You are on page 1of 14

^

Fractured Continuum Approach to Stochastic


Permeability Modeling
S. A. McKenna
P. C. Reeves
Sandia National Laboratories
Albuquerque, New Mexico, U.S.A.

ABSTRACT
A new kind of fracture-permeability model called the fractured continuum
model (FCM) is presented that incorporates the advantages of discrete fracture
network models and the processing speed of effective continuum representations
of fracture permeability. Observations of fracture orientation, length, frequency,
and transmissivity from boreholes and outcrops are used as input to the FCM
approach. Geostatistical realizations of fracture connectivity, represented by the
coordination number of a local percolation network, and fracture frequency are
combined with object-based simulations of high- and low-permeability classes in
the model domain. At each location, these three spatially variable properties are
combined into an effective grid-block permeability using an approximation based
on the effective medium theory. The resulting realizations of fracture permeabili-
ty, containing greater than 106 elements, are used as input to a single-phase flow
model. A parallel computer platform coupled with a unique groundwater flow
code is used to efficiently solve steady-state pressure fields on multiple reali-
zations. The ability to solve many realizations in a short amount of time allows for
the evaluation of the effects of two different conceptual models of fracture per-
meability on particle traveltimes.

INTRODUCTION fracture networks for fluid-flow problems have been


developed. One of these techniques is the stochastic
The function of fractures in subsurface fluid-flow continuum model (SCM). Stochastic continuum mod-
problems has long been recognized and has become els have been proposed as a means of representing the
an especially important issue throughout the past two spatially heterogeneous nature of fracture permeabil-
decades in the areas of petroleum-reservoir engineer- ity at the scale of continuum grid blocks (see Hsieh,
ing and long-term storage of nuclear waste. Several 1998). In general, the discrete behavior of individual
different techniques for modeling the permeability of fractures is represented by an average property over a

Copyright n2006 by The American Association of Petroleum Geologists.


DOI:10.1306/1063815CA53234

173
174 McKenna and Reeves

representative elementary volume (REV). The vol- resentation of the fracture network. The homoge-
ume necessary to define a REV within a fracture sys- nization employed in SCM approaches may not be
tem can be quite large. The SCM approach has been able to reproduce potential extremes of ground-
used previously to model fracture systems at ex- water transport behavior. An approach to fracture-
perimental facilities for nuclear waste programs in permeability modeling that can capitalize on the
the United States (see Altman et al., 1996; Finsterle, strengths of both DFN and SCM approaches and
2000) and in Sweden (Follin and Thunvik, 1994; Tsang can directly use information measured on discrete
et al., 1996). In these studies, geostatistical-simulation fractures is presented herein.
algorithms were used to model continuum properties This chapter describes a new approach to sto-
of discrete fracture systems at the scale of numerical chastic modeling of fracture permeability the frac-
model grid blocks (ones to tens of meters). The geo- tured continuum model (FCM). The approach is
statistical realizations of these properties were then based on a two-step process: critical properties of the
used as input to flow and transport models to pre- DFN are considered to be spatially random variables
dict the advective traveltime of particles through and are modeled with a combination of geostatis-
the heterogeneous systems. tical and object-based simulation. These simulations
Another technique for modeling fluid flow in define the number of fractures, the connectivity of
groundwater and petroleum systems is the discrete the fractures, and the relative permeability class of
fracture network (DFN) model, in which each indi- the fractures in each grid block of the model domain.
vidual fracture is modeled as a discrete one- or two- The effective permeability for each grid block is de-
dimensional (1-D or 2-D) feature within a 2-D or rived from the fracture properties using a local ef-
three-dimensional (3-D) domain, respectively. Initial fective medium approximation. The simulations also
work with DFNs focused on techniques for deriv- define the larger scale connectivity of the perme-
ing the equivalent permeability of areas or volumes ability values throughout the model domain through
of fractured rock (e.g., Long et al., 1982; Long and the correlation of properties across adjacent grid
Witherspoon, 1985). Recent developments in DFN blocks. Such a modeling approach is made feasible
models allow for the simulation of large, complex by the use of a fine-scale grid discretization that
discrete fracture systems in 2-D and 3-D domains necessitates a parallel-flow solver. An example prob-
(e.g., Dershowitz et al., 1998). lem demonstrating this modeling approach as ap-
Both the DFN and SCM approaches to modeling plied to the prediction of groundwater flow away
fracture permeability have strengths and weakness- from a proposed nuclear waste repository is pre-
es. The relative merits of the different modeling sented. In this example, single-phase fluid flow and
approaches are discussed by Dershowitz et al. (1998). streamline particle tracking are calculated directly on
The DFN models can explicitly model the fractures the fine-scale permeability models using a recently
as 2-D objects in a 3-D domain. These models can be developed parallel groundwater flow simulator.
conditioned to fracture observations and generally
reproduce the distributions of observed fracture at-
tributes. However, stochastic realizations of a DFN EFFECTIVE MEDIUM APPROXIMATION
model require computationally expensive regrid-
ding of the finite-element mesh for each realization, The first step in the FCM approach is to deter-
and the fracture network must commonly be ab- mine the numerical model grid-block-scale fracture
stracted to a series of 1-D pipes for solute transport permeability given the number of fractures, a mea-
modeling in groundwater problems. The compu- sure of fracture connectivity and the permeability of
tationally efficient regular grid employed in SCM each individual fracture contained within any grid
models allows for multiple realizations and complex block.
solute transport processes to be modeled. Because of A constitutive approximation based on effective
the necessary upscaling to the REV-size grid blocks, medium theory (EMT) is employed to derive the
explicit conditioning of SCMs to observed fracture effective permeability of each grid block in the
data can be problematic. In addition, abstractions of model domain. This approximation is derived from
discrete fracture patterns to a mean, variance, and work done in percolation theory (Kirkpatrick, 1973)
spatial covariance of permeability, as commonly done and has been previously used to model effective
in the SCM approach, overly homogenize the rep- permeability in fractured rock (David et al., 1990;
Fractured Continuum Approach to Stochastic Permeability Modeling 175

Zimmerman and Bodvarsson, 1996). A brief over- shown in Figure 1, where the fractures are the result
view of this approximation, following the approach of multiple tectonic events producing several orien-
of Zimmerman and Bodvarsson (1996), is presented tations of fractures.
below. The effective medium approach begins with main-
Consider a 3-D, cubic, numerical model grid block taining the irregular geometry of the fracture net-
with all three dimensions equal to a length, L, which work but replacing the individual fracture conduc-
contains a single set of parallel, evenly spaced frac- tances with a single effective conductance, C* (upper
tures spanning the block in one direction. Along each right image, Figure 1). The geometry of the fracture
fracture, a conductance, or transmissivity, Ci[L2/T], is system is held constant, with the dead-end fractures
defined here by invoking the assumption of flow removed, and the macroscopic conductivity of the
through parallel plates (Snow, 1969): fracture system remains the same for both the orig-
inal conductances and the single effective conduc-
gb3 r
Ci 1 tance. C* can be found by solving the implicit equa-
12m
tion (Kirkpatrick, 1973)
where g[L/T2] is the gravitational acceleration, m[M/
X
N
C  Ci
LT] is the viscosity, r[M/L3] is the density of water, 0 5
and b[L] is the fracture aperture. The length and i1
zi =2  1C Ci
width of the fractures are equal and span the grid
block in both directions. where the coordination number, zi, is the number of
By placing a pressure difference,  D H[L], across conductive fracture branches emanating from each
the grid block parallel to the fracture set, the vol- fracture intersection.
umetric flux, Q[L3/T], through the block can be
defined as

Q Nf C H 2

where Nf is the number of parallel vertical fractures


and C* is an effective conductance value calculated
as an average of the Nf values of Ci.
Volumetric flux through the cubic block can also
be calculated by applying Darcys law. The volu-
metric flux through the face of the grid block or-
thogonal to the fracture orientation is

Q Kb LH 3

where Kb[L/T] is the continuum hydraulic conduc-


tivity for the grid block in the direction parallel to
the fracture set. By setting equations 2 and 3 equal to
one another, it is possible to solve for Kb of the cubic
block using only geometric information on the frac-
ture network, the dimensions of the numerical grid
block, and the effective conductance of the fracture
segments:
Nf C
Kb 4
L Figure 1. Conceptual example of determining the av-
erage fracture conductance for an irregular discrete frac-
For natural fracture systems, multiple fracture sets ture network in a numerical model grid block. Fracture
may be intersecting and variably connected. There- junctions are the intersection of any two fractures, and
fracture segments are the straight-line fracture traces
fore, derivation of C* requires a sophisticated av- connecting the junctions. See text for details. The gray
eraging technique. Consider a 2-D fracture map as scale of the fractures indicates the relative permeability.
176 McKenna and Reeves

As pointed out by Zimmerman and Bodvarsson conductance segments in the square lattice is 1 
(1996), equation 5 was originally developed for net- (z*/4), and the fraction of segments with C* is z*/4. A
works where the coordination number does not vary second application of equation 5 to the square lat-
between intersections. However, equation 5 can also tice leads to the relation
be used with an average coordination number for
 
topologically irregular networks. For the example z
C  1 C 6
shown in Figure 1, eight intersections are present, 2
and the average coordination number, z*, is 3.0.
After the effective conductance value, C*, is de- This relationship between the effective conduc-
termined through equation 5 and defined for each tance calculated for the irregular network and the
fracture segment, the irregular discrete fracture map effective conductance of the resulting square network
is transformed to a square network where the mean demonstrates that at average coordination numbers
fracture spacing and mean coordination number are less than or equal to 2.0, the resulting effective con-
maintained (Figure 1, middle image). The square ductivity of the grid block is zero. At this point, the
network requires that z* = 4.0. The transition from an scalar value of the effective hydraulic conductivity of
irregular lattice with z* < 4.0 to a square lattice with the numerical grid block can be calculated using C**
z* = 4.0 is accomplished by filling in the missing as the effective conductance in equation 4.
segments with zero conductivity segments (Figure 1, A strong advantage of using the effective media
dashed lines in the bottom right image). theory is that three parameters controlling the hy-
A final effective hydraulic conductivity must be draulic conductivity in any grid block are derived
calculated for the grid block containing the square from readily obtainable fracture observations. These
fracture network with the zero-conductance seg- parameters are summarized in Table 1. Another
ments. Note that in the example shown in Figure 1, advantage of the EMT approach is that fracture con-
a total of eight fracture intersections with the grid- nectivity, parameterized as the coordination num-
block boundaries are present, or an average of two ber, is explicitly considered in the derivation of the
intersections per side. Maintaining this average re- effective permeability. This is in contrast to other
sults in two vertical and two horizontal fractures approaches commonly used to derive effective per-
through the grid block. C* is treated in this work as meability from discrete fracture information (e.g.,
a scalar quantity that applies to flow through the Snow, 1969; Oda, 1985). It is noted that fracture
block in all three principal directions. More complex length does not explicitly enter the effective medium
approaches to determining anisotropic measures of formulation, but indirectly, fracture length will in-
conductance and the number of horizontal and ver- fluence the observed fracture spacing and connec-
tical fractures have been evaluated by Hestir and tivity (see Renshaw, 1999).
Long (1990); however, the relatively simple approach
taken here follows the work of Zimmerman and
Bodvarsson (1996).
The effective value of fracture transmissivity in STOCHASTIC MODELING
the grid block can now be calculated by recognizing
that the value of C* derived from equation 5 must The effective grid-block conductivity (equation 4)
be modified to consider that the actual coordina- is a function of three fracture parameters defined lo-
tion number is less than 4.0. The fraction of zero- cally at the scale of the numerical grid block: (1) the

Table 1. Listing of the parameters used to define the EMT approximation to permeability and the
corresponding parameters observable in discrete fracture systems along with general data sources.
Effective Media Effective Media Discrete Fracture
Parameter Symbol Parameter Data Sources

Fracture spacing Nf Fracture spacing Borehole and outcrop observations


Coordination number z Fracture connectivity Outcrop observations
Conductivity Ci Fracture transmissivity Downhole pressure testing
Fractured Continuum Approach to Stochastic Permeability Modeling 177

average conductance of the individual fractures; Permeability Classes


(2) the number of fractures in the grid block; and
(3) the average coordination number. A second step The Poisson distribution of fracture frequency
in the modeling approach is necessary to populate dictates that only 40% of the grid blocks will contain
a larger 3-D domain and address spatial variability fractures. The fracture length distribution shows that
in fracture permeability. These three fracture parame- 97.8% of the fractures are below the 7-m (23-ft) mea-
ters are considered spatial random functions and surement truncation limit. With such a high propor-
are modeled using Boolean and geostatistical simu- tion of small fractures, it is reasonable to conclude
lation. The development of this second step of the that all grid blocks will have some fracture perme-
modeling approach is demonstrated with an exam- ability. To accommodate the truncated small-radius
ple problem. Results of the models are described in fractures, two permeability classes (high and low)
terms of permeability [L2]. are defined. Boolean simulation is used to define the
spatial distribution of the two permeability classes.
The Boolean simulation algorithm is used to gen-
Available Data erate zones of high-permeability grid blocks that
occupy 60% of the model domain and are centered
For the example problem, information regarding at random locations throughout the model domain.
the shape, radius, frequency, orientation, transmis- These zones represent high permeability resulting
sivity, and aperture of two sets of discrete fractures from a single fracture or a heavily fractured zone
have been summarized by K. Oyamada and H. extending across multiple grid blocks. The Boolean
Takase (1998, 1999, personal communication). The simulation code is an extension of ellipsim (Deutsch
fractures are disk shaped, and the fracture radius is and Journel, 1998) that allows the radii and orien-
defined by a power-law distribution. For this ex- tations of the elliptical objects to be defined by Mon-
ample problem, the power-law distribution was trun- te Carlo draws from user-defined distributions.
cated at 7 and 300 m (23 and 984 ft) (minimum and The radii of the elliptical objects is taken from the
maximum) (K. Oyamada and H. Takase, 1999, per- truncated power-law distribution of fracture radii
sonal communication). The power-law distribution with 97.8% of the radii set equal to the lower trun-
results in a large number of small fractures, 97.8% of cation limit of 7 m (23 ft), and the orientation of the
which are less than 7-m (23-ft) truncation length. zones is drawn from a triangular distribution ap-
These smaller fractures are not modeled explicitly proximation of the Fisher distribution of deviations
and are considered to contribute to the background about the mean orientations of north south and
permeability. east west.
The frequency of the fractures over the entire An example Boolean simulation is shown in
radius distribution is described by a Poisson dis- Figure 2 (upper left image). Note that most of the
tribution with an expected value of 0.52 fractures conducting grid blocks (red in Figure 2, upper left
per 2 m (6.6 ft). The frequency is given in these units image) are arranged in small clusters, but a few
to be consistent with the grid-block size used in the large high-permeability zones span the model do-
following models. The Poisson distribution dictates main. This pattern results from the truncated power-
that only 40% of the grid blocks in the model will law distribution of the fracture radii. Most of the
contain a fracture. clusters are formed using radii at the truncation limit
Two fracture sets with mean orientations of north (7 m; 23 ft); however, a small fraction of the features
south and eastwest (0.0 and 90.08) are present. Vari- have much longer lengths up to the upper truncation
ation in the orientation about these mean directions limit of the distribution (300 m; 984 ft). The Boolean
is governed by a Fisher distribution. The log10 frac- model employs a constant anisotropy ratio that re-
ture transmissivities (m2/s) are defined by a normal sults in the width of the features being proportional
distribution with a mean of 8.99 and a standard to the length of the features. By creating the Boolean
deviation of 1.07. Fracture aperture, b, is necessary objects using the distributions of radii and orienta-
to calculate fracture porosity in each grid block and tions measured on the discrete fractures, the spatial
is defined by an empirical relationship
p with the frac- correlation of intergrid-block effective permeability
ture transmissivity, T: b 2  T (K. Oyamada and caused by fracture lengths greater than the grid-
H. Takase, 1999, personal communication). block size is ensured.
178 McKenna and Reeves

Figure 2. Two-dimensional example of the stochastic simulations of permeability class, coordination number, and
fracture frequency. These three realizations are combined to produce the fracture permeability and porosity shown in
the two images on the right.

Fracture Frequency tions of Nf are done using the indicator geostatistical-


simulation algorithm sisim (Deutsch and Journel,
The number of fractures in a grid block, Nf, is 1998). In these simulations, the minimum value of Nf
defined as an integer value for use in equation 4. for a 2-m (6.6-ft) grid block in the high-permeability
Because of the isotropy assumption, Nx = Ny = Nf, class is 1, and the maximum is constrained to 5. The
and therefore, only Nf must be simulated. The simula- probability of occurrence for the different values of
Fractured Continuum Approach to Stochastic Permeability Modeling 179

Nf (1, 2, 3, 4, or 5) is taken from the Poisson fracture portion of fractures that are longer than the 2-m
frequency distribution defined by K. Oyamada and (6.6 ft) grid-block dimension.
H. Takase (1998, personal communication). These Simple graphical constructs were examined to de-
probabilities of occurrence define the proportion termine that coordination number is a much stronger
of each value of Nf in the indicator geostatistical function of fracture length than of fracture frequen-
simulations. cy. For this example, the average coordination num-
The truncated power-law distribution of fracture ber, z*, and the fracture frequency, Nf, are considered
radii shows approximately 98% of the fractures to to be independent variables and are created through
be equal to or less than the truncation limit of 7 m independent geostatistical simulations. The multi-
(23 ft). Based on this observation, the integral scale Gaussian geostatistical-simulation algorithm, sgsim
of the model of spatial correlation for Nf is taken to (Deutsch and Journel, 1998), is used to create reali-
be 14 m (46 ft) (length of fracture is twice the radius). zations of the coordination number. No information
An isotropic, exponential semivariogram model is regarding the spatial correlation of the coordination
used with an effective range of 52 m (170 ft) (three number is present, and an isotropic, spherical semi-
times the integral scale). The resulting indicator geo- variogram with a range of 5 m (16 ft) (2.5 grid blocks)
statistical simulations show a mosaic pattern of ho- is used to create the coordination number simula-
mogeneous zones of fractures per grid block (Nf) tions (Figure 2, center left image). The values of the
(Figure 2, lower left image). coordination number range from above 2 to 4, and
the univariate distribution across the 2-D domain
Average Coordination Number matches the triangular distribution defined above. The
short range (5 m; 16 ft) employed in the coordina-
The average coordination number for a fracture tion number simulations creates the somewhat spa-
system in a grid block is a complex function of the tially random appearance of the simulation (Figure 2,
fracture lengths, orientations, and frequency in that center left image).
grid block. For the basically orthogonal system con-
sidered here, a maximum upper bound on the aver- Permeability and Porosity
age coordination number is 4. This situation corre-
sponds to every fracture being continuous across The final information necessary to create the final
the grid block and thus intersecting every fracture image of effective grid-block permeability is the trans-
from the opposite set (those in the orthogonal di- missivity of the individual fractures in each grid
rection). Given the specified fracture length distri- block. For this study, these transmissivities are drawn
bution (K. Oyamada and H. Takase, 1999, personal independently from two different log-normal distri-
communication), the proportion of fractures in a 2-m butions. For grid blocks in the high-permeability class,
(6.6-ft) grid block that are longer than 2 m (6.6 ft), as defined by the Boolean simulation, the fracture
thus leading to a coordination number of 4, is high transmissivities are drawn randomly from the log-
but less than 1. The absolute lower bound of the av- normal distribution (mean = 8.99, standard devia-
erage coordination number is 2 corresponding to the tion = 1.09) defined in K. Oyamada and H. Takase
case of straight lines cutting across the grid block in (1998, personal communication). Log10 (m2/s) frac-
a single direction or to elbow joints in a set of rough- ture transmissivities for the low-permeability class
ly orthogonal fractures. The probability of an aver- grid blocks are drawn from a second normal distri-
age coordination number being 2 for this data set is bution of lower fracture transmissivities with a mean
negligible. Given the theoretical constraints on the of 14.0 and a standard deviation of 0.5. The combi-
coordination number of a square lattice as 2 and 4 for nation of the Boolean model of permeability classes,
the minimum and maximum, respectively, the cu- along with the geostatistical simulations of coordina-
mulative distribution function (CDF) is constructed tion number and Nf, as well as the transmissivity of
with a triangular distribution to define the univariate the individual fractures using equation 4, produces
distribution of coordination number. The minimum, the final grid-block permeability map (Figure 2, up-
mode, and maximum of the distribution are set at 2.0, per right image).
3.5, and 4.0, respectively. The value of the mode, Fracture porosity in the high-permeability class is
being closer to the theoretical maximum instead of calculated as the total volume of the fractures per
the minimum, is caused by the relatively high pro- grid-block volume. Because this modeling approach
180 McKenna and Reeves

does not explicitly define the orientation and connec- (e.g., 2 m [6.6 ft]) that fracture properties can be
tivity of the individual fractures in each grid block, assumed constant over the size of the grid block.
it is necessary to assume that each fracture is fully An example permeability realization on a 202 
connected across the grid blocks (z* = 4.0) for the 200  202-m (662  656  662-ft) domain is shown
porosity calculations. This assumption will slightly in Figure 3.
increase the fracture porosity of the grid blocks
relative to an explicit porosity calculation using
the same fracture parameters. Porosity in the low- Modeling Stages
permeability class is set to a constant value of
5  105. An example porosity image is shown in In the development of the FCM, the intersection
the lower right image of Figure 2. From Figure 2, of multiple, high-permeability features in a single
it is seen that porosity is a stronger function of frac- numerical model grid block does not have an ex-
ture frequency than permeability. plicit effect on the fracture permeability assigned to
In summary, the 2-D effective media approxima- that grid block. To examine the effect of increased
tion for permeability of fracture networks in a single high-conductivity regions in a grid block on per-
grid block can be employed with Boolean and meability, a second stage of modeling (stage 2) is
geostatistical simulation to develop 3-D models of created with correlation between the high- and low-
fracture permeability. This development is made by conductivity regions of the model and the fracture
assuming that the 2-D system in any numerical grid frequency and coordination number. Instead of just
block can be extended uniformly in the third dimen- assuming that the high-conductivity regions are
sion up to the size of that grid block. Three factors caused solely by the local presence of relatively
allow this extension to three dimensions: (1) the two higher transmissivity fractures, a stage 2 model con-
fracture sets specified for this work are in the X and siders that the high-conductivity grid blocks are also
Y planes, and no third fracture set is present in the the result of higher fracture frequency and greater
Z plane; (2) the fracture sets have a vertical orien- fracture connectivity.
tation (zero dip angle); and (3) the size of the nu- In stage 2 models, the values of both the number
merical model grid blocks can be kept small enough of fractures and the average coordination number

Figure 3. Example
realization (stage 1
approach) of the frac-
ture permeability in
log10 (m2) in the
model domain.
Fractured Continuum Approach to Stochastic Permeability Modeling 181

Table 2. Relationship between proportion of the flow modeling transport results between the
high-conductivity grid blocks in the template two stages. A comparison of stage 1 and stage 2 per-
and fracture frequency. meability models is given in Figures 4 and 5.
In general, the stage 2 approach creates a more
Proportion of Template in Assigned Fractures
High-conductivity Features per Grid Block distinct permeability difference between the areas in
the Boolean objects and the background material.
<0.222 1 Note the greater contrast in permeability values be-
0.222 0.444 2 tween the two classes in stage 2 (Figure 4, lower im-
0.444 0.666 3 age) relative to stage 1 (Figure 4, upper image). The
0.666 0.889 4 extreme high and low permeability values are of
>0.889 5 greater magnitude in the stage 2 models (Figure 5).
Two ensembles of 50 realizations each (stages 1 and 2)
are constructed for the 202  200  202-m (662  656 
662-ft) 3-D volume.
drawn from the respective distributions are con-
nected to the locations of the high-conductivity fea-
tures through the use of a 3  3  3 grid-block
template. To assign properties to the central block of
the template, the fraction of high-permeability grid
blocks defined by the Boolean model in the template
is calculated. This fraction is used to determine the
cumulative probability of the fracture frequency and
the average coordination number when drawing
from those distributions. Therefore, in stage 2, the
probability value applied to the CDFs of fracture
frequency and average coordination number is not
randomly determined as in stage 1 but is defined as
the fraction of the blocks in the nine-point template
that contains high-conductivity features. The tem-
plate size of 3  3  3 grid blocks limits the influence
of the surrounding model grid blocks to only those
that are adjacent to the current grid block being
calculated.
For stage 2 permeability models, a nearly uni-
form distribution of the number of fractures in a
grid block is used. Table 2 lists the number of frac-
tures assigned to each grid block as a function of the
proportion of the 3  3  3 template in the high-
conductivity zone. The remainder of stage 2 mod-
eling is identical to that used for stage 1, with the
exception that the assignment of the average coor-
dination number is based on the proportion of the
template in the high-permeability class. Stage 2 mod-
els require less computational effort than do the
stage 1 models because it is not necessary to create
the geostatistical simulations of average coordina-
tion number and fracture frequency. The Boolean
simulations used to define the regions of high and low Figure 4. Permeability models created with the stage 1
conductivity are identical to those used in stage 1. approach (upper image) and the stage 2 approach (lower
image). Both of these images show the 30th layer of
Maintaining a constant Boolean template for each realization number 1. The color scale indicates perme-
pair of realizations allows for direct comparison of ability in log10 (m2).
182 McKenna and Reeves

The proportion of the model domain in the Bool-


ean objects is compared to the target proportion of
grid blocks containing conductive fractures. Orien-
tations of the two sets of Boolean objects are com-
pared to the orientations of the two fracture sets, and
the proportion of the Boolean objects with lengths
less than or equal to the truncation limit of 7 m
(23 ft) is compared to the same proportion as speci-
fied in the truncated power-law length distribu-
tion (Table 3). Comparative statistics (minimum,
mean, and maximum) are computed using the re-
sults from all 50 realizations. Recall that the same
50 Boolean model realizations are used for both stage
Figure 5. Stage 1 and stage 2 permeability values along 1 and stage 2. Model evaluation results shown in
a transect from (0.0,100) to (0.0,+100.0) along the Table 3 indicate that these three fracture attributes
middle layer (layer 51) of realization 1. are accurately reproduced by the Boolean model
across the 50 realizations.

MODEL APPLICATION
Flow Model and Particle Tracking
Two ensembles of 50 realizations each are created
using stage 1 and stage 2 modeling approaches. The The 100 realizations of fracture permeability and
results of the permeability modeling are evaluated porosity created in stages 1 and 2 are used as input
for their ability to reproduce the observed fracture to a parallel flow code, POR-SALSA (Martinez et al.,
data used as input. These models then serve as in- 1997), that is used to obtain a single-phase, steady-
put to a set of groundwater flow and streamline state solution to the pressure field with prescribed
transport calculations. pressures at two ends and no flow boundaries on
the sides of the model (Figure 6).
Model Evaluation A total of 600 particles are introduced into the
flow field uniformly along the central 100 m (328 ft)
The FCM realizations are checked by comparing of a three-element-wide disturbed zone surround-
the frequency, orientations, and lengths of the Bool- ing a backfilled tunnel that extends through the do-
ean objects against the distributions of the same main (Figure 6). The tunnel is horizontal and cen-
parameters specified in the input data for the frac- tered along the lateral face of the domain midway
tures. These three parameters provide the only data between the upstream and downstream boundaries.
against which the models can be directly compared; Particles are tracked along their respective stream-
there are no values of permeability or porosity mea- lines to the downstream boundary of the model do-
sured at the 2-m (6.6-ft) grid-block scale. main. The average hydraulic gradient is 8.0  103,

Table 3. Comparison of the Boolean model results to the target values of fracture frequency,
orientation, and length.
Parameter Target Value Minimuma Meana Maximuma

Proportion of high-permeability zones 0.405 0.406 0.406 0.409


Orientation: Set 1 (8) 0.00 0.51 0.02 0.49
Orientation: Set 2 (8) 90.00 89.57 89.96 90.41
Radii less than 7 m (23 ft): Set 1 (%) 97.8 97.7 98.0 98.2
Radii less than 7 m (23 ft): Set 2 (%) 97.8 97.7 98.0 98.1
a
The minimum, mean, and maximum values are calculated across all 50 realizations. The two different fracture sets are defined by the
two different target orientation values of 0.0 and 90.08.
Fractured Continuum Approach to Stochastic Permeability Modeling 183

Figure 6. Schematic diagram of the groundwater flow model setup and the particle starting locations. The particle
starting locations are shown as white dots in the expanded view and are not drawn to scale.

and the total number of 2  2  2-m (6.6  6.6  shown. Stage 2 permeability models produce longer
6.6-ft) elements in the flow model is 1,020,100. minimum and median arrival times than stage 1
The flow solutions for this study were accom- models. Relative to stage 1 models, stage 2 models
plished using 20 Digital Equipment Corporation have higher permeability in the Boolean objects and
Alpha processors (612 MHz clock speed) with 4 GB lower permeability in the background material. The
of shared random-access memory. The flow code increase in traveltimes in stage 2 models is caused
required 15 25 min to achieve a stable pressure field by the increase in fracture porosity in the high-
for each realization. Using a single processor, less permeability parts of the domain defined by the Bool-
than 10 sec were required to track the 600 particles ean simulation.
through each realization. The speed at which the The increase in fracture porosity in the high-
pressure solutions are obtained allowed flow and permeability materials is caused by the functional
particle tracking across 50 realizations to be accom- dependence of the fracture apertures on the mea-
plished overnight. An obvious advantage of the par- sured fracture transmissivity and the more uniform
allel approach is that multiple realizations of several fracture frequency distribution employed in stage 2
different conceptual models of the fracture system models. Both of these factors increase the cross sec-
can be explored in a reasonably short amount of time. tional area available for flow, thus decreasing the
The results of the particle traveltime calculations flow velocities in the model grid blocks relative to
for stage 1 and stage 2 are shown in Figure 7. For the stage 1 approach. The result is an increase of
both stages, the cumulative distributions of the min- roughly a factor of 2 in the minimum and median
imum, median, and maximum arrival times are arrival times.
184 McKenna and Reeves

Figure 7. Cumulative dis-


tributions of minimum, me-
dian, and maximum particle
arrival times at the down-
stream face of the model do-
main for both stage 1 and
stage 2 models.

The porosity of the low-permeability class mate- on discrete fracture systems into a constitutive model
rial is set to the same constant value as in the stage 1 of fracture permeability. The FCM serves as a prac-
models (5  105). Given the similarity of stage 1 tical bridge between DFN and SCM approaches. The
and stage 2 maximum traveltime distributions in combination of approximating the local permeability
Figure 7, it is inferred that the pathways that pro- with an effective medium approximation, based on
duce the maximum traveltimes are predominantly readily observable fracture attributes and spatial
within the low-permeability class material. The spa- simulation techniques, provides a traceable path from
tial arrangement of this material did not change discrete fracture observations to final permeability
from stage 1 to stage 2. models. In addition, the FCM approach avoids the
The standard deviation of the 600 traveltimes in large-scale averages commonly employed in the SCM
each realization was also calculated. One-to-one approach in deference to local averages that better
comparison of the standard deviations for each in- honor the behavior of the underlying DFN. The FCM
dividual realization shows a strong correlation be- approach is flexible enough to simulate different
tween stage 1 and stage 2 results. This result indi- conceptual models of the fracture permeability.
cates that the changes in the approach to modeling Flow modeling over multiple realizations of the
the permeability and porosity from stage 1 to stage FCM is simplified by the use of a uniform orthog-
2 do not affect the variability of traveltimes in a given onal mesh. For the example problem of groundwater
realization. The spatial distribution of the Boolean flow from a hypothetical nuclear waste repository,
objects remains the same in both modeling stages, these results show that the range, maximum to mini-
and we conclude that the distribution of these ob- mum, of groundwater traveltimes is approximately
jects is the primary control on traveltime variability one order of magnitude for both modeling stages.
within a realization. The second-stage models increase the porosity of
the high-permeability regions relative to stage 1 mod-
els and thus increase the minimum and median trav-
eltimes. The distribution of the longest ground-
CONCLUSIONS water traveltimes and the variation of times in a
single realization do not vary between modeling
This chapter documents the FCM approach to stages and indicate that these two results are con-
creating stochastic realizations of fracture perme- trolled by the spatial distribution of the Boolean
ability. The FCM approach provides for the quan- objects, which is held constant between both mod-
titative inclusion of multiple data types as collected eling stages.
Fractured Continuum Approach to Stochastic Permeability Modeling 185

Extensions of FCM original draft of this chapter was completed and sent to
the editors in November 1999. Sandia is a multiprogram
laboratory operated by Sandia Corporation, a Lockheed
The example application presented here was
Martin Company, for the U.S. Department of Energys
based on a relatively simple fracture system, and National Nuclear Security Administration under contract
several extensions of the FCM would be possible for DE-AC04-94AL85000.
more complex models. If nonvertical fracture sets
were included in the modeling, it would be neces-
sary to use an alternative approach (e.g., Chen et al.,
1999) to determining the grid-block-scale effective- REFERENCES CITED
permeability tensor. The current EMT approach to Altman, S. J., B. W. Arnold, R. W. Barnard, G. E. Barr, C. K.
calculating grid-block permeability can be used to Ho, S. A. McKenna, and R. R. Eaton, 1996, Flow cal-
calculate anisotropic fracture permeability. Aniso- culations for Yucca Mountain groundwater travel time
tropic calculations would be based on a different (GWTT-95), SAND96-0819: Albuquerque, New Mex-
number of fractures and average coordination num- ico, Sandia National Laboratories, 170 p., http:
//infoserve.sandia.gov/sand_doc/1996/960819.pdf
ber in each direction. If specific fracture information is
(accessed November 15, 1999).
available, the centers of the Boolean objects can be Chen, M., M. Bai, and J.-C. Roegiers, 1999, Permeability
conditioned to specific locations in the model domain, tensors of anisotropic fracture networks: Mathematical
and the FCM approach can be extended to condition Geology, v. 31, p. 355 374.
the final permeability models to local measurements David, C., Y. Gueguen, and G. Pampoukis, 1990, Effective
of permeability made on specific fractures. medium theory applied to the transport properties of
rock: Journal of Geophysical Research, v. 95, p. 6993
7007.
Model Comparison Dershowitz, B., P. LaPointe, T. Eiben, and L. Wei, 1998,
Integration of discrete fracture network methods with
The discrete fracture data provided for the ex- conventional simulator approaches: Society of Petro-
ample problem described in this chapter were also leum Engineers Annual Technical Conference and Ex-
hibition, New Orleans, Louisiana (September 27 30),
provided to six other modeling groups. Each of these
SPE Paper 49069, 9 p.
modeling groups employed different approaches to Deutsch, C. V., and A. G. Journel, 1998, GSLIB: Geostatis-
modeling fracture permeability, including several tical software library and users guide, 2d ed.: New
variations of the DFN and SCM approaches. The York, Oxford University Press, 369 p.
results of the seven different modeling approaches Finsterle, S., 2000, Using the continuum approach to model
with respect to several groundwater flow and trans- unsaturated flow in fractured rock: Water Resources
Research, v. 36, p. 2055 2066.
port performance measures have been compared
Follin, S., and R. Thunvik, 1994, On the use of continuum
(Sawada et al., 2000). This comparison shows that approximations for regional modeling of groundwater
the DFN models produce higher bulk permeability flow through crystalline rocks: Advances in Water Re-
values and faster groundwater flow velocities than sources, v. 17, p. 133 145.
do the SCM and FCM models. The bulk permeability Hestir, K., and J. C. S. Long, 1990, Analytical expressions
for the permeability of random two-dimensional
calculated for the SCM and FCM models is close to
Poisson fracture networks based on regular lattice per-
that of the harmonic mean of the fracture perme- colation and equivalent media theories: Journal of
ability distribution, whereas the bulk permeability of Geophysical Research, v. 95, no. B13, p. 21,565 21,581.
the DFN models is better approximated by the arith- Hsieh, P. A., 1998, Scale effects in fluid flow through
metic average permeability. These results appear to fractured geologic media, in G. Sposito, ed., Scale de-
be caused by the DFN modeling groups making the pendence and scale invariance in hydrology: New
York, Cambridge University Press, p. 335 353.
assumption of a constant spatially homogeneous per-
Kirkpatrick, S., 1973, Percolation and conduction: Reviews
meability value in each individual fracture. of Modern Physics, v. 45, p. 574 588.
Long, J. C. S., and P. A. Witherspoon, 1985, The relationship
of the degree of interconnection to permeability in
ACKNOWLEDGMENTS fracture networks: Journal of Geophysical Research,
v. 90, no. B4, p. 3087 3098.
Funding for this work was provided by the Japanese Long, J. C. S., J. S. Remer, C. R. Wilson, and P. A.
Nuclear Cycle Development Institute. This chapter was Witherspoon, 1982, Porous media equivalents for net-
improved through reviews by Steve Carle, Eileen Poeter, works of discontinuous fractures: Water Resources Re-
Norm Warpinski, and two anonymous reviewers. The search, v. 18, p. 645 658.
186 McKenna and Reeves

Martinez, M. J., P. L. Hopkins, and J. N. Shadid, 1997, rock, in Proceedings of the 30th Fossil Energy Sym-
LDRD final report: Physical simulation of nonisother- posium (in Japanese): Japanese Society for Civil Engi-
mal multiphase multicomponent flow in porous media, neering Monograph 17, p. 123 127.
SAND97-1766: Albuquerque, New Mexico, Sandia Na- Snow, D., 1969, Anisotropic permeability of fractured
tional Laboratories, 65 p., http://infoserve.sandia.gov media: Water Resources Research, v. 5, p. 1273 1289.
/sand_doc/1997/971766.pdf (accessed November 15, Tsang, Y. W., C. F. Tsang, F. V. Hale, and B. Dverstop,
1999). 1996, Tracer transport in a stochastic continuum model
Oda, M., 1985, Permeability tensor for discontinuous rock of fractured media: Water Resources Research, v. 32,
mass: Geotechnique, v. 35, p. 483 495. p. 3077 3092.
Renshaw, C. E., 1999, Connectivity of joint networks with Zimmerman, R. W., and G. S. Bodvarsson, 1996, Effective
power law length distributions: Water Resources Re- transmissivity of two-dimensional fracture networks:
search, v. 35, p. 2661 2670. International Journal of Rock Mechanics, Mineral
Sawada, A., M. Shiotsuki, K. Oyamada, and H. Takase, Science and Geomechanical Abstracts, v. 33, p. 433
2000, Study of flow model comparison in fractured 438.

McKenna, S. A., and P. C. Reeves, 2006,


Fractured continuum approach to stochas-tic permeability modeling,

J. M. Yarus, and R. L. Chambers, eds., Sto-chastic modeling and geostatistics: Princi-


ples, methods, and case studies, volume II: AAPG Computer Applications in Geology 5,
p. 173 186.

You might also like