Published Online March 2015 (http://www.cuet.ac.bd/merj/index.

html)
M
E R
Mechanical Engineering Research Journal J

Dept. of Mech. Eng.

Vol. 9, pp. 100103, 2013
ISSN: 1990-5491
CUET

PERMEABILITY MODELING FROM WELL LOGS USING

ARTIFICIAL NEURAL NETWORKS
M. Z. B. Sultan1* and M. F. Howladar2

1
Department of Petroleum and Mining Engineering, Chittagong University of Engineering and Technology,
Chittagong-4349, Bangladesh
2
Department of Petroleum and Mining Engineering, Shahjalal University of Science and Technology,
Sylhet-3114, Bangladesh

Abstract: Permeability, the ability of a rock to transmit fluid, is one of the most important properties of
hydrocarbon bearing formations. Formation permeability is generally measured in laboratory from cores. It can
also be evaluated from well test data. This study discussed a rigorous approach for permeability determination from
well logs of the Habiganj Gas Field (HGF). This is a “Virtual Measurement” method and makes use of Artificial
Neural Networks (ANN). ANN can predict permeability values for entire wells without prior exposure to their log
and core data with accuracies. The ability of neural networks to learn from experience and then to generalize these
learning for solving new problems sets it apart from all conventional methods. One developed well of HGF was
considered under this study. The permeability model was developed from the data set consisting of core
permeability and logs from the well. The model was then applied to that well and the result (i.e. predicted
permeability) was compared with the core permeability.

Keywords: Permeability, ANN, Well log, HGF

NOMENCLATURE bo = Bias weights of the input layer

K = Permeability
ANN = Artificial Neural Network
n1 = Dimension of the input vector
BD = Bulk density log
n2 = Number of hidden neurons
BGFCL = Bangladesh Gas Field Company Ltd.
w = Connection weight of the hidden layer
BPNN = Back Propagation Neural Network
wij = Weight for the i neuron in the hidden layer j
DI = Deep induction log
woj = Bias weights of the hidden layer
GR = Gamma ray log
xi = Input variables
HGF = Habiganj Gas Field
y = Output variable
HBJ#7 = Well No. 7 of Habiganj Gas Field
X = Input vector of one dimension for any input variable
MVR = Multiple Variable Regression
Y = Output variable
Neu = Neutron log
= Corrected log porosity
PMML = Predictive Model Markup Language
por = Corrected log porosity
SPSS = Statistical Package for the Social Sciences
1. INTRODUCTION
SNNS = Stuttgart Neural Network Simulator
Sonic = Sonic compressional transit time There are several ways to measure formation permeability
XML = Extensible Markup Language [1,2]. Three methods are available to determine permeability
b = Connection weight of the input layer from well logs namely Empirical, Multiple Variable Regression
(MVR) and Artificial Neural Networks (ANN) [35].
*
Corresponding author: E-mail: [email protected]
 M. Z. B. Sultan and M. F. Howladar /Mech. Eng. Res. Journal, Vol. 9 (2013) 101

Empirical models were developed based mainly on porosity and The output, y, depends on the particular transfer function
water saturation of a formation. Thus it is often limited to that is chosen. All the values of the input variables and target
provide accurate estimation of permeability [6]. To overcome variable are normalized or scaled in the range of (–1, 1). In this
such limitation MVR was introduced and frequently used to study, the input variables are different well logs ( , GR, Sonic,
develop permeability model from well log [7]. But MVR is not Neu, BD, DI) and the output variable is the target which is
able to predict permeability for entire domain of interest. ANN permeability (K).
shows the consistency to predict permeability from well logs Neural network simulators are software applications that
[8,9]. are used to simulate the behavior of artificial or biological neural
networks. They focus on one or a limited number of specific
2. ARTIFICIAL NEURAL NETWORKS types of neural networks. They are typically stand-alone and not
intended to produce general neural networks that can be
Artificial Neural Networks (ANN) is a computational tool
integrated in other software. Commonly used artificial neural
modeled on the interconnection of the neuron in the nervous
network simulators include the Stuttgart Neural Network
systems of the human brain and that of other organisms. ANN is
Simulator (SNNS), Emergent, Java NNS and Neural Lab. The
the type of non-linear processing system that is ideally suited for
majority implementations of neural networks available are
a wide range of tasks, especially tasks where there are no
however custom implementations in various programming
existing algorithm for task completion. ANN can be trained to
languages and on various platforms. Basic types of neural
solve certain problems using a teaching method and sample data.
networks are simple to implement directly. There are also many
In this way, identically constructed ANN can be used to perform
programming libraries that contain neural network functionality
different tasks depending on the training received. With proper
and that can be used in custom implementations. A common
training, ANN is capable of generalization, the ability to
language is necessary for neural network models to be shared by
recognize similarities among different input patterns. ANN is
different applications. Recently, the Predictive Model Markup
essentially simple mathematical model defining a function f:X
Language (PMML) has been proposed to address this need.
→Y or a distribution over X or both X and Y, but sometimes
PMML is an XML-based language which provides a way
models also intimately associated with a particular learning
for applications to define and to share neural network models
algorithm or learning rule. One of the more popular algorithms is
(and other data mining models) between PMML compliant
the back-propagation algorithm [10,11].
applications. PMML applications provide a vendor-independent
A typical back propagation neural network (BPNN) is
method of defining models so that proprietary issues and
composed of three layers namely input, hidden and output layers.
incompatibilities are no longer a barrier to the exchange of
Each layer has a number of processing elements which are called
models between applications. It allows users to develop models
neurons. Neurons of different layers are connected together by a
within one vendor's application, and use other vendors'
simple weighted link (Fig. 1).
applications to visualize, analyze, evaluate or otherwise use the
models. Previously, this was very difficult, but with PMML, the
Hidden Layer exchange of models between compliant applications is now
straightforward. A range of products are being offered to
produce and consume PMML. SPPS is one of them and was
Input Layer preferred to develop permeability model.
Output Layer
3. RESERVOIR DESCRIPTION
Habiganj Gas Field (HGF) is located in the north-eastern
part of Bangladesh. This area belongs to Madhobpur Upazilla
under the District Habiganj & Division Sylhet. Habiganj
structure lies in the north-eastern part of Bengal Foredeep and in
the south central part of Surma Basin. It is the northern
culmination of Baramura anticline of Tripura (India), is
separated by a saddle. The structure is almost symmetrical and
Fig. 1: Typical Neural Network Architecture. trending SSE. Seismic and well data of Habiganj structure have
proved that the structure has entirely separate closure. Including
BPNN uses the following mathematical function: Baramura it is a giant structure of about 130 km long. In
(1) Habiganj structure, the exposed rock is Dupi Tila and the
topographical height is about 20 m above mean sea level. No
where y is the output variable, xi are the input variables, b and w fault has been detected on flanks.
are the connection weights, n1 is the dimension of the input Habiganj structure has one of the largest producing gas
vector, n2 is the number of hidden neurons, b0 and woj are called fields of the country. This field comprises two sands namely:
the bias weights. Upper Gas Sand (1320–1530 m) in Bokabil and Lower Gas Sand
102 M. Z. B. Sultan and M. F. Howladar /Mech. Eng. Res. Journal, Vol. 9 (2013)

(3014–3022 m) in Bhuban formation. The upper structure is

about 10 km long towards North-South and 5 km wide in the 4.2 Application to the HGF
direction of East-West. In the crest, reservoir has very good Core and log data from HBJ#7 were used to construct the
permeable sand. The Upper Gas Sand is fine grained, clean, network model. A total of 31 core measurements for porosity
unconsolidated and well sorted. Lack of clay matrix and and permeability and their corresponding well logging responses
presence of glauconite in Upper Gas Sand indicate the were available for network training and testing. The well log
deposition in high-energy littoral to sub littoral environment. responses that have been used include bulk density (BD),
The sand indicates a beach and barrier bar characteristic. gamma ray (GR), sonic compressional transit time, neutron
porosity (Neu), and deep induction (DI). The permeability
4. DEVELOPMENT OF PERMEABILITY MODEL model is given by

For developing the permeability model using ANN, K = f ( , GR, Sonic, Neu, BD, DI) (2)
information about the field were collected. Required data were
selected to be used in the ANN model. Section 4.1 describes the 23 data samples were chosen by a random number
data used in ANN model. The developed permeability model is generator for network training. The remaining 8 samples were
given in section 4.2. put aside to be used for testing the network’s integrity and
robustness. The available data were randomly divided into two
4.1 Data Description groups. The first group was used for the process of network
Permeability is usually modeled using ANN based on the training which represent 74.2% of the total sample points, and
core and log data. Core data of only well no. 7 of HGF (HBJ#7) the rest 25.8% were used for network testing. Finally, the
for an interval of 1500–1518 m depth were available during this training process data were divided into validation process (30%)
study. Fig. 2 shows log and core data of HBJ#7. Permeability and forward training process (70%). The final architecture of the
modeling requires the following well logs – neural network to predict permeability contained six input
 Density log variables and one hidden layer with 2 neurons (as shown in Fig.
 Neutron log 3). This configuration and the proper use of the validation set
 Sonic log were sufficient to ensure fast convergence after about 60
 Gamma ray log and iterations.
 Resistivity log

Fig. 2: Log and core data of HBJ#7 (courtesy: BGFCL).

 M. Z. B. Sultan and M. F. Howladar /Mech. Eng. Res. Journal, Vol. 9 (2013) 103

The process can help engineers to acquire a better handle on the

5. RESULTS AND DISCUSSION reservoir characteristics. This is especially beneficial for the
fields which have many newly drilled producing wells.
ANN has a correlation coefficient of 0.813 (Fig. 4), where
1.0 is a perfect match. Permeability models need both core and REFERENCES
log data for several wells to give a consistent result. But, in this
study permeability model was developed based on only one well [1] U. Ahmed, S. F. Crary and G. R. Coates, “Permeability
data. The fact that core data of only one well of HGF was Estimation: The Various Sources and Their
available during this study. This model can be updated using Interrelationships”, Journal of Petroleum Technology, Vol.
more than one well data whenever core data of the wells of 43, pp. 578-587, 1991.
interest are available. It can be done such a way that the n (where [2] C. Y. Yao and S. A. Holditch, “Estimating Permeability
n = 2, 3, 4 …) developed well of HGF will be considered under Profiles Using Core and Log Data”, SPE 26921, Eastern
this study. The permeability model will be developed from the Regional Conference, Pittsburgh, Pennsylvania, USA, pp.
data set consisting of core permeability and logs from n-1 well. 317322, 1993.
The model will then be applied to the nth well. However, the [3] A. K. Verma, B. A. Cheadle, A. Routray, W. K. Mohanty
permeability model based on HBJ#7 can be applied in other and L. Mansinha, ‘Porosity and Permeability Estimation
wells of this field to determine permeability from well logs. using Neural Network Approach from Well Log Data’,
2014. [Online] Available :
Bias http://www.searchanddiscovery.com/pdfz/documents/201
4/41276verma/ndx_verma.pdf.html (February 25, 2014)
por
[4] E. M. EL-M. Shokir, A. A. Alsughayer and A. Al-Ateeq,
“Permeability Estimation from Well Log Responses”,
GR Bias
Journal of Canadian Petroleum Technology, Vol. 45, pp.
Sonic H(1:1) K 4146, 2006.
[5] B. Balan, S. Mohaghhegh and S. Ameri, “Formation
GR H(1:2)
Determination from Well Log Data”, SPE Formation
Evaluation, pp. 170174, 1997.
BD
[6] G. E. Archie, “The Electrical Resistivity Log as an Aid in
DI Some Reservoir Characteristics”, SPE Journal, Vol. 146,
pp. 5462, 1942.
Fig. 3: SPSS generated Neural Network Architecture. [7] N. R. Draper and H. Smith, Applied Regression Analysis,
3rd ed., pp. 217232, John Wiley & Sons, New York, 1998.
Correlation coefficient = 0.813
[8] S. Mohaghegh, A. Reza, S. Ameri and D. Rose, “Design
and Development of an Artificial Neural Network for
Estimation of Formation Permeability”, SPE Journal, Vol.
7, pp. 151154, 1995.
[9] S. Mohaghegh, A. Reza, S. Ameri and M. H. Hefner, “A
Methodological Approach for Reservoir Heterogeneity
Characterization Using Artificial Neural Networks”, SPE
28394, SPE Annual Technical Conference and Exhibition,
New Orleans, Louisiana, pp. 337346, 25-28 September,
1994.
[10] M. Smith, Neural Networks for Statistical Modeling, 1st
ed., pp. 113132, Van Nostrand Reinhold, New York,
Fig. 4: ANN permeability vs. core permeability of well HBJ#7. 1993.
[11] J. E. Dayhoff, Neural Network Architectures: An
6. CONCLUSIONS Introduction, 1st ed., pp. 83112, Van Nostrand Reinhold,
New York, 1990.
A rigorous methodology has been presented here which has
the potential to significantly reduce the cost of permeability
determination from well logs. This methodology uses the
conventional well logs and generates predictive model for all the
wells in a field. The development process requires a handful of
wells in a field. Then the log data generated from those wells are
used to develop an intelligent, predictive permeability model.