Accepted Manuscript

Minimum energy multiple crack propagation. Part III: XFEM computer imple-
mentation and applications

Danas Sutula, Pierre Kerfriden, Tonie van Dam, Stéphane P.A. Bordas

PII: S0013-7944(17)30155-8
DOI: http://dx.doi.org/10.1016/j.engfracmech.2017.08.004
Reference: EFM 5638

To appear in: Engineering Fracture Mechanics

Received Date: 3 April 2017

Revised Date: 26 July 2017
Accepted Date: 1 August 2017

Please cite this article as: Sutula, D., Kerfriden, P., van Dam, T., Bordas, S.P.A., Minimum energy multiple crack
propagation. Part III: XFEM computer implementation and applications, Engineering Fracture Mechanics (2017),
doi: http://dx.doi.org/10.1016/j.engfracmech.2017.08.004

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
 Minimum energy multiple crack propagation.
Part III: XFEM computer implementation and applications
Danas Sutulaa,b , Pierre Kerfridenb , Tonie van Dama , and Stéphane P.A.
Bordasc,a,*
a
Institute of Computational Engineering, University of Luxembourg, 6 Avenue de
la Fonte, 4362 Esch-sur-Alzette, Luxembourg
b
School of Engineering, Cardiff University, Queen’s Building, The Parade,
Cardiff, CF24 3AA, UK
c
Visiting Professor, Institute of Research and Development, Duy Tan University,
K7/25 Quang Trung, Danang, Vietnam
*
[email protected]

July 17, 2017

Contents
1 Introduction 2

2 Aspects of implementation 4
2.1 Crack intersections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Enrichment tracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Numerical integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3 Assembly of equations 12
3.1 Initiation of the discrete system . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Updating the discrete system . . . . . . . . . . . . . . . . . . . . . . . . . 14

4 Numerical benchmarks 16
4.1 Verification of enrichment updating . . . . . . . . . . . . . . . . . . . . . . 16
4.2 Computational speed-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Discussion and general remarks . . . . . . . . . . . . . . . . . . . . . . . . 17

5 Convergence of fracture paths 23

5.1 Square plate with 10 random cracks . . . . . . . . . . . . . . . . . . . . . 23
5.2 Rectangular plate with 10 parallel cracks . . . . . . . . . . . . . . . . . . . 23
5.3 A numerical improvement to the growth direction . . . . . . . . . . . . . . 30
5.4 Comparison of fracture paths by different criteria . . . . . . . . . . . . . . 30

1
6 Summary 31

7 Supplementary material 31

References 39

Abstract

The three-part paper deals with energy-minimal multiple crack propagation in a linear
elastic solid under quasi-static conditions. The principle of minimum total energy, i.e.
the sum of the potential and fracture energies, which stems directly from the Griffith’s
theory of cracks, is applied to the problem of arbitrary crack growth in 2D. The proposed
formulation enables minimisation of the total energy of the mechanical system with
respect to the crack extension directions and crack extension lengths to solve for the
evolution of the mechanical system over time. The three parts focus, in turn, on (I) the
theory of multiple crack growth including competing cracks, (II) the discrete solution by
the extended finite element method using the minimum-energy formulation, and (III)
the aspects of computer implementation within the Matlab programming language. The
key contributions of Part-III of the three-part paper are as follows: (1) implementation
of XFEM in Matlab with emphasis on the design of the code to enable fast and efficient
computational times of fracture problems involving multiple cracks and arbitrary crack
intersections, (2) verification of the minimum energy criterion and comparison with
the maximum tension criterion via multiple benchmark studies, and (3) we propose a
numerical improvement to the crack growth direction criterion that gives significant gains
in accuracy and convergence rates of the fracture paths, especially on coarse meshes.
Finally, the open-source Matlab code, documentation, benchmarks and other example
cases are included as supplementary material.

1 Introduction

In the previous two parts of this three-part paper we focussed on the theory and the
formulation of energy-minimal crack growth in the context of linear elastic fracture me-
chanics (LEFM). We described three solution strategies within a discrete framework for
the prediction of the fracture paths. We recommended a gradient-descent based solution
approach due to its better robustness in tackling unstable crack growth problems. In
the actual implementation of the proposed solution approach we discovered that a large
amount of thought and planning was required to achieve a robust computer implemen-
tation. Indeed, such an implementation is a sine qua non for the success of simulations
involving arbitrary multiple crack growth with intersects. Here, in Part-III of our three-
part paper, we describe the lessons learnt from building such an implementation.

2
We apply the proposed energy minimisation routines within the framework of the ex-
tended finite element method (XFEM) [5, 29, 4, 19, 18]. We chose XFEM because it
offers certain advantages in modelling problems with non-smooth solutions; in partic-
ular, cracks can be modelled independently of the underlying finite element mesh.1 In
XFEM, the classic finite element approximation is enriched with additional functions
that can capture the essential features of discontinuities, high gradients and singularities
associated with cracks in LEFM. In turn, XFEM can offer accurate and optimally con-
vergent numerical solution for the elastic field [25, 22, 23, 2, 1]. The reader is referred
to Part-II of our three-part paper for details on the XFEM approximation we use.
Even though XFEM offers mesh independent modelling of cracks, the enrichment, which
is introduced via the partition of unity method [27, 12], needs to be updated as cracks
evolve over time. The XFEM implementation poses some practical challenges, especially
concerning crack growth with intersections or when the crack extension solution needs
to be determined iteratively, such as in pursuing the energy-minimal fracture solution.
To this end, a robust enrichment tracking strategy needs to be devised to facilitate the
efficient updating of the discretised mechanical system with respect to any topological
changes in the enrichment. The current XFEM implementation relies heavily on available
computational memory for storing various enrichment data. The idea is to strive for
faster computations at the expense of a higher memory demand.
Nowadays, many XFEM implementations exist in C++ [8, 24, 31] and Fortran [7, 36,
35]. Since these are compiled programming languages, fast computations are generally
possible. An interpreted language, such as Matlab, is generally slower (despite using
many pre-combined libraries); however, faster development times can be achieved, which
can be beneficial particularly for research purposes. Some example open-source research
codes written in Matlab include an (extended) isogeometric analysis code [32] (available
here) and an (enriched) meshless method library [33] (available here).
Concerning simulations of multiple crack growth within the XFEM framework, it can be
noted that only small topological changes take place with each time-step in the enrich-
ment topology. Provided the enriched part of the global system of equations is updated
only where needed and not reassembled from scratch (for simplicity’s sake), the bottle-
neck in every time-step will tend to lie in the solution to the linear system of equations
as opposed to updating of the discrete system or post-processing of the solution. Our
way of speeding up the computational times is to avoid recomputations of quantities
that tend to be reused regularly. We consider a fair compromise to store each enriched
element’s shape functions, shape function derivatives and quadrature weights at Gauss
points at the expense of a higher memory demand. This, in turn, can speed up the
solution post-processing and updating times since certain expensive recomputations can
be avoided (especially those that can not be easily expressed by vectorisation [20, 3, 26]).
For example, the post-processing stage of the solution normally involves computing do-
main integrals to evaluate the crack tip stress intensity factors or the energy release rates.
1
There are, however, some discretisation constraints related to XFEM, which will be addressed later.

3
Having the enriched elements’ shape functions at one’s disposal the domain integrals can
be evaluated readily. Moreover, updating the global system of equations after each time-
step requires certain enriched elements’ contributions to be subtracted from the global
system of equations. Since the enriched elements’ shape functions are pre-computed, the
subtraction of the elemental equations can be carried out faster. The benchmark prob-
lems of multi-crack growth that we solve clearly show such a performance profile of the
current XFEM implementation within Matlab. Since Matlab uses built-in pre-compiled
solvers, which are fast and robust, the total computational time of a large multi-crack
growth simulation can be comparable to that of a compiled language.
The aim of this paper is to show the implementation of XFEM in Matlab for the solution
of arbitrary multi-crack growth problems with the application of the minimum energy
crack growth criterion. The salient features of this implementation are as follows:
1. arbitrary number of enrichment layers over an element,
2. efficient updating of the system of equations,
3. crack growth with arbitrary intersections,
4. minimum energy crack growth solution,
5. internal pressure driven crack growth.
A significant advantage of updating the enrichment topology only where the fracture
geometry changes is that the non-linear problem of determining the crack extensions by
the minimum energy criterion can be solved computationally faster. On the other hand,
the limitation of the current XFEM implementation is that crack growth is not entirely
mesh independent because the crack extension length needs to be at least as big as the
tip enrichment radius. Although this is a general limitation of enriched discretisation
strategies, some heuristic treatments can be found in the literature, e.g. mapping the
crack tip enrichment functions along curved cracks [16, 5, 37]. The reader is referred to
[34] for a review of the partition of unity treatments of crack growth in 2D and 3D.

2 Aspects of implementation

2.1 Crack intersections

Generally, a Heaviside enrichment for a crack is introduced at a node if the nodal support
is completely cut by the crack. Care needs to be taken to determine if a node is enriched
in the vicinity of a crack junction; as two Heaviside enrichments will exist at some nodes,
it is important that the resulting enriched shape functions are unique and that there are
enough of them to describe the displacements in the vicinity of the crack junction.
A crack junction enrichment was first developed by [13, 4] and used in [10] to model
multi-crack growth with intersections. The enrichment took the form of a superposition

4
of two Heaviside enrichments – one for the intersecting (slave) crack and another for
the intersected (master ) crack. However, the Heaviside for the slave crack had to be
modified; the enrichment function assumed the value of the Heaviside of the slave crack
if a queried point lay on the side of the intersection and the Heaviside value of the master
crack if the queried point lay on the other side of the intersection. The first downside of
such an enrichment construct is the need to explicitly track the crack junction elements
and the modified Heaviside enrichment function for the slave crack. The second downside
is the lack of robustness of the enrichment method if intersection take place very close to
the tip of the master crack. Depending on which end of the master crack is intersected,
it may not be possible to enrich a node whose support is cut by both the slave and
the master crack since the discontinuity from the enrichment would then extend past
the crack tip of the master crack. However, the full enrichment of the nodes of a
junction element with the Heaviside for the slave crack is essential in order to adequately
reproduce the discontinuous displacements in the vicinity of the crack junction.
The present approach to the junction enrichment is in most cases equivalent to the
one used in [10]. However, the present strategy does not require a conceptually differ-
ent Heaviside enrichment and, consequently, the explicit tracking of the crack junction
elements. Instead, the standard Heaviside function is used to simulate the junction en-
richment for the slave crack. The key idea is to consider an imaginary branch of the
slave crack that is deflected along the master crack. The problem is then to determine
along which direction this imaginary branch should be deflected so that the junction
enrichment can be constructed optimally, i.e. all the nodes of the junction elements can
be enriched. Thus, the Heaviside-enriched nodes for the slave crack are selected as the
nodes whose support is cut by both the real and the imaginary branches of the slave
crack. On the other hand, the Heaviside enrichment for the master crack is unchanged.
Figure 1 shows a schematic diagram of the crack junction enrichment. The deflection
of the slave crack is normally chosen to be along the direction that yields an absolute
change in the angle relative to the crack increment direction of less than π/2. An ex-
ception to this rule is when the crack intersection is close to the tip of the master crack,
in which case the deflection of the slave crack is forced away from the tip of the master
crack. This way, the junction enrichment can be properly constructed, i.e. with all the
nodes of the crack junction elements enriched. On the other hand, the elements that
are cut only by the imaginary branch and that are the nodes of the junction elements
are the Heaviside blending elements for the slave crack since only some (not all) of these
elements’ nodes can be enriched. For the present purposes of junction enrichment, an
element can be regarded as a junction element if it is simultaneously cut by both the real
and the imaginary branches of the slave crack such that all of the elements nodes will be
enriched. Note that this can mean that the crack intersection point may lie outside some
of the junction elements (depending on the geometry of the crack intersection).
Concerning initially crossing cracks (i.e. ‘X’-type crack intersections), the superposition
of two Heaviside enrichments (for each of the two cracks) is insufficient to kinematically
decouple the four quadrants of the intersected region. One crack in the ’X’-type intersec-

5
 another slave crack
deflected along the
master crack

Heaviside enriched nodes for

the slave crack. Nodal support
is split by the real and the
imaginary branches of the
slave crack
Heaviside blending nodes for
the slave crack. The nodal
support is split only by the
salve crack’s imaginary
slave crack deflected branch; hence, the enriched
along master crack DOF’s will be set to zero

intersecting
(slave) crack

intersected
(master ) crack

Figure 1: An example of two minimum-distance crack intersections. Consider the inter-

secting (slave) crack on the left. A junction enrichment is introduced by first deflecting
the slave crack along the master crack. Subsequently, a Heaviside enrichment for the
slave crack is introduced at the nodes whose support is cut by both the real (i.e. in-
tersecting) and the imaginary (i.e. deflected) branches of the slave crack. The darker
squares indicate the nodes whose supports are fully enriched whereas the lighter squares
indicate the nodes whose supports are only partially enriched.

tion needs to be converted to two slave cracks, each of which needs to be deflected along
the master crack as discussed previously. The Heaviside junction enrichment is then
introduced for each of the two slave cracks. For propagating cracks, the following two
criteria are applied for determining when a crack intersection needs to be created:
ˆ The minimum distance criterion: intersection between crack A and crack B is
created once the distance between A’s tip and B’s surface (including B’s crack tip)
becomes less than a prescribed tolerance (e.g. A’s tip enrichment radius); crack A
is then extended normal to B’s surface.
ˆ The intersection criterion: intersection between a propagating crack A and
another crack B is created once it is detected that A’s tip extension crosses B’s
surface thereby creating an ‘X’-type intersection; crack A is then pulled back so
that its tip lies on B’s surface.
If an intersection point happens to lie very close to a crack node the intersection point is
snapped to the crack vertex. If the crack intersection occurs very close to the tip of the
master crack (e.g. the distance from the tip is smaller than the tip enrichment radius),
the tip enrichment is annihilated since its benefit would otherwise be diminished.

6
2.2 Enrichment tracking

Different types of enriched elements are tracked separately in order to facilitate the
enrichment updating process (e.g. extending or annihilating a certain enrichment) and to
facilitate assembly routines routines that are specific to different element types, such as:
tailoring element quadratures (e.g. for integrands that are: continuous/discontinuous,
low/high order, singular/non-singular), and introducing the enrichment functions within
an element in a particular way (e.g. using corrected branch enrichment functions [17]).
Therefore, the different element types of each enrichment function are identified as:
ˆ Heaviside enriched elements
1. standard (all nodes are enriched) elements:
cHvStd elm = cell(nCrack,1)
2. blending (not all nodes are enriched) elements:
cHvBln elm = cell(nCrack,2)
ˆ Branch enriched elements
1. standard crack-tip elements (partially cut by crack):
cBrStd eTp = cell(nCrack,2)
2. standard crack-split elements (fully cut by crack):
cBrStd eSp = cell(nCrack,2)
3. standard non-split (full) elements:
cBrBln eFl = cell(nCrack,2)
4. blending crack-split elements:
cBrBln eSp = cell(nCrack,2)
5. blending non-split (full) elements:
cBrBln eFl = cell(nCrack,2)
The element storage structure that we use is known as a cell in Matlab. We use the cell
to store a vector (a 1D array) of enriched elements of a particular type. For example,
the vector of standard Heaviside elements of crack i crk is stored in cHvStd elm{i crk}
whereas the blending branch split elements of crack i crk and crack tip i tip are stored
in cBrBln eSp{i crk,i tip}. In general, we refer to an element as an enriched element
if any of the element’s nodes is enriched. An enriched standard element is an element
that has all of its nodes enriched, where as an (enriched) blending element has only some
(not all) of its nodes enriched. In the proposed book-keeping of the different enriched
element types there is no need to explicitly track the Heaviside crack junction elements;
the crack junction enrichment is resolved naturally by considering that the slave crack
has an imaginary branch that is deflected along the master crack on intersection.

7
 0.5

0.4

0.3

0.2
y (mm) 0.1

-0.1

-0.2 tip element (branch enrichment)

non-split el. (branch standard enr.)
-0.3 non-split el. (branch blending enr.)
split el. (Heaviside std. enrichment)
-0.4 split el. (Heaviside std. & branch bln.)
split el. (Heaviside std. & branch std.)
-0.5 split el. (Heaviside bln. & branch std.)

0 0.2 0.4 0.6 0.8 1

x (mm)

Figure 2: Different types of enriched elements. The disk-like patch of branch-enriched

elements is centred at the crack tip. The branch enrichment is introduced at all the nodes
of these elements. As in the corrected XFEM approximation [17], the branch enrichment
is weighted by a ramp function that smoothly decreases over the last ring branch-enriched
elements. The ramp equals to unity in the interior of the branch-enriched domain and
zero on its boundary. Refer to Part-II for details on the XFEM approximation.

Of course, some elements will contain multiple overlapping enrichments. For instance,
the Heaviside and the branch enrichments generally appear along some portion of the
crack tip segment. Also, when two crack tips are in a sufficiently close proximity it is
possible for an element to be simultaneously enriched with two sets of branch functions.
Finally, when two cracks intersect, the intersected elements (i.e. the crack junction
elements) will need to be enriched with two Heaviside functions, e.g. refer to Figure 1.
In general, an allowance must be made for an element to be enriched an arbitrary number
of times. Hence, it is necessary to track each enriched element’s enrichment functions
and the generalised enriched nodes (for the element’s enriched shape functions). For this
purpose, the cell variable cLEnDt = cell(nElemn,1) (cell eLement ENrichment Data) is
introduced. The enrichment information of a particular enriched finite element i elm is
tracked by the matrix cLEnDt{i elm} = [3-by-n] where the number of columns in the
matrix (n) corresponds to the number of enrichments in the finite element and where
each column of the matrix lists the following information about the enrichment:
1. cLEnDt{i elm}[1,i enr] the crack identity of enrichment layer i enr
2. cLEnDt{i elm}[2,i enr] the enrichment function identity
3. cLEnDt{i elm}[3,i enr] the preferred quadrature scheme identity

8
Another cell structure cLNodE = cell(nElemn,1) (eLements’ Nodes Enriched) is used
to store the generalised enriched nodes of the enriched shape functions of each finite
element. An element of the cell is left empty if the finite element is not enriched. The
enriched nodes are permanently tied to the enriched shape functions. Thus, when some
enriched shape functions (and their enriched nodes) are annihilated during the updating
of the discrete system, the remaining enriched nodes are unaffected. Also, the generalised
enriched nodes are not recycled, i.e. new enriched shape functions are always tied to new
enriched nodes. This is done for practical reasons in order to avoid having to globally
update the cell cLNodE, which would be cumbersome in Matlab. The purpose of the
enriched nodes is to serve as indices into the columns of the generalised nodal degrees of
freedom matrix mNDofE = [2-by-nNdEnr max], where loosely speaking nNdEnr max =
max(cLNodE{1:nElemn}) is the largest enriched node number. The generalised DOFs
(unlike the enriched nodes) have to be updated in response to the annihilation of any
enriched shape functions (and their enriched nodes). This is because the enriched DOFs
are indices into the columns/rows of the stiffness matrix (and into the elements of the
force vector) and, as such, the DOFs must be numbered sequentially without gaps.
Consequently, the enriched DOFs are recycled every time there is an update in the
discrete system. Note that recycling old DOFs while introducing new enriched nodes
means that the matrix mNDofE tends to become sparser with increasing number of time
steps. However, the updating process of the matrix mNDofE is relatively simpler: the
columns of mNDofE corresponding to the annihilated enriched nodes are zeroed out,
mNDofE is enlarged by the number of columns equal to the number of new enriched nodes,
the DOFs corresponding to all current enriched nodes are numbered sequentially from
nNdStd*2+1 to nNdEnr*2, where the factor 2 is the number of DOFs per node (which is
2 in 2D). In particular, if vNdEnr del is the vector of deleted enriched nodes and vNdEnr
is the vector of current enriched nodes (where the vector is ordered sequentially with
vNdEnr(end) = nNdEnr max), then updating mNDofE is basically as follows:
1. mNDofE(:,vNdEnr del) = 0;
2. mNDofE(1,vNdEnr) = [(nNdStd*2+1):2:(nNdEnr*2-1)];
3. mNDofE(2,vNdEnr) = [(nNdStd*2+2):2:(nNdEnr*2)];
The computational times can be further improved in the updating of the system of
equations and in the post-processing of the numerical solution by pre-computing the
shape functions at quadrature points of the enriched finite elements. To this end, the
following cells are used: (suppose i elm is the enriched finite element, nGauss is the
number of quadrature points in the element, nLNodS is the number of nodes in the
element, and nLNodE is the number of generalised enriched nodes in the element.)
1. standard shape functions:
cGsEnr omgShS = cell(nElemn,1), where
cGsEnr omgShS{i elm} = [nGauss-by-nLNodS]
2. enriched shape functions:

9
 cGsEnr omgShE = cell(nElemn,1), where
cGsEnr omgShE{i elm} = [nGauss-by-nLNodE]
3. derivatives of standard shape functions:
cGsEnr omgDvS = cell(nElemn,1), where
cGsEnr omgDvS{i elm} = [2*nGauss-by-nLNodS]
4. derivatives of enriched shape functions:
cGsEnr omgDvE = cell(nElemn,1), where
cGsEnr omgDvE{i elm} = [2*nGauss-by-nLNodE]
5. weights of quadrature points:
cGsEnr omgWgt = cell(nElemn,1), where
cGsEnr omgWgt{i elm} = [nGauss-by-1]
The shape functions of element i elm are evaluated based on the quadrature-governing
enrichment. The identities of the preferred quadrature rules for each of the occurring
enrichments inside the finite element i elm are stored in vector cLEnDt{i elm}[3,:].
Among the preferred quadrature rules, which are defined for all enriched element types,
the critical rule is selected as the one that will result in no (or, at worst, minimal)
under-integration of the XFEM element stiffness matrix and force vector.

2.3 Numerical integration

Different types of elements of the enrichment functions (refer to Figure 2) require spe-
cialised quadrature schemes in order to perform the integration of the elemental equa-
tions adequately and efficiently [6, 38, 30] For example, integration over sub-cells should
be used if an element is cut by a crack since there will be a discontinuity in the integrands
across the crack interface. If the element contains a crack tip, an integration scheme that
is better suited for the singular integrands should be used [25, 11], such as: polar [25],
almost-polar [14] or the parabolic [28] integration scheme. Below is a summery of the
quadrature schemes used in the current implementation for triangular and quadrilateral
finite elements: (note that the schemes are enumerated in the order of precedence)
ˆ Linear triangle (T3)
1. Crack tip element with branch enrichment: use polar integration with the
number of points per sub-cell in the radial and angular directions nGsBrn pol=[13,7]
2. Completely split element with branch enrichment: use a standard quadrature
rule with the number of points per integration sub-cell nGsBrn sub=33
3. Completely split element with Heaviside enrichment: use a standard quadra-
ture rule with the number of points per integration sub-cell nGsHvi sub=1
4. Non-split element with branch enrichment: use a standard quadrature rule
with the number of points per element nGsBrn omg=33

10
 ˆ Bi-linear quadrilateral (Q4)
1. Crack tip element with branch enrichment: use polar integration with the
number of points per sub-cell in the radial and angular directions nGsBrn pol=[13,7]
2. Completely split element with branch enrichment: use a standardquadrature
rule with the number of points per integration sub-cell nGsBrn sub=33
3. Completely split element with Heaviside enrichment: use a standard quadra-
ture rule with the number of points per integration sub-cell nGsHvi sub=3
4. Non-split element with branch enrichment: use a standard quadrature rule
with the number of points per element nGsBrn omg=36
Concerning the elements with branch enrichment, the above quadrature schemes showed
good accuracy. Increasing the quadrature order did not notably change the strain energy
for the discretisations used in the numerical benchmarks. Note that we apply the same
quadrature scheme to both the standard and the blending branch-enriched elements.
Concerning the elements with multiple overlapping enrichments, a single quadrature
scheme is applied for computing all the enriched parts of the elemental equations. The
enrichment tracking variable cLEnDt lists the preferred quadrature rules for each en-
richment inside each element. Specifically, the vector cLEnDt{i elm}[3,:] contains the
preferred quadrature schemes for the element i elm. When there are multiple quadrature
schemes for an element with multiple layers of enrichment, the critical scheme is selected
based on the most demanding enrichment function (see the enumerations above).
For every discontinuous enrichment the crack intersection points with an element need to
be known in order to prepare the element for integration over sub-cells. Since, in general,
an element can be intersected multiple times, all crack intersections with each element
need to be found before the integration sub-cells can be computed. We use a cell variable
cXsElm = cell(nElemn,1) to track the intersection points between elements and cracks.
Recall that an intersection point is generalised to be any of these instances:
ˆ a point on a boundary of an element that is crossed by a crack
ˆ a point inside an element that is a crack vertex
ˆ a point inside an element that is a crack tip
Note that a crack intersection point inside an element does not need to tracked explicitly
because, in the current implementation, a slave (i.e. the intersecting) crack is deflected
along a master (i.e. intersected) crack such that the point of deflection is actually a crack
vertex point inside an element (i.e. the second bullet-point in the foregoing list).
Once all the intersection points between elements and cracks have been determined,
the integration sub-cells can be obtained based on a Delaunay triangulation. An im-
portant requirement of a triangulation is to respect crack boundaries. Matlab (2015b)
offers a built-in function that can perform a triangulation with respect to a set of edge
constraints. These can be specified in terms of the crack segments themselves.

11
We store the quadrature weights of enriched elements in a cell variable cGsEnr omgWgt.
The weights are actually scaled by a factor that is equal to the ratio of the Jacobian
at a quadrature point in the sub-cell to that in the element. The advantage is that the
numerical integration only involves multiplying the (adjusted) quadrature weight by the
determinant of the Jacobian of the element to obtained the discrete volume differential.
Consequently, the integrals can be evaluated using the same algorithm for all types of
enriched elements or quadrature schemes, which is very similar to the classic FEM.

3 Assembly of equations

3.1 Initiation of the discrete system

The elements to be enriched with the crack tip branch functions are relatively simple to
identify; as the current implementation assumes the enrichment domain to be a disk of
a fixed radius, the elements that have any of their nodes within this disk are enriched
at all their nodes. The elements that lie on the periphery of the enrichment domain
are classified as the blending elements (refer to Figure 2). The blending elements are
different from the other (standard ) branch-enriched elements in that the enrichment
function is weighted by a ramp function [17]. The ramp function is a smooth function
that is equal to unity at the nodes within the disk and zero outside the disk. Refer to
Part-II Appendix A.1 for details on the construction of this XFEM approximation.
The robust selection of elements for Heaviside enrichment can be more challenging,
especially in the presence of crack intersections. In general, a node is enriched with the
Heaviside function if its support is completely split by a crack. On the contrary, if the
nodal support covers the crack tip then the nodal support is not completely split by the
crack; hence, Heaviside enrichment at that node should not be introduced. When two
cracks intersect, the so-called slave (i.e. interesting) crack is deflected along the master
(i.e. intersected) crack. The deflected branch is called the imaginary branch and the
remainder of the crack is called the real branch. The imaginary branch is a fictitious
part of the slave crack that serves to smoothly blend the Heaviside enrichment of the
slave crack onto the master crack. A node is enriched with the Heaviside function of the
slave crack if the nodal support is completely split by the real branch or by both the real
and the imaginary branches of the slave crack. On the contrary, a node is not enriched
if its support is split only by the imaginary branch or if the support of the node covers a
crack tip.2 Finally, the Heaviside enrichment for the master crack is unchanged.
Once the elements to be enriched with a particular enrichment are found, the generalised
enriched nodes of these elements need to be determined. The first step is to reduce the
standard nodal connectivity matrix of the set of elements to be enriched to a reference
nodal connectivity matrix where the nodes are renumbered sequentially (without gaps)
2
In practice, an element is enriched at all its nodes but the unwanted enriched DOFs are set to zero.

12
starting from one. The second step is to offset the nodes in the reference nodal con-
nectivity matrix by the current largest enriched node number in the discrete system,
which we call nNdEnr max. This yields the enriched nodal connectivity matrix mLNodE
for the patch of enriched elements. Finally, the vector of the generalised enriched nodes
mLNodE[i elm,:] for each element i elm in the enriched patch needs to be appended
to the element’s global vector of all generalised enriched nodes cLNodE{i elm}. This
routine is repeated for any other enrichment functions over the same patch of elements,
such as in the case of branch enrichment where four enrichment functions are used.
Apart from tracking the generalised enriched nodes of the enriched elements, the gen-
eralised enriched DOFs also need to be tracked. For this we use a matrix mNDofE =
[2-by-nNdEnr max] whose rows correspond to the DOFs in the two spatial dimensions.
The nodal DOFs can be retrieved simply by using the generalised enriched nodes as
column indices. In our implementation, the matrix mNDofE tends to become more sparse
with increasing number of time-steps. This is because we do not recycle any enriched
nodes once they are deleted during the enrichment updating stage; instead, when a new
enrichment is introduced, the numbering of the new generalised enriched nodes is con-
tinued from the maximum enriched node number mNdEnr max. This means there will be
gaps in the numbering of the enriched nodes and that the nodes will need to be remapped
to their corresponding DOFs. This is relatively easy to do in comparison to renumbering
the generalised enriched nodes in the cell cLNodE for each enriched element.
The enrichment information of each element is stored in the cell variable cLEnDt. Every
time a new layer of enrichment is introduced, the enriched element’s enrichment tracking
matrix cLEnDt{i elm} is updated by appending a column vector that contains the fol-
lowing information about the new enrichment: (1) the crack identity, (2) the enrichment
function identity, and (3) the preferred quadrature scheme identity.
If an enriched element is cut by a crack, each intersection point between the crack and the
element needs to be stored. A cell variable cXsElm keeps track of the intersection points
for each element. For an element i elm the matrix cXsElm{i elm} is updated for each
new intersection point by appending the intersection point coordinates to the matrix.
Recall that a valid intersection point is generalised to be: (1) an intersection point on a
boundary of an element, (2) a crack vertex inside an element, or (3) a crack tip inside an
element. In the end, the matrix cXsElm{i elm} will be used to generate a discontinuity-
conforming Delaunay triangulation for numerical integration purposes.
The preferred integration scheme for each enrichment occurring in an element is specified
by an integer value. The integer value (also called the integration identity) is associated
to the required order of quadrature and to either a continuous or a discontinuous type of
integration. Once all enrichments over an element are known, the governing quadrature
scheme is selected based on the precedence of the schemes (refer to Section 2.3).
After the governing numerical integration scheme is identified, the element’s quadrature
points and weights need to be determined. To apply the same algorithm for the numerical
integration to elements that are either subdivided into sub-cells or not, all quadrature

13
points and weights need to be computed with respect to the parent coordinates of the
element (not the sub-cell). Thus, if an element requires integration over sub-cells, the
quadrature points and weights will need to be mapped from the parent coordinates of
each sub-cell to the parent coordinates of the element. Concerning the quadrature point
weights, the weights are computed by scaling each weight in the sub-cell by the ratio of
the determinant of the Jacobian of the sub-cell to that of the element. The scaled weights
are then stored in the cell variable cGsEnr omgWgt for the enriched element.
The next step is to evaluate each enriched element’s standard shape functions and shape
function derivatives at the quadrate points and store them in cell variables cGsStd omgShS
and cGsStd omgDvS respectively. Subsequently, the enriched shapes can be constructed
by computing the products of the standard shapes with the value of the enrichment
function for each quadrature point. This is repeated for any number of enrichment func-
tions that occur inside an element. The assembled enriched shapes are stored in the
cell variables cGsEnr omgShS and cGsEnr omgDvS. It is important to assemble the en-
riched shape functions in the order specified by the element’s enrichment tracking matrix
cLEnDt{i elm}; specifically, the information provided by cLEnDt{i elm}[1:2,:] gives
a matrix whose column vectors denote the crack and the enrichment identities.
The final step is to assembly the stiffness matrix and the generalised force vector. Con-
cerning the enriched elements, the assembly is similar to the classic FEM because each
enriched element is accompanied by the quadrature point weights and all the shape
functions and shape function derivatives pre-evaluated at the quadrature points.

3.2 Updating the discrete system

When a crack is extended, the previous tip enrichment needs to be annihilated. The
procedure of clearing the former branch enrichment involves the following steps:
1. Deleting the branch-enriched elements previously stored (refer to Section 2.2).
2. Deleting the subsets of the generalised enriched nodes and the enriched shape
functions of each element belonging to the former branch enrichment.
3. Deleting the rows and columns of the global equation system corresponding to the
enrichment DOFs at the former branch-enriched nodes.
4. Renumbering all enrichment DOFs to match the remaining enriched nodes.
After the former branch enrichment is annihilated, the branch enrichment needs to be
reintroduced considering the new crack tip configuration. This is essentially a repetition
of the steps described in the previous subsection. Updating the Heaviside enrichment,
on the other hand, is different since the new enrichment needs to be consistent with the
existing Heaviside enrichment. More specifically, the Heaviside enrichment for a crack
extension needs to be connected to the existing Heaviside enrichment along the remainder

14
of the crack. This is done by assigning the same generalised enriched nodes at the
interface between the existing and the new Heaviside-enriched patches of elements.
If a crack tip extension cuts an element that is already cut by the same crack, the cut
element’s stiffness matrix and force vector contributions to the global system of equations
need to be recomputed since the Heaviside enrichment is effectively different now. This
can be done by first subtracting the enriched parts of the element’s stiffness matrix
and force vector from global system of equations; then, considering the updated crack
geometry, recomputing the quadrature points and weights (based on the new integration
sub-cells), recomputing the standard and the enriched shape functions, recomputing the
enriched parts of element’s stiffness matrix and force vector and, finally, adding them
to the global system of equations. More generally, when a new enrichment needs to
be introduced to an element that is already enriched, we adopt the simplest approach
which is to recompute the entire enrichment from scratch for that element. Therefore,
our procedure of enriching an already enriched element involves the following steps:
1. Subtracting the element’s enriched parts of the stiffness matrix and and force vector
from the global system of equations.
2. Determining (if necessary) the enriched nodes for the new/updated enrichment
and appending them to the element’s vector of all enriched nodes.
3. Determining the appropriate integration strategy for the element and (if necessary)
recomputing the quadrature points and weights.
4. Re-evaluating the standard shape functions at the quadrature points and then
recomputing the enriched shape functions for each of the enrichment functions.
5. Computing the new enriched parts of the element’s stiffness matrix and force vector
and adding them to the global system of equations.
Even though it would be more efficient to compute only what is required, specifically: the
new enriched parts of the stiffness matrix and force vector due to the new enrichment,
the entire computational routine is not so trivial. More often than not, introducing a
new enrichment changes the preferred quadrature scheme; hence, it is usually necessary
to recompute the quadrate points and weights, re-evaluate the standard shape functions
and then all of the enriched shape functions. Both the standard and the enriched shape
functions need to be recomputed because the interactions between the standard shape
functions, the enriched shape functions of the existing enrichments, and the enriched
shape functions of the new enrichment appear in the enriched parts of the element’s
stiffness matrix. Resolving these interactions can be cumbersome in the general case.
Considering that a lot of computational work is necessary in updating the enriched parts
of the stiffness matrix anyway, we can avoid the last bit of cumbersomeness in computing
the interactions between the element’s shape functions in the element’s stiffness matrix
by simply recomputing the entire enriched part of the element’s stiffness matrix from
scratch using Matlab’s efficient vectorisation for matrix multiplications [20, 3, 26].

15
4 Numerical benchmarks

The numerical benchmarks have two aims: to verify our implementation of the enrich-
ment updating strategy, which we call the efficient implementation, and to assess its
effectiveness in terms of the computational speed-up relative to reassembling the entire
enrichment from scratch, which we call the basic implementation.
The so-called efficient approach updates the enrichment and the enriched part of the
stiffness matrix only where there are changes in the enrichment topology, such as due to
the growth and intersection of cracks; furthermore, it does so consistently with all exist-
ing enrichments. The so-called basic approach updates the enrichment by reassembling
all enrichments and the enriched part of the stiffness matrix from scratch, as in the initial
assembly (refer to Section 3.1). The basic approach is attractive mainly because achiev-
ing a robust implementation is much simpler in comparison to the efficient approach.
However, the basic implementation can lead to a major bottleneck in the computational
time when simulating multi-crack growth. Therefore, the efficient implementation is
particularity desirable for multi-crack growth as it can offer a significant speed-up in the
overall computational time even on moderate size meshes.
The benchmark cases assume a rectangular 10 × 1 plate with a roughly uniform distribu-
tion of cracks along its length. The domain is discretised using a uniform grid of bilinear
quadrilateral elements (Q4). We consider 4 meshes: 300 × 30, 600 × 60, 900 × 90 and
1200 × 120, and 6 randomly generated crack distributions of: 5, 10, 20, 30, 40 and 50
cracks of length 0.25. The rectangular plate is subjected to fixed-grip vertical extension
loading conditions. The maximum hoop stress crack growth criterion [15] is used.

4.1 Verification of enrichment updating

Firstly, the computer implementations of the efficient enrichment updating approach

is verified by showing that the numerical solutions to the fracture paths are close (in
terms of machine precision) to the fracture paths obtained using the basic enrichment
updating approach. Figure 3 shows some example fracture paths that were obtained for
the different initial crack distributions. Figure 4 shows the convergence of the fracture
paths as the mesh is refined for the test case of 50 randomly distributed cracks.
The maximum distance dmax between the fracture profiles obtained by the efficient and
the basic enrichment updating implementations is computed using equation (1):

r  
dmax = max max (∆xji )2 + (∆yij )2 (1)
j∈{1,2,...,ncrk } i∈{1,2,...,nvtx }

where the term in the parenthesis represents the square of the separation distance be-
tween corresponding vertex points of each crack. Table 1 summarises the results for the

16
maximum separation distance dmax between the final fracture profiles of each test case.
The discrepancies between the two implementations are close to machine precision.

Table 1: Maximum distance dmax between the fracture profiles obtained by the two
enrichment updating strategies, namely: the efficient and the basic approaches.

number of cracks
mesh size
5 10 20 30 40 50
300 × 30 3.99E-15 1.95E-15 2.13E-15 8.00E-15 2.67E-15 3.74E-15
600 × 60 9.17E-15 1.31E-14 7.26E-15 2.17E-14 2.15E-14 6.63E-15
900 × 90 2.85E-14 2.02E-14 1.08E-14 6.75E-14 8.97E-14 4.77E-14
1200 × 120 1.77E-14 2.37E-14 1.33E-14 8.45E-14 2.54E-13 4.84E-14

4.2 Computational speed-up

The total computational time of a simulation is composed of 5 parts, namely: pre-

processing, initial assembly, updated assembly, solution of the linear system and post-
processing. The first two parts are done only once whereas the last three need to be
repeated with each time-step. Therefore, in simulating multi-crack growth over many
times-steps the total computational time will be primarily composed of the solution, post-
processing and updating times. For the solution of the linear system we use Matlab’s
built-in direct solver (by calling the backslash operator “\”), which uses a Cholesky
decomposition that is optimised for sparse symmetric positive definite matrices. The
post-processing step involves mainly the computations of the stress intensity factors using
the domain-form interaction integral approach [40, 39, 21], identification of the crack tips
to grow and the management of crack coalescence. Our aim is to assess the computational
speed-up of the efficient implementation compared to the basic implementation.
Figure 5 shows the total computational time spent in updating the discrete system (tupd )
using the basic approach relative to the total computational time (ttot ) of the simulation.
Similarly, Figure 6 shows the total time spent in updating the discrete system using
the efficient approach. Figure 7 gives the speed-up factor of the total simulation time
achieved by the efficient approach relative to the basic approach. Finally, Figure 8 gives
the speed-up factor as a function of the number of time-steps (or crack increments).

4.3 Discussion and general remarks

The numerical results show that the computational benefits of the efficient enrichment
updating approach can be significant relative to the basic approach, especially for coarser
meshes or for larger numbers of cracks. While the basic implementation leads to a major
computational bottleneck in the total simulation time (refer to Figure 5), the efficient
implementation, in comparison, cuts down significantly on this time (refer to Figure 6)

17
giving a considerable speed-up factor in the total simulation time. In the studied cases,
the speed-up factor was from around 2 to 5 (refer to Figure 7). Although the speed-up
factor tends to diminish with mesh refinement, the computational benefit is considerable
even for moderate size problems of around a few hundred thousand DOFs.
We considered benchmark problems of multi-crack growth within a relatively thin strip
such that for the different mesh sizes and numbers of cracks that we studied there was a
large number of enriched elements relative to the total number of elements. This, in turn,
gave rise to more dramatic speed-up factors. However, with mesh refinement the number
of enriched elements will increase at a slower rate than the total number of elements.
Consequently, the times spent updating the enrichment and post-processing the solution
will further diminish relative to the time spent solving the linear system of XFEM
equations. In other words, with mesh refinement, the total computational time will be
dominated by the solution to the linear system, regardless of the implementation of the
enrichment updating. Nonetheless, for solving practical fracture mechanics problem the
efficient implementation is still very much desirable despite the diminishing return.
Finally, recall that our implementation relies a lot on computational memory to achieve
faster computations. Various enrichment related data are pre-computed to speed up the
enrichment updating and solution post-processing times. Although excessive reliance on
computational memory can create a bottleneck by itself, the proposed implementation
scales quite favourably with mesh refinement in terms of the memory usage. The ratio
of the number of enriched elements to the total number of elements will decrease as the
mesh is refined and, thus, the stored enrichment data will not grow at an outstanding
rate. For example, the memory required by large discrete system will be predominantly
used to store the global stiffness matrix followed by the typical mesh related data.

18
 Final fracture profile for nck = 5, nel = 300×30
0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Final fracture profile for nck = 10, nel = 300×30

0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Final fracture profile for nck = 20, nel = 300×30

0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Final fracture profile for nck = 30, nel = 300×30

0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Final fracture profile for nck = 40, nel = 300×30

0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Figure 3: Fracture profiles for: ncrk = {5, 10, 20, 30, 40}, nelm = 300×30. The sub-figures
show the type of fracture solutions obtained for each crack distribution. Note that the
crack extension length is proportional to the mesh size.

19
 Final fracture profile for nck = 50, nel = 300×30
0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Final fracture profile for nck = 50, nel = 600×60

0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Final fracture profile for nck = 50, nel = 900×90

0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Final fracture profile for nck = 50, nel = 1200×120

0.5
y (units)

−0.5
0 1 2 3 4 5 6 7 8 9 10
x (units)

Figure 4: Illustration of the convergence of the fracture profiles for the case ncrk = 50.
Note that the crack extension length is proportional to the mesh size.

20
 0.9

Time spent re−assmbling equations, tupd / ttot 0.85

0.8

0.75

0.7

0.65

0.6

0.55 300 × 30
600 × 60
0.5 900 × 90
1200 × 120
0.45
0 10 20 30 40 50
Number of cracks, ncrk

Figure 5: Total time spent updating the discrete system relative to the total computa-
tional time of the fracture simulation obtained by the basic implementation.

0.35
300 × 30
600 × 60
Time spent updating equations, tupd / ttot

0.3
900 × 90
1200 × 120
0.25

0.2

0.15

0.1

0.05

0
0 10 20 30 40 50
Number of cracks, ncrk

Figure 6: Total time spent updating the discrete system relative to the total computa-
tional time of the fracture simulation obtained by the efficient implementation.

21
 5.5

4.5
Speed up factor, ttot / t*tot

3.5

2.5 300 × 30
600 × 60
2 900 × 90
1200 × 120
1.5
0 10 20 30 40 50
Number of cracks, ncrk

Figure 7: Speed-up in the total simulation time obtained by the efficient implementation
relative to the basic one. (The superscript ”∗” denotes the basic implementation).

3.5
Speed up factor, ttot / t*tot

2.5
n = 50
crk

2 ncrk = 40
ncrk = 30
ncrk = 20
1.5
ncrk = 10
ncrk = 5
1
20 40 60 80 100 120 140 160 180
Number of crack increments, ncrk

Figure 8: Speed-up in the simulation time as a function of the number of time-steps. The
finest mesh (1200×120) is used. (The superscript ”∗” denotes the basic implementation).

22
5 Convergence of fracture paths

The maximum stress criterion is useful for solving linear-elastic fracture mechanics prob-
lems mainly because it is simple to apply. In Part-I and Part-II of this three-part paper
we developed the minimum energy approach to multi-crack growth within the XFEM
framework. It is of both theoretical and practical interest to compare the fracture paths
by the two criteria and to assess the converge of the solutions with mesh refinement.
This is the aim of the following benchmark cases involving multi-crack growth.

5.1 Square plate with 10 random cracks

The present test case is a square plate with 10 randomly distributed cracks where the
crack distribution is taken from [10]. Two types of boundary conditions are considered.
The first case assumes a plate subjected to a bi-axial extension. The second case assumes
internal pressure driven crack growth. In both cases, all cracks are allowed to grow at
the same rate. The fracture paths obtained for the bi-axially loaded plate are shown in
Figure 9. The apparent convergence of the fracture paths by the two criteria is plotted in
Figure 11 in terms of the L2 -norm of the distance between the crack surfaces. Similarly,
the fracture paths for the internal pressure driven cracks are shown in Figure 10. The
apparent convergence of the fracture paths by the two criteria is plotted in Figure 12.
Interestingly, the fracture paths obtained by two criteria appear to convergence towards
the same solution with mesh refinement. The bi-axially loaded case yields an average
convergence rate of 0.73 whereas the internal pressure loaded case yields 0.99.

5.2 Rectangular plate with 10 parallel cracks

We consider a rectangular plate with 10 initially non-overlapping parallel cracks dis-

tributed randomly within a narrow width. The plate is simply supported and subjected
to a uniform vertical tensile load. For the assumed crack distribution, crack growth is
found to occur from left to right leading to complete horizontal splitting of the plate.
The fracture solutions by the maximum hoop stress and the minimum energy criteria
are computed for two types of finite element meshes, namely: quadrilateral (Q4) and
triangular (T3) meshes. The fracture profiles obtained on Q4 and T3 meshes of similar
numbers of DOFs are shown in Figures 13 and 14 respectively. The convergence plots of
the fracture paths towards the same solution for Q4 and T3 meshes are shown in Figures
15 and 16, where the convergence rates are found to be 1.04 and 0.93 respectively.

23
 Fracture paths by different criteria Fracture paths by different criteria
(square plate subjected to a bi−axial extension with 10 randomly distributed cracks)
(square plate subjected to a bi−axial extension with 10 randomly distributed cracks)
2 Max hoop stress 2 Max hoop stress
Global energy min. Global energy min.
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
y

y
0.8 0.8
0.6 0.6

0.4 0.4

0.2 0.2
nmesh = 300×300, ∆a ∝ he 0 nmesh = 600×600, ∆a ∝ he
0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
x x

(a) Q4 mesh 300 × 300. (b) Q4 mesh 600 × 600.

Fracture paths by different criteria
(square plate subjected to a bi−axial extension with 10 randomly distributed cracks)

2 Max hoop stress

Global energy min.
1.8
1.6
1.4
1.2
1
y

0.8
0.6
0.4
0.2
0 nmesh = 1200×1200, ∆a ∝ he

0 0.5 1 1.5 2
x

(c) Q4 mesh 1200 × 1200.

Figure 9: Fracture paths by the maximum stress and the minimum energy criteria. The
plate is subjected to a bi-axial extension loading conditions. All cracks are allowed to
grow at the same rate.

24
 Fracture paths by different criteria Fracture paths by different criteria
(square plate with 10 randomly distributed cracks that are subjected to internal pressure)
(square plate with 10 randomly distributed cracks that are subjected to internal pressure)
2 Max hoop stress 2 Max hoop stress
Global energy min. Global energy min.
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
y

y
0.8 0.8
0.6 0.6

0.4 0.4

0.2 0.2
nmesh = 300×300, ∆a ∝ he 0 nmesh = 600×600, ∆a ∝ he
0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
x x

(a) Q4 mesh 300 × 300. (b) Q4 mesh 600 × 600.

Fracture paths by different criteria
(square plate with 10 randomly distributed cracks that are subjected to internal pressure)

2 Max hoop stress

Global energy min.
1.8
1.6
1.4
1.2
1
y

0.8
0.6
0.4
0.2
0 nmesh = 1200×1200, ∆a ∝ he

0 0.5 1 1.5 2
x

(c) Q4 mesh 1200 × 1200.

Figure 10: Fracture paths by the maximum stress and the minimum energy criteria. The
cracks are subjected to a uniform pressure loading conditions. All cracks are allowed to
grow at the same rate.

25
 Convergence to same fracture path by hoop−stress and energy−min. criteria
(square plate subjected to a bi−axial extension with 10 randomly distributed cracks)

L2−norm of distance between fracture surfaces

best fit (slope = 0.73)

−2
10

−3
10 −3 −2
10 10
element size, he

Figure 11: Convergence of the fracture paths by maximum stress and the minimum
energy criteria towards the same solution. The test case is a square plate with 10
randomly distributed cracks. The plate is subjected to a bi-axial extension loading
conditions. All cracks are allowed to grow at the same rate.

Convergence to same fracture path by hoop−stress and energy−min. criteria

(square plate with 10 randomly distributed cracks that are subjected to internal pressure)
L2−norm of distance between fracture surfaces

best fit (slope = 0.99)

−2
10

−3
10 −3 −2
10 10
element size, he

Figure 12: Convergence of the fracture paths by maximum stress and the minimum
energy criteria towards the same solution. The test case is a square plate with 10
randomly distributed cracks. The cracks are subjected to a uniform pressure loading
conditions. All cracks are allowed to grow at the same rate.

26
 Fracture paths by different criteria
(simply supported rectangular plate in vertical tension with 10 cracks in a narrow band)
0.3
Max hoop stress
0.2 Global energy min.

0.1

0
y

−0.1

−0.2
∆`inc ≈ 0.015
0 0.2 0.4 0.6 0.8 1
x
(a) Q4 mesh
Fracture × 200.criteria
100different
paths by
(simply supported rectangular plate in vertical tension with 10 cracks in a narrow band)
0.3
Max hoop stress
0.2 Global energy min.

0.1

0
y

−0.1

−0.2
∆`inc ≈ 0.008
0 0.2 0.4 0.6 0.8 1
x
(b) Q4 mesh
Fracture × 400.criteria
200different
paths by
(simply supported rectangular plate in vertical tension with 10 cracks in a narrow band)
0.3
Max hoop stress
0.2 Global energy min.

0.1

0
y

−0.1

−0.2
∆`inc ≈ 0.004
0 0.2 0.4 0.6 0.8 1
x
(c) Q4 mesh 400 × 800.

Figure 13: Fracture paths by the maximum stress and the minimum energy criteria. The
plate is simply supported and subjected to a uniform vertical tension loading conditions.
Q4 meshes are used.

27
 Fracture paths by different criteria
(simply supported rectangular plate in vertical tension with 10 cracks in a narrow band)
0.3
Max hoop stress
0.2 Global energy min.

0.1

0
y

−0.1

−0.2
∆`inc ≈ 0.002
0 0.2 0.4 0.6 0.8 1
x
(a) T3 mesh
Fracture × 200.criteria
100different
paths by
(simply supported rectangular plate in vertical tension with 10 cracks in a narrow band)
0.3
Max hoop stress
0.2 Global energy min.

0.1

0
y

−0.1

−0.2
∆`inc ≈ 0.001
0 0.2 0.4 0.6 0.8 1
x
(b) T3 mesh
Fracture × 400.criteria
200different
paths by
(simply supported rectangular plate in vertical tension with 10 cracks in a narrow band)
0.3
Max hoop stress
0.2 Global energy min.

0.1

0
y

−0.1

−0.2
∆`inc ≈ 0.0005
0 0.2 0.4 0.6 0.8 1
x
(c) T3 mesh 400 × 800.

Figure 14: Fracture paths by the maximum stress and the minimum energy criteria. The
plate is simply supported and subjected to a uniform vertical tension loading conditions.
T3 meshes are used.

28
 Convergence to same fracture path by hoop−stress and energy−min. criteria
(simply supported rectangular plate in vertical tension with 10 cracks in a narrow band)

L2 -norm of distance between fracture surfaces

best fit (slope = 1.04)
−2
10

−3
10

−4
10 −3 −2
10 10
Crack increment length, ∆`inc

Figure 15: Convergence of the fracture paths by maximum stress and the minimum
energy criteria towards the same solution. The test case is a simply supported plate
in vertical tension with 10 narrowly distributed parallel and initially non-overlapping
horizontal cracks. Q4 meshes were used.

Convergence to same fracture path by hoop−stress and energy−min. criteria

(simply supported rectangular plate in vertical tension with 10 cracks in a narrow band)
L2 -norm of distance between fracture surfaces

best fit (slope = 0.93)

−3
10

−4
10 −3
10
Crack increment length, ∆`inc

Figure 16: Convergence of the fracture paths by maximum stress and the minimum
energy criteria towards the same solution. The test case is a simply supported plate
in vertical tension with 10 narrowly distributed parallel and initially non-overlapping
horizontal cracks. T3 meshes were used.

29
5.3 A numerical improvement to the growth direction

The preceding numerical results indicate that the solutions to the fracture path by the
maximum stress and the minimum energy criteria converge towards very similar solutions
when compared on the global length scale. For the range of XFEM discretions that were
studied such convergence behaviours of the fracture paths were clearly observed. These
results were obtained by the two criteria despite their inherent differences in the solutions
to the incipient crack tip kink angles under general mixed-mode loading conditions, as
demonstrated by the benchmark cases in Part-II of this three-part paper.
From inspection of the discrete solutions (including the solution to the benchmark cases
from Part-II) it appears that the maximum stress criterion consistently underestimates
the crack kink angles whereas the minimum energy solution tends to overestimate them.
Refining the XFEM discretisation (and, thus, decreasing the crack extension length)
leads the fracture solutions by the two criteria to convergence from opposite directions
towards each other and towards a solution that lies close to the middle of the solutions
obtained by the two criteria on coarser discretisations. Motivated by these observations,
it may be possible to numerically improve the crack growth direction by assuming the
average of the directions obtained by the stress and the energy based criteria.
It turns out that the accuracy and the convergence rate of the numerical solution to the
fracture paths can be significantly improved by this approach. For the sake of brevity,
we will refer to the proposed modification as the bi-section method. Satisfactory perfor-
mance of this method was obtained in all the test cases that were attempted. The fracture
solutions by the different criteria, namely: the maximum stress, the minimum energy
and the bi-section method are presented in the subsequent benchmark studies.

5.4 Comparison of fracture paths by different criteria

Several 2D benchmark cases are proposed for the comparison the fracture paths by the
three criteria, namely: maximum stress, minimum energy, and the proposed bi-section
method which assumes the average direction obtained by maximum stress and minimum
energy criteria at each time-step. The fracture paths for each test case are computed
for different XFEM discretisation densities. It is found in every case that the maxi-
mum stress and minimum energy criteria converge towards very similar fracture paths;
however, the bi-section method appears to converge to this solution the fastest.
The subsequent results of the benchmark cases verify that the bi-section method can be
useful for improving the accuracy and the convergence rate of the fracture paths with
mesh refinement. The bi-section method is most effective for fracture evolutions that
do not involve crack intersections. The reason is that the position of a crack intersec-
tion tends to have a strong influence on the subsequent fracture paths (particularly on
courser discretisations). The benefits of the bi-section approach are less apparent if the

30
discretisation is already well refined because the fracture paths by the stress and the
energy based criteria tend to already be in a very close proximity, e.g. Figure 20.

6 Summary

It was shown via solutions to multiple benchmark problems of multi-crack growth with
coalescence that the efficient enrichment updating approach was implemented within
Matlab correctly as the fracture solutions coincided within machine precision with those
obtained by the basic enrichment updating approach, which essentially involved reassem-
bling the entire enrichment and recomputing the entire enriched part of the stiffness
matrix from scratch at each time-step. With regard to the computational times, it was
shown that the so-called basic implementation created a major computational bottleneck
in the total simulation time, e.g. Figure 5. On the other hand, the so-called efficient
implementation resulted in significantly faster computational times, e.g. Figure 6.
One of the contributions of this paper is the comparison of different crack growth criteria.
The maximum stress and minimum energy criteria – despite their inherent differences in
the incipient crack tip kink angles under general mixed-mode loading conditions – were
found to yield very similar fracture solutions when compared on the global length scale.
From the multiple benchmark cases that were studied, the fracture solutions converged
to very similar fracture paths and at similar rates but from opposite directions.
Finally, a numerical improvement to the crack growth direction was proposed. The so-
called bi-section method averaged the crack growth directions obtained by the maximum
stress and minimum energy criteria. Despite the method’s lack of a sound physical basis,
it was found to be useful for numerically improving the accuracy of either criterion and for
speeding-up the convergence of the fracture paths with discretisation refinement.

7 Supplementary material

The open-source code XFEM Fracture2D and supporting material can be found here:
ˆ XFEM Fracture2D: https://figshare.com/s/0b4394e8fab7191d2692
ˆ competing cracks: https://figshare.com/s/4a7dd5fb0a8634c9fae4
ˆ demo screenshots: https://figshare.com/s/6397737c78beb59f3b58
ˆ demo movies: https://figshare.com/s/73d7b50a7729070c2173

31
Double cantilever problem

Fracture paths by different criteria

(double cantilever problem with an edge crack offset by 0.01 above the x−axis)

0.5 Max hoop stress

Global energy min.
0.4
Averaged direction
0.3

0.2
y

0.1

0
nelm = {30×60, 60×120, 120×240, 240×480}, ∆a ∝ he
−0.1
0 0.2 0.4 0.6 0.8 1 1.2
x

Figure 17: Fracture paths by different growth criteria for the double cantilever problem
with the initial crack positioned 0.01 above the x-axis. The prying action is exerted by
prescribed displacements on the left edge.

Two edge crack problem (simple tension loading)

Fracture paths by different criteria

(simply supported square plate in vertical tension with two edge cracks: ∆x=0.6, ∆y=0.1)
0.25
Max hoop stress
0.2 Global energy min.
Averaged direction
0.15

0.1

0.05
y

−0.05

−0.1

−0.15
nelm = {60×60, 120×120, 240×240}, ∆a ∝ he
−0.2
0.2 0.3 0.4 0.5 0.6 0.7 0.8
x

Figure 18: Fracture paths by different criteria for a simply supported square plate (1×1)
in simple vertical tension with two initial edge cracks (a1 = a2 = 0.2). Crack-tip
separation: ∆x = 0.6, ∆y = 0.10.

32
Three crack problem (pressure loaded centre crack)

Fracture paths by different criteria

(simply supported cracked square plate with a pressure loaded center crack)
0.3
Max hoop stress
0.2 Global energy min.
Averaged direction
0.1

0
y

−0.1

−0.2
nelm = {50×50, 100×100, 200×200}, ∆a ∝ he

0 0.2 0.4 0.6 0.8 1

x

Figure 19: Fracture paths by different growth criteria for a simply supported square
plate with three pre-existing cracks, where the centre crack is subjected to a pressure
load acting normal to the crack surface.

Two cracks protruding from holes (vertical tension load)

Fracture paths by different criteria

(simply supported square plate in vertical tension with two cracks protruding from holes)

0.3 Max hoop stress

Global energy min.
0.2 Averaged direction

0.1

0
y

−0.1

−0.2

−0.3 nnod = {3k, 15k, 50k}, ∆a ∝ he

2 2.2 2.4 2.6 2.8 3

x

Figure 20: Fracture paths by different growth criteria for a simply supported square
plate with a pair of initial cracks protruding from holes.

33
Two edge crack problem (crack pressure loading)
Fracture paths by different criteria
(simply supported square plate with two edge cracks loaded with pressure: ∆x=0.6, ∆y=0.12)

0.5 Max hoop stress

Global energy min.
0.4
Averaged direction
0.3

0.2

0.1

0
y

−0.1

−0.2

−0.3

−0.4

−0.5 nelm = {50×50, 100×100, 200×200, 400×400}, ∆a ∝ he

0 0.2 0.4 0.6 0.8 1
x
Figure 21: Fracture paths by different criteria for a simply supported square plate (1×1)
with two initial edge cracks (a1 = a2 = 0.2) that are loaded by pressure. Crack-tip
separation: ∆x = 0.6, ∆y = 0.12.
Fracture paths by different criteria
(simply supported square plate with two edge cracks loaded with pressure: ∆x=0.6, ∆y=0.08)

0.5 Max hoop stress

Global energy min.
0.4
Averaged direction
0.3

0.2

0.1

0
y

−0.1

−0.2

−0.3

−0.4

−0.5 nelm = {50×50, 100×100, 200×200, 400×400}, ∆a ∝ he

0 0.2 0.4 0.6 0.8 1
x
Figure 22: Fracture paths by different criteria for a simply supported square plate (1×1)
with two initial edge cracks (a1 = a2 = 0.2) that are loaded by pressure. Crack-tip
separation: ∆x = 0.6, ∆y = 0.08.

34
Two-edge crack problem (crack pressure loading) [cont.]

Fracture paths by different criteria

(simply supported square plate with two pressure loaded edge cracks: ∆x=0.6, ∆y=0.04)

Max hoop stress

0.2
Global energy min.
Averaged direction
0.1

0
y

−0.1

−0.2 n = {100×100, 200×200, 400×400}, ∆a ∝ h

elm e

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

x

Figure 23: Fracture paths by different criteria for a simply supported square plate (1×1)
with two initial edge cracks (a1 = a2 = 0.2) that are loaded by pressure. Crack-tip
separation: ∆x = 0.6, ∆y = 0.04.

Fracture paths by different criteria

(simply supported square plate with two pressure loaded edge cracks: ∆x=0.6, ∆y=0.02)

0.15
Max hoop stress
Global energy min.
0.1 Averaged direction

0.05

0
y

−0.05

−0.1

nelm = {100×100, 200×200, 400×400}, ∆a ∝ he

−0.15
0.3 0.4 0.5 0.6 0.7 0.8
x

Figure 24: Fracture paths by different criteria for a simply supported square plate (1×1)
with two initial edge cracks (a1 = a2 = 0.2) that are loaded by pressure. Crack-tip
separation: ∆x = 0.6, ∆y = 0.02.

35
The PMMA beam with a bottom slit (case-1)
5

y (in)
0

(−5,−2.5)
initial crack
expermental
−5
−10 −5 0 5 10
x (in)
(a) Schematic of the PMMA beam with a bottom crack; A concentrated point
load is applied in the middle of the top face.
Test case: a = 1.5 (in), b = 5 (in)

0.9

0.8
y (in)

0.7

0.6
max hoop
energy min.
0.5
averaged

−4 −3.9 −3.8 −3.7 −3.6 −3.5 −3.4 −3.3 −3.2

x (in)

(b) A close-up view of the fracture paths around the middle hole of the beam.
Test case: a = 1.5 (in), b = 5 (in)
−0.85

−0.9

−0.95
y (in)

−1

−1.05

−1.1 max hoop

energy min.
averaged

−4.4 −4.35 −4.3 −4.25 −4.2 −4.15 −4.1 −4.05 −4

x (in)

(c) A close-up view of the fracture paths around the bottom hole of the beam.
Figure 25: A simply supported 4 × 10 (in) PMMA beam with an initial vertical slit
of length a = 1.5 (in) and a point load mid-way the top-edge. The fracture paths by
different criteria are compared.

36
The PMMA beam with a bottom slit (case-2)
5

y (in)
0
(−6,−1.5)

initial crack
expermental
−5
−10 −5 0 5 10
x (in)
(a) Schematic of the PMMA beam with a bottom crack; A concentrated point
load is applied in the middle of the top face.
Test case: a = 2.5 (in), b = 6 (in)

2.6

2.5

2.4
y (in)

2.3

2.2

max hoop
2.1 energy min.
averaged

−4.1 −4 −3.9 −3.8 −3.7 −3.6 −3.5

x (in)
(b) A close-up view of the fracture paths around the top hole of the beam.
Test case: a = 2.5 (in), b = 6 (in)

1.2

1.1

1
y (in)

0.9

0.8

max hoop
0.7 energy min.
averaged

−4.7 −4.6 −4.5 −4.4 −4.3 −4.2 −4.1 −4

x (in)
(c) A close-up view of the fracture paths around the middle hole of the beam.
Figure 26: A simply supported 4 × 10 (in) PMMA beam with an initial vertical slit
of length a = 2.5 (in) and a point load mid-way the top-edge. The fracture paths by
different criteria are compared.

37
A pre-cracked plate with two holes

(a) Schematic diagram of paths

Fracture the pre-cracked part. (Source: [9])
by different criteria
(rectangular plate with two holes and two edge cracks subjected to vertical extension)
6
Max hoop stress
4 Global energy min.
Averaged direction
2

0
y

−2

−4
nnod = {20k, 80k, 320k}, ∆a ∝ he
−6
−10 −5 0 5 10
x

(b) Outline of fracture paths by different criteria.

Fracture paths by different criteria
(rectangular plate with two holes and two edge cracks subjected to vertical extension)
1.5
Max hoop stress
Global energy min.
Averaged direction
1

0.5
y

nnod = {20k, 80k, 320k}, ∆a ∝ he

−0.5
4.5 5 5.5 6 6.5 7 7.5
x

(c) Close-up view of sub-figure (b) around the hole on the right.

Figure 27: Fracture paths by different criteria for a rectangular plate with two holes and
two edge cracks subjected to a vertical extension.

38
References

[1] K. Agathos, E. Chatzi, and S. P. A. Bordas. “Stable 3D extended finite elements

with higher order enrichment for accurate non planar fracture”. In: Computer
Methods in Applied Mechanics and Engineering 306.October 2015 (2016), pp. 19–
46. issn: 00457825.
[2] K. Agathos et al. “A well-conditioned and optimally convergent XFEM for 3D
linear elastic fracture”. In: International Journal for Numerical Methods in Engi-
neering 105.9 (2016), pp. 643–677. issn: 00295981.
[3] I. Anjam and J. Valdman. “Fast MATLAB assembly of FEM matrices in 2D and
3D: Edge elements”. In: Applied Mathematics and Computation 267.13 (2015),
pp. 252–263. issn: 00963003. arXiv: arXiv:1409.4618v1.
[4] T. Belytschko et al. “Arbitrary discontinuities in finite elements”. In: International
Journal for Numerical Methods in Engineering 50.4 (2001), pp. 993–1013. issn:
0029-5981.
[5] T. Belytschko and T. Black. “Elastic crack growth in finite elements with minimal
remeshing”. In: International Journal for Numerical Methods in Engineering 45.5
(1999), pp. 601–620. issn: 0029-5981.
[6] T. Belytschko, R. Gracie, and G. Ventura. “A review of extended/generalized
finite element methods for material modeling”. In: Modelling and Simulation in
Materials Science and Engineering 17.4 (2009), p. 043001. issn: 0965-0393.
[7] E. Benvenuti et al. “Variationally consistent eXtended FE model for 3D planar
and curved imperfect interfaces”. In: Computer Methods in Applied Mechanics
and Engineering 267 (2013), pp. 434–457. issn: 00457825.
[8] S. P. A. Bordas et al. “An extended finite element library”. In: International
Journal for Numerical Methods in Engineering 71.6 (2007), pp. 703–732. issn:
00295981.
[9] P. Bouchard, F. Bay, and Y. Chastel. “Numerical modelling of crack propagation:
automatic remeshing and comparison of different criteria”. In: Computer Meth-
ods in Applied Mechanics and Engineering 192.35-36 (2003), pp. 3887–3908. issn:
00457825.
[10] E. Budyn et al. “A method for multiple crack growth in brittle materials without
remeshing”. In: International Journal for Numerical Methods in Engineering 61.10
(2004), pp. 1741–1770. issn: 0029-5981.
[11] A. Cano and C. Moreno. “Transformation Methods for the Numerical Integration
of Three-Dimensional Singular Functions”. In: Journal of Scientific Computing
(2016). issn: 0885-7474.
[12] J. Chessa, H. Wang, and T. Belytschko. “On the construction of blending elements
for local partition of unity enriched finite elements”. In: International Journal for
Numerical Methods in Engineering 57.7 (2003), pp. 1015–1038. issn: 0029-5981.
[13] C. Daux et al. “Arbitrary branched and intersecting cracks with the extended finite
element method”. In: International Journal for Numerical Methods in Engineering
48.12 (2000), pp. 1741–1760. issn: 0029-5981.

39
[14] M. G. Duffy. “Quadrature Over a Pyramid or Cube of Integrands with a Singular-
ity at a Vertex”. In: SIAM Journal on Numerical Analysis 19.6 (1982), pp. 1260–
1262. issn: 0036-1429.
[15] F. Erdogan and G. C. Sih. “On the Crack Extension in Plates Under Plane Loading
and Transverse Shear”. In: Journal of Basic Engineering 85.4 (1963), p. 519. issn:
00219223.
[16] M. Fleming and Y. A. Chu. “Enriched element-free Galerkin methods for crack tip
fields”. In: International Journal for Numerical Methods in Engineering 40.Jan-
uary 1996 (1997), pp. 1483–1504.
[17] T. P. Fries. “A corrected XFEM approximation without problems in blending
elements”. In: International Journal for Numerical Methods in Engineering 75.5
(2008), pp. 503–532. issn: 00295981.
[18] T. P. Fries. “Overview and comparison of different variants of the XFEM”. In:
PAMM 14.1 (2014), pp. 27–30. issn: 16177061.
[19] T. P. Fries and T. Belytschko. “The extended/generalized finite element method:
An overview of the method and its applications”. In: International Journal for
Numerical Methods in Engineering (2010), n/a–n/a. issn: 00295981.
[20] E. Gekeler. Mathematical Methods for Mechanics. 2008, p. 636. isbn: 9783540302674.
arXiv: arXiv:1011.1669v3.
[21] V. F. González-Albuixech et al. “Domain integral formulation for 3-D curved and
non-planar cracks with the extended finite element method”. In: Computer Meth-
ods in Applied Mechanics and Engineering 264 (2013), pp. 129–144. issn: 00457825.
[22] V. Gupta et al. “Stable GFEM (SGFEM): Improved conditioning and accuracy of
GFEM/XFEM for three-dimensional fracture mechanics”. In: Computer Methods
in Applied Mechanics and Engineering 289 (2015), pp. 355–386. issn: 00457825.
[23] V. Gupta and C. A. Duarte. “On the enrichment zone size for optimal conver-
gence rate of the Generalized/Extended Finite Element Method”. In: Computers
& Mathematics with Applications (2016). issn: 08981221.
[24] R. Huang, N. Sukumar, and J. H. Prévost. “Modeling quasi-static crack growth
with the extended finite element method Part II: Numerical applications”. In:
International Journal of Solids and Structures 40.26 (2003), pp. 7539–7552. issn:
00207683.
[25] P. Laborde et al. “High-order extended finite element method for cracked do-
mains”. In: International Journal for Numerical Methods in Engineering 64.3
(2005), pp. 354–381. issn: 0029-5981.
[26] S. Linge and H. P. Langtangen. Programming for Computations - MATLAB/Octave.
Vol. 14. Texts in Computational Science and Engineering. Cham: Springer Inter-
national Publishing, 2016. isbn: 978-3-319-32451-7.
[27] J. Melenk and I. Babuška. “The partition of unity finite element method: Basic
theory and applications”. In: Computer Methods in Applied Mechanics and Engi-
neering 139.1-4 (1996), pp. 289–314. issn: 00457825.

40
[28] H. Minnebo. “Three-dimensional integration strategies of singular functions in-
troduced by the XFEM in the LEFM”. In: International Journal for Numerical
Methods in Engineering 92.13 (2012), pp. 1117–1138. issn: 00295981.
[29] N. Moës, J. Dolbow, and T. Belytschko. “A finite element method for crack growth
without remeshing”. In: Int. J. Numer. Meth. Engng 150.February (1999), pp. 131–
150. issn: 00295981.
[30] S. Natarajan. “Enriched finite element methods: advances & applications”. PhD
thesis. Cardiff University, 2011.
[31] V. P. Nguyen. “An object-oriented approach to the extended finite element method
with applications to fracture mechanics”. MSc Thesis. Hochiminh City University
of Technology, 2005.
[32] V. P. Nguyen et al. “Isogeometric analysis: An overview and computer implemen-
tation aspects”. In: Mathematics and Computers in Simulation 117 (2015), pp. 89–
116. issn: 03784754.
[33] V. P. Nguyen et al. “Meshless methods: A review and computer implementation
aspects”. In: Mathematics and Computers in Simulation 79.3 (2008), pp. 763–813.
issn: 03784754.
[34] T. Rabczuk, S. Bordas, and G. Zi. “On three-dimensional modelling of crack
growth using partition of unity methods”. In: Computers & Structures 88.23-24
(2010), pp. 1391–1411. issn: 00457949.
[35] H. Talebi, M. Silani, and T. Rabczuk. “Concurrent multiscale modeling of three
dimensional crack and dislocation propagation”. In: Advances in Engineering Soft-
ware 80.C (2015), pp. 82–92. issn: 09659978.
[36] H. Talebi et al. “A computational library for multiscale modeling of material fail-
ure”. In: Computational Mechanics 53.5 (2014), pp. 1047–1071. issn: 0178-7675.
[37] G. Ventura, E. Budyn, and T. Belytschko. “Vector level sets for description of
propagating cracks in finite elements”. In: International Journal for Numerical
Methods in Engineering 58.10 (2003), pp. 1571–1592. issn: 0029-5981.
[38] G. Ventura, R. Gracie, and T. Belytschko. “Fast integration and weight function
blending in the extended finite element method”. In: International Journal for
Numerical Methods in Engineering 77.1 (2009), pp. 1–29. issn: 00295981.
[39] M. C. Walters, G. H. Paulino, and R. H. Dodds. “Interaction integral procedures for
3-D curved cracks including surface tractions”. In: Engineering Fracture Mechanics
72.11 (2005), pp. 1635–1663. issn: 00137944.
[40] J. F. Yau, S. S. Wang, and H. T. Corten. “A Mixed-Mode Crack Analysis of
Isotropic Solids Using Conservation Laws of Elasticity”. In: Journal of Applied
Mechanics 47.2 (1980), p. 335. issn: 00218936.

41
Highlights

* We describe the implementation of XFEM in Matlab with emphasis on the design of the code to
enable fast and efficient computational times of fracture problems involving many cracks and
arbitrary crack intersections. The efficiency and robustness of the implementations are tested via
the solutions to several benchmark problems involving multi-crack growth with coalescence.

* We carry out verification studies of the implementation of the minimum energy criterion and give
comparisons with the maximum tension criterion. The comparisons reveal that both criteria
converge to virtually the same fracture solutions albeit from opposite directions. In other words, it is
found that the converged fracture path tends to lie in between those obtained by each criterion on
coarser meshes.

* From the convergence trends observed in the numerical solutions by different crack growth
criteria, we propose a numerical improvement to the crack growth direction criterion which is to
assume the average direction of the directions obtained by the maximum tension and the minimum
energy criteria. The fracture paths obtained by the modified criterion show significant improvements
in accuracy and convergence rates of the fracture paths, as observed from a multitude of test cases.

* Finally, we make available the open-source Matlab code including documentation, benchmark and
example cases as supplementary material for the verification of our results and with the hope that
the code can be useful in general.