SPE-180439-MS

Pseudo Density Log Generation Using Artificial Neural Network

Wennan Long, University of Southern California; Di Chai, University of Kansas; Fred Aminzadeh, University of
Southern California

Copyright 2016, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Western Regional Meeting held in Anchorage, Alaska, USA, 2326 May 2016.
This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers 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 SPE copyright.

Abstract
Reservoir characterization is often a demanding and complicated task due to the nonlinear and heterogeneous physical properties of the subsurface environment. Those issues can be overcome accurately and
efficiently by the use of computer-based intelligence methods such as Neural Network, Fuzzy Logic and
Genetic Algorithm. This paper will describe how one integrates a comprehensive methodology of data
mining techniques and artificial neural network (ANN) in reservoir petrophysics properties prediction and
regeneration. Density log, which acts as a powerful tool in petrophysical properties indication, is often run
over just a small portion of the well due to economic considerations, the borehole environment or
operation difficulties. Furthermore, missing log data is common for old wells, and wells drilled by other
companies. Working towards a resolution to these challenges, we will demonstrate successfully constructed automatic system which includes well logging data preprocessing, data mining technologies and
ANN prediction. Based on one field case study, this methodology was proficient and stable in pseudodensity log generation.

Introduction
In reservoir engineering, well log involves three significant aspects reservoir identification, oil-inplace (OIP), and recoverable oil in place. Each is a key element of the reservoir description; efficiency is
lost when the well log information is missing or incomplete. The classic method of handling this problem
is to re-log the well using various logging tools. However, most openhole log measurements are very
difficult to obtain once a well has been cased, which leads to unnecessary economic issues, especially
when shutting down the well is inevitable. Also, any well log data mismatch, (the discrepancy between
the original and re-logged data), due to petrophysical properties changes, can cause serious problems
when estimating reservoir characteristics, such as porosity and permeability from those well logs.
Consequently, several alternative technologies have been applied to generate pseudo well logs since 1980.
ANN, a machine-learning model that provides the potential to establish a multidimensional, nonlinear
and complex model, can be a powerful tool with which to analyze experimental, industrial and field data
(Aminzadeh, 1996). The task of predicting petrophysical properties such as porosity and permeability
dramatically affect the results of reservoir characterization. Using multivariate statistical analysis and
nonparametric regression, electrofacies were characterized and permeability was predicted in carbonate

SPE-180439-MS

reservoirs (Lee et al., 1999). Associated with fuzzy logic, intelligent software was developed (Nikravesh
and Aminzadeh, 2001a) that is able to identify the nonlinear relationship between seismic attributes and
well logs. Distributed in a spatially nonuniform and nonlinear geologic manner, permeability and porosity
are estimated from well log data using ANN, and validated using core data (Verma et al., 2012). In
addition, the probabilistic-neural-network (PNN) facies analysis was used to predict facies from well log
data, as it has the potential to delineate the nonlinear relationship between lithofacies and well log data
(Tang et al., 2011). However, a high degree of unknown uncertainty somehow has a great impact on the
predicted result. In order to quantify uncertainty, one study introduced Adaptive Neuro Fuzzy Inference
System (ANFIS) to predict porosity and sand fraction value from well logs, and solidify its advantage over
ANN (Chaki et al., 2013).
Because of strong sensitivity of input data in ANN, we propose to screen input. For example, different
rocks have different petrophysical responses. It is crucial to find the optimal data from one well to build
the model using ANN for pseudo well log generation of a target well. Thus, the pattern recognition is
inevitable. Manual stratigraphic interpretation is regarded as one approach. However, it is both labor
intensive and time consuming to identify the patterns of well logs. Data mining techniques is another
applicable approach to automatically process data associated with nonlinearity by using a statistical
method to discover the data patterns (Nikravesh and Aminzadeh, 2001b). One application in petrophysics
is facies (or electrofacies) classification, which is widely used to divide well log data in order to obtain
target information, such as porosity-permeability relation (Nikravesh et al., 1999). An adjunct to artificial
intelligence, clustering analysis can determine electrofacies and categorize lithological profile quite
efficiently (Hassibi and Ershaghi, 1996). Nikravesh and Aminzadeh (2003) explain how the reservoir
characterization is solved by mining and fusing of reservoir data in an intelligent system built by soft
computing techniques. Other applications, such as permeability estimation, have been easily proceeding
under cluster analysis, multivariate analysis, classification trees and feedforward neural network (Kumar
et al., 2006). For additional details on the use of neural networks and other soft computing techniques in
the petroleum industry see Aminzadeh and de Groot, (2006).
Porosity, one of the more important petrophysical properties, can be obtained from density log. The
essence of this paper is density log generation, so in this study, a three step approach was produced. First,
we applied preprocessing the log data using standardization, dimension reduction (Principal Component
Analysis). Second, we proposed clustering (Model Based Clustering) to recognize the specific pattern and
interpret stratigraphic information. Finally, we chose a similar pattern as input to generate target
pseudo-density log using ANN. A field containing eight wells was utilized as the case study to affirm and
prove the stability, efficiency and accuracy of merging data mining with ANN.

Workflow & Methodology

The three step approach will be shown in a workflow (Figure 1). It comprises three components: data
preprocessing, data mining and data postprocessing. Data preprocessing contains normalization and
principal component analysis (PCA). In the data mining part, clustering using Gaussian mixture model
(GMM) is applied to give a 2-D qualitative indication of prediction performance across two different wells
in a block. Model based clustering (MBC) is then used to fully merge the data of two good correlated
wells, selected from the previous step, into separate electrofacies, which have some inner correlations with
lithofacies. If a large dataset is used for clustering, which is a time consuming process, the discriminant
analysis (DA) is taken into account. The primary purpose of DA is to overcome the problem of long
computation time and high memory demand when MBC is applied to a large dataset. In a comparison of
electrofacies graphs between two wells for validation, the best related wells were chosen if similar trends
were observed. Finally, in the data postprocessing (pairwise well prediction), back-propagation neural
network (BPNN) is used to train and predict actual density log in two wells.

SPE-180439-MS

Figure 1Systematic pseudo density log generation workflow

Well Log Data Preprocessing

Data preprocessing is the first stage of systematic pseudo-density log generation. It aims to normalize well
log data and reduce the data size. It is of paramount importance based upon the facts that: (a) the quality
of data has a significant impact on prediction results through ANN; and (b) the quantity of data will
substantially increase the efficiency of the prediction.
Normalization
Well log data is constructed in a matrix form, whereby each row represents the depth record, and each
column is the different type of well logs. Well log matrix (X) is a m n dimension, and the number of
rows m is much larger than that of the column n. Each well constructs one well log matrix and one field
which has multiple wells constructs a big dataset. The initial step in the first stage is to normalize well
log data. This normalization is necessary because different types of well log data have different units. For
instance, the unit of Spontaneous Potential Log is millivolt, whereas the unit of Gamma Ray is API unit.
Following is the normalization expression:
(1)
where xi,norm is the norm of attribute i; i and i are mean and standard deviation of attribute i; ||xi||
is L-infinity norm to scale well log data in [-1,1].

SPE-180439-MS

Principal Component Analysis

The principal component analysis is a statistical procedure that uses an orthogonal transformation to
convert a set of well log vectors (attributes) of possibly correlated variables into a set of values of linearly
uncorrelated variables called principal components (PC) (Aminzadeh et al., 2000). Then, principal
components constructed a new matrix which has much lower dimensions than the original. In this paper,
we propose to use singular value decomposition (SVD). We consider the SVD of well log covariance
matrix: XTX UVT, where is a diagonal matrix (r r) with the r non-null singular values (
diag(d1,d2, . . .,dr)) sorted in decreasing order (d1 d2 . . . dr 0). The column vectors in Unr and
in Vrn are linearly independent of each other, which can be expressed as UTU VTV I. Hence, PCA
allows the reduction of the dimensionality (number of the columns) of the well log data but retains most
of the variability of that data. Hence, the principal components of well log matrix are generated based
upon the following formula:
(2)
where XPCA is mr matrix and r n. In the case study, we chose k number of principal components,
where k is much smaller than r, and we have proved that k is an optimal number (Appendix). Figure 2
shows scatter plot of PC 1, PC 2 and PC3 from Well #1.

Figure 2Scatter plot of pricinpal components of Well #1

Well Log Data Mining

Well logs can be described as a record of rock and formation properties against depth. They play an
important role in understanding the petrophysical properties of the reservoir rocks. Well logs, however,
are not a straightforward representation of formation, and the amount of well log data is usually extremely
large. In addition, stratigraphic interpretation and classification are necessary to select appropriate wells
for data postprocessing. Data mining is the computation process of discovering patterns in large data sets
involving methods at the intersection of artificial intelligence, machine learning, statistics and database
systems (Figure 3). It, therefore, develops into an ideal tool to take the place of conventional manual
analysis, and becomes a keystone in the automatic pseudo-density log generation system.

SPE-180439-MS

Figure 3Data mining technologies

Lithofacies & Electrofacies

Lithofacies is a body of rock with specified characteristics. It is a distinctive rock unit that forms under
certain conditions of sedimentation, reflecting a particular process or environment and petrophysical
properties of that rock unit. Rock characteristics differ from one to another; therefore, different types of
lithofacies have an internal different signal response on well logs. For example, gamma ray log of
sandstone can reflect the clay content. If we use gamma ray log of sandstone to build a model, and
subsequently use that model to predict gamma ray log of shale, severe mismatch problems arise that will
ensure prediction failure. So it is of paramount importance to have correct lithofacies. Lithofacies
identification is classified as three general approaches:

Core data analysis

Knowledge-based well log analysis by the expert system
Electrofacies

Compared with the first and second methods, electrofacies analysis is relatively inexpensive and more
efficient. The term electrofacies was first introduced by Serra and Abbott (1980). It has been defined as
the set of log responses which characterize a bed (layer) and permits this to be distinguished from others.
Electrofacies are not equal to lithofacies, but they have a correlation with each other. Figure 4 illuminates
this inner correlation. Well log data is determined by lithology to describe formation petrophysical
properties. Then lithofacies are manually obtained by well log data based upon experience, while
electrofacies are calculated based upon statistical analysis. Finally, electrofacies hereby have a correlation
with lithofacies.

SPE-180439-MS

Figure 4 Correlation between lithofacies and electrofacies

Gaussian Mixture Model

We assume one well log matrix is one mixture model from the statistical point of view. A mixture model
is a probabilistic model representing the presence of a subpopulation within an overall population. Here,
the subpopulation is well log data from the specific type of rock. The mathematical expression of the
mixture model is below:
(3)
where P(x) is probability density function (PDF) of the overall population, f(x) is the PDF of
subpopulation and n is the weight. Because different well log data from different types of rock may have
different statistical properties, f(x) can fall into different PDF categories. However, in this paper, for
simplicity, we automatically transform well log data into Gaussian distribution. When the data is
converted into Gaussian distribution, the mixture model becomes Gaussian mixture model, which is a
parametric probability density function represented as a weighted sum of Gaussian component densities.
Thus, the f(x) is transformed into the following expression:
(4)

where *j and *j are mean and covariance of subpopulation j. So the Gaussian mixture model
expression is as follows:
(5)
Equation (5) shows that the Gaussian mixture model has three unknowns. It is parameterized by the
mean vectors (*j), covariance matrices (*j) and mixture weights (j) from all component densities. We
usually use the following symbol to collect these parameters together:
(6)
There are multiple ways to determine these parameters of the GMM, , which is the cornerstone on
which to build the cluster for each different rock. In this paper, we propose to apply the expectationmaximization (EM) algorithm for estimating these parameters. The objective function of EM algorithm
is the likelihood of the GMM given the training data. The well log data after PCA is both training data
and observation data. For a sequence of T training vectors XPCA {x1, x2, . . ., xT}, the GMM likelihood
can be written as the formula below:

SPE-180439-MS

(7)
where z is latent variables and P is PDF of some unknown parameters . Therefore, the goal of
likelihood maximization is shown as the following expression:
(8)
However, equation (8) is usually difficult to maximize, but by applying numerical method, EM
algorithm can give a relatively accurate approximation after several iterations. It consists of two steps:
(9)
(10)
where t is the time steps (t 0,1,2, . . ., convergence). Figure 5 shows the clustering results after using
GMM. For visual convenience, we only chose the first and second principal components (PC1 and PC2)
as inputs (observation) and overall three clusters as outputs. Thus, three different colors represent three
clusters, but each color in a different figure may have a different meaning. The upper figure shows the
clustering result of Well #1 and the one below shows the result of combined well log data from Well #1
and Well #2. Based on the shape of a single well and combined wells clustering results, the GMM enable
us to have a qualitative indication of the similarity between the two wells. For example, Figure 6 includes
all the clustering results. The first row shows single well clustering results, and the second row shows
results for combined wells. In the Figure, A1 and B2 have a similar shape. It means that Well #1 has a
higher probability of having a large number of the same cluster with Well #2. Then we can continue to
the next stage. If the shape is obviously different, we should go back and pick another well. The main
reason of the qualitative indication is that the first two principal components contain about thirty percent
data information.

Figure 5Well #1 GMM result and Well #1 & Well #2 GMM result

SPE-180439-MS

Figure 6 Eight wells gaussian mixture model clustering results

Model-Based Clustering
As with the GMM, the foundational assumption of model-based clustering is that the data is generated by
a mixture of probability distributions (or mixture models) in which each part is a different cluster or
model. In the GMM, the number of clusters has to be defined by the user. However, this might not be the
case, because we scarcely have the luxury of accessing accurate rock properties before processing the well
log data. So the easy way to set this imperative parameter in GMM is to add two components: (1) four
basic models of the covariance matrix, and (2) the agglomerative algorithm. The new upgraded method
is the main structure of model-based clustering.
Before applying these two components, we need to extend the likelihood function into multivariate
form:
(11)
where i is the set of data of observations that belongs to i th group, j is the total number of clusters
and n is the dimension of well log data matrix. Then we concentrate the log likelihood as:
(12)
where
is the cross-product matrix for the i th group, Ni is total
number of data in i th group and c is constant which equal to 0.4.
The four basic models are scenarios which have four related criteria (Table 1). Based on the table, we
have the following conclusions:

SPE-180439-MS

Table 1Four basic models in MBC

The first model has diagonal and equal covariance matrices and same value in diagonal elements.
The second model has diagonal covariance matrices but the same value in each diagonal element
of the individual covariance matrix.
The third model has equal covariance matrices which have non-zero off-diagonal elements.
The fourth models covariance matrices can vary among components.
The agglomerative algorithm is implemented by using single linkage, which is the similarity of the
closest pair as the following expression:
(13)
where
and
are two different clusters or groups. If two clusters are close enough, the
two will be merged.
Well Correlation
It enables us to do well correlation by generating electrofacies after applying model- based clustering.
Before using MBC, it is necessary to change well log matrix X to combined well log matrix X*, which
is the data from all interested wells. However, the dimension of matrix X* is much larger than that of X
so it would be a serious problem for well correlation due to the low efficiency of expectation maximization
algorithm. For this paper, we have proposed two steps to solve the problem: (1) sampling and (2)
discriminant analysis. In the sampling part, the size of the sample (nsample) should be determined by
following:
(14)
where Np is the number of total population and c0 is sampling constant which is equal to 384.16. Then,
based on the stratigraphic property of the formation, we utilize systematic sampling method to select
sample points. In the systematic random sampling - first, randomly choose the first item or subject from
the defined population; then select proper interval (l) for sampling. The formula of the whole process is:

10

SPE-180439-MS

(15a)
(15b)

After taking a sample from combined well log matrix X*, we implement model-based clustering
method to generate electrofacies from sampling data. For the rest of the dataset, we suggest using
discriminant analysis to put data into the existing electrofacies generated from the previous step.
In terms of the discriminant analysis (DA), it is a classification method where groups or clusters from
populations are known to be a priori and other new observations are classified into one of them based upon
the measured characteristics. It assumes that different classes generate data based on different Gaussian
distributions. For this study, we chose quadratic discriminant analysis, in which the covariance matrix can
be different for each class. The classification is determined by:
(16a)
(16b)
We will classify the sample into the groups that have the largest quadratic score function. Figure 7
shows the result of well correlation of eight wells. From the Figure, the total number of electrofacies is
three. So we use three different colors to represent different electrofacies. With computer assistance, the
efficiency of well correlation has substantially increased. To facilitate future improvement, we can use
core analysis to calibrate well correlation results.

Figure 7Well correlation result

SPE-180439-MS

11

Well Log Data PostprocessingIn the data postprocessing part, we have already selected the optimal
well to build a model using the artificial neural network. ANN is a biologically inspired dynamic
computation system that processes data, and learns in an inherently parallel and distributed fashion. By
discerning the nature of the dependency between input and output variables, it is capable of extracting and
recognizing the underlying dominant patterns and structural relationships among data. Once properly
trained, the network can implicitly classify new patterns and generalize an output based upon the learned
patterns.
Typically, ANN is arranged into three types of layers, namely: input, hidden and output (Figure 8).
Each layer is comprised of a different number of processing elements (PE) which are massively
interconnected. There are a number of designed factors which must be considered when constructing a
neural network model; they include the selection of neural network architecture, the learning rule, the
number of processing elements in each layer, and the type of transfer function. In terms of the architecture
of ANN, feedforward back-propagation is a common scheme for training the network. It attempts to
repeatedly update the matrix of weights of the inputs until the global minimum of the output error is finally
achieved. The output error is described by mean squared error function (MSE) and the training algorithm
is Levenberg-Marquardt, which is the fastest back-propagation algorithm.

Figure 8 Structure of artificial neural network

Field Application & Results

In our case study, we have a total of eight wells in one field, each of which has 38 different types of well
logs. Table 2 shows the list of those logs. After the data mining process, we have an ideal well to use as
a training tool for the artificial neural model; we subsequently employ the trained neural network model
to generate density log from the target well. This process is called pairwise well prediction. The learning
rule is the gradient descent with momentum weight and bias learning function. The momentum () is an
added parameter to the generalized delta rule to prevent the learning process from converging to a local

12

SPE-180439-MS

minimum. It normally varies within 0 to l. The learning speed of a neural network is decided by the
parameter of learning rate (). Generally, it is case specific, and smaller than 1. In the most cases, the
number of PEs in the input and output layers are given by the number of each dimension. But the number
of hidden layers and number of PEs in each one of them is somewhat arbitrary. One rule of thumb states
that the number of hidden layer neurons should be about 75% of the input variables (Al- Fattah et al.,
2001). Incorporating that tenet with trial and error, we decided to use one-hidden-layer architecture
(Figure 9) and specified 40 neurons for that single hidden layer.

Table 2List of types of well logs

Abbreviation

Description

SP
GR
CALI
A010
A020
A090
AT10
SRES
AT20
AT30
MRES
AT60
AT90
DRES
COND
AF10
AF20
AF30
AF60
AF90
TNPH
RHOB
PEF
RMIN
RMNO
RXO
RSO
RMUD
TENS
TOHC
ADT SHALLOW TR.
ADT NEAR MED TR.
ADT FAR MED TR.
ADT DEEP TR.
ADT SHALLOW LG.
ADT NEAR MED LG.
ADT FAR MED LG.
ADT DEEP LG.

Spontaneous Potential
Calibrated Gamma-Ray
Caliper
Resistivity, One Foot
Resistivity, One Foot
Resistivity, One Foot
Resistivity, Two Foot
Resistivity, Two Foot
Resistivity, Two Foot
Resistivity, Two Foot
Resistivity, Two Foot
Resistivity, Two Foot
Resistivity, Two Foot
Resistivity, Two Foot
Resistivity, Two Foot
Resistivity, Four Foot
Resistivity, Four Foot
Resistivity, Four Foot
Resistivity, Four Foot
Resistivity, Four Foot
Thermal Neutron Porosity
Standard Resolution Formation Density
Standard Resolution Formation Photoelectric Factor
Micro Inverse Resistivity
Micro Normal Resistivity
Standard Resolution Invaded Zone Resistivity
Standard Resolution Resistivity Standoff
Mud Fully Calibrated
Cable Tension
Open Hole Contact Temperature
Dielectric Scanner Transverse Shallow Water Filled Porosity
Dielectric Scanner Transverse Near-Medium Water Filled Porosity
Dielectric Scanner Transverse Far-Medium Water Filled Porosity
Dielectric Scanner Transverse Deep Water Filled Porosity
Dielectric Scanner Transverse Shallow Water Filled Porosity
Dielectric Scanner Transverse Near-Medium Water Filled Porosity
Dielectric Scanner Transverse Far-Medium Water Filled Porosity
Dielectric Scanner Transverse Deep Water Filled Porosity

SPE-180439-MS

13

Figure 9 Architecture of one layer neural network

After data preprocessing, the input and output numerical values are normalized within 1 to 1.
Initially, weights of inputs are assigned randomly but are updated after each iteration. A weighted sum of
input variables at each PE is then modified by hyperbolic tangent sigmoid transfer function, given by:
(16)
where is the input data, Si jwjij and w is the weights. The dataset introduced into the neural
network is further divided into training subset, validation subset and test subset in order to avoid
overfitting during the training process. The ratio for each is 0.7, 0.15 and 0.15, respectively. After building
up the model, the next critical part is prediction or pseudo well log generation using an existing model.
The formula to generate well log is as follows:
(17)
where y is the output, b is the bias and W is weight matrix. We have implemented a comprehensive
prediction based on the obtained well log data. Figure 10 shows the correlation matrix of pairwise
prediction results. The well number in the blue box is derived from the training well, while those in the
red box come from target wells. For example, the number in the green circle means that using Well #2
as a training model to predict Well #3, the correlation is 0.85. The diagonal number indicates whether the
training model has been built successfully. The average correlation is above 0.7 which is suitable for
pseudo well log analysis. The diagonal number is the validation result. It is strictly one means that the
model has been built successfully. Additionally, the correlation matrix has symmetric properties. For
example, the correlation result of using Well #2 to predict Well #3 is similar to the result of using Well
#3 to predict Well #2. This property can be further used to validate the prediction result. Figure 11 is the

14

SPE-180439-MS

pseudo-density log of Well #3 using both the model of Wells #1 and #2 from left and right. The orange
line is the prediction and the blue line is the reference. From Fig. 11, we can realize small variances
between different models (Well #1 and Well #2).

Figure 10 Pairwise well prediction results

Figure 11Pseudo density log of Well #3

SPE-180439-MS

15

Summary & Conclusions

In our paper, we created a systematic approach to generating pseudo-density log, which has been proven
to work well. The system consists of three components: data preprocessing, data mining and data
postprocessing. In data preprocessing, we first normalized the data, and then suggested using principal
component analysis. In the next stage, we proposed the initial use of Gaussian mixture model, so as to
have an intuitive scope of the property of the dataset. We subsequently introduced model-based clustering
to devise a proper result without setting up the number of clusters in GMM. For well correlation, we
advocated a sampling method and discriminant analysis which increase the efficiency of model-based
clustering in processing a large dataset. Finally, one hidden layer with back-propagation algorithm neural
network was utilized to build up the model for prediction.
The following conclusions can be drawn from this study:
1. Principal Component Analysis can be a powerful tool in dimension reduction of well log data. The
approach is simple and intuitive to apply. The number of principal components has been proven
in the Appendix.
2. Gaussian Mixture Model can be a natural indicator of selecting potential wells for building up a
model. If clustering shape out of two wells data bears a close resemblance to the shape out of any
single well data, a good relationship may exist between them and vise versa. For ease of
visualization, the input should be the first two principal components after PCA.
3. Model-based clustering method can be an updated version of GMM. It incorporates the algorithm
to merge the clusters if they fall into the same distribution. In MBC, instead of guessing the
number of clusters, we only need to organize an upper boundary. This study proves the stability
of MBC, which is of great importance in well correlation.
4. Sampling method and discriminant analysis are applied to increase the efficiency of MBC in well
correlation. Theoretically, different sampling methods may result in slightly different outcomes.
Herein, the sampling method we suggested is simple to implement, and works well.
5. The parameters and the structure of neural network may vary case by case. It could have some
influence on the prediction result, but there would not be much disparity if the input is the same.
We suggest having increased tests on the neural network before implementation.

Acknowlegdement
Authors would like to thank L. J. (Jim) Lederhos, Lederhos Engineering Consulting, Inc. for giving
us the professional interpretation of well log dataset by applying conventional petrophysical analysis. Mr.
Lederhos has also provided us the software to testify the results. Additionally, we would like to thank
Professor Donald Hill, USC for giving us the idea of pairwise well prediction. It is of paramount
importance to our paper.

Nomenclature
ANN
ARD
CPV
DA
EM
GMM
GR
MBC
MSE

Artificial Neural Network

Auto Relevance Determination
Cumulative Percent Variance
Discriminant Analysis
Expectation Maximization
Gaussian Mixture Model
Gamma Ray
Model-Based Clustering
Mean Squared Error

16

SPE-180439-MS

PC
PCA
PE
SVD

Pricinpal Component
Pricinpal Component Analysis
Processing Element
Singular Value Decomposition

References
1. Abou-Sayed, A. 2012. Data Mining Applications in the Oil and Gas Industry. Journal of Petroleum Technology.
10.2118/1012-0088-JPT.
2. Al-Alwani, M., Britt, L. K. and Dunn-Norman, S. 2015. Data Mining and Statistical Analysis of Completions in an
Unconventional Play: The Canadian Montney Formation. Unconventional Resources Technology Conference, San
Antonio, Texas, USA, 20-22 July. 10.2118/178586-MS.
3. Al-Fattah, S.M., Startzman, R.A. 2001. Neural Network Approach Predicts, U.S. Natural Gas Production. SPE
Production and Operations Symposium, Oklahoma City, Oklahoma, 24-27 March. 10.2118/67260-MS.
4. Aminzadeh, F. and de Groot, P., 2006, Neural Networks and Soft Computing Techniques, with applications in the oil
industry EAGE Publishing.
5. Aminzadeh, F. 1996. Future Geophysical Technology Trends. Proceedings of Stratigraphic Analysis, Pages 16.
6. Aminzadeh, F., Barhen, J., Glover, C. W.N. et al. 2000. Reservoir Parameter Estimation Using a Hybrid Neural
Network. Computers & Geosciences, Volume 26, Issue 8, Pages 869 875. DOI: 10.1016/S0098-3004(00)00027-3.
7. Aminzadeh, F. and Chatterjee, S. 1984. Applications of Clustering in Exploration Seismology. Geoexploration, 23,
Pages 147159.
8. Banchs, R. E. and Michelena, R. J. 2000. Well Log Estimates and Confidence Intervals by Using Artificial Neural
Networks. SEG Annual Meeting, Calgary, Alberta, 6-11 August.
9. Banfield, J. D. and Raftery, A. E. 1993. Model-Based Gaussian and Non-Gaussian Clustering. Biometrics, Volume
49, No.3, Pages 803821.
10. Chaki, S., Verma, A. K., Routray, A. et al. 2013. Prediction of Porosity and Sand Fraction from Well Log Data Using
Neural and Neuro-fuzzy Systems: A Comparative Study. 10th Biennial International Conference and Exposition,
Kochi, India, 23-25 November.
11. Chakra, N. C., Song, K., Gupta, M. M. et al. 2013. An Innovative Neural Forecast of Cumulative Oil Production from
A Petroleum Reservoir Employing Higher-order Neural Networks (HONNs). Journal of Petroleum Science and
Engineering, Volume 106, Pages 18 33. 10.1016/j.petrol.2013.03.004.
12. De Jonge, G., Stundner, M. and Zangl, G. 2003. Automated Reservoir Surveillance through Data Mining Software.
Offshore Europe, Aberdeen, United Kingdom, 2-5 September. 10.2118/83974-MS.
13. Fraley, C. and Raftery, A. E. 1998. How Many Clusters? Which Clustering Method? Answers via Model-Based
Cluster Analysis. The Computer Journal, Volume 41, No. 8, Pages 578 588.
14. Fraley, C. and Raftery, A. E. 2002. Model-Based Clustering, Discriminant Analysis, and Density Estimation. Journal
of the American Statistical Association, Volume 97, No. 458, Pages 611631.
15. Grujic, O., Silva, C. D. and Caers, J. 2015. Functional Approach to Data Mining, Forecasting, and Uncertainty
Quantification in Unconventional Reservoirs. SPE Annual Technical Conference and Exhibition, Houston, Texas,
USA, 28-30 September. 10.2118/174849-MS.
16. Hassibi, M. and Ershaghi. 1999. Reservoir Heterogeneity Mapping Using an Automated Pattern Recognition
Approach. SPE Annual Technical Conference and Exhibition, Houston, Texas, 3-6 October. 10.2118/56818-MS.
17. Hassibi, M., Ershaghi, and Aminzaeh, F., 1996. High Resolution Reservoir Heterogeneity Characterization Using
Recognition Technology: Development in Petroleum Science, Vol. 51, Pages 471497, Elsevier Science, B.V.
18. Holdaway, K. R. 2013. Data Mining Methodologies Enhance Probabilistic Well Forecasting. SPE Middle East
Intelligent Energy Conference and Exhibition, Manama, Bahrain, 28-30 October. 10.2118/167428-MS.
19. Hotelling, H. 1933. Analysis of A Complex of Statistical Variables into Principal Components. J. Educ. Psychol.,
Volume 24, Issue 6, Pages 417441.
20. Jain, A. K., Murty, M. N. and Flynn, P. J. 1999. Data Clustering: A Review. ACM Computing Surveys (CSUR),
Volume 31 Issue 3, Pages 264 323. 10.1145/331499.331504.
21. Kass, R. E. and Raftery, A. E. 1993. Bayes Factors and Model Uncertainty. Technical Report 254, University of
Washington.
22. Kishore, K., Sharma, P. and Khanapurkar, P. 2014. State-of-the-Art Emerging Technology for Technical Data Mining
and Analysis. SPE Large Scale Computing and Big Data Challenges in Reservoir Simulation Conference and
Exhibition, Istanbul, Turkey, 15-17 September. 10.2118/172985-MS.

SPE-180439-MS

17

23. Kumar, B. and Kishore, M. 2006. Electrofacies ClassificationA Critical Approach. Proceedings of the 6th
International Conference & Exposition on Petroleum Geophysics, Kolkata, India.
24. Lee, S. H., Datta-Gupta, A. 1999. Electrofacies Characterization and Permeability Predictions in Carbonate Reservoirs: Role of Multivariate Analysis and Nonparametric Regression. SPE Annual Technical Conference and Exhibition, Houston, Texas, 3-6 October. 10.2118/56658-MS.
25. Lim, J., Kang, J. M. and Kim, J. 1997. Multivariate Statistical Analysis for Automatic Electrofacies Determination
from Well Log Measurements. SPE Asia Pacific Oil and Gas Conference and Exhibition, Kuala Lumpur, Malaysia,
14-16 April. 10.2118/38028-MS.
26. Minka, T. P. 2000. Automatic Choice of Dimensionality for PCA. Advances in Neural Information Processing
Systems 13, Pages 598 604, MIT Press.
27. Mohaghegh, S. D. 2009. Artificial Intelligence and Data Mining: Enabling Technology for Smart Fields. TWA SPE
Journal Paper. 10.2118/0309-014-TWA.
28. Nikravesh, M. 1998. Neural Network Knowledge-Based Modeling of Rock Properties Based on Well Log Databases.
SPE Western Regional Meeting, Bakersfield, California, 10-13 May. 10.2118/46206-MS.
29. Nikravesh, M. and Aminzadeh, F. 2001a. Mining and Fusion of Petroleum Data with Fuzzy Logic and Neural Network
Agents. Journal of Petroleum Science and Engineering, Volume 29, Issues 3-4, May 2001, Pages 221238.
10.1016/S0920-4105(01)00092-4.
30. Nikravesh, M. and Aminzadeh, F. 2001b. Past, Present and Future Intelligent Reservoir Characterization Trends.
Journal of Petroleum Science and Engineering, Volume 31, Issues 2-4, November 2001, Pages 6779. 10.1016/
S0920-4105(01)00121-8.
31. Nikravesh, M., Aminzadeh, F. and Zadeh, L. A. 2003. Intelligent Data Analysis for Oil Exploration. Developments
in Petroleum Science, Volume 51, ISBN: 978-0-444-50685-6.
32. Nikravesh, M., Levey, R. A., Adams, R. D. et al. 1999. Soft Computing: Tools for Intelligent Reservoir Characterization (IRESC). SEG Annual Meeting, Houston, Texas, 31 October-5 November.
33. Pearson, K. 1901. On Lines and Planes of Closest Fit to Systems of Points in Space. Phil. Mag., Series 6, Volume 2,
Issue 11, Pages 559 572.
34. Serra, O. and Abbott, H. T. 1980. The Contribution of Logging Data to Sedimentology and Stratigraphic, 55th Annual
Fall Technical Conference and Exhibition, Dallas, Texas.
35. Sharma, P., Mamgain, G., Bahuguna, V. K. et al. 2011. Improved Permeability Estimates in Carbonate Reservoirs
Using Electrofacies Characterization: A Case Study of Mumbai High South. The 2nd South Asain Geoscience
Conference and Exhibition, New Delhi, India.
36. Shokir, E.M. El-M, Alsughayer, A.A., and Al-Ateeq, A. 2005. Permeability Estimation from Well Log Responses.
Canadian International Petroleum Conference, Calgary, Alberta, 7-9 June. 10.2118/2005-012.
37. Tang, H., Meddaugh, W. S. and Toomey, N. 2011. Using an Artificial-Neural-Network Method to Predict Carbonate
Well Log Facies Successfully. SPE Reservoir Evaluation & Engineering. 10.2118/123988-PA.
38. Tipping, M. E. and Bishop, C. M. 1997. Probabilistic Principal Component Analysis. Technical Report NCRG/97/010,
Neural Computing Research Group, Aston University.
39. Valle, S., Li, W. and Qin, S. J. 1999. Selection of the Number Principal Components: The Variance of the
Reconstruction Error Criterion with a Comparison to Other Methods. Industrial & Engineering Chemi stry Research.
10.1021/ie990110i.
40. Verma, A. K., Cheadle, B. A., Routray, A. et al. 2012. Porosity and Permeability Estimation using Neural Network
Approach from Well Log Data. GeoConvention, Calgary, Alberta, Canada, 14-18 May 2012.
41. Zangl, G. and Oberwinkler, C.P. 2004. Predictive Data Mining Techniques for Production Optimization. SPE Annual
Technical Conference and Exhibition, Houston, Texas, 26-29 September. 10.2118/90372-MS.

18

SPE-180439-MS

Appendix
The concept of principal component analysis (PCA) was first introduced by Karl Pearson in 1901 and
independently developed by Harold Hotelling in the 1930s. Since then it has been a powerful tool to
reduce dataset dimensionality by transforming an original dataset to a new one in a lower dimension
subspace. In this method, choosing dimensionality in the projected space is a key issue because less
information will be forthcoming if fewer principal components (PCs) than required are retained, which
may lead to underfitting of the training model. In contrast, if an excess of required PCs are used, it will
unavoidably increase computation expenses, and potentially generate overfitting problems in training. In
the past, numerous approaches have been developed to choose the optimal number of PCs, including, but
not limited to the Kaiser rule, elbow rule on scree plot, cumulative percent variance (CPV), parallel
analysis (PA), Bayesian Model Selection, RRN/RRU, and auto relevance determination (ARD). Each
method has shown strengths in different cases. However, at this writing, how to determine the optimal
number of PCs remains controversial. Among previous approaches, CPV, elbow test, Kaiser Rule and
parallel analysis can be deemed as empirical methods. CPV determines the number of PCs based on the
cumulative percent of the variance, but the benchmark of selectivity is subjective. The elbow rule is to
locate the number of PCs by picking an evident turning point on the on scree plot. But sometimes a
relatively smooth curve, or the presence of several elbows may make selectivity of numbers ambiguous.
Kaiser Rule is a popular approach which retains PCs if corresponding eigenvalues are above 1. But
generally speaking, it may overestimate the numbers of PCs. In parallel analysis, a Monte-Carlo based
simulation is implemented wherein a dataset of random numbers with the same size and number of
variables as the observed dataset is generated for analysis. Their eigenvalues will then be calculated and
saved. This process is repeated many times prior to performing a statistical analysis to obtain the variance
for each PC at a 95% confidence interval. However, this method can be quite costly in terms of
computation time. Bayesian model selection, RRN/RRU and ARD have solid statistical foundations.
Bayesian model selection method has been commonly applied to probabilistic PCA model since Tipping
and Bishop (1997) related PCA to maximum likelihood density estimation. To solve the estimate
analytically, Laplace approximation and its simplification BIC approximation (Kass and Raftery, 1993)
have been adopted. That method offers a fast and statistically reliable way to pick correct dimensionality
of the data (Minka, 2000) while the assumption is only Gaussian distribution. RRN/RRU method
incorporates two models with Gaussian density and uniform density, respectively. But it is not prone for
those two methods to get true dimensionality for a large dataset because they use a restrictive model for
the subspace. ARD is also a classical method based on Bayesian inference. Nevertheless, that method
seems to be computation intensive for many applications.
In our paper, we compared results of optimal PC numbers given by the above methods. In order to get
an appropriate number of PCs as input to the ANN, we try to find a balance between empirical and
statistical methods in terms of accuracy and computation expense. Therefore, we try to screen the
candidates given those two considerations. Figure 12 and Figure 13 show simulated graphs, respectively,
using our dataset with 1861 observations and 37 features. In Figure 12, x-axis value of intersection point
will give us the PC number. In Figure 13, the PC number is obtained according to the evident turning point
location. It is noted that the score keeps increasing with the dimensionality in (a) and (b), because
corresponding eigenvalues dont level off, but continue downward. So in this case, we pick the largest
possible dimensionality. All the results are listed in Table 3, from which one notices that the result from
CPV may be our target. In order to validate that choice, we later carried out an experiment. Given different
optimal PC numbers from previous models, we adjust our input dataset dimensions of all 8 wells to reflect
that. Then we use this dataset as the input of ANN to make a pairwise prediction of bulk density. For
example, provided that optimal number of PCs is 34, we pick out 34 components from the transformed
dataset of all 8 wells as the input of ANN. Then we use log dataset in Well #1 as the training dataset to

SPE-180439-MS

19

predict the bulk density log value in Well #2, Well #3 and so on. Then we calculated an average
correlation between predicted values and real values of the 8 wells. In Figure 14, we see the correlation
between average correlation coefficients and PC numbers. We found that the PC number according to the
approximate turning point matches our earlier choice. Therefore, given the computation cost and
prediction accuracy, the CPV method offered the best result.

Figure 12Principal component selection

Figure 13Principal component selection

20

SPE-180439-MS

Table 4
Methods

Number of PCs

Methods

Number of PCs

Cumulative Variance Proportion

Elbow Test
Kaiser Rule
Parallel Analysis

15
5
7
6

BIC Approximation
Laplace Approximation
RRU
ARD

34
33
3
10

Correlation K34

Well 1

Well 2

Well 3

Well 4

Well 5

Well 6

Well 7

Well 8

1.00
0.88
0.84
0.80
0.83
0.69
0.79
0.69
0.81

0.87
1.00
0.83
0.90
0.79
0.78
0.85
0.68
0.84

0.82
0.80
1.00
0.78
0.81
0.78
0.73
0.64
0.79

0.79
0.89
0.79
1.00
0.84
0.78
0.79
0.72
0.82

0.82
0.81
0.82
0.83
1.00
0.85
0.77
0.63
0.82

0.72
0.79
0.79
0.76
0.83
1.00
0.84
0.69
0.80

0.81
0.83
0.72
0.80
0.76
0.84
1.00
0.74
0.81

0.72
0.66
0.65
0.73
0.66
0.71
0.75
1.00
0.74

Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 7
Well 8
Average
K15

Well 1

Well 2

Well 3

Well 4

Well 5

Well 6

Well 7

Well 8

Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 7
Well 8
Average

1.00
0.86
0.85
0.67
0.78
0.67
0.79
0.71
0.79

0.85
1.00
0.77
0.83
0.84
0.74
0.83
0.68
0.82

0.78
0.74
1.00
0.67
0.80
0.73
0.68
0.65
0.76

0.64
0.80
0.67
1.00
0.84
0.83
0.79
0.72
0.79

0.75
0.80
0.75
0.80
1.00
0.82
0.73
0.61
0.78

0.66
0.70
0.70
0.81
0.80
1.00
0.84
0.69
0.78

0.77
0.80
0.67
0.80
0.71
0.74
1.00
0.82
0.79

0.65
0.65
0.60
0.68
0.63
0.65
0.73
1.00
0.70

K10

Well 1

Well 2

Well 3

Well 4

Well 5

Well 6

Well 7

Well 8

Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 7
Well 8
Average

1.00
0.83
0.77
0.66
0.79
0.69
0.58
0.74
0.76

0.80
1.00
0.83
0.77
0.78
0.77
0.81
0.69
0.81

0.75
0.80
1.00
0.57
0.50
0.46
0.66
0.59
0.67

0.64
0.75
0.50
1.00
0.81
0.75
0.61
0.44
0.69

0.80
0.77
0.51
0.79
1.00
0.79
0.70
0.68
0.76

0.66
0.79
0.48
0.74
0.76
1.00
0.69
0.73
0.73

0.62
0.78
0.64
0.62
0.70
0.67
1.00
0.73
0.72

0.70
0.68
0.57
0.43
0.63
0.71
0.74
1.00
0.68

K7

Well 1

Well 2

Well 3

Well 4

Well 5

Well 6

Well 7

Well 8

Well 1
Well 2
Well 3
Well 4
Well 5
Well 6
Well 7
Well 8
Average

1.00
0.78
0.61
0.55
0.58
0.56
0.56
0.50
0.64

0.80
1.00
0.67
0.73
0.50
0.53
0.53
0.56
0.66

0.56
0.64
1.00
0.33
0.69
0.43
0.49
0.51
0.58

0.53
0.66
0.30
1.00
0.68
0.70
0.64
0.49
0.63

0.55
0.52
0.64
0.67
1.00
0.61
0.78
0.71
0.69

0.56
0.49
0.42
0.68
0.65
1.00
0.74
0.80
0.67

0.56
0.49
0.53
0.66
0.75
0.71
1.00
0.64
0.67

0.48
0.53
0.54
0.52
0.74
0.78
0.62
1.00
0.65

SPE-180439-MS

21

Figure 14 Average correlation between predicted values and real values of eight wells