Comput. Methods Appl. Mech. Engrg.

237–240 (2012) 244–259

Contents lists available at SciVerse ScienceDirect

Comput. Methods Appl. Mech. Engrg.

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

A finite element approach coupling a continuous gradient damage model and

a cohesive zone model within the framework of quasi-brittle failure
Sam Cuvilliez a,b,⇑, Frédéric Feyel c, Eric Lorentz a,b, Sylvie Michel-Ponnelle a,b
a
Laboratoire de Mécanique des Structures Industrielles Durables, UMR EDF-CNRS-CEA 2832, 1 av. Général de Gaulle, 92141 Clamart cedex, France
b
EDF R&D, 1 av. Général de Gaulle, 92141 Clamart cedex, France
c
ONERA, BP72 – 29 avenue de la Division Leclerc, 92322 Châtillon cedex, France

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

Article history: A coupled continuous/discontinuous approach involving a gradient damage model and a cohesive zone
Received 12 December 2011 model is proposed to simulate crack propagation in quasi-brittle materials. First, a semi analytical one-
Received in revised form 24 April 2012 dimensional study is carried out on a bar submitted to tensile loading (in the case of inhomogeneous
Accepted 26 April 2012
symmetrical damage localisation) in order to identify a set of cohesive laws allowing to switch from dam-
Available online 16 May 2012
age growth to cohesive crack opening. This continuous/discontinuous approach is constructed so that
energetic equivalence between both models remains ensured whatever the damage level reached when
Keywords:
switching. This strategy is then extended to the finite element framework, in the case of 2D (and 3D)
Gradient damage model
Cohesive zone model
crack propagation under mode I loading conditions. Both non local and cohesive models are combined
Energetic equivalence using a critical damage criterion which is checked at the end of each converged time step.
Finite element method Ó 2012 Elsevier B.V. All rights reserved.

1. Introduction – the ‘‘integral’’ non local model [27] and the ‘‘implicit gradient’’
model [26] (also called differential non local model) which both
The prediction of failure mechanisms in large quasi-brittle rely on the regularisation of state variables;
structures, such as encountered in civil engineering, still remains – the introduction of higher order gradients of the displacement
a challenge from a computational point of view as several model- field [7];
ling steps are involved, namely: damage inception in the fracture – the introduction of gradient of internal variables [10,1].
process zone, propagation of traction free macro-cracks and crack
path tracking. Finite element (FE) approaches dedicated to failure In the present work we will focus on the latter, more precisely
analysis can be classified in two main categories, which are unable on the gradient damage model proposed in [20] because it requires
to properly treat all these steps: a continuous one, named the con- a reduced number of additional unknowns compared to higher or-
tinuum damage mechanics (CDM) [18], where fracture is consid- der continuum models, and because the spurious growth of the
ered as the ultimate stage of damage accumulation in a fully damaged band met with both integral and differential non lo-
continuum; a discontinuous one, where displacement discontinu- cal models is avoided. However, the inability of regularised dam-
ity surfaces propagate inside an elastic media, ranging from Grif- age models to describe cracks as displacement discontinuity
fith’s theory to cohesive zone models (CZM) [2]. surfaces remains inherent to their continuous formulation, which
An important difficulty met with the continuous description is may be an issue for instance when dealing with problems involving
that the strain localisation occurring in its local version leads to potential fluid leakage through cracks, or contact subsequent to
an ill-posed problem, with severe numerical consequences such possible crack closure.
as a strong mesh dependence. The well-posedness can be retrieved On the other hand, the discontinuous approach, within the lin-
by means of some regularisation technique. Among those which ear elastic fracture mechanics (LEFM) framework, is able to deal
consist in adding some non locality in the constitutive equations with stable crack propagation according to the Griffith criterion,
(related to an internal length) we can recall: and can also encompass crack initiation with the enhancements
provided by cohesive zone models. Unfortunately, its straightfor-
ward application to the standard FE method is restricted to prede-
fined crack paths because potential discontinuities surfaces must
⇑ Corresponding author at: Laboratoire de Mécanique des Structures Industrielles conform to the mesh elements boundaries. Several numerical tech-
Durables UMR EDF-CNRS-CEA 2832, 1 av. Général de Gaulle, 92141 Clamart cedex, niques have been developed to circumvent this limitation: reme-
France. Tel.: +33 1 47 65 52 03; fax: +33 1 47 65 41 18.
shing strategies, use of a discontinuous interpolation of the
E-mail address: sam.cuvilliez@edf.fr (S. Cuvilliez).

0045-7825/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.cma.2012.04.019
 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259 245

displacement field, at a global level as in the extended finite ele- bringing an additional level of validation: simple crack opening,
ment method (X-FEM) [3,25] or at a local one with embedded dis- stable crack propagation, unstable propagation of two cracks, and
continuity finite elements (E-FEM) [15,16]. However, the 3D plane crack propagation. This section mainly aims at validating
treatment of the crack path tracking has still to rely on external that, regardless of the critical damage value, the FE extension re-
physical criteria and remains a difficult numerical operation. mains consistent with both CDM and CZM stand-alone approaches,
Consequently, combining both continuous and discontinuous and that it does not generate additional CPU cost compared to the
descriptions in the same FE approach seems an appealing tech- non local model. Finally, some concluding remarks in Section 7
nique as the whole failure process can be described in a single FE close the paper. FE computations have been conducted with Co-
analysis. The transition from regularised damage models to discon- de_Aster, [30] an open source FE software.
tinuous crack models has been actively studied during the last fif-
teen years; the following review is far from exhaustive and only 2. Gradient damage model
aims to recall the main contributions. In [23] the concept of LEFM
equivalent crack is derived from an integral non local damage The behaviour of the bulk material is described by a brittle iso-
model by enforcing the equality between increments of dissipated tropic gradient damage model as proposed in [20]. In this section
energy related to both models. In [17] a transition from non local we first briefly present the main features of the model; its
damage to embedded cohesive cracks (E-FEM) is proposed in order closed-form solution in a one dimensional setting is then recalled.
to provide a crack orientation criterion determined from the direc-
tion of the maximum principal non local strain. In [9] the transition 2.1. Constitutive law
from non local damage to cohesive zone model is performed (un-
der mode I loading conditions) by means of the X-FEM technique. The material state is defined by the strain tensor e and a scalar
Discontinuities are introduced when the damage exceeds a critical damage variable a. The stress–strain relation is brittle elastic,
value related to the mesh element size within the damage band- where the damage progressively weakens the initial stiffness:
width, and the corresponding cohesive law is endowed with the to-
tal surface energy which corresponds to the energy not yet
r ¼ AðaÞE : e ð2-1Þ
dissipated by the non local model within the band. In [6], a strategy The stress is denoted r, E is Hooke’s tensor, and 0 6 A(a) 6 1 is the
allowing to identify a cohesive law that is equivalent to a differen- stiffness function. In particular, a = 0 corresponds to the sound
tial non local model is developed in a one-dimensional setting. The material and a = 1 corresponds to the ultimate damage state:
cohesive crack then plays the role of a localisation limiter as it is A(0) = 1, A(1) = 0. The stress–strain relation is linear, which implies
coupled with a local damage model and becomes active in the that the model does not take into account crack closure effects. The
localisation regime (after a diffuse damage stage). In the references damage driving force derives from the elastic energy which reads:
mentioned above, a portion of the fracture energy is transferred
1
from the continuous to the discontinuous model because the tran- Y ¼  A0 ðaÞe : E : e ð2-2Þ
sition can be triggered before the ultimate damage stage. In several 2
works this feature is circumvented by introducing traction free dis- In order to control the localisation of strain and damage, the cou-
continuities when damage almost reaches the unity as in [24,29]. It pling between neighbour material points is modelled through the
also allows to avoid the spurious growth of the damage band as introduction of the damage Laplacian r2a into the yield function:
differential non local damage models are involved in these
f ðY; r2 aÞ ¼ Y þ cr2 a  k ð2-3Þ
contributions.
In this paper, a coupled continuous – discontinuous approach is where k > 0 is a yield threshold and c > 0 is a parameter which gov-
developed in order to switch from a continuous gradient damage erns the intensity of the coupling, hence the strength of the non lo-
model to a discrete failure model. In order to keep options open cal effects. The Kuhn–Tucker consistency condition takes its usual
(transition at the early or late stages of damage) it is constructed form:
so that the switch can occur at any damage level, which raises
the issue of energy conservation between both models. The choice f 6 0; a_ P 0; _ ¼0
af ð2-4Þ
made is to describe the discontinuous part through a cohesive zone
where the dot denotes time differentiation. Finally, additional
model in order to permit an energy transfer from the non local
boundary conditions have to be postulated. They include a regular-
model. The paper is organised as follows. In Section 2, the main
ity condition on the damage fields and natural boundary conditions.
features of the gradient damage model and its closed-form solution
If @ X denotes the boundary of the body domain X, n its outer nor-
for a one-dimensional problem are recalled. Cohesive zone models
mal, C a potential surface of discontinuity, m its normal and st the
are then briefly reviewed in Section 3. In Section 4, the coupled ap-
discontinuity of a quantity across C, then the interface and bound-
proach is first established in a one-dimensional setting on the basis
ary conditions read:
of the non local closed-form solution. A traction-separation law
parameterized by the critical damage value beyond which the sat ¼ 0 and scratm ¼ 0 across C; ð2-5Þ
transition is triggered is identified in order to preserve a strict
equivalence with the non local model. As 2D analytical solutions
ra  n ¼ 0 on @ X: ð2-6Þ
seem out of reach, the extension of the approach to crack propaga- The model belongs to the class of generalised standard materi-
tion has to be done in the FE framework: the implementation is als, extended to gradient enhanced constitutive laws. The corre-
then detailed in Section 5. In order to asses the feasibility of the ap- sponding Helmoltz free energy and dissipation potential
proach, the study is limited to mode I crack propagation so as to respectively read (with IRþ the indicator function enforcing
permit a straightforward extension of the one dimensional results. a_ P 0Þ :
Moreover, the questions of the detection of the crack path from the Z Z Z
damage distribution, and the gradual insertion of the cohesive ele- c
Fðe; aÞ ¼ AðaÞ/el ðeÞdX þ ðraÞ2 dX; _ ¼
DðaÞ _ X;
wðaÞd
ments with remeshing strategies are not tackled: straight crack X 2 X X

propagations are considered so as to place the cohesive elements 1 0 if a_ P 0;
with /el ðeÞ ¼ e : E : e; _ ¼ ka_ þ IRþ ðaÞ;
wðaÞ _ _ ¼
IRþ ðaÞ
into the discretisation prior to the computation. Numerical results 2 þ1 if a_ < 0:
are then presented in Section 6 trough four test cases, each of them ð2-7Þ
246 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259

2
2.2. Closed-form solution in a one-dimensional setting d au
c 2
¼ k; ð2-14Þ
dx
Starting from the gradient damage model, let us recall the main
results obtained in [20], where a closed-form solution is developed with au the ultimate damage profile. The continuity condition (Eq.
for the inhomogeneous problem of a brittle one-dimensional bar of (2-10)) on the damage derivative at point x = 0 is not fulfilled any-
length 2L, submitted to a tensile displacement at both ends. The more, and au is parabolic:
damage is assumed to localise, regardless of stability conditions rffiffiffiffiffiffi
 x 2 2c
[4], and the bar is taken sufficiently large so that its boundaries au ðxÞ ¼ 1  with D ¼ ; ð2-15Þ
D k
do not interfere with the damage process zone.
A symmetrical solution for the localisation band can be con- with 2D = 2b(1) the ultimate band width.
structed, without loss of generality. The symmetry point is chosen
at the center of the bar (which is then defined as the origin axis 2.3. Parameters of the model
x = 0), so that the studied domain X can be reduced to X = ]0, L[.
For a strictly increasing load level (no unloading), and once the The uniaxial parameters of the model can be split in two cate-
homogeneous elastic regime has been overstepped, one can para- gories. On the one hand observable macroscopic ones, which are
meterise the prescribed displacement by a0, the current top dam- the young modulus E and the peak stress ry. On the other hand
age value at the center of the localisation band (x = 0). Damage internal parameters, such as the dissipation potential slope k, the
growth occurs in a (half) localisation band [0, b], whose boundary softening parameter c, and the regularisation weight c. It as been
x = b is a function of the load parameter a0. Thanks to the consis- shown that the internal parameters are functions of ry and E,
tency condition (Eq. (2-4)), this reads: and two additional observable parameters which are the ultimate
 localisation band width 2D (Eq. (2-15)) and the fracture energy
f ¼ 0 and a_ P 0 in 0; b½;
ð2-8Þ Gf. The fracture energy can be interpreted as the energy to supply
f 6 0 and a_ ¼ 0 in b; L½: in order to break the bar. It is the sum of the stored energy
For a strictly increasing bandwidth, and assuming that each point of (through the gradient of damage) and the dissipated energy (Eq.
the localisation band is subjected to a continuous loading i.e. a_ > 0, (2-7)) which correspond to the ultimate damage distribution (Eq.
the irreversibility condition is automatically fulfilled in [0, b] and (2-15)). A calculation leads to:
the differential system (Eq. (2-8)) can be integrated in an easier 4
way. In that case, the damage field remains zero in the elastic part Gf ¼ kD: ð2-16Þ
3
[b, L] while the damaged zone [0, b] propagates in X according to the
Finally, the yield function (Eq. (2-3)) provides the missing expres-
equation f = 0. It has been shown in [20] that for specific expressions
sion that involves the softening parameter c. Indeed, when the
of the stiffness function A(a), the gradient damage model provides a
damage threshold is first reached at the end of the elastic regime,
strictly increasing bandwidth. From now on, we choose one of these
all the fields remain homogeneous (r = ry, e = ry/E and a = 0) and
expressions:
the condition f = 0 reads, according to the expression of the stiffness
 2
1a function (Eq. (2-9)):
AðaÞ ¼ ; c > 0: ð2-9Þ sffiffiffiffiffiffiffiffiffiffiffiffi
1 þ ca
0 r2y kE
The last step consists in solving the second order non linear ODE A ð0Þ  k ¼ 0 ) ry ¼ : ð2-17Þ
2E 1þc
related to the yield function (Eq. (2-3)) in the interval ]0, b[ taking
into account the symmetry conditions at point x = 0 and the inter- The relations (Eqs. 2-15–2-17) can be inverted in order to express
face conditions (Eq. (2-5)) at point x = b: internal parameters as function of the observable ones:
(
aða0 ; bÞ ¼ 0; da ða0 ; bÞ ¼ 0; 3 Gf 3 3 EGf
f ¼ 0 in 0; b½; with the conditions dx k¼ ; c¼ Gf D; c¼  1: ð2-18Þ
aða0 ; 0Þ ¼ a0 ; da ða0 ; 0Þ ¼ 0;
4 D 8 4 r2y D
dx

ð2-10Þ The condition c > 0 in (Eq. (2-9)) leads to the following restriction
on the observable parameters:
where a(a0, x) is the damage field for a given load level a0. Equation
f = 0 can then be parameterised either by a0 or by the stress r 3 EGf
D< : ð2-19Þ
(which is homogeneous according to the equilibrium equation): 4 r2y
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2
d a A0 ðaÞ r2 2Eka0 Thereafter, material parameters values representative of concrete
c 2
  k ¼ 0; with r¼ : ð2-11Þ (tensile damage) are chosen (Table 2-1). Fig. 2-1(a) and (b) show
dx A2 ðaÞ 2E A1 ða0 Þ  1
respectively for this set of values the damage profiles across the
Some calculation leads to the following implicit relation, where the localisation band and the global stress vs. prescribed displacement
abscissa ^
xða0 ; aÞ is the inverse function of the damage field a(a0, x): response, for a bar of length 2L = 200 mm. One can notice on Fig. 2-
Z a0 1(a) that the hypothesis made in Section 2.2 of no elastic unloading
^xða0 ; aÞ ¼ Gða0 ; sÞ1=2 ds;
with Gða0 ; aÞ within the band ða_ > 0Þ is confirmed. However, this does not lead to
a a spurious widening of the fully damaged band as high values of a_
!
2ka0 a AðaÞ1  1
¼  : ð2-12Þ
c a0 Aða0 Þ1  1
Table 2-1
Material parameters representative of concrete.
In particular, the increasing width of the localisation band is given
by: Observable parameters Internal parameter
E = 3  104 MPa k = 1.5  103 MPa
bða0 Þ ¼ ^xða0 ; 0Þ: ð2-13Þ ry = 3 MPa c=4
Gf = 0.1 N/mm c = 1.875 N
When the ultimate stage a0 = 1 is reached, the bar breaks: the stress
2D = 100 mm
is then zero and (Eq. (2-11)) reduces to:
 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259 247

Fig. 2-1. (a) Damage profile across the localisation band at several load levels; (b) global stress – prescribed displacement response.

are met in a narrowing zone, thus resulting in the sharp profile cor- with mesh refinement, resulting in a strong mesh dependency (see
responding to the ultimate damage field au (Eq. (2-15)). [13] for instance). Consequently, crack path prediction should not
be treated only with cohesive zone models in their numerical appli-
3. Cohesive zone models cations, but has to be postulated a priori (surfaces C no more de-
pend on u).
Cohesive zone models can be considered as an enhancement of
the Griffith theory, especially regarding the inability of the latter to 4. Transition from gradient damage model toward cohesive
deal with crack initiation and to take into account a fracture pro- zone model in a one-dimensional setting
cess zone (whose size is not negligible in the case of quasi-brittle
materials such as concrete) ahead of the crack tip. These models In [19] a convergence analysis of the damage model toward a
rely on the assumption that the process zone can be described cohesive zone model was performed: all the macroscopic parame-
through a fictitious interface along which the displacement field ters remain constant while the length scale goes to zero. Here, we
may admit discontinuities while transferring cohesive forces. Such first want to identify a cohesive law that would be energetically
an approach dates back to Barenblatt’s [2] and Dugdale’s [11] equivalent to the gradient damage model (Fig. 4-1(a)), as it has
works, in which a cohesive process zone was introduced in order been done in [6] for an implicit gradient damage model based on
to remove the stress field singularity near the crack tip predicted strain regularisation. Then, a strategy that enables to switch from
by linear elastic fracture mechanics. Hilleborg et al. [14] enhanced the non local damage model to a cohesive zone model (Fig. 4-
these models by introducing the fracture energy Gf (total surface 1(b)) at an arbitrary load level a0 while preserving the equivalence
energy required to completely separate the interface at a given will be presented.
point) and the critical stress ry (the crack tip stress value from
which the crack is assumed to propagate), that are two parameters
4.1. Identification of an equivalent cohesive law
characterising the cohesive law (relation between the cohesive
stress vector and the displacement jump across the interface).
In this section, we consider the same bar, with an elastic mate-
Cohesive zone models have then been studied in a variational
rial behaviour and a potential displacement discontinuity located
setting by Charlotte et al. [8] and Bourdin et al. [5], where fracture
on its centre. Three of the four macroscopic parameters of the
is described in terms of energy minimisation. If X denotes the do-
non local model should be preserved as they are usual parameters
main body (whose mechanical behaviour is only described in this
of cohesive zone models: the young modulus E to describe the bulk
section by linear elasticity) in which the displacement field u
elasticity, the peak stress ry and the fracture energy Gf. The missing
may admit discontinuities sut across surfaces C(u), the total en-
parameter is the shape of the cohesive law. For given values of
ergy of the structure is then expressed as the sum of the elastic
(ry, Gf) it will only depend on the localisation band width D. Enforc-
strain energy Eel(u), the cohesive energy Eco(sut) and the work of
ing the equivalence then consists in identifying this shape so that
external forces Wext(u):
the cohesive law provides the same stress-displacement global re-
EðuÞ ¼ Eel ðuÞ þ Eco ðsutÞ  W ext ðuÞ sponse as the one obtained by means of the gradient damage
Z Z model.
with Eel ðuÞ ¼ /el ðeðuÞÞdX; Eco ðsutÞ ¼ PðsutÞdC; Let us call UCDM and UCZM the prescribed displacements at the
XnCðuÞ CðuÞ
abscissa x = L, respectively for the non local and the cohesive ap-
ð3-1Þ
proaches, with eCDM and eCZM the corresponding strain fields:
with P the surface energy density, which also depends on a history Z L Z L
variable field in order to take into account irreversibility of fracture. r dx
U CDM ¼ eCDM dx ¼ ;
For a given load, the displacement field is defined as a local mini- 0 E 0 AðaðxÞÞ
Z L
ð4-1Þ
miser of the total energy, and fracture is considered reversible. CZM d CZM r
Moreover, as all potential discontinuities are considered through U  ¼ e dx ¼ L;
2 0 E
the dependence of C on the displacement field u, the subsequent
numerical minimisation in a finite element framework would a pri- where d = sutx=0 denotes the displacement jump at the interface
ori require to allow potential discontinuities anywhere, hence for x = 0. Thereafter, the load level is parameterised in the non linear re-
practical reasons at the interface between all continuous finite ele- gime by the homogeneous stress r, which decreases from ry to 0.
ments. It is now well known that such a strategy does not converge By enforcing the equality between UCDM and UCZM at each load level
248 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259

Fig. 4-1. (a) Equivalence between non local damage model and cohesive zone model. (b) Switching from non local model to cohesive zone model.

r, one can extract the expression of the displacement discontinuity tion by inserting a cohesive interface at the center of the observed
d(r): localisation band while maintaining the stiffness distribution re-
lated to the damage field atcr ðxÞ. It is important to highlight that:
U CDM ¼ U CZM
Z  Z b ! – at time t = tcr, the displacement jump d(r) has to be zero in
L
2r dx 2r dx
) dðrÞ ¼ L ¼ b ; order to preserve equilibrium;
E 0 AðaðxÞÞ E 0 AðaðxÞÞ
dðrtcr Þ ¼ 0; ð4-3Þ
ð4-2Þ
– once this transition is triggered, dissipative mechanisms are
with b the half width of the localisation band under the stress level only embedded in the cohesive zone model, consequently, dam-
r. It is then possible to compute incrementally the corresponding age does not evolve anymore:
cohesive law (d, r(d)): for a given stress value r such as 0 < r < ry, 8t P tcr ; at ðxÞ ¼ atcr ðxÞ: ð4-4Þ
the damage field a(x) and the corresponding localisation band
width 2b are known (as seen in the previous section). d(r) is then
computed by means of (Eq. (4-2)) using a numerical integration The notation U CDM
t is preserved to denote the prescribed dis-
procedure (here a trapezoidal rule was used). Fig. 4-2 shows the placement seen by the non local approach at time t, while U TRA
t des-
equivalent cohesive law calculated with the material parameters ignates the prescribed displacement at time t > tcr once the
used in Table 2-1, and the solutions obtained independently with transition has been triggered (where exponent ‘‘TRA’’ refers to
both models in term of global response. ‘‘transition’’). These two quantities read:
Z L
rt dx
4.2. Switching from non local damage model to cohesive zone model 8t > tcr ; U CDM
t ¼ ;
E 0 Aðat Þ
Z L ð4-5Þ
dTRA rt dx
As an equivalent cohesive law has been identified, it is interesting 8t > tcr ; U TRA
t  t
¼ :
to investigate in what way it is possible to switch from one model to 2 E 0 Aðatcr Þ
another. In that sense, the following strategy is proposed. First, the The energetic equivalence still relies on the same condition (Eq. (4-
non linear regime is only described with the non local approach, un- 1)), which consists in enforcing the equality between the two pre-
til an arbitrary critical step is reached (denoted by tcr). At this stage, scribed displacements related to both models for each stress level
we switch from a continuous to a discontinuous failure representa- rt, here for t P tcr:

Fig. 4-2. (a) Equivalent cohesive law. (b) Stress versus displacement global responses obtained independently with both models.
 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259 249

Z  
8t > t cr ;U CDM ¼ U TRA 2r t bt
1
t t
Z   8t > t cr ; dREF ðrt Þ ¼  1 dx; ð4-9Þ
dTRA ðrt Þ rt L 1 1 E 0 Aðat Þ
) ¼  dx: ð4-6Þ
2 E 0 Aðat Þ Aðatcr Þ – the second one, denoted by dCOR
t cr , is a linear function of the stress
rt. The dependence of this corrective term on the time when the
(Eq. (4-6)) can be rewritten: model switches is highlighted by the subscript tcr
8t > t cr ; Z btcr  
Z   Z   2r t 1
dTRA ðrt Þ rt L
1 rt L 1 8t > t cr ; dCOR
t cr ðrt Þ ¼  1 dx; ð4-10Þ
¼  1 dx þ 1 dx: ð4-7Þ E 0 Aðatcr Þ
2 E Aðat Þ E 0 Aðatcr Þ
0
8t > t cr ; dTRA
t cr ðrt Þ ¼ d
REF
ðrt Þ  dCOR
tcr ðrt Þ: ð4-11Þ
If bt and btcr , respectively denotes the half width of the localisation
band at time t and tcr, the first integrand in the equation (Eq. (4-7)) Thus, the cohesive law (dREF, r) appears to be a master curve,
is zero in the elastic part [bt, L] and the second one too in the elastic only depending on the non local model and
 its parameters values,
part ½btcr ; L: and from which any cohesive law dTRA t cr ; r ensuring the energetic
equivalence with the continuous/discontinuous approach can be
8t > t cr ; calculated according to (Eq. (4-11)) given a critical parameter tcr.
Z   Z   Moreover, the corrective term dCOR
2rt bt
1 2rt btcr 1 t cr ðrÞ vanishes when r ? 0, thus
dTRA ðrt Þ ¼  1 dx   1 dx: the displacement jump corresponding to zero stress is the same
E 0 Aðat Þ E 0 Aðatcr Þ
whatever the critical parameter tcr:
ð4-8Þ
This equation means that the displacement jump can be expressed 8tcr ; dTRA
t cr ð0Þ ¼ d
REF
ð0Þ: ð4-12Þ
as the difference between two positive contributions:
Up to now the loading parameter was a(0) = a0 (the current top
REF
– the first one, denoted by d , is the displacement discontinuity damage inside the localisation band). From now on we choose
that belongs to the previously called ‘‘equivalent cohesive law’’ acr ¼ atcr ð0Þ as the critical parameter, which is more meaningful
and detailed in (Eq. (4-2)). The exponent ‘‘REF’’ refers to the fact than time when dealing with quasi-static problems. Fig. 4-3(a)
that this cohesive law is perfectly equivalent to the gradient shows the shape of two induced cohesive laws corresponding to
damage model, which is here the reference model the critical values acr = 0.3 and 0.7; Fig. 4-3(b,c) shows the associ-
ated global responses obtained by switching the model when the
loading parameter a(0) reaches its critical value acr.

Fig. 4-3. (a) Example of two induced cohesive laws for the transition criteria acr = 0.3 and 0.7. (b,c) Respective global responses.
250 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259

It is important to highlight that unlike the approaches proposed (iv) finally, fields (u, a, k) (where u is the displacement field), are
in [6,9], the construction of the cohesive law leads to an explicit discretised with classical continuous Lagrange shape func-
expression (d, r(d)). It even leads to a closed-form expression as tions whereas the discretisation of the auxiliary field a is
long as the integrals (Eqs. (4-9) and (4-10)) can be expressed ana- achieved on collocation points that coincide with the Gauss
lytically and the resulting expressions d(r) (Eq. (4-11)) is invert- points related to the quadrature rule.
ible. As the chosen non local constitutive equations do not
provide such properties, numerical integration procedures are used Steps (i), (ii), and (iii) turn the minimisation of the potential en-
and the resulting cohesive laws are stored as tabulated curves. One ergy into finding the saddle-point of the following augmented
could also has chosen to fit this results with simple analytical func- Lagrangian, so that the problem reads:
tions, but in this case, the subsequent FE implementation would
Find ðunþ1 ; anþ1 ; knþ1 ; anþ1 Þ realising min min max min ½Lr ðu; a; k; aÞ;
have become dependent on the fitted analytical formulas (that de- u a k a
pends on the chosen material parameter values). ð5-4Þ
Finally, as this one-dimensional study is led only in the case of
increasing monotonic load, no attention is paid to the unloading Wext(u) being the potential of external work and r > 0 the penalty
behaviour of the coupled model. In order to stay consistent with parameter. Step (iv) permits to eliminate the auxiliary field a from
the non local damage model which unloads linearly, a linear the global formulation at a local level: minimisation with respect to
unloading branch has to be chosen for the cohesive law. a is then very close to the integration of the purely local damage
law (a is the unique root of a scalar equation at each Gauss point).
Once a is computed at the integration points, the stress tensor r is
5. Finite element implementation
determined trough (Eq. (2-1)). It is then possible to compute the
residual vectors related to the unknown nodal fields (u, a, k) and
In this section, we first recall the spatial discretisation dedicated
the consistent tangent matrix associated to the global nonlinear
to the gradient damage model and the cohesive zone model (see
system (the solution is obtained using a global Newton–Raphson
respectively [20,21]) which both rely on mixed finite elements.
method, see [20] for more details).
Integration with respect to time is performed by means of an im-
The resulting mixed finite element relies on the following
plicit Euler scheme. Then, the coupling algorithm that enables
choice for the shape functions and the quadrature rule: the dis-
the extension of the preliminary one-dimensional results to a bi-
placement u is interpolated with P2-continuous polynomials, the
dimensional FE context is presented.
damage field a and the Lagrange multiplier field k with P1-contin-
uous polynomials, and the local damage field a with P1-discontin-
5.1. Gradient damage model
uous polynomials. In a bi-dimensional setting, the latter
corresponds to a collocation method on a set of three points (trian-
The problem is formulated as a minimisation problem at the
gle) or four points (quadrangle) which are respectively chosen as
structure scale. After time discretisation, it reads at the end of each
the set of integration points to provide the quadrature rule.
time step [tn, tn+1]:

unþ1 ¼ arg min Eðu; anþ1 Þ; anþ1 ¼ arg min Eðunþ1 ; aÞ; 5.2. Cohesive zone model
u2KA a
ð5-1Þ
n
Eðu; aÞ ¼ FðeðuÞ; aÞ þ Dða  a Þ  W ext ðuÞ; In this part we consider a structure X crossed by a potential dis-
where E is the incremental potential energy, F and D are potentials placement discontinuity surface C which represents an a priori
already given in (Eq. (2-7)), and KA is the set of kinematically admis- postulated crack path. The bulk XnC is supposed linear elastic
sible displacements while the crack evolution is modelled through a cohesive zone
As E is highly nonlinear and non differentiable due to the local model. Here again, the problem is formulated as a minimisation
dissipation potential w (Eq. (2-7)), a decomposition-coordination problem: at each load level (see Section 3), the displacement field
technique is used in order to confine the non differentiability at a u is regarded as a local minimum of the potential energy E(u),
local level (Gauss point): which is expressed as the sum of the elastic strain energy, a surface
cohesive energy, and the potential of external work, see (Eq. (3-1)).
(i) an auxiliary field a assumed to be equal to the damage field In order to describe accurately the behaviour of the cohesive
a is introduced as an argument of the non regular part of the crack, some important points must be verified. Mainly, irreversibil-
potential energy; ity of fracture, a contact condition along C that precludes interpen-
Z etration, and a non-zero stress threshold for the crack initiation
Eðu; a; aÞ ¼ ½AðaÞ/el ðeðuÞÞ þ wða  an ÞdX stress criterion. The latter implies a perfect initial adherence (i.e.
X
Z an infinite initial stiffness for the cohesive law (sut, r)) which ex-
c cludes the use of cohesive energy regularisation. The resulting
 W ext ðuÞ þ ðraÞ2 dX; ð5-2Þ
X 2 non-differentiability of P leads to a multi-valued cohesive law,
which requires a special treatment. As in the previous part, a
(ii) the formulation is augmented by adding a quadratic penalty
decomposition – coordination technique is used, this time by intro-
term to the potential energy, (with r > 0 the penalty param-
ducing an auxiliary field d (assumed to be equal to the displace-
eter) without any effect while the condition a = a is met;
ment discontinuity sut across C) as an argument of the non-
(iii) the constraint a = a is enforced through dualisation, leading
differentiable energy density P. The same steps lead to the follow-
thus to the following augmented Lagrangian with k the
ing augmented Lagrangian (see [21] for more details):
Lagrange multiplier field;
Z Z Z
Lr ðu; a; k; aÞ ¼ n
½AðaÞ/el ðeðuÞÞ þ wða  a ÞdX  W ext ðuÞ Lr ðu; t; dÞ ¼ /el ðeðuÞÞdX þ PðdÞdC  W ext ðuÞ
XnC C
Z X
Z Z Z
c r
þ ðraÞ2 dX þ kða  aÞdX þ t  ðsut  dÞdC þ ðsut  dÞ2 dC; ð5-5Þ
2 X X C 2 C
Z
r 2 where t is the Lagrange multiplier field which actually represents
þ ða  aÞ dX: ð5-3Þ
2 X the cohesive force vector.
 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259 251

Thanks to the restrictive hypothesis of a postulated crack path, pate the amount of transferred energy). In order to prevent
the interface C can be discretised (still in a bi-dimensional context) this kind of spurious damage evolution, all the previously
with degenerated quadrangles (zero thickness). The bulk is meshed mentioned layers of non linear elements are switched to lin-
with triangles or quadrangles, in accordance with the previous ear elastic elements while preserving the field of rigidity loss
interface discretisation, so that the nodes of both sides of C (C A(a) (Eq. (2-1)) due to the damage field a reached just before
and C+) coincide. A quadratic interpolation with P2-continuous the transition. The detection of such elements is easier with
elements is adopted for the elastic bulk. On the other hand, the structured meshes. In several simulations led without pro-
mixed interface finite element relies on the following choices: viding this specific treatment, damage growth was still
the displacement u is interpolated with P2-continuous polynomi- observed in the vicinity of activated cohesive elements (at
als, the Lagrange multiplier field t with P1-continuous polynomi- least during the first time step subsequent to their activa-
als, and the local auxiliary field d with P1-discontinuous tion). Even if those solid elements ended up unloading elas-
polynomials. The latter corresponds to a collocation method on a tically while the displacement jump was increasing, it has
set of two points which is chosen as the set of integration points led in some cases to a lack of convergence.
to provide the quadrature rule on the degenerated quadrangle.
The coupling algorithm relies on a simple criterion which states
Remark 1. For the sake of simplicity the cohesive crack model is that damage growth can only occur as its value a0 at the center of
presented in the case of linear elastic bulk behaviour, but in the the band does not exceed a critical one, denoted acr. Otherwise, the
following part both formulations are coupled. The material transition to CZM has to be triggered:
behaviour of the bulk is described by the gradient damage model
and its associated spatial discretisation (the augmented Lagragian If a0 P acr ; then switch ð5-6Þ
(Eq. (5-4))), thus leading to a more complex FE formulation.
From a practical point of view, the proposed method (see Fig. 5-
1(a)) is explicit as the integration points of the cohesive interface
5.3. Coupling algorithm C (which is initially in a perfect adherent regime) are updated at
the end of any converged time step fulfilling the switch criterion.
Under restrictive hypotheses such as straight crack propagation At the integration point level, if the condition (Eq. (5-6)) is fulfilled
under mode I loading conditions, the preliminary one-dimensional a cohesive law that enables to dissipate the appropriate amount of
results related to the tensile bar can be easily extended in a bi- energy is calculated according to the above mentioned point (i). At
dimensional FE context. Indeed, the knowledge of the crack path the following time step, the cohesive elements containing at least
C (corresponding to the centre of the localisation band, where a an activated integration point no longer remain in their state of en-
is maximum) allows to take into account the cohesive interface forced adherence, thus requiring to freeze the damage evolution in
into the discretisation prior to the FE computation. A mesh that their direct vicinity according to the above mentioned point (ii). Fi-
conforms to the straight interface discretisation (degenerated nally, an additional point to be emphasised is that a special treat-
quadrangles along C) is then built, as explained in the previous ment has to be applied to duplicated non local degrees of
section. freedom (DOFs) carried by coincident vertex nodes located along
It has been shown in [20] that even in 2D or 3D problems, the C (Fig. 5-1(b)). Indeed, in the case of initial adherent regime, as
gradient damage model provides a localisation band that remains cohesive elements ensure the equality between each displacement
confined in a thickness of about 2D (in agreement with the one- DOF corresponding to a any pair of coincident nodes located along
dimensional prediction (Eq. (2-15))) during the propagation. Tak- C (i.e. sut = 0), the condition sat = skt = 0 has to be enforced on the
ing advantage of that property, a structured mesh composed only pairs of coincident vertex nodes (with a,k respectively the P1-con-
of quadrangles is used within the expected localisation band, in or- tinuous fields of damage and Lagrange multiplier (Eq. (5-4))). These
der to provide a simpler and more accurate treatment of the tran- conditions are enforced through dualisation in adherence regime,
sition, mainly for two reasons: and once a cohesive element becomes active it is no longer applied
since damage does not evolve anymore in both neighbouring vol-
(i) in order to extend 1D results to straight crack propagation ume elements. Consequently, the gradient damage model does
with mode I decohesion, we make the assumption that for not require particular boundary condition treatment when the
any point of the potential discontinuity interface, the dam- cohesive crack opens.
age and stress fields along the normal to the crack path are Numerical implementation of such a strategy leads to a global
the same (inside the localisation band) than in the tensile description of the crack that can be divided into three levels during
bar for a corresponding top damage value at the centre of its propagation (see Fig. 5-2):
the band. With this kind of mesh, each cohesive integration
point along C and the two closest integration points in the – a non local damage process zone (DPZ), which relies on a phys-
bulk are located on the same normal to the crack path (see ical background as it allows to describe the early stage of crack
Fig. 5-1(b)). It is then possible to identify for each Gauss inception, as well as microscopic degradation mechanisms that
point of C an appropriate cohesive law when switching, occur upstream of the crack tip in the propagation regime;
using the uniaxial relation (Eq. (4-11)). The corrective term – a cohesive process zone (CPZ), which characteristics should be
in (Eq. (4-11)) is then calculated according to the mean value directly related to the switch criterion acr as the transition sep-
of the additional damage field a met at the two closest vol- arations laws are an approximation of the non local constitutive
ume integration points (Fig. 5-1(b)), which is assumed to be law;
the top damage value a0 at the center of the band. The latter – a traction free discontinuity (TFD), corresponding to complete
assumption leads to an approximation whose effect should decohesion and characterised by total surface damage.
become negligible with mesh refinement.
(ii) Moreover, in order to trigger accurately the transition from 6. Numerical validation
CDM to CZM, damage must not evolve anymore in the vol-
ume elements located inside the localisation band from both This section aims to assess the efficiency of the coupled ap-
sides of an ‘‘activated’’ cohesive element (which has to dissi- proach through three bi-dimensional numerical examples:
252 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259

Fig. 5-1. (a) Explicit coupling algorithm. (b) Sketch of a cohesive element and its volume vicinity.

width of 2D = 100 mm then remains reasonable at the structure

scale. Moreover, the Poisson ratio is set to 0.2, except for the tensile
bar problem (for this test case, it is zero in order to stay in a one-
dimensional context). Material instabilities are expected because
of the quasi-brittle behaviour of the concrete, resulting in potential
snap-backs in the global response. In order to follow them, a path-
following technique proposed in [22,31] is used when necessary. It
mainly consists in controlling the maximum damage increment
Da = an+1  an met at least at one integration point within the
structure body, so as to select only a dissipative solution. As dam-
age should only grow upstream from the cohesive crack tip, its
application to the coupled approach obviously requires to identify
and not to treat the non local elements that have been successively
switched to elastic ones (located from both sides of the opening
lips).
Fig. 5-2. Sketch of the coexisting descriptions of the crack.
6.1. Tensile bar

– A brittle bar submitted to tensile loading, for which the local- The first example concerns the validation of the FE implementa-
ised solution branch is followed. This test case is a first step tion through the tensile bar test-case. As a closed-form solution is
the purpose of which is to validate the numerical implementa- available for the inhomogeneous response, the reproducibility by
tion on the simplest configuration possible (only crack opening, FE computations is assessed: first with the CDM approach only,
i.e. no propagation). and then with the coupled CDM/CZM one. Even if the FE model it-
– A trapezoidal double cantilever beam (DCB), providing rectilin- self is 2D, the inhomogeneity is only expected in the longitudinal
ear stable crack propagation. The sensitivity to the mesh refine- direction so as to provide crack opening (no propagation).
ment and to the switch criterion acr is studied. We consider here a bar of length 2L = 2 m and of thickness
– A perforated plate, providing rectilinear unstable multi-crack t = 200 mm. Boundary conditions consist in a tensile displacement
propagation, which allows checking the robustness of the prescribed at both ends in the longitudinal direction, while rigid
method. body modes in the transversal one are suppressed. Five different
structured meshes are used, obtained by successively refining up
As we focus on large concrete structures, their characteristic to five times an initial (coarse) structured mesh composed of quad-
size is about several meters and the material parameters are cho- rangles. Each refinement step is carried out on a narrowing region
sen according to Table 2-1: the choice of an ultimate damage band- centered on the expected localisation band, where the element size
 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259 253

Fig. 6-1. (a) Finest adopted structured mesh (h = 2D/64). (b) CDM: force–displacement responses observed over mesh refinement. (c) CDM: damage profile observed across
the localisation band on the finest mesh at several load levels.

is halved. If h denotes the characteristic element size in the region and global response is specific to 1D problems, but this can give us a
of interest for a given mesh, and 2D the localisation bandwidth, h first insight about the interest of switching to CZM. Given a charac-
ranges from h = 2D/4 on the first mesh to h = 2D/64 on the finest teristic element size h within the localisation band, the switch
one (see Fig. 6-1(a)). In order to catch the localised solution, an ini- criterion acr could be defined as the top damage value beyond which
tial damage field has to be introduced prior to the first time step the damage gradient is no longer accurately described.
because Newton’s method would select the homogeneous branch: Fig. 6-2(b) illustrates a coupled CDM/CZM FE computation led
all the integration points that belong to both central elements lay- with a mesh size and a switch parameter respectively set to
ers are assigned a damage value a = 105, while a is set to zero h = 2D/16 and acr = 0.3. Accordingly with Section 5.3, a cohesive
everywhere else. Once the dissipative regime is reached, damage interface is introduced at the center of the mesh prior to the com-
then grows at most in the band [D, D] while the bar elastically un- putation, and all the cohesive integration points become active
loads everywhere else. As its length 2L is chosen ten times greater once the switch criterion (Eq. (5-6)) is verified, while the surround-
than it was in Section 2.3, the force–displacement response may ing non local elements are converted to elastic ones. A very good
admit a snap-back requiring the use of the above mentioned agreement with the semi analytical solution is observed, especially
path-following technique. at the late stage of softening (see the corresponding error on the
Fig. 6-1(b) compares the global force–displacement responses CDM global response, Fig. 6-2(a)).
observed on the five different meshes with the closed-form solu- Because of its 1D character, this test case turns out to be very
tion depicted in Section 2.2 (dashed orange1 line). The latter is re- convenient in order to validate some basic parts of the implemen-
ported to this pseudo-2D case with 2L = 2 m and by multiplying tation (such as automatic identification and integration of the
the stress by the thickness t to obtain the resultant force. One can cohesive law). Nevertheless, as it deals with pure crack opening,
observe that at the beginning of the snap-back all computed solu- further investigations should be conducted for real crack propaga-
tions are coincident with the semi analytical one. As damage grows, tion cases that are more relevant to assess the capabilities and
a gap gradually develops and they eventually become asymptotic interests of the approach.
when the closed-form solution predicts that the ultimate prescribed
displacement remains finite. The coarser the mesh is, the earlier this
6.2. Trapezoidal DCB
gap begins developing, meaning that beyond a certain damage level
the element size does not provide an accurate enough description of
The trapezoidal double cantilever beam test is described in
the damage gradient. Indeed, Fig. 6-1(c) shows, for five given stress
Fig. 6-3(a). The notched specimen is submitted to prescribed open-
states, the comparison between the theoretical damage profiles and
ing displacements at the mouth of the pre-crack, thus providing
the computed ones (with h = 2D/64): they all perfectly match ex-
mode I loading conditions. In order to avoid a possible bifurcation
cepted for the sharp one corresponding to almost zero stress into
from the rectilinear crack path, a trapezoidal shape is chosen rather
the bar, for which the FE computation leads to a very slight overes-
than the classical rectangular one. As it is specified in Section 5.3, a
timation (maximum relative error of about 2%). The relative error
structured sub-mesh composed of quadrangles is adopted in the
observed for the corresponding point of the global response
expected damage band location, h denoting the size of these quad-
h = 2D/64 (Fig. 6-1(b), green curve) is distinctly amplified (relative
rangles, and a cohesive interface is set up in its center. The rest of
error of about 25%). Such a correspondence between local fields
the mesh remains unstructured (see Fig. 6-3(b) for h = 2D/4).
The sensitivity to mesh refinement is first studied on the basis
1
For interpretation of color in Figs. 6-1, 6-4, 6-6 and 6-8, the reader is referred to of five meshes, but unlike the previous test case, dividing by two
the web version of this article. the characteristic element size h at each refinement step here
254 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259

Fig. 6-2. Observed force–displacement responses with h = 2D/16: (a) CDM. (b) Coupled CDM/CZM with acr = 0.3.

Fig. 6-3. Trapezoidal DCB: (a) specimen geometry and loading conditions. (b) Coarsest adopted mesh (h = 2D/4).

quickly leads to an exploding number of DOFs, so that refinement for h = 2D/12. The irregularities originate from a rough description
levels such as h = 2D/64 really become prohibitive in terms of of the damage field, resulting in identified cohesive laws that do
computational time. Consequently, FE computations are performed not fulfil accurately enough the energy balance at each integration
with meshes corresponding to 4, 6, 8, 10, and 12 elements across point of a sudden crack extension. Consequently, an increase of the
the bandwidth, resulting in a number of DOFs ranging from switch criterion value must be accompanied with an increase of
45,470 to 153,830. The convergence analysis is performed with a the mesh refinement.
switch criterion arbitrarily set to acr = 0.5. The results relative to Fig. 6-4(b) compares the global force–displacement response
the five meshes are plotted in terms of force versus displacement obtained on the finest mesh h = 2D/12 with CDM, CZM, and the
on Fig. 6-4(a). The grey dashed line corresponds to the first load coupled approach. The CZM response actually corresponds to the
increments during which the damage spreads only at the very switch criterion acr = 0, in the sense that only the ‘‘reference’’ cohe-
end of the notch. The other plotted points start when cohesive sive law (Eq. (4-9)) is involved during the computation. The ob-
crack growth is first triggered so that both CDM and CZM begin served discrepancy between CDM (black curve) and CZM (blue
to coexist in the model. One can observe that for the coarsest curve) responses mostly stems from the propagative nature of
meshes the responses show bumps that gradually fade out with the test, but this slight gap diminishes when refining the mesh.
mesh refinement, and eventually converge toward a smooth curve Concerning the coupled approach response (acr = 0.5, red dots),
 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259 255

Fig. 6-4. Trapezoidal DCB: (a) global force–displacement responses observed over mesh refinement. (b) Global force–displacement responses observed on the finest mesh
(h = 2D/12).

Fig. 6-5. Trapezoidal DCB: damage distribution reported on the deformed configuration (displacement amplified 800 times), with acr = 0.5 and h = 2D/12.

the obtained curve first fits the CDM response, and gradually both sides of the cohesive crack lips. Upwards of the cohesive crack
comes between the CDM and CZM curves in the post-peak regime. tip, the damage process zone contains several elements in which
Despite these slight differences, a good agreement with the non lo- the switch criterion (Eq. (5-6)) is met, so that a crack extension will
cal response is obtained. occur at the following time step.
From a qualitative point of view, the damage profiles related to The next step of the study consists in evaluating the influence of
this computation (acr = 0.5 and h = 2D/12) are plotted in Fig. 6-5, the switch parameter acr. Computations are led with a mesh corre-
and reported on the deformed configuration with an amplification sponding to h = 2D/10 as it provides sufficiently smooth responses,
factor so that crack opening can be appreciated. As expected, the and acr is set successively to acr = 0.1, 0.3, 0.5, 0.7.
bandwidth remains constant along the damage zone with a maxi- First, Fig. 6-6(a) shows the force versus displacement responses
mum value acr = 0.5 reached in each layer of elements located from related to the four values of acr, again compared with the CDM and

Fig. 6-6. Trapezoidal DCB (mesh parameter h = 2D/10): (a) influence of acr on the global force–displacement responses; (b) evolution of the cohesive process zone length and
the traction-free discontinuity length with respect to the structural compliance (abscissa in logarithmic scale).
256 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259

Table 6-1
Trapezoidal DCB (mesh parameter h = 2D/10, required volume damage increment per time step Da = 0.1): comparison between non local and coupled computations in terms of
computational cost (number of Newton iterations per time step).

Number of Newton iterations per time step Gradient damage model Coupled, acr = 0.1 Coupled, acr = 0.3 Coupled, acr = 0.5 Coupled, acr = 0.7
Maximum 10 10 10 9 9
Minimum 7 7 6 6 6
Average 7, 7 8, 5 7, 8 7, 5 7, 6

the CZM responses. The grey curve associated to acr = 0.1 remains lution curve shows three stages, first a rising part that corresponds
the closest to the cohesive one while the orange one associated to the cohesive crack initiation (which ends once an integration
to acr = 0.3 presents the most important deviation. Red and green point of the interface is fully damaged), then a plateau value
curves (0.5 and 0.7) are very close at the peak and almost reflecting a steady propagation regime that ends in a decreasing
superimposed with the grey one (0.1) in the steady propagation part (that would not exist in the case of a semi-infinite medium)
regime. We think that these deviations can be attributed to 2D as the crack tip moves closer to the edge of the structure. TFD
effects that are not taken into account in the construction of the length evolution curves show two stages: a zero plateau during
mode I switch cohesive laws. Indeed they are constructed to crack inception and an increasing part during its propagation.
strictly reproduce the 1D opening behaviour provided by the One can notice that for acr = 0 (CZM only, blue curves), the CPZ
continuous model. In 2D propagation cases, a slight approximation length identified during the steady propagation part in Fig. 6-
is made while not taking into account the gradient of damage in 6(b) is about 500 mm that is 1/6 of the ligament length. The order
the longitudinal direction (parallel to the crack path). However, of magnitude is in accordance with the expression provided by
these results show a quite weak dependence on acr, with a maximal Rice in [28] (Eq. (6-1)), that leads to 1/9 of the ligament length:
discrepancy (about 5%) observed at the peak, which significantly
decreases in the post-peak part. One can also observe that
EGf
whatever the value of acr, the force peak is reached for the same LCZM  ! N:A:LCZM  345 mm ð6-1Þ
prescribed displacement level than for the CDM response.
r2y ð1  m2 Þ
The dependence on the parameter acr is then studied in terms of
cohesive crack evolution. The evolution of the cohesive process For the coupled approach, one has seen that the chosen value of acr
zone (CPZ) length and the traction-free discontinuity (TFD) length determines the size of the damage process zone (DPZ): the later the
versus structural compliance is plotted in Fig. 6-6(b) for the four transition is triggered, the longer the DPZ is. Consequently, a consis-
values of acr and for CZM only. In this figure, the structural compli- tent model should provide, for a positive value of acr, a CPZ length
ance is used instead of the prescribed displacement as the load lower than the one obtained with acr = 0, and this length should also
parameter because in this case, the CPZ and TFD length curves decrease as acr increases. This is what we observe in Fig. 6-6(b) for
would not have started at the same abscissa: indeed, the latter acr = 0.1, 0.3, 0.5 and 0.7. Moreover, an important result is that for
the first switch is triggered, the higher the prescribed displace- each of the five CPZ curves, the three stages (cohesive crack initia-
ment. However, as volume damage has already contributed to in- tion, steady-state propagation, and structure edge effect) are de-
crease the structural compliance, cohesive process zones growth fined on the same ranges of structural compliance, and at the
start at the same abscissa. The CPZ and TFD lengths are computed same time the TFD curves are perfectly coincident. These results
as the sum of the cohesive elements lengths in which the surface prove that the coupled approach preserves the properties of the
damage is in ]0, 1[ (CPZ), and equal to 1 (TFD). Each CPZ length evo- CZM in terms of propagation kinetics whatever the switch criterion

Fig. 6-7. Perforated plate: (a) specimen geometry and loading conditions. (b) Coarsest adopted mesh (h = 2D/4).
 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259 257

Fig. 6-8. Perforated plate (mesh parameter h = 2D/10): (a) damage distribution reported on the deformed configuration (displacement amplified 800 times). (b) Global force–
displacement responses.

Fig. 6-9. 3D Trapezoidal DCB: damage distribution reported on the deformed configuration, with acr = 0.6 and h = 2D/8.

value acr, and that the implemented framework involves simulta- in the expected localisation band and a priori discretisation of the
neously CZM and non local CDM in a consistent manner. cohesive interface), see Fig. 6-7(b).
The influence of the parameter acr is finally considered in terms Numerical simulations are performed with a mesh size set to
of computational cost. It is a priori possible that a sudden extension h = 2D/10. Non local and CZM computations are first carried out;
of the cohesive crack requires more Newton iterations in the
subsequent time step than with the non local model alone, because
an additional non linearity is introduced in the CPZ whose length is
directly related to acr. However, such cost increases are not ob-
served in Table 6-1: an average number of 8 Newton iterations
per time step is required to reach convergence, for the non local
model as for each value of acr.

6.3. Perforated plate

We consider here a perforated plate subjected to tension on its

opposite faces, see Fig. 6-7(a). As the hole is located at the center of
the plate, two cracks are expected to propagate simultaneously in
both ligaments until the complete failure occurs. Nevertheless a
possible loss of symmetry can be encountered, mainly because of
mesh imperfections. Moreover, an unstable propagation resulting
in snap-backs is expected. This test aims to assess the robustness
of the approach in the case of complex propagations involving mul-
ti-cracking in the presence of snap-backs, thus requiring the use of
the path following technique [22]. As in the previous section, the Fig. 6-10. Trapezoidal DCB: comparison between 3D and 2D computation in terms
same mesh construction procedures are applied (structured mesh of global force–displacement response.
258 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259

results are plotted in terms of global response in Fig. 6-8(b). The  a straightforward extension of the approach to 3D plane crack
first snap-back corresponds to the propagation of both cracks propagation is now in progress.
(during which a slight gap is observed between non local and cohe-
sive responses), ending once the process zones reach the end of the The validation of the approach on these quite particular cases is
ligaments. The second snap-back is the result of complete failure of a necessary first step before considering an extension to more
both ligaments: symmetry is preserved until the end with the complex situations. In its current state, this approach has some
cohesive zone model (blue curve), while it is lost just before com- limitations that must be overcome in order to open up new pros-
plete failure with the non local model (black curve), so that the lig- pects, mainly:
aments successively break (resulting in the last bifurcated branch).
A simulation involving the coupled approach is then performed  the implementation of a crack-path tracking algorithm, such as
with a switch parameter acr = 0.5. Qualitatively, both cracks propa- those used in [24] (based on the spatial distribution of damage)
gates simultaneously during the first snap back (Fig. 6-8(a)), and or in [12] (based on the spatial distribution of the gradient of
the results are quantitatively in good agreement with the non local damage), in order to overcome the hypothesis of postulated
and CZM approaches (Fig. 6-8(b)). Nevertheless, convergence is- crack paths;
sues are encountered once the damage process zones get too close  the gradual insertion of the cohesive discontinuities within the
from the ligaments end (Fig. 6-8(a) shows the crack configuration spatial discretisation, with the X-FEM technique or with a reme-
at the last converged time step). shing tool, which is the next step of the previous point;
 the use of mesh adaptivity techniques such as for instance the
damage rate based remeshing/transfer procedure proposed in
6.4. Extension to 3D plane crack propagation
[24] could reduce the computational cost, by refining the fracture
process zone throughout its propagation (thus providing a better
In order to show the ability of the method to deal with 3D crack
description of the damage field) an by coarsening areas where
propagation, a straightforward extension of the 2D implementation
material separation is complete (no more cohesive stress).
detailed in Section 5.3 to a 3D problem has been carried out. This
 the extension to mixed mode propagation which seems to be a
time, plane crack propagation is considered and the local activation
tricky task.
criterion still remains applicable if a planar cohesive interface is in-
serted in the discretisation before the beginning of the computation.
Preliminary results have been obtained on the trapezoidal DCB stud-
References
ied in paragraph 6.2: as 3D computations are more expensive in
terms of CPU cost, the specimen dimensions have been reduced in [1] S. Andrieux, E. Lorentz, M. Joussemet, Une classe de modèles de
order to keep a certain margin for its extrusion thickness value that comportements avec gradients de variables internes: problème d’évolution,
has been set to 150 mm. For the same reason, a relatively coarser C. R. Acad. Sci. Paris 323 (Série II b) (1996) 571–578.
[2] G.I. Barenblatt, The mathematical theory of equilibrium cracks in brittle
mesh has been chosen, with a characteristic size of h = 2D/8. The crit- fracture, Adv. Appl. Mech. 7 (1962) 55–129.
ical damage has been set to acr = 0.6. The corresponding damage map [3] T. Belytschko, T. Black, Elastic crack growth in finite elements with minimal
is presented in Fig. 6-9 in the deformed configuration with the cohe- remeshing, Int. J. Numer. Methods Engrg. 45 (1999) 601–620.
[4] A. Benallal, J.J. Marigo, Bifurcation and stability issues in gradient theories with
sive crack opening. The comparison in terms of global response with softening, Model. Simul. Mater. Sci. Engrg. 15 (2007) 283–295.
a 2D plane strain simulation with the same dimensions is plotted in [5] B. Bourdin, G. Francfort, J.J. Marigo, The Variational Approach to Fracture,
Fig. 6-10. In order to compare the two global responses, the resultant Springer, 2008.
[6] F. Cazes, M. Coret, A. Combescure, A. Gravouil, A thermodynamic method for
force has been divided by the specimen thickness: one can observe
the construction of a cohesive law from a nonlocal damage model, Int. J. Solid
that the two curves are nearly superimposed, thereby validating Struct. 46 (2008) 1476–1490.
these preliminary results. [7] R. Chambon, D. Caillerie, T. Matsushima, Plastic continuum with
microstructure, local second gradient theories for geomaterials: localization
studies, Int. J. Solid Struct. 38 (2001) 8503–8527.
7. Conclusion [8] M. Charlotte, J. Laverne, J.J. Marigo, Initiation of cracks with cohesive force
models: a variational approach, Eur. J. Mech. A/Solids 25 (2006) 649–669.
[9] C. Comi, S. Mariani, U. Perego, An extended FE strategy for transition from
A finite element approach which relies on the combination of a continuum damage to mode I cohesive crack propagation, Int. J. Numer. Anal.
gradient damage model with a cohesive zone model has been pro- Methods Geomech. 31 (2007) 213–238.
[10] R. de Borst, H.B. Mühlhaus, Gradient-dependent plasticity: formulation and
posed in order to simulate crack propagation within large quasi- algorithmic aspects, Int. J. Numer. Methods Engrg. 35 (1992) 521–539.
brittle structures. It is based on the local extension (at the integra- [11] D.S. Dugdale, Yielding of steel sheets containing slits, J. Mech. Phys. Solids 8
tion point level) of an exact equivalence between both models con- (1960) 100–104.
[12] S. Feld-Payet, Amorçage et propagation de fissures dans les milieux ductiles
structed for a one-dimensional tensile problem. The consecutive non locaux, Ph.D. thesis, École Nationale Supérieure des Mines de Paris, 2010.
transfer of fracture energy is performed through the definition of [13] F. Fraternali, Free discontinuity finite element models in two-dimensions for
a cohesive law, which in a bi-dimensional context is only valid in-plane crack problems, Theoret. Appl. Fract. Mech. 47 (2007) 274–282.
[14] A. Hillerborg, M. Modeer, P.E. Petersson, Analysis of crack formation and crack
for mode I loading conditions. The numerical assessment of the growth in concrete by means of fracture mechanics and finite elements,
proposed approach shows promising results: Cement Concrete Res. 06 (1976) 773–782.
[15] M. Jirásek, Comparative study on finite elements with embedded
discontinuities, Comput. Methods Appl. Mech. Engrg. 188 (2000) 307–330.
 it shows a very slight dependence on the critical damage value
[16] M. Jirásek, T. Zimmermann, Embedded crack model: I. Basic formulation, Int. J.
beyond which the transition is triggered; Numer. Methosd Engrg. 50 (2001) 1269–1290.
 it remains consistent with both non local and cohesive stand- [17] M. Jirásek, T. Zimmermann, Embedded crack model. Part II: Combination with
alone models, from an energetic point of view as well as in smeared cracks, Int. J. Numer. Methods Engrg. 50 (2001) 1291–1305.
[18] J. Lemaitre, J.L. Chaboche, Mécanique des matériaux solides, Dunod, 1988.
terms of crack propagation kinetics; [19] E. Lorentz, S. Cuvilliez, K. Kazymyrenko, Convergence of a gradient damage
 the additional non linearity introduced with the cohesive crack model toward a cohesive zone model, C. R. Méc. (2010), http://dx.doi.org/
has no effects in terms of computational cost; 10.1016/j.crme.2010.10.010.
[20] E. Lorentz, V. Godard, Gradient damage models: toward full-scale
 it is able to deal with brutal crack propagation (involving snap- computations, Comput. Methods Appl. Mech. Engrg. 200 (2011) 1927–1944.
backs in quasi-static analysis) when used with a path following [21] E. Lorentz, A mixed interface finite element for cohesive zone models, Comput.
technique; Methods Appl. Mech. Engrg. 198 (2008) 302–317.
 S. Cuvilliez et al. / Comput. Methods Appl. Mech. Engrg. 237–240 (2012) 244–259 259

[22] E. Lorentz, P. Badel, A new path-following constraint for strain-softening finite [27] G. Pijaudier-Cabot, Z.P. Bazant, Nonlocal damage theory, Engrg Mech., ASCE
element simulations, Int. J. Numer. Methods Engrg. 60 (2004) 499–526. 113 (1987) 1512–1533.
[23] J. Mazars, G. Pijaudier-Cabot, From damage to fracture mechanics and [28] J.R. Rice, The mechanics of earthquake rupture, in: A. M Dziewonski, E. Boschi
conversely: a combined approach, Int. J. Solid Struct. 33 (1996) 3327–3342. (Eds.), Physics of the Earth interior, Proc. Int. School of Physics ‘‘Enrico Fermi’’,
[24] J. Mediavilla, R.H.J. Peerlings, M.G.D. Geers, An integrated continuous- course 78, Italian Physical Society and North-Holland Publ., 1979, pp. 555–649.
discontinuous approach towards damage engineering in sheet metal forming [29] A. Simone, G.N. Wells, L.J. Sluys, From continuous to discontinuous failure in a
processes, Engrg. Fract. Mech. 73 (2006) 895–916. gradient-enhanced continuum damage model, Comput. Methods Appl. Mech.
[25] N. Moës, J. Dolbow, T. Belytschko, A finite element method for crack growth Engrg. 192 (2003) 4581–4607.
without remeshing, Int. J. Numer. Methods Engrg. 46 (1999) 131–150. [30] Code_Aster, FE software distributed by EDF, Open-source. <http://www.code-
[26] R.H.J. Peerlings, R. de Borst, W.A.M. Brekelmans, J.H.P. de Vree, Gradient- aster.org>.
enhanced damage for quasibrittle materials, Int. J. Numer. Methods Engrg. 39 [31] Méthodes de pilotage du chargement, Documentation de Référence du
(1996) 3391–3403. Code_Aster R5.03.80, <http://www.code-aster.org>.