0 0

has to be speci ed. The key requirement for the design of energy–momentum methods is that
condition (40) for energy conservation has to be ful lled by the quadrature rule for (49). For
k = 1 condition (40) reads
Z 1
nP
node
A
 
Vint (q2 ) − Vint (q1 ) − Fint d · q2A − q1A = 0 (50)
A=1 0

Here Vint (q2 ) = Vint (q( (tn )) is the total strain energy at time tn and Vint (q1 ) = Vint (q( (tn−1 )) is the
total strain energy at time tn−1 . An alternative local statement of (50) can be written as
Z 1
nP
node   
W (C2 ) − W (C1 ) − SAB d qBh 12 · q2A − q1A = 0 (51)
A; B=1 0

where (48) has already been incorporated. Here W (C) denotes the strain energy density and CI ,
I = 1; 2, denotes the deformation tensor at time tn−1 and tn , which, according to (4), is given by
nP
node
CI = q IA · qBI ∇NA ⊗ ∇NB for I = 1; 2 (52)
A; B=1

With regard to the symmetry of SAB , (51) may also be written in the form
Z 1
nPnode 1 
W (C2 ) − W (C1 ) − SAB d q2A · qB2 − q1A · q1B = 0 (53)
A; B=1 0 2

Substitution from (49) into (53) yields


Z 1
nPnode 1  
W (C2 ) − W (C1 ) − S d : ∇NA ⊗ ∇NB q 2A · qB2 − q1A · q1B = 0 (54)
A; B=1 2 0

which in view of (52) can also be written as


Z 1
1
W (C2 ) − W (C1 ) − S d : [C2 − C1 ] = 0 (55)
2 0

Accordingly, if the time integration of the second Piola–Kirchho stress tensor satis es condition
(55) the corresponding time-stepping scheme conserves total energy as well as angular momentum.

4.3.1. Assumed strain method in time. Condition (55) for energy conservation can be enforced
by introducing the ‘assumed strain method in time’. Our approach was motivated by the so-called
‘assumed strain method’ which has been developed for low-order space nite elements, see e.g.
Reference [29]. In this connection it is well known that appropriate assumed strain modi cations
of the strain eld furnished by pure displacement nite element formulations can eliminate locking
behaviour. This approach has been especially successful in the context of low-order shell nite
elements, see e.g. Reference [32] and references therein.

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1943

For the low-order (k = 1) time nite element formulation at hand we propose the following
assumed strain approach : Use the assumed strain modi cation
P
2

C( ) = MI ( )CI (56)
I =1

for the calculation of the stress tensor according to S := 2DW (C)


 such that the time integrals
under consideration take the form
Z 1 Z 1
S d = 2 
DW (C( )) d (57)
0 0

Accordingly, instead of calculating the deformation tensor at the con guration qh ( ) via (4), the
assumed strain modi cation rests on the nodal values CI according to (52). We now show that
formula (57) leads to an ecient implementation of energy–momentum schemes. Recall that the
fundamental theorem of calculus implies the relationship
Z 1
W (C2 ) − W (C1 ) − 
DW (C( )) d : [C2 − C1 ] = 0 (58)
0

Comparison of (58) with (55) leads to the conclusion that the assumed strain modi cation
incorporated according to formula (57) implies energy conservation.

Example 4.1. We now illustrate the application of formula (57) for the design of energy-
momentum methods. Let us consider St. Venant–Kirchho material with strain energy density
function given by
W (C) = 12 E : D : E with E = 12 [C − I] (59)
Here E denotes the material strain tensor and D is the P fourth-order elasticity tensor. Note that the
assumed strain interpolation (56) implies E( )  2
= I =1 MI ( )EI with nodal values EI =
[CI − I]=2; I = 1; 2. Then the right-hand side of (57) can be exactly calculated according to

Z 1 Z 1
2 
DW (C( )) d = D : 
E( ) d
0 0
"Z Z #
1 1
=D: M1 ( ) d E1 + M2 ( ) d E2
0 0

= D : 12 [E1 + E2 ] (60)

In summary, the non-standard quadrature rule resulting from (48), (57) in conjunction with (60)
reads
Z 1 Z
A
nP
node
Bh 1

Fint d ≈ q 2 ∇NA ⊗ ∇NB : D : 12 [E1 + E2 ] dV (61)
0 B=1 B

Substitution from (61) into (42) eventually speci es the time-stepping scheme which conserves
angular momentum as well as energy. It is interesting to note that for St. Venant–Kirchho material

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1944 P. BETSCH AND P. STEINMANN

the present scheme is equivalent to the energy–momentum di erence method proposed by Simo
and Tarnow [8].

Remark 4.1. Concerning the evaluation of the time integrals associated with the linear nite
element formulation (k = 1) at hand, the mid-point rule can be considered the natural choice.
Interestingly, application of the mid-point rule to the time integration in (60) suces to get the
exact result. Thus, for St. Venant–Kirchho material, it is merely the assumed strain modi cation
(56) that enables energy conservation while retaining conservation of angular momentum.

Concerning more elaborate stored energy functions for large elastic strains such as Ogden-type
material, see Reference [33], the time integration in (57) can be performed numerically. In this
case (48) together with (57) yield
Z 1 Z Z 1
nP
node 
A
Fint d ≈ 2 Bh 1
q 2 ∇NA ⊗ ∇NB : 
DW (C( )) d dV (62)
0 B=1 B 0

where a suciently accurate quadrature rule has to be applied to the time integral in (62) to
enforce algorithmic energy conservation. Note that the time integration in (62) is con ned to the
constitutive routine which facilitates a very ecient implementation, see Section 4.4.2 for details.
In addition to being inexpensive this approach has the virtue that the material part of the tangent
operator remains unaltered, see Section 4.4.2 for further details. This is in contrast to the alternative
energy–momentum scheme based on the discrete gradient method considered in the next section.

4.3.2. Discrete gradient method. We next discuss a second noteworthy possibility for the calcula-
R1
tion of the time integral 0 S d which leads to an exact ful llment of condition (55) for arbitrary
strain energy density functions. The approach relies on the introduction of a discrete gradient. For
a detailed treatment of discrete gradients for Hamiltonian systems we refer to Gonzalez [34] and
references therein.
In the present case the gradient of the strain energy density function, i.e. DW (C) is replaced
by the discrete version

 1
 W (C2 ) − W (C1 ) − DW C 12 : [C2 − C1 ]
DW (C1 ; C2 ) := DW C 2 + [C2 − C1 ] (63)
||C2 − C1 ||2

where our notation according to (56) means C 12 = [C1 + C2 ]=2. It can be easily veri ed that the
relation DW (C1 ; C2 ) : [C2 − C1 ] = W (C2 ) − W (C1 ) holds. Note also that the discrete gradient in
(63) merely depends on the deformation tensors C1 and C2 at times tn−1 and tn , respectively. The
discrete gradient motivates the design of the following non-standard quadrature rule :
Z 1  
 W (C2 ) − W (C1 ) 

S d ≈ 2DW C 2 + 2 1 
− M : 2DW C 2 1
M (64)
0 ||C2 − C1 ||
where
C2 − C1
M= (65)
||C2 − C1 ||
has been introduced. It can be easily checked that condition (55) for energy conservation is exactly
satis ed when (64) is applied. The discrete scheme resulting from inserting (64) and (48) into
(42) coincides with the energy–momentum method due to Simo and Gonzalez [35].

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1945

In summary, the calculation of the relevant time integral in (42) amounts to the approximation
Z 1 Z Z 1
A
nP
node 
Fint d ≈ qBh 12 ∇NA ⊗ ∇NB : S d dV (66)
0 B=1 B 0
R1
with 0
S d being replaced by (64).

4.4. Numerical implementation



To solve for the unknowns at time tn , i.e. z2A = q 2A ; p2A ; A = 1; : : : ; nnode , we proceed along the
lines of Part I. Referring to the nite element formulation (42) we introduce the residual vector
R = (R1 ; : : : ; Rnnode ) with
Z 1
2 nP node    A 
RA = MAB qB2 − q1B − 2p1A + hn A
Fint − Fext d for A = 1; : : : ; nnode (67)
hn B=1 0

For simplicity we consider dead loads such that the time integration of the prescribed external
R1 A
loading, i.e. the calculation of 0 Fext d , can be easily accomplished. On the other hand, the
R1 A
evaluation of 0 Fint d can be performed by employing one of the quadrature rules investigated
above. Then R(q 2 ) constitutes an algebraic system of non-linear equations which can be solved by
means of the Newton–Raphson method. Table II contains an outline of the main steps pertaining
to the iterative solution procedure.

Remark 4.2. Due to the compact support of the nodal shape functions NA : B → R,
A = 1; : : : ; nnode corresponding to the space nite element method, the tangent operator K in Table
II is sparse, that is, many of the submatrices KAB are identical with the ndim × ndim zero matrix.

The quadrature rule used for the evaluation of the time integrals in Table II speci es a particular
member of the family of time-stepping schemes generated by the time nite element method at
R1 A
hand. Moreover, it becomes evident from Table II that the speci c quadrature rule for 0 Fint d
plays an important role concerning the calculation of the tangent operator de ned by K = @R=@q 2 .
We next investigate how particular quadrature rules a ect the implementation.

4.4.1. Standard quadrature rules. We rst consider the application of standard quadrature rules
R1 A
to the evaluation of the time integrals 0 Fint d . As has been shown in Part I well-known nite
di erence schemes can be extracted from the linear time nite element method at hand. For
example, employing the one-point Gauss quadrature rule yields the mid-point time-stepping scheme.
On the other hand, due to the inherent energy conservation of the present time nite element
formulation algorithmic energy conservation can be simply enforced by increasing the number of
Gauss points. Applying standard (interpolatory) quadrature rules, the time integrals in question can
be calculated according to
Z 1 Z
A P
lint nP
node
Fint d ≈ wl qCh (l ) ∇NA ⊗ ∇NC : S|l dV (68)
0 l=1 C=1 B

where wl are the weights and l ∈ [0; 1] are the co-ordinates of the integration points l = 1; : : : ; lint
on the master element. Furthermore, S|l symbolizes the stress calculation at qh (l ) via (4) and (5).

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1946 P. BETSCH AND P. STEINMANN

Table II. Non-linear elastodynamics: summary of computations for


one typical time step in the case k = 1 corresponding to linear time
nite elements.
Given:initial conditions: q1A ; p1A ; A = 1; : : : ; nnode
time step size: hn
set iteration counter: l = 1

Find: nodal unknowns q 2A and p2A ; A = 1; : : : ; nnode


(a) initialization
(1)
nP
node
−1 B
qA2 = q1A + hn MAB p1
B=1
(b) compute residual
h (l) i Z 1h i
(l) 2 nPnode
A(l) A
RA = MAB qB2 − q1B − 2p1A + hn Fint − Fext d
hn B=1 0
(l) 1 2 nnode (l)
R = {R ; R ; : : : ; R }
(l)
if ||R ||¿ goto (c) else goto (d)
(c) compute tangent
Z 1 (l)
A
(l) 2 @Fint
KAB = MAB Indim + hn d
hn 0 @qB2
 11 1nnode (l)
K ··· K
(l) . .. .. 
K ..
= . . 
Knnode 1 · · · Knnode nnode
n
solve for increment q 2 = {q12 ; : : : ; q2node }
−1
q 2 = − K(l) R(l)
update nodal position vectors
(l+1) (l)
qA2 = qA2 + q 2A
goto (b) with l = l + 1
(d) update linear momenta
(l) 2 nPnode h (l) i
p2A = MAB qB2 − q1B − p1A
hn B=1

Consistent linearization of the right-hand side of (68) yields the contribution to the submatrices
KAB of the tangent operator in Table II. One may write
Z 1 A
@Fint AB AB
B
d ≈ Kgeo + Kmat (69)
0 @q 2

where the result of the linearization process has been additively decomposed into the geometric part
Z
AB P
lint
Kgeo = wl M2 (l ) ∇NA ⊗ ∇NB : S|l dV Indim (70)
l=1 B

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1947

and the material part


Z
AB P
lint nP
node
Kmat = wl M2 (l ) qCh (l ) ⊗ qDh (l ) ∇NA ⊗ ∇NC : 4D2 W (C)|l : ∇ND ⊗ ∇NB dV (71)
l=1 C; D=1 B

Note that the loop over the time integration points includes the summation over the nodes corre-
sponding to the space nite element discretization. This fact leads to a comparatively expensive
computer implementation for lint ¿1.

4.4.2. Non-standard quadrature rules. Assumed strain method. We next focus on the computa-
tional implications of the assumed strain method in time developed in the above. With regard to
(62), for general strain energy densities, the method is based on the non-standard quadrature rule
given by
Z 1 Z l 
nP
node  Pint
A
Fint d ≈ qBh 12 ∇NA ⊗ ∇NB :  l )) dV
wl 2DW (C( (72)
0 B=1 B l=1

As before wl and l are the weights and abscissae for [0; 1]. Note that the numerical time integration
in (72) is con ned to the stress calculation. In analogy to (69) the consistent linearization of the
right-hand side of (72) yields the geometric part
Z l 
1 Pint
AB
Kgeo = ∇NA ⊗ ∇NB : 
wl 2DW (C(l )) dV Indim (73)
2 B l=1

and the material part


Z  
nP
node  P
lint
KAB
mat = qCh 12 ⊗ q 2D ∇NA ⊗ ∇NC : 2 
wl M2 (l )4D W (C(l )) : ∇ND ⊗ ∇NB dV (74)
C; D=1 B l=1

Again the numerical time integration in (74) can be explicitly performed, that is, independent of
the summation over the nodes corresponding to the space nite element discretization. Essentially,
merely the constitutive terms in (72)– (74) enter the numerical time integration which facilitates
an ecient implementation. Consequently, the computational cost of increasing the number of time
integration points lint turns out to be negligible.
Note that in the case of St. Venant–Kirchho material only one Gauss point, i.e. lint = 1, 1 = 12
and w1 = 1, is sucient to obtain a time-stepping scheme which exactly conserves energy as well
as angular momentum.

4.4.3. Non-standard quadrature rules. Discrete gradient method. We next consider the
non-standard quadrature rule (64) emanating from the introduction of the discrete gradient in
(63). In view of (66) one obtains
Z 1 Z
A
nP
node 
Fint d ≈ qBh 12 ∇NA ⊗ ∇NB :  dV (75)
0 B=1 B

with
 
 W (C2 ) − W (C1 ) 
 = 2DW C 1
2 + 2 − M : 2DW C 1
2 M (76)
kC2 − C1 k

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1948 P. BETSCH AND P. STEINMANN

where the second-order tensor M has been de ned in (65). Proceeding as before we obtain the
geometric part
Z
AB 1
Kgeo = ∇NA ⊗ ∇NB :  dV Indim (77)
2 B
Moreover, a tedious but straightforward calculation yields the material part
nP
node
Z
AB
Kmat = qCh ( 12 ) ⊗ q2D ∇NA ⊗ ∇NC : 2@=@C2 : ∇ND ⊗ ∇NB dV (78)
C; D=1 B

with

"  #

 DW (C2 ) − DW C 12 
2 1
@=@C2 = D W C 2 + M ⊗ 2 − M : D2 W C 1
2
kC2 − C1 k
"  #
W (C2 ) − W (C1 ) DW C 12
+2 −M: [I − 2M ⊗ M] (79)
kC2 − C1 k2 kC2 − C1 k

Here (I)ijkl = 12 [ik jl + il jk ] denotes the fourth-order identity tensor. It is obvious from (79)
that (@=@C2 )ijkl is symmetric in ij and kl (minor symmetry condition). Note, however, that the
major symmetry condition (D2 W (C))ijkl = (D2 W (C))klij does not hold for @=@C2 anymore.

5. NUMERICAL INVESTIGATIONS

In this section we investigate the numerical performance of various time-stepping schemes furnished
by the time nite element formulation at hand. We consider essentially two distinct types of
motion: (i) Quasi-rigid motion related to a large bulk modulus which can be considered as penalty
parameter for the enforcement of the rigidity constraint. (ii) Large-strain motion associated with
a comparatively exible material behaviour. The following time-stepping schemes associated with
linear time nite elements (k = 1) will be employed in the calculations:
FE(Nl ) Standard time nite elements relying on Nl Gauss points (see Section 4.4.1).
AS(Nl ) Assumed strain method in conjunction with Nl Gauss points (see Sections 4.3.1 and 4.4.2).
DGM Discrete gradient method (see Sections 4.3.2 and 4.4.3).
Our numerical investigations focus in particular on the accuracy, energy conservation and con-
servation of angular momentum of the various time-stepping schemes. In this connection we sum-
marize in Table III the investigated conservation properties of the time-stepping schemes under
consideration. These properties are con rmed by our numerical experiments as documented in the
next section.

5.1. Motion of L-shaped block


The initial mesh con guration of a L-shaped block in two dimensions (ndim = 2) is shown in
Figure 2. The spatial discretization relies on 36 displacement-based isoparametric 4-node nite
elements. Starting at rest the block is subjected to spatially xed external loads (see Figure 2)

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1949

Table III. Conservation properties of time-stepping schemes associated with linear


time nite elements (i.e. nite elements based on the cG(1) method). Note that
for FE(Nl ) and AS(Nl ), Nl denotes the (problem-dependent) minimum number of
integration points which is required for algorithmic energy conservation to hold.

Angular Conservation of total energy


Scheme momentum St. V.-K. material General elastic material

FE(1) Yes No No
FE(Nl ) No Yes Yes
AS(1) Yes Yes No
AS(Nl ) Yes Yes Yes
DGM Yes Yes Yes

Figure 2. Initial mesh con guration of Figure 3. Time history of external loading for
L-shaped block. quasi-rigid case.

that are applied according to the load history depicted in Figure 3. After t ¿ 1 the external loads
vanish such that the total energy as well as the total angular momentum have to be conserved.
In order to provide uniform loading conditions we generally apply a time step size of hn = 0:01
during the loading interval 0¡t¡1. Then the time step size is adjusted as documented below.

5.1.1. Quasi-rigid motion. We rst elaborate on the quasi-rigid motion of the block governed
by St. Venant–Kirchho material with material constants E = 107 and  = 0:3. The reference mass
density is assumed to be %R = 1. Figure 4 illustrates the quasi-rigid motion of the block by showing
a sequence of con gurations calculated with AS(1) and hn = 0:2.
Accuracy considerations. We next calculate the mean-square norm of the error in the displace-
ments according to
Z tb nnode 1=2
P  A 2
||eq ||L2 (ta ; tb ) = q (t) − q ∗A (t) dt (80)
ta A=1

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1950 P. BETSCH AND P. STEINMANN

Figure 4. Quasi-rigid motion: sequence of con gurations corresponding to t = 0; 1; 2; 3; 4; 5; 6.

Figure 5. Quasi-rigid motion: mean-square norm of Figure 6. Quasi-rigid motion. FE(5): sensitivity
the error in the displacements. with respect to time step change, hn = 0:01 for
0¡t 6 2, then change to, respectively, hn = 0:02,
hn = 0:025 and hn = 0:04 for 2¡t 6 4.

Here q∗A (t) denotes the reference solution that has been calculated by using hn = 0:0001. Moreover,
we set ta = 1 and tb = 2 in (80). The results are depicted in Figure 5. Accordingly, the rate of
convergence of the energy–momentum scheme AS(1) and the mid-point rule FE(1) is two with
respect to the mean-square norm. Note, however, that AS(1) yields more accurate results than
FE(1). The results of FE(5), also shown in Figure 5, are rather disappointing. Similar results
have been documented in Part I for the N-body problem and again one may conclude that al-
gorithmic energy conservation alone, e ectively obtained with FE(5), does not imply a good
numerical performance. In fact, as has been shown in the above, enforcement of the inherent en-
ergy conservation property of the continuous Galerkin method by using a higher-order quadrature
rule (e.g. 5 Gauss points) is at the expense of conservation of angular momentum. Thus the lack
of algorithmic conservation of angular momentum seems to be responsible for the deterioration of
accuracy shown in Figure 5.

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1951

Figure 7. Quasi-rigid motion. FE(1): sensitivity Figure 8. Quasi-rigid motion. Insensitivity of en-
with respect to time step change, hn = 0:01 for ergy–momentum scheme AS(1) with respect to
0¡t 6 2, then change to, respectively, hn = 0:02, time step change: hn = 0:01 for 0¡t 6 2, then
hn = 0:025 for 2¡t 6 4. change to hn = 0:04 for 2¡t 6 3, then change to
hn = 0:2 for 3¡t 6 4.

Time step changes and conservation properties. An important property of one-step methods,
especially in view of time adaptivity, is that the time step size can be easily adjusted (see e.g.
Reference [4]). We next examine the performance of the various time-stepping schemes with
respect to time step changes. To this end we use hn = 0:01 for 0¡t 6 2 and then change the time
step size.
First let us consider FE(5) which is e ectively energy conserving. It becomes evident from
Figure 6 that the time step change at t = 2 induces severe oscillations in the angular momentum
component L3 .
Next, we consider the symplectic mid-point rule FE(1) which conserves angular momentum. As
can be seen from Figure 7 the time step change at t = 2 leads to severe oscillations in the total
energy which eventually result in numerical blow up.
In contrast to the time step sensitivity of FE(1) and FE(5) documented above the energy–
momentum scheme AS(1) turns out to be entirely insensitive with respect to time-step changes.
It can be seen from Figure 8 that both total angular momentum and total energy are not a ected
by the serious time step changes at t = 2 and 3.

5.1.2. Large-strain motion. We next consider the L-shaped block in Figure 2 with constitutive
behaviour now governed by compressible Neo-Hooke material with associated strain energy density
function given by
 
W (C) = [tr C − 3] + (ln J )2 −  ln J (81)
2 2

where J = det C. The material constants are assumed to take the values  = 1000,  = 500 and
%R = 0:5. Again the external loading is given in form of a hat function depicted in Figure 9. During
the loading period a uniform time-step size of hn = 0:01 is used.

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1952 P. BETSCH AND P. STEINMANN

Figure 9. Time history of external loading for large-strain case.

Figure 10. Large-strain motion: sequence of con gurations corresponding to t = 0; 1; 2; 3; 4.

Figure 11. Large-strain motion: numerical veri ca- Figure 12. Large-strain motion: time history of the
tion of conservation of total angular momentum and total energy for time-step size hn = 0:04.
conservation of total energy for AS(5) and DGM
(hn = 0:01 for 0¡t 6 1, then hn = 0:04 for t¿1).

Figure 10 shows a sequence of deformed con gurations corresponding to t = 0; 1; 2; 3; 4. It can


be observed that the block undergoes large deformations accompanied by nite strains.
Conservation properties. Figure 11 contains a numerical veri cation of the algorithmic conser-
vation properties, i.e. conservation of energy and angular momentum, for AS(5) and DGM. On

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1953

Figure 13. Large-strain motion: mean-square norm of the error in the displacements.

the other hand FE(1) again exhibits a lack of energy conservation as documented in Figure 12.
In the present case one may rather speak of energy drift than energy blow up.
Concerning time-step changes we did not observe the pronounced oscillation behaviour as re-
ported for FE(1) and FE(5) in the previous case of quasi-rigid motion. Furthermore, there does
not appear to be a big di erence concerning the accuracy of the various schemes:
Accuracy considerations. As in the case of quasi-rigid motion we calculate the mean-square
norm of the error in the displacements according to (80) with ta = 1 and tb = 2 . Again the
reference solution q∗A (t) has been calculated with hn = 0:0001. The results are shown in Figure 13.
Accordingly, FE(1), FE(5), AS(5) and DGM yield almost equal results. The rate of convergence
of all the schemes is about two with respect to the mean-square norm.

6. CONCLUSIONS

We elaborated on a fully discrete nite element Galerkin method for non-linear elastodynamics.
The continuous Galerkin (cG) method applied to the semidiscrete equations of non-linear elas-
todynamics in Hamiltonian form turned out to be well suited for the development of conserving
time-stepping schemes. Rather than enforcing the inherent energy conservation property of the
cG method by employing high-order quadrature formulas we developed new quadrature rules that
(i) retain algorithmic energy conservation, (ii) minimize the computational e ort and (iii) simul-
taneously enable algorithmic conservation of angular momentum.
In analogy to the well-known assumed strain method for space nite elements we introduced
the ‘assumed strain method in time’. For St. Venant–Kirchho material and linear time elements
the assumed strain method in time directly leads to an energy–momentum conserving scheme that
is identical to the algorithm proposed in Reference [8]. Furthermore, we saw that the assumed
strain approach can be conveniently applied to the case of arbitrary strain energy densities. In this
connection we saw that the discrete gradient method provides an alternative tool for the design of
quadrature formulas for energy–momentum schemes. The corresponding time-stepping algorithm
coincides with the scheme proposed in Reference [35].

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
1954 P. BETSCH AND P. STEINMANN

In summary the time nite element approach advocated herein appears to provide a number of
advantages when compared with customary nite di erence methods:
1. Arbitrary order time nite elements can be systematically developed. In the present paper
we focused on linear time elements which seem to be most important for non-linear elasto-
dynamics. The formulation of higher-order time elements can be performed in a straightfor-
ward way, see e.g. Part I [1] where also quadratic time elements are treated in detail.
2. A family of time-stepping schemes can be easily implemented. A particular scheme results
simply from the selection of the corresponding quadrature rule. For example, the implementa-
tion described herein (see Table II) embraces the aforementioned energy–momentum schemes
as well as schemes based on standard quadrature formulas such as the average acceleration
method and the mid-point rule.
3. Finite element technologies that have been developed to improve nite elements in space
may be applied to the design of time nite elements by way of analogy. One example
for this approach is the assumed strain method treated herein. In addition to that, well-
known terminology from space nite elements can also be used for time nite elements. For
example, the term ‘locking’ is often employed for the bad numerical performance of low-
order elements in highly constrained situations such as the incompressible limit or thin-shell
bending. Our numerical simulations indicate that the often observed energy-blow-up behaviour
of schemes such as the mid-point rule can be interpreted as ‘locking’ of the low-order time
nite element formulation. In this case the bulk modulus can be identi ed as penalty term
for the enforcement of the rigidity constraint.
4. Galerkin methods facilitate the computation of residual based error estimates that can be used
for adaptive error control. The insensitivity of energy–momentum schemes with respect to
time-step changes, as documented in our numerical examples in Section 5.1.1, obviously is
of crucial importance when it comes to time adaptivity.

REFERENCES
1. Betsch P, Steinmann P. Conservation properties of a time nite element method. Part I: Time-stepping schemes for
N-body problems. International Journal for Numerical Methods in Engineering 2000; 49(5):599–638.
2. Simo JC, Tarnow N, Wong KK. Exact energy–momentum conserving algorithms and symplectic schemes for nonlinear
dynamics. Computer Methods in Applied Mechanics and Engineering 1992; 100:63–116.
3. Hughes TJR. Stability, convergence and growth and decay of energy of the average acceleration method in nonlinear
structural dynamics. Computers and Structures 1976; 6:313–324.
4. Wood WL. Practical Time-Stepping Schemes. Oxford University Press: New York, 1990.
5. Hughes TJR, Caughey TK, Liu WK. Finite-element methods for nonlinear elastodynamics which conserve energy.
Journal of Applied Mechanics 1978; 45:366 –370.
6. Kuhl D, Ramm E. Constraint energy momentum algorithm and its application to nonlinear dynamics of shells. Computer
Methods in Applied Mechanics and Engineering 1996; 136:293–315.
7. Ortiz M. A note on energy conservation and stability of nonlinear time-stepping algorithms. Computers and Structures
1986; 24:167–168.
8. Simo JC, Tarnow N. The discrete energy–momentum method. Conserving algorithms for nonlinear elastodynamics.
Zeitschrift fner Angewandte Mathematik and Physik (ZAMP) 1992; 43:757–792.
9. Cris eld MA, Shi J. A co-rotational element=time-integration strategy for non-linear dynamics. International Journal
for Numerical Methods in Engineering 1994; 37:1897–1913.
10. Kuhl D, Cris eld MA. Energy-conserving and decaying algorithms in non-linear structural mechanics. International
Journal for Numerical Methods in Engineering 1999; 45:569–599.
11. Marsden JE, Ratiu TS. Introduction to Mechanics and Symmetry. Springer: Berlin, 1994.
12. Sanz-Serna JM. Symplectic integrators for Hamiltonian problems: an overview. Acta Numerica 1992; 1:243–286.
13. Bauchau OA, Theron NJ. Energy decaying scheme for non-linear beam models. Computer Methods in Applied
Mechanics and Engineering 1996; 134:37–56.

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955
CONSERVATION PROPERTIES OF A TIME FE METHOD—PART II 1955

14. Bauchau OA, Bottasso CL. On the design of energy preserving and decaying schemes for exible, nonlinear multi-body
systems. Computer Methods in Applied Mechanics and Engineering 1999; 169:61–79.
15. Armero F, Romero I. Dissipative integration algorithms for nonlinear elastodynamics. In Proceedings of the European
Conference on Computational Mechanics (ECCM’99), Munich, Wunderlich W. (ed.), 1999.
16. Marsden JE. Lectures on Mechanics. Cambridge University Press: Cambridge, 1992.
17. Hulme BL. One-step piecewise polynomial Galerkin methods for initial value problems. Mathematics of Computation
1972; 26(118):415– 426.
18. Eriksson K, Estep D, Hansbo P, Johnson C. Computational Di erential Equations. Cambridge University Press:
Cambridge, 1996.
19. Lasaint P, Raviart PA. On a nite element method for solving the neutron transport equation. In Mathematical Aspects
of Finite Elements in Partial Di erential Equations, de Boor C. (ed.), Academic Press: New York, 1974; 89–123.
20. Hughes TJR, Hulbert GM. Space-time nite element methods for elastodynamics: Formulations and error estimates.
Computer Methods in Applied Mechanics and Engineering 1988; 66:339 –363.
21. Li XD, Wiberg N-E. Structural dynamic analysis by a time-discontinuous Galerkin nite element method. International
Journal for Numerical Methods in Engineering 1996; 39:2131–2152.
22. Hulbert GM. A uni ed set of single-step asymptotic annihilation algorithms for structural dynamics. Computer Methods
in Applied Mechanics and Engineering 1994; 113:1–9.
23. Bauchau OA, Joo T. Computational schemes for non-linear elasto-dynamics. International Journal for Numerical
Methods in Engineering 1999; 45:693–719.
24. Aharoni D, Bar-Yoseph P. Mixed nite element formulations in the time domain for solution of dynamic problems.
Computational Mechanics 1992; 9:359 –374.
25. Peters DA, Izadpanah AP. hp-version nite elements for the space-time domain. Computational Mechanics 1988;
3:73–88.
26. Bottasso CL. A new look at nite elements in time: a variational interpretation of Runge–Kutta methods. Applied
Numerical Mathematics 1997; 25:355–368.
27. Borri M, Bottasso C. A general framework for interpreting time nite element formulations. Computational Mechanics
1993; 13:133–142.
28. Betsch P, Steinmann P. Inherently energy conserving time nite elements for classical mechanics. Journal of
Computational Physics 2000; 160:88–116.
29. Simo JC, Hughes TJR. On the variational foundations of assumed strain methods. Journal of Applied Mechanics 1986;
53:51–54.
30. Oden JT. Finite Elements of Nonlinear Continua. McGraw-Hill: New York, 1972.
31. Hughes TJR. The Finite Element Method. Prentice-Hall: New Jersey, 1987.
32. Betsch P, Stein E. A nonlinear extensible 4-node shell element based on continuum theory and assumed strain
interpolations. Journal of Nonlinear Science 1996; 6:169 –199.
33. Ogden RW. Non-Linear Elastic Deformations. Ellis Horwood: Chichester, 1984.
34. Gonzalez O. Time integration and discrete Hamiltonian systems. Journal of Nonlinear Science 1996; 6:449– 467.
35. Simo JC, Gonzalez O. Recent results on the numerical integration of in nite-dimensional Hamiltonian systems. In
Recent Developments in Finite Element Analysis, Hughes TJR, Onate E, Zienkiewicz OC (eds), CIMNE: Barcelona,
1994; 255–271.

Copyright ? 2001 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2001; 50:1931–1955

You might also like