You are on page 1of 8

Rock Mechanics for Natural Resources and Infrastructure

SBMR 2014 – ISRM Specialized Conference 09-13 September, Goiania, Brazil


© CBMR/ABMS and ISRM, 2014

3D Geological Modeling of Subsurface for Drilling Purposes Using


Neural Networks and Fuzzy Logic
Alvaro Gustavo Talavera Lopez
PUC-Rio, Rio de Janeiro, Brasil, alvaro@ele.puc-rio.br

Vivian Rodrigues Marchesi


PUC-Rio, Rio de Janeiro, Brasil, vivianrm@puc-rio.br

Débora Lopes Pilotto Domingues


PUC-Rio, Rio de Janeiro, Brasil, deborapilotto@puc-rio.br

Sergio Augusto Barreto da Fontoura


PUC-Rio, Rio de Janeiro, Brasil, fontoura@puc-rio.br

Clemente José Gonçalves


Petrobras, Rio de Janeiro, Brasil, clemente@petrobras.com.br

Marcos Fonseca Alcure


Petrobras, Rio de Janeiro, Brasil, marcosalcure@petrobras.com.br

SUMMARY: 3D geological modeling is an important part of geomechanical models. Therefore, the


knowledge of spatial distribution algorithms of geological properties is essential for the proper
characterization of the subsurface. Generally, geostatistical and neural networks can be used as
forecasting strategies of geological characteristics. One problem in forecasting logs and
geomechanical properties is that they are not linearly related to several decision variables. Through
neural networks it is possible to describe the behavior of complex systems and to estimate relevant
correlations between a given set of inputs and another of outputs using a nonlinear function.
Alternatively, fuzzy logic is a knowledge-based system that allows adding rules, which can be used
to infer the result of a forecast. This work presents a methodology for 3D geological modeling of oil
fields using artificial neural networks for forecasting geological properties, and a Fuzzy logic system
– employed in order to improve the performance of neural network through the inclusion of geological
rules. It was concluded that, using both computational intelligence techniques, it is possible to obtain
reliable models based on numerical results.

KEYWORDS: Neural networks, Fuzzy logic, properties forecast.

1 INTRODUCTION performance of the predictive models of standard


statistical methods.
One problem in predicting logs and 3D Neural networks have great potential for the
geomechanical properties is that they are related precise description of complex systems
to several decision variables, such as seismic behavior. Neural networks make it possible to
attributes, lithofacies, information core and well approximate a relation between an input set of
logs. This large database generates highly dimension n and an output set of dimension m
nonlinear relations regarding the variables to be through a nonlinear function. This method
forecast; thus, contributing to the low allows for an efficient representation of highly

SBMR 2014
nonlinear relations and performs well in variables. A principal component analysis was
situations in which the mathematical model carried out in order to reduce the dimension of
related to the problem to be solved is very data space variables. The variables were chosen
complex to obtain (Haykin, 1999; Mohaghegh, among seismic attributes, well logs and
2000). geological properties of reservoir. This strategy
Artificial neural networks are becoming has the advantage of assigning neural networks
increasingly popular in areas such as geology for each spatial region of the reservoir for more
and geophysics. They are used, for instance, in accuracy in prediction.
the characterization of petroleum reservoirs, for In general, the prediction of geological
geological properties forecasting and well logs properties by neural networks or by any other
prediction, for seismic inversion and fractures prediction model are subject to errors due to
modeling (Mohaghegh and Ameri, 1995; Van possible criteria disregarded in the forecast, such
der Baan and Jutten, 2000; Calderón-Macías et as geological faults, fractures, data quality, very
al., 2000; Mohaghegh, 2005; Baddari et al. thin facies layers and heterogeneity in the field.
(2010) and Darabi et al., 2010). To correct possible errors in the prediction, a
Huang et al. (1996) used back-propagation system with fuzzy logic is used to infer the result
artificial neural networks (BP-ANN) to model of the prediction model.
the interrelations between spatial position, six Lin and Cunningha (1995) have proposed the
different well logs, and permeability. The BP- use of fuzzy logic to model the identification of
ANN produced permeability values that nonlinear systems. Subsequently, other studies
compared well to measured values in the cored used this strategy for reservoir characterization
intervals. Fung et al. (1996) used neural (Bhuiyan, 2001; Weiss et al., 2001). Lim and
networks and Learning Vector Quantization Kim (2004) proposed a model that uses fuzzy
(LVQ) for forecasting rock properties. Saggaf et logic to select the best reservoir properties,
al. (2003) used BP-NN to estimate the which will be used by the neural network for the
distribution of reservoir porosity using 3D prediction of permeability and porosity. Taghavi
seismic in the field of Ghawar, Saudi Arabia. (2005) used linear regression and fuzzy logic in
Wang et al. (1998) modeled porosity the prediction of permeability and concluded that
distribution through the integration of techniques the fuzzy logic produced better results.
of neural networks and kriging in the field of Kadkhodaie-Ilkhchi et al. (2009) used seismic
A'nan, north of China. The association of attributes as decision variables in a committee
geostatistics and neural network can also be fuzzy inference system for estimating water
found in Niu et al. (2000) and Nakayama (2000). saturation and porosity parameters. The model
Studies involving the relation between reservoir proposed in this paper is divided in two stages.
properties and seismic attributes to estimate 3D The first stage is the prediction of properties
reservoir parameters can be found in Arzuman using neural networks and the second stage is to
(2009). Although seismic attributes assist in use fuzzy logic to correct possible errors in the
predicting the geological properties, they add prediction, using expert knowledge. Next, we
more complexity to the forecasting operation, present the theoretical framework of the theories
due to nonlinearities of seismic attributes and used herein such as neural networks, fuzzy logic
large amount of seismic 3D data. and principal component analysis. The proposed
Most studies mentioned above use a single method is described afterwards followed by the
neural network for predicting properties in the description of the application of the method to
well and 3D field. This paper proposes a model predict sonic logs at a field case. Finally, we
of multi-neural networks in which each neural present some conclusions regarding the method,
network is designed for each geological zone of its potential and possible developments for its
the subsurface. For this purpose, a study of the application are drawn.
different types of input variables in each neural
network appropriate for each zone was carried
out. This study implied working with many

SBMR 2014
2 THEORY Neural network modeling process comprises
three basic steps related to experimental data:
2.1 Neural Networks training, testing and predicting. Network training
is the process of adjusting the network
The Rumelhart-Hinton-Williams multilayer connection weight so that explanatory and
network that we consider here is a feed-forward response values match the data as closely as
type network with connections between possible. Testing data are used for checking the
adjoining layers only. Networks generally have training. Finally, validating data are used for
hidden layers between the input and output studing model accuracy. A comprehensive
layers. Each layer consists of computational reference on neural networks can be found in
units as it can be seen in Figure 1. Haykin (1999).
The input-output relationship of each unit is
represented by inputs 𝑥𝑖 , output 𝑦, connection 2.2 Fuzzy Logic
weights 𝑤𝑖 , threshold 𝜃, and differentiable
function 𝜑 as follows: Fuzzy logic is a logic based on fuzzy set theory
which aims to model the approximate mode of
𝑦 = 𝜑(∑𝑘𝑖=1 𝑤𝑖 𝑥𝑖 − 𝜃) (1) reasoning, mimicking human ability to make
decisions in an environment of uncertainty and
Usually, the function φ is called the activation imprecision (Zadeh, 1976; Pedrycz and Gomide,
function, which is a bounded and monotone 1998). It allows intelligent control and decision
differentiable function such as the sigmoid support systems to deal with imprecise
function represented by Equation (2), where 𝑎 is information. A fuzzy logic system is able to
the slope parameter of the function (Funahashi, simultaneously process numerical data and
1989). linguistic knowledge (Mendel, 1995). In fuzzy
set theory, each element may belong to a set of
1 degree (μ), which can take values ranging from
𝑓(𝑥) = (2)
1+𝑒 −𝑎𝑥 0 to 1. Each fuzzy set is represented by a
membership function (MF). MFs can present
The learning rule of the neural network is several types, such as Gaussian, triangular,
known as backpropagation algorithm trapezoidal and sigmoidal, see Figure 2.
(Rumelhart et al., 1986) and consists in the use
of the gradient descent method to find a set of
weights so that the error between the desired
output and the output signal network is
minimized.

Figure 2. Examples of four classes of parameterized MFs

We aim, with this work, at developing a fuzzy


Figure 1. Schematic Multi-Layer Perceptron (MLP) inference system FIS, which is a process that
network (Darabi, et al. 2010). maps a set of input data into a set of output data

SBMR 2014
using fuzzy logic. A FIS has 4 main parts, The process of Mamdani method is
presented in Figure 2, which are (a) illustrated in Figure 4 and more details on this
fuzzification, (b) rules, (c) inference engine and method can be obtained from Kadkhodaie-
(d) defuzzification. Ilkhchi (2009).

Figure 4. Illustration of Fuzzy Inference Systems.

2.3 Principal Components Analysis

The principal component analysis (PCA) is a


Figure 3. Main parts of a Fuzzy Inference Systems. linear transformation of data to a new coordinate
system (space features). This transformation is
The fuzzification consists in mapping out the designed so that the data set can be represented
accurate inputs or non-fuzzy (crips) numbers by a reduced number of data ensuring higher
(resulting measurements) into fuzzy sets. This intrinsic information contained in the data set.
process is needed in order to activate rules which Therefore, the data set is reduced dimensionally.
are expressed in terms of linguistic variables and In this sense, PCA can be defined as an optimal
have fuzzy sets associated to them. Rules can be linear transformation in order to preserve the
provided by experts or extracted from the subspace that has the highest variance. Principal
numerical data (this method is particularly useful component analysis is a technique commonly
in problems of forecasting time series) in the used in pattern recognition and signal processing
form of linguistic sentences, e.g., “IF 𝑢1 is very applications.
warm and 𝑢2 is quite low THEN turn 𝑣 In petroleum field, it is widely used especially
somewhat to the right”. The inference process in the area of geo-modeling and in modeling 3D
are used to combine fuzzy IF-THEN rules from properties where the amount of data is too large
to fuzzy rule base into a mapping from fuzzy e.g. seismic attributes. In this sense, it is used to
input set to fuzzy output set. Once the fuzzy reduce the dimension of the data. Hagen (1982)
output set is obtained through inference (modus describes the application of PCA to model
ponens generalized), the stage of defuzzification seismic stratigraphy and Brito (2010) also
is carried out in order to interpret the mentions seismic data and seismic attribute
information. There are several defuzzification volume to improve seismic stratigraphy using
methods in the literature, two of the most used PCA. Guo et al. (2006) describe the use of PCA
ones are the center of gravity and the mean of the to reduce the dimension of the spectral
maximum. components in seismic images to delineate
In this work, the fuzzy inference used is stratigraphic features of interest.
called the method of Mamdani in which the
output membership functions are fuzzy sets.
After the aggregation process, there is a fuzzy set 3 PROPOSED MODEL
for each output variable that needs
defuzzification. For example, consider the rule: This paper proposes a model using neural
networks and fuzzy logic for predicting
𝑅𝑖 : 𝑖𝑓 𝑥 𝑖𝑠 𝐴1 𝑎𝑛𝑑 𝑦 𝑖𝑠 𝐵1 𝑡ℎ𝑒𝑛 𝑧 𝑖𝑠 𝐶𝑖, 𝑖 = 1,2, … , 𝑛 geological properties and electric logs in

SBMR 2014
petroleum reservoirs. Figure 5 shows the block
diagram of the proposed model. The model is
divided into three steps. The first step is called
data preparation; it includes the division of the
model field in 𝑛 geological zones and the
evaluation of decision variables that the neural
network will use in predicting properties in each
zone. The decision variables include the seismic
attributes due to its 3D nature (seismic cube),
which helps predicting properties (Taner, 2001),
the facies observed at the drilled wells that
provide the geological lithology in each zone, the
well log at drilled wells and the normalized
coordinates x, y, z for each well. Finally, data
dimensionality is reduced using PCA. Figure 6. Neural Networks Strategy.

Once the prediction by neural network is


carried out, a fuzzy logic system is applied. Since
it is possible to enter if-then rules with a geologic
sense in fuzzy system in order to eliminate errors
due to possible criteria disregarded in the
forecast, such as geological faults, fractures, low
data quality noises, very thin facies layers, there
is a reduction of errors in forecasting, especially
in thinner layers of zones where the neural
network does not have good performance in
prediction, as found by Tang et al. (2011).
The third step is to regroup all the results of
Figure 5. Proposed Model.
neural networks in each zone to generate the 3D
model of the field. To evaluate the predicted
The second step is the development of the outcome the results of all networks in each zone
model using neural networks and fuzzy logic. In are added together and compared to the same
this step, neural networks are designed for each property in well validation (i.e. a well whose data
geological zone in order to predict some property were not used at all stages).
̂. The advantage of using this strategy is that it
𝑥
is easy to identify which neural network had the 4 APPLICATION OF THE PROPOSED
lowest or highest performance in predicting and MODEL
then change one decision variable or the
structure of the neural network. On the other In this section, the proposed model is applied for
hand, choosing the input variables for each the studied field. The methodology used is as
neural network is complex because each network follows: (i) the studied field has eight wells that
will not necessarily have the same entries. The will be referred to as 𝑝1 , 𝑝2 , 𝑝3 , 𝑝4 , 𝑝5 , 𝑝7 and 𝑝8 ,
strategy used for setting is shown in Figure 6. (ii) well 𝑝4 is chosen as well validation, i.e. it
will only be used to compare with the final result
of the prediction made by the proposed model.
The studied field is divided into seven geological
zones so seven neural networks will be
projected.

SBMR 2014
In this first example, the objective is to design according to equation 3, where 𝑥 is the value of
the proposed model to predict the cube of property DTC in well validation 𝑝4 , 𝑥̂ is the DTC
compressional sonic log, DTC, in the field. As estimated by the proposed model. Using this
explained in the previous section, the set of input error measure, a MAPE of 6.10% was obtained.
variables for each neural network is selected
|𝑥−𝑥̂|
using studies such as Taner (2000), Arzuman 𝑀𝐴𝑃𝐸 = × 100 (3)
𝑥
(2009) and Azevedo et al. (2009) that show the
usefulness of each seismic attribute for detecting The same exercise was done for the shear
geological properties. The selected variables are sonic log, DTS, following the same procedure.
presented in Table 1, where A is amplitude, AI is Table 2 shows the input variables for the neural
the acoustic impedance, IV is interval velocity, F networks. The result of the prediction model is
is the facies and ‖𝑧‖ is the normalization of shown in Figure 8, with a MAPE regarding well
depth. validation of 6.80%.
The input variables of each network is shown
in Table 1 and the results of the proposed model Table 2. Selecting inputs by neural networks - DTS.
of DTC prediction in the 3D field is shown in Zone NN A AI IV F ‖𝑍‖
Figure 7, where the vertical arrow is the location
of the well validation 𝑝4 1 1 X - X X X
2 2 X X - X X
3 3 - X X X X
4 4 X X - X X
Table 1. Selecting inputs by neural networks - DTC. 5 5 - X X X X
Zone NN A AI IV F ‖𝑍‖ 6 6 - X X X X
7 7 X X - X X
1 1 X - X X X
2 2 X - X X X
3 3 X - X X X
4 4 X X - X X
5 5 - X X X X
6 6 - X X X X
7 7 X X - X X

Figure 8. Forecast DTS by Neural Network at coordinate


J.

It can be seen in Figure 8, the formation of


pockets, which can be interpreted as errors in
forecasting, especially in very thin layers where
the neural network cannot get a good prediction,
as in Tang et al. (2011). One way of solving this
Figure 7. Prediction of DTC at coordinate J. problem is to use a filter to eliminate these
disturbances. In this case, it is advisable to use
To validate this model the Mean Absolute Fuzzy Logic. To do this, rules for the behavior of
Percentage Error (MAPE) is calculated properties of well profiles such as DTC, Gamma

SBMR 2014
Ray and DTS, were generated as follows: Applying the fuzzy system in prediction made
by the neural network in thinner layers can
1. If RG is low and RHOB is high then DTS is reduce the approximation error of the neural
low. network. This can be seen in Figure 11. Using
2. If RG is high and RHOB is medium then this strategy, the MAPE is reduced to 5.94%.
DTS is medium.
3. If RG is low and RHOB is medium then DTS
is medium.

It is important to highlight that these rules are


just examples and should not be applied as
general truth.
The fuzzy sets were designed according to the
expert's knowledge and statistics of each profile,
as shown in Figure 9. The fuzzy sets and fuzzy
inference system were developed using the
Petrel calculated by a program *. mac.

Figure 11. Filtered using Fuzzy Logic.

It can also be seen in Figure 11 that the neural


network has a lower performance on thin layers
due to the limited information in these regions.

5 CONCLUSIONS
Figure 9. Fuzzy Sets Gamma Ray, respectively low,
medium and high GR. This paper presents a method for 3D geological
modeling of oil fields using artificial neural
The fuzzy system accepts two inputs: the networks for forecasting geological properties
Gamma Ray GR and Resistivity RHOB. Then, and a system of fuzzy logic to improve the
the input data are normalized (fuzzification prediction performance.
step), followed by the application of fuzzy The proposal for designing one neural
inference system and denormalization of data network for each different geological unit
output (desfuzzication), resulting in a DTS real improves the evaluation of the performance of
value. This process can be seen in Figure 10. each neural network as well as facilitates the
selection of input parameters for each geological
zone. The use of fuzzy logic in order to make the
prediction of properties more accurate indicated
that the technique is quite promising. As future
work, it is proposed to study de fuzzy inference
rules that can help predict specific properties;
possibly rules that include the degree of rock
fracturing.

ACKNOWLEDGEMENTS

Figure 10. Fuzzy inference system.


The authors thank Petrobras for the data
provided and Schlumberger for the academic

SBMR 2014
license of the Petrel software. Nakayama, K. (2000). "Estimation of Reservoir Properties
by Monte Carlo Simulation ." Yokohama: SPE Asia
Pacific Conference on Integrated Modelling for Asset
REFERENCES Management.
Niu, Y., P. M. Wong, and L. Chen (2000). "Sequential
Azevedo, L. (2009). Seismic Attributes in Hydrocarbon Neural Simulation: A New Approach for Stochastic
Reservoirs Characterization. MSc. Dissertation. Reservoir Modelling, 24-25 Outubro ." Paris: SPE
Departamento de Geociencias, Universidade de European Petroleum Conference.
Oviedo. Pedrycz W. and Gomide F. (1998). “An Introduction to
Arzuman, S. (2009). Comparasion of Geoestatistics and Fuzzy Sets: Analysis and Design”. MIT Press Complex
Artificial Neural Networks in Reservoir Property Adaptive Systems.
Estimation. Turkey: Theses-Geologial Engineering Rumelhart, D. E., G. E. Hinton, and R. J. Williams. (1986);
Deparment. The Midle East Technical University. Learning internal representations by error propagation.
Baddari, K., Noureddine D., Tahar A., and Jalal F. (2010). Vol. 1, in Parallel distributed processing: explorations
"Acoustic impedance inversion by feedback artificial in the microstructure of cognition, by D. E. Rumelhart,
neural network." Journal of Petroleum Science and J. L. McClelland, & PDP Research Group (Eds.),
Engineering 71, no. 3–4 (April 2010). edited by Rumerlhart, 318-362. Cambridge MA: MIT
Bhuiyan, M. H. (2001). An Intelligent System’s Approach Press.
to Reservoir Characterization in Cotton Valley. Master Saggaf, M., Nafi Toksoz M., and Mustafa H. (2003)
theses. Department of Petroleum and Natural Gas “Estimation of reservoir properties from seismic data.”
Engineering, 2001. Geophysics 68:1969–1983.
Calderón-Macías, C.; Mrinal K. S., and P. L. Stoffa Taghavi, A. A. (2005). "Improved Permeability
(2000). "Artificial neural networks for parameter Estimation through use of Fuzzy Logic in a Carbonate
estimation in geophysics." Geophysical Prospecting Reservoir from Southwest, Iran." Kingdom of Bahrain:
48 21–47. SPE Middle East Oil and Gas Show and Conference.
Darabi, H., Kavousi A., M. Moraveji, and M. Masihi. Taner, M. T. (2000). Attributes Revisited. Rock Solid
(2010). "3D fracture modeling in Parsi oil field using Images.
artificial intelligence tools." Journal of Petroleum Tang, H.; Toomey, N. and Meddaugh, S. (2011). "Using
Science and Engineering 71, no. 1-2: 67–76. and Artificial-Neural-Network Method To Predict
Funahashi, K. (1989). "On the approximate realization Carbonate Well Log Facies Successfully" SPE Journal
ofcontinuous mappings by neural networks." Neural Paper
Networks 2: 183-192. Tanscheit, R. Sistemas Fuzzy. Rio de Janeiro:
Fung, C. C., Kok W. W., Halit E., Charlebois, R. and Departamento de Engenharia Elétrica, PUC- Rio,
Crocker, H. (1996). "Modular Artificial Neural 2004.
Network for Prediction of Petrophysical Properties Van der Baan, M., and C. Jutten. (2000) "Neural networks
from Well Log Data." IEEE Instrumentation and in geophysical applications." Geophysics 65: 1032-
Measurement Technology Conference. Brussels. 1047.
Haykin, S. (1999). "Neural Networks: A comprehensive Wang, L., P.M. Wong , and S.A.R. Shibli . (1998).
foundation"Second Edition, Prentice Hall, New "Modelling Porosity Distribution in the A'nan Oilfield:
Jersey. Use of Geological Quantification, Neural Networks
Kadkhodaie-Ilkhchi, A., M. R. Rezaee, H. Rahimpour- and Geostatistics." SPE International Conference and
Bonad, and A. Chehrazi. (2009). "Petrophysical Data Exhibition in China. SPE 48884.
Prediction from Seismic Attributes Using Committee Weiss, W. W. , Shaochang Wo, and J.W. Weiss. (2011).
Fuzzy Inference System."Computers & Geosciences: "Data Mining at a Regulatory Agency to Forecast
2314-2330. Waterflood Recovery." Colorado: SPE Rocky
Lin, Y. and G. A. Cunningha. (1995). "A New Approach Mountain Petroleum Technology Conference.
to Fuzzy-Neural System Modeling." IEEE Zadeh, L.A. (1976). "Fuzzy Sets as a Basis for a Theory of
Transactions on Fuzzy Systems 3, no. 2: 190-198. Possibility". Fuzzy Sets andSystems, Vol. 1: 3-28.
Mohaghegh , S. and S. Ameri. (1995). "Artificial Neural Zehui, Huang; Shimeld, John; Williamson, Mark and
Network As A Valuable Tool For Petroleum Katsube, John. (1996)."Permeability prediction with
Engineers." Society of Petroleum Engineers. SPE artificial neural network modeling in the Venture gas
29220. field, offshore eastern Canada."Geophysics 61: 422-
Mohaghegh, S. (2000). “Virtual-Intelligence Applications 436.
in Petroleum Engineering: Part I - Artificial Neural
Networks.” Journal of Petroleum Technology,
Distinguished Author Series, September: 64-72.
Mohaghegh, S. (2005) "Recent Developments in
Application of Artificial Intelligence in Petroleum
Engineering." Journal of Petroleum Technology,
Distinguished Author Series, April: 86-91.

SBMR 2014

You might also like