Crack patterns in directional drying: numerical simulations

using a variational approach and experiments.

C. Maurini

, B. Bourdin

, G. Gauthier

V. Lazarus

July 20, 2011

Abstract
Cracks sometimes form beautiful, complex, intriguing, but hard to predict patterns. Stable
propagation of preexisting simple cracks is nowadays quite well described by the traditional
linear elastic fracture mechanics theory. However, crack nucleation, unstable propagation,
or complicated fracture pattern identication are still poorly understood. The variational ap-
proach to brittle fracture provides a mathematically sound model based on Grifths idea of
competition between bulk and fracture energy and deals naturally with these issues, without
resorting to a priori assumptions on fracture patterns. Here, we focus on shrinkage cracks
obtained in the directional drying of a colloidal suspension conned in a capillary cell. We
propose a variational model, describe its approximation and numerical implementation. We
achieve qualitative and quantitative agreement between numerical simulations, experiments,
and known closed form solutions.
1 Introduction
Giants Causeway (DeGraff & Aydin 1987), Port Arthur tessellated pavement (Branagan &
Cairns 1993), Bimini Road (Shinn 2009), Mars polygons (Mangold 2005), septarias (Seilacher
2001), fracture networks in permafrost (Plug & Werner 2001), muds or thin lms (coatings,
paints) (Lazarus & Pauchard 2011, Colina & Roux 2000) are some more or less known ex-
amples of self-organized crack patterns that have intrigued people throughout history. The
spiderweb-like fracture networks observed during impacts (Corbett et al. 1996) or indenta-
tion (Rhee et al. 2001) of plates are further, engineering harmful, examples of complex crack
patterns. The mechanism of formation of all those complex crack paths is still poorly under-
stood and no fracture mechanics tool has been shown, at present, to be able to quantitatively
predict the development of these complicated patterns without making any a priori assump-
tion on their geometry.
The traditional approach to fracture mechanics derived from pioneer works of Grifth
(1920) and Irwin (1958) is able to predict the propagation conditions (critical loading, path)
of a preexisting crack in some situations. Modeling typical features of the complex cracking
phenomena mentioned above such as nucleation, initiation, instantaneous propagation, and
branching however poses fundamental, theoretical, and numerical hurdles. The variational

UPMC Univ Paris 6, UMR 7190 and CNRS, UMR 7190, Institut Jean Le Rond dAlembert, Boite courrier 161-2,
4 Place Jussieu, F-75005, Paris, France

Department of Mathematics and Center for Computation & Technology, Louisiana State University, Baton Rouge
LA 70803, U.S.A.

UPMC Univ Paris 6, UMR 7608, Univ Paris-Sud, UMR 7608, and CNRS, UMR 7608. FAST, Bat 502, Campus
Univ, F-91405, Orsay, France

UPMC Univ Paris 6, UMR 7608, Univ Paris-Sud, UMR 7608, and CNRS, UMR 7608. FAST, Bat 502, Campus
Univ, F-91405, Orsay, France
1
approach to fracture mechanics proposed by Francfort & Marigo (1998) appears as a promis-
ing framework to address these issues. It extends the energetic theory of Grifth and treats
the crack geometry as a genuine unknown, making no a priori assumptions on its geometry
or temporal evolution. Instead, it postulates that the deformation and crack conguration of
a body at a given time is obtained by minimising the sum of its elastic and fracture ener-
gies, among all admissible crack sets and (discontinuous) displacement elds. In the simplest
model, the fracture energy is proportional to its surface in two space dimensions and its area
in three dimensions (Grifth fracture energy).
Of course such a minimisation problem is challenging in particular because it is techni-
cally not possible to test all crack congurations. One possibility is to suppose that the crack
shape is known and to restrict the minimisation on the subset formed by these morphologies.
But how can one be sure that the optimum solution belongs to this subset? Alternatively, to
answer this question, numerical regularized approaches can be used. Roughly, they consist
in replacing the minimisation on all possible crack congurations corresponding to a discon-
tinuity surface by a minimisation on a continuous scalar eld more suitable for numerical
purposes.
More precisely, their analysis borrows tools from the calculus of variations and free-
discontinuity problems (Ambrosio et al. 2000). Their numerical implementation relies on
the concept of variational approximation by elliptic functional (Braides 1998), where approx-
imated functionals depending on a regularization parameters are constructed. Froma technical
standpoint, the approximation takes place in the sense of convergence (Braides 2002) i.e.
one can prove that as the regularization parameter goes to 0, the minimisers of the regularized
functionals approaches that of the total energy. The specic regularized functional we focus
on here resembles gradient damage laws or phase eld approximations of sharp interfaces
models (Hakim & Karma 2009, Corson et al. 2009). It is very similar to the one proposed
by Ambrosio & Tortorelli (1992) where the crack set is represented by a secondary smooth
variable and the displacement eld is also approximated by a smooth function. The main ad-
vantages of this approach is that it eliminates the issue of representing discontinuous elds
when their discontinuity set is not known a priori. It also reduces energy minimisation with
respect to any admissible crack geometry to minimisation with respect to a smooth eld, a
much simpler problem. In addition, it can be discretized numerically using standard continu-
ous nite elements. The rst numerical implementation of the variational fracture mechanics
is reported in Bourdin et al. (2000). Further developments and applications may be found
in Bourdin et al. (2008), Chambolle et al. (2009), Del Piero et al. (2007), Lancioni & Royer-
Carfagni (2009), Amor et al. (2009), Freddi & Royer-Carfagni (2010). To our knowledge, no
precise quantitative comparison between this approach and experiments has been done so far.
One aim of this paper is to ll this gap.
For this, we focus on the complex cracking phenomena encountered in directional drying
experiments of colloidal suspensions conned in capillary tubes. During drying, the suspen-
sion gradually transforms into a drained porous solid matrix. Further drying induces a natural
shrinking of the solid matrix. In capillary cells, the shrinking is prevented by the strong adhe-
sion to the wall of the cell, and gives rise to high tensile stresses in the matrix. These stresses
are at the origin of complex crack patterns, whose shapes depend on the geometry of the cell
and the drying conditions. This kind of experiments has been proposed rst in at rectangular
capillary tubes (Allain & Limat 1995, Dufresne et al. 2003, 2006) and later on, in circular
ones (Gauthier et al. 2007, 2010). They are extended here to squared cell shapes. Crack prop-
agation in at specimens has been extensively studied by the traditional approach for instance
by Ba zant et al. (1979), Nemat-Nasser et al. (1980), Hofmann et al. (2006), Bahr et al. (2010).
Gauthier et al. (2007, 2010) recently showed that the observed crack patterns can be correctly
predicted by energy minimisation amongst a given family of cracks, namely arrays of paral-
lel cracks or star shaped cracks. Here, our aim is to compare the fracture patterns observed
experimentally to those found by numerical energy minimisation according to the variational
approach to fracture introduced above.
We show that the regularized form of the variational approach is able to predict the shape
of the crack patterns as a function of the cell shape, without any a priori hypotheses.
The outline of the paper is as follows. In Section 2, the experimental setup is described.
2
Various crack shapes are obtained by changing the suspension, the cell geometry and the
drying velocities. In the aim to predict the crack shape, we model (Section 3) the experiments
by a 2D linear elasticity problem in the cross-section, the drying loading being given by a
tensile isotropic inelastic strain. Then we suppose that the fracture conguration for a given
loading, material and cell geometry minimises the sum of the elastic and fracture energies.
This problem is rst solved by restricting the crack shapes to radial cracks and searching their
number that minimises the total energy (Section 4). The question then is to show if star-shape
cracks are energetically optimal. To answer it, the minimisation is performed numerically
using the regularizated form of the variational approach mentioned above. In Section 5, the
principle of this method is described and the simulation results are given and shown to be in
agreement with the direct minimisation for not too high loadings. For higher loadings more
complex crack shapes seems to be preferred to radial cracks. The close agreement with the
experiments is shown in Section 7. This demonstrates that (i) this simple 2D elastic model
captures the physics of the fracture in suspension drying phenomenon and (ii) the efciency
of the variational approach to predict complex crack morphologies without any preliminar
assumption on the shape.
2 Experiments
Fractured gel
Air satured with water
Suspension
Evaporation surface C
r
a
c
k
i
n
g

a
n
d

d
r
y
i
n
g

d
i
r
e
c
t
i
o
n
Crosssection view
(a) Experimental setup. (b) Pictures of some
cross section cuts (the
colors depend on the
light used).
Figure 1: A vertical glass capillary is lled with a colloidal suspension; the single bottom open
edge allows for evaporation of the water in a surrounding maintained at a constant relative humidity
(RH) and temperature (T). The cross-sectional shape of the cracks depend on the cell shape and
size and on the drying conditions.
The experiments are similar to those of Gauthier et al. (2007, 2010). They are carried
out using aqueous suspensions of mono disperse silica spherical particles (Ludox

SM30 of
3
radius r 3.5 nm or Ludox

HS40 of radius r 6 nm) and volume fraction 0.3.

To investigate unidirectional drying, vertical circular or square glass capillary tubes are used
here. The radius of the cross-section of the circular cells are R = 0.05 mm, R = 0.15 mm,
R = 0.5 mm or R = 0.75 mm; the edge length of the square cells is 2R = 1 mm. The lengths
are given with a precision of 10 %.
The tube is lled with the colloidal suspension (Figure 1(a)) and placed in a controlled
environment maintained at a constant relative humidity (RH) and temperature T. Experiments
are performed either:
1. at room temperature T 20C and relative humidity maintained below 10% using
desiccant;
2. at room temperature and RH kept over 90% by introducing water in the chamber;
3. at fridge temperature of T 3Cand relative humidity kept below10% using desiccant.
The top is closed and the bottom is open allowing the water to evaporate. The tube is only par-
tially lled with the suspension, so that the air and water vapor located above the suspension
can expand to compensate the loss of water during desiccation. More technical details about
the experiments can be found in Gauthier et al. (2007, 2010).
As the sample loses water, particles aggregating at the open edge form a growing drained
solid porous medium (Figure 1(a)). High negative capillary pressure in the draining uid
generates high tensile stresses in the gel (Dufresne et al. 2003). This causes crack formation
along the drying direction. Their sectional shape can be visualized either by transparency with
a camera or by cutting a tube to see the cross-section. The later manipulation is delicate and
doesnt work for very small tubes. Some pictures are given in Figure 1(b). We observe that
the crack shape for a given suspension and tube geometry depends on the drying conditions
through the drying velocity only. Indeed drying at T 3 C, RH 10 % or at T 20
C, RH 90 % gives the same crack tip velocities and the same crack patterns. Thus in the
sequel, experiments performed at T 3 C, RH 10 % or at T 20 C and RH 90 %
will be gathered together under slow velocity (SV) experiments and those performed at T
20C and RH 10 % under higher velocity (HV) experiments.
A summary of the shapes obtained can be found in Figure 1(b). For square cells, two
perpendicular cracks appear along the diagonals of the square. For circular cells, the cross-
section forms mostly a star-shape crack breaking up the circular cross section in n circular
sectors (with a central angle 2/n). The number of sectors n is observed to increase with the
cell size R and the drying velocities. A summary of the panel of experiments made can be
found in table 3 (section 7).
3 Model
3.1 Basic hypotheses
Our aim is to predict for a given tube geometry, suspension and drying condition, the shape of
the cracks appearing in the tube. For this matter, we make the following simplifying hypothe-
ses:
H1. The solid adheres to the cell walls, and cracks cannot grow along the walls. This is
required to create the tensile stresses that lead to crack formation.
H2. The solid medium behaves as a linear isotropic homogeneous material. This is reason-
able since the porous matrix is formed by relatively hard silica particles.
H3. Changes in material properties induced by the drying phenomenon are negligible.
H4. The drying phenomenon induces isotropic inelastic strains
0
=
0
1, where 1 denotes
the 3 3 identity matrix and
0
< 0. The inelastic strains are taken as independent of
the deformation in the material, according to a simplied view of the full poroelastic
problem (Wang 2000).
Concerning the fractures, we make the following hypotheses:
4
H5. The material is perfectly brittle.
H6. Crack surfaces are stress-free. This is reasonable since it has been shown (Dufresne
et al. 2003) that the crack surfaces are dry.
H7. Growing cracks have no impact on the value of the inelastic strain
0
. This allows us
to rst determine
0
independently of the crack conguration, then to solve the crack
propagation problem.
We focus on situations where the cracks grow only along the z-axis and their cross-
sectional geometry remains unchanged (Figure 1(a)). Instead of solving the full three-dimensional
evolution problem, for any given loading and geometry we look for the cross-sectional crack
morphology as a two-dimensional static problem. Since the solid is perfectly bounded to
the wall, the reduced two-dimensional model is derived under the plane strain assumption.
Following the variational approach to brittle fracture (Francfort & Marigo 1998):
H8. For a given loading, the deformation state and fracture conguration correspond to a
minimum of a total energy dened by the sum of the bulk elastic energy and a fracture
energy.
H9. The fracture energy is of Grifths type. The energy S() per height unit, associated to
a crack set is proportional to its length given by:
S() := G
c
L(), (1)
where G
c
is the fracture energy of the material, and L denotes the length of the crack.
We invite the reader more interested in the physics of the drying and fracturation of col-
loidal suspensions than in the technicalities of the model and its numerical implementation to
switch directly to Section 6 for a presentation of the numerical simulations and to Section 7
for a comparison with experiments.
3.2 Variational fracture model
Let us introduce the following equivalent 2d inelastic strain dened by
2d
0
= (1 + )
0
1
2
,
where 1
2
is the 2 2 identity matrix. With this notation and the aforementioned hypotheses,
the strain energy can be written under the following the form:
w(,
0
) :=
E
2(1 +)
_

(1 2)
tr
2
(
2d
0
) + (
2d
0
) (
2d
0
)
_
(2)
where E and are the Young modulus and the Poisson ratio of the material, is the symmet-
ric second-order 2 2 matrix representing the linearized plane strain and tr denote the trace
operator, and the dot is used for the scalar product. In linear elasticity, kinematical compati-
bility implies that (u) =
1
2
(u + u
T
), where u is the displacement eld, the gradient
operator, and the superscript T denotes the transpose operator. We parametrise the inelastic
strain
2d
0
representing the drying loading by a non dimensional drying intensity , dened by
(see the rst remark at the end of this section)

2d
0
:=
_
G
c
ER
1
2
(3)
where Ris a characteristic length associated with the cross-section, typically its radius. Hence,
the potential energy P

of the cross-section occupying the open set C and associated to a
displacement eld u and a crack set for a loading parameter is given by
P

(u, ) :=
_
C/
w

((u))dS, with w

() := w
_
,
_
G
c
/ER1
2
_
. (4)
The total energy is dened as the sum of the potential energy and the surface energy required
to create the cracks:
E

(u, ) := P

(u, ) +S(). (5)
5
For any given loading parameter , we seek to nd the crack set and displacement eld u
as the global minimiser of (5) amongst any admissible crack set and kinematically admissible
displacement elds. The admissible crack sets consist of all possible curves or sets of curves
inside the boundary of C. For any given crack set , the space of the admissible displacements
is
U() := {u H
1
(C \ ; R
2
), u = 0 on C}, (6)
i.e. it consists of all vector valued elds satisfying the adhesion boundary condition and suf-
ciently smooth (square integrable with square integrable rst derivatives) on the uncracked
domain.
More precisely, the global minimality condition can be expressed as:
Find C, u U() : E

(u, ) E

(u

),

C, u

U(

). (7)
3.3 Remarks
The model presented above deserves several remarks:
The scaling factor
_
G
c
/ER in (3) renders all the results, presented in terms of in the
rest of the paper, independent of the material constants and cross-sectional dimension.
Other choices for the relevant non dimensional parameter are possible. In particular, as
in Gauthier et al. (2010), one could also chose to parameterize the loading in terms of
the Grifth length L
c
:= EG
c
/
2
0
, where
0
is a prestress. After some calculations,
one can relate and L
c
by
L
c
=
R

2
(1 2)
2
(1 +)
2
. (8)
This relation will be useful in Section 7 as it will allow us to estimate the value of for
various experiments.
We consider the problem of nding the optimal displacement eld and crack pattern in a
cross section of the tube of Figure 1 for a given value of the loading parameter, indepen-
dently of the previous history or irreversibility conditions on preexistent crack patterns.
We refer to this problem as a static formulation of the fracture mechanics problem, in
opposition to the quasi-static setting, where one need to account for the previous his-
tory and the irreversible nature of crack propagation through unilateral minimisation as
in Francfort & Marigo (1998).
Perfect bonding to the walls of the cell is accounted for in the minimisation principle (7)
by imposing null-displacement boundary conditions on u, and that the admissible cracks
be included in the open set C. This in particular proscribes cracks along C.
The admissible displacement elds are potentially discontinuous across cracks , but
the location of the potential discontinuities themselves is not known a priori. This ren-
ders the numerical minimisation of (5) challenging as most numerical methods such
as cohesive, discontinuous or extended nite element methods require at least some a
priori knowledge of the crack path or of its topology. Indeed, this problem falls into
the broader class of free discontinuity problems for which a wealth of mathematical and
numerical literature now exists. In the following, we solely focus on the numerical im-
plementation using an extension of that proposed in Bourdin et al. (2000) and Bourdin
(2007), and inspired from Ambrosio and Tortorellis results on the approximation of
the Mumford-Shah functional by means of elliptic functionals Ambrosio & Tortorelli
(1990, 1992). We refer the reader interested in the analysis of the model to Francfort &
Larsen (2003), Dal Maso et al. (2005), Bourdin et al. (2008), and references within.
4 Simple illustration: star-shaped cracks in circular cells
Although the main strength of the variational approach to fracture is that it does not require any
a priori hypotheses on crack geometry, the following basic computation provides a valuable
6
insight into our approach. It is essentially equivalent to that of Gauthier et al. (2010) with
the difference that the loading parameter we consider here is the inelastic strain instead of the
prestress. We consider a circular cell with radius R and star shaped cracks. By
n
, (n > 1),
we denote a curve consisting of the union of n equi-distributed radial segments partitioning
the cell into n polar regions. By analogy, we write
1
= . Our motivation for considering
such geometries comes from the fact that they are frequently observed in the experiments, at
least for small values of n.
Note rst that for a given crack pattern , the potential energy can computed by solving a
linear elasticity problem, and we write
P

() := min
uU()
P

(u, ), (9)
the potential energy of the equilibrium displacement. It is then easy to see that the form of
the strain energy density (2) implies that P

() =
2
P
1
(), so that we can rewrite the total
energy in the form
E

() =
2
P
1
() +S(). (10)
Furthermore, for a star-shaped crack
n
, using eq. (2), one has
E

(
n
) =
2
P
1
(
n
) +nG
c
R. (11)
For n = 1 the problem can be solved in closed form, the elastic equilibrium is achieved for
u = 0 and the total energy is E

() = P

() =
2
G
c
R/(1 + )(1 2). For n > 1,
P
1
(
n
) can be computed by a simple nite element computation. In this setting, for a given
loading parameter , energy minimisation reduces to a discrete minimisation problem with
respect to n. And the total energy of the solution as a function of the loading parameter can
be obtained by taking the lower envelope of the family of energy curves associated to each
conguration. Note that this construction is essentially equivalent to that of the backtracking
algorithm proposed in Bourdin (2007).
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0
2
4
6
8
10
2.15 1.55 0.97 0.73

R
G
c
n1
n2
n3
n4
n5
(a) Total energy associated with
n
for n = 1, 2, 3, 4, 5
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0
2
4
6
8
10
2.15 1.55 0.97 0.73

R
G
c
(b) The lower envelope of the family of energy curves
gives the energy of the minimiser
Figure 2: Energy minimisation amongst star-shaped cracks.
Figure 2(a) represents the total energy associated with
n
as a function of the loading
parameter for n = 1, 2, 3, 4, 5 and = 0.3. Using this graph and the global minimality
principle (7), it is easy to identify the optimal crack conguration associated with a given
load (the branch of the energy with the smallest value at ) as well as the bifurcation points
upon which the geometry of the optimal crack set changes (the crossing points upon which
the energy branch achieving minimality changes). We obtain that there exists a family (0 =

0
,
1
,
2
, ) of critical loadings such that for
i1
< <
i
, i = 1, 2, , the optimal crack
conguration is any curve in the family
i
. Of course, in the absence of defects or impurities,
7
the solution for a given loading parameter is unique up to a rotation. The numerical values
of the critical loadings are
1
0.73,
2
0.97,
3
1.55,
4
2.15, and
5
2.76.
Figure 2(b) shows the energy associated with the optimal conguration, obtained by taking
the lower envelope of the family of curves plotted in the left.
Before closing this simple example, we stress again that this analysis is based upon the
assumption that the optimal crack geometry is a star-shaped pattern. As we will see in the fol-
lowing sections, relaxing this hypothesis allows to show that the star-shape cracks are indeed
optimum for not too high loading and can yield to more complicated but energetically least
costly crack patterns for higher loading.
5 General case: minimisation over arbitrary crack geome-
tries
The numerical implementation we use is essentially similar to the one described in Bourdin
et al. (2000, 2008), Bourdin (2007). We briey recall its properties here and refer the reader
to the literature for further details.
5.1 Regularization by elliptic functionals
The main idea of our approach was originally developed by Ambrosio & Tortorelli (1992)
for an image segmentation (Mumford & Shah 1989) problem, and adapted to fracture me-
chanics by Bourdin et al. (2000). One introduces a small parameter with dimension of a
length, a secondary variable taking its values in [0, 1] and representing the crack set, and the
regularized functional
E
()

(u, ) :=
_
C
((1 )
2
+k

)w

((u))dS +
3G
c
8
_
C
_

+
_
dS. (12)
Hence, one approximates the solution of (7) by those of the following minimisation problem:
min
uU,A
E
()

(u, ) (13)
where U = {u H
1
(C; R
2
), u = 0 on C} and A = {0 1, H
1
(C; R), =
0 on C}.
This regularized functional can be shown to converge in the sense of convergence to
the total energy (5). This implies that for any , the minimisers of E
()

converges as 0
to minimisers of E

, and that each term in (12) converges to the matching one in (5). The
parameter k

is a residual stiffness introduced mainly for numerical purposes which is known

to have very little impact on the minimisers. The convergence result is valid provided that k

=
o(). We refer the interested reader to Braides (2002), Dal Maso (1993) for more details on
convergence and to Braides (1998), Bourdin et al. (2008) for details on the approximation
of E

by E
()

. Formally, as goes to 0, remains close to 0 away from the cracks, and
approaches 1 along the cracks. For small but non-zero values of , both arguments u and of
E
()

are continuous functions with high gradients (of order 1/) in bands of width .
From a mechanical point of view, the functional (12) can be interpreted as the energy
functional of a non-local gradient damage model, where plays the role of the damage eld
and of the internal length (Pham, Amor, Marigo & Maurini 2011). We also point out that
the regularized energy we use here is slightly different from the one in Bourdin et al. (2000)
and Bourdin et al. (2008). The form used here has some advantages from a numerical and the-
oretical standpoints, which are not discussed here. The interested reader is referred to Pham,
Amor, Marigo & Maurini (2011) for further details on this point. Despite their apparent sim-
ilarities, there are signicant differences between our numerical approach and the phase-eld
fracture models (see e.g. Hakim & Karma 2009). The regularized energy (12) resembles a
phase-eld equation. However, even in the static or quasi-static case, phase-eld models are
formulated as a rate-dependent evolution equation and stated as fracture models per se. Here,
8
we do not see our regularized formulation as a fracture model, but merely as a numerical
approximation of the total energy of the static or quasi-static variational approach of Franc-
fort & Marigo (1998). This approximation is deeply rooted in the mathematical literature on
free-discontinuity problems (Braides 1998). In particular the minimisation principle for the
regularized energy is derived from that of the variational model in the static case.
5.2 Numerical implementation
The numerical minimisation of (12) is implemented in a way similar to that described in Bour-
din (2007). We discretize the regularized energy by means of linear Lagrange nite ele-
ments over an unstructured mesh. As long as the mesh size h is such that h = o(), the
convergence property of (12) to (5) is also true for the discretization of the regularized en-
ergy (see Bellettini & Coscia (1994), Bourdin (1999), Burke et al. (2010) for instance). This
compatibility condition leads to ne meshes, which are better dealt with using parallel super-
computers. We use PETSc (Balay et al. 1997, 2010, 2011) for data distribution, parallel linear
algebra, and TAO (Benson et al. 2010) for the constrained optimization. In order to avoid
prefered directions in the mesh, we use the Delaunay-Voronoi mesh algorithm implemented
in Cubit, from Sandia National Laboratories.
Due to the size of the problems, global minimisation algorithms are not practical. Instead,
we notice that although (12) is non convex, it is convex with respect to each variable indi-
vidually. We alternate minimisations with respect to u and , an algorithm akin to a block
Newton method or a segregated solver. Note that minimisation with respect to u is equivalent
to solving a simple linear elasticity problem, but that minimisation with respect to [0, 1]
requires an actual box-constrained minimisation algorithm. Of course, as the total energy is
non convex, one cannot expect convergence to a global minimiser. However, one can prove
that the alternate minimisation process is unconditionally stable and globally decreasing and
that it leads to a stationary point of (12) which may be a local (or global) minimiser or a
saddle point of the energy. From a practical standpoint we observe that the algorithm is quite
robust with respect to the mesh discretization, provided that the regularization length is large
enough compared to the mesh size. However it can be quite sensitive with respect to the initial
value of u and . Different choices of the initial guess or of the regularization parameter
can lead to convergence to different solutions
In the following section, we present numerical experiments performed using the method
as described above, highlight its shortcomings and illustrate how to decrease the sensitivity
with respect to the choice of initial conguration and regularization parameter.
6 Numerical simulations
6.1 Selection of crack shapes
We rst illustrate our numerical approach on circular cells. Experimental evidence suggests
that for small values of , the actual fracture pattern resembles the star-shaped cracks from
Section 4. We can use this feature to perform partial verication of our numerical approach.
As dimensional analysis highlights the dependency of the fracture energy (5) on a single
loading parameter, , it is natural to replace the regularization length with a non-dimensional
parameter

= /R.
Figure 3 presents the eld obtained by numerical minimisation of (12) for various
choices of the parameters and

with a cell of radius R = 1. The material parameters (E,
G
c
) were set to 1 without loss of generality, and the Poisson ratio to 0.3. In each computation,
the mesh size was h = 0.025 and the residual stiffness was set to k

= 10
6
. The alternate
minimisation algorithm was initialized with = 0, u = 0. The value 1 (corresponding to
cracks) of is encoded in red and the value 0 (the un-cracked material) in blue. A rst glance
at the table highlights the wide variety of crack geometries obtained, and that the complexity
of the fracture pattern increases with the loading parameter. Again we stress that no hypothe-
sis on this geometry is made in the model and that the shape of the crack patterns is purely an
9

0.2 0.1 0.05

0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
2.5
3.0
4.5
5.0
(a) Direct numerical simulations. Each prob-
lem was solved independently initializing the
alternate minimizations algorithm with the un-
cracked solution = 0.

0.2 0.1 0.05

0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
2.2
2.5
3.0
4.5
5.0
(b) Numerical results obtained using

continuation. Each row corresponds to a set of

computations, each taking the one at its left as a
rst guess for u and .
Figure 3: Numerical results by minimisation of functional (12) for circular cross-sections. The
results are obtained using uniform Delaunay-Voronoi unstructured meshes with size h = 0.025 on
disk of radius R = 1. The material properties are E = 1, G
c
= 1, = 0.3.
10
outcome of the minimisation of the regularized energy. We observe that the width of the tran-
sition zone from 1 to 0 decreases as

goes to 0, which is consistent with the convergence
property stated in Section 5. For large values of

, when the width of the transition zone
is of the order of the diameter of the cell and as increases, one cannot distinguish between
neighboring cracks (see for instance the case

= 0.2 for = 5.0).
A closer look at Figure 3(a) highlights the dependency of the crack pattern upon the reg-
ularization parameter

. See for instance how for a loading parameter = 1.2, we obtain a
triple junction for

= 0.2, but a complex crack made of two triple junctions for

= 0.1, and
no cracks at all for

= 0.05. Each of these conguration correspond to a critical point of the
energy (12) (likely local minimisers). Indeed, below a critical load

) depending on

, the
un-cracked conguration = 0 is a stable critical point of (12). In a simpler 1d setting, it is
known that

) = O(1/

) (Pham, Marigo & Maurini 2011). Here, we observe that

)
increases as

0. When alternate minimisations iterates escape the un-cracked solution,
they converge to the nearest critical point which may or may not be the global minimiser of
the energy.
If the regularized model (12) is seen as a gradient damage model with internal length

(see
Pham, Amor, Marigo & Maurini 2011, Pham, Marigo & Maurini 2011), and if one focusses
on criticality instead of global minimality, this behavior is consistent with a scale effect linking
the critical load and the ratio of the structural dimension and the internal length (Ba zant 1999).
In the present study, we stand by the interpretation that the regularization length is an articial
numerical parameter and we seek global minimisers of the regularized energy. We avoid

dependency of our solution by adopting a

continuation technique. For a given loading
and mesh size h, we minimise the total energy E
(

)
for decreasing values of the regularization
length, initializing each computation with the eld obtained at the previous one. Figure 3(b)
represents the outcome of such a series of computations highlighting, how the continuation
approach eliminates the sensitivity of the numerical result on the regularization length

and
allows us to retrieve an accurate representation of the crack geometry without the need for
extremely ne meshes. From a qualitative standpoint, one observes that the crack geometry
remains unchanged as the regularization length is decreased. Table 1 provides a quantitative
comparison of the quality of the solutions. In each case, the congurations obtained using

continuation (last column of Figure 3(b)) have a lesser energy than the one obtained through a
direct computation (Figure 3(a)). Since we are interested in energy minimising conguration,
these are the ones we retain. To obtain accurate values of the fracture energies with this choice
of h = O(

), and following the discussion in Bourdin et al. (2008) (Sec 8.1.1 p.103), all the
results are obtained by replacing the fracture toughness by G
(num)
c
= G
c
/(1 + 3h/8

). This
is especially important when comparing multiple computations obtained with different ratios
h/

.
It is possible to further select the optimal crack geometry as a function on the basis of
the global minimality principle (13) by using a method similar to that in Section 4. Noticing
that for a given -eld (i.e. crack geometry) the elastic part of the total energy (12) scales
quadratically with the loading, one can calculate the total energy that each of the crack patterns
obtained numerically for a given loading

would have for any . Figure 4(a) represents

the total energy obtained in this way for each of the crack patterns in the last column of
Figure 3(b). From this gure, for each , is possible to select the best crack pattern as the one
with the lowest energy level. Amongst all the curves, the ones attaining the lowest energy for
some value of the loading parameter are plotted in black and thicker line width. Their lower
envelope is the continuous black line reported in Figure 4(b), together with the associated
optimal crack geometries.
Although it is of course never possible to prove global optimality, the crack geometries
depicted in Figure 3(b) are the lowest energy congurations we were able to attain, and the
ones which we will compare with star-shaped cracks and experiments in the sequel.
11
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0
2
4
6
8
10
2.65 2.01 1.51 0.96 0.71

R
G
c
(a) Energies associated with the crack geometries identied in Figure 3(b) as a func-
tion of the loading parameter. Thick black lines distinguishes energy curves corre-
sponding to congurations attaining the minimal energy for some value of the loading
parameters. The gridlines marks the critical loading for the passage of one optimal
curve to the next. Note that for large the identication of critical loadings becomes
difcult.
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0
2
4
6
8
10
2.65 2.01 1.51 0.96 0.71

R
G
c
(b) Comparison of the optimal energy obtained using minimisation over star shaped
cracks (dashed line) and numerical simulation (continuous line). The continuous line
is the lower envelope of the curves in gure 4(a) . The pictures represent the optimal
crack shapes in each range of the loading parameter delimited by the gridlines.
Figure 4: Minimisation over star-shaped cracks vs. minimisation of functional (12). As the loading
increases, our numerical method identies crack congurations with much lower energetically than
star-shaped cracks.
12
Table 1: Energies of the numerical solutions in Figure 3 for

= 0.05 without (left) or with (right)

continuation.
Elastic Surface Total
0.6 2.2 0 2.2
0.8 3.9 0 3.9
1.0 6.0 0 6.0
1.2 8.7 0 8.7
1.4 1.6 3.9 5.5
1.6 2.0 4.0 6.0
1.8 1.9 4.9 6.8
2.0 2.0 5.3 7.3
2.2 2.1 5.7 7.8
2.5 2.1 6.4 8.5
3.0 2.1 7.3 9.4
4.5 2.3 9.6 11.8
5.0 2.4 9.8 12.2
Elastic Surface Total
0.6 2.2 0 2.2
0.8 1.5 1.8 3.4
1.0 1.3 2.8 4.1
1.2 1.7 2.9 4.6
1.4 2.0 3.3 5.3
1.6 2.0 3.9 5.9
1.8 2.0 4.4 6.4
2.0 2.0 4.9 6.9
2.2 2.0 5.4 7.4
2.5 2.1 5.9 8.0
3.0 2.1 6.9 9.1
4.5 2.3 9.2 11.5
5.0 2.4 9.9 12.3
6.2 Comparison with star-shaped cracks
Figure 4(b) compares the total energy of these congurations with the energy of the star-
shaped cracks taken from Figure 2 (dashed line). For small values of the loading parameter,
we obtain similar geometries and critical loading. The surface energy obtained is close to the
number of branches, and the critical loading upon which we obtain a single straight crack is
0.71 (vs. a theoretical value of
2
= 0.73). Bifurcation between straight and Y-shaped cracks
take place at 0.94 (vs. a theoretical value of
2
= 0.97).
More interestingly, for larger values of

, our numerical simulations have identied mul-
tiple congurations that are energetically close to each other but always less expensive than
star-shaped cracks. In particular, we show that perfect 5-branch stars are never optimal and
that congurations consisting of either two triple junctions very close to each others (see for
instance = 1.6 in Figure 3(b)), a 4-branch star whose branches split in two near the cell
boundary (see for instance = 1.8, 2.0, 2.2 in Figure 3(b)), or a more complicated patterns
like the stick gure looking 5 cracks conguration that we obtain for = 2.5 have lesser
energy. Of course, that the local geometry near the crack crossing resembles 2 triple junctions
near each others rather that an X does not really come up as a surprise. As mentioned earlier,
the fracture energy (5) resembles the Mumford-Shah energy for edge segmentation (Mumford
& Shah 1989). Therefore, it seems natural to expect that if they posses some form of regu-
larity, optimal crack geometries satisfy the Mumford-Shah conjecture which rules out crack
crossing, kinks and only allows cracks to meet at 120

triple junctions, locally.

These quantitative and qualitative properties of the numerical solution provide a powerful
illustration of the strength of the variational approach and our numerical methods by high-
lighting that global minimisation over unrestricted crack shapes can lead to congurations
that are energetically favored over simple ones. In this context, the choice of the family of
cracks one has to minimize the total energy over in classical approaches is neither trivial nor
innocent.
6.3 Square cells
Finally, we performed another set of numerical simulations on unit square cells. The results
obtained by the same

-continuation method, as in Figure 3(b) for circular cells, are depicted
on Figure 5. The materials parameters and mesh size are unchanged (E=1, G
c
=1, = 1,
h = 0.025). Whereas for circular cells, starshaped cracks are natural candidates, there were
13

0.2 0.1 0.05

0.6
0.8
1.0
1.2
1.4
1.6

0.2 0.1 0.05

1.8
2.0
2.2
2.5
3.0
4.5
Figure 5: Numerical results for square cells by minimisation of functional (12) using the

-
continuation method.
no obvious family of cracks in this case. This geometry also leads to a rich variety of crack
patterns and highlights the strength of the proposed method in identifying complex crack pat-
terns without a priori hypothesis. Some of the quantitative properties of the optimal cracks
highlighted in the case of circular cells are still observed. Again, cracks seem to split near
the edges of the cells. Triple junctions seem to be favored over crack crossing, although in
the case of two diagonal or longitudinal cracks, the resolution of our numerical experiments
does not allow us to clearly identify the conguration. As for the circular cell, one can further
post process the numerical result in order to identify the range of loadings for which each
of the identied conguration is optimal. This is presented in Figure 6. Again, for small
values of the loading parameters, simple and somewhat predictable crack geometries are ob-
tained. For larger values of , more complex and less intuitive patterns are energetically more
advantageous.
7 Comparison between experimental and numerical results
7.1 Identication of the loading parameter
Dimensional analysis shows that the model relies on a single parameter, the value of which
needs to be estimated in order to perform quantitative comparison between experiments and
numerical simulation. As depends on experimental conditions, colloidal suspension type,
and tube geometry, one solution is to try to measure separately
0
, E, and G
c
appearing in the
denition (3) of . One may obtain the material constants E, G
c
by indentation (Malzbender
et al. 2002) and the mismatch strain
0
by beam deection technics (Tirumkudulu & Russel
2004, Chekchaki et al. 2011) from a thin lm drying experiments, for instance. However,
such direct measurements are difcult, and transposing the values obtained from one type
of experiments (thin lm drying) to another (directional drying) is questionable. Indeed, the
parameters may depend on the type of experiments and even evolve in time. For example,
14
0.0 0.5 1.0 1.5 2.0 2.5 3.0
0
2
4
6
8
10
2.18 1.6 1.38 1.0 0.68

R
G
c
Figure 6: Range of parameters in which each of the conguration identied in Figure 5 is optimal
the material constants E and G
c
of the porous medium may depend on the microstructure,
inuenced by formation dynamics.
Instead of performing such difcult measurements, whose relevance to our problem may
be questioned, we used the method presented in Gauthier et al. (2010), which we briey
summarize. The basis of the method is to consider a directional drying experiment in thin
rectangular cells (Allain & Limat 1995). In this geometry, an array of parallel tunneling cracks
is obtained and the cracks spacing can be correlated to the Grifth length L
c
:= EG
c
/
2
0
(
0
being the prestress induced by the lms drying). Using an energy minimisation principle
similar to the one in Section 4, one can show that the spacing is proportional to

L
c
t, t
being the cells thickness and in particular, for = 0.3, one obtains 3.1

L
c
t. For a given
material and drying parameter, the value of L
c
can therefore be deduced from measurements
of . Table 2 presents the value of L
c
for Ludox

SM30 (r 3.5 nm) and Ludox

HS40
(r 6 nm) under high velocity and slow velocity conditions.
L
c
Ludox

SM30 (r 3.5 nm) Ludox

HS40 (r 6 nm)
HV 34 10 40 10
SV 60 18 45 15
Table 2: Values of Grifths length L
c
(in m) for several Ludox

suspensions and drying rates

(SV=[T 3C and RH 10% or at T 20C and RH 90%] and HV=[T 20C and RH
10%])
We assume that the Grifth length L
c
is a well-dened material parameter for a given
suspension and drying condition, and that it is independent of the cross-sectional geometry of
the directional drying experiment. Hence, from the values of L
c
in Table 2, we estimate the
value of in the directional drying of circular and square cells of different diameters using the
relation (8), which gives = 0.52
_
R/L
c
for = 0.3.
7.2 Results and analysis
Table 3 reports on the series of experiments on circular cells described in Section 2. From
a qualitative standpoint we observe that star shaped appear above a critical load, and that
the number of branches increases with the loading, which is consistent with the analysis in
Section 4 and the numerical simulations of Section 6.
15
Ludox

HV/SV L
c
(m) R (m) n
a SM30 HV 34 10 50 5 1 0.6 0.12
b SM30 SV 60 18 150 15 2 0.8 0.16
c SM30 SV 60 18 500 50 3 1.5 0.30
d HS40 SV 45 15 500 50 4 1.7 0.37
e HS40 HV 40 10 500 50 4 1.8 0.27
f SM30 HV 34 10 500 50 4 2.0 0.39
g HS40 HV 40 10 750 75 5 2.3 0.41
h SM30 HV 34 10 750 75 5 2.4 0.48
Table 3: Experimental results on circular cells. The value of n corresponds to the number of
sectors delimited by the cracks as in Section 4.
In order to perform a quantitative comparison, we summarize all the results obtained in
the case of circular cells in Figure 7. The rst row corresponds to the outcome of the semi-
analytical minimization over star shaped cracks. The critical values of the loading parameters
computed in Section 4 upon which bifurcation between different morphologies take place
is printed in red letters, and represented by red dashed vertical lines. The second row cor-
responds to the numerical experiments without a priori hypotheses on the crack path. The
critical loads extracted from Figure 4(a) are printed in black letters, and represented by ver-
tical solid black lines. As highlighted in Figure 4(b), the critical loads obtained in the case
of the bifurcation from
1
to
2
, then
3
and
4
are very close. This part of the table can be
seen as a verication of the numerical implementation, i.e. as evidences that the computed
solutions are indeed solution of the variational fracture model. The third row summarizes the
outcome of the experiments from Table 3. For each experiment, the value of the loading pa-
rameter is shown together with the accuracy of the measurement. When available, photos of
the cross sections are also displayed. We observe that for every single choice of , the crack
geometry predicted by our approach matches the one observed in the experiment. This acts
as a validation of the variational fracture model as a predictive tool in the setting of drying of
colloidal suspension.
a n1
b n2 c n3
d n4
e n4
f n4
g n5
h n5
0.71 0.96 1.51 2.01
0.73 0.97 1.55 2.15
Figure 7: Comparison between semi-analytical, numerical and experimental results for circular
cells.
We also did a single experiment on a square cell, for an estimated value 1.8 of the
16
loading parameter for which we obtained two diagonal cracks (see the bottom image in Fig-
ure 1b). Again, the numerical simulation in this case matches the experiment (see Figures 5
and 6).
Despite the modeling simplifying assumption, the complexity of the numerical technique,
and the uncertainty of the measurement of the parameters, the agreement between analysis,
simulation, and experiments is excellent. Our model correctly captures the essential physics
of the crack formation giving credit to the idea that crack growth can be predicted by minimi-
sation of the sum of elastic and surface energy over all possible crack path. In order to further
justify this idea, one will need to compare experiments and simulations for higher loading
parameters in which case numerical simulations identify complex crack patterns with signif-
icantly lower energy than classical star-shaped solutions. For instance, better quality imaging
will be required to unambiguously determine if the the 5 cracks congurations we observe ex-
perimentally (see Figures 1(b) or 7) resembles a stick gure as in our numerical simulation
(cf. Figure 3 for = 2.5), a regular 5-branch star, or something completely different.
8 Conclusions and future work
In this paper, we have presented some directional drying experiments of colloidal suspensions
realized in capillary cells where solvent evaporation leads to the formation of a growing porous
solid media. Due to shrinkage prevented by adhesion on the wall of the cells, high tensile
stresses appear that give rise to cracks of various morphologies depending on the cell geometry
and the drying velocities.
We proposed a simple model based on the assumption that when the crack cross sec-
tional geometry does not evolve, this problem can be reduced to a 2d static one. We showed
that changes in crack geometry due to different materials and experimental condition can be
accounted for by a single dimensionless parameter which can represent the intensity of the
tensile strain or stress induced by drying or the energetic cost of growing cracks. We used
the variational approach to fracture to account for crack propagation in the porous medium as
a function of this parameter without any underlying assumptions on crack geometry. Under
additional assumptions on the crack geometry, we computed the range of loading parameters
for which optimality may be achieved by star-shaped cracks. We then presented a numeri-
cal method and some simulations allowing us to predict crack patterns without any a priori
knowledge. Finally, we achieve qualitative and quantitative agreement between numerical
simulation, semi-analytical solutions and experiments.
At this point, though, we were not able to perform qualitative or quantitative comparisons
for higher values of the loading parameters, where the virtue of the variational approach to
fracture over more conventional ones requiring at least some a priori knowledge of the crack
path becomes more striking. Such experiments will require additional work to deal with larger
cells for instance. In these situations, the main difculty is the post-mortem analysis of the
crack geometry. Microphotography though the sides of the cells becomes hard to interpret,
and cutting the tubes without perturbing the cracks geometry is difcult. Perhaps the solution
lies in full 3d imaging of the cells and post-processing in order to highlight the location of
the cracks. From the modeling perspective, for larger cell size, a full three-dimensional linear
poroelasticity model initiated by Biot (1941) will become necessary. Finally, from a physico-
chemical point of view, the link between the drying velocity and the macroscopic signature
will have to be explored.
Acknowledgements
The work of V. Lazarus and G. Gauthier was partially supported by the ANR Program JC-
JC ANR-05-JCJC-0029 Morphologies. C. Maurini gratefully acknowledges the funding of
the French National Research Council (CNRS) for a PICS bilateral exchange program with
B. Bourdin and a grant of the University Pierre et Marie Curie EMERGENCES-UPMC.
17
B. Bourdins work was supported in part by the National Science Foundation under the grant
DMS-0909267. Some of the numerical experiments were performed using the National Sci-
ence Foundation TeraGrid resources (Grandinetti 2007) provided by TACC at the University
of Texas under the Teragrid Resource Allocation TG-DMS060014N and the resources of the
Institut du D eveloppement et des Ressources en Informatique Scientique (IDRIS) under the
DARI 2011 allocation 100064.
References
Allain, C. & Limat, L. (1995), Regular patterns of cracks formed by directional drying of a
colloidal suspension, Phy. Rev. Lett. 74, 29812984.
Ambrosio, L., Fusco, N. & Pallara, D. (2000), Functions of bounded variation and free dis-
continuity problems, Oxford Mathematical Monographs, Oxford Science Publications.
Ambrosio, L. & Tortorelli, V. (1990), Approximation of functionals depending on jumps by
elliptic functionals via -convergence, Comm. Pure Appl. Math. 43(8), 9991036.
Ambrosio, L. & Tortorelli, V. (1992), On the approximation of free discontinuity problems,
Boll. Un. Mat. Ital. B (7) 6(1), 105123.
Amor, H., Marigo, J.-J. & Maurini, C. (2009), Regularized formulation of the variational
brittle fracture with unilateral contact: Numerical experiments, J. Mech. Phys. Solids
57(8), 1209 1229.
Bahr, H., Weiss, H., Bahr, U., Hofmann, M., Fischer, G., Lampenscherf, S. & Balke, H.
(2010), Scaling behavior of thermal shock crack patterns and tunneling cracks driven by
cooling or drying, J. Mech. Phys. Solids 58(9), 1411 1421.
Balay, S., Brown, J., Buschelman, K., Eijkhout, V., Gropp, W., Kaushik, D., Knepley, M.,
Curfman McInnes, L., Smith, B. & Zhang, H. (2010), PETSc users manual, Technical
Report ANL-95/11 - Revision 3.1, Argonne National Laboratory.
Balay, S., Brown, J., Buschelman, K., Gropp, W., Kaushik, D., Knepley, M., McInnes, L. C.,
Smith, B. & Zhang, H. (2011), PETSc Web page.
URL: http://www.mcs.anl.gov/petsc
Balay, S., Gropp, W., Curfman McInnes, L. & Smith, B. (1997), Efcient management of
parallelism in object oriented numerical software libraries, in E. Arge, A. M. Bruaset &
H. P. Langtangen, eds, Modern Software Tools in Scientic Computing, Birkh auser Press,
pp. 163202.
Ba zant, Z. (1999), Size effect on structural strength: a review., Arch. Appl. Mech. 69, 703
725.
Ba zant, Z., Ohtsubo, H. & Aoh, K. (1979), Stability and post-critical growth of a system of
cooling or shrinkage cracks, Int. J. Fracture 15(5), 443456.
Bellettini, G. & Coscia, A. (1994), Discrete approximation of a free discontinuity problem,
Numer. Funct. Anal. Optim. 15(3-4), 201224.
Benson, S., Curfman McInnes, L., Mor e, J., Munson, T. & Sarich, J. (2010), TAO user man-
ual (revision 1.10.1), Technical Report ANL/MCS-TM-242, Mathematics and Computer
Science Division, Argonne National Laboratory.
URL: http://www.mcs.anl.gov/tao
Biot, M. A. (1941), General theory of 3-dimensional consolidation, J. Appl. Phys. 12, 155
164.
18
Bourdin, B. (1999), Image segmentation with a nite element method, M2AN Math. Model.
Numer. Anal. 33(2), 229244.
Bourdin, B. (2007), Numerical implementation of the variational formulation of quasi-static
brittle fracture, Interfaces Free Bound. 9, 411430.
Bourdin, B., Francfort, G. & Marigo, J. J. (2000), Numerical experiments in revisited brittle
fracture, J. Mech. Phys. Solids 48(4), 797826.
Bourdin, B., Francfort, G. & Marigo, J.-J. (2008), The variational approach to fracture, J.
Elasticity 91(1), 5 148.
Braides, A. (1998), Approximation of Free-Discontinuity Problems, Vol. 1694 of Lecture
Notes in Mathematics, Springer.
Braides, A. (2002), -convergence for beginners, Vol. 22 of Oxford Lecture Series in Mathe-
matics and its Applications, Oxford University Press, Oxford.
Branagan, D. & Cairns, H. (1993), Tessalated pavements in the Sydney region, New South
Wales, J. Proc. Roy. Soc. New South Wales 126(1), 6372.
Burke, S., Ortner, C. & S uli, E. (2010), An adaptive nite element approximation of a varia-
tional model of brittle fracture, SIAM J. Numer. Anal. 48(3), 9801012.
Chambolle, A., Francfort, G. & Marigo, J.-J. (2009), When and how do cracks propagate?,
J. Mech. Phys. Solids 57(9), 1614 1622.
Chekchaki, M., Frelat, J. & Lazarus, V. (2011), Analytical and 3D nite element study of the
deection of an elastic cantilever bilayer plate, J. Appl. Mech.-T. ASME 78(1), 011008.
Colina, H. & Roux, S. (2000), Experimental model of cracking induced by drying shrinkage,
Eur. Phys. J. E 1(2-3), 189194.
Corbett, G., Reid, S. & Johnson, W. (1996), Impact loading of plates and shells by free-ying
projectiles: A review, Int. J. Impact. Eng. 18(2), 141 230.
Corson, F., Adda-Bedia, M., Henry, H. & Katzav, E. (2009), Thermal fracture as a framework
for quasi-static crack propagation, Int. J. Fracture 158(1), 114.
Dal Maso, G. (1993), An introduction to -convergence, Birkh auser, Boston.
Dal Maso, G., Francfort, G. & Toader, R. (2005), Quasistatic crack growth in nonlinear
elasticity, Arch. Ration. Mech. Anal. 176(2), 165225.
DeGraff, J. & Aydin, A. (1987), Surface morphology of columnar joints and its signicance
to mechanics and direction of joint growth, Geol. Soc. Am. Bull. 99, 600617.
Del Piero, G., Lancioni, G. & March, R. (2007), A variational model for fracture mechanics:
Numerical experiments, J. Mech. Phys. Solids 55(12), 25132537.
Dufresne, E., Corwin, E., Greenblatt, N., Ashmore, J., Wang, D., Dinsmore, A., Cheng, J.,
Xie, X., Hutchinson, J. & Weitz, D. (2003), Flow and fracture in drying nanoparticle
suspensions, Phys. Rev. Lett. 91(22), 224501.
Dufresne, E., Stark, D., Greenblatt, N., Cheng, J., Hutchinson, J., Mahadevan, L. & Weitz, D.
(2006), Dynamics of fracture in drying suspensions, Langmuir 22(17), 71447147.
Francfort, G. & Larsen, C. (2003), Existence and convergence for quasi-static evolution in
brittle fracture, Comm. Pure Appl. Math. 56(10), 14651500.
Francfort, G. & Marigo, J.-J. (1998), Revisiting brittle fracture as an energy minimization
problem, J. Mech. Phys. Solids 46, 13191342.
19
Freddi, F. & Royer-Carfagni, G. (2010), Regularized variational theories of fracture: A uni-
ed approach, J. Mech. Phys. Solids 58(8), 1154 1174.
Gauthier, G., Lazarus, V. & Pauchard, L. (2007), Alternating crack propagation during direc-
tional drying, Langmuir 23(9), 47154718.
Gauthier, G., Lazarus, V. & Pauchard, L. (2010), Shrinkage star-shaped cracks: Explaining
the transition from 90 degrees to 120 degrees, EPL 89, 26002.
Grandinetti, L., ed. (2007), TeraGrid: Analysis of Organization, System Architecture, and
Middleware Enabling New Types of Applications, Advances in Parallel Computing, IOS
Press, Amsterdam.
Grifth, A. (1920), The phenomena of rupture and ow in solids, Philos. T. R. Soc. Lon.
221, 163198.
Hakim, V. & Karma, A. (2009), Laws of crack motion and phase-eld models of fracture, J.
Mech. Phys. Solids 57(2), 342 368.
Hofmann, M., Bahr, H., Linse, T., Bahr, U., Balke, H. & Weiss, H. (2006), Self-driven tun-
neling crack arraysa 3D-fracture mechanics bifurcation analysis, Int. J. Fracture 141(3-
4), 345356.
Irwin, G. (1958), Fracture. Hand. der Physik, Vol. IV, Springer, Berlin.
Lancioni, G. & Royer-Carfagni, G. (2009), The variational approach to fracture mechanics.
a practical application to the french Panth eon in Paris, J. Elasticity 95(1-2), 130.
Lazarus, V. & Pauchard, L. (2011), From craquelures to spiral crack patterns: inuence of
layer thickness on the crack patterns induced by desiccation, Soft Matter 7(6), 25522559.
Malzbender, J., den Toonder, J. M. J., Balkenende, A. R. & de With, G. (2002), Measuring
mechanical properties of coatings: a methodology applied to nano-particle-lled sol-gel
coatings on glass, Mat. Sci. Eng. R. 36(2-3), 47103.
Mangold, N. (2005), High latitude patterned grounds on mars: Classication, distribution
and climatic control, Icarus 174(2), 336359.
Mumford, D. & Shah, J. (1989), Optimal approximations by piecewise smooth functions and
associated variational problems, Comm. Pure Appl. Math 42, 577685.
Nemat-Nasser, S., Sumi, Y. & Keer, L. (1980), Unstable growth of tension cracks in brittle
solids - stable and unstable bifurcations, snap-through, and imperfection sensitivity, Int. J.
Solids Structures 16(11), 10171035.
Pham, K., Amor, H., Marigo, J.-J. & Maurini, C. (2011), Gradient damage models and their
use to approximate brittle fracture, Int. J. Damage Mech. 20(4), 618652.
Pham, K., Marigo, J.-J. & Maurini, C. (2011), The issues of the uniqueness and the stability
of the homogeneous response in uniaxial tests with gradient damage models, J. Mech.
Phys. Solids 59(6), 1163 1190.
Plug, L. & Werner, B. (2001), Fracture networks in frozen ground, J. Geophys. Res.-Earth
106(B5), 85998613.
Rhee, Y., Kim, H., Deng, Y. & Lawn, B. (2001), Contact-induced damage in ceramic coatings
on compliant substrates: Fracture mechanics and design, J. Am. Ceram. Soc. 84(5), 1066
1072.
Seilacher, A. (2001), Concretion morphologies reecting diagenetic and epigenetic path-
ways, Sediment. Geol. 143(1-2), 4157.
20
Shinn, E. (2009), The mystique of beachrock, in P. Swart, G. Eberli & J. McKenzie, eds,
Perspectives in Carbonate Geology, Vol. 41 of Special publication of the International
Association of Sedimentologists, Wiley-Blackwell, pp. 1928.
Tirumkudulu, M. & Russel, W. (2004), Role of capillary stresses in lm formation, Lang-
muir 20(7), 29472961.
Wang, H. F. (2000), Theory of linear poroelasticity with applications to geomechanics and
hydrogeology, Princeton University Press, Princeton and Oxford.
21