Journal of Tethys: Vol.1, No.

2, 156-163 ISSN: 2345-2471 ©2013

Estimation of water saturation from petrophysical logs using radial basis function neural
network
Amir Mollajan1*, Hossein Memarian2

1- M.Sc at Mining Exploration Engineering, School of Mining Engineering, University College of Engineering,
University of Tehran, Tehran, Iran.
2- Professor of Geo–Engineering, School of Mining Engineering, University College of Engineering, University of
Tehran, Tehran, Iran.
* Corresponding Author: [email protected]

Abstract

Estimation of reservoir water saturation (Sw) is one of the main tasks in well logging. Many
empirical equations are available, which are, more or less, based on Archie equation. The present
study is an application of Radial Basis Function Neural Network (RBFNN) modeling for estimation
of water saturation responses in a carbonate reservoir. Four conventional petrophysical logs (PLs)
including DT, LLd, RHOB and NPHI related to four wells of an oil field located in southwest of
Iran are taken as inputs and Sw measured from core analysis as output parameter of the model. To
compare performance of the proposed model with empirical equations, the same database was
applied. Superiority of the RBFNN model over empirical equations was examined by calculating
coefficient of determination and estimated root mean squared error (RMSE) for predicted and
measured Sw. For the RBFNN model, R2 and RMSE are equal to 0.90 and 0.031, respectively,
whereas for the best empirical equation, they are 0.81 and 0.042, respectively.

Keywords: Water saturation, Petrophysical logs, Radial Basis Function Neural Network, Iran.

1–Introduction saturation in shaly–sand formations which can

be categorized into two groups:
Water saturation is one the most important
petrophysical properties of a hydrocarbon The first group includes models that import
reservoir that is mainly used to estimate the shale volume in their equations (Hossin 1960;
volume of hydrocarbon in place and Simandoux 1963; Poupon and Leveaux 1971,
determining pay zones. This parameter can be Dual–water 1977) and suppose that the
measured directly from Routine Core Analyses additional conductivity is related to the volume
(RCAL) or estimated by petrophysical methods. of the shale.
Various equations have been developed based
on petrophysical models like Archie equation The second group includes models that consider
for clean sand formations (1942).The presence the influence of clay mineral types in addition to
of clay mineral in sand formations the shale volume (Waxman and Smits 1968; and
(Worthington, 1985) or irregular distribution of Dual–water 1984). These models that have been
pore sizes in carbonate rocks (Van Golf–Rocht, developed based on the concepts of cation
1982) causes an additional conductivity while exchange capacity and ionic double–layer, have
Archie equation presupposes that the matrix of some disadvantages. For example, the
rock doesn't have electrical conductivity. This laboratory tests to obtain the input variables of
additional conductivity causes overestimation in these models are usually non–economic and
prediction of water saturation. Diverse time consuming. In addition different laboratory
equations have been presented to estimate water techniques result different answers for same
core sample and it also depends on precision of
Mollajan and Memarian, 2013 156 Available online at http://jtethys.org
Journal of Tethys: Vol.1, No. 2, 156-163 ISSN: 2345-2471 ©2013

the operator. Hence, some methods have been be necessary to find a new method for carbonate
developed to overcome this limitation like new reservoirs with respect to their special
LSU model (2002). properties. There are just a few works on the
prediction of water saturation in carbonate
Beside the above mentioned equations, some reservoirs (Obeida et al., 2005; Lucia, 2007;
other activities have been done. Balch et al. Mollajan et al., 2013). The present study
(1999) predicted the water saturation in a focuses on Sw prediction from PLs in a
sandstone reservoir in Mexico using artificial carbonate reservoir using Radial Basis Function
intelligence and seismic attributes. Kamel and Neural Network (RBFNN).
Mabrouk (2002) introduced an equation for
estimating water saturation in clean formations 3–Data set
utilizing resistivity and sonic logs. The effect of
To commence this study, four conventional logs
water and gas saturation on P and S wave
including DT, LLd, RHOB, and NPHI related to
velocity values in sandstone samples have been
four wells of an oil field in southwestern Iran
studied by Kitamura et al. (2006). Al–Bulushi et
were used as input data. As satisfactory
al. (2009) developed an artificial neural network
estimation results could only beobtained
to estimate water saturation and fluid
through the selection of appropriate data, the
distribution. A new predictive capillary pressure
water saturation measured from core analysis
function have presented by Tillero for Better
were used as output of the model and well test
Estimation of permeability and water Saturation
responses were also employed to verify the
(2012).
results. Scatter plots between selected PLs and
However, all of these models are appropriate for Swfor well No.1 is shown in Fig.1.
shaly–sand formation. Consequently, it seems to

Figure 1) Scatter plot of selected PLs versus water saturation in well No. 1. (a) As resistivity decreases, Sw
consistently increases. (b) DT has decreased in corresponding with increase in S w. (c) Increase in Sw
(hydrogen content) causes dropping of NPHI. (d) Increasing rate of RHOB, resulting in Sw increasing.

Mollajan and Memarian, 2013 157 Available online at http://jtethys.org

Journal of Tethys: Vol.1, No. 2, 156-163 ISSN: 2345-2471 ©2013

3–Methodology
( ) ∑ ( ) (1)
3–1–Radial Basis Functions
As is stated before, in this study the Gaussian
Radial Basis Functions Neural Network
function is employed as commonly used basis
(RBFNN) is an excellent tool for prediction or
function in Eq.1 ( ) and can be expressed as:
interpolation used as an alternative for MLP
neural networks. A particular RBFNN consists ‖ ‖
of three layers namely: input layer, hidden ( ) (2)
layer(s) and output layer. The input layer is a
where μj and σj are the center and width
buffer that presents data to the network and
parameter, respectively.
contains the input variables while the hidden
layer is composed of a number of RBF nodes In the learning process, the network is presented
with radial Gaussian activation functions. The with a pair of patterns and network computes its
output layer is the following layer in the own output. Afterward, the actual output is
network, which presents the output response to compared with target values or the desired
a given input and is connected to the previous output. So, the error can be calculated at any
nodes in the hidden layer by linear weights output in layer j as follow:
(Bishop, 1995).
E= yj– tj (3)
The Figure 2 shows the typical RBF neural
network architecture used in this study. In Where tj is the desired output and yj is the actual
Figure 2, x1, x2…, xN represents the number of output. The total error function is given by:
input nodesϕ1,ϕ2,ϕ3,....,ϕM represents the basis
∑ ∑ [ ( ) ] (4)
functionnodes,w0 is the weight of the bias node
(optional), and w1, w2,……, wM are connection
weights between hidden nodes and output node. 3–2– Empirical equations

As mentioned earlier, water saturation is

conventionally estimated using several
suggested equations which are exclusively
developed for shaly–sand formations. In this
study, Archie equation in addition to two
frequently used equations from Vsh models has
been proposed as predictors for Sw. Selected
equations are shown in Table 1.

In these equations, shale volume ( ) and

resistivity of water formation ( ) can be
obtained through the following equations,
respectively:
( ) (5)
Figure 2) Schematic representation of a radial basis
function neural network.
( ) (6)
The net output from hidden layer (y) will be:
where is the formation temperature and is
the salinity of water formation. For Dual–water
Mollajan and Memarian, 2013 158 Available online at http://jtethys.org
Journal of Tethys: Vol.1, No. 2, 156-163 ISSN: 2345-2471 ©2013

equation (Clavier et al., 1977), parameters are

defined as follow: * ( )+ (9)
It is noted that the value of Rsh can be calculated
( ) (7)
from resistivity logs at shaly intervals.
(8)

Table1) Different conventional predictors

Name Equation
Archie (1942)

Hossin (1960)

Dual Water (1977) [ ( )]

4– Implementation process presented. Minimum and maximum values of

the PLs used in the models along with their
In this section, the results obtained from the symbols are shown in Table 2.
above mentioned methodologies are briefly

Table 2) Description of input and output parameters in models.

Type of data Well log Symbol Min Max

Input Neutron NPHI 0.0056 0.31
Deep Resistivity LLd 2.89 196.79
Density RHOB 2.24 2.705
Sonic DT 162.401 279.19
Output Water Saturation SW 0.06 0.93
implementation process and results of each of
In this study, to validate the predictive models the models.
based on the comparing predicted and measured
values, R2 (Eq.10) is used. Also, RMSE (Eq.11) 4–1– Sw estimation using RBFNN
is used to compare the results of RBFNN and To predict the water saturation in carbonate
empirical models. intervals, the RBFNN model was used. In this
regard, four conventional logs including DT,
∑ ( ̅ )( ̅ )
[ ] (10) LLd, RHOB and NPHI were considered as
√∑ ( ̅ ) ∑ ( ̅ )
inputs and Sw measured from core analysis as
output of the models. In order to build the
( ) √ ∑ ( ) (11)
proposed model, the bellow steps have been
followed:
Where is the ith measured element,
is the ith estimated element, ̅ and ̅ are (1) Existing wells were divided into two
average of measured and estimated values, groups; one for model construction
respectively, and n is the number of data used. (including wells 1 to 3) and one for
examining the generalization
The following sections describe the
capability of the model (well No.4).

Mollajan and Memarian, 2013 159 Available online at http://jtethys.org

Journal of Tethys: Vol.1, No. 2, 156-163 ISSN: 2345-2471 ©2013

(2) Model construction data set randomly

were subdivided into two data sets
namely training data, with 70% of the
data points and testing data with the
remaining 30%.
(3) To check the generalization ability of
the models, data associated to the well
No 4 that were not entered in model
construction were used.

Figure 4) Results of empirical equations.

It was observed that Dual–water predictor is the

most appropriate predictor among the
conventional models, better fits the data and has
a greater determination coefficient whereas,
Archie predictor is the poorest one (Fig. 4).
Also, the values of estimated error for all
methods have been offered in Table 3.
According to this table, for Dual–water equation
Figure 3) Linear Regression results for RBFNN. the value of determination coefficient is the
highest one and its calculated RMSE is lower
The trial and error method was used to establish than other empirical equations that are
the number of hidden neurons. Using this considered.
strategy the optimal number of hidden neurons
Table 3) error values in employed equation
is the one that produces the least network error.
Model RMSE R2
Applying this method, the 4–10–1 network
Archie (1942) 0.042 0.78
structure is considered to be the desired Hossin (1960) 0.040 0.80
network. The radial basis function parameters Dual-water (1977) 0.042 0.81
(i.e. μ and σ) were obtained from normalized
input data using method of random sampling. 5–Discussion
Also, the weights are computed to minimize the
error function (Eq.4). The results of this study are shown in Figure 5
Figure 3 shows the results obtained from and Table 4.As can be seen, RBFNN model is
RBFNN model in well No.4. As seen, the closer to the measured Sw, whereas estimation
determination coefficient (R2) between by conventional equations has wide variation. In
predicted and measured Sw values and measured addition, these empirical models are appropriate
RMSE for this well are0.90 and0.031, for clean sands as well as shaly–sand formations
respectively. and are not efficient enough for carbonates due
to their specific properties as mentioned in
4–2– Sw estimation using empirical equations introduction.
To evaluate performance of the empirical
equations, the same datasets used for RBFNN
model were applied.

Mollajan and Memarian, 2013 160 Available online at http://jtethys.org

Journal of Tethys: Vol.1, No. 2, 156-163 ISSN: 2345-2471 ©2013

Figure 5) Comparison of the estimated Sw with measured.

According to the results, because of the very according to the obtained results, it is obvious
low shale volume, the results of empirical that there is no need to calculate the complex
relations are very close to each other. However, coefficients such as cementation factor,
the proposed model has considerably high tortuosity factor, saturation exponent and etc.
performance in estimation of Sw. It is concluded
that, at least in carbonate reservoirs using this Acknowledgments:
method for calculating water saturation is more
The authors would like to thank Dr. B.
appropriate. Tokhmechi and Dr.S.A. Ouadfeul for their kind
and careful comments that made the manuscript
Table 4) comparison between error values
improved. We also would like to express their
Model RMSE R2
sincere thanks to the Exploration Directorate of
Dual-water (1977) 0.042 0.81
RBFNN model 0.031 0.90
National Iranian Oil Company (NIOC) for
monetary supporting and their assistance in
providing data and information in this study.
6–Conclusions
References:
This paper presents a new approach based on
Archie, G.E. 1942. The Electrical Resistivity
Radial Basis Function Neural Network Log as an Aid in Determining Some
(RBFNN), for equations of water saturation Reservoir Characteristics. Transactions of the
from PLs in a carbonate reservoir. The proposed American Institute of Mechanical Engineers:
model was constructed by using data related to 146: 54–62.
three wells and its performance was examined
Balch, R.S., Stubbs, B.S., Weiss, W.W., Wo, S.
by a well which were not incorporated in the
1999. Using Artificial Intelligence to
model development. Also, to compare the Correlate Multiple Attributes to Reservoir
results of the RBFNN model with empirical Properties. Paper SPE. 56733. 10 pp.
equations, the same database was applied. It was
concluded that performance of the proposed Bishop, C.M. 1995. Neural Networks for
model is considerably better than the empirical Pattern Recognition. Oxford University
Press: pages 165–171.
models. For the RBFNN model R2 and RMSE
were equal to 0.90 and 0.031, respectively and Clavier, C., Coates, G., Dumanoir, J. 1977. The
for the best model of empirical models (Dual– theoretical and experimental bases for the “dual
water) were 0.81 and 0.042, respectively. Also, water” model for the interpretation of shaly
Mollajan and Memarian, 2013 161 Available online at http://jtethys.org
Journal of Tethys: Vol.1, No. 2, 156-163 ISSN: 2345-2471 ©2013

sands: Society of Petroleum Engineers Journal Simandoux, P. 1963. Dielectric measurements

24, 153–168. on porous media application to the
measurement of water saturations: study of
Hossin, A. 1960. Calcul des saturations en eau the behavior of argillaceous formations:
par la methode du cimentargileux Revue de l’Institut Francais du Petrole,
(formuled’Archiegeneralisee). Bulletin Supplementary Issue:18, 193–215.
Association Francaise des Techniciens du
Petrole. 140. Schlumberger.1998. Log interpretation
principles/Application. Texas: Seven
Kamel, M.H., Mabrouk, W.M. 2002. An printing,March sugar Land.
Equation for Estimating Water Saturation in
Clean Formations Utilizing Resistivity and Tillero, E. 2012. Stepping Forward: An
Sonic Logs: Theory and Application. Journal Automated Rock Type Index and a New
of Petroleum Science and Engineering: 36 Predictive Capillary Pressure Function for
159–168. Better Estimation of Permeability and Water
Saturation. Case Study, Urdaneta–01 Heavy
Kitamura, K., Masuda, K., Takahashi, M., Oil Reservoir SPE Latin America and
Nishizawa, O. 2006.The Influence of Pore Caribbean Petroleum Engineering
Fluids on Seismic Wave Velocities under Conference, 16–18 April 2012, Mexico City,
High Temperature and High Pressure Mexico.
Conditions: Development of a New
Technique with Gas Apparatus at AIST, Van Golf–Rach, T.D. 1982. Fundamentals of
Japan. Earth Planets Space:58, 1515–1518. fractured Reservoir Engineering, Elsevier.

Lucia, F.J. 2007. Carbonate Reservoir Waxman, M.H, Smits, L.J.M. 1968. Electrical
Characterization, an Integrated Approach, conductivities in oil–bearing shaly sands.
Second Edition, Springer. Society of Petroleum Engineers
Journal:8,107–122.
Mollajan, A., Memarian, H., Jalali, M.R.
2013.Prediction of reservoir water saturation Worthington, P.F. 1985.The evolution of shaly–
using support vector regression in an Iranian sand concepts in reservoir evaluation. The
carbonate reservoir. 47thUS Rock Mechanics/ Log Analyst:26, 23–40.
Geomechanics Symposium, San Francisco,
CA, USA, 23–26 June. Received: 07July2013 / Accepted: 10September2013 /
Published online: 17September2013
Nabil Al–Bulushi, N., King, P., Blunt, M.,
Kraaijveld, M. 2009. Development of EDITOR–IN–CHIEF:
artificial neural network models for
predicting water saturation and fluid Dr. Vahid Ahadnejad:
Payame Noor University, Department of Geology .PO
distribution. Journal of Petroleum Science BOX 13395–3697, Tehran, Iran.
and engineering:68, 197–208. E–Mail: [email protected]

Obeida, T.A., Al–Mehairi, Y.S., Suryanarayana, EDITORIAL BOARD:

K. 2005. Calculations of Fluid Saturations
from Log–Derived J–Functions in Giant Dr. Jessica Kind:
ETH Zürich Institut für Geophysik, NO H11.3, Sonnegg
Complex Middle–East Carbonate Reservoir, strasse 5, 8092 Zürich, Switzerland
E–J..Petrophysics:1, 1–9. E–Mail: [email protected]

Poupon, A., Levaux, J. 1971. Evaluation of Prof. David Lentz

water saturation in shaly formations University of New Brunswick, Department of Earth
Proceeding of SPWLA 12th Annual Logging Sciences, Box 4400, 2 Bailey Drive
Fredericton, NB E3B 5A3, Canada
Symposium. E–Mail: [email protected]
Mollajan and Memarian, 2013 162 Available online at http://jtethys.org
Journal of Tethys: Vol.1, No. 2, 156-163 ISSN: 2345-2471 ©2013

Dr. Anita Parbhakar–Fox E–Mail: [email protected]

School of Earth Sciences, University of Tasmania, Private
Bag 126, Hobart 7001, Australia Dr. Alexander K. Stewart
E–Mail: [email protected] Department of Geology, St. Lawrence University, Canton,
NY, USA
Prof. Roberto Barbieri E–mail: [email protected]
Dipartimento di Scienze della Terra e Geoambientali,
Università di Bologna, Via Zamboni 67 – 40126, Dr. Cristina C. Bicalho
Bologna, Italy Environmental Geochemistry, Universidade Federal
E–Mail: [email protected] Fluminense – UFF, Niteroi–RJ, Brazil
E–mail: [email protected]
Dr. Anne–Sophie Bouvier
Faculty of Geosciences and Environment, Institut des Dr. Lenka Findoráková
science de la Terre, Université deLausanne, Office: Institute of Geotechnics, Slovak Academy of Sciences,
4145.4, CH–1015 Lausann, Switzerland Watsonova 45,043 53 Košice, Slovak Republic
E–Mail: Anne–[email protected] E–Mail: [email protected]

Dr. Matthieu Angeli Dr. Mohamed Omran M. Khalifa

The Faculty of Mathematics and Natural Sciences, Geology Department, Faculty of Science, South Valley,
Department of Geosciences, University of Oslo Qena, 83523, Egypt
Postboks 1047 Blindern, 0316 OSLO, Norway E–Mail: [email protected]
E–Mail: [email protected]
Prof. A. K. Sinha
Dr. Miloš Gregor D.Sc. (Moscow), FGS (London). B 602,
Geological Institute of Dionys Stur, Mlynska Dolina, Vigyan Vihar, Sector 56, GURGAON 122011,
Podjavorinskej 597/15 Dubnica nadVahom, 01841, NCR DELHI, Haryana
Slovak Republic E–Mail: [email protected]

Mollajan and Memarian, 2013 163 Available online at http://jtethys.org