SPE 84870
Application of Artificial Intelligence to Characterize Naturally Fractured Reservoirs
A. Kouider El Ouahed, Sonatrach Inc. Djebbar Tiab, Oklahoma University. Amine Mazouzi, Sonatrach Inc., Sarfraz A.
Jokhio, University of Oklahoma
Copyright 2003, Society of Petroleum Engineers Inc.
This paper was prepared for presentation at the SPE International Improved Oil Recovery
Conference in Asia Pacific held in Kuala Lumpur, Malaysia, 20–21 October 2003.
This paper was selected for presentation by an SPE Program Committee 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 Petroleum Engineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the Society of Petroleum Engineers 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 whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.
Abstract
In highly heterogeneous reservoirs classical
characterization methods often fail to detect the location and
orientation of the fractures. Recent applications of Artificial
Intelligence to the area of reservoir characterization have made
this challenge a possible practice. Such practices consist of
seeking the complex relationship between the fracture index
and some geological and geomechanical drivers (Facies,
porosity, permeability, bed thickness, proximity to faults,
slopes and curvatures of the structure) in order to obtain a
fracture intensity map using Fuzzy Logic and Neural Network.
This paper shows the successful application of Artificial
Intelligence tools such as Artificial Neural Network and Fuzzy
Logic to characterize naturally fractured reservoirs. A 2D
fracture intensity map and fracture network map in a large
block from Hassi Messaoud field have been developed using
Artificial Neural Network and Fuzzy Logic.
This was achieved by first building the geological model
of the permeability, porosity and shale volume using
stochastic conditional simulation. Then by applying some
geomechanical concepts first and second structure directional
derivatives, distance to the nearest fault, and bed thickness
were calculated throughout the entire area of interest. Two
methods were then used to select the appropriate fracture
intensity index. In first method well performance was used as
a fracture index. In the second method, which consists of a
new proposed approach, a Fuzzy Inference System (FIS) was
built. With such system static to dynamic data was coupled to
reduce the uncertainty and resulted in a more reliable Fracture
Index. The different geological and geomechanical drivers
were ranked with the corresponding fracture index for both
methods using a Fuzzy Ranking algorithm. Only much
important data were selected to be mapped with the
appropriate fracture index using a feed forward Back
Propagation Neural Network (BPNN). The neural network
was then used to obtain a fracture intensity maps throughout
the entire area of interest. A mathematical model based on
“the weighting method” was then applied to obtain fracture
network maps, which resulted in a deep insight about the
major fracture trends.
The obtained maps were compared in the end and the
results show that the proposed approach is a feasible
methodology to map the fracture network.
Introduction
Naturally fractured reservoirs represent a significant
percentage of oil reservoirs throughout the world and have
been given considerable research attention. As of our
understanding, such reservoirs behave like two media of
different porosity: the matrix and the fracture network. The
productivity of the wells in these low permeability reservoirs
is attributed to fracture network. Naturally fractured reservoirs
were initially homogeneous and became fractured under
certain circumstance of rock deformation and/or physical
diagenesis. A thorough description of fracture connectivity,
orientation and location is a key point to fractured
reservoir characterization.
Researchers have attempted to model natural fractures with
idealized approaches such as the sugar cube and the slab
models13. Classical geostatistics methods often fail to predict
the distribution of the fractures. Recent trends in reservoir
characterization tend to use more sophisticated tools such as
artificial intelligence. A key element of using such techniques
is that reservoir characterization is more complex to model
with traditional methods. Artificial intelligence has the ability
of combining several tools from several sources thereby,
reducing the uncertainty. The most common known artificial
intelligence techniques are Artificial Neural Network (ANN)
and Fuzzy Logic.
Neural Networks are systems constructed to use some
organizational principles resembling those of the human brain.
They are composed of simple elements operating in parallel
employing a set of linear and non-linear activation functions
that do not require a prior selection of a mathematical model.
A network can be trained to perform a particular function by
adjusting the values of the connection’s weights between the
elements.3 Usually neural networks are adjusted or trained so
that a particular input leads to a specific target output.
Fuzzy logic uses the benefit of fuzzy reasoning. Under this
type of reasoning, decisions are made on the basis of fuzzy
linguistic variables such as low, good and high, with fuzzy set
operators such as ‘AND’ and ‘OR’. This process simulates
2 SPE 84870
the human expert reasoning process much more realistically
than do the conventional systems15. Fuzzy set theory is an
efficient tool for modeling the kind of uncertainty associated
with vagueness, imprecision and/or a lack of information
regarding a particular element of the problem at the hand.
This study aims to apply the above discussed techniques to
characterize naturally fractured reservoirs and Hassi Messaoud
field Algeria is used as a case study.
Literature Review
The application of Artificial Intelligence tools such as
Fuzzy Logic and Neural Network is evolving as an oilfield
technology. During the last few years several studies have
been conducted in the field of petroleum engineering by
applying artificial intelligence. The major applications are
seismic data processing and interpretation, well logging,
reservoir mapping and engineering.
Recent applications of artificial intelligence tend to focus
on reservoir characterization. Big efforts have been devoted to
the area of naturally fractured reservoirs. Balch14 used the well
data set consisting of traditional logs and tuned-to-core
estimates of secondary porosity from FMI image to train an
artificial neural system to predict the secondary porosity.
Neural network has also been used to provide porosity and
permeability estimates across a study area by training on
seismic amplitude information. Sadiq and Nashawi12 presented
a neural network model to predict fracture gradient pressure;
the results indicated that neural network is not only feasible
but yields quite accurate results.
Ouenes et. al6,7,8,9,10,11 introduced a collection of artificial
Intelligence tools to model the interwell region of fractured
reservoirs. Assuming that fractures can be represented by well
performance at grid block scale, Ouenes used AI tools to map
between different geological and geomechanical drivers and
the well performance. Eduardo13 proposed a Fuzzy algorithm
for fracture detection using conventional well logs and drilling
information. He also suggested the use of Fuzzy Curvature
Index for application to naturally fractured reservoirs.
Ouenes and Lee9 presented an integrated fractured
reservoir model using both discrete and continuum approach.
Gauthier19 applied successfully such Integrated Fractured
Reservoir Characterization to a case study in North Africa
Hassi Messaoud Field description
Hassi Messaoud (HMD) field is situated in the
Algerian part of Sahara desert and was discovered in 1956.
With its reserves (OOIP) of several billion m3 in place, the
HMD field is considered a giant one. Mainly solution-gas
driven, the productive formation is Cambrian sandstones
3400m deep with an average thickness of 300m and 3
productive layers (R2, Ra, and Ri from bottom to top).
The R2 and Ri layers are characterized by poor
petrophysical properties. The important layer is Ra and is
characterized as heterogeneous with an average porosity of 8%
and a permeability of 5 mD. The large extension of the field
and the presence of fault barriers and the facies lateral
variation have confirmed the nonuniform depletion throughout
the field, which has led to its zonation, Figure 1.
Zone of Interest
The focus of this study was the area limited by the
rectangle (x=820000 y=130000, x=828000, y=140000), in the
northern part of Hassi Mesaoud field (Figure1). Basically this
area includes the northern part of zone 13, the southern part of
zone 8 and the interzone in between them. Some parts of zones
6, 9 and 10 were also taken into account (Figure 2)
The interpretation of recent image logs recorded from
some vertical and horizontal wells, core analysis and pressure
buildup tests indicate that this area is naturally fractured.The
zone of interest; especially the interzone mentioned above; is
now an object of a development plan by drilling new
horizontal wells. The study presented here aims to obtain an
image of the most probable zones to be naturally fractured.
The first step consists of building a stochastic model of
permeability, porosity and shale volume (Vsh).
Data Collection
Geologically this area is located on the top of the reservoir.
Basically, it includes three layers R2, ID, and D1. The ID was
eroded and was not observed in several wells, for such reason
geologists treat the ID and D1 as a single layer Dh.
Data were collected from the two layers together. Core
permeabilities and porosities were collected from 36 vertical
wells with sampling frequency of four cores per meter.
Gamma ray data were taken from 20 vertical wells and two
horizontal wells lying in the interzone 8-13 (OMPZ333 and
OMPZ34).
We also gathered CMR results from these two horizontal
wells in terms of permeability and porosity. Theses two
horizontal wells helped in constructing horizontal variograms
and selecting the appropriate model type to be used later in the
geostatistical software.
Statistical Analysis
The statistical analysis carried out in the present
study showed a great heterogeneity in reservoir properties in
terms of permeability and shale volume. The coefficient of
variation, Cv, for the case of permeability varies between 1.2
and 3 and the average mean value varies from few millidarcies
up to more than 100mD. Vsh exhibit less heterogeneity with Cv
range between 0.4 to 0.8 whereas porosity does not change too
much with Cv between 0.22 and 0.55 and a mean value of
about 8%.
Establishing PDF’s and CDF’s
Using 7090 samples of permeability and porosity and
8000 sample of Vsh, PDF’s and CDF’s were built for each of
the three parameters, (Figure 3, 4 and 5). Examining the three
figures, we can notice that:
• Vsh, calculated from GR log, shows a negatively skewed
distribution where most of data lie to the low end of the
data range.
• Core porosities follow very well a normal distribution. In
fact after taking a cut off of 25 values from the amount of
7090 available data, all the remaining 7065 lie on the range
of 0 and 0.20 giving a nice Gaussian shaped distribution
• Permeability, however, follows a log normal distribution.
The corresponding lognormal histogram shows basically
SPE 84870
two clusters with almost the same percentage suggesting
the presence of two groups of data. These two clusters
correspond to the layers R2 and Dh respectively.
Areal and Vertical Variability (Variograms Construction
and Analysis)
While parameters of vertical semivariograms are
usually available from the well data, the lateral parameters
must be established from the normally sparse well data. It is
important that uncertainties in these parameters be
acknowledged in the modeling schemes.
Areal Variability
The purpose of presenting areal semivariograms driven
from horizontal well data in this study is to diagnose the
lateral variability and sketch out the appropriate model type
(i.e., Spherical, Gaussian or Exponential).This model will be
used later on to fit the sparse well data that characterize the
lateral variability in four different directions.
Areal variability is available from horizontal well logs
interpretation in the northwest to southeast direction.
Horizontal semivariograms were calculated using
permeabilities and porosities obtained from the results of
CMR interpretation and Vsh obtained from the GR logs.
Figures 6-8 show horizontal semivariograms of permeability,
porosity, and shale content respectively. They all illustrate a
combination of exponential and spherical variability. They
also reveal some periodicities represented by a succession of
structures of high and low property value, or proportional
build up related to sequences of about five meters.
These variograms cannot be directly used to stochastically
model the reservoir properties because of two reasons: (1) they
only show a variability in one direction and (2) they only
represent a length of 400-800 meter whereas the area to be
modeled is 8 Km ×10 Km. However, with some geological
knowledge and assumptions about the depositional
environment, they can be used if a reservoir simulation history
matching process can be conducted side by side with the
stochastic model
Stochastic Modeling
The aim of the present section is to build a 2D model
which will be used as a basis for inputs of the Neural Network
modeling of fractures. To get better results, however, we used
a 3D grid system model of 80 × 100 × 30 cells. The results
were then upscaled to obtain a 2D model and exported to an
ASCII file which then was used to redraw the corresponding
maps with the help of “SURFER 7”.
This work has been conducted with the help of a
geostatistical software called GRID STAT. It is based on the
conditional simulation concept. The conditioned simulation
honors the following points.
• Reproducing CDF’s and PDF’s functions,
• Matching data at well locations, and
• Matching the varigrams models.
Results and Discussion
Figures 9, 10, and 11 show a two dimensional
permeability, porosity and Vsh distribution from the Hassi
3
Mesaoud sandstone reservoir generated using conditional
simulation and spherical areal variograms. The reddish color
indicates good characteristics (high k, high φ and low Vsh),
whereas the bluish color indicates poor characteristics (low k,
low φ or high Vsh). All of the three maps represent the
distribution of the properties in a succession of good sand
bodies distributed among a poor sandstone medium, into
which are the dispersed discontinuous siltstones which have
low porosity and permeability. They also show the distinction
of the inter zone13-8 of low petrophysical characteristics
which confirm the actual zonation in this area.
Geomechanical drivers
In the previous section, the geological model was
built for the reservoir properties of permeability, porosity and
shale volume. These parameters will serve as a set of inputs
for the upcoming neural network. In this section we will deal
with the calculation and analysis of structure shape (slopes and
curvatures), distance to faults and, bed thickness (Figure 13).
Slopes and Curvatures
Stress in rocks depends on the curvature. As such,
the shape of the formation structure top can have a significant
role on fracture development. Curvature analysis is inherent
from Maurry’s model that relates fracture porosity and
permeability to structural curvatures.
∂ 2z  ∂2z 
2
φ = C1 T , K = C 2  T 2 
∂x 2  ∂x 
where φ is the porosity, K is the permeability, T is the
∂2z
thickness is the curvature and C1 and C2 are constants
∂x 2
Maurry’s model includes unrealistic assumption that fractures
occur only in the direction perpendicular to the radius of the
folding. It fails, for example, to recognize stress as a tensor
property13. Ouenes6~11used Maurry’s observations about
curvature and the presence of fractures in folded structure to
calculate structural first and second derivatives and extend it
to four directions. These first and second derivatives were then
used as inputs for his Neural Network model.
Structure Top
Evaluating reservoir structure may seem an easy task.
However, the fractured reservoir description approach requires
a detailed and accurate structure. There are two possibilities
for obtaining a meaningful structure: 3D seismic and
interpolation between well data. 3D seismic, unfortunately, is
not available for the study area. Thus structure map based on
well data and the old 2D seismic data is used. In such case,
uncertainties must be acknowledged in zones where wells are
very scattered.
First and second derivatives were discretely calculated in
each grid node throughout the entire gridbloks without
breaking discontinuities (faults). In other words if the line,
along which derivatives are to be calculated, crosses the fault
lines at one or many points, the calculation process should be
achieved in several segments separated by the
discontinuity lines.
4 SPE 84870
Distance to Faults
Distance to faults is a crucial parameter to be taken
into account when looking to natural fracture indicators from a
geomechanical point of view. It is well known in
geomechanics that under certain circumstances of stress
distribution, fractures develop perpendicularly to the major
faults in a decreasing frequency and chance of occurrence
away from the faults. Therefore, introducing this distance as
an input for the upcoming Neural Network may reveal some
additional information about fractures and thus a
faster training.
Bed Thickness
Reservoir bed thickness is intimately related to fracturing.
Depending on the bed thickness and lithology that exist at
each point, as well as the stress magnitude and directions, the
rock will either fracture or resist by rearranging itself9. In
general thick reservoirs are more likely to be fractured than
thin ones.
As we are interested in developing a 2D model, bed
thickness was obtained for both layers Dh and R2 as one
single layer. The problem of obtaining accurate bed thickness
map is the same as for structure map. As mentioned before,
3D seismic data were not available, and thus structure tops and
bed thickness were obtained simply from the old 2D seismic
and well log data interpolation.
Results
As summary, the following drivers have been calculated
and/or mapped throughout the zone of interest using some
Matlab programs:
• First directional derivatives (slopes): SN, EW, SWNE,
NWSE
• Second directional derivatives (curvatures): SN, EW,
SWNE, NWSE.
• Distance to the nearest fault.
• Bed thickness.
The Fracture Index, Formulation of the Problem
Fracture index is the variable that will be mapped with the
previously prepared drivers and indicators through a Neural
Network. Fracture index can be obtained either from core or
log data or, recently, from the image log. When using core
data or the imagery log one may account for the number or
frequency of fractures as a fracture index. If classical
resistivity log is used the following formula can estimate the
fracture index5.
1 1
− High frequency of high permeability obtained from
R R fo
FII = fi
1 1
−
R mf R w
The above indicators have a common weakness, they are
only available within a very small area. The question that can
arise is that: do really these very small discrete cracks reflect
the image of the reservoir? To avoid such upscaling problem,
Ouenes et al.6~11 looked to those natural fissures from a
dynamic point of view. They used the well performance as
fracture index. According to their proposed model either early
productivity index, the maximum rate, the cumulative
production or the ultimate recovery can be used. Early
productivity index, however, has been proven to have good
results and has attracted the attention of many authors.
Ouenes et al.6~11 started from the assumption that well
performance is directly (linearly) related to natural fractures.
This assumption, however, doesn’t always hold. In fact,
naturally fractured reservoirs can have a high productivity
index. High productivity, however, doesn’t necessarily mean
the presence of NFR. In terms of well performance, thick
reservoirs with no presence of fractures may behave as thin
fractured formation. Several parameters can lead to a high well
performance, from the small effect of low viscosity or
negligible capillary pressure to the enormous effect of thick
large fractured reservoir under a strong active aquifer.
No single tool can bring full information about fractures.
Several indicators from different sources (well performance,
well test, core analysis and well logging), if combined
together, may lead to a reliable fracture index estimation.
Current section is an attempt to make such combination. By
building a Fuzzy Inference System (FIS), we could couple
static to dynamic parameters to obtain a Fuzzy Logic based on
Fracture Index. This index gathers the following indicators
and leaves doors open for others to be included.
• Frequency of samples of high permeability.
• Fractured intervals obtained from core description.
• Class indicator describing the goodness of those open
fractures observed from core description. This indicator
has also been obtained via a small Fuzzy
Inference System.
• Productivity index normalized to net thickness.
• Skin value obtained from well test analysis.
Both techniques, Ouenes et al.6~11 model and the current
proposed approach, will be used for the Neural Network
Modeling. Discussion and comparison will be made at the end
Well Performance
Early productivity index and skin values were
collected from 46 vertical wells almost all wells have a
negative skin. The early PI will be used as fracture index in
the case of Ouenes model use. For our approach, however, it
should be normalized to the net thickness in order to omit the
effect of h from the productivity. Horizontal well performance
for (OMPZ 34, OMPZ111, and OMPZ 331) are illustrated in
Table 1, we notice that these four wells has not been used,
they were left just as eyewitnesses.
Permeability
core data might be an indictor for the presence of fractures. In
fact, if there is presence of open and well connected fractures,
we expect to have several samples of high permeability.
Core Description
We started from the archived old core description reports.
Such reports contain a detailed description. In fact each core is
drawn as it was observed. Each fractured length was assigned
a rectangle equal in length to the fractured interval itself. If the
fractures are open, the rectangle is hachured. If they are close,
SPE 84870
the rectangle is filled with black color. Four classes have been
distinguished A, B, C and D, where class A presents the best
characteristics and class D presents the poorest ones.
This study covered 24 wells where sufficient data were
available. The descriptive information has been converted into
digital data presented in Table 12.4. It was noted that class D
was absent almost in all the wells. For such reason, those very
rare fractures of class D were added to class C. Open fractures
represent 10 to 96 % among the whole fracture system with an
average of 65%.
Class Indicator
The class indicator (CI) introduced here consist of
describing how good the fractures are, which means it is
dependent on the dominant class: high for Class A, Medium
for B and Low for C. It might be obtained by simply applying
a weighting method. CI = ∑
i
λ i C i . Where λi is the weight
assigned to the Class Ci. However, this raises the question of
what value should be assigned to each weight?. To avoid such
question, we have found that this problem could be a simple
application of a simple Fuzzy Inference System (FIS). This
FIS is based on the following three rules:
if Class is Class A then, CI is High
if Class is Class B then, CI is Medium
if Class is Class C then, CI is Low
Such FIS has been obtained with the help of a Matlab Tool
Box for Fuzzy Logic4.
Fuzzy Logic Based Fracture Index
After obtaining all the variables which will contribute in
the fracture index, the Fuzzy Inference System illustrated in
Figure 16 has been built. It is based on the following 8
if-then rules.
1. if FL is high and CI is high and High_k is high then
FI is very high
2. if FL is high or High_k is high then FI is high.
3. if FL is low or High_k is low then FI is low.
4. if FL is low and CI is low and High_k is low then FI
is very low.
5. if NPI is high or skin is low then FI is high
6. if NPI is low or skin is high then FI is low.
7. if NPI is high and skin is low then FI is very high.
8. if NPI is low and skin is high then FI is very low.
Where:
FL: is the fraction of the open fractured interval obtained
from core data.
CI: is the class indicator,
High_k: is the frequency of samples of high
permeabilities
Skin: is the skin value obtained from well test analysis,
NPI: is the normalized productivity index
100 × PI and,
NPI =
net. thick
FI: is the fracture Index.
We note that all these fuzzy rules have been assigned the
same weight.
5
Input Fuzzification
This fuzzy inference system has been built on 8 rules, each
rule depends on resolving the inputs into a number of different
fuzzy linguistic sets: skin is low, fractured length is high, class
indicator is low and so on. Before the rules can be evaluated
the inputs must be fuzzified according to each of the linguistic
sets. As an example, the Figure 15 shows that to what extent
the skin is low. In that manner all the inputs are fuzzified and
overall the qualifying membership functions required by
the rules.
Fuzzy Operators
Figure 14 is an example of the OR operator. Here we
have evaluated rule number two for the fracture index
calculation. The different pieces of the antecedent (‘FL is
high’ and ‘High k is high’) yielded the fuzzy membership
values 0.3 and 0.4, the fuzzy rule simply calculates:
probor (a , b) = a + b − a × b .
probor (0.3,0.4) = 0.3 + 0.4 − 0.3 × 0.4 = 0.58
Implication Method
The implication function modifies the fuzzy set to the
degree specified by the antecedent. For the current application
minimum (min) implication method has been chosen. This
method truncates the output fuzzy set. Figure 14 shows the
min method applied to the previous fuzzy rule
Output Aggregation
There are several aggregation methods among them
we can name: max method (maximum), probar (probabilistic
or) and sum (simply the sum). We chose simple sum method
for this study. The right lower corner of Figure 16 shows the
aggregation results of the 8 implemented rules.
Deffuzification
A centroid diffuzification method which calculates
the center of area under the curve has been used. The result
deffuzified number is the fracture index. Table 3 shows the
deffuzified fracture index for 24 wells. This index varies from
a lowest value of 0.06 at well #22 to a highest value of 0.78
obtained for the case of #16.
Drivers Ranking
The ranking aims to analyze the effect of any driver on the
chosen fracture index. The benefit from this is to better
understand what primary drivers are and to eliminate, if
needed, those with low effect or contribution. There are a
number of methods that can be used to rank a set of inputs.
The simple cross-plotting of each input against its
corresponding output can give an indication of the quality of
linear or multiple linear models which can be formed. Such
simple tools, however, do not provide adequate solution for
complex petroleum engineering problems. To rank our drivers,
a fuzzy logic algorithm16,17 was used. This algorithm
statistically determines how well a particular input could
resolve a particular output with respect to any other input
using Fuzzy Curve Analysis, details on such method are
discussed in references16,17,18
Table 4 shows the result of the fuzzy ranking for both
Productivity Index model (PI) and the Fuzzy Logic based
6 SPE 84870
Fracture Index model (FFI). We notice that the major drivers
for the first case are those related to lithology (permeability
and porosity) and then come the structural drivers whereas for
the second case it is those geomechanics related drivers which
come first followed by lithology drivers.
Neural Network Design
After trying several architectures with different
number of layers, neurons and transfer functions, the best
results have fallen on the one described in Figure 17. This
network has two hidden layers with logarithmic transfer
functions implemented in all neurons. The first hidden layer
has eight neurons. Each neuron has a bias and is fully
connected to all inputs, which means 8×8+8 =72 connections.
The second hidden layer has also eight neurons where each
one has a bias. The connection between this layer and the
previous one is ensured by neuron to neuron connection (each
neuron from the second hidden layer is connected to one and
only one neuron from the first hidden layer and vise versa).
This layer has then 8+8 =16 connections. All the neurons of
the second hidden layer are connected to the output layer. The
output layer has only one neuron without bias which means 8
more connections. Finally we have 96 connections to be
adjusted by a proper training algorithm.
Inputs Selection
Selecting the appropriate number of inputs is crucial, yet
difficult task. However, we have found that taking the first
eight inputs among the available thirteen leads to good results.
The number of eight inputs has neither been chosen arbitrarily
nor under certain precise rules. It is the result of the
following observations:
• We only have few data available for the training.
Taking all the thirteen inputs with a similar architecture results
in 221 connections and thus very little chances to ovoid
the overtraining.
• At least the first layer of the network should have a
full connection with the inputs.
• The more inputs we introduce, the better modeling
we could achieve.
• To better model the problem we should have inputs
from various sources
• Some inputs have a strong relationship with the
fracture index than others do.
Taking these five remarks into account and the fuzzy
ranking results of the drivers, we have chosen to take the first
eight inputs.
Neural Network Modeling
A key element in using Artificial Neural Network in
this study is that the fracture distribution is a complex process
to be modeled by classical regression. A neural network
modeling process comprises three basic steps, training, testing
and predicting.
Training: The Overfitting Problem
One of the most common problems in the training process
is the overfitting. This happens when the error on the training
set is driven to a very small value, but when new data is
presented to the network the error is large. This problem
occurs mostly in case of large networks with only few
available data.
There are a number of ways to avoid overfitting3.
However, with just fixing the error and the number of epochs
to an adequate level (not too low , not to high) and dividing
the data into two sets; training and testing; one can ovoid such
problem by making several realizations and selecting the best
of them.
Training Design
For both models 75% of the available data were used
to train the neural network. The other 25% were left aside for
testing purpose. Prior to any modeling all data were scaled to
the range [0 1]. Three training algorithms were tested:
conjugate gradient, scaled conjugate gradient and Levenberg
Marquadt. We have found, however, that the second and the
third ones are more suitable for our case. Training stop criteria
was chosen to a maximum number of epochs of 150 and a
mean squared error of 0.005.
Testing
Several realizations were conducted where
sensitivities were made on the training algorithm parameters:
the training algorithm itself, the stopping criteria and the wells
used in the training and testing sets. Once the training process
converges, the testing data set must be presented to the
network. If the testing present good agreement between the
actual and the predicted fracture index, the bias and weight
matrices must be saved and kept aside. If not, the realization
is cancelled. Such process has been repeated several times
until a satisfied number of realizations with good testing
results were achieved.
Predicting
Once the training leads to satisfactory results, the
network is considered to be trained and generalized well. It is
then ready to predict the fracture index on the other grid
blocks. From those best realizations obtained in the previous
two sections, the one that produce a fracture map
corresponding to intervals where we do have high confidence,
is selected
Neural Network Results: PI model
For this model, productivity index was taken as a
fracture index assuming that well performance is linearly
related to the fractures. Only early PI’s were selected, these
PI’s range from 0.001 to 1.68 m3/h/kg/cm². Since the output
transfer function in the Neural Network model is sigmoid
which produces an output in the range of [0 1], we must
provide our data by fracture indexes in the same range [0 1]. If
we take FI = 0.5 × PI, we get our fracture index in the range of
[0 0.84] leaving a degree of freedom of 0.16 for the network to
predict higher values (if they could be). From the 46 available
wells, 10 were chosen from different locations and left aside
for testing purpose. The training/testing and predicting process
has been conducted using a Matlab program based on BPNN.
Figure 18 shows the training testing results. It may seem that
the predicted values are somehow different from the actual
values, but they can be considered in a good agreement for a
SPE 84870 7

complex rock failure process. Had we forced the training to a 10*

FIi , j
low deviation error, the Network would have been trapped in a length = 1.5
max FIi , j

memorization problem (overfitting) and predicted

erroneous outputs.
After testing the network, it was used to predict the fracture Results and Discussion
Maps in Figures 22 and 23 show the fracture network
index in other grid blocks. Results are shown in a 2D map,
for both models. These maps have been generated from the
Figure 19. Comparing this map with the structure top, Figure
fracture intensity maps using a Matlab program and the
12, we can notice that the major fractured intervals are those
previously discussed weighting method. They help giving an
which are located near the major faults. We could also notice
insight about nested areas and major trends of the fractures.
that the inter zone 8-13, Figure 2, is of low fractures except in
Figures 24 and 25 show the rose diagrams for the two models.
the upper middle part following the extension of the fault
We can see that the major trend is SN for the PI model and
passing between zones 9 and 10.
NWSE for the FFI model; we can also notice that these trends
are not well distinguished especially for the case of PI model.
Fuzzy Logic Based Fracture Index Model (FFI model)
We believe that the current proposed model is much more
restricted and reliable in situations of high complexity such as Conclusions
Following conclusions are drawn from this
reservoir heterogeneity. However, the application conducted
research study:
up to this point suffers somehow from data scarcity. Further
1. Artificial Intelligence was successfully applied to predict
investigations with large amount of data may improve
2D fracture intensity maps within a large block in Hassi
the results.
Mesaoud field.
From the 26 available wells, six were left for the testing
2. Two models have been presented and applied. The first
process. A careful training has been conducted using the
one uses single tool to derive fracture index, the second
remaining 18 wells. The same procedure and programs were
uses many.
used. The training/testing results are shown in Figures 20. We
3. By building a fuzzy inference system in the second
notice that both training and testing sets present almost the
model, we reduced the uncertainty, coupled dynamic to
same differences between actual and predicted values. In other
static data and came up with a more reliable
words, despite the data scarcity, the network generalizes very
Fracture Index.
well. This may be due to the representative fracture index
4. Fracture network maps have also been developed. In the
which gathers several indicators.
first model the major trend of fractures is South to North,
The map in Figure 21 shows the predicted FFI over the
in the second it is Northwest to Southeast.
entire gridblocks. Comparing this map with the structure top
Future Work
map and well-position map (zone of interest) in Figures 12 and
For future research, following recommendations
2 and taking a quick glance on Table 1, one can make the
are made:
following observations:
1. To conduct interference well test in those fractured areas
• Fractures are located in both near faults and
prior using the results of this study for any development
folded regions.
scheme, this will increase the chance and confidence of
• Even though horizontal wells were not used, the map drilling successful wells.
confirms to a certain degree the channel crossing the well 2. To improve the Fuzzy Logic approach and extend it to a
OMPZ 34 obtained from well test analysis, Table 1. 3D model and integrate other types of data. The following
• Zone near OMPZ 111 presents a high level of points will help achieving such improvements.
fractures, which confirms the high productivity of this well • Use all available wells to build more reliable 3D
which may be due to the presence of NFR. stochastic geological model (For permeability, porosity
and shale volume). For missing intervals or uncored
Fracture network wells, Neural Network can be used to predict
To visualize the main trends of the fractures, Ouenes et permeability and porosity from well log data.
al.14 proposed a mathematical model based on the “weighting
• Use 3D seismic data if available. 3D seismic data will
method”and the following assumptions:
help building more reliable stochastic model with
• Each grid block on the fracture intensity map is associated sequential simulation and having a meaningful structure
with a fracture top and bed thickness.
• This fracture has an angle and an intensity represented by • Extend well performance data along the wellbore by
a line. using some basic reservoir engineering equations and/or
• The fracture intensity as represented by a line is a integrating PLT if available.
combination of the aperture, fracture population in the • Integrate other data from other sources such as the
gridblock and fracture length. imagery log from horizontal well logging and loss
This model helps deriving the fracture network map from circulation from drilling data.
the fracture intensity map predicted by Neural Network. It
calculates the angle from the eight surrounding gridblocks and
the length from a power formula:
8 SPE 84870

Nomenclature 18. Weiss, W.W. and Balach, R.S.: “How Artificial Intelligence
Rfi : Borehole corrected short normal resistivitly, Ω-m. Methods can Forecast Oil Production”, SPE 75143
Rfo : Borehole corrected log normal resistivity, Ω-m. 19. Gauthier et al.: “Integrated Fractured Reservoir
Rmf : Mud filtrate resistivity, Ω-m. Characterization: a Case Study in a North Africa Field.”, SPE
Rma : Matrix resistivity, Ω-m and, 65118 .
Rw : Water resistivity,Ω-m.
20. Aissaoui, K.: “Contribution of Horizontal Well Logs and
Pressure Data in Stochastic Modeling.” MS Thesis, Norman,
References Oklahoma 2001.
1. Haykin, S.: “Neural Networks,” McMaster University Hamilton,
Ontario, Canada. Prentice Hall International, Inc. Table 1 Horizontal wells performance
2. Hagan, M.T.: Howard B. Demuth, Mark Beal,: “Neural Network
Design”, PWS Publishing Company.
WELL PI
Skin kx ky kz
NAME m3/h/kg/cm2 mD mD
mD
3. Demuth, H. and Beale, M: “Neural Network Toolbox for use
with MATLAB”, User’s Guide, Version 4, MATLAB 6.0. OMPZ34 1.14 -2.7 1.9 200 82
OMPZ111 3.4 1.7 17 35 0.7
4. Demuth, H. and Beale, M.: “Fuzzy Logic Toolbox for use with
MATLAB”, User’s Guide, Version 2, MATLAB 6.0. OMPZ331 0.412 -1.9 1.75 6.12 1.98

5. Tiab, D.: “Advances in Petrophysics, Volume 1- Flow Units.”

Table 2 Fractured intervals (obtained from core description)
Norman, Oklahoma, U.S.A. 2000.
6. Richradson, Q.S. and Weiss, W.W.: “Fractured Reservoir Fractured Interval (%)

Well Name
Characterization and Performance Forecasting Using
Geomechanics and Artificial Intelligence,” paper SPE 30572 Class A Class B Class C

Open

Close

Open

Close

Open

Close
7. Zellou, A.M., Ouenes, A. and Banik, A.K: “Improved Fractured
Reservoir Characterization Using Neural Networks,
Geomechanics and 3-D Seismic,” paper SPE 30722
# 01 2.84 0 3.78 0.95 0 0
8. Ouenes, A., Zellou, A.M., Nova, T., Basinski, P.M. and Head, # 02 0 0 10.51 2.97 0 0
C.F.: “Practical Use of Neural Networks in Tight Gas Fractured
Reservoirs: Application to the San Juan Basin.” Paper SPE # 03 0 0 15.49 1.46 0 0
39965 # 04 0 0 10.00 7.50 0 0
9. Ouenes, A. and Hartley, L.: “Integrated Fractured Reservoir # 05 0 0 9.53 6.32 0 2.08
Modeling Using Both Discrete and Continuum Approaches.” # 06 0 0 5.71 6.86 0 0
Paper SPE 62939
# 07 0 0 23.56 0.61 0 0
10. Ouenes, A.: “Practical application of Fuzzy Logic and Neural # 08 0 0 13.09 0.15 0 0
Networks to Fractured Reservoir Characterization”, Computers
and Geosiences, Vol.26, pp. 953-962, 2000. # 09 0 0 6.75 27.58 0 0
11. Barman, A., Ouenes, A. and Wang, M.: “Fractured Reservoir # 10 0 0 8.85 0.77 0 0
Characterization Using Streamline-Based Inverse Modeling and # 11 1.08 0 28.85 9.19 0 0
Artificial Intelligence Tools.” Paper SPE 63067. # 12 0 0 11.54 13.08 0 0
12. Sadiq, T. and Nashawi, I.S.: “Using Neural Networks for # 13 1.88 0 24.38 7.15 0 0
Prediction of Formation Fracture Gradient.” Paper SPE 65463.
# 14 0 0 5.43 27.32 0 0
13. Eduardo, Quintero, J., Martinez, L.P. and Gupta, A: # 15 0.60 0.2 18.10 5.40 0 0
“Characterization of Naturally Fractured Reservoirs Using
Artificial Intelligence.” paper SPE 67286. # 16 26.60 6.60 22.98 24.04 0 0
14. Balch, R., Weiss, W. and Wo, S.: “Correlation of Traditional # 17 0 0 2.23 2.05 0 0
Wire-Line Logs to FMI Estimated Secondary Porosity”, For: # 18 4.41 2.06 2.94 2.65 0 0
Yates Petroleum Corporation –April 1998. # 19 0 1.79 0.22 0.11 0 0
15. Soto, R. B., Garcia, J.C., Torres, F. and Perez, G.S.: # 20 0 0 6.09 3.85 13.40 0
“Permeability Prediction Using Hydraulic Flow Units and
Hybrid Soft Computing Systems,” SPE 71455. # 21 0 0 24.29 5.71 0 0
# 22 0 0 5.56 4.44 0 0
16. Weiss, W.W., Wo, S., Weiss, J.W. and Weber, J.: “Data Mining
at a Regulatory Agency to Forecast Waterflood Recovery” SPE # 23 2.01 5.97 7.78 0.42 0 0
71057 # 24 0 0 22.83 0.83 0 0
17. Balch, R.S., Broadhead, R. F., Weiss, W.W. and Hart, D. M.:
“Regyonal Data Analysis to Better Predict Drilling Success:
Brushy Canyon Formation”, SPE 75145
SPE 84870 9

Table 3 Fuzzy logic based Fracture Index (FFI).

Welle name

(fraction)

High_K

Skin
NPI
(%)

FFI
FL

CI

# 01 0.07 0.63 29.08 0.38 -4.54 0.42

# 02 0.11 0.5 16.20 0.39 -1.60 0.21
# 03 0.15 0.5 6.17 0.20 -2.34 0.27
# 04 0.10 0.5 23.91 0.61 -3.26 0.41
# 05 0.10 0.48 14.86 0.98 -2.65 0.41
# 06 0.06 0.5 6.18 0.19 -2.00 0.18
# 07 0.24 0.5 3.28 0.04 -2.07 0.30
# 08 0.13 0.5 49.59 0.31 0.00 0.25
Zone of interest
# 09 0.07 0.5 32.77 0.80 -3.58 0.48
Figure1. Map showing the 26 zones of Hassi Messaoud field
# 10 0.09 0.5 6.90 0.25 -3.49 0.31
and the location of the zone of interest.
# 11 0.30 0.51 21.32 0.13 9.70 0.23
140,000
# 12 0.12 0.5 10.20 0.51 -3.96 0.38 CMLPZ173

# 13 0.26 0.51 3.14 0.58 -2.71 0.39

138,000 CMPZ201
# 14 0.05 0.5 30.20 0.21 -1.93 0.28 N
# 15 0.19 0.51 14.34 0.26 -3.69 0.39
# 16 0.50 0.65 36.56 3.17 -3.77 0.78
136,000
# 17 0.02 0.5 18.81 2.12 -2.38 0.44 OMPZ34

# 18 0.07 0.71 3.76 1.68 -0.22 0.33

OMPZ331
0.01 0.83 0.93 0.22 -0.90 0.10 134,000
# 19
# 20 0.19 0.29 8.75 1.30 -0.14 0.32
# 21 0.24 0.5 7.01 1.15 0.00 0.34
132,000
0.06 0.5 5.20 0.00 -0.22 0.06 OMPZ111
# 22
# 23 0.10 0.69 28.13 1.31 -3.34 0.59
# 24 0.23 0.5 8.45 1.07 -1.69 0.40 130,000
820,000 822,000 824,000 826,000 828,000

Figure 2. Map showing the zone of interest.

Table 4 Fuzzy Ranking Results.
Strength Rank 1000 125%
Frequency
Driver % Sum
Model

Model

Model

Model
FFI

FFI

800 100%
PI

PI

600 75%
Frequency

% Sum
Permeability 100 57.70 1 4
2
88.52 23.29 2 8 400 50%
Porosity
Vsh 53.52 36.30 10 5 200 25%
Slope SN 59.98 28.72 8 7
0 0%
Slope WE 44.74 19.12 12 11
0.04

0.12

0.2

0.28

0.36

0.44

0.52

0.6

0.68

0.76

0.84

0.92

Slope SWNE 64.10 21.63 5 9 Class

Slope NWSE 41.39 13.91 13 13

Figure 3. Histogram & cumulative distribution function
Curvature SN 62.78 94.98 6 2 "CDF" of Vsh.
Curvature WE 54.42 15.03 9 12
Curvature SWNE 48.65 19.90 11 10
Curvature NWSE 78.44 100 3 1
Bed thickness 71.60 83.88 7 3
Distance to the
76.88 30.07 4 6
nearest fault
10 SPE 84870

1000 125%
Frequency
% Sum 25
800 100%

20
600 75%
Frequency

% Sum
15

Variance
400 50%
10

200 25%
5

0 0%
1 3 5 7 9 11 13 15 18 More 0
0 50 100 150 200 250 300 350 400 450 500
Class
Lag distance
Figure 4. Histogram & cumulative distribution function
"CDF" of porosity.
800 125% Figure 8. Average horizontal semivariogram of Vsh.
Frequency
% Sum
100%
600

75%
Frequency

% Sum

400
50%

200
25%

0 0%
-2 -1 0 1 2 3 4 5 6 7 8
Class

Figure 5. Histogram & cumulative distribution function

"CDF" of permeability.

4000

3000
Variance

2000

1000

0
0 25 50 75 100 125 150 175 200
Lag distance

Figure 6. Average horizontal semivariogram of permeability.

Figure 9. .Map showing log normal permeability distribution.
8

6
Variance

0
0 25 50 75 100 125 150 175 200
Lag distance

Figure 7..average horizontal semivariogram of porosity.

SPE 84870 11

Figure 12. Map showing the structure top over which the
directional derivatives were calculated

Figure 10. Map showing porosity distribution.

Bed Thickness

Slopes and
Curvatures
x

Distance to
Faults

Figure 13. Geomechanical drivers

Figure 11. Map showing shale distribution.
12 SPE 84870

Figure 14. Fuzzy rules of evaluation

Figure 17. Neural network architecture.

1.0
Training
Testing
Predicted Fracture Index

0.8

0.5

0.3

0.0
0 0.2 0.4 0.6 0.8 1
Figure 15. Input fuzzification. Actual Fracture Index

Figure 18. Predicted versus actual fracture index (PI model).

1
Trainig
Testing
Predicted Fracture Index

0.8

0.6

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1

Actual Fracture Index

Figure 20. Predicted versus actual fracture index (FFI model).

Figure 16. Fuzzy inference diagram
SPE 84870 13

5
x 10
1.4

1.39

1.38
High
1.37

1.36

1.35

1.34

1.33

1.32
Low
1.31

1.3
8.2 8.21 8.22 8.23 8.24 8.25 8.26 8.27 8.28
5
x 10

Figure 19. Fracture intensity map based on the productivity Figure 22. Fracture Network Map (PI model).
index (PI model).
5
x 10
1.4

1.39

1.38
High
1.37

1.36

1.35

1.34

1.33

1.32
Low
1.31

1.3
8.2 8.21 8.22 8.23 8.24 8.25 8.26 8.27 8.28
5
x 10

Figure 21.. Fracture intensity map based on the fuzzy logic

Fracture Index (FFI model). Figure 23. Fracture network map (FFI model).
14 SPE 84870

Figure 24. Rose diagram showing fractures orientation

(PI model).

Figure 25. Rose diagram showing fractures orientation (FFI

model).