Energy Sources, Part A: Recovery, Utilization, and

Environmental Effects

ISSN: 1556-7036 (Print) 1556-7230 (Online) Journal homepage: http://www.tandfonline.com/loi/ueso20

Using Well Logs to Predict a Multimin Porosity

Model by Optimized Spread RBF Networks

M. Baneshi, M. Behzadijo, M. Rostami, M. Schaffie & H. Nezamabadi-pour

To cite this article: M. Baneshi, M. Behzadijo, M. Rostami, M. Schaffie & H. Nezamabadi-

pour (2015) Using Well Logs to Predict a Multimin Porosity Model by Optimized Spread RBF
Networks, Energy Sources, Part A: Recovery, Utilization, and Environmental Effects, 37:22,
2443-2450, DOI: 10.1080/15567036.2011.628362

To link to this article: http://dx.doi.org/10.1080/15567036.2011.628362

Published online: 05 Nov 2015.

Submit your article to this journal

Article views: 14

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=ueso20

Download by: [University of Western Ontario] Date: 29 March 2016, At: 19:53
 Energy Sources, Part A: Recovery, Utilization, and Environmental Effects, 37:2443–2450, 2015
Copyright © Taylor & Francis Group, LLC
ISSN: 1556-7036 print/1556-7230 online
DOI: 10.1080/15567036.2011.628362

Using Well Logs to Predict a Multimin Porosity Model by

Optimized Spread RBF Networks

M. Baneshi,1,2 M. Behzadijo,2 M. Rostami,3 M. Schaffie,3 and H. Nezamabadi-pour3

Downloaded by [University of Western Ontario] at 19:53 29 March 2016

1
National Iranian Oil Company (NIOC), Tehran, Iran
2
Energy and Environmental Engineering Research Centre (EERC), Shahid Bahonar University of
Kerman, Kerman, Iran
3
Department of Oil and Gas Engineering, Shahid Bahonar University of Kerman, Kerman, Iran

An accurate porosity index prediction is one of the most important requirements in the oil industry. But
preparing and gathering of information is very time consuming and expensive and the evaluating
process needs highly skilled experts. Therefore, it is very economical to design a model that can predict
porosity with good accuracy, which takes less time and cost. In this study, about 10,000 accredited log
data and evaluated porosity (obtained by multimin model), which are related to one of the southern
fields of Iran, are available. Optimal neural networks structures for predicting porosity index have been
investigated. At first, neutron, sonic, and density logs were used as input variables. Half of the data was
used as a training set and other data were used as a test data set. By implementing various radial basis
function networks the best structure was determined and ultimately porosity index was predicted with a
high percentage of precision.

Keywords: multimin porosity model, optimal spread, radial basis function, well log

INTRODUCTION

New technologies are creating ways to reduce cost of exploration, production, and reservoir
management. Use of computers, employing expert models in decision, digital control, and display
equipment, that long ago were the newest technologies are now considered to be traditional
methods. Since the underlying layers are not visible, the ways to get information in the drilling
industry are sampling, geophysical, and well-testing operations. These operations are very expen-
sive and time consuming. Lack of information or inaccuracies make operations be faced with
plenty of problems and running all of the tests would not be cost effective. The majority of
engineers and programmers try to get adequate information from the environment using the
conventional and lower cost laboratory tests. Expert models can provide comprehensive informa-
tion about the underlying environment with minus tests (Baneshi et al., 2010).

Address correspondence to M. Baneshi, Energy and Environmental Engineering Research Centre (EERC), Shahid
Bahonar University of Kerman, Kerman, Iran. E-mail: [email protected]
Color versions of one or more of the figures in the article can be found online at www.tandfonline.com/ueso

2443
 2444 M. BANESHI ET AL.

Porosity is volume of empty space rather than the total volume of rock. More accurate
assessment of the reservoir can be obtained with having porosity and other parameters, such as
saturation and permeability. Using petrophysical tests and well logging data are common
methods in calculating the porosity index. Gathering all information is a very expensive and
time-consuming endeavor. Many formula computation and correlation charts are needed too.
By computer technology improvement and expanding its application in the oil industry, all
necessary correction operations and calculations can be done for a lot of data in a short time,
although the process is played as straight, final achieved error will be high (Baneshi et al.,
2013).
Finding the knowledge between log data can limit implementation of more expensive and
difficult logs in other places. Meanwhile, most of the time well log has not been implemented.
Downloaded by [University of Western Ontario] at 19:53 29 March 2016

Therefore, in these cases by using fewer logs one can estimate porosity. Artificial neural networks
are capable of finding the knowledge between log data and petrophysical factors of a reservoir.
These intelligent tools can find this knowledge for new data too. If log data obtained from several
wells and petrophysical tests’ information are carefully specified, well log data is charged to the
network as input data and formation factors as output data. Then, the network finds any knowledge
during the learning and training process. At last, the network is able to predict formation factors by
a good estimate every time new log data are charged to the network. Finally, without spending
much more money and time for providing petrophysical data in laboratories, formation factors are
obtained just using log data (Baneshi, 2014).
The main objectives of this study are as follows:
● To introduce a comprehensive relationship between petrophysical indexes and well logging;
● To introduce a comprehensive methodology for developing radial basis function neural
network models for predicting porosity index;
● To investigate the ability of ANNs to predict above factor and choosing input variables;
● To optimization of network’s structure and finding the best network;
● To compare the performance of an ANN model with that of other, more traditional models.
In the subsequent sections, an overview of the theory of artificial neural networks (ANNs) is
presented, followed by a description of the data available for model development. Next, the
development of the ANN model is described, as well as the results obtained and the conclusions
of the study.

COMPREHENSIVE RELATIONSHIP BETWEEN PETROPHYSICAL INDEXES AND

WELL LOGGING

Porosity can be defined as the ratio of voids to the total volume of rock (φ). The amount of internal
space or voids in a given volume of rock is a measure of the amount of fluid a rock will hold.
Although each produces a porosity value from basic measurements, none actually measures
porosity directly. Two such logs, the density and neutron, are nuclear measurements. A third log,
the sonic, uses acoustic measurements (Asquith and Krygowski, 2004).
When log data are ready, several petrophysical tests need to be done. Then, with the corporation
of laboratory experiments and well logging data, petrophysical factors are obtained. But this action
is very difficult and takes a lot of time. The log interpreter should notice each layer and correct the
log. Some corrections should be done in comparison with the kind of rock. After all of the
corrections and using several equations they can give petrophysical indexes, such as porosity
(Baneshi, 2015).
 PREDICTING A MULTIMIN POROSITY MODEL 2445

ARTIFICIAL NEURAL NETWORKS (ANNs)—AN OVERVIEW

The fundamental component in an ANN is called a neuron. Each artificial neuron is connected to a
certain number of other neurons, each of which has a changing degree of connectivity. The strength
of each connection is represented by a connection weight allocated to it. Learning in the ANNs is
brought about through training with examples or via exposing the system to an existing set of
input/output data. During the learning process the training algorithm updates iteratively the
connection weights and offers the appropriate result as output. Radial basis networks may require
more neurons than standard feed-forward back propagation networks, but often they can be
designed in a fraction of the time it takes to train standard feed-forward networks. There are two
variants of radial basis networks: generalized regression networks (GRNN) and probabilistic neural
Downloaded by [University of Western Ontario] at 19:53 29 March 2016

networks (PNN). Radial basis networks can be designed with either newrbe or newrb and GRNN
can be designed with newgrnn.
Each bias in the first layer is set to 0.8326/SPREAD. This gives radial basis functions that cross
0.5 at weighted inputs of +/– SPREAD. This determines the width of an area in the input space to
which each neuron responds. The only condition we have to meet is to make sure that SPREAD is
large enough so that the active input regions of the radbas neurons overlap enough so that several
networks with different spread should be investigated (Demuth and Beale, 2007).

DATA AVAILABILITY

The aim of the research presented in this article is to develop an intelligent system using ANNs to
predict porosity index for an Iranian onshore hydrocarbon basin. To achieve this purpose, an
extensive data set has been gathered from a number of wells. The data used in this article are: sonic
log, density log, and neutron log, and evaluated porosity index (in a range of 0.005 to 0.21). The
available data set covers about 8,000 data points collected.

RADIAL BASIS FUNCTION (RBF) NETWORK MODEL DEVELOPMENT—STAGES AND

METHODOLOGY

The first and foremost objective in designing any prediction model is to develop a model that
is capable of solving the problem under investigation to a desired degree of accuracy. In order
to accomplish this goal when developing a neural network model, a number of steps in the
model development process need to be considered. These steps for RBF networks, summarized
below, will be considered in developing an ANN model for the prediction porosity index, as
follows:
1. Determination of models inputs;
2. Data handling and preprocessing;
3. Data division;
4. Network architecture design (choosing the best spread);
5. Weight optimization (training);
6. Model validation.

Determination of Model Inputs

Understanding the influencing input parameters is of primary concern when developing ANN
models. Introducing more input parameters than required will result in a large network size
 2446 M. BANESHI ET AL.

and, consequently, decrease learning speed and efficiency. The determination of relevant
input parameters was achieved by train and error method or considering the economical
conditions for reducing the cost of data gathering and accurate answers simultaneously
(Baneshi, 2015).

Data Handling and Preprocessing

Before supplying the available data to the neural network, it is crucial to preprocess the data.
Data preprocessing helps to speed up the learning process and ensure that every parameter
receives equal attention by the network, improving overall network performance.
Preprocessing involves two fundamental elements: data scaling and data transformation. It is
Downloaded by [University of Western Ontario] at 19:53 29 March 2016

always necessary to scale the output and input parameters so that it is within the limits of the
activation function used in the output layer. In the current study, the available data have been
normalized into the range of 0 to 1.

Data Division
A final aspect is the ratio between the degree of freedom of the model and the number of data
points in the training set. If the number of model parameters (weights and bias) is large compared
with the number of data points in the training set, the model can easily lose the ability to
generalize, which might lead to over fitting of the training data set.

Network Architecture Design

Once the available data have been preprocessed and split into representative subsets the network
architecture has to be determined. The literature gives no universal guideline for determining the
optimal spread.

Model Validation
Validating of the ANN model is the final stage in the development sequence. An independent
data set is used for this purpose, in order to check whether the model has the ability to
generalize or has just memorize the specific features contained in the training data. If the
model is believed to perform adequately on the validation set, it is then ready to be used in
“real-world” situations. To facilitate assessment of the model’s generalization ability, a number
of error measures may be considered. The most popular set of error measures are the set from
Shahin et al. (2002).

RESULTS AND DISCUSSION

At first it is determined that using just three logs data are enough to accurately predict porosity and
reduce the cost of providing necessary information. These logs are used to evaluate porosity by a
common method, which is deterministic. For example, it was not necessary to use resistance or
gamma ray logs because this log did not have any specific correlation with porosity. For each
network, first the data was normalized. Then, half of them were chosen as training data, and half as
testing data. The networks used in this article are radial basis functions with Newrb and Newrbe
functions and generalized regression networks.
Several networks were run to find the best network containing the smallest error. The structures
of them were similar to each other but they had a different spread. Table 1 shows networks with
 PREDICTING A MULTIMIN POROSITY MODEL 2447

TABLE 1
Comparison of GRNN and RBF Networks with Newrb, Newrbe Functions, and
Different Spread for Predicting Porosity by Using DT, RHOB, and NPHI Logs
as Input Variable

Test Error (%) Train error (%) Spread

Newrb
130 141 0.01
23.22 21.56 0.5
10.141 7.015 1
9.208 4.682 1.25
8.788 3.918 1.5
Downloaded by [University of Western Ontario] at 19:53 29 March 2016

8.276 3.156 2
7.54 3.494 2.5
8.837 4.541 3.5
12.63 8.96 5
35.54 29.36 10
Newrbe
321 1.47 0.5
95.12 1.872 1
24.939 1.964 2
17.869 2.207 3
9.854 2.389 5
9.746 2.488 10
9.046 3.549 25
10.416 4.571 50
12.23 5.94 100
18.36 7.42 200
GRNN
75.69 0.015 0.01
13.748 0.806 0.04
12.916 1.247 0.05
9.314 1.782 0.07
12.517 2.234 0.09
13.622 3.887 0.1
16.285 9.414 0.2
18.859 23.79 0.5
33.279 50.873 1
67.88 101.5 2

different function and spread. Minimum square error and relative error percentage is written in this
table. The networks try to learn the training data set. After that, a simulation process is done by
simulating training and test data set. In each state, the error of each set is plotted versus spread.
Figures 1 to 3 show the error of training and test data versus different spread.
After the simulation, the graph of trained network for test data set is drawn. In this graph
network outputs are plotted versus the targets as open circles. The best linear fit is indicated by a
dashed line. The perfect fit (output equal to targets) is indicated by the solid line. The ratio between
these two lines is shown in Figure 4. Any number that is closer to 1, shows that the network’s
output values are more real.
In the present article, relative error percentage and minimum square error are used to compare a
network’s output and target data.
 2448 M. BANESHI ET AL.

GRNN
100

Train set data

80
Relative error percentage

Test set data

60

40
Downloaded by [University of Western Ontario] at 19:53 29 March 2016

20

0
0.01 0.1 1 10
logarithmic Spread

FIGURE 1 Error of training and test data sets vs, Spread (GRNN network).

Newrbe function
8 1000
logarithmic Relative error percentage
7 Train set data
Relative error percentage

6 Test set data

100
5
4
3
10
2
1
0 1
0.1 1 10 100 1000
logarithmic Spread

FIGURE 2 Error of training and test data sets vs. Spread (RBF network with Newrbe function).

The error percentage and minimum square error of each output set data are written in tables.
And the best minimum square error is equal to 0.000027 for test data set. Another way to compare
output data and target data is to plot both of them versus depth as it is shown in Figure 5.

CONCLUSION

Comparison between targets and a network’s output data shows that radial basis function networks
are suitable tools for predicting reservoir parameters, such as porosity, consuming less time, and
reducing cost.
 PREDICTING A MULTIMIN POROSITY MODEL 2449

Newrb function

logarithmic Relative error percentage

100
Train set
data
Test set data

10
Downloaded by [University of Western Ontario] at 19:53 29 March 2016

1
0 2 4 6 8 10
Spread

FIGURE 3 Error of training and test data sets vs. Spread (RBF network with Newrb function).

FIGURE 4 In this graph network outputs are plotted vs. the targets as open circles. The best linear fit is indicated by
a dashed line. The perfect fit (output equal to targets) is indicated by the solid line. The ratio between these two lines
is shown as R.

At last, in this article, by using just three logs as DT, NPHI, and RHOB that have more reliable
information, an important reservoir parameter was predicted well. Therefore, preparing of data for
interpretation will be cheaper and faster.
Two points are important in good performance of a network. One of them is selection of input
variable. In this article extra logs, such as gamma ray, depth, caliper, etc., were not used because
 2450 M. BANESHI ET AL.
Downloaded by [University of Western Ontario] at 19:53 29 March 2016

FIGURE 5 Comparing networks output data (ANN.PREDICTED) and targets data for test data.

extra logs reduce the performance and increase the cost. The second point is choosing a generalized
network by finding a suitable structure.

ACKNOWLEDGMENT

The authors would like to thank the National Iranian Oil Company (NIOC) for supporting this
study.

REFERENCES

Asquith, G., and Krygowski, D. 2004. Basic Well Log Analysis. Tulsa, Oklahoma: American Association of Petroleum
Geologists.
Baneshi, M. 2014. Determination of lithofacies using an optimized neural network and well log data. Pet. Sci. Technol.
32:897–903.
Baneshi, M. 2015. Using ANFIS and neural networks to predict the volume percentage of matrix and fluid. Energ. Source.
Part A (37:2247–2258).
Baneshi, M., Schaffie, M., Nezamabadi-pour, H., and Behzadijo, M. 2010. Reservoir petrophysical index prediction using
neural network with selection suitable feature. National Conference of Energy and Environment, Kerman, Iran, October
10–13.
Baneshi, M., Schaffie, M., Nezamabadi-pour, H., and Behzadijo, M. 2013.Predicting log data by using artificial neural
networks to approximate petrophysical parameters of formation. Pet. Sci. Technol. 31:1238–1248.
Demuth, H., and Beale, M. 2007. Neural Network Toolbox for Use With MATLAB. User’s Guid 6th ed. Natick, MA: The
Math Works, Inc.
Shahin, M. A., Maier, H. R., and Jaksa, M. B. 2002. Predicting settlement of shallow foundations using neural networks. J.
Geotech. Geoenviron. Eng. 128:785–793.