A discontinuous Galerkin method for strain

gradient-dependent damage: Study of interpolations

and convergence
Luisa Molari
a
, Garth N. Wells
b
, Krishna Garikipati
c,
*
, Francesco Ubertini
a
a
DISTART, Universita` di Bologna, Viale Risorgimento 2, 40136 Bologna, Italy
b
Faculty of Civil Engineering and Geosciences, Delft University of Technology, Stevinweg 1, 2628 CN Delft, The Netherlands
c
Department of Mechanical Engineering, University of Michigan, Ann Arbor, MI 48109-2125, USA
Received 8 October 2004; received in revised form 27 February 2005; accepted 10 May 2005
Abstract
A discontinuous Galerkin method has been developed for strain gradient-dependent damage. The strength of this
method lies in the fact that it allows the use of C
0
interpolation functions for continuum theories involving higher-order
derivatives, while in a conventional framework at least C
1
interpolations are required. The discontinuous Galerkin for-
mulation thereby oers signicant potential for engineering computations with strain gradient-dependent models.
When using basis functions with a low degree of continuity, jump conditions arise at element edges which are incorpo-
rated in the weak form. In addition to the formulation itself, a detailed study of the convergence properties of the
method for various element types is presented and an error analysis is undertaken. Numerical results of some one-
and two-dimensional problems are presented and discussed.
2005 Elsevier B.V. All rights reserved.
Keywords: Strain gradient theory; Mixed methods; Stabilised methods; Damage
1. Introduction
The nucleation and growth of cracks can be described by continuum damage mechanics. In this ap-
proach, an internal variable is included in the constitutive law to represent the evolution of microstructural
0045-7825/$ - see front matter 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2005.05.026
*
Corresponding author. Fax: +1 734 936 0414.
E-mail address: krishna@engin.umich.edu (K. Garikipati).
Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
www.elsevier.com/locate/cma
damage [13]. The material moduli degrade with the increase of this parameter, which could be either scalar
or tensorial in character. A scalar internal variable is commonly used to model damage degradation for
cases in which the material remains isotropic. A tensorial form of the internal variable is used for the aniso-
tropic case.
Damage degradation can manifest itself in progressive material softening, for which reason numerical
results based upon classical continuum mechanics are characterised by a pathological mesh dependence
[23,5]. As the material softens, deformations localise within bands of nite thickness, which is viewed as
the smearing out of a crack [17,6]. However, the width of bands is intimately related to the element size.
As the mesh is rened the band width decreases, tending to zero in the limit. The result of this is seen in
loaddisplacement response with slopes that are increasingly negative with mesh renement in the softening
regime. The origin of this pathology lies in the loss of ellipticity of the material tangent modulus tensor for
softening inelastic continuum models in the absence of rate eects [18].
The last two decades have witnessed a great deal of work in regularised continuum models to avoid the
loss of ellipticity in the presence of material softening. Among these are the strain gradient models
[2,9,21,16], nonlocal models [7,11,8], localisation limiters [14], and Cosserat models [10], to identify but a
fraction of a vast body of literature. By various techniques, each of these models introduces an intrinsic
length scale to the continuum theory. When the material softens, the localisation band width is controlled
by this length scale, rather than the element size. As a result, the mesh size-dependency, described above, is
eliminated.
In this paper we aim to address numerical issues associated with some strain gradient models. This
class of models involves strains and strain gradients as kinematic terms. In some models work conjugate
couple stresses are also identied. Dimensional considerations lead to the introduction of intrinsic length
scales associated with the strain gradients. The presence of strain gradient terms alters the mathematical
character of the governing dierential equations, elevating them to a higher order (at least locally). A
numerical complication arises from the higher order character of the governing dierential equations,
as standard C
0
interpolation functions often do not furnish sucient continuity. The obvious approach,
that is to employ interpolation functions with higher-order continuity (C
1
and higher), has not proven
fruitful. Finite element methods with these interpolations are expensive and dicult to construct. There
exists some work with mixed formulations for this purpose [19,24]. However, the very simplest two-
dimensional elements involve up to 28 degrees of freedom, placing them beyond the realm of practicabil-
ity. Some authors have applied element-free Galerkin methods, on the basis of the arbitrary degree of
continuity that they can provide [4]. However, this class of methods introduces complications in the appli-
cation of boundary conditions, is somewhat less ecient than the nite element method, and its imple-
mentation involves a reformulation of the dominant nite element architecture, making its widespread
use less attractive. In addition, just as high-order continuity can be dicult to achieve, excessive continu-
ity can also be a drawback for many models, particularly at internal boundaries where the order of the
governing dierential equation changes.
In this work we seek to expand upon recent developments in discontinuous Galerkin methods for
strain gradient-dependent theories [12,22]. The great advantage of this class of methods is the ability
to use C
0
interpolation functions for the displacement when solving continuum theories involving
higher-order derivatives. In cases where other considerations make it advantageous to also interpolate
strain-like quantities, the approach can allow interpolated elds that are discontinuous across elements
[22].
The content of this paper is organised as follows. The strain gradient damage model of interest is de-
scribed in Section 2, with special attention paid to the boundary conditions. The Galerkin formulation
for this model is presented in Section 3. The convergence behaviour of the model in one dimension is
examined both analytically and numerically in Section 4, and the section is concluded with one- and
two-dimensional examples. Conclusions are then drawn in Section 5.
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1481
2. Gradient damage
Consider a body X, which is an open subset in R
n
, where n is the spatial dimension. The boundary of the
body is denoted by C = oX, and the outward unit normal to C is denoted by n. We consider quasi-static
equilibrium of the body, which is governed by
r r f 0 in X; 1
rn h on C
h
; 2
u g on C
g
; 3
where r is the stress tensor, f is the body force, h is the prescribed traction on the boundary C
h
and g is the
prescribed displacement on the boundary C
g
. The boundary C is partitioned such that C C
g
[ C
h
and
C
g
\ C
h
= ;. For isotropic damage, it is sucient to introduce a single scalar damage variable, D, satisfying
0 6 D 6 1. Considering that D = 0 implies no damage and D = 1 corresponds to a completely damaged
material point, the stressstrain relation is of the form
r 1 D C : e; 4
where C is the elasticity tensor and e is the strain tensor, e = $
s
u. The scalar damage variable, D, is related
to a scalar strain measure, e, through the KuhnTucker conditions:
e j 6 0; _ j P0; _ je j 0; 5
where D = D(j). The scalar strain measure, e, of interest here arises from a so-called explicit gradient dam-
age formulation. In this approach, the strain measure is dened to be
e e
eq
c
2
r
2
e
eq
; 6
where c is an intrinsic length scale, and e
eq
is a suitably chosen invariant of the local strain tensor.
For material points at which damage is developing, the governing equation is nonlinear and fourth-order
in the displacement. In undamaged or unloading regions ( _ j 0), the governing equation is linear and sec-
ond-order. Recognising the fourth-order form of the dierential equation (at least in sub-regions of X) leads
to the question of appropriate boundary conditions. Possible boundary conditions (in addition to the usual
boundary conditions on the displacement and traction) include:
e
eq
q on C
e
; 7
re
eq
n
d
r on C
e
0 ; 8
r re
eq
_ _
n
d
_ _
n
d
s on C
e
00 ; 9
where C
e
[ C
e
0 [ C
e
00 C
d
, C
e
\ C
e
0 ;, C
e
\ C
e
00 ;, C
e
0 \ C
e
00 ;, and n
d
indicates the unit outward nor-
mal to the boundary of the damaged domain, C
d
. These boundary conditions supplement the standard dis-
placement and traction boundary conditions on C. If the entire body X is damaging, then C
d
= C and the
suitable conditions should be specied among Eqs. (7)(9), together with the values for q and/or r and/or s.
Apart from thermodynamically consistent damage models, there is not yet a clear and physically-based
interpretation to serve as a guide. Common assumptions are: condition (8) with r = 0 or condition (9) with
s = 0. Commonly however, damage is localised and interface conditions arise on the boundaries which are
internal to X (the boundaries to sub-domains where _ j > 0 and _ j 0 meet). When the body force is smooth,
continuity of u and rn
d
are natural continuity conditions. At the interface between damaging and elastic
regions, an additional condition is required on the damage domain. In this case Eqs. (7)(9) should be re-
garded as transition conditions and quantities q, r and s directly descend from the solution on the elastic
region. The need for extra boundary conditions is made particularly clear by the analytical solution pre-
sented in Section 4.2.1. In the adopted formulation, this extra condition will come from implied continuity
of $e
eq
n
d
across the damage-elastic boundary.
1482 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
3. Discontinuous Galerkin formulation
A discontinuous Galerkin approach based on a two-eld formulation is advocated here to numerically
solve the outlined damage problem. The rationale for this choice lies in the serious diculties which are
inevitably encountered by operating within a conventional framework. In particular, a weak formulation
involving only the displacement eld as independent variable is made possible by inserting the constitutive
equations into the equilibrium equation and applying integration by parts twice. However, the resultant
nite element procedure would be specic to the chosen damage law, which is typically complex, and
not suited to easily handle a moving elastic-damage boundary (which is the interface between second-
and fourth-order solutions). Hence Eq. (6) is treated separately from the equilibrium Eq. (1) and the strain
measure e is taken as an independent variable. In addition, the most obvious approach based on a contin-
uous mixed Galerkin formulation demands C
1
continuity of displacement interpolation, due to the presence
of strain gradient terms to be computed from displacements. To circumvent the well-known diculties
engendered by higher-order continuity, a discontinuous Galerkin formulation is adopted by imposing
the required degree of continuity of the displacement eld in a weak sense. A wider discussion of the con-
ventional nite element approach can be found in [22].
The body X is partitioned into n
el
non-overlapping elements, which satisfy standard nite element
requirements. The typical element is denoted by the open set X
e
. For later convenience it is useful to intro-
duce a domain

X which does not include element boundaries and the interior boundary

C:

X
_
n
el
e1
X
e
;

C
_
n
b
i1
C
i
; 10
where C
i
is the ith interior element boundary and n
b
is the number of internal inter-element boundaries. In
addition, jump and averaging operations are dened. The jump in a eld a across a surface is given by
sat a
1
n
1
a
2
n
2
; 11
where the subscripts denote the side of the surface and n is the outward unit normal vector (n
1
= n
2
in the
geometrically linear case). The average of a eld a across a surface is given by
hai
a
1
a
2
2
. 12
On such a partition, the following function spaces are considered:
S
h
u
h
i
2 H
1
X j u
h
i
j
X
e
2 P
k
1
X
e
8e; u
i
g
i
on C
g
_ _
; 13
V
h
w
h
i
2 H
1
X j w
h
i
j
X
e
2 P
k
1
X
e
8e; w
i
0 on C
g
_ _
; 14
W
h
q
h
2 L
2
X j q
h
j
X
e
2 P
k
2
X
e
8e
_ _
; 15
where P
k
represents the space of polynomial nite element shape functions of order k. The spaces S
h
and
V
h
represent usual, C
0
continuous nite element shape functions. Note that the space W
h
contains discon-
tinuous functions.
A weak form of the gradient enhanced damage problem in terms of displacement u and strain e can be
now stated as follows [22]: nd u
h
2 S
h

n
and e
h
2 W
h
such that
_
X
r
s
w
h
: 1 D e
h
_ _ _ _
C : r
s
u
h
dX
_
C
h
w
h
hdC
_
X
w
h
f dX

_
C\C
d
a
2
h
2
r r rw
h
_ _
eq
_ _ _ _
n Ec
2
r re
h
eq
_ _ _ _
ndC 0 8w
h
2 V
h
_ _
n
; 16
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1483
_
X
q
h
e
h
dX
_
X
q
h
e
h
eq
dX
_
~
X
rq
h
c
2
re
h
eq
dX
_
C
q
h
c
2
re
h
eq
ndC
_
~
C
sq
h
t c
2
hre
h
eq
i dC

_
~
C
hrq
h
i c
2
se
h
eq
tdC
_
~
C
ac
2
h
sq
h
t se
h
eq
tdC 0 8q
h
2 W
h
. 17
Except for the last term, the rst equation is the standard Galerkin problem for the equilibrium Eq. (1). It is
worth emphasising that u
h
is assumed to be C
0
continuous. The additional term attempts to impose the
eventual higher-order boundary condition ($($e
eq
)n) n = 0 on C \ C
d
through the penalty parameter a
2
,
with E being the Youngs modulus. The enforcement of this boundary condition is not in the spirit of dis-
continuous Galerkin methods, as the resulting Galerkin problem is not consistent, in the same sense that
penalty methods are not consistent (the restoration of consistency would require an approach analogous
to Nitsches method). However, this is of no consequence for the considered problem due to a fortuitous
choice of interpolation, as will be shown later. Preserving consistency while weakly imposing the non-stan-
dard boundary condition is a complicated task due to the nonlinear dependency on e. This is a topic requir-
ing further investigation. In particular, no systematic numerical tests have been carried out on such a weak
condition. Extra boundary conditions on elastic-damage boundaries (C
d
C) are not explicitly applied
since the necessary conditions are implied implicitly in the formulation, as will be shown in examining con-
sistency of Eq. (17).
The second equation is constructed to solve for e starting from a weak form of Eq. (6) [22], a being a pen-
alty-like parameter and h the element dimension. Unlike standard formulations, no gradients of e
h
eq
, e
h
or q
h
appear in terms integrated over X (which includes interior boundaries), hence u
h
can be C
0
continuous across

C, and e
h
and q
h
can be discontinuous, according to the assumed nite element spaces (see Eqs. (13)(15)).
Notice that equation (17) resembles the interior penalty method, which can be interpreted within the wider
framework of discontinuous Galerkin methods [3]. Moreover, if the space W
h
is restricted to C
0
continuous
functions, the formulation would then resemble a continuous/discontinuous Galerkin method [12].
Consistency of Eq. (17) can be easily proved by employing standard variational arguments. In particular
the following conditions are weakly enforced [22]:
e e
eq
c
2
r
2
e
eq
0 in

X; 18
c
2
se
eq
t 0 on

C; 19
c
2
sre
eqn
t 0 on

C. 20
The rst condition corresponds to Eq. (6) over element interiors. The other two conditions impose conti-
nuity of the corresponding elds across element boundaries. Owing to these continuity conditions, the for-
mulation naturally provides for the extra boundary conditions on elastic-damage interfaces.
The discrete damage problem formulated above requires the simultaneous solution of Eqs. (16) and (17),
which are nonlinear and coupled through the dependency of D on e
h
and the dependency of e
h
on u
h
. Lin-
earisation of these equations is discussed in [22].
4. Analysis and examples
The formulation outlined in Section 3 has been analysed, and tested for various problems and element
types. Elements are dierentiated on the basis of the interpolation order for u
h
(which is always C
0
contin-
uous), the interpolation order for e
h
, and the continuity of e
h
. For example, an element with cubic shape
functions for u
h
and C
0
continuous, quadratic shape functions for e
h
is denoted P
3
/P
2
(C
0
). An element with
quadratic shape functions for u
h
and discontinuous, linear shape functions for e
h
is denoted P
2
/P
1
(C
1
).
For one-dimensional examples, e
eq
= e.
1484 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
4.1. Convergence for the elastic case in one dimension
For an elastic problem, there exists a one-way coupling between the two weak Eqs. (16) and (17). In
this section, the approximation of e, for given a solution to Eq. (16), is examined. For an elastic bar,
the fundamental problem is a second-order dierential equation, and the second weak equation provides
a projection of the solution to the equilibrium equation (and its relevant derivatives) onto the basis
for e.
4.1.1. Error analysis for the mixed strain eld
Here, an analysis for the L
2
-error in e
h
, ke
h
e c
2
e
;xx
k, where e is the exact strain eld, is performed. In
order to compare with numerical solutions, we have solved an elastic problem with a forcing function cho-
sen such that e has a localised character to it. The analysis rests upon the crucial result that the one-dimen-
sional nite element displacement eld, u
h
, is nodally-exact when the shape functions are derived from
Lagrange polynomials [20].
Consider the interpolation combinations P
k+1
/P
k
(C
0
) and P
k+1
/P
k
(C
1
). Let the nodal values of e
h
cor-
responding to element e be denoted by d
k(e1)+1
, . . . , d
ke+1
for the C
0
case, and by d
(k+1)(e1)+1
, . . . , d
(k+1)e
for the C
1
case. The values of u
h
at the nodes corresponding to this element are d
(k+1)(e1)+1
, . . . , d
(k+1)e+1
.
The nodal displacements coincide with the exact solution. The discrete matrix form of Eq. (17) reveals the
stencil relating the d- and d-values. In the C
0
case, {d
k(e1)+1
, . . . , d
ke+1
} depend on {d
(k+1)(e3)+1
, . . . ,
d
(k+1)(e+2)+1
}, and in the C
1
case, {d
(k+1)(e1)+1
, . . . , d
(k+1)e
} depend on {d
(k+1)(e3)+1
, . . . , d
(k+1)(e+2)+1
};
i.e., on 5k + 6 d-values.
Consider a nodally-exact, 5(k + 1)th-order polynomial interpolant, u
p
, of the exact solution, u.
Fig. 1a shows a patch of ve elements, nodes for the d- and d-degrees of freedom, u and u
p
for the P
2
/
P
1
(C
0
) interpolation combination and Fig. 1b depicts the situation for the P
2
/P
1
(C
1
) interpolation
combination.
Let e
p
denote the strain eld arising from u
p
. Using the triangle inequality the error can be bounded as
follows:
ke
h
ek
X
6 ke
h
e
p
c
2
e
p
;xx
k
X
ke
p
c
2
e
p
;xx
ek
X
; 21
where e e c
2
e
;xx
. Since u
p
is of order 5(k + 1), and is nodally-exact over X, nite element interpolation
theory [15] leads to the result
ke
p
c
2
e
p
;xx
ek
X
6 Ch
5k12
; 22
where C is a constant, independent of h.
The rst term on the right-hand side in (21) is estimated using the stencil for each interpolation combi-
nation. The 5(k + 1)th-order polynomial interpolant, u
p
(x), can be written as
u
p
x

5k1
n0
A
n
x e
1
2
_ _
h
_ _
n
; 23
where A
n
are coecients and (e + 1/2)h is the midpoint of X
e
with the rst node of the mesh at x = 0. The
nodal exactness of u
h
allows {d
k(e3)+1
, . . . , d
k(e+2)+1
} to be written in terms of A
n
, n = 0, . . . , 5(k + 1). On
combining with the stencil, {d
k(e1)+1
, . . . , d
ke+1
} (for the C
0
case) and {d
(k+1)(e1)+1
, . . . , d
(k+1)e
} (for the
C
1
case) can therefore be expressed in terms of A
n
. Using the interpolation functions for e
h
, it is then a
trivial matter to exactly evaluate ke
h
e
p
c
2
e
p
;xx
k
X
e
.
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1485
For convenience, ke
h
e
p
c
2
e
p
;xx
k
2
X
e
was evaluated for each interpolation combination. The results are:
P
3
/P
2
(C
0
)
ke
h
e
p
c
2
e
p
;xx
k
2
X
e

9
500
A
2
3
2A
3
A
5
c
2

500
9
A
2
5
c
4
_ _
h
5
Oh
7
; 24
P
2
/P
1
(C
0
)
ke
h
e
p
c
2
e
p
;xx
k
2
X
e

21
20
A
2
3
22A
3
A
5
c
2
120A
2
5
c
4
_ _
h
5
Oh
7
; 25
P
3
/P
2
(C
1
)
ke
h
e
p
c
2
e
p
;xx
k
2
X
e

1936
27

704
27
a
64
27
a
2
_ _
A
2
4
c
4
h
3
Oh
5
; 26
P
2
/P
1
(C
1
)
ke
h
e
p
c
2
e
p
;xx
k
2
X
e

4
27
A
2
2

4
3
A
2
A
4
c
2
3A
2
4
c
4
_ _
h
3

4
3
A
2
A
4
c
2
6A
2
4
c
4
_ _
ah
3
3A
2
4
c
4
a
2
h
3
; and
27
P
1
/P
0
(C
1
)
ke
h
e
p
c
2
e
p
;xx
k
2
X
e
36 1 2a a
2
_ _
A
2
3
c
4
h Oh
3
. 28
For each case listed above, the higher-order terms in h have a nite maximum power, O(h
l
), and
h 6 m(X), where m(X) is a measure of the length of the total domain. Therefore it follows that there exist
constants

C
1
,

C
2
,

C
3
,

C
31
,

C
4
,

C
5
,

C
51
, such that for the P
3
/P
2
(C
0
) element
(a)
(b)
Fig. 1. Element patch for error analysis of (a) P
2
/P
1
(C
0
) and (b) P
2
/P
1
(C
1
) formulations.
1486 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
ke
h
e
p
c
2
e
p
;xx
k
2
X
e
6

C
1
h
5
; 29
for the P
2
/P
1
(C
0
) element
ke
h
e
p
c
2
e
p
;xx
k
2
X
e
6

C
2
h
5
; 30
for the P
3
/P
2
(C
1
) element
ke
h
e
p
c
2
e
p
;xx
k
2
X
e
6

C
3
h
5
if a 5:5;

C
31
h
3
otherwise;
_
31
for the P
2
/P
1
(C
1
) element
ke
h
e
p
c
2
e
p
;xx
k
2
X
e
6

C
4
h
3
8a; 32
and for the P
1
/P
0
(C
1
) element,
ke
h
e
p
c
2
e
p
;xx
k
2
X
e
6

C
5
h
3
if a 1;

C
51
h otherwise;
_
33
where the constants

C
1
; . . . ;

C
51
are independent of h (if u is suciently regular) and element number e
(since the constants can be chosen to be the maximum over all elements). For elements of a uniform size,
in each mesh there are n
el
= m(X)/h elements. Therefore,
ke
h
e
p
c
2
e
p
;xx
k
2
X
6
mX
h
max
e
ke
h
e
p
c
2
e
p
;xx
k
2
X
e
; 34
leading to, for the P
3
/P
2
(C
0
) element
ke
h
e
p
c
2
e
p
;xx
k
2
X
6 mX

C
1
h
4
; 35
for the P
2
/P
1
(C
0
) element
ke
h
e
p
c
2
e
p
;xx
k
2
X
6 mX

C
2
h
4
; 36
for the P
3
/P
2
(C
1
) element
ke
h
e
p
c
2
e
p
;xx
k
2
X
6
mX

C
3
h
4
if a 5:5;
mX

C
31
h
2
otherwise;
_
37
for the P
2
/P
1
(C
1
) element
ke
h
e
p
c
2
e
p
;xx
k
2
X
6 mX

C
4
h
2
8a; 38
and for the P
1
/P
0
(C
1
) element
ke
h
e
p
c
2
e
p
;xx
k
2
X
6
mX

C
5
h
2
if a 1;
mX

C
51
h
0
otherwise.
_
39
Since k P0 in (22), each of Eqs. (35)(39) can be combined with (21) and (22), and the convergence rate in
terms of ke
h
ek
X
can be determined. The results are summarised in Table 1.
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1487
4.1.2. Observed convergence rates
The convergence rate estimates from the previous section are now compared with observed rates. To do
this, the convergence of e is examined for the quadratically tapering bar shown in Fig. 2. Considering that
the origin is located as the centre of the bar, the cross-section A is given by
Ax A
1
c
2
A
1
x
2
L
2
; 40
where c controls the ratio between A
1
and A
2
(c
2
= (A
2
A
1
)/A
1
). This problem can be equivalently formu-
lated as a rod with unit cross-section and loading f which varies along the rod, which allows the conver-
gence analysis from the previous section to be carried over for this case. The exact solution for e along
the bar is
e
PL
2
EA
1
L
2
c
2
x
2
_ _ 41
and the exact solution of e
,xx
is
e
;xx

2PL
2
c
2
3c
2
x
2
L
2
_ _
EA
1
L
2
c
2
x
2
_ _
3
. 42
The term e
h
involves both e and e
,xx
. The e component involves the standard L
2
projection of e
h
(coming
from the solution of the equilibrium equation) onto the basis for e
h
. This projection does not involve
any of the element interface terms that have arisen in the discontinuous Galerkin formulation. Of special
interest is the approximation of the e
,xx
term. To examine numerically the convergence of this term, the
weak form corresponding to
e e
;xx
0 43
Table 1
Analytical convergence rates in terms of ke
h
ek
X
for various elements
Element type Convergence rate
P
1
/P
0
(C
1
) (a = 1) O(h
1
)
P
1
/P
0
(C
1
) (a 5 1) O(h
0
)
P
2
/P
1
(C
0
) O(h
2
)
P
2
/P
1
(C
1
) O(h
1
)
P
3
/P
2
(C
0
) O(h
2
)
P
3
/P
2
(C
1
) (a = 5.5) O(h
2
)
P
3
/P
2
(C
1
) (a 5 5.5) O(h
1
)
Fig. 2. Quadratically tapering bar.
1488 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
is solved, as it is the e
,xx
term that dominates the convergence rate. However, in practice, the convergence
rate of e may appear to dominate due to a large dierence in the constants in the error inequality. Taking
Eq. (17) and removing terms related to the projection of e
h
, we solve the following problem: nd u
h
2 S
h
and e
h
2 W
h
such that
_
X
w
h
;x
Eu
h
;x
dX w
h
Pj
xL
0 8w
h
2 V
h
; 44
_
X
q
h
e
h
dX
_
~
X
q
h
;x
e
h
;x
dX
_
C
q
h
e
h
;x
ndC
_
~
C
sq
h
the
h
;x
i dC
_
~
C
hq
h
;x
ise
h
tdC
_
~
C
a
h
sq
h
tse
h
tdC 0
8q
h
2 W
h
. 45
For the parameters A
1
= 1 mm
2
, A
2
= 0.1 mm
2
, L = 50 mm, E = 1 MPa and P = 1 N, the convergence
behaviour is examined for a range of elements. The L
2
-norm of the computed error for a range elements
is shown in Fig. 3 for the case a = 1. If the error is computed by integrating over the entire bar
(50 < x < 50), as in Fig. 3, the results are polluted by errors at the boundaries of the bar. In Fig. 3, the
convergence behaviour represents primarily how rapidly the computed solution approaches the exact solu-
tion at the boundaries. However, when simulating a tapered damaging bar, the error at the boundaries of
the bar is of little consequence as damage develops at the centre, and e at the boundaries plays no role (pre-
suming that the deviation from the exact solution is not sucient to induce spurious damage development).
Excluding the error at the end of the bar by integrating over 40 < x < 40, the convergence behaviour is
more predictable, as can be seen in Fig. 4. From Fig. 4, it can be concluded that the convergence rate
for all elements is consistent with the predicted rates (as summarised in Table 1).
The eect of a on convergence for the P
2
/P
1
(C
1
) element is shown in Fig. 5. As expected, the conver-
gence rate is unaected by a. For the P
3
/P
2
(C
1
) element, the convergence behaviour for dierent a is
shown in Fig. 6. From Fig. 6, it is clear the convergence rate increases by one order for a = 6. This is close
to the predicted value of a = 5.5.
4.2. Convergence for the inelastic case in one dimension
The formulation is now examined for the inelastic case. Again, the quadratically tapering bar is consid-
ered, which leads to damage development at the centre of the bar. For a particular relationship between D
Fig. 3. Convergence for the problem e e
;xx
with a = 1 on the domain 50 < x < 50.
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1489
Fig. 4. Convergence for the problem e e
;xx
with a = 1 on the domain 40 < x < 40.
Fig. 5. Convergence of the P
2
/P
1
(C
1
) element for dierent values of a on the domain 40 < x < 40 for the problem e e
;xx
.
Fig. 6. Convergence of the P
3
/P
2
(C
1
) element for dierent values of a on the domain 40 < x < 40 for the problem e e
;xx
.
1490 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
and j, it is possible to solve the problem analytically, which provides the basis for numerical convergence
tests.
4.2.1. Analytical solution
Consider again the bar shown in Fig. 2, where the shaded region indicates the damaged zone and x = w
is the location of the damage-elastic boundary. For convenience, the force P is expressed as
P 1 b
2
_ _
EA
1
j
0
; 46
where b governs the magnitude of the applied load and j
0
is the value of j at which damage is rst induced.
In the undamaged part of the bar, the strain response is given by
e
P
EA
47
and within the damaged zone by
e
P
1 D EA
. 48
With the intention of nding an analytical solution, a simple damage law is assumed,
D
0 if j 6 j
0
;
1
j
0
j
if j > j
0
.
_
49
The relationship in Eq. (49) is the analogy of perfect plasticity for the case c = 0, in the sense that it yields a
plateau in the loaddisplacement response once the elastic limit has been exceeded. Assuming that no
unloading takes place in the damaged zone, j e, that is
j e c
2
e
;xx
; w < x < w. 50
Inserting Eqs. (50) and (49) into Eq. (48), the following ordinary dierential equation is obtained:
1
P
j
0
EA
_ _
e c
2
P
j
0
EA
e
;xx
0; w < x < w; 51
which governs the response in the damaged zone. The general solution to the above equation for x > 0 is
given by
e C
1
M
b
2
L
4cc

1 b
2
_ ;
1
4
;
cx
2
Lc

1 b
2
_
_
_
_
_
_
_

x
p C
2
W
b
2
L
4cc

1 b
2
_ ;
1
4
;
cx
2
Lc

1 b
2
_
_
_
_
_
_
_

x
p ; 52
where M and W are Whittakers functions [1], and C
1
and C
2
are integration constants. The integration
constants can be obtained by considering boundary conditions at x = w. The rst considered condition
is symmetry about x = 0. Expanding Eq. (52) in a Taylor series about x = 0, and requiring that odd terms
vanish, the following relation is obtained:
C
1
C
2
2p
C
1
4

b
2
L
2
cc

1 b
2
_
4c

1 cb
2
_
_
_
_
_
_
_
; 53
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1491
where C is the Gamma function. The second considered condition is
se
;x
t 0 at x w; 54
which implies
e
;x

P
EA
_ _
;x
at x w. 55
Finally, the location of the elastic-damage boundary can be determined by requiring that
j j
0
at x w. 56
For the sake of brevity, the expressions for C
1
, C
2
and w have been omitted.
4.2.2. Damage convergence results
The convergence behaviour of various elements for the outlined test problem is now examined. For the
tests, the following parameters are adopted: c = 3, b = 5/29, L = 100 mm, E = 200 10
3
MPa, c = 1 mm
and A
1
= 1 mm
2
and j
0
= 1 10
3
. In calculating the error, the Gamma function has been computed
numerically.
The results of the error analysis, for various elements, are shown in Fig. 7. For the P
1
/P
0
(C
1
), P
2
/
P
1
(C
1
) and P
3
/P
2
(C
1
) elements, a = 1, a = 4 and a = 6, respectively. These choices draw upon the ob-
served convergence results for the elastic case. For all the elements, the solution converges. Once a level
of renement has been reached, the convergence rate is constant. Moreover, for a given interpolation order,
continuous and discontinuous interpolations yield similar results.
4.2.3. Non-trivial damage response
The quadratically tapering bar is now examined for a non-trivial softening relationship. For the tapered
bar the relevant parameters are: c = 3, A
1
= 1 mm
2
, Youngs modulus E = 20 10
3
MPa, j
0
= 1 10
4
,
j
c
= 0.0125, and c = 1 mm. The functional form of the damage variable is specied to be
D
0 if j 6 j
0
;
1
j
0
j
c
j
jj
c
j
0

if j
0
< j < j
c
;
1 if j Pj
c
;
_

_
57
Fig. 7. Error in e
h
for the damaging bar.
1492 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
which leads to a linear softening relationship for c = 0. The numerical performance of the formulation has
previously been demonstrated in Wells et al., [22] for continuous, piecewise linear u
h
and constant e
h
, dis-
continuous across element boundaries (the P
1
/P
0
(C
1
) element). The formulation was shown numerically
to converge to a benchmark solution. Here, the formulation is extended for a range of dierent element
types.
The motivation behind the considered strain gradient-dependent model is regularisation in the presence
of strain softening. Without strain gradient eects (c = 0), computed results are pathologically mesh-depen-
dent; the result of which is manifest in the loaddisplacement responses. Therefore, each of the elements
which to this point have been examined are tested, and the loaddisplacement responses reported for
meshes with 20, 40, 80, 160 and 320 elements. For the P
3
/P
2
(C
1
) element, the numerical tests are per-
formed with meshes of 20, 40, 80 and 160 elements. To provide a reference solution, the response computed
using 160 P
3
/P
2
(C
0
) elements is included in all gures.
The loaddisplacement responses for the two elements using a continuous interpolation of e
h
are shown
in Fig. 8. As the mesh is rened, the computed response for both element types converges towards to the
reference solution. The loaddisplacement responses for three elements using a discontinuous interpolation
of e
h
are shown in Fig. 9. Again, for the P
1
/P
0
(C
1
) element, a = 1, for the P
2
/P
1
(C
1
) element, a = 4, and
for the P
3
/P
2
(C
1
) element, a = 6. It is clear, for all elements, that the loaddisplacement response
(a)
(b)
Fig. 8. Loaddisplacement response for (a) P
2
/P
1
(C
0
) and (b) P
3
/P
2
(C
0
) elements.
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1493
converges to the reference solution with mesh renement. To further examine the computed results, the
damage proles for the two continuous elements and the three discontinuous elements are compared in
Figs. 10 and 11, respectively. For all cases, the 160 element mesh is considered. Clearly, the damage proles
are nearly identical for all element types.
(a)
(b)
(c)
Fig. 9. Loaddisplacement response for (a) P
1
/P
0
(C
1
), (b) P
2
/P
1
(C
1
) and (c) P
3
/P
2
(C
1
) elements.
1494 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
From the numerical results, it can not be stated conclusively that the continuous formulation is superior
to the discontinuous formulation, or vice versa. From the stand-point of implementation and eciency, the
continuous formulation has an advantage. However, situations can be foreseen in which a discontinuous
formulation, at least in parts of the domain, would be more faithful to the mechanical reality.
4.3. Three-point bending test
To conclude the numerical validations, a three-point bending test is performed using a P
2
/P
1
(C
0
) ele-
ment. The element is triangular, with degrees of freedom for u
h
located at the vertexes and at the mid-sides,
and degrees of freedom for e
h
located only at the vertexes of the element. The choice of a quadratic inter-
polation of u
h
means that the penalty term for imposing the non-standard boundary condition in Eq. (16)
vanishes, and the non-standard boundary condition is satised by construction (see Eq. (16)). The equiv-
alent strain is taken as the trace of the strain tensor,
e
eq
tracee. 58
Fig. 10. Damage proles for the e
h
-continuous elements, computed using 160 elements.
Fig. 11. Damage proles for the e
h
-discontinuous elements, computed using 160 elements.
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1495
This choice does not reect a strong physical motivation, rather it is chosen for illustrative purposes as it
allows for relatively simple linearisation of the method (which can become extremely complex when c 5 0).
The three-point bending test is performed for two dierent meshes with c 5 0 and c = 0. The adopted
geometry for the three-point bending test is shown in Fig. 12, and the adopted material parameters are:
Fig. 14. Loaddisplacement responses for two dierent meshes.
(a) (b)
Fig. 13. Damage contours for two meshes (a) with gradient eects (c 5 0) and (b) without gradient eects (c = 0).
Fig. 12. Three-point bending specimen.
1496 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498
Youngs modulus E = 20 10
4
MPa, Poissons ratio m = 0, j
0
= 1 10
4
, and j
c
= 1.25 10
2
. For gradi-
ent-dependent simulations, c = 8 10
2
mm. Computations are stopped when damage reaches unity at any
point in the mesh, and damage development at the supports is prevented. The computed damage contours
for the four cases are shown in Fig. 13. From the damage contours, it is clear that the computed results are
similar for the two meshes with c 5 0. In the absence of regularising eects (c = 0), the result is clearly af-
fected by the discretisation. The loaddisplacement responses for the various cases are shown in Fig. 14.
Recall that a computation is halted when damage reaches unity at a material point. For the gradient-depen-
dent case, the responses for the two meshes are similar. For the case c = 0, the responses are also similar,
which is somewhat in contrast to what is normally expected for a strain softening problem. The responses
are similar in this case due to the spurious development of two cracks for the ner mesh (in contrast to the
single main crack for the coarse mesh). This is evident from the damage contours in Fig. 13, and is itself a
manifestation of pathological results with the standard formulation (c = 0).
5. Conclusions
A discontinuous Galerkin formulation for a strain gradient-dependent damage model has been investi-
gated for a range of dierent nite elements. The model allows the numerical solution of a continuum prob-
lem which would classically require C
1
interpolations with a simple C
0
or even discontinuous basis.
Examples demonstrate robust performance for a range of polynomial orders and degrees of continuity
of the interpolation functions, and are supported by rigorous error analysis. Specically, lower-order inter-
polations perform well and are relatively simple to construct. The convergence properties of the proposed
method have been examined for the elastic case, for which the observed rates are consistent with the theo-
retically predicted rates. The formulation has been observed numerically to converge also for damage prob-
lems. Finally, the formulation was applied successfully to a two-dimensional problem. While the approach
is promising, several issues remain. Diculties which must be resolved for other gradient models include the
eective imposition of boundary conditions on the xed boundary of a body, and at moving boundaries
internal to a body. The development of thermodynamically consistent models would assist in this sense,
as the higher-order kinematic gradients have a natural partner in the energetic sense.
Acknowledgements
LM and FU acknowledge the support of University of Bologna, GNW acknowledges the support of the
Netherlands Technology Foundation (STW), and KG acknowledges support from the US National Science
Foundation by way of Grant No. CMS0087019, and from Sandia National Laboratory. The support from
Sandia National Laboratory includes a Presidential Early Career Award.
References
[1] M. Abramowitz, I.A. Stegun (Eds.), Handbook of Mathematical Functions, Dover Publications, Inc., New York, 1965.
[2] E.C. Aifantis, On the microstructural origin of certain inelastic models, J. Engrg. Mater. Technol. 106 (1984) 326334.
[3] D.N. Arnold, F. Brezzi, B. Cockburn, D. Marini, Unied analysis of discontinuous Galerkin methods for elliptic problems, SIAM
J. Numer. Anal. 39 (5) (2002) 17491779.
[4] H. Askes, J. Pamin, R. de Borst, Dispersion analysis and element-free Galerkin solutions of second- and fourth-order gradient-
enhanced damage models, Int. J. Numer. Methods Engrg. 69 (6) (2000) 811832.
[5] Z.P. Bazant, Mechanics of distributed cracking, Appl. Mech. Rev. 39 (1986) 675705.
[6] Z.P. Bazant, B. Oh, Crack band theory for fracture of concrete, RILEM Mater. Struct. 16 (93) (1983) 155177.
L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498 1497
[7] Z.P. Bazant, G. Pijaudier-Cabot, Nonlocal continuum damage, localization instability and convergence, J. Appl. Mech. 55 (1988)
287293.
[8] G. Borino, B. Failla, F. Parrinello, A symmetric nonlocal damage theory, Int. J. Solids Struct. 40 (2003) 36213645.
[9] B.D. Coleman, M.L. Hodgdon, On shear bands in ductile materials, Arch. Ration. Mech. Anal. 90 (1985) 219247.
[10] R. de Borst, L.J. Sluys, Localisation in a Cosserat continuum under static and dynamic loading conditions, Comput. Methods
Appl. Mech. Engrg. 90 (1991) 805827.
[11] J.H.P. de Vree, W.A.M. Brekelmans, M.A.J. van Gils, Comparison of non-local approaches in continuum damage mechanics,
Comput. Struct. 55 (4) (1995) 581588.
[12] G. Engel, K. Garikipati, T.J.R. Hughes, M.G. Larson, L. Mazzei, R.L. Taylor, Continuous/discontinuous nite element
approximations of fourth-order elliptic problems in structural and continuum mechanics with applications to thin beams and
plates, and strain gradient elasticity, Comput. Methods Appl. Mech. Engrg. 191 (34) (2002) 36693750.
[13] L.M. Kachanov, On creep rupture time, IZV Akad Nauk SSSR Otd. Tech. Nauk 8 (1958) 2631.
[14] D. Lasry, T. Belytschko, Localization limiters in transient problem, Int. J. Solids Struct. 24 (6) (1988) 581597.
[15] J.T. Oden, G.F. Carey, Finite Elements: Mathematical Aspects, vol. IV, Prentice-Hall, Englewood Clis, NJ, 1984.
[16] R.H.J. Peerlings, R. de Borst, A.M. Brekelmans, J.H.P. de Vree, Gradient enhanced damage for quasi-brittle materials,
Int. J. Numer. Methods Engrg. 39 (1996) 33913403.
[17] Y.R. Rashid, Ultimate strength analysis of prestressed concrete pressure vessels, Nucl. Engrg. Des. 7 (1968) 334344.
[18] J.R. Rice, The localization of plastic deformation, in: W.T. Koiter (Ed.), Theor. Appl. Mech., North Holland Publishing Co.,
1976, pp. 207220.
[19] J.Y. Shu, W.E. King, N.A. Fleck, Finite elements for materials with strain gradient eects, Int. J. Numer. Methods Engrg.
44 (1999) 373391.
[20] G. Strang, G.J. Fix, An Analysis of the Finite Element Method, Prentice-Hall, New Jersey, 1973.
[21] N. Triantafyllidis, E.C. Aifantis, A gradient approach to localization of deformation: I Hyperelastic materials, J. Elast. 16 (1986)
225237.
[22] G.N. Wells, K. Garikipati, L. Molari, A discontinuous Galerkin formulation for a strain gradient-dependent continuum model,
Comput. Methods Appl. Mech. Engrg. 193 (3335) (2004) 36333645.
[23] K. Willam, Experimental and computational aspects of concrete fracture, in: R. Owen, E. Hinton, Bicanic (Eds.), Proc. Int. Conf.
Comp. Aided Anal. and Design of Concrete Structures, Pineridge Press, Swansea, UK, 1984, pp. 3370.
[24] A. Zervos, P. Papanastasiou, I. Vardoulakis, A nite element displacement formulation for gradient elastoplasticity, Int. J.
Numer. Methods Engrg. 50 (2001) 13691388.
1498 L. Molari et al. / Comput. Methods Appl. Mech. Engrg. 195 (2006) 14801498