Copyright 2000, Society of PetroleumEngineers Inc.

This paper was prepared for presentation at the SPE European PetroleumConference held in
Paris, France, 2425 October 2000.
This paper was selected for presentation by an SPE ProgramCommittee following review of
information contained in an abstract submitted by the author(s). Contents of the paper, as
presented, have not been reviewed by the Society of PetroleumEngineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of PetroleumEngineers, its officers, or members. Papers presented at
SPE meetings are subject to publication review by Editorial Committees of the Society of
PetroleumEngineers. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the Society of PetroleumEngineers is
prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300
words; illustrations may not be copied. The abstract must contain conspicuous
acknowledgment of where and by whomthe paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.
ABSTRACT
In fractured reservoirs, data directly related to fractures are
scarce and uni-dimensional (i.e. cores and image logs). Other
types of data are better distributed and have proved to be
related to fracturing but only indirectly (e.g. lithology or large
scale structure).
In such reservoirs, however, one has to understand
fracture distribution and behavior at the field scale. A
methodology has been developed within TotalFinaElf to define
the relationships of all sources of data to fracturing and to
integrate them and compared to another independent
published method. To that end, a systematic work flow which
goes from 1D to 2D and from static to dynamic data has been
defined and various technologies tested.
A field case in North Africa is taken to illustrate this
methodology. In this field, fracture data from image logs have
been related to: 1) production data; 2) 3D seismic attributes
(coherency, amplitude, structural curvature) and fault
interpretation and strain; 3) log data such as porosity,
thickness and lithology index. The former type of data is used
to understand the contribution of each fracture set to flow. The
latter two types of data are used to better map fracture
distribution at the field scale. Ultimately, this mapping is
calibrated with the production data of the other wells where
fracturing data are not available and is then used to validate
the specific role of fracturing in this field.
A better reservoir simulation and infill well planning can
be subsequently achieved.
Introduction
Fractured reservoirs are by nature highly heterogeneous. In
such reservoirs, fracture systems control permeability and can
also control porosity. Fracture modeling is therefore a key
development issue and requires an integrated approach from
geology to reservoir simulation and well planning. The
geometry (i.e. static model) of the fracture network is
generally defined from well data (i.e. cores or image logs)
using conventional structural geology techniques. Then,
fracture permeability can be assessed by relating the fracture
aperture to the fracture excess conductivity measured on
electrical image logs
1
and/or to critically stressed fractures
within the present day stress field
2
. However, it is the authors
opinion that such approaches can only give, in the best case, a
relative estimate of the fracture permeability. A quantitative
modeling of fracture flow behavior is therefore required (i.e.
dynamic model). At the well scale, this can be done by
constructing Discrete Fracture Networks (DFN)
3-4
through
which flow is modeled and which are matched to well test
data
5
. Ultimately, these models can help in determining the
fracture parameters required in dual porosity / dual
permeability reservoir flow simulation
6,7
. However, if these
DFN models are appropriate for reservoir sector models, their
application to full field simulation is somewhat difficult since
their extrapolation outside the well scale can be limited by the
heterogeneous vertical and lateral distribution of the fracture
networks. The modeling of the spatial distribution of
fracturing at the scale of the entire field and its calibration to
well data is the purpose of this paper.
Methodology
Methods have been presented to model fracture distribution
based solely on fracture density measured in wells
8
. We
believe that the scarcity of wells where fracturing data are
available makes difficult and very uncertain such direct
mapping, at the field scale, of these data. Fracturing process
can be shown, however, to be related to geological parameters
that constrain the mechanical behavior of fractured rocks.
Among others, one can mention lithology, bed thickness,
stress and strain states, large scales features (e.g. faults,
doming). These parameters are often better sampled at the
field scale (especially when derived from seismic attributes).
The relationships between the fracturing variable and such
geological parameters (so called drivers throughout this paper)
can be exploited to model the spatial distribution of fracturing
within a field. Although only one parameter may be needed to
characterize a fractured reservoir
9-12
, it is often the lack of a
methodology to integrate the combined effects of structure,
thickness, and lithology that leads geologists to only focus on
the most important factor. However, complex reservoirs
require a more comprehensive description to allow a reliable
SPE 65118
Integrated Fractured Reservoir Characterization: a Case Study in a North Africa Field
B.D.M. Gauthier; A.M. Zellou, SPE; A. Toublanc, TotalFina Elf; M. Garcia, FSS, J -M. Daniel, IFP
2 B. GAUTHIER, A.M. ZELLOU, A. TOUBLANC, M. GARCIA, J -M. DANIEL SPE 65118
fracturing prediction. Ericsson et al,
13
published an empirical
and deterministic approach to derive a fracture density index
as a function of other indices related to reservoir variables
such as the structural curvature, the crestal distance and facies
types. Similarly, Argawal et al,
14
related fracture intensity to
geological parameters on one hand and to effective
permeabilities on the other hand to model the field
permeability distribution. Their approach is less empirical but
remains fully deterministic. Based on scarce data, the
relationships between fracturing and controlling geological
parameters are uncertain. Therefore deterministic approaches
are not appropriate (i.e. there is not one single solution). Only
probabilistic modeling can help in better assessing these
uncertainties. Such a probabilistic approach has been
published
15-17
where fracture indices are assumed to be
correlated to some geological parameters of the field (e.g.
reservoir thickness, structural curvature) and rock petrophysics
(e.g. sand resistivities) by some complex, non linear
relationships. These relationships are established using fuzzy
logic techniques to rank the relative importance of the drivers
upon fracturing and are then modeled throughout the entire
field with neural network systems (see below). Independently,
TotalFinaElf and FSS International have developed in the last
two years another probabilistic approach where the
relationships between fracturing and fracture drivers are
established through a multivariate discriminant analysis and
the fracture sets (if several) are modeled jointly using
sequential simulation techniques (see below). These two
different approaches are here tested, compared and ultimately
calibrated against well production data in a North Africa field.
Fracture Index (FI)
The fracture index is the variable that we relate to the fracture
driver. In published work
15-17
well performances in tight
fractured reservoirs are assumed to be directly (i.e. linearly)
related to fracturing and are used as a fracture index. It should
be noted that using well performance data as an index to
fracturing may be erroneous for reservoirs where the matrix
also contributes to flow. In our workflow, the fracture
distribution is modeled solely from fracture data acquired in
image logs or cores and from drivers that are known to control
their initiation and propagation (Figure 1). The relation
between fracturing and well performance is made at a second
stage through calibration. Therefore in this paper, the fracture
frequency which is defined as the number of fractures per unit
length measured perpendicularly to the fracturing plane is the
fracturing variable used as FI. Fracture frequencies are
calculated for families of orientations (i.e. fracture sets) and
geometrically corrected from fracture densities along wells
which are well direction-dependent
18
.
Methodology One: Fractured Reservoir Modeling Using
Fuzzy Logic and Neural Networks
The complexity and the numbers of fracturing drivers in this
study justifies an integrated approach to characterize natural
fracturing. A thorough description of this integrated approach
using artificial intelligence tools is given in Ouenes.
19
In this
paper, we will limit ourselves to a short summary of the
methodology based on the three following steps:
1. Ranking the fracturing drivers: when considering
several possible fracturing drivers, the relative importance of
each driver (the input) and the impact on the FI (the output) is
critical. Given the dataset, many different techniques can be
used to rank the fracturing drivers. In this methodology, we
employ a fuzzy logic tool that has proven very efficient and
fast in ranking geologic parameters. In this specific study, we
dealt initially with over 15 fracturing drivers. After ranking
and eliminating the less influential drivers, the geologist and
the engineer have to analyze, understand and validate this
ranking and the effect of each geological parameter on the
fracturing. The description of the fracturing drivers for this
specific study is given in the application section of this paper.
2. Training and testing: the above ranked (and
validated) drivers are used to establish the complex, non
linear relationship relating the FI to these drivers. This is done
using a neural network technique. To that end, the set of
available data, where the inputs and output are known, is
divided into two subsets: a training set and a testing set. Each
regression model is derived by selecting randomly (or
according to some rule) the training set. The neural network
modeling process consists of adjusting some weights until the
actual FIs match the estimated ones. Once the matching
process is done, we can assume that a model is available and
can be used for testing and cross-validation. Depending on the
ability of the model to predict the fracture index at testing
locations, the model can be kept for further use or discarded.
3. Simulation process: the stochastic aspect of this
approach is related to the uncertainty about the above
regression models. Therefore, several realizations or
simulations are generated cell by cell over the entire field
where the fracturing drivers are known to predict the FI. With
a large number of realizations, a probability and average
model is estimated over the entire field.
Methodology Two: Geostatistical Approach using Sequential
Simulation
The other solution to the data integration problem of modeling
directional fracture sets is through a geostatistical approach
that is based on both multivariate statistical analysis and
sequential simulation. Like in the previous methodology, the
aim is to characterize the spatial distribution of a fracturing
variable (the FI) and to quantify the uncertainty about it.
Unlike the previous methodology, which relies on multivariate
non linear regression and addresses the uncertainty on the
regression model, the objective here is to generate fracture
frequency realizations that honor fracturing data at data
locations and reproduce different statistical models inferred
from the data. These models include: the global distribution of
the FI, variogram models that specify the spatial correlation of
the FI and its cross-correlation with a geological component
capitalizing the geological information (see hereafter the
discriminant analysis section), and a bivariate distribution
model of the FI and the geological component. The bivariate
distribution model provides the conditional distribution of the
SPE 65118 INTEGRATED FRACTURED RESERVOIR CHARACTERISATION: A CASE STUDY IN A NORTH AFRICA FIELD 3
FI for any geological component value. The geological
component being known everywhere over the field, this model
provides the a priori FI distribution (prior to simulating) at any
estimation point according to the collocated geological
component value. In that sense, it provides information about
spatial fracturing trends.
1. Discriminant Analysis: As discussed above, it is
several fracture drivers (explicative variables) taken all
together that allow identifying spatial statistic trends. To
integrate them in the geostatistical approach, discriminant
analysis is used to determine the linear combination of the
drivers that distinguishes at best selected FI classes (e.g. low,
medium and high fracturing classes). Because the trend may
not be linear, all drivers are first transformed into distribution
probabilities (cumulative distribution function values). From
the discriminant analysis, only the first component (i.e. the so
called geological component) is retained, provided it indeed
shows a monotonic relationship with the FI. The scatterplot of
FI versus the geological component typically shows an S-
shape cloud with two plateaus for the small and high FI
values.
2. Sequential Simulation: To better represent the rather
complex (non-linear and heteroscedastic) relationship between
the fracture frequency and the geological component,
sequential indicator simulation is used to simulate the fracture
frequency. The integration of the geological component
(secondary information) is done by cokriging through a
Bayesian formalism in which the geological information is
converted into soft (probability-like) data
20,21
. A Markov
model is adopted, which leads to a collocated cokriging
simulation approach with the cross-variogram (between the
fracture index and the geological component) being written as
a function of one of the two auto-variograms
22
.
Application to a North Africa Field
Geology of the Field
The case study is an oil bearing field. Porosity ranges from 10
to 35 % (average 20%) . The matrix permeability measured
from core plugs ranges from 0 to as high as 1000 mD (several
tens of mD in average) whereas the well test permeability
ranges from 10 to 110 mD. The ratio between test and plug
permeability, which is generally interpreted as an indicator of
fracturing effects varies from 1 to 10. Therefore, this field
cannot be considered strictly as a fractured reservoir since a
minimum ratio of ten between these two types of permeability
measurements is considered as an indicator of dual
permeability behavior. Subsequently, single porosity single
permeability reservoir simulations give generally good well
history matches. However, bad matches observed in some
locations would be due to high permeability streaks related to
small scale faults and/or fracture swarms. In other locations,
fracturing is believed to homogeneously improve the matrix
performance.
In this field, a carbonate sequence of Late Paleocene age
subdivides an oil producing upper reservoir unit and a mostly
water bearing lower unit. These two units are separated by
non-reservoir shale/carbonate alternations. The upper reservoir
unit is subdivided into 4 layers 5 to 40 feet thick. They consist
in tight limestone, variably argillaceous, to good porosity
limestone (locally inter-bedded with dolomitic streaks) and
calcareous shale. Our study focuses on the good porosity
limestone layers.
The structural history of the field is quite complex but is
generally interpreted as a transtensional basin, initialised in the
Cretaceous and reactivated during the Tertiary. However, the
tectonic phase(s) post-dating reservoir deposition is (are)
poorly documented. Figure 2 shows the structural map at top
reservoir with the fault pattern at a deep level overlaid. One
can note the change in the main strike of the seismically
defined faults with depth. At top reservoir, the faults are
mainly oriented N120-130. This fault network developed just
above the deep main trends which strike preferentially N170.
Subsequently, the field structure can be explained by
reactivation of deep basement faults within an oblique
extension regime. This structural style results at top reservoir
in the dominant oblique NW-SE normal faults and in N170
secondary faults and flexures. From the fault pattern in the top
teservoir, the deformation seems to be concentrated on the
flanks of the structure just above the deep faults. Indeed, the
top fault density and continuity increase in the southern part of
the field where the two main deep faults join.
Fracture Network Characterization
In faulted/fractured reservoirs, it is now well known that both
3D seismic and horizontal wells provide key data
23
. Two main
points are highlighted in the following:
How can the fault map be used to define fault control on
small scale fracturing?
How can horizontal wells provide key argument to define
the origin of fracture clusters?
1. 3D Seismic: The fault map is classically used to
control the reservoir structural map and sometimes to modify
inter-block transmissibilities in the flow simulation grid when
faults control the reservoir dynamic behavior. Less
conventionally, various attributes of the fault network can be
computed to map the areas where faults could influence
reservoir properties. This should be done when faults are
suspected to induce significant strain accommodated by
fractures or pressure solution, or, when an heterogeneous
strain is accommodated by a combination of faulting,
fracturing and pressure solution. At first, fault orientation and
distance-to-fault maps contain useful information. In addition,
the non-stationarity of the fault network can be characterized
by a fault density map computed by summing the fault length
in a moving window. The main drawback of this map is that a
single long fault with a large throw is not differentiated from a
lot of short faults with small throws. To overcome this
limitation, instead of summing lengths, the vertical throw can
be integrated along the fault trace in the moving window.
Ws dl VTi S
N i
i
i / ) * (
1
=
=
=
N=number of fault segments
dli =length of segment i
VTi=average vertical throw along dli
4 B. GAUTHIER, A.M. ZELLOU, A. TOUBLANC, M. GARCIA, J -M. DANIEL SPE 65118
Ws=Surface of the moving window
This measure has the same dimension as strain. It weights the
faults according to their throw. As a first approximation, it will
be called strain in the following, even if an exact measure of
strain requires more information than the throw defined on the
mapped faults. In that sense, a strain map can be made.
(Figure 3).
In this study, the faults are mapped from 3D seismic data.
At first, faults are detected using mainly the displacement of
the top reservoir horizon. In addition, interactive shadings of
the horizon support the picking of tiny features that are
laterally coherent. The simultaneous analysis of several
seismic attribute maps demonstrates that these features also
correspond to low amplitude and low coherence lineaments, a
standard fault signature
24
. This interpretation has also been
validated using horizontal wells (see below). The final map
(Figure 1) includes faults with a vertical throw less than 10
meters. In this study, horizon shadings are able to reveal more
subtle lineaments than any of the other methods because they
are not disturbed by seismic noise affecting the amplitude
whereas they efficiently highlight geometric correlation. Using
this approach, and the results of the image log interpretations,
one can also demonstrate that fractures tend to be more
numerous near sub seismic scale faults.
2. Horizontal wells: Although some 60 wells were
drilled in this field, the only available fracturing data come
from image logs along 2 vertical wells and 5 horizontal wells.
Two of them are located in the northern part, four in the
central part and one in the southern part of the field. The five
horizontal wells, drilled in the main reservoir unit, have
various orientations with some of them turning 90 along hole
(multidrains). This dataset can therefore be considered as
representative of the field fracture network. The quality of the
image logs is good to relatively poor. Figure 4 shows the
overall fracture interpretation. One can note that the dominant
fracture trend strikes parallel to sub-parallel to the seismic
fault trend. Two other secondary preferential orientations can
also be locally noticed: N-S and E-W. Figure 5A shows the
orientation distribution of the fracture in each studied well
where the size of the rose diagram radii are proportional to the
apparent average fracture density. Figure 5B shows one
example of the distribution of the fracture along a horizontal
well. This simple interpretation showed that i) the fracture
frequency is much higher in the southern drain compared to
the other horizontal wells; ii) the fractures are generally
distributed in a background fracturing (0 to 2 fractures/m) and
in fractures swarms (3 to 9 fractures/m).
In order to better define the parameters that control the
observed fracture distribution, horizontal wells can also be
used further. Indeed, these wells contain high resolution data
about logs, facies and fracture density. Their 3D analysis in a
geomodeler together with reservoir layers constrained by the
seismic survey and with picked faults offer a powerful
integrated picture of the reservoir. Highlighting correlation
along the well is very efficient to refine the picking of intra-
reservoir layers from logs, especially when horizontal drains
cross repeatedly the same reservoir marker. In addition, the
interactive edition of marker and layer surface associated with
fault visualization is a very good way to measure the reliability
of the fault map. Such a visualization can also be used to
understand the occurrence of fracture clusters.
Along a well, highly fractured intervals reflect the
intersection of the well with either fracture swarms or highly
fractured layers (Figure 6). The very contrasted impacts of
these two configurations on the fracture modeling strategy and
on the hydraulic behavior requires to recognize the prevailing
configuration. Generally, when fracture clusters are correlated
with a facies, the second hypothesis is retained. However, this
argument does not account for the fact that fracture swarm
can enhance diagenesis. 3D modeling provides a unique
opportunity to demonstrate that the correlation between
fracture cluster and dolomite is mainly due to the fact that
some wells repeatedly intersect thin dolomitic layers
particularly in the southern part of the field.
Factors affecting fracturing in this specific field
Nelson
25
documented the geological parameters controlling the
fracture frequency in sedimentary layered rock: composition,
grain size, porosity, bed thickness and structural position. To
incorporate all those information, but the grain size, and to be
the most comprehensive (i.e. geologically validated) with the
parameters influencing the fracturing, four main sources of
data are used here:
- seismic derived attributes, i.e. amplitude and coherence, top
and deep fault maps;
- log derived data, i.e. porosity and gross thickness;
- core related data: dolomite, calcite and illite content;
- computed data: strain (as defined in the previous section),
maximum structural curvature and distance to top faults.
All these fracturing drivers are mapped over the entire field,
driver values being assigned or calculated at the center of a
100*100 meter cell (with a total of 16534 cells).
This paper only focuses on the modeling of the dominant
NW-SE fracture set interpreted in the 4 out of 5 horizontal
wells (vertical wells and NW-SE oriented C-H3 are deemed
not to be representative).
Application of Methodology One
On each of the 24 cells crossed by the 4 horizontal wells, the
FI is calculated as the average fracture frequency over the
section of the well drain crossing that cell.
Ranking the fracturing drivers: Using the initial 24 control
cells, where the drivers (the input) and the FI (the output) are
defined, a fuzzy logic tool
26
is used to rank the fracturing
drivers. A prior elimination process of irrelevant drivers, based
on the knowledge of the field, is performed using this fuzzy
logic tool. Figure 7 shows the final result of this ranking which
confirms that the two main controls of fracturing in this field
are the lithology (with the dolomite and the calcite content)
and the tectonic event (with the strain and the Deep faults).
This ranking process allows a better understanding of the
SPE 65118 INTEGRATED FRACTURED RESERVOIR CHARACTERISATION: A CASE STUDY IN A NORTH AFRICA FIELD 5
fracturing and has the benefit of improving the full field
fracture modeling by keeping only the most relevant drivers.
In this case, the first eight are selected.
Training and testing: Different models are derived starting
with a high number of training cells (20 out of 24) to test the
modeling limit of this technique. The final minimum number
of cells used for training and needed to build a consistent
model is 16. 8 tested cells are hidden to validate the model.
Once the predicted and the actual FI values for the training set
and the testing set are higher than a user-defined correlation
coefficient (0.8 and 0.7 respectively) the realization is kept.
Simulation results: 75 realizations of FI are performed and
a first qualitative screening is done to select the best
realizations based on the correlation coefficient and the
realization map. After this first screening is done, 40
realizations are kept to compute average and probability maps
over the entire field. Figure 10A shows the probability map at
a given threshold. This map can be compared to the one
obtained with the second methodology (see below).
Application of Methodology Two
For the geostatistical approach, fracture frequency data are
calculated every 25 m, within a moving window of 25 m. The
number of data so calculated is 55. Among the eight drivers
used with the previous methodology, six only are retained
here, the two others (calcite content and seismic coherency)
being strongly correlated to another one and thus redundant.
The discriminant analysis and simulation results can be
summarized as follows.
Discriminant analysis: The number and ranges of fracture
frequency classes influence the results. Different class
definitions may lead indeed to different discriminant functions
(i.e. geological components), which may be correlated
differently with the drivers. Two extreme cases are evaluated.
One involves a few (3) classes representing low, medium and
high fracture frequency values. The other relies on many
(about 20) classes, most classes containing no more than three
or four data. This second class definition allows distinguishing
the few highest fracture frequency data, the latter belonging to
the same class (with other data) in the three-class definition.
With three classes, the geological component is dominated by
the seismic amplitude and the deep faults, with a positive
correlation, and by the gross-thickness and the distance to
faults with a negative correlation. With twenty classes, the
dominating drivers become the deep faults, the strain and the
dolomite content, with a positive correlation, and the distance
to faults with a negative correlation. Figure 8 shows the
corresponding correlation plots with the two first discriminant
components represented by the x and y-axes. The scatterplots
of the fracture frequency versus the first geological component
is given in Figure 9 for the two class definitions. Though the
two trends look very similar, one will notice that the 3-class
component better separates at least some of the very small and
high fracture frequencies, i.e. cases from the two extreme
classes. The reason can be found in the number of classes: it is
just easier to distinguish a few classes than many of them.
Sequential simulation: The drivers are known at grid-nodes
where the two previous geological components (with 3 and 20
classes) can be calculated. Each geological component is used
independently as secondary (soft) information to simulate the
NWSE fracture frequency with the sequential indicator
simulation approach presented previously. For each geological
component, 100 realizations are generated from which
probability maps can be derived. Figure 10B-C shows the
probability maps that the NWSE fracture frequency be greater
than or equal to 0.5 frac/m.
Relation Between Fracturing and Production
Data
The relation between fracturing as seen in the image log data
and from our modeling is qualitatively interpreted in this paper
and should be considered as a first analysis only.
Figure 11 shows the distribution of the productivity index
(PI) over the entire field and for both horizontal and vertical
wells. The following comments can be made: i) the PI
distribution is log normal which is indicative of wells drilled
in an heterogeneous reservoir; ii) although the best two
producers are horizontal wells, the performance of this type of
wells is not systematically better than vertical wells; iii) wells
located in the southern part of the field show better production
performances than those located in the central and northern
parts; iv) the three regions show the same type of PI
distribution.
It could be argued that this type of distribution is related to
the fractured nature of the field. However, the combined effect
of the fracturing and of a heterogeneous matrix behavior is
here most likely. In order to qualitatively assess the role of
fracturing, we calculated the ratios of the four central and
northern drains to the most fractured well S-H1 taken as a
reference for the following parameters: PI, average fracture
density for all fractures, average fracture density for the
dominant NW-SE set only, and wellbore length. Figure 12
shows the results of this exercise. It can be noticed that if the
southern reference well is two to seven time shorter than the
four other wells, it produces 3 to 15 times more. Two factors
can control this behavior: i) the matrix property with more
dolomite in the well SH-1; ii) the greater abundance of open
fractures and particularly of those striking NW-SE (the
absence of theses fractures have a strong influence on the PI
ratio). Since we showed that dolomite streaks tend to be more
fractured both factors may play a role.
Finally, the particular role of the seismic faults and/or
fracture corridors in conducting flow is still difficult to assess
at this stage. Indeed, a pilot waterflood strongly suggests a
high flow capacity channel in NW-SE direction that can be
interpreted as the effect of a fault and/or fracture swarm.
In order to see (i.e. to calibrate) how fracturing could be
related to well behaviour, fracturing probabilities of Figure 10
were extracted at all well locations of Figure 11 and
subsequently cross plotted with their respective PI values
(Figure 13). One can note that there is no clear relationship
between predicted fracturing and PI. All the wells with a
6 B. GAUTHIER, A.M. ZELLOU, A. TOUBLANC, M. GARCIA, J -M. DANIEL SPE 65118
relative PI larger than 2 tend to have a high probability to be
fractured at the chosen threshold. However, the converse is not
true. Looking at the geographical location of the wells, it
seems that fracturing plays a more important role for the
southern wells (and to a less extent for the northern wells)
whereas the central wells do not show any correlation with
fracturing. These results reinforce the idea that in this field, the
matrix has a key role.
Discussion
Beyond selecting a stochastic method for full field fracture
modeling, the use of geological drivers as indirect fracturing
information relies on two critical points or decisions:
The first one is the choice of the fracture index which
needs to be directly related to fracturing and which is not
influenced by the well orientation or length. Fracture
frequency is therefore adequate. However, obtained from cores
or image logs, it depends on the core recovery (poorly
recovered cored zones might be a sign of a highly fractured
zone) or on the quality of the image logs. Additionally,
fracture frequency may also be subjective to the interpreter.
The possibility of working with classes of fracturing can partly
solve this problem.
The second one is about data extrapolation (Figure 14). An
extrapolation model is required at grid-cells where one at least
of the drivers (methodology 1) or the geological component
(methodology 2) is outside the experimental range for which
fracturing data are available. Figure 15 shows the
extrapolation and interpolation zones for methodology 1.
These extrapolation zones represent up to 2/3 of the entire
field area. Although this problem is common to both
methodologies, the way to solve it is different. In methodology
1, the extrapolation is imposed by the neural network through
a complex non linear regression function that cannot be
controlled by the user. Part of the data only being used for
training in the simulation process, the extrapolation can be
partly evaluated by looking at the locations of the other data
(testing data) corresponding to extrapolation situations. If
some predicted values at these locations show large departures
from the actual fracturing data, the realization can be rejected.
In methodology 2, the extrapolation is controlled at the
bivariate statistical model level (e.g. in Figure 9). It is the
users decision how to extrapolate the bivariate model beyond
the experimental range of the geological data. This decision
may be based on observable trends and on the geologist
judgement. In all cases, it is clearly established and can easily
be changed if deemed inappropriate against new data.
The probability maps (Figure 10) obtained with the two
methodologies calls for some comments.
First of all, the maps look very similar in terms of trends
and show similar high and low probability areas. The
significant differences from one map to another are in regions
where model extrapolation situations occur (Figure 15). The
northern part of the field is the most illustrative: methodology
1 and methodology 2 with the 20-class geological component
both give, in some regions, a high probability of finding
NWSE fracture frequencies greater than 0.5 frac/m. This is not
apparent with the 3-class geological component. The latter
relies on a class definition that gathers all fracture frequencies
greater than 0.5 frac/m into the same class without any
distinction between them. The northern high probability
regions can therefore be seen as influenced (in extrapolation
conditions) by the high fracture frequency data that the non-
linear regression function of methodology 1 and the 20-class
discriminant analysis function of methodology 2 try to fit.
These data being located in the southern part of the field, their
influence in the North can be questioned.
Looking at the probability values, methodology 1 tends
show higher discontinuity and sharper variation (i.e. more
probabilities very close to either 0 or 1). This means that
methodology 1 introduces less uncertainty about fracturing
that methodology 2, at least for a threshold of 0.5 frac/m. With
methodology 2, the uncertainty is governed by the input
statistical models that can be inferred from the data. With
methodology 1, the uncertainty is controlled by playing
(subjectively) with two correlation coefficients that determine
which realizations are accepted or rejected from training and
testing data sub-sets. Decreasing one or both correlation
coefficients allows accepting a wider variety of realizations
but also leads to higher departures between predicted and
actual values. Conversely, increasing them leads to over
constrained models with much less variability between
realizations.
Looking at the exploitation of results, if the aim is to
characterize large scale fracturing trends from several
(explicative) geological drivers, methodology 1 is easy to use
and should be efficient. This requires, however, that enough
data are available to allow deriving reliable non-linear
multivariate regression models and that the available drivers
are enough to explain all kinds of fracturing factors. The latter
point is certainly the most critical. Uncertainty about
fracturing trends can also be addressed provided the models
are not too much constrained as discussed above. If all the
previous conditions are not met or if the aim is to produce
(conditional) fracture frequency fields to be used for reservoir
simulation or for other purposes, methodology 2 is more
appropriate though more demanding in man-efforts for data
exploratory and analysis. Especially, the realizations generated
by sequential simulation can be seen as realistic in the sense
they honor the fracturing data at their locations and reproduce
different statistical models representative of the actual spatial
distribution of fracturing as inferred from the data or deemed
acceptable from a geological standpoint.
The application of our methodology to our case study showed
that although this field cannot be considered strictly as a
fractured reservoir, fracturing could play a positive role in the
southern part and to a less degree in the northern part. We
demonstrated that fracturing is mainly controlled by lithology
(dolomitic streaks) and by faulting. These two parameters
being more present in the south, they both contribute to the
better productivity of the southern wells.
SPE 65118 INTEGRATED FRACTURED RESERVOIR CHARACTERISATION: A CASE STUDY IN A NORTH AFRICA FIELD 7
Conclusions
The following conclusions are derived from the results
presented in this paper:
The use of fuzzy logic or discriminant analysis allows to
define the drivers controlling fracturing;
Using these drivers, the fracture frequency is modeled
over the entire field.
The maps derived from this analysis can be used in infill
drilling and to improve the success of horizontal wells;
These maps can be used to control the local discrete
fracture networks (DFN) involved in well test
simulations
3-8
at any location.
Acknowledgments
We wish to thank TotalFinaElf and Subsidiaries for
authorizing the publication of this paper. The help and support
of Ahmed Ouenes with (RC)
2
for the use of the ResFrac
TM
software were appreciated. Providing this study with key
facies interpretations in horizontal wells, Olivier Lerat is
sincerely thanked. Comments by Xavier Britsch, Frederic
Duval, Caterine J avaux and J ean Pouzet, improved earlier
version of this manuscript.
References
1. Sibbit, A.M. : "Quantifying porosity and estimating permeability
from well logs in fractured basement reservoirs," paper SPE
30157 presented at the 1995 SPE PetroVietnam, Ho Chi Min,
March 1-3.
2. Barton, C.A., Moos,D. Peska, P. and Zoback, M.D. : "Utilizing
wellbore image data to determine the complete stress tensor:
application to permeability anisotropy and wellbore stability,"
The log analyst, Nov.-Dec. 1997, p.21-33.
3. Swaby,P.A. and Rawnsley, K.D.: An interactive 3D fracture
modelling environment, paper SPE 36004. presented at the
1996 SPE PetroleumComputer Conference, Dallas, J une 2-5.
4. Cacas, M. C., Letouzey, J . & Daniel, J . M.: Object-Oriented
Modelling of J ointed Oil Reservoirs, Paper 4-03 presented at the
1998 60th Eage Conf, Leipzig,
5. Wei, L., Hadwin, J ., Chaput, E., Rawnsley, K. and Swaby, P.:
Discriminating fracture patterns in fractured reservoirs by
pressure transient tests, paper SPE 49233 presented at the 1998
SPE Annual Technical Conference and Exhibition, New
Orleans.
6. Ding, Y. and Chaput, E.: Simulation of hydraulically fractured
wells paper 007 presented at the 1997 9th European
Symposiumon Improved Oil Recovery, The Hague, Oct. 20-22.
7. Bourbiaux, B., Cacas, M. C., Sarda, S. & Sabathier, J . C.:A fast
and efficient methodology to convert fractured reservoir images
into a dual-porosity model, paper SPE 38907, presented at the
1997 SPE Annual Technical Conference and Exhibition, San
Antonio, Oct. 5-8.
8. Guerreiro, L. Costa Silva, A., Alcobia, V. and Soares, A,
Integrated reservoir characterisation of a fractured carbonate
reservoir, paper SPE 58995, presented at the 2000 SPE
PetroleumComputer Conference, Villahermosa, Feb. 1-3.
9. Harris, J ., Taylor, G., Walper, J .: Relation of deformational
structures in sedimentary rocks to regional and local structure,
AAPG Bulletin, (1960) v. 44 p.1853-1873.
10. Lisle J . L.: Detection of zones of abnormal strains in structures
using Gaussian curvature analysis, AAPG Bulletin, (Dec.
1994), v. 78, no. 12, p.1811-1819.
11. McQuillan H.: "Small scale fracture density in Asmari formation
southwest Iran and its relation to bed thickness and structural
setting," AAPG Bulletin, (1973) v. 57, no. 12, p.2367-2385.
12. Murray, G.: Quantitative fracture study- Sanish Pool, McKenzie
County, North Dakota, AAPG Bulletin, (J an. 1968), v. 52, no.
1, p.57-65.
13. Ericsson, J .B., McKean H.C. and Hooper R.J .: Facies and
Curvature Controlled 3D Fracture Models", Geological Society
of London Special Publication (1998) vol. 147 p.299-312.
14. Argawal, B. Allen, L.R. and Farrell, H.E.: Ekofisk Field
reservoir characterization: mapping permeability through facies
and fracture intensity, paper SPE 35527, presented in the 16-17
April 1996 NPF/SPE European 3D.
15. Zellou, A.M, Ouenes, A. and Banik, A.K: "Improved fractured
reservoir characterization using neural network, geomechanics
and 3-D seismic," paper SPE 30722 presented at the 1995 SPE
Annual Technical Conference and Exhibition, Dallas, Oct. 22-
25.
16. Ouenes, A., Richardson, S., and Weiss,W.W.: Fractured
Reservoir Characterization and Performance Forecasting Using
Geomechanics and Artificial Intelligence," paper SPE 30572
presented at the 1995 SPE Annual Technical Conference and
Exhibition, Dallas, Oct. 22-25..
17. Ouenes, A., Zellou, A.M., Basinski, P.M. and Head, C.F.:
Pratical use of neural networks in tight gas fractured reservoirs:
application to the San Juan Basin, paper SPE 39965 presented
at the 1998 Rocky Mountain Regional Low Permeability
Reservoirs Symposium, Denver, CO, Aug..
18. Gauthier, B.D.M, Franssen, R.C.W.M & Drei, S.: Fracture
networks in Rotliegend gas reservoirs of the Dutch offshore:
implications for reservoir behaviour," Geologie en Mijnbow
(2000) 79 (1) p.45-57.
19. Ouenes, A.: Practical application of fuzzy logic and neural
networks to fractured reservoir characterization, Computer and
Geosciences, Sahab Mohageg (Ed.), (2000) v. 26, no 7.
20. Gooverts P., Geostatistics for natural resources evaluation,
Oxford University Press (1997).
21. Zhu H. and J ournel A.G. "Formatting and Integrating Soft Data:
Stochastic Imaging via the Markov-Bayes Algorithm," Geostat
Troia (1992), ed. Soares, Kluwer Publ., p.1-12.
22. Schmaryan L.E. and J ournel A.G.: Two Markov models and their
applications, Math Geology, (1999) vol 31, N8, p.965-988.
23. Auzias, V., Rives, T., Rawnsley, K.D. and Petit, J .P. et al:
Fracture orientation modelling in the vicinity of a horizontal
well, Bull. Centre Rech. Elf Explor. Prod., (1997) 21, p.381-
397.
24. Townsend, C., Firth, R., Westerman, R., Kirkevollen, L., Harde,
M. & Andersen, T.: Small seismic-scale fault identification and
mapping. In: J ones, G., Fisher, Q. J . & Knipe, R. J . (Eds),
Faulting, Fault sealing and Fluid flow in Hydrocarbon
Reservoirs , Geological Society, Special Publications , p.1-25
(1998).
25. Nelson, R.: Geologic Analysis of Naturally Fractured Reservoirs,
Gulf Publishing Company, Houston, TX (1985) 320.
26. http:\\www.rc2.com\products\ResFrac.html
8 B. GAUTHIER, A.M. ZELLOU, A. TOUBLANC, M. GARCIA, J -M. DANIEL SPE 65118
G
e
o
l
o
g
i
c
a
l
C
o
n
t
r
o
l
Fracture Drivers
Fractures
Well Performance
(EUR, PI, Kh)
Calibration
Figure 1: Fractured Reservoir Modeling Concept. In this
paper, we focus on the geological modeling of fractures.
-1200 -1100 -1000 -900 -800 -700
(m)
(
m
)
Faults at Deep level
Faults at Top reservoir
Figure 2: Structural map at top reservoir with deep faults
overlaid.
500 3000 5500 8000
2
0
0
0
4
5
0
0
7
0
0
0
9
5
0
0
1
2
0
0
0
1
4
5
0
0
1
7
0
0
0
1
9
5
0
0
(
m
)
High low
N
Figure 3 Strain map. This map is one of the fracture driver
used in the modeling. See details in the text.
N
N=966 Sch. Low.
Figure 4: Stereoplot and strike histogram of raw open fracture
orientation interpreted in image log data.
SPE 65118 INTEGRATED FRACTURED RESERVOIR CHARACTERISATION: A CASE STUDY IN A NORTH AFRICA FIELD 9
N-V1
(Vertical)
S-H1
(E-W)
N
N
N-H1
(NW-SE & NE-SW)
C-V1
(Vertical)
N
C-H1
(NW-SE & NE-SW)
C-H2
(NW-SE & NE-SW)
N
C-H3
(NW-SE)
0
50
100
0 100 200 300 400 500 600 700 800 900
D
e
p
t
h

(
f
t
)
Fract. Freq.
Cluster Zone
Offset (m)
A
B
Figure 5: A) Fracture orientation in each studied well. The
rose diagrams are normalized according to relative fracture
density in each well. Well orientation is given in brackets. B)
example of along hole fracture frequency distribution
Figure 6: Possible interpretations of fracture density peaks in
horizontal wells.
All Fractures NWSE Fractures
S
i
g
n
i
f
i
c
a
n
t

D
r
i
v
e
r
s
I
n
s
i
g
n
i
f
i
c
a
n
t
D
r
i
v
e
r
s
Dolomite Content
Calcite Content
Strain
Structure Max Curv.
Seis. Amplitude
Deep Struc. Fault
Thickness
Dist. To Fault
Illite Content
Seis. Coherence
Porosity
Dolomite Content
Calcite Content
Strain
Illite Content
Porosity
Deep Struc. Fault
Seis. Amplitude
Dist. To Fault
Seis. Coherence
Thickness
Structure Max Curv.
Figure 7: Methodology One: Ranking of the drivers. Here the
case of all the fractures combined is also considered. It can be
noted that both rankings are consistent.
Deep Faults
GrossThick
Dolomite
Strain
Dist2Faults
Amplitude
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
First Discriminant Component
S
e
c
o
n
d

D
i
s
c
r
i
m
i
n
a
n
t

C
o
m
p
o
n
e
n
t
3 Class Component

Deep Fault
GrossThick
Dolomite
Strain
Dist2Faults
Amplitude
-1
-0.75
-0.5
-0.25
0
0.25
0.5
0.75
1
-1 -0.75 -0.5 -0.25 0 0.25 0.5 0.75 1
First Discriminant Component
S
e
c
o
n
d

D
i
s
c
r
i
m
i
n
a
n
t

C
o
m
p
o
n
e
n
t
20 Class Component
Figure 8: Correlation plots corresponding the two first
discriminant components for different class definitions.
10 B. GAUTHIER, A.M. ZELLOU, A. TOUBLANC, M. GARCIA, J -M. DANIEL SPE 65118
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
-2 -1 0 1 2 3
Geol ogical component (dimensi onl ess)
N
W
S
E

F
r
a
c
t
u
r
e

F
r
e
q
.

(
f
r
a
c
.
/
m
)
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
-2 -1 0 1 2 3 4
Geol ogical component (dimensi onl ess)
N
W
S
E

F
r
a
c
t
u
r
e

F
r
e
q
.

(
f
r
a
c
.
/
m
)
3 Class Component
20 Class Component
Figure 9 Scatterplots of the NWSE fracture frequency vs. the
geological component for different class definitions
Method 1
A
Method 2
20 class
Method 2
3 class
B C
Figure 10: Probability map of having fracture frequency
larger than 0.5 frac/m for method 1 (A) and for 20 and 3
fracturing classes for method 2 (B and C respectively). See
discussion in the text.
SPE 65118 INTEGRATED FRACTURED RESERVOIR CHARACTERISATION: A CASE STUDY IN A NORTH AFRICA FIELD 11
Wells
R
e
l
a
t
i
v
e

P
I
S-H1
C-H2
C-H1
N-V1
N-H1
C-H3
South
Central
North
Horizontal
Vertical
Well with imagelog data
Figure 11: Relative PI distribution in the studied field for all
tested wells.
0
0.2
0.4
0.6
0.8
1
S-H1 C-H2 C-H1 C-H3 N-H1
6 6
1 1
2 2
2 2
7 7
PI
All fracture density
NW-SE fracture density
1 1 Wellbore length
Figure 12: Relation between the SH-1 well taken as a
reference well and the other horizontal wells with image log
data for various parameters. See details in the text.
A) Method 1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 1 2 3 4 5 6
Relative PI
P
r
o
b
(F
ra
c
>
0
.5
frc
/m
)
Central
North
South
B) Method 2
3 Fracturing Classes
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 1 2 3 4 5 6
Relative PI
P
r
o
b
(F
r
a
c
>
0
.5
fr
c
/m
)
Central
North
South
C ) Method 2
20 Fracturing Classes
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 1 2 3 4 5 6
Relative PI
P
ro
b
(F
r
a
c
>
0
.5
fr
c
/m
)
Central
North
South
Figure 13: Relation between relative PI and predicted
fracturing in wells without image log data.
F
r
a
c
t
u
r
e

I
n
d
e
x
Driver amplitude
FieldMapRange
Well
Control
data range
E
x
t
r
a
p
o
l
a
t
i
o
n
Interpolation
E
x
t
r
a
p
o
l
a
t
i
o
n
Figure 14: The extrapolation issue: definition
E
x
t
r
a
p
o
l
a
t
i
o
n
I
n
t
e
r
p
o
l
a
t
i
o
n
Figure 15: Map showing the area where at least one out of the
4 main drivers (dolomite, strain, max. curv. & seism.
Amplitude) is in extrapolation.