Authors Accepted Manuscript

A new equi-dimensional fracture model using

polyhedral cells for microseismic data sets

Omar Al-Hinai, Rencheng Dong, Sanjay

Srinivasan, Mary F. Wheeler

www.elsevier.com/locate/petrol

PII: S0920-4105(17)30419-9
DOI: http://dx.doi.org/10.1016/j.petrol.2017.04.004
Reference: PETROL3944
To appear in: Journal of Petroleum Science and Engineering
Received date: 31 March 2016
Revised date: 28 March 2017
Accepted date: 5 April 2017
Cite this article as: Omar Al-Hinai, Rencheng Dong, Sanjay Srinivasan and Mary
F. Wheeler, A new equi-dimensional fracture model using polyhedral cells for
microseismic data sets, Journal of Petroleum Science and Engineering,
http://dx.doi.org/10.1016/j.petrol.2017.04.004
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 galley proof before it is published in its final citable 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.
A new equi-dimensional fracture model using polyhedral cells for microseismic data
sets

Omar Al-Hinaia,, Rencheng Donga , Sanjay Srinivasanb , Mary F. Wheelera

a Center for Subsurface Modeling, University of Texas at Austin, United States

b Department of Energy and Mineral Engineering, Pennsylvania State University, United States

Abstract

We present a method for modeling ow in porous media in the presence of complex fracture networks. The
approach utilizes the Mimetic Finite Dierence (MFD) method. We employ a novel equi-dimensional approach for
meshing fractures. By using polyhedral cells we avoid the common challenge in equi-dimensional fracture modeling of
creating small cells at the intersection point. We also demonstrate how polyhedra can mesh complex fractures without
introducing a large number of cells. We use polyhedra and the MFD method a second time for embedding fracture
boundaries in the matrix domain using a cut-cell paradigm. The embedding approach has the advantage of being
simple and localizes irregular cells to the area around the fractures. It also circumvents the need for conventional
mesh generation, which can be challenging when applied to complex fracture geometries. We present numerical results
conrming the validity of our approach for complex fracture networks and for dierent ow models. In our rst
example, we compare our method to the popular dual-porosity technique. Our second example compares our method
with directly meshed fractures (single-porosity) for two-phase ow. The third example demonstrates two-phase ow for
the case of intersecting ellipsoid fractures in three-dimensions, which are typical in microseismic analysis of fractures.
Finally, we demonstrate our method on a two-dimensional fracture network produced from microseismic eld data.
Keywords: Mimetic Finite Dierence method, Discrete fracture model, Microseismic

1. Introduction

Fractures play a signicant role in porous media uid ows. As a result, there is an enormous body of literature
devoted to ow modeling in fractured media. The most common approaches rely on homogenization or upscaling of
the fracture network. The most notable example of upscaling is the dual-porosity/dual-permeability method (DPDP)
(Arbogast et al., 1990; Warren and Root, 1963). In the DPDP model, fractures are represented as a continuous medium
within a computational cell and communicate with the rock matrix through interaction terms and shape factors. The
DPDP approach has been widely used in reservoir simulation, as it is fast and relatively simple to adopt. The main
challenge of DPDP is in incorporating complex physical processes. Much like other upscaling approaches, it is often
dicult to upscale non-linear eects such as gravity and capillary pressure.
In contrast, discrete fracture models (DFM) directly represent fracture networks in the computational mesh, ac-
counting for fracture geometry and connectivity. By doing so, DFM can solve for complex physical processes without
the need to upscale those eects. In the DFM setting, one typically has the option of either an equi-dimensional or a
lower-dimensional representation of fractures. In the equi-dimensional case, fractures are represented directly in the
computational mesh with the same dimensionality of the overall problem. Examples of this approach can be found in
(Hgland et al., 2009; Ochs et al., 2002). Maintaining the dimensionality of the fracture domain has the advantage of

Corresponding author
Email addresses: ohinai@gmail.com (Omar Al-Hinai), rencheng_dong@utexas.edu (Rencheng Dong), sanjays@psu.edu (Sanjay
Srinivasan), mfw@ices.utexas.edu (Mary F. Wheeler)

Preprint submitted to Journal of Petroleum Science and Engineering

 (a) Equi-dimensional (b) Lower-dimensional (c) Our approach

Figure 1: Three dierent approaches for discrete fracture modeling. (a) The equi-dimensional approach represents fractures directly in the
computational mesh while maintaining the same dimensionality as the surrounding matrix. (b) The lower-dimensional method
represents fractures at the faces of adjacent cells. (c) Our approach embeds fractures at the cell faces, and represents the ow in
the fracture using a full dimensional mesh with special boundary conditions.

directly representing the full geometry of the fracture, including complex features such as curvature and intersections.
On the other hand, the lower-dimensional approach relies on a lower-dimensional representation of fractures at mesh
faces. Examples of this approach can be found in (Ahmed et al., 2015; Alboin et al., 2002; Hoteit and Firoozabadi,
2008; Karimi-Fard et al., 2004). An illustration of these two approaches is shown in [Fig. 1(a, b)].
Many authors advocate for the lower-dimensional representation, citing simpler meshes and the ability to avoid
small cells at fracture intersections. In this work, we make the case for a modied version of the equi-dimensional
representation of fractures. Our approach can avoid the creation of small cells at the intersection by taking advantage
of polyhedral cells in the mesh. The equi-dimensional representation allows us to consider the fracture network in its
entirety when constructing the mesh. As a result, our approach opens new meshing possibilities that can overcome
certain challenges associated with complex fracture networks. In this work, we use the Mimetic Finite Dierence
(MFD) method (Brezzi et al., 2005), as it is locally conservative, it strongly relates to nite volumes methods, and it
works for a general set of polyhedral cells.
Fractures are coupled to the matrix as lower-dimensional boundary conditions enforced at internal faces of the
mesh. In this regard, it is similar to the approach used in lower-dimensional DFM. An illustration of this can be seen
in [Fig. 1(c)]. The internal boundary faces are generated using a cut-cell paradigm that takes advantage of polyhedral
cell capabilities of the MFD method. Similar ideas have been used in work by (Mallison et al., 2014). Such an approach
avoids the need to generate a tetrahedral or hexahedral mesh that directly conforms to the fracture network and rock
properties. Instead, a rectangular mesh is rst constructed, and then fracture boundaries are added by dividing the
cells at the fracture planes. As a result, most cells remain rectangular, and irregular cells are localized to the area
around the fractures.
The target application area is fracture networks resulting from seismic and microseismic data. Seismic analysis
plays a signicant role in fracture characterization (Sil and Srinivasan, 2009; Viruete et al., 2001). During the process
of hydraulically fracturing a well, very small seismic activities are generated around the fractures. These are often
referred to as microseismic events and are becoming increasingly important for understanding the properties of fracture
networks. The analysis of seismic data often produces a network of fractures as pixel-maps in two dimensions, or elliptic
shaped fracture planes in three dimensions (Srinivasan and Anupam, 2012). These fractures have varying orientations,
apertures, conductivities, as well as overall fracture density and locations. The techniques outlined in this work are
designed to fully represent such data sets as a DFM.
Earlier work in this area based on the Mimetic Finite Dierence (MFD) method and Multi-Point Flux Mixed Finite
Elements (MFMFE) was presented in (Al-Hinai et al., 2013). The MFD method has been used by other researchers
to model ow in fractures. The open source MATLAB Reservoir Simulation Toolbox (MRST) has been utilized for
solving problems with faults and fractures using the MFD method (Lie et al., 2012). In the work of (Formaggia et al.,
2014), the MFD method is compared to the XFEM method for fracture problems. Recently, the work of (Antonietti
et al., 2015) demonstrates coupling the MFD method and the Finite Volume method, with some theoretical results
concerning the approach. A similar approach to ours can be found in the recent work of (Yan et al., 2016), in which a
similar cut-cell approach is used for fracture embedding. Further theoretical results for cases with poroelasticity were
published in (Girault et al., 2014) and in (Girault et al., 2016) for the mixed nite element method. We remark that

2
the analysis for the MFE and MFD methods is similar. Also of note is the work found in (Berrone et al., 2016), in
which XFEM methods are used to tackle DFM. In addition, their work in (Berrone et al., 2013, 2015) uses a novel
constrained PDE optimization method to avoid challenges imposed by small cells.

2. Problem Formulation

We denote the matrix domain as m Rd (d = 2 or 3), as seen in [Fig. 2(a)]. The domain m has both external
boundaries (m ) and internal boundaries (f ), which correspond to the surface of the fracture. We distinguish the
top and bottom sides of the fracture boundary by +
f and f . The fracture domain is denoted by f R ,
d

as shown in [Fig. 2(b)], and has the same dimensionality as the matrix domain. We assume Darcy ow in both the
matrix and fracture. The fracture domain and matrix domain are coupled via boundary conditions and source terms.
The fracture domain imposes a pressure boundary on the reservoir, and the reservoir in turn imposes forcing terms on
the fracture. The system of equations for single-phase steady state ow are:

Km pm = qm in m
{+,}
p m = pf on f
(1)
Kf pf = qf ql in f
Kf pf n = 0 on f .

With ql representing the leakage term and is dened as,

ql = [Km pm ]f := Km pm |f + n+ + Km pm |f n .

Following the standard formulation of the MFD method found in (Brezzi et al., 2005), the fracture problem becomes:
Find vm Xm and pm Qm such that,

[vm , um ]Xm [pm , DIV um ]Qm = pf um dS, um Xm ,
{+,}
f

[DIV vm , wm ]Qm = [qm , wm ]Qm , wm Qm ,

and nd vf Xf and pf Qf such that,

[vf , uf ]Xf [pf , DIV uf ]Qf = 0, uf Xf ,

[DIV vf , wf ]Qf = [qf ql , wf ]Qf , wf Qf .

With ql dened as,

ql = [vm ]f := vm |+ + vm | .
f f

The space Q represents pressure, and corresponds to a single piece-wise constant pressure over every cell. The
space X represents ux, and corresponds to a single ux unknown for each face in the mesh. In this formulation, the
mesh cells can be made of a range of possible polyhedral cells (Brezzi et al., 2005). Pressure boundary conditions
are dened at the mesh faces. An interior boundary is created by duplicating an interior face and associating each
of the duplicates to an adjacent cell. In our approach, we do not rene the fracture mesh in the direction normal to
its surface, that is, we use a single cell whose thickness is equal to the fracture aperture. We have only stated the
formulation for the case of steady-state, single-phase incompressible ow. The method can be readily extended to time
dependent and multi-phase ow. Details about the two-phase ow can be found in the appendix.

3
 (a) Matrix (b) Fracture

Figure 2: Dene two separate domains m and f . On the left is the matrix domain with an interior boundary condition representing the
fracture. On the right is the fracture domain with no-ow boundary conditions.

(a) Hexahedral Cells (b) Polyhedral Cells

Figure 3: (a) Hexahedral cells typically produce a very small cell at the fracture intersection junction. (b) Using polyhedral cells, we can
represent intersecting fractures without creating a small cell at the intersection junction.

3. Resolving Intersecting Fractures

The advantages of an equi-dimensional representation of fractures are mainly present when considering fracture
intersections. This is because intersections extend the fracture network into the full dimensions of its space, both
geometrically and logically. In a lower-dimensional representation, intersections are handled by conning each fracture
to a separate plane, and then coupling the fractures at the intersection lines. On the other hand, equi-dimensional
fractures are meshed directly as a single connected domain.
It has been noted in the literature that equi-dimensional fracture meshes typically result in the creation of very
small cells at the intersection junction (Ahmed et al., 2015; Karimi-Fard et al., 2004). An illustration of this problem
can be seen for the two-dimensional case in [Fig. 3(a)]. However, this issue is the result of meshing fractures with
quadrilateral or triangular meshes. We suggest a means of avoiding the creation of small cells by using polyhedral
meshes. An example of using such polyhedra is shown in [Fig. 3(b)]. Note that we are fully capturing the geometry
of the two fractures without creating small cells at the junction. The algorithm for generating the mesh for a two-
dimensional case is illustrated in [Fig. 4]. The method preserves all the geometric characteristics of the fractures,
including aperture, fracture volume and the shape of the intersection junction. Examples of these cells can be seen in
[Fig. 5]. While the exact detail of the intersection junction may not have a large impact for linear Darcy type ows,
there is evidence suggesting there is an impact on more complex ow regimes (Kosakowski and Berkowitz, 1999; Park
et al., 2001).
Another advantage of equi-dimensional fracture representation is that it allows us to mesh the fracture network

4
 (2) We locate the inter-
(1) An initial set of inter-
section vertices as well as
secting fractures.
the tips of the fractures.

(3) At each vertex, we

generate two points on
both sides normal to the (4) The nal cells are
fracture. At intersec- constructed from con-
tions, two neighboring necting the appropriate
points are merged and points generated.
the normals are aver-
aged.

Figure 4: Algorithm for constructing intersecting fracture cells in the two-dimensional problem. In this current work, we focused on
fractures with a constant aperture. However, the approach can be readily generalized for cases in which aperture varies over the
length of the fracture.

Figure 5: By using general polyhedra and polygons, we can resolve dierent types of fracture intersections, including T-junctions and cases
with multiple fractures meeting at a point.

as a whole, rather than the sum of multiple fracture segments. An example of this point is illustrated in [Fig. 6(a)],
where we nd that for the three intersecting fractures shown, the lower-dimensional representation forces a small cell
at segment A. One way to avoid a cell for segment A for the lower-dimensional case would be by merging points.
However, merging fractures can be problematic, as it modies the fracture connectivity and may cascade into collapsing
portions of the fracture network. In addition, such merging is non-trivial especially in a three-dimensional setting. The
equi-dimensional representation can directly avoid the small segment at A without modifying the fracture network.
One way of accomplishing this is by taking advantage of non-convex polyhedral cells, which are allowed by the MFD
method. An example of such a mesh is illustrated in [Fig. 6(b)]. In the example, we are meshing the same three
intersecting fractures, but can directly avoid the creation of a small cell at segment A.
The same principle can be applied to three-dimensional intersecting fractures. For two fully-penetrating fractures,
the two-dimensional case can be simply extruded. However, fractures are rarely fully penetrating. A more common
occurrence is the case seen in [Fig. 7]. The rst step is to divide the two fractures along the line of intersection, giving
us four separate segments. The second step involves extruding the segments into three-dimensional cells giving them
depth and allowing them to connect appropriately. This is accomplished by generating a sloping, beveled edge at the
location of the intersection, and a at surface at the connection between the two cells. The resulting geometry is
illustrated in [Fig. 8].
After generating the fracture cells, one is free to create a separate mesh inside the fracture, perhaps using conven-
tional meshing methods. However, in our work we avoid meshing inside fractures. That is, we only divide the fracture
when necessary to accommodate an intersection. This can be reasonable in cases with a large number of fractures, as
any further renement would render the model intractable. However one can still ask in general, how to further rene

5
 (a) Lower-dimensional representation (b) Equi-dimensional representation

Figure 6: We compare an important dierence between the lower-dimensional and the equi-dimensional representation for three intersecting
fractures as show in (a). The lower-dimensional mesh does not allow many options for representing the conguration without
creating a very small cell for segment A. On other hand, the equi-dimensional representation allows for a mesh shown in (b). In
this case, we took advantage of non-convex cells to connect the fractures without creating a small cell at segment A.

Figure 7: Two partially penetrating ellipse shaped fractures in three dimensions. The rst step in generating the fracture mesh is to divide
each fracture along the line of intersection. The red line represents the location of overlap between the two fractures.

the mesh. One approach we have relied on is to intersect the fracture mesh with the matrix mesh, and force nodes at
the intersection points. This naturally renes the fracture mesh as the matrix is rened.

4. Embedding Fracture Boundaries

In order to couple the fractures with the surrounding matrix, we need to incorporate the internal boundaries in
the matrix mesh. That is, the internal boundaries (+
f , f ) must be located at mesh faces. If the fracture happened
to be located on an existing face, no further mesh manipulation is needed, as internal boundaries can be assigned
directly to these faces. However, if the fracture surface is located at any other location, new faces must be generated.
In order to create these faces, we divide the cell along the plane of the fracture. Two new cells are formed, and the
new face between them becomes the internal boundary condition for the fracture. A two-dimensional illustration can
be seen in [Fig. 9]. The red line represents the location of the fracture. One segment of the fracture is already located
at the boundary of two adjacent cells, meaning we do not need to change anything about the mesh for that segment.
However, the second segment is located diagonally through an existing cell and requires some work. In this case, we
divide the cell into two triangular cells, and the boundary between them becomes the internal boundary surface for
the fracture. Note that with multiple cell divisions, arbitrary polygons and polyhedra can be formed. It is here that
we rely again on the MFD methods ability to solve using general polyhedra. This is often referred to as a cut-cell
paradigm which has been explored previously in the literature for modeling complex reservoir features (Mallison et al.,
2014).

6
 (a) (b)

Figure 8: The geometry of the mesh cells for the fractures shown in [Fig. 7]. Figure (a) shows the nal form of the intersecting fractures.
Here, all four cells join together to fully resolve the intersection. Note that the nal form accurately represents fracture volumes
as well as the interaction region between them. Figure (b) shows a close up of the two polyhedra for cell 1 and cell 2.

Figure 9: The plot on the left represents the initial matrix mesh (in solid black lines) and the fracture we wish to incorporate (in dashed
red lines). The fracture interior boundary must occur at a mesh face. This requirement is satised for the vertical portion of
the fracture. However, the diagonal portion (in the top left cell) requires some work. Since the MFD method allows for general
cells, we can create a face that corresponds to the fracture by dividing the top left square into two triangles as is shown in the
illustration on the right.

The same principle can be used in three dimensions. Given an ellipse shaped fracture, we can divide elements along
the plane of the ellipse as in [Fig. 10]. Notice that the fracture surface will not always fully penetrate the cell. In this
case, cells that intersect fractures even partially are fully divided by the plane of the fracture. We can then subdivide
the resulting face into a region that corresponds to the fracture boundary, and the rest that becomes a regular internal
face in the matrix domain. By doing so, we can correctly account for the surface area contact between the fracture
and the matrix. The approach of dividing cells at the fracture boundaries naturally generalizes for multiple fractures.
The algorithm works incrementally, by dividing the cells for the rst fracture, and then dividing the resulting cells by
the next one and so on.
We tested embedding a larger number of fractures. We start with a (212121) square mesh, giving a total of 9261
cells and 29106 faces. After embedding 35 fractures, the total number of cells increased by 5% to 9727, and the number
of faces increased by 7% to 31205. There are many advantages to this approach. Firstly, we avoid time consuming
conventional mesh generation. Secondly, each cell in the nal mesh can be associated with an original rectangular cell,
allowing for a direct correspondence in cell properties to the original grid. This means we can more easily incorporate
fractures into an existing geological grid. Thirdly, most cells remain square, localizing the unstructured polyhedral
cells to the area around the fractures.

7
Figure 10: Dividing cells in order to create internal fracture boundaries extends to the three-dimensional case. On the left we have an
illustration of an ellipse shaped fracture dividing a square domain across its plane. Cells that are only partially penetrated by
the fracture are still divided all the way across the plane of the fracture. The resulting surface can be seen in the gure on
the right. Note the faces on the fracture plane have been subdivided based on the shape of the fracture. Faces on the fracture
become internal boundary conditions, while the rest are interior mesh faces.

5. Numerical Results

We present numerical experiments demonstrating and validating the techniques in this work. In the rst ex-
ample, we use our approach to calculate shape factors for fracture networks and compare them to standard shape
factors derived for dual-porosity models. The second example compares the methods we propose with a standard
equi-dimensional representation of intersecting fractures. In the third example, we demonstrate a solution to three-
dimensional intersecting ellipsoid fractures with a two-phase ow model. In the nal example, we apply our method
to a fracture network derived from a microseismic study of a real reservoir.

5.1. Comparison with Dual-Porosity Methods

The dual-porosity approach assumes two separate volumes in the domain, a fracture volume and a matrix volume.
Darcy ow is assumed in the matrix and the fracture. Exchanges of uids between the two are governed by,

q = Km (pm pf ). (2)

Here, , known as the shape factor, has the dimension of reciprocal area and depends on the geometry of the fractures
as well as the physics of the ow. We use our approach to numerically estimate the shape factor for the single-phase
case and compare the resulting number to estimates in the literature. The four estimates are by (Warren and Root,
1963), (Kazemi et al., 1976), (Coats, 1989) and (Lim and Aziz, 1995). The variable N indicates the number of fracture
sets, which can be one set in one direction (N = 1), two sets in orthogonal directions (N = 2) or three sets in
orthogonal directions (N = 3) with spacing denoted by L.
The rst comparison (Case 1) uses the setup in [Fig. 11(a)], which corresponds to N=1 and Ly = 0.25m. For
the second comparison (Case 2), we use the setup in [Fig. 11(b)], which corresponds to N=2 and Lx = Ly = 0.25m.
The outside boundaries are set to no-ow conditions except for the left most fracture tips, which are set to a pressure

8
 (a) (b)

Figure 11: Problem setup for the dual-porosity comparison cases. We use the square A as the representative block and calculate average
pressure and total exchange of uids over A. (a) Setup for Case 1 with N=1 and Ly = 0.25m. (b) Setup for Case 2 with N=2
and Lx = Ly = 0.25m.

specied boundary of p = 500Pa. System properties are:

reservoir thickness = 1m
Kf = 1 108 m2
Km = 1 1014 m2
fracture width = 0.01m
porosity = 0.3
p(t = 0) = 1000Pa
total compressibility = 1 108 Pa1

We can numerically estimate the shape factors by noting that,

q
= , (3)
Km (pm pf )
1

where pm = |A| p is the average pressure over the representative matrix block A. Square A is selected at the
A m
center of the domain to avoid the boundary eects. The leakage term q is calculated directly from the MFD method
over the boundary of the fracture inside the representative volume A. We plot the non-dimensional form of the shape
factors (L2 ) in [Fig. 12]. After an initial phase, the system results in a constant shape factor. Notice that most shape
factors in the literature ignore this initial transient phase. In both cases, our method produces a shape factor that is
consistently very close to the one estimated in (Lim and Aziz, 1995). For Case 1, our estimate settled on (L2 = 9.75)
and the Lim-Aziz shape factor is (L2 = 2 2 9.87). For Case 2, our estimate settled on (L2 = 19.53) and the
Lim-Aziz shape factor is (L2 = 2 19.74).

5.2. A Two-Dimensional Comparison

In this example we solve a two-dimensional problem with two intersecting fractures. The problem setup can be
seen in [Fig. 13]. Two dierent methods are used to solve the problem. The rst approach (Method 1) resolves the
fractures by creating small cells in a structured rectangular mesh, see [Fig. 14]. The second approach (Method 2)

9
 (a) (b)

Figure 12: A comparison of the dimensionless form of the shape factor (L2 ) computed using the MFD method with values found in the
literature. At early time, the calculated value of the shape factor decreases until the system results in a constant shape factor.
(a) The calculated value of L2 vs time for Case 1, with N = 1 and Ly = 0.25m. (b) The calculated value of L2 vs time for
Case 2, with N = 2 and Lx = Ly = 0.25m. Note that in both cases our approach most resembles the shape factor calculations
from (Lim and Aziz, 1995).

uses the techniques suggested in this work. Both use standard IMPES for two-phase ow and the MFD method for
resolving the pressure problem.
The meshes used can be seen in [Fig. 14]. Notice how the rst method creates a very small cell at the location of
the fracture intersection. The cell at the intersection has a volume of (0.01m3 ), compared to the rest of the fracture
cells, which are (0.4m3 ), a factor of 40. In the method we suggest, no such small cell is created. Since the CFL
condition is inversely proportional to the smallest cell volume, this means that we can take a time step on the order
of 40 times larger with our method. The time step restriction for Method 1 is the reason we set the fracture width
to an unrealistic value (0.1m). A smaller fracture width would have taken far too long to compute using that method.
The overall rock properties for the problem are,

reservoir thickness = 1m
Kf = 1m2
Km = 1 106 m2
fracture width = 0.1m
qm = 104 m3 /s
porosity = 0.3

The uid properties are as follows:

initial water saturation = 0 sro = 0.2

w = 8.9 104 Pa s srw = 0
o = 8.9 104 Pa s sw srw
3 se =
water reference density (pref = 0) = 1000kg/m 1 srw sro
oil reference density (pref = 0) = 1000kg/m
3 krw = se2
water compressibility = 1 108 Pa1 kro = (1 se)2
oil compressibility = 1 108 Pa1 pc = 0

10
Figure 13: Problem setup for the two-phase comparison results. There are two intersecting fractures in the domain and a rate specied
injector well in the lower right corner. The boundaries are no-ow on all sides except the top boundary that is set to p = 0.
Fractures are set to a width of 0.1m.

The comparison of saturation results at dierent times can be seen in [Fig. 15]. The two approaches produce almost
identical solutions. This conrms that our approach gives us the correct answer. However as mentioned, our method
allows for a much larger time step compared with standard equi-dimensional models.

5.3. Three-Dimensional Case

We present an example of two intersecting elliptic shaped fractures. The problem setup can be seen in [Fig. 16].
Such shapes are typical in microseismic analysis of fracture networks (Srinivasan and Anupam, 2012). This example
illustrates how such fractures can be directly modeled with a small number of elements and without resorting to
conventional mesh generation. The domain m = [0, 1]3 , with two pressure boundary conditions, (p = 10Pa) at
(x = 0) and (p = 0Pa) at (x = 1) and no-ow conditions on all other boundaries. The domain is initially meshed with
a square grid that is [11 11 11]. The fractures are meshed using the same method shown earlier in [Fig. 8]. Both

(a) Method 1 (b) Method 2

Figure 14: (a) The mesh used for Method 1, the fracture is directly included in the rectangular mesh. Notice the very small cell created
at the fracture intersection. (b) The meshes used for Method 2, our suggested approach. The dots in the mesh are the location
of the internal pressure boundary. The mesh on the right is used for the fracture solution. In this illustration, the fracture
thickness has been altered for better visibility.

11
Figure 15: Saturation at t=5000s. The plot on the left corresponds to the solution from Method 1, and the plot on the right is from
Method 2. Both methods produced a similar result. The dierence is that our approach (Method 2) accepts a much larger
time step due to avoiding small cells created by the more conventional equi-dimensional technique (Method 1).

fractures have a width of (0.002m). Rock properties are,

Kf = 1m2
Km = 1 106 m2
fracture width = 0.002m
porosity = 0.3

We use the same uid properties from the previous example. The saturation solution at dierent times can be
seen in [Fig. 17]. Note that many of the same ideas we used for the two-dimensional case apply directly to the three-
dimensional example. In fact, we use the same code for all these cases. This means that validation results obtained
from simpler cases extend directly to the more complex fracture problems. Notice that in three-dimensions, the contact
between the fractures and the rock is far more complex than that in two-dimensions. However, by using polyhedral
cells in our meshes, we are able to exactly reproduce the contact region between them.

5.4. A Microseismic Derived Fracture Dataset

In our nal example, we demonstrate our methods for modeling a fracture network produced from microseismic
eld data. We use the data set found in (Fisher et al., 2002). The fracture mesh and a matrix mesh are constructed
as shown in [Fig. 18]. In order to test the mesh, we solved a slightly compressible single-phase ow problem with the
following parameters:

reservoir thickness = 1m w = 0.001m

production rate = 0.26m /day3
cf = 1 108 Pa1
initial pressure = 1 108 Pa = 8.9 104 Pa s
3
Km = 1 1017 m2 ref = 1000kg/m (pref = 0)
Kf = 1 109 m2 = 0.3

The domain is 730m900m, with a point source producer placed at (343.1m, 477.0m, 0.5m). The resulting pressure
solution in the matrix is shown in [Fig. 19]. Note that we are able to directly apply our techniques to eld scale model.
This is because the meshes we generate can be scaled directly to t any eld size needed with no dependence on
fracture width.

12
Figure 16: Problem setup for two elliptic shaped fractures intersecting in three dimensions. We impose a pressure gradient across the x-axis
and no-ow boundaries at all other faces.

Figure 17: Saturation distribution for two intersecting fractures in three dimensions. The saturations are recorded at times t = [0.4, 1.0,
1.8]s.

13
Figure 18: The matrix mesh (left) generated for the microseismic eld data found in (Fisher et al., 2002). The mesh was generated by rst
constructing a square grid, and then using the cut-cell approach to create internal fracture faces. On the right we nd the mesh
generated for the fracture ow using polyhedral cells.

Figure 19: Pressure solution in matrix at times = [0.059, 0.462, 1.516] days for the microseismic eld data example.

14
6. Conclusions and Future Work

In this work, we have presented a modied version of equi-dimensional fracture modeling using the Mimetic Fi-
nite Dierence method. We take advantage of the MFD methods ability to solve over general polyhedral cells.
This property is used for meshing intersecting fractures without the need for creating very small cells. In our new
equi-dimensional representation, fracture networks are represented as a single connected domain, allowing many geo-
metrically consistent options when meshing them. Using polyhedral cells, we can mesh complex fractures with a small
number of cells. This is very important especially when considering systems of hundreds of thousands of fractures.
We used the polyhedral cells a second time when embedding fracture boundaries into rectangular and hexahedral
domains. We noted how employing a cut-cell paradigm allows us to avoid conventional mesh generation around the
fractures, an often time-consuming and challenging task. We have also demonstrated how this approach keeps most
elements square, and localizes the unstructured nature of the mesh to the area adjacent to the fractures.
While we have presented some three-dimensional examples, there are many more possible fracture networks to
mesh. Our work presents some novel principles of how to accomplish this. However, more work is needed to develop
robust algorithms for tackling the geometry of three-dimensional fractures.
The MFD method is not a single discretization, but rather a family of methods. Within that family exists standard
Finite Volume methods such as the two-point ux approximation (Al-Hinai et al., 2015; Droniou et al., 2010) as well as
multi-point ux methods (Lipnikov et al., 2009). Using the relationship between the MFD method and standard nite
volume methods, we can establish a connection between the methods outlined here and some of the other approaches
in the literature.

7. Nomenclature

X = Discrete velocity space

Q = Discrete pressure space
E = Cell E in mesh
|E| = Volume of cell E
e = Face e in mesh
|e| = Area of face e
n = Unit normal vector
= External normal direction (-1, 1)
= Density
= Porosity
c = Compressibility
q = Source or sink term (producer or injector)
K = Permeability tensor
kr = Relative permeability of phase
p = Pressure
pc = Capillary pressure
s = Saturation
= Fluid viscosity
t = Time
w = Fracture width

8. Subscripts

= Phase (water or oil)

E = Cell E in the mesh
ref = Reference

15
9. Superscripts

e = Face e in the mesh

f = Fracture variable
m = Matrix variable
T = Transpose
1 = Inverse

10. Acknowledgments

The code used to generate the results was written in Python and used the SciPy and NumPy libraries (Jones
et al., 2001). The Python package Shapely was used for two-dimensional polygon calculations. Visualization was done
using ParaView (Henderson, 2007). For cell volume and centroid computations, the code uses the algorithm dened
in (Mirtich, 1996). The authors would like to acknowledge help from Dr. Ivan Yotov, Dr. Mark Mear, and Dr. Xin
Yang. A portion of this research was supported by the King Abdullah University of Science and Technology Academic
Excellence Alliance and Saudi Aramco.

References

Ahmed, R., Edwards, M. G., Lamine, S., Huisman, B. A. H., Pal, M., 2015. Control-volume distributed multi-point ux
approximation coupled with a lower-dimensional fracture model. Journal of Computational Physics 284, 462489.

Al-Hinai, O., Singh, G., Almani, T., Pencheva, G., Wheeler, M., 2013. Modeling multiphase ow with nonplanar
fractures. In: SPE Reservoir Simulation Symposium. Society of Petroleum Engineers.
Al-Hinai, O. A., Wheeler, M. F., Yotov, I., 2015. A generalized Mimetic Finite Dierence Method and Two-Point Flux
schemes over Voronoi diagrams. ESAIM: Mathematical Modelling and Numerical Analysis.

Alboin, C., Jare, J., Roberts, J., Serres, C., 2002. Modeling fractures as interfaces for ow and transport in porous
media. In: Fluid Flow and Transport in Porous Media, Mathematical and Numerical Treatment: (South Hadley,
MA, 2001). Vol. 295 of Contemp. Math. Amer Mathematical Society, pp. 1324.

Antonietti, P. F., Formaggia, L., Scotti, A., Verani, M., Verzotti, N., 2015. Mimetic nite dierence approximation of
ows in fractured porous media. ESAIM: Mathematical Modelling and Numerical Analysis.

Arbogast, T., Douglas, Jr, J., Hornung, U., 1990. Derivation of the double porosity model of single phase ow via
homogenization theory. SIAM Journal on Mathematical Analysis 21 (4), 823836.

Berrone, S., Pieraccini, S., Scialo, S., 2013. A PDE-constrained optimization formulation for discrete fracture network
ows. SIAM Journal on Scientic Computing 35 (2), B487B510.
Berrone, S., Pieraccini, S., Scialo, S., 2016. Towards eective ow simulations in realistic discrete fracture networks.
Journal of Computational Physics 310, 181201.

Berrone, S., Pieraccini, S., Scialo, S., Vicini, F., 2015. A parallel solver for large scale DFN ow simulations. SIAM
Journal on Scientic Computing 37 (3), C285C306.

Brezzi, F., Lipnikov, K., Shashkov, M., 2005. Convergence of the mimetic nite dierence method for diusion problems
on polyhedral meshes. SIAM Journal on Numerical Analysis 43 (5), 18721896.

Coats, K. H., 1989. Implicit compositional simulation of single-porosity and dual-porosity reservoirs. In: SPE Sympo-
sium on Reservoir Simulation. Society of Petroleum Engineers.

16
Droniou, J., Eymard, R., Gallouet, T., Herbin, R., 2010. A unied approach to mimetic nite dierence, hybrid nite
volume and mixed nite volume methods. Mathematical Models and Methods in Applied Sciences 20 (02), 265295.

Fisher, M., Wright, C., Davidson, B., Goodwin, A., Fielder, E., Buckler, W., Steinsberger, N., 2002. Integrating
fracture mapping technologies to optimize stimulations in the barnett shale. In: SPE Annual Technical Conference
and Exhibition. Society of Petroleum Engineers.

Formaggia, L., Antonietti, P., Panli, P., Scotti, A., Turconi, L., Verani, M., Cominelli, A., 2014. Optimal techniques to
simulate ow in fractured reservoir. In: ECMOR XIV-14th European conference on the mathematics of oil recovery.

Girault, V., Wheeler, M., Ganis, B., Mear, M., 2014. A lubrication fracture model in a poro-elastic medium. Mathe-
matical Models and Methods in Applied Sciences.

Girault, V., Wheeler, M. F., Kumar, K., Singh, G., February 2016. Mixed formulation of a linearized lubrication
fracture model in a poro-elastic medium. ICES REPORT 16-06, The Institute for Computational Engineering and
Science, The University of Texas at Austin.

Hgland, H., Assteerawatt, A., Dahle, H., Eigestad, G., Helmig, R., 2009. Comparison of cell-and vertex-centered
discretization methods for ow in a two-dimensional discrete-fracturematrix system. Advances in water resources
32 (12), 17401755.

Henderson, A., 2007. The ParaView guide: A parallel visualization application. , Kitware Inc.

Hoteit, H., Firoozabadi, A., 2008. An ecient numerical model for incompressible two-phase ow in fractured media.
Advances in Water Resources 31 (6), 891905.

Jones, E., Oliphant, T., Peterson, P., 2001. SciPy: Open source scientic tools for Python.
URL http://www.scipy.org/

Karimi-Fard, M., Durlofsky, L., Aziz, K., 2004. An ecient discrete-fracture model applicable for general-purpose
reservoir simulators. SPE Journal 9 (02), 227236.
Kazemi, H., Merrill Jr, L. S., Portereld, K. L., Zeman, P. R., 1976. Numerical simulation of water-oil ow in naturally
fractured reservoirs. Society of Petroleum Engineers Journal 16 (06), 317326.

Kosakowski, G., Berkowitz, B., 1999. Flow pattern variability in natural fracture intersections. Geophysical research
letters 26 (12), 17651768.
Lie, K., Krogstad, S., Ligaarden, I., Natvig, J., Nilsen, H., Skaestad, B., 2012. Open-source matlab implementation
of consistent discretisations on complex grids. Computational Geosciences 16 (2), 297322.

Lim, K., Aziz, K., 1995. Matrix-fracture transfer shape factors for dual-porosity simulators. Journal of Petroleum
Science and Engineering 13 (34), 169 178.
Lipnikov, K., Shashkov, M., Yotov, I., 2009. Local ux mimetic nite dierence methods. Numerische Mathematik
112, 115152.

Mallison, B., Sword, C., Viard, T., Milliken, W., Cheng, A., et al., 2014. Unstructured cut-cell grids for modeling
complex reservoirs. SPE Journal 19 (02), 340352.

Mirtich, B., 1996. Fast and accurate computation of polyhedral mass properties. Journal of Graphics Tools 1 (2),
3150.

Ochs, S., Hinkelmann, R., Neunhauserer, N., Suess, M., Helmig, R., Gebauer, S., Kornhuber, R., 2002. Adaptive
methods for the equidimensional modelling of ow and transport processes in fractured aquifers. In: 5th International
Conference on Hydro-Science &-Engineering, Warsaw, Poland.

17
Park, Y., Lee, K., Berkowitz, B., 2001. Eects of junction transfer characteristics on transport in fracture networks.
Water Resources Research 37 (4), 909923.
Sil, S., Srinivasan, S., 2009. Stochastic simulation of fracture strikes using seismic anisotropy induced velocity anomalies.
Exploration Geophysics 40 (3), 257264.
Srinivasan, S., Anupam, A., 2012. Multiscale modeling of fracture network in a carbonate reservoir. In: Abrahamsen,
P., Hauge, R., Kolbjrnsen, O. (Eds.), Geostatistics Oslo 2012. Vol. 17 of Quantitative Geology and Geostatistics.
Springer Netherlands, pp. 185196.
Viruete, J., Carbonell, R., Jurado, M., Mart, D., Prez-Estan, A., 2001. Two-dimensional geostatistical modeling and
prediction of the fracture system in the Albala Granitic Pluton, SW Iberian Massif, Spain. Journal of Structural
Geology 23 (12), 2011 2023.
Warren, J., Root, P., 1963. The behavior of naturally fractured reservoirs. Old SPE Journal 3 (3), 245255.
Yan, X., Huang, Z., Yao, J., Li, Y., Fan, D., 2016. An ecient embedded discrete fracture model based on mimetic
nite dierence method. Journal of Petroleum Science and Engineering 145, 1121.

Appendix A Two-Phase Flow

In this formulation, we neglect the eect of capillary pressure and gravity, but allow for slightly compressible uids
and no compressibility for the rocks. Starting with the two-phase equations,

v = kr Kp ,

s
= v + q ,
t
pc = po pw = 0,
sw + so = 1, (4)

with the mobility of phase dened as

kr
= . (5)

By adding the saturation equations we get,

vt = (w w + o o )Kpo , (6)
(w sw + o (1 sw ))
= vt + qt . (7)
t
The variable vt denotes he total velocity, and qt = qo + qw is the total source term.
In order to use the MFD method, we replace the exact gradient and divergence operators in the equations with
discrete approximations, as follows:

G (w w + o o )K,
DIV . (8)

The operators DIV and G share the adjoint relation,

[Gph , uh ]X = [ph , DIV uh ]Q (9)

The weak form of the problem is stated as:

18
 Find sw , po Q and vt X, such that

[vt , uh ]X [po , DIV uh ]Q = 0, uh X (10)

(w sw + o (1 sw ))
[ , wh ]Q + [DIV vt , wh ]Q = [Qt , wh ]Q , wh Q. (11)
t
We denote (10) as F1 (po , vt ) = 0. Since we will be using backward Euler, we set the variables to the next time step,
giving,

F1 (pn+1
o , vtn+1 ) = 0. (12)

Assuming slight compressibility of the uid, we have

(p) = 0 (1 + c p), (13)

using the backward Euler approximation for the time derivative, we get for (11),
 
 
w (po )sw + o (po )(1 sw ) , wh
n+1 n+1 n+1 n+1
t Q
 
(14)
(w (pnw )snw + o (pno )(1 snw )) , wh
t Q

+ DIV vtn+1 , wh Q [qtn+1 , wh ]Q = 0,

which we will denote as

F2 (pn+1
o , vtn+1 ) = 0. (15)

We use the Newton-Raphson method for solving the non-linear system of equations,



F1 (pn+1
o , vtn+1 ) 0
= (16)
F2 (pn+1
o , vtn+1 ) 0

The resulting Jacobian matrix is

 F1 F1


vtn+1 pn+1
o
M DIV
= , (17)
F2 F2
DIV C
vtn+1 pn+1
o

with M , DIV and DIV constructed as in usual MFD fashion. For simplicity, we can lag the coecient (w w +o o )
to time step n when building M . The matrix C is diagonal, and is dened for each cell E as,

|E|  0 
C(E) = w cw sn+1
w + 0o co (1 sn+1
w ) . (18)
t

Next, we look at the saturation update. The water velocity is computed using a fractional ow function calculated
from the upwinded mobility,

vw = fw vt . (19)

Applying the weak form directly to the saturation equation gives us,
 
w sw
, wh = [DIV vw
n+1
, wh ]Q + [Qn+1
w , wh ] Q . (20)
t Q

19
Using backward Euler, we have
 
 
w (pn+1
w )s n+1
w (p n n
w w w)s , w h
n+1
= [DIV vw , wh ]Q + [Qn+1
w , wh ]Q . (21)
t Q

We can now apply the standard IMPES method by rst solving the non-linear pressure equation, updating the
saturation explicitly and so on.

20