Fracture Modeling in Computer Graphics

Lien Muguercia, Carles Bosch

1
, Gustavo Patow
ViRVIG-UdG, Universitat de Girona, Spain
Abstract
While object deformation has received a lot of attention in Computer Graphics in recent years, with several good surveys that
summarize the state-of-the-art in the eld, a comparable comprehensive literature review is still needed for the related problem of
crack and fracture modeling. In this paper we present such a review, with a special focus on the latest advances in this area, and a
careful analysis of the open issues along with the avenues for further research. With this survey, we hope to provide the community
not only a fresh view of the topic, but also an incentive to delve into and explore these unsolved problems further.
Keywords: crack modeling, fracture modeling, simulation, survey
1. Introduction 1
Physically plausible object deformation and fracture have 2
been of central importance in many elds, and particularly in 3
Computer Graphics since more than 25 years ago [1, 2]. Dier- 4
ent areas, such as architecture and fabrication, usually require 5
very precise simulations, for which numerical models have been 6
devised using a combination of continuum mechanics, dynam- 7
ics, dierential geometry, calculus and Computer Graphics, among 8
others. As a body can undergo many physical phenomena, frac- 9
tures are essential to the movie and video game industries be- 10
cause of the explosions or shattering bodies required. In gen- 11
eral, the phenomena we study in this survey can be considered 12
as ubiquitous, as can be observed in almost every structure, 13
from crystals to entire buildings. 14
In spite of its importance, the study of fractures is still a 15
non-closed problem, with several open issues to be dealt with, 16
most of which result from the many approximations and sim- 17
plications introduced to simulate the intrinsic complexities of 18
this phenomenon. Advances in this eld would open new fron- 19
tiers for applications such as simulation and prototyping of frag- 20
ile objects, resistance assessment and model resilience studies. 21
There are a number of good reviews on deformable models 22
in Computer Graphics [1, 3] as well as aging/weathering tech- 23
niques [4] that touch on the topic of fracture processes. How- 24
ever, we feel that a deep review of the current state-of-the-art 25
in crack and fracture modeling techniques is missing. Thus, in 26
this paper we aim to ll this gap with a comprehensive review 27
of the work done thus far, and, to improve understanding and 28
strengthen the relationships among the dierent works carefully 29
classifying them according to several criteria. 30
2. Overview 31
After an introduction to the mathematical background needed 32
to understand the basic principles of object deformation and the 33
phenomenon of fracture (Section 3), we present our principal 34
classication of the dierent methods involved in the fracture 35
process: 36
Physically-based methods (Section 4), are those that fol- 37
low a simulation-based approach to compute the fracture 38
opening, propagation and appearance. Among these, we 39
sub-classify the state-of-the-art in the eld into: 40
Mass-spring models (Section 4.1), where the object 41
is approximated by a nite set of masses, pairwise 42
joined by springs, each with its own dening pa- 43
rameters. 44
Finite element methods (Section 4.2), that partition 45
the object into a set of disjoint elements (e.g., tetra- 46
hedrons) joining at discrete points. When the prob- 47
lem is formulated in terms of these points, then it is 48
converted into a set of simpler algebraic equations, 49
which are then solved to establish the behavior of 50
the system. 51
Meshless methods (Section 4.3), where the model is 52
approximated with a set of unconnected calculation 53
points that are simulated. The value for any other 54
point in the model is obtained by interpolation. 55
Other approaches (Section 4.4) cover those, which 56
do not fall into the previous categories, but rather 57
follow physical principles for their simulations of 58
the fracture process. 59
Geometry-based methods (Section 5), also known as pro- 60
cedural methods, seek plausible patterns but are not inter- 61
ested in a physically accurate phenomena description. 62
Example-based methods (Section 6) try to mimic real- 63
world fractures by copying the behavior observed in real 64
phenomena. These methods, which build on both Com- 65
puter Vision and Computer Graphics techniques, usually 66
extract parameters from images and then apply these to 67
generate a new fracture. 68
Preprint submitted to Computers & Graphics October 16, 2014
Finally, in our conclusions (Section 7), we present compar- 69
isons and further classication schemes to ensure the reader has 70
a comprehensive view of the most recent developments in this 71
area. This includes Table 2, which provides further details of 72
the main techniques reviewed in this survey. Some avenues for 73
future research are also outlined at the end. 74
3. Background 75
In this section we will briey introduce the main physical 76
concepts behind generating and propagating fractures. 77
3.1. Stress and Strain 78
In continuum mechanics, we dene the physical quantities 79
of an object as a continuous function in space (and time). In 80
general, we dene the rest shape of an undeformed object as 81
the connected subspace M R
3
[1, 3, 5]. Each point x M 82
has its own properties dened at its coordinates x inside the ob- 83
ject, called material coordinates. When we deform the object, 84
we apply forces that move its x points to their new positions x. 85
With the old and new positions we can dene the displacement 86
vector eld on M as u(x) = x x, which represents the posi- 87
tional dierences between the current point and rest positions. 88
Refer to Figure 1 for a graphical representation. 89
Figure 1: The displacement vector eld u(x).
We usually measure deformation in terms of the so-called
strain, which we often dene as a normalized measure of the
body deformation. This measure represents the displacement
between particles in the body relative to a reference length. Ba-
sically, the strain measures the local deviation of a given de-
formation from a rigid-body deformation. As the deformation
in dierent directions might be dierent, the strain is generally
expressed as a tensor. In three dimensions, this tensor is of or-
der 2. Given the eld u(x), we can compute the elastic strain
of a point at a given time, simply by relating it to the gradi-
ent u. Observe that u is a 3 3 matrix of the derivatives
(u)
i j
=
j
u
i
. In Computer Graphics the strain is usually de-
ned for small deformations as one of

G
=
1
2
(u + [u]
T
+ [u]
T
u)

C
=
1
2
(u + [u]
T
)
The rst one (i.e.,
G
) is called the Green-Lagranges strain ten- 90
sor and
C
the Cauchy strain tensor, is its linearized counterpart. 91
Based on the strain tensor, we can compute the stress tensor
R
33
, which provides information about the forces acting
on a point when the body is deformed. Of course, this relation-
ship strongly depends on the properties of the material, and can
be quite complex. In Computer Graphics it is customary to use
Hooks law:
= E
where E is a rank 4 tensor which relates both tensors and
in a linear way, which is useful for small deformations. On
the other hand, other denitions of stress are required for large
deformations, such as the Piola-Kirchho stress tensor, which
expresses the stress relative to the reference conguration (in
contrast to the Cauchy stress tensor that expresses the stress rel-
ative to the present conguration); the Biot stress tensor, which
expresses the forces due to stretch only applied in the unde-
formed body per unit undeformed area; or the Kirchho stress
tensor, which is widely used when there is no change in volume
during plastic deformation [6]. Another possibility is the Saint
Venant-Kirchho model:
= Tr(
G
)I
3
+ 2
G
where and are constants specic to each material and de- 92
ne the way it is deformed, and I
3
is the 3 3 identity ma- 93
trix. For more information on these tensors, we refer the inter- 94
ested reader to the works authored by Bonet and Wood [6] and 95
Chakrabarty [7]. 96
In general, the stress tensor is a symmetric 3 3 matrix, 97
so it has 3 real eigenvalues. These eigenvalues correspond to 98
the stresses in the principal directions, represented by its re- 99
spective eigenvectors to the principal stress directions. Positive 100
eigenvalues indicate tension, while negative eigenvalues repre- 101
sent compression. 102
It is possible to compute the body force f for each point
from as:
f (x, t) = (x, t)
whose elements are f
i
=

j

j

i j
. With these forces f (x, t),
we can model deformation using the equations of motion. In
general, these equations are posed in terms of the density of
the material, which again is a function of position x and time t:
(x, t)

2
t
2
x = f
i
(x, t)
However, for general cases it is almost impossible to nd an 103
analytical solution, and we must resort to numerical methods, 104
which are the subject of the following sections. 105
2
Figure 2: Brittle (a) vs. ductile (b) fracture, from OBrien et al. [8]. (image
copyright ACM 2002)
3.2. Brittle and ductile fractures 106
We can dene an elastic material as one that will return to 107
its original shape when the external forces on it cease to exist. 108
To the contrary, a plastic material, will not go back to its orig- 109
inal conguration. Real materials usually have a limited elas- 110
tic behavior, and if deformed beyond a certain threshold (called 111
elastic limit or yield point), they will undergo a plastic deforma- 112
tion. If the material is deformed further, there is another limit, 113
called the failure threshold
max
, which is the point at which the 114
material fractures. This failure threshold is a material-specic 115
parameter. If the elastic limit and the failure threshold are close 116
to each other, then the material will undergo a small (almost 117
negligible) elastic deformation before fracturing, and the ma- 118
terial is termed brittle. A brittle fracture releases most of its 119
elastic energy thus allowing the crack to proceed further into 120
the material and reducing the energy required to break it. As a 121
consequence, brittle objects generally shatter. 122
On the other hand, if the thresholds are adequately sepa- 123
rated, such that the failure threshold is much larger than the 124
elastic limit, then the object undergoes an elastic deformation, 125
followed by a plastic one, before being fractured. In this case 126
the material is termed ductile (Refer to Figure 2). 127
Examples of ductile materials include structural steel, as 128
well as many alloys of other metals. They usually exhibit a 129
very linear stress-strain relationship up to the well dened yield 130
point, as shown in Figure 3. The stress value plotted in curve 131
(A) was obtained by dividing the load P by the cross-sectional 132
area A
0
of the specimen measured before any deformation had 133
taken place (compare to its tensorial version dened above). 134
The linear portion of the curve shown in the left is the elastic 135
region, and the slope is the modulus of elasticity or Youngs 136
Modulus. The curve typically decreases slightly after passing 137
the yield point (Point 2 in the gure). As deformation contin- 138
ues, the stress increases on account of strain hardening until 139
it reaches the ultimate strength (Point 1 in the gure). Until 140
this point, the cross-sectional area decreases uniformly and ran- 141
domly. However, beyond this point a neck forms where the lo- 142
cal cross-sectional area decreases more quickly than the rest of 143
the sample, which results in an increased true stress. Since the 144
cross-sectional area of the specimen decreases as P increases, 145
the stress plotted in curve (A) does not represent the true stress 146
in the specimen, shown as curve (B). The dierence between 147
this engineering stress (P/A
0
) and the true stress (P/A) becomes 148
apparent in ductile materials after yield has started. While the 149
engineering stress, which is directly proportional to the load 150
P, decreases with P during the necking phase, the true stress, 151
which is proportional to P but also inversely proportional to A, 152
is observed to keep increasing until fracture of the specimen oc- 153
curs (at point 3 in the gure). The actual fracture point is in the 154
same vertical line as the visual fracture point. It is easily appre- 155
ciated that ductile fractures require greater amounts of energy 156
to break as the result of the considerable energy used for plas- 157
tic deformation, so in fact these objects usually tear [9]. For- 158
mally, we can say that fracture will happen, in both cases, when 159
the maximum stress is greater than the failure threshold
max
. 160
Other factors, such as material density and damping factors, are 161
usually taken into account to simulate fracture processes. 162
Figure 3: Approximate stress vs. strain. Left: typical curve for steel. Here,
(A) is the apparent curve (computed with = P/A
0
, where P is the load and
A
0
the cross section of the sample before any deformation) and (B) is the real
curve (computed with = P/A, with A the real cross section of the samble
being deformed). Right: brittle material [9]. In the gures, numbered regions
represent: 1. Ultimate Strength; 2. Yield Strength; 3. Fracture point (at
max
);
4. Strain hardening region; 5. Necking region. Adapted from [10]
To simplify computations, it is customary to split the total
strain into its elastic and plastic parts, as =
e
+
p
. As a
consequence, one option commonly used for the yield point is
the von Mises yield criterion [11], dened as

1
<

Tr(
e
)
3
I
3

(1)
3
where
1
is the yield threshold, Tr(.) is the trace of a matrix,
and ||.|| is the Frobenius norm. Basically, this last expression
measures the deviation of the elastic strain from an initial ref-
erence conguration. The second term, to the right, essentially
eliminates any dilation eect, concentrating only on shape dis-
tortions. For strains beyond this point, plastic deformation oc-
curs. To handle the plastic part of the deformation, OBrien [8]
proposes measuring the change in the plastic strain as

p
=
||

||
1
||

||

where ||

|| is the right hand side of Equation 1. Then, the limit

on the maximum plastic deformation can be enforced by updat-
ing the plastic strain with the threshold
2
as

p
:= (
p
+
p
)min

1,

2
||
p
+
p
||

Thus, the dierence between the total strain and the plastic 163
strain is the current elastic strain. By introducing
2
we are 164
limiting the plastic strain, but not the total strain, which can be 165
larger and produce fracture. 166
Both brittle and ductile fracture have been widely addressed 167
in Computer Graphics, although their dierent properties often 168
led to the use of dierent methods to simulate their behavior. 169
Refer to Section 4 for further details. 170
3.3. Visual aspects of fracture simulation 171
The presence of cracks and fractures on objects may provide 172
a great deal of information on their properties, interactions with 173
other objects, usage or history. Objects being impacted by other 174
objects are expected to show deformation and/or fracture, and 175
crack patterns are also very common on aged environments sub- 176
ject to fatigue conditions. Properly simulating this phenomenon 177
is thus important to improving the perceived realism of virtual 178
environments. 179
Physically-based simulation is especially eective for mod- 180
eling and animating fractured objects, which tend to have many 181
degrees of freedom. Animating these phenomena with keyfram- 182
ing and motion capture would be much more dicult [12]. Most 183
simulation techniques reviewed in this survey drawheavily from 184
the eld of fracture mechanics and its literature, which abstracts 185
the micro-scale nature of fracture to a macroscopic level based 186
on a continuum model. The requirements of engineering ap- 187
plications, however, are dierent from the ones in graphics ap- 188
plications. In Computer Graphics, simulation techniques rely 189
on simplications that would be unacceptable in an engineer- 190
ing context. In the continuum approach, the scale of the ef- 191
fects being modeled is signicantly greater than the scale of the 192
materials composition. Even if macroscopic fractures can be 193
signicantly inuenced by eects occurring at the microscale, 194
ignoring the microscopic eects in favour of a macroscopic de- 195
scription is often reasonable, as our interest is on the graphical 196
appearance rather than its physical correctness. Similarly, nu- 197
merical accuracy is less important compared to issues such as 198
visual appearance, ease of use, or computational eciency. 199
In order to achieve good visual quality, we must rely on 200
representations able to capture realistic fracture patterns. A 201
common limitation present in earlier approaches is to limit the 202
propagation of fractures to the boundaries of the initial mesh, 203
which tends to create visible artifacts. Irregular shaped shards 204
can only be achieved by allowing fractures to propagate in ar- 205
bitrary directions. Rendering realistic fractures from the results 206
of a simulation may also pose its own problems. In particular, 207
when the fracture is not represented by the boundaries of the 208
original mesh elements, then we have a dissociation that results 209
in two dierent meshes being used: a simulation mesh and a 210
rendering mesh. In such situations, special care is needed to 211
maintain consistency between both representations. 212
3.4. Related reading material 213
For a more in-depth review of the related background, the 214
interested reader is referred to some previous surveys. In partic- 215
ular, the works by Gibson and Mirtich [1] and Nealen et al. [3] 216
present the most signicant contributions for physically-based 217
deformable models in recent decades. Akin to our classica- 218
tion, they also focus on nite element methods, mass-spring- 219
based methods, meshfree methods, coupled particle systems 220
and reduced deformable models based on modal analysis. They 221
also make a particular connection with the fundamental con- 222
cept of time discretization, and present a number of application 223
areas such as cloth, hair, virtual surgery, deformation and frac- 224
ture. The background in the recent thesis by Glondu [5] should 225
also be taken into account, as model deformation is presented 226
and previous work is carefully classied. Finally, the survey by 227
Merillou and Ghazanfarpour [4] presents a very detailed sur- 228
vey on aging and weathering techniques where methods dealing 229
with crack formation are also reviewed. 230
4. Physically-based methods 231
Eective simulation of fracture processes is one of the most 232
accurate ways of approaching the problem. The main tech- 233
niques usually used to solve this complex problem are mass- 234
spring models, nite element methods and meshless methods. 235
In general, an analytical solution is not possible, so we must 236
resort to numerical approximations to solve the equations for 237
the material. One option is explicit integration schemes, which 238
provide explicitly the values for a given time step in function of 239
the previous step. These are usually not unconditionally stable: 240
they converge only for small values of the integration step. Two 241
examples are the explicit Euler integration [13] and the Runge- 242
Kutta integration [14]. To the contrary, implicit schemes pose 243
a system of equations that must be solved, but provide stability 244
for larger time steps, some even arbitrarily large. An example 245
of these is the implicit Euler method [13]. The choice made be- 246
tween the two has implications not only on the stability of the 247
system but also on its computational time. 248
4.1. Mass-spring models 249
Mass-spring models are one of the simplest ways to model
a deformable body. Mass-spring models are characterized by
discretizing M into a nite set of particles {p
i
; 1 i n}. Each
particle p
i
has its own mass m
i
and position x
i
, and are pairwise
4
connected with springs, each with its own stiness, damping
factor and rest length. Each particle is set to a classic equilib-
rium equation relating the internal forces (the springs) and the
external forces (gravity, collisions) with its mass:

2
x
i
t
2
= f
i
,
where f
i
is the sum of the forces acting on the particle p
i
, and

i
is the mass associated with the i-th particle. In general, f
i
is described as the sum of two main terms: the external forces
(e.g., gravity or collisions) and the internal forces, which come
from the springs attached to the particle p
i
. In general, springs
follow Hookes law [15], which can be stated as:
f
i
= k
(|x
i j
| l
i j
)x
i j
||x
i j
||
,
with k being the spring constant that characterizes its stiness, 250
l
i j
its original or rest length, and x
i j
= x
i
x
j
its current length, 251
measured as the dierence between the particle positions. As 252
all the particles are connected through the springs, the equa- 253
tions lead to a systemof coupled ordinary dierential equations, 254
which is solved through numerical integration methods. When 255
the limit of a spring exceeds a given threshold, the breakage is 256
simply performed by removing the corresponding connection 257
between the particles. 258
In Computer Graphics, some of the rst approaches for sim- 259
ulating fracture relied on mass-spring representations [16, 17, 260
18, 19, 20]. Norton et al. [16], for instance, place lattices of 261
cubic cells around the surface of objects and connect them with 262
several spring congurations. They simulate brittle fracture due 263
to impacts and stretching eects by removing an entire cube of 264
springs at once, which may produce visible artifacts. For the 265
nal surface, they classify the faces of the model as belonging 266
to the interior or the fracture surface, generating and rendering 267
a polygon in the latter. 268
Hirota et al. [17, 18] focus on the simulation of drying ma- 269
terials and their temporal cracking. A set of shrinking and 270
contraction eects modify the physical properties of a material 271
based on a set of measurements, which results in the opening 272
and propagation of cracks (Refer to Figure 4 for an example). 273
For this, both a bi-layered representation [17] and a 3D mass- 274
spring model [18] have been examined. For visualization, Hi- 275
rota et al. rely on a marching tetrahedra-like algorithm that uses 276
the broken points on the springs to compute the crack shape. 277
Aoki et al. [21] later incorporate a moisture diusion model to 278
guide the drying process. Federl and Prusinkiewicz [19, 20] 279
also use a similar bi-layered approach for modeling tree bark, 280
but based on wedge elements and element removal. 281
Mazarak et al. [22] simulate fractures and debris produced 282
by explosions based on a similar strategy, where objects are 283
modeled with voxels connected through rigid links, represent- 284
ing innitely sti springs (Refer to Figure 5). Based on the 285
strength of a spherical blast wave and a set of heuristics, the 286
yield limit of each link is subsequently decreased until fracture 287
appears. The spring-based model is later improved on [23] by 288
using adaptive voxel shapes and including rotational forces. 289
Figure 4: Comparison between the cracks obtained on a real drying clay (top)
and the corresponding simulation (bottom), from [18]. (image copyright
Springel-Verlag 2000)
Figure 5: Graph representation of the connected voxel model, from Mazarak et
al. [22]. (image copyright Mazarak 1999)
Smith et al. [24] propose a systemthat connects point-masses 290
with distance-preserving constraints rather than a grid of sti 291
springs. Instead of computing displacements, they evaluate the 292
forces that these constraints exert in response to the impulses, 293
and use strength thresholds to indicate when and where objects 294
will break. In contrast to explicit methods and the use of elas- 295
tic meshes, these rigid constraints coupled with a quasi-static 296
approach allow them to quickly compute a solution. Figure 6 297
shows a table broken using this approach. 298
Figure 6: Breaking table from Smith et al. [24]. (image copyright CGF 2001)
Mass-spring models and similar approaches are sometimes 299
preferred because of their simplicity and faster results, which 300
is especially true for real-time fracture. However, these ap- 301
5
proaches suer from several limitations in terms of mechanical 302
behavior and visual quality. They do not provide a direct way 303
of resisting shear or bending, and it is dicult to express im- 304
portant material properties such as the stress-strain relationship. 305
Furthermore, the exact location and orientation of the fracture 306
are unknown, and fracture surfaces are often restricted to the 307
boundaries in the initial mesh structure, exhibiting directional 308
artifacts (Refer to Figure 4). As a result, these techniques can 309
only realistically model eects that occur on a scale much larger 310
than the inter-node spacing. Continuous models such as the 311
Finite Element Method, as described in the following section, 312
directly account for most of these issues. 313
4.2. Finite element methods 314
The Finite Element Methods (FEM) partition the model M 315
into a set of disjoint elements (usually tetrahedrons). These el- 316
ements join at discrete points, usually called node points. The 317
FEM method proceeds by dening the problem on the nite set 318
of node points, instead of on the original mesh, resulting in a 319
set of algebraic equations that are solved numerically. Rather 320
than solving a continuous problem, these methods look to solve 321
the discrete positions of the node points. First, the unknown x 322
is approximated by x =

i
x
i
b
i
, where x
i
are the positions of 323
the node points and b
i
are xed nodal basis functions that are 324
1 at node i and 0 everywhere else. While replacing x in the 325
above equations for motion results in a set of simpler algebraic 326
equations, no sets of basis functions will yield the solution in 327
the general case, so all methods aim at reducing the error in- 328
troduced by the approximations employed. This approximation 329
error is measured by replacing the resulting approximation into 330
the original equations. One example is the Galerkin method 331
[25] that treats nding x
i
as an optimization problem where the 332
error is minimized. Refer to Swenson and Ingraea [26] for 333
further theoretical details about Galerkin methods. 334
In their pioneering work, Terzopoulos and Fleisher [2, 27] 335
model viscoelastic, plastic and fracture behaviors based on a 336
continuous formulation, where both FEM and nite dierences 337
are proposed for their discretization. One of their formulations 338
is based on a hybrid decomposition of the model into rigid and 339
deformable components which, in turn, improves the numeri- 340
cal conditioning of rigid objects. This enables a wide range of 341
deformable behaviors from highly elastic to nearly rigid to be 342
covered, while fracture is shown on torn paper and cloth sheets. 343
Later, OBrien and Hodgins [28] propose a model that uses 344
FEM with the theory, which has found widespread use in the 345
graphics literature, developed by Grith [29] and Irwin [30]. 346
Their approach is based on linear elastic fracture mechanics and 347
non-linear nite element analysis, where 3D objects are repre- 348
sented as tetrahedral elements, refer to Figure 7. A separation 349
tensor is used to decide where and how fractures should appear 350
or propagate. The separation tensor is built from tensile and 351
compressive forces obtained from the stress tensor. Remeshing 352
is nally applied in order to enable unconstrained crack paths, 353
thus achieving high visual quality. 354
The same approach has been applied to simulate shattering 355
due to explosions [31] and later extended to incorporate plastic 356
deformation [12, 8]. By means of an additive model, the total 357
Figure 7: FEM discretization into tetrahedra. Surface (a) and object interior
(b); OBrien [28]. (image copyright ACM 1999)
strain is separated into elastic and plastic components, easily 358
allowing the simulation of ductile fracture. Muller et al. [32] 359
include a further co-rotational formulation that avoids visible 360
artifacts during large deformations. As a result, the rotation on 361
a per-element basis is factored out, allowing for linear deforma- 362
tion analysis independent of rotation. 363
For real-time purposes, common solutions rely on instanta- 364
neous fracture models, which combine rigid bodies with quasi- 365
static stress analysis [33, 34, 35]. Such a hybrid approach treats 366
objects as rigid bodies between collisions, while using static 367
FEM during collision events. Instantaneous fracture models 368
can be appropriate for sti materials, which tend to exhibit 369
small deformations. However, the elastic stiness matrix can 370
have null spaces (rigid translations and rotations that do not de- 371
formthe model) that may aect computations. Muller et al. [33] 372
address this issue by anchoring elements far from the collision 373
event, while Bao et al. [34] directly identify and eliminate them 374
during the solution; typically based on the conjugate gradient 375
method. Refer to Figure 8 for an example of FEM-based frac- 376
ture from [34]. In the context of sound simulation, Zheng and 377
James [35] propose a sparse least-squares solver to produce 378
fracture sounds at near audio rates, which they combine with 379
Voronoi-based fracture patterns. Iben and OBrien [36] use a 380
quasi-static system for simulating elastic relaxation and shrink- 381
age after cracking, while the stress eld is dened heuristically, 382
thus providing substantial control for the user in terms of the 383
position and patterns of the cracks. 384
Figure 8: An example of a FEM-based fracture with copious small shards; Bao
et al. [34]. (image copyright IEEE 2007)
Methods based on quasi-static analysis are much cheaper 385
than a fully dynamic simulation, but ignoring the elastic en- 386
ergy released during fracture, which might produce unrealistic 387
fracture patterns. Parker and OBrien [37] propose a simplied 388
version of previous FEM techniques [8, 32] for use in video- 389
games and real-time simulations. Their approach relies on a 390
6
linearized semi-implicit solver and a parallelized implementa- 391
tion of the conjugate gradient method, where objects are parti- 392
tioned into islands and processed independently. This leads to 393
a fast and robust simulation of fracture on sti and soft mate- 394
rials. Another common approach is to use a simplied model, 395
which can be embedded into the original mesh for the simula- 396
tion [32, 38, 37], which allows the use of lower resolutions and 397
simpler representations in order to obtain faster results without 398
compromising visual quality. Refer to Figure 9 for a represen- 399
tation of an embedded mesh. 400
Figure 9: An embedded object mesh representation; Parker and OBrien [37].
The simulated mesh is based on tetrahedrons, while a triangle-based represen-
tation models the nal fracture patterns. (image copyright ACM 2009)
While continuous remeshing oers high visual quality it has 401
many limitations in terms of performance and simulation sta- 402
bility. Stability can be guaranteed by limiting the directions of 403
cracks [28], but the model can easily grow in complexity and 404
become computationally expensive. Wicke et al. [39] propose 405
a conservative local re-meshing algorithm that tries to replace 406
as few tetrahedra as possible. This maintains high tetrahedron 407
quality and limits the accumulation of numerical error. For vi- 408
sualization (and collision detection), they keep track of a sec- 409
ond mesh which shares the topology with the simulation mesh, 410
but introduce dierent nodal positions. The rendering mesh 411
also needs to be remeshed to maintain an appropriate geometric 412
quality. Refer to Figure 10 to see an example of re-meshing. 413
Other approaches simply split the mesh along element bound- 414
aries [33] or even remove whole elements [20, 40], which, while 415
they can be very fast, compromise visual accuracy. Embed- 416
ded methods may help in mitigating these eects, as the qual- 417
ity of the simulation does not need to match the visual quality 418
[37, 32, 38]. Molino et al. [41] propose a virtual node algorithm 419
that relies on a lattice for modeling the extra geometry, which 420
serves to avoid the generation of ill-conditioned elements. The 421
material within an element is fragmented by creating several 422
Figure 10: Adaptive mesh renement helps a ball to crash through dierent
ductile plates; Wicke et al. [39]. (image copyright ACM 2010)
replicas of the element and assigning a portion of material to 423
each one. These elements can then be directly used for vi- 424
sualization purposes. However, the geometry increases with 425
both the subdivision and the added hidden nodes, thus aect- 426
ing the performance. The approach by Molino et al. can be 427
seen as a generalization of extended FEM (XFEM) [42], and 428
has been used in other works [34] (see Figure 8). Level sets 429
can be also very eective for handling topological changes, as 430
demonstrated by Hegemann et al. [43]. Here, the level set is de- 431
ned in the undeformed conguration of the object, and evolves 432
in material space to represent the transition from undamaged to 433
failed material. The tetrahedralization is done in a regular lat- 434
tice and the elements are duplicated to handle both damaged 435
and healthy regions. Refer to Figure 11 for an embedded mesh 436
representation. 437
Figure 11: Embedded Lagrangian mesh stages for their material; Hegemann et
al. [43]. (image copyright SCM 2013)
4.3. Meshless methods 438
Meshless methods appear as an alternative to enhance FEMs,
where M is discretized in a set of points without connectivity
information between them. In these methods, the value at any
point in the interior of the body is retrieved through the inter-
polation of these calculation points. Basically, if we have com-
puted certain values
i
at the calculation points x
i
, we can obtain
this value at any other point x by computing
(x) =

r
i
(x)
i
,
where the
r
i
are a set of kernel functions that allow us to com- 439
pute the interpolation at point x from the values
i
at points x
i
. 440
Usually, these kernel functions depend on a parameter r that is 441
1 at the evaluation point and goes to 0 at range r. 442
7
There are a number of features of these methods that make 443
them favorable for fracture simulation, such as avoiding the 444
complex re-meshing operations and the problems associated to 445
element cutting and mesh alignment common to FEM, as de- 446
scribed in the work by Belytschko et al. [44]. In the eld of 447
mechanics, Sukumar et al. [45] proposed a particle-based ap- 448
proach to model the physical behavior around a crack; while 449
Belytschko and Tabbara [46] resort entirely to meshless meth- 450
ods. The methods later developed in the eld of computer 451
graphics were based on these seminal works. 452
Muller et al. [47] introduce a meshless framework for the 453
animation of elastic and plastic deformation, which supports 454
a wide range of materials. The spatial derivatives of the dis- 455
placement eld are computed using a Moving Least Squares 456
(MLS) [48] procedure based on a linear basis. These deriva- 457
tives are subsequently used to obtain strains, stresses and elastic 458
forces at each simulated point. Topological changes are handled 459
with resampling, in order to deal with undersampled and over- 460
sampled regions. Pauly et al. [49] build from the method by 461
Muller et al. to include brittle and ductile fractures, which is 462
achieved by continuously adding surface samples during crack 463
propagation and dynamically adapting the shape functions around 464
them. Resampling is also conducted around the fracture sur- 465
faces, as shown in Figure 12. 466
Figure 12: Volume sampling: Octree decomposition (a), initial adaptive oc-
tree sampling (b), sampling after local repulsion (c), and dynamic re-sampling
during fracture process (d); Pauly et al. [49]. (image copyright ACM 2005)
Liu et al. [50] propose a quasi-static solution for sti ma- 467
terials by treating brittle objects as fully rigid bodies. The lo- 468
cal Petrov-Galerkin method (MLPG) [25] is used in order to 469
avoid dealing with large neighborhoods of point masses, and a 470
simple damage-based fracture model is proposed. This model 471
clusters particles previously classied as damaged or undam- 472
aged to generate the new fragments. Chen et al. [51] focus on 473
a similar problem but using Smoothed Particle Hydrodynamics 474
(SPH) [52]. Their formulation is used to analyze local stress 475
tensors induced by collisions. Based on a tetrahedralization of 476
the material, particles are linked with up to four neighborhood 477
particles and fracture is simulated by breaking these links ap- 478
propriately. A clustering approach is also conducted to group 479
damaged points, in order to resample the point set before gener- 480
ating the fracture surface. Even with the resampling, directional 481
artifacts are easily spotted. 482
Eciently dealing with proximity information is paramount 483
for meshless approaches. Preliminary approaches proposed the 484
use of spatial hashing to retrieve neighboring points within a 485
specic distance, or a local caching scheme for exploiting tem- 486
poral coherence [47, 49]. Steinemann et al. [53] use a visibility 487
graph instead, which is applied both for cutting and fracturing 488
deformable objects. Figure 13 shows an object splitted using 489
this approach. 490
Figure 13: Pumpkins splitted into pieces, from Steinemann et al. [53]. (image
copyright ACM 2006)
In terms of rendering, Pauly et al. [49] adapt a CSG ren- 491
dering technique for point-sampled surfaces [54], while Liu et 492
al. [50] embed a detailed triangle mesh into the particles, where 493
the triangles move with the particles using the same rigid body 494
parameters. 495
Despite all the advantages of Meshless methods, the fact 496
that essential boundary conditions cannot be as straightforwardly 497
implemented as mesh-based methods can be a problem given 498
that its shape functions are not interpolating. Eventually, the 499
computational cost would be higher than in FEM. One thing 500
these methods have in common is that again we nd the disasso- 501
ciation between the simulation representation and the geometry 502
used for rendering (e.g., the rendering mesh). In general, these 503
methods have to generate that mesh from the set of sampling 504
points, and the number of points can grow in an unpredictable 505
manner, which can pose serious problems to the process. 506
4.4. Other approaches 507
This section covers fracture methods that rely on physically- 508
based approaches but do not belong to any of the above classi- 509
cations. 510
Glondu et al. [55, 56], for instance, use modal analysis to 511
simulate real-time brittle fracture. They base the fracture initi- 512
ation method on this analysis and use a fast energy-based frac- 513
ture propagation algorithm. To generate the fragments and their 514
geometric surfaces, they rely on an implicit representation that 515
eciently models the surfaces obtained. Refer to Figure 14 for 516
a representacion of a crack propagation, from [55]. The gen- 517
erated mesh is constructed by evaluating the type of fractured 518
surface in a way very similar to a marching-tetrahedra algo- 519
rithm; mean while the deviations of the cracks are handled with 520
a noise function. As modal analysis uses an analytical solution 521
for computing deformation, no integration method is necessary. 522
As a result, modal analysis leads to very fast deformable simu- 523
lations, although the range of deformation tends to be limited. 524
Ning et al. [57] handle heterogeneous materials using a mov- 525
able cellular automata (MCA) approach [58]. The object is dis- 526
cretized into spherical particles, called the MCA, which react 527
to the deformation. In order to obtain accurate simulations, a 528
large number of automata are often needed, which results in a 529
signicant computational cost. Ning and colleagues use CUDA 530
to achieve real-time performance. 531
8
Figure 14: Propagation of one crack in a coarse mesh. The colored section rep-
resents a fracture that cut the body straight through the physical mesh; Glondu
et al. [55]. (image copyright IEEE 2012)
5. Geometry-based methods 532
Geometry-based methods, also known as procedural tech- 533
niques, are able to produce patterns of cracks and fractures that 534
are visually plausible but without relying on the actual physical 535
process or its accurate simulation. Physical simulation is of- 536
ten considered to be computationally demanding and does not 537
provide sucient control over fracture propagation. Procedural 538
methods, on the other hand, rely on tools that oer considerable 539
control of the patterns of cracks obtained as well as the size and 540
shape of the fragments. This often translates in the generation 541
of customized patterns by means of a simple set of parameters. 542
Figure 15: Real (left) and synthetic (right) bark; Lefebvre and Neyret [67].
(image copyright The Eurographics Association 2002)
In the context of fragmentation induced by explosions, Ne 543
and Fiume [59] propose a simple recursive pattern generator 544
that divides a planar region into polygonal shards, where cracks 545
fork according to a user-specied angle. Gobron and Chiba 546
[60] propose a semi-physical approach to simulate cracks on 547
layered surfaces. These layers are divided into cells and a 2D 548
directional stress distribution is assigned to each one based on a 549
set of geometric properties. Cracks then open based on a clas- 550
sical stress threshold and propagate using a relaxation process 551
and a set of heuristics. For rendering, the crack segments are 552
projected onto the cells. Then, each cell crossed by a segment 553
is divided into micro-cells, and these are used for anti-aliased 554
rendering while taking into account shading and refraction ef- 555
fects. A similar strategy has been used for propagating cracks 556
on layered surfaces in the context of paint cracking and peeling 557
[71]. Lefebvre and Neyret [67] also rely on a semi-physical ap- 558
proach for bark generation, refer to Figure 15. They model the 559
bark with a set of strips parallel to the growing direction, where 560
the epidermal elements are treated as semi-rigid. The fracture 561
criterion is based on the relative lengthening of elements and 562
which corresponds to the Grith energetic approach [29]. Once 563
generated, the bark surface can be rendered directly from any 564
viewpoint. 565
Some authors focus on non-photorealistic (NPR) techniques 566
for simulating cracking on images. Wyvill [70], for instance, 567
simulates crack patterns on images of wax paintings, where 568
their placement is based on a Distance Transformation along 569
with a set of heuristics. Mould [72] simulates cracks by means 570
of Voronoi regions and weighted edges. For rendering cracks, 571
two approaches are proposed: texture transfer, which involves 572
extracting the texture from one image for its application into 573
another; and the modulation of an existing texture based on the 574
distance from the crack locations. Figure 16 shows two crack 575
patterns generated by this method. 576
Figure 16: Dierent crack patterns by manipulating the site distributions, as
shown in [72]. (image copyright ACM 2005)
Martinet et al. [61] propose an interactive approach for mod- 577
eling both cracks and fractures on 3Dobjects. Their patterns are 578
modeled using a set of skeletons that are applied onto the object 579
by means of a set of Boolean operations. Cracks also include 580
an associated prole curve that is swept along the pattern before 581
its subtraction from the object. Once generated, fragments are 582
converted into polygons for fast rendering. Refer to Figure 17 583
for a comparison between a real object and a synthetic model. 584
A similar representation is also used in [64], but storing the pat- 585
terns into an atlas for their later application and animation. 586
Figure 17: A real clay vase (left) and a synthetic model (right). Martinet et al.
[61]. (image copyright IEEE 2004)
For real-time fracture, especially in computer games, a com- 587
mon approach is to rely on pre-fracturing, also known as pre- 588
scoring. Objects are fractured during the modeling stage, and 589
9
Behavior 2D 3D Pre-fractured Not pre-fractured Voronoi Semi-physical
Brittle [59] [60] [61] [62] [63] [64] [65] [63] [66] [59] [60] [62] [63] [60] [67]
[67] [68] [64] [65] [66] [62] [69] [68] [67] [61] [66] [65] [68]
Ductile [70] - - [70] - -
Table 1: Summary of geometry-based methods exposed in Section 5.
the pre-fractured versions replace the original objects based on 590
impacts and similar events happening at run-time. This ap- 591
proach is very fast and provides a lot of control for the artists, 592
but the obtained patterns tend to be xed. Most of these pat- 593
terns habitually rely on a Voronoi partitioning of the object 594
[65, 63, 66, 62], and are readily available in software such as 595
Maya [73] or Houdini [74]. 596
In order to handle non-centered events, pre-fractured pat- 597
terns can be aligned with specic locations or even adapted 598
to a specic object or bounding-box [69, 62]. These patterns 599
ease the modeling stage, as only a small set of (user-provided) 600
generic fracture patterns is necessary, but it does require a dy- 601
namic cutting of the mesh. In order to cut the mesh, Muller et 602
al. [62] propose fast boolean operations while Oh et al. resort to 603
recursive spatial decompositions [66]. The approach from Oh 604
et al. further interconnects fragments as in [24], where connec- 605
tions can be dynamically broken based on relative velocities. 606
With respect to the visualization aspects, as with other meth- 607
ods, Muller et al. [62] further relie on a second mesh for render- 608
ing, represented by an approximate convex hull decomposition. 609
Su et al. [69] incorporate a level set into the fractured object 610
that is combined with the target object for collision purposes. 611
The level set is triangulated by either fast dual-contouring or 612
by marching cubes. Figure 18 shows an overview of the pre- 613
fractured patterns aligniation proposed by [62]. 614
Figure 18: Overview of Muller et al. [62] fracture algorithm. (image copyright
ACM 2013)
Along the same lines, Valette et al. [68] use pre-scoring for 615
modeling cracks on soils. Cracks propagate inside a voxelized 616
representation of the terrain according to a precomputed pattern 617
of paths, where shrinkage is incorporated to guide the progres- 618
sion. Rendering is performed by transforming the terrain into 619
a mesh, and then incorporating the cracks as surface polygons. 620
An interesting aspect of this work is its validation against real 621
patterns. The validation is based on a set of statistics taken over 622
area densities and the length and connectivity of these patterns, 623
which could be used for validating other methods. 624
In order to avoid expensive physical computations authors 625
tend to bank on a procedural technique. However, the freedom 626
to make some assumptions and provide the user with tools to 627
control and guide the processes, can bring physical limitations 628
with non-real behaviors and visual limitations. To avoid some 629
of these assumptions, an example-based method could be ap- 630
plied, where the idea is to extract information from examples in 631
order to reproduce or guide the fracture. Refer to Section 6 for 632
further details. 633
Given the large variety of procedural solutions, Table 1 sum- 634
marizes the various papers that were presented in this section 635
and classied according to dierent criteria. This includes sim- 636
ulated behavior (brittle vs. ductile), whether they are 2D or 3D, 637
pre-fractured or not, based on a Voronoi diagram, and whether 638
they rely on a semi-physical approach. Notice how ductile frac- 639
ture has barely been addressed in this context. 640
6. Example-based methods 641
Example-based techniques build from both Computer Vi- 642
sion and Computer Graphics elds. In the case of cracks and 643
fractures, the purpose is to use real examples, often in the form 644
of photographs, to obtain similar patterns or to guide the fractur- 645
ing process itself. For a comprehensive review of image-based 646
techniques we refer the reader to Shum and Kang [75]. 647
The direct mapping of crack patterns onto 3D models has 648
been addressed by Hsien-Hsi and Wen-Kai [76]. Given an im- 649
age of a crack pattern, they vectorize the image and extract the 650
corresponding pattern in the form of a graph. The graph can 651
then be interactively projected onto the surface from a parabolic 652
bounding volume, rendered as two projected quads for each 653
crack edge, and enhanced with bump mapping. Note that other 654
procedural techniques could also be useful for mapping crack 655
patterns [61, 64]. Wang et al. [77] focus on reconstructing 656
tree bark from an input image, where bark features are seg- 657
mented using texton analysis. An interactive system is pro- 658
posed to transform these features into a height-eld represen- 659
tation, which is then used for rendering. 660
Figure 19: Paint cracking, with context of paint thickness on a frog; [78]. (im-
age copyright ACM 2007)
Texture synthesis is a standard practice for mapping gen- 661
eral example-based textures [79]. Synthesis techniques may 662
rely on many types of input data and can be performed in both 663
space and time. Enrique et al. [80], for instance, rely on time- 664
lapse images to synthesize temporal phenomena including paint 665
10
Figure 20: Overview of the optimization process from [81]. Parameter vector
p is optimized iteratively. After each simulation, the statistics of the simulation
are compared to the statistics of the reference image. (image copyright The
Authors 2012)
cracking. Lu et al. [78] resort to similar data but to analyze 666
how these phenomena might be inuenced by a set of context 667
parameters. These parameters are measured during the acqui- 668
sition process and include local geometrical properties of the 669
object. Given a target 3D object, these kinds of properties serve 670
to guide the placement of features like cracks. Refer to Figure 671
19 for an example. 672
More recently, Glondu et al. [81] proposed using real ex- 673
amples to guide a physical simulation of fracture. Rather than 674
matching the input pattern, they focus on matching a set of 675
statistics captured from the example. By means of a user study, 676
they evaluate which statistics are more relevant for visual simi- 677
larity, using measures similar to [68]. Using an extension from 678
[55], these statistics are then used in an inverse procedure to 679
estimate the simulation parameters that generate such patterns. 680
Refer to Figure 20 for an overview of this process. 681
7. Conclusions 682
It is well known that creating good and realistic motion, 683
whatever the eld of application, is a challenging task due to 684
its complexity. Our survey addresses the particular problem of 685
simulating and animating cracks and fractures, and provides a 686
survey of the most relevant techniques proposed in the litera- 687
ture. 688
There is no single generic solution or representation that 689
can cover all cases well because the behavior of each mate- 690
rial can vary in various ways, depending on multiple factors. 691
Thus, our classication is based on the way the simulation of 692
these phenomena is performed. One section was dedicated to 693
physically-based works, which can be classied according to 694
the simulation method. A further section was dedicated to those 695
approaches that renounce physical accuracy preferring to select 696
a procedural technique in order to gain a more interactive ap- 697
plication. The nal classications focus on approaches based 698
on examples (i.e., data, images, etc). Theses approaches are 699
looking to replicate observed patterns, consequently avoiding 700
tedious physical experiments and computations. 701
We can see that FEM methods are some of the most pop- 702
ular techniques, while meshless approaches are gaining more 703
and more relevance. This is surely because of their ability to 704
correctly represent cracks and fracture patterns. Models derived 705
from the equations of continuum mechanics are important as it 706
is possible to quantify the precision of the simulation, since the 707
parameters of the model can be obtained from experiments con- 708
ducted on real world objects, even if it is not trivial to obtain all 709
the necessary quantities. Some authors, on the other hand, still 710
prefer to work on representations that limit their accuracy but 711
provide the user with more freedom and interactivity. Geomet- 712
ric approaches are one such example. 713
In Table 2, we provide a comparison of the main techniques 714
reviewed in this paper; which can help the reader to easily eval- 715
uate the dierences between them. For each method, we report 716
the dimensions the method operates on, the underlying repre- 717
sentation, the simulated behavior, the method or solver used 718
by the technique, and the estimated performance. The latter 719
represents the speed of a method on simulating a fracture pat- 720
tern, where a higher number of dots represents a better perfor- 721
mance/speed. This classication is roughly based on timings 722
provided by the authors as well as other considerations such 723
as model size or hardware specications. Methods with higher 724
performance (5 dots) are able to simulate fractures in the order 725
of a few milliseconds, while slower methods (1 dot) may take 726
several seconds or minutes. Please note that this performance 727
can only be considered as a very rough indication of the real 728
performance. The timings reported by these methods depend 729
on many aspects, including the parameters chosen for the simu- 730
lations, the numerical approach, the number of iterations, or any 731
code optimizations. The main purpose of this score is thus to 732
provide an intuitive way to compare their behavior rather than 733
a precise estimation. 734
Finally, we devise some current challenges and open prob- 735
lems regarding the simulation of fracture in Computer Graph- 736
ics. 737
Scalability. Despite recent advances in the eld, physically- 738
based methods are still too time consuming to deal with large 739
scenes. Real-time approaches either rely on fast geometric ap- 740
proaches or on reducing the physical accuracy to obtain greater 741
performance. Even with these approximations, the complexity 742
of the simulated objects is still limited, so there is a need to deal 743
with very complex scenes such as outdoor scenarios. 744
Parameter tweaking. Obtaining a specic pattern often 745
implies trial and error of the parameters oered by each method, 746
which can be tedious for artists. Image-based methods are able 747
to directly map real patterns onto the objects, although they 748
often lack physical validity. There is a need to provide more 749
intuitive ways to edit these patterns and to capture them from 750
measured data. A recent approach [81] represents a rst step 751
towards combining both approaches and capturing such param- 752
eters from images, although there is still a great deal of room 753
for improvement. 754
Fracture Surface. Generating fracture surface is an impor- 755
tant challenge that needs specic solutions. Little of the pre- 756
11
Type/Method Dim. Representation Behavior Method/Solver Performance
Mass-spring
[16] 3D deformed lattices brittle Euler n/a
[17] 2D two-layer brittle n/a
[18] 3D lattice (cubic cells) brittle n/a
[21] 3D tetrahedra brittle nite di (moisture) + quantum mechanics
[22] 3D interconnected voxels brittle Euler
[23] 3D arbitrarily shaped voxels brittle Euler
[24] 3D lattice w/ rigid constraints brittle quasi-static
FEM
[2] 2D mesh ductile semi-implicit n/a
[28, 8] 3D tetrahedra brit./duct. Euler or 2nd-order Taylor
[33] 3D tetrahedra brittle rigid body (Euler) + quasi-static
[20] 2.5D wedge elements brittle static
[41] 2D/3D thin sells/tetrahedra ductile Newmark
[32] 3D tetrahedra ductile implicit Euler
[38] 3D surface + cube mesh brittle implicit Euler
[34] 3D mesh/tetrahedra + level set ductile rigid body + quasi-static
[36] 2D triangle mesh brittle Euler
[37] 3D tetrahedra ductile backward Euler
[35] 3D tetrahedra brittle quasi-static
[43] 3D tetrahedra + level set ductile Euler
Meshless
[47] 2D/3D point-based ductile Leap-Frog / implicit
[49] 3D point-based + surfels ductile Leap-Frog
[50] 3D point-based + mesh brittle quasi-static (Local Petrov-Galerkin)
[51] 3D point-based + tetrahedra brittle SPH
Others
[57] 3D spherical particles brittle movable cellular automata
[55] 3D mesh + implicit brittle modal analysis
Geometry-based
[59] 2D at mesh + grid brittle iterative propagation
[60] 2.5D mesh + multi-layer cells brittle cellular automata
[67] 2.5D bark slices brittle procedural + quasi-static
[72] 2D image - Voronoi
[61] 2D/3D mesh + implicit brittle manual + boolean operations
[64] 2.5D mesh + implicit brittle manual + boolean operations
[70] 2D image ductile distance transform
[68] 2.5D height map + implicit brittle quasi-static + precomputed paths
[69] 3D mesh + level set brittle pre-scoring
[66] 3D mesh brittle recursive decomposition + SPH
[62] 3D mesh brittle pre-scoring + convex decomposition
Example-based
[77] 2.5D image - texton analysis + user assistance
[76] 2D image - vectorization + projection n/a
[78] 2D mesh + texture - texture synthesis
[81] 3D mesh + implicit brittle optimization + simulation
Table 2: Comparison between dierent fracture methods. Refer to the text for a detailed description.
vious work has resolved this problem in an elegant way. For 757
instance, in FEM methods, element cutting and local remesh- 758
ing are time consuming and may lead to sliver elements. Even 759
the use of virtual nodes cannot be regarded as a perfect deni- 760
tive solution. In general, they all have artifacts that appear and 761
that are related to the size of elements. Hence, there is still a 762
pending issue crucial to this point. 763
Validation. Fracture methods proposed in Computer Graph- 764
ics are rarely validated against real physical data or even from a 765
perceptual point of view. This is common for most simulation 766
approaches as well as aging/weathering techniques [4]. Some 767
of the very few works that deal with validation are Valette et 768
12
al. [82], who compare their patterns obtained in lab conditions 769
against real patterns, or Ramanarayanan et al. [83], who pro- 770
vided a framework to link physically accurate to visually accu- 771
rate models using a Visual Equivalence Predictor. Having a set 772
of standard tests based on real-life examples could clearly help 773
here. 774
Implementation. Developing a fracture technique often 775
implies starting from scratch and the techniques available are 776
rarely made public. This also makes the comparison with pre- 777
vious approaches much more dicult. Given the complexity 778
of some of these methods, promoting the availability of such 779
techniques and the reusability of libraries would help to further 780
advance the eld. 781
Fracture simulation is still a very active research area, and 782
we believe this will continue to be so in the future. Given its 783
applications in many areas, there is a lot of interest in develop- 784
ing new techniques that oer more realism and greater interac- 785
tivity. In this survey, we have revised the state-of-the-art in the 786
eld and we have provided insights on its present circumstances 787
as well as on some of the problems that still have to be faced. 788
We hope this eld will further evolve and provide us with more 789
exciting results. 790
Acknowledgements 791
We would like to thank the anonymous reviewers for their 792
valuable comments. This work was partially funded by the 793
TIN2013-47137-C2-2-P project from Ministerio de Economa 794
y Competitividad, Spain, and a Beatriu de Pin os grant from the 795
Catalan Government, Spain. 796
Bibliography 797
[1] Gibson, S.F., Mirtich, B.. A survey of deformable modeling in computer 798
graphics. Tech. Rep. TR-97-19; MERL, Cambridge, MA; 1997. 799
[2] Terzopoulos, D., Fleischer, K.. Modeling inelastic deforma- 800
tion: Viscolelasticity, plasticity, fracture. SIGGRAPH Comput Graph 801
1988;22(4):269278. 802
[3] Nealen, A., M uller, M., Keiser, R., Boxerman, E., Carlson, M.. Physi- 803
cally Based Deformable Models in Computer Graphics. Computer Graph- 804
ics Forum 2006;25(4):809836. 805
[4] M erillou, S., Ghazanfarpour, D.. A survey of aging and weathering phe- 806
nomena in computer graphics. Computers & Graphics 2008;32(2):159 807
174. 808
[5] Glondu, L.. Physically-based and real-time simulation of brittle fracture 809
for interactive applications. Ph.D. thesis; INRIA Rennes - Ens Cachan, 810
France; 2012. 811
[6] Bonet, J., Wood, R.. Nonlinear ContinuumMechanics for Finite Element 812
Analysis. Cambridge University Press; 2008. ISBN 9781139467544. 813
[7] Chakrabarty, J.. Theory of Plasticity. Elsevier Science; 2006. ISBN 814
9780080481364. 815
[8] OBrien, J.F., Bargteil, A.W., Hodgins, J.K.. Graphical model- 816
ing and animation of ductile fracture. ACM Transactions Graphics 817
2002;21(3):291294. 818
[9] Beer, F., Jr. Johnston, E., DeWolf, J., Mazurek, D.. Mechanics of 819
Materials. McGraw-Hill Education; 2011. ISBN 9780073380285. 820
[10] Wikipedia, . Stressstrain curve. 2014. [accessed 14-August-2014]; URL 821
http://en.wikipedia.org/wiki/Stress-strain curve. 822
[11] von Mises, R.. Mechanik der festen K orper im plastisch- deformablen 823
Zustand. Nachrichten von der Gesellschaft der Wissenschaften zu 824
G ottingen, Mathematisch-Physikalische Klasse 1913;1913:582592. 825
[12] OBrien, J.F., Hodgins, J.K.. Animating fracture. Commun ACM 826
2000;43(7):6875. 827
[13] Hairer, E., Nrsett, S., Wanner, G.. Solving Ordinary Dierential Equa- 828
tions I Nonsti problems. Springer; 2000. 829
[14] Atkinson, K.E.. An Introduction to Numerical Analysis. New York: 830
Wiley; 1978. 831
[15] Petroski, H.. Invention by design : How Engineers Get from Thought to 832
Thing. Harvard University Press,; 1996. 833
[16] Norton, A., Turk, G., Bacon, R., Gerth, J., Sweeney, P.. Animation of 834
fracture by physical modeling. The Visual Computer 1991;7:210219. 835
[17] Hirota, K., Tanoue, Y., Kaneko, T.. Generation of crack patterns with a 836
physical model. The Visual Computer 1998;14(3):126137. 837
[18] Hirota, K., Tanoue, Y., Kaneko, T.. Simulation of three-dimensional 838
cracks. The Visual Computer 2000;16:371378. 839
[19] Federl, P., Prusinkiewicz, P.. A texture model for cracked surfaces, 840
with an application to tree bark. In: In Proceedings of Western Computer 841
Graphics Symposium. 1996, p. 2329. 842
[20] Federl, P., Prusinkiewicz, P.. Modelling fracture formation in bi-layered 843
materials, with applications to tree bark and drying mud. In: In Proceed- 844
ings of Western Computer Graphics Symposium. 2002, p. 2935. 845
[21] Aoki, K., Dong, N.H., Kaneko, T., Kuriyama, S.. Physically based sim- 846
ulation of cracks on drying 3d solids. Computer Graphics International 847
Conference 2004;:357364. 848
[22] Mazarak, O., Martins, C., Amanatides, J.. Animating exploding objects. 849
In: In Proceedings of Graphics Interface. Morgan Kaufmann Publishers 850
Inc; 1999, p. 211218. 851
[23] Martins, C., Buchanan, J., Amanatides, J.. Visually believable explo- 852
sions in real time. Computer Animation, 2001 The Fourteenth Conference 853
on Computer Animation Proceedings 2001;. 854
[24] Smith, J., Witkin, A.P., Bara, D.. Fast and controllable simulation of 855
the shattering of brittle objects. Comput Graph Forum 2001;20(2):8190. 856
[25] Mikhlin, S.. Variational methods in mathematical physics. International 857
series of monographs in pure and applied mathematics; Pergamon Press; 858
[distributed by Macmillan, New York]; 1964. 859
[26] Swenson, D.V., Ingraea, A.R.. Modeling mixed-mode dynamic crack 860
propogation using nite elements- theory and applications. vol. 3. 1988, 861
p. 381397. 862
[27] Terzopoulos, D., Fleischer, K.. Deformable models. The Visual Com- 863
puter 1988;4(6):306331. 864
[28] OBrien, J., Hodgins, J.. Graphical model and animation of brittle frac- 865
ture. In: In Proceedings of SIGGRAPH99. 1999, p. 137146. 866
[29] Grith, A.. The Phenomena of Rupture and Flow in Solids. Philosophi- 867
cal Transactions of the Royal Society, London 1921;A221:163198. 868
[30] Irwin, G.R.. Analysis of Stresses and Strains Near the End of a Crack 869
Traversing a Plate. J Appl Mech 1957;. 870
[31] Yngve, G.D., OBrien, J.F., Hodgins, J.K.. Animating explosions. In: 871
Proceedings of ACM SIGGRAPH 2000. 2000, p. 2936. 872
[32] M uller, M., Gross, M.. Interactive virtual materials. In: GI 04- Pro- 873
ceedings of Graphics Interface 2004. 2004, p. 239246. 874
[33] M uller, M., McMillan, L., Dorsey, J., Jagnow, R.. Real-time simu- 875
lation of deformation and fracture of sti materials. In: Proceedings of 876
the Eurographic workshop on Computer animation and simulation. New 877
York, NY, USA: Springer-Verlag New York, Inc. ISBN 3-211-83711-6; 878
2001, p. 113124. 879
[34] Bao, Z., Hong, J.M., Teran, J., Fedkiw, R.. Fracturing rigid materials. 880
IEEE Trans Vis Comput Graph 2007;13(2):370378. 881
[35] Zheng, C., James, D.L.. Rigid-body fracture sound with precom- 882
puted soundbanks. ACM Transactions on Graphics (Proceedings of SIG- 883
GRAPH 2010) 2010;29(3). 884
[36] Iben, H.N., OBrien, J.F.. Generating surface crack patterns. Graph 885
Models 2009;71:198208. 886
[37] Parker, E.G., OBrien, J.F.. Real-time deformation and fracture in a game 887
environment. In: Symposium on Computer Animation. ACM; 2009, p. 888
165175. 889
[38] M uller, M., Teschner, M., Gross, M.. Physically-based simulation 890
of objects represented by surface meshes. In: Proc. Computer Graphics 891
International. 2004, p. 156165. 892
[39] Wicke, M., Ritchie, D., Klingner, B.M., Burke, S., Shewchuk, J.R., 893
OBrien, J.F.. Dynamic local remeshing for elastoplastic simulation. 894
ACM Trans Graph 2010;29:4914911. 895
[40] Federl, P., Prusinkiewicz, P.. Finite element model of fracture formation 896
on growing surfaces. In: In Proceedings of International Conference on 897
Computational Science. 2004,. 898
13
[41] Molino, N., Bao, Z., Fedkiw, R.. A virtual node algorithm for changing 899
mesh topology during simulation. ACM Trans Graph 2004;23:385392. 900
[42] Moes, N., Dolbow, J., Belytschko, T.. A nite element method for crack 901
growth without remeshing. International Journal for Numerical Methods 902
in Engineering 1999;:131150. 903
[43] Hegemann, J., Jiang, C., Schroeder, C., Teran, J.M.. A level set 904
method for ductile fracture. In: Proceedings of the 12th ACM SIG- 905
GRAPH/Eurographics Symposium on Computer Animation. SCA 13; 906
New York, NY, USA: ACM; 2013, p. 193201. 907
[44] Belytschko, L.T., Organ, D., Krongauz, Y.. A coupled nite element- el- 908
ementfree galerkin method. Computational Mechanics 1995;17(3):186 909
195. 910
[45] Sukumar, N., Moran, B., Black, T., Belytschko, L.T.. An element- 911
free galerkin method for three-dimensional fracture mechanics. Compu- 912
tational Mechanics 1997;:170175. 913
[46] Belytschko, L.T., Tabbara, M.. Element-free galerkin method for wave 914
propogation and dynamic fracture. Computer Methods in Applied Me- 915
chanics and Engineering 1995;126(1-2):131153. 916
[47] M uller, M., Keiser, R., Nealen, A., Pauly, M., Gross, M., Alexa, M.. 917
Point based animation of elastic plastic and melting objects. In: Proceed- 918
ings of the 2004 ACM SIGGRAPH/Eurographics symposium on Com- 919
puter animation. SCA 2004; Eurographics Association. ISBN 3-905673- 920
14-2; 2004, p. 141151. 921
[48] Levin, D.. The approximation power of moving least-squares. Math 922
Comput 1998;67(224):15171531. 923
[49] Pauly, M., Keiser, R., Adams, B., Dutr e, P., Gross, M., Guibas, 924
L.J.. Meshless animation of fracturing solids. ACM Trans Graph 925
2005;24(3):957964. 926
[50] Liu, N., He, X., Li, S., Wang, G.. Meshless simulation of brittle 927
fracture. Journal of Visualization and Computer Animation 2011;22(2- 928
3):115124. 929
[51] Chen, F., Wang, C., Xie, B., Qin, H.. Flexible and rapid anima- 930
tion of brittle fracture using the smoothed particle hydrodynamics for- 931
mulation. Journal of Visualization and Computer Animation 2013;24(3- 932
4):215224. 933
[52] Gingold, R.A., Monaghan, J.J.. Smoothed particle hydrodynamics - 934
Theory and application to non-spherical stars. Mon Not Roy Astron Soc 935
1977;181:375389. 936
[53] Steinemann, D., Otaduy, M.A., Gross, M.. Fast arbitrary split- 937
ting of deforming objects. In: Proceedings of the 2006 ACM SIG- 938
GRAPH/Eurographics symposiumon Computer animation. SCA06; Eu- 939
rographics Association; 2006, p. 6372. 940
[54] Wicke, M., Teschner, M., Gross, M.H.. Csg tree rendering for point- 941
sampled objects. In: Pacic Conference on Computer Graphics and Ap- 942
plications. IEEE Computer Society; 2004, p. 160168. 943
[55] Glondu, L., Marchal, M., Dumont, G.. Real-time simulation of brittle 944
fracture using modal analysis. IEEE Transactions on Visualization and 945
Computer Graphics 2013;19(2):201209. 946
[56] Glondu, L., Schvartzman, S.C., Marchal, M., Dumont, G., Otaduy, 947
M.A.. Fast collision detection for fracturing rigid bodies. IEEE Transac- 948
tions on Visualization and Computer Graphics 2013;99(PrePrints):1. 949
[57] Ning, J., Xu, H., Wu, B., Zeng, L., Li, S., Xiong, Y.. Modeling 950
and animation of fracture of heterogeneous materials based on cuda. The 951
Visual Computer 2013;29(4):265275. 952
[58] Psakhie, S., Horie, Y., Korostelev, S., Smolin, A., Dmitriev, A., 953
Shilko, E., et al. Method of movable cellular automata as a tool for 954
simulation within the framework of mesomechanics. Russian Physics 955
Journal 1995;38(11):11571168. 956
[59] Ne, M., Fiume, E.. A visual model for blast waves and fracture. In: In 957
Proc. of Graph. Interface. 1999, p. 193202. 958
[60] Gobron, S., Chiba, N.. Crack pattern simulation based on 3d surface 959
cellular automaton. The Visual Computer 2001;17:287309. 960
[61] Martinet, A., Galin, E., Desbenoit, B., Hakkouche, S.. Procedural mod- 961
eling of cracks and fractures. In: Shape Modelling International (Short 962
Paper ). Genova, Italy; 2004, p. 346349. 963
[62] M uller, M., Chentanez, N., Kim, T.Y.. Real time dynamic fracture 964
with volumetric approximate convex decompositions. ACM Trans Graph 965
2013;32(4):115:1115:10. 966
[63] Hellrung, J., Selle, A., Shek, A., Sifakis, E., Teran, J.. Geometric 967
fracture modeling in bolt. In: ACM SIGGRAPH Talks (Sketch). 2009,. 968
[64] Desbenoit, B., Galin, E., Akkouche, S.. Modeling cracks and fractures. 969
The Visual Computer 2005;21(8-10):717726. 970
[65] Raghavachary, S.. Fracture generation on polygonal meshes using 971
voronoi polygons. In: ACM SIGGRAPH 2002 Conference Abstracts and 972
Applications. New York, NY, USA: ACM; 2002, p. 187187. 973
[66] Oh, S., Shin, S., Jun, H.. Practical simulation of hierarchical brittle 974
fracture. Journal of Visualization and Computer Animation 2012;23(3- 975
4):291300. 976
[67] Lefebvre, S., Neyret, F.. Synthesizing bark. In: Gibson, S., De- 977
bevec, P.E., editors. 13th Eurographics Workshop on Rendering Tech- 978
niques (EGSR 02); vol. 28. Pisa, Italy: Eurographics Association; 2002, 979
p. 105 116. 980
[68] Valette, G., Pr evost, S., Lucas, L., L eonard, J.. A dynamic model of 981
cracks development based on a 3d discrete shrinkage volume propagation. 982
Comput Graph Forum 2007;27(1):4762. 983
[69] Su, J., Schroeder, C., Fedkiw, R.. Energy stability and fracture for 984
frame rate rigid body simulations. In: Proceedings of the 2009 ACM SIG- 985
GRAPH/Eurographics Symposium on Computer Animation. New York, 986
NY, USA: ACM; 2009, p. 155164. 987
[70] Wyvill, B., van Overveld, K., Carpendale, S.. Rendering cracks in batik. 988
In: Proceedings of the 3rd international symposiumon Non-photorealistic 989
animation and rendering. NPAR 04; New York, NY, USA: ACM. ISBN 990
1-58113-887-3; 2004, p. 61149. 991
[71] Paquette, E., Poulin, P., Drettakis, G.. The simulation of paint cracking 992
and peeling. In: Graphics Interface 2002. 2002, p. 5968. 993
[72] Mould, D.. Image-guided fracture. In: GI 05- Proceedings of Graphics 994
Interface 2005. 2005, p. 219226. 995
[73] Fracture FX. 2014. URL http://www.fracture-fx.com. 996
[74] Houdini. 2014. URL http://www.sidefx.com. 997
[75] Shum, H., Kang, S.B.. Review of image-based rendering techniques. 998
Proc SPIE 2003;4067:213. 999
[76] Hsien-Hsi, H., Wen-Kai, T.. A straightforward and intuitive approach on 1000
generation and display of crack-like patterns on 3d objects. In: Computer 1001
Graphics International. 2006, p. 554561. 1002
[77] Wang, X., Wang, L., Liu, L., Hu, S., Guo, B.. Interactive modeling of 1003
tree bark. 11th Pacic Conference on Computer Graphics and Applica- 1004
tions 2003;. 1005
[78] Lu, J., Georghiades, A.S., Glaser, A., Wu, H., Wei, L.Y., Guo, B., 1006
et al. Context-aware textures. ACM Trans Graph 2007;26(1):3. 1007
[79] Wei, L.Y., Lefebvre, S., Kwatra, V., Turk, G., et al. State of the art in 1008
example-based texture synthesis. In: Eurographics 2009, State of the Art 1009
Report, EG-STAR. 2009, p. 93117. 1010
[80] Enrique, S., Koudelka, M., Belhumeur, P., Dorsey, J., Nayar, S., 1011
Ramamoorthi, R.. Time-varying textures: denition, acquisition, and 1012
synthesis. In: ACM SIGGRAPH 2005 Sketches. ACM; 2005, p. 130. 1013
[81] Glondu, L., Muguercia, L., Marchal, M., Bosch, C., Rushmeier, H., 1014
Dumont, G., et al. Example-based fractured appearance. Computer 1015
Graphics Forum 2012;31(4):15471556. 1016
[82] Valette, G., Herbin, M., Lucas, L., L eonard, J.. A preliminary approach 1017
of 3d simulation of soil surface degradation by rainfall. In: NPH. 2005, 1018
p. 4149. 1019
[83] Ramanarayanan, G., Ferwerda, J., Walter, B., Bala, K.. Visual equiv- 1020
alence: Towards a new standard for image delity. ACM Trans Graph 1021
2007;26(3). 1022
14