IPTC-19678-Abstract

Well Log Interpretation Using Deep Learning Neural Networks

Aarushi Gupta and Utkarsh Soumya, University of Petroleum and Energy Studies

Copyright 2020, International Petroleum Technology Conference

This paper was prepared for presentation at the International Petroleum Technology Conference held in Dhahran, Saudi Arabia, 13 – 15 January 2020.

This paper was selected for presentation by an IPTC Programme 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 International Petroleum Technology Conference and are subject to correction by the author(s). The
material, as presented, does not necessarily reflect any position of the International Petroleum Technology Conference, its officers, or members. Papers presented at
IPTC are subject to publication review by Sponsor Society Committees of IPTC. Electronic reproduction, distribution, or storage of any part of this paper for commercial
purposes without the written consent of the International Petroleum Technology Conference 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, IPTC, P.O. Box 833836, Richardson, TX 75083-3836, U.S.A., fax +1-972-952-9435.

Abstract
As artificial intelligence continues to revolutionize the industries around with innovations that make work
easy and fast, its’ use in the Oil and Gas Industry has been quite limited. With the technology available
to us, artificial intelligence ought to revolutionize the Oil and Gas Industry, mainly by time-saving and
increasing the ease of operation. This study uses deep learning to predict the lithology of the formations
encountered with the use of open-hole log data. It also works to calculate the porosity of the hydrocarbon
bearing formation using the density and neutron values. The trained program analyses and assimilates the
log values to indicate the lithology, reducing the time of the operation during the exploration stage. If used
as a real-time indicator, the program ought to ease the drilling operation. With drilling being one of the
costliest operations, any ease in the process that saved time would be a boon to the Industry. The paper gives
a summary of the technique used and discusses the work that could further strengthen the operation at hand.

Introduction
Well Logging is a method which is used to determine reservoir properties and identify various components
like oil, gas and water. Well logging methods are based on the fact that different rocks have different physical
properties. Well Logging methods include formation evaluations methods used to identify lithology and
reservoir property of various formations and production logging methods used for well operation control.
Currently more than thirty logging methods are available. Well logging technique include electric, magnetic,
radioactive, thermal, acoustic and other methods. Logging tools are run in the wellbore on wirelines. They
measure certain parameters depending on physical properties of rock such as electrical resistance. Analog
or digital signal from the tolls are sent to the surface by the wireline. At the surface they are recorded by the
special equipment. Electric methods measure formation resistance the electric current flowing from the rock,
because oil does not conduct electricity. The formation with oil shows high resistance while water saturated
parts shows low resistance thus oil and water zone can be identified which is called oil water contact.
The interpretation of these well log is a mammoth task which need highly experienced engineers to
interpret the readings measured by the logging tools and convert it into meaningful information. The
conventional methods for this process involves comparing the values of the logs which is varying with depth
to identify different lithology and reservoir properties, however these methods have some limitation as they
2 IPTC-19678-Abstract

are expensive and different wells in a field might have different physical properties and lithology which can
be a very time consuming task to process and identify using conventional methods. The classical methods
also face problems in determining accurately the relevant properties. Application of Artificial Intelligence
to interpret the logging data can provide much more accurate, efficient and economic method for well log
interpretation.
The use of Artificial Intelligence has led to revolutionary innovations over the past years. However, its use
in the Exploration and Production (E&P) Industry has been quite limited. Its application has only grown over
the years with the advent of modern drilling and production technologies. Since well log interpretation has a
log of scope of using artificial intelligence we propose the use of artificial neural network for identification
of lithology and calculation of porosity using the well log data.
This paper discusses the use Artificial Neural Network Tool to compute and analyse data collected during
the current logging methods. The program was, then, trained so as to analyse a well log data and hence, find
the best ways for identifying different lithology like shale, sandstone, carbonate or coals. The paper adds
another drop to the ever- developing ocean of knowledge and development in terms of Artificial Intelligence.
Having trained the software, it will help us reduce and overcome the challenges faced by us during every
well log interpretation. The expected result spectra shall vary starting from the initial drilling process to all
minor or major difficulties that we might come across when operating. We shall not only prevent a good
investment, but also reduce our effort by a good percent. Artificial Neural Networks have been particularly
successful in pattern recognition and phase behaviour prediction of petroleum engineering interests.
This work aims at designing a solution to well log interpretation problems where some data are not
immediately available. Embracing chemical, mechanical and other wellbore effects, it was expected to
address the shortcomings of uncertainty in the input data needed to run the analysis of efficient and robust
wellbore concerns.

Neural Network Approach

Neural Network forms of base of Deep Learning, a subset of machine learning which have an algorithm
inspired by human brain. Neural network takes in data, train themselves to recognise the patterns in this
data and then predict the output for a new set of similar data.
Neural networks are made up of layers of neurons, these neurons are core processing unit of the network.
First we have the input layer which receives the input, the output layer predicts the final output, in between
exists the hidden layers which performs most of the computations required by the network. Neuron of
one layer are connected to neurons of the next layer through channels, each of the channels is assigned a
numerical values known as weight. The input is multiplied with the corresponding weights and their sum is
sent as an input to the neurons to the hidden network. Each of these neurons is associated with a numerical
values called bytes which is then added to the input sum, this value is then passed to a threshold function
called the activation function.
The result of the activation function determines if the particular neuron will get activated or not. An
activated neurons transmit data to the neurons of the next over the channels and this manner the data
propagated through the network. This is called forward propagation. In the output layer the neuron will the
highest values finally determines the output. The values are basically a probability.
The network is trained to predict the correct output, during the training of the neural network the network
is feed also with the output along the input. The predicted output is compared with the actual output of the
training data to find out the error, the error is then feed back to the network known as backpropagation. Now
based on this the values of weights are adjusted, this cycle of forward-propagation and back-propagation is
iteratively performed with multiple inputs. This process continues until the weights are assigned such that
the networks predicts the output correctly.
IPTC-19678-Abstract 3

Figure 1—Feed Forward Neural Network Architecture

Methodology
Step 1: Dataset preparation
Before training the program, the first step is to normalize the dataset. Of all the methods available for
normalization, we choose the inverse method. The dataset is converted in an interval of 0 to 1.
For example, the gamma ray values are distributed in an interval of 25.45 to 114.60. For a value of 39.54,
normalization acts as:
Normalized value = (39.54 – 25.45)/ (114.60 – 25.45) = 0.158
Therefore, normalized value = (X – Xmin)/ (Xmax – Xmin)
This is used here because it easies the interpretation for the program. Neural networks are not frequently
used in the Industry as we need to first train them using field specific data for best results. When we use
normalized data coming from logs, the trend the logs follow remains unchanged on any field. Hence, it can
be used even in new projects that are currently under exploration stage.
For non-numeric dataset, the principle of numbering classes is used. According to this principle, integers
are assigned to this class of data. Here, the principle of numbering classes is used to denote lithology. In
this case, we have used the following denotion:

• Sandstone (water bearing) – 1

• Shale – 2

• Sandstone (oil bearing) – 3 Etc.

The data used is for the following logs:

1. Gamma Ray
2. Resistivity Log
4 IPTC-19678-Abstract

3. Neutron Log
4. Density Log
For porosity calculation, the program needs actual values from the neutron and density logs, hence, apart
from the normalized values from the above four logs, we also use the actual values of neutron and density
as inputs.

Step 2: The Program

Because of the program is able to model complex functions and decipher patterns, a multi-layer perceptron
is used with supervised learning. A forward feed network is created with six input neurons, the normalized
values from the gamma ray log (I1), the resistivity log (I2), the neutron log(I3), the density log (I4), and the
actual values from neutron log (I5) and the density log (I6); and two output neurons, the lithology (O1) and
the porosity (O2). The input neurons have variable weights defined in the limits [0, 1]. For the first output
I.e. lithology, the inputs I5 and I6 are assigned the weights equal to zero. For the second output, i.e. the
porosity, the weights for inputs I1 through I4 is zero. Porosity calculation is done only for the hydrocarbon
bearing zones.

Step 3: Training the program

The dataset is divided randomly such that 70% of the data is used for training while the rest is the test
dataset. The training of the program depends on three factors, namely, Learning rate, Number of Iterations,
and the Target Error Percentage. The more the training dataset, better the output results. When abundance
dataset is available, the program might memorize the values, instead of pattern recognition, especially in
case of porosity calculations. To avoid this, a part of the dataset may be used as validation dataset, using
which the error percentage suddenly increases if the training set is overwhelming.
The acceptable error percentage, in this case was set to 0.1% for O1, i.e. lithology and 0.01% for O2,
i.e. porosity. When tested for the first time, the program could not achieve the desired precision and hence,
making little changes to bias weights, the program was re-trained. The program was then re-tested. After
a few such attempts, a suitable bias weight value was assigned to the input neurons which lead to output
that met the desired precision expectations.

Results and Conclusion

In this study, the lithology of the formation is predicted using Gamma Ray log, Resistivity log, Neutron
log and density log. With the developed network, high accuracy results were obtained. This can help
petro-physicists in predicting the lithology, reducing the time invested in the entire process. This shall
be particularly helpful in the exploration stage where these estimates are of utmost importance. In the
development stage, this can be used with the LWD dataset and can help us locate the bit while drilling.
Overall, the time reduction would be value addition to the project.
With the porosity calculation for the hydrocarbon zone, the precision is dependent on the log data
precision and final porosity can be calculated as the weighted average value of all the obtained values.
The overall aim is to reduce the time of operation and also, to cut down the human error associated with
the project. With machines taking over most tedious jobs, this program would be able to emphasize the
importance of time and precision of the project.

Future Work
The work can be further improved by addition of other open hole log data and basic calculations made using
it. The use of SP Log, for example, can help us calculate volumes. Sonic log data can be used for porosity
calculations and verification.
IPTC-19678-Abstract 5

As the number of input and output neurons increase, the precision may be hindered, if the operator isn't
careful. As stated, the program might just memorize the data and corresponding results. For best results,
training is most important.

References
1. Liping Zhu, Hongqi Li, Zhongguo Yang, Chengyang Li, Yile Ao. Intelligent Logging
Lithological Interpretation With Convolution Neural Networks, Document ID- SPWLA-2018-
v59n6a5
2. Zhang Xiangjun, Li Youming, Liu Hon. Hydrocarbon Prediction Using Dual Neural Network,
Document ID- SEG-2000-1440
3. Michael P. McCormack (Arco Oil & Gas Co.). Neural Networks In the Petroleum Industry,
Document ID SEG-1991-0728
4. David A. Ford (Diamond Geoscience Research Corp.) | Michael C. Kelly (Diamond Geoscience
Research Corp.). Using Neural Networks to Predict Lithology From Well Log, Document ID-
SEG-2001-0666
5. Gerardo Garcia (Walt Whitman Software Inc.) | W.W. Whitman (Walt Whitman Software Inc.).
Inversion of a Lateral Log Using Neural Networks, DOI- https://doi.org/10.2118/24454-MS
6. Petroleum production engineering: a Computer-assisted approach. Elsevier Science
&Technology Books, Guo, B.; Lyons, W. C.; Ghalambor
7. Naseri et al., A neural network model and an updated correlation for estimation of dead crude oil
viscosity.
8. P. An (University of Manitoba, Canada) | W.M. Moon (University of Manitoba, Canada),
Reservoir Characterization Using Feedforward Neural Networks, Document ID-
SEG-1993-0258
9. Yi Guo (Center for Potential Fields Studies, Colorado School of Mines) | R.O. Hansen (Center
for Potential Fields Studies, Colorado School of Mines) | Norman Harthill (Center for Potential
Fields Studies, Colorado School of Mines), Artificial Intelligence I Neural Networks In
Geophysics, Document ID- SEG-1992-0001
10. Motaz Alfarraj (Center for Energy and Geo Processing (CeGP), Georgia Institute of Technology)
| Ghassan AlRegib (Center for Energy and Geo Processing (CeGP), Georgia Institute of
Technology), Petrophysical property estimation from seismic data using recurrent neural
networks, Document ID- SEG-2018-2995752
11. Kou-Yuan Huang (National Chiao Tung University) | Jiun-De Pissarenko (National Chiao
Tung University) | Kai-Ju Chen (National Chiao Tung University) | Hung-Lin Lai (National
Chiao Tung University) | An-Jin Don (National Chiao Tung University). Neural Network For
Parameters Determination And Seismic Pattern Detection. Document ID- SEG-2006-2285
12. Moustafa Elshafei (King Fahd University of Petroleum & Minerals) | Gharib M. Hamada (King
Fahd University of Petroleum & Minerals). Neural Network Identification of Hydrocarbon
Potential of Shaly Sand Reservoirs, Neural Network Identification of Hydrocarbon Potential of
Shaly Sand Reservoirs. DOI- https://doi.org/10.2118/110959-MS.
13. Al-Bulushi (Imperial College London) | M. Araujo (Shell Oil Company) | M. Kraaijveld
(Petroleum Development of Oman (PDO)) | X.D. Jing (Shell Technology Oman). Predicting
Water Saturation Using Artificial Neural Networks (ANNs). Document ID- SPWLA-
MERS-2007-W.
14. Odutola Toyin Olabisi (Department of Petroleum and Gas Engineering, University of Port
Harcourt.) | Ajienka Joseph Atubokiki (Department of Petroleum and Gas Engineering,
University of Port Harcourt.) | Ojedapo Babawale (Department of Petroleum and Gas
6 IPTC-19678-Abstract

Engineering, University of Port Harcourt.). Artificial Neural Network for Prediction of Hydrate
Formation Temperature. DOI https://doi.org/10.2118/198811-MS.
15. H. Tang (Louisiana State University). Improved Carbonate Reservoir Facies Classification Using
Artificial Neural Network Method. DOI- https://doi.org/10.2118/2008-122.
16. Chris Carpenter (JPT Technology Editor), Pseudodensity Log Generation by Use of Artificial
Neural Networks. DOI https://doi.org/10.2118/0517-0061-JPT.
17. Liping Zhu (China University of Petroleum-Beijing) | Hongqi Li (China University of Petroleum-
Beijing) | Zhongguo Yang (China University of Petroleum-Beijing) | Chengyang Li (China
University of Petroleum-Beijing) | Yile Ao (China University of Petroleum-Beijing). Intelligent
Logging Lithological Interpretation With Convolution Neural Networks. Document ID
SPWLA-2018-v59n6a5.
18. Yufeng Gu (China University of Petroleum, Beijing; PetroChina Research Institute of Petroleum
Exploration and Development) | Zhidong Bao (China University of Petroleum, Beijing;
PetroChina Research Institute of Petroleum Exploration and Development) | Zhenhua Rui
(Independent Project Analysis Inc.). Complex Lithofacies Identification Using Improved
Probabilistic Neural Networks. Document ID- SPWLA-2018-v59n2a9