SPE-191827-18ERM-MS

Application of Machine Learning Algorithms for Optimizing Future

Production in Marcellus Shale, Case Study of Southwestern Pennsylvania

Alireza Shahkarami, Saint Francis University; Kimberly Ayers, Ayers Petroleum Consulting; Guochang Wang, Saint
Francis University; Alivia Ayers, Ayers Petroleum Consulting

Copyright 2018, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Eastern Regional Meeting held in Pittsburgh, Pennsylvania, USA, 7 - 11 October 2018.

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
The Marcellus Shale has more than a decade of development history. However, there are many questions
that still remain unanswered. What is the best inter-well spacing? What are the optimum stage length,
proppant loading, and cluster spacing? What are the ideal combinations of these completion parameters?
And how can we maximize the rate return on our investment? This study proposes innovative tools that
allow researchers to answer these questions. We build these set of tools by utilizing the pattern
recognition abilities of machine learning algorithms and public data from the Southwestern Pennsylvania
region of the Marcellus Shale.
By means of artificial intelligence and data mining techniques, we studied a database that includes
public data from more than 2,000 wells producing from the aforementioned study area. The database
contained completion, drilling, and production history information from various operators active in
Allegheny, Greene, Fayette, Washington, and Westmoreland counties located in the Southwestern
Pennsylvania. Extensive preprocessing and data cleansing steps were involved to prepare the database.
Various machine learning techniques (Linear Regression (LR), Support Vector Machines (SVMs),
Artificial Neural Networks (ANNs), and Gaussian Processes (GP)) were applied to understand the non-
linear patterns in the data. The objective was to develop predictive models that were trained and validated
based on the current database. The predictive models were validated using information originating from
numerous wells in the area. Once validated, the model could be used in reservoir management decision-
making workflows to answer questions such as what are the best drilling scenarios, the optimum
hydraulic fracturing design, the initial production rate, and the estimated ultimate recovery (EUR). The
workflow is purely based on field data and free of any cognitive human bias. As soon as more data is
available, the model could be updated. The core data in this workflow is sourced from public domains,
and therefore, intensive preprocessing efforts were necessary.

Introduction
The current era of the Marcellus Shale play development started on October 20, 2004, when Range
Resources completed a vertical and single stage well (Renz Unit #1) in Washington County, Pennsylvania
(Ventura, et al. 2013). Range Resources was inspired by the success of Barnett Shale play in Texas and
2 SPE-191827-18ERM-MS

saw the similarity of Marcellus Shale play to Barnett Shale as an opportunity (Ventura, et al. 2013). Over
370,000 pounds of proppant and one million gallons of water were used to complete the Renz Unit #1
well (Ventura, et al. 2013). The initial rate of 300 MCF/Day was considered a success compared to the
Barnett Shale play examples (Ventura, et al. 2013). After the Renz Unit #1, Range Resources drilled and
completed several vertical and horizontal wells in order to find the most efficient and economical
techniques to make the Marcellus Shale play worth exploring. It was not until 2007 that Range Resources
drilled and completed three horizontal wells, yielding initial rates of 3.7, 4.3, and 4.7 MMCF/Day
(Beckwith 2013).
Over a decade later, the Marcellus Shale play is the largest natural gas reservoir in the United States.
Based on the February 2018 report from the U.S. Energy Information Administration (U.S. Energy
Information Administration 2018), the Marcellus Shale play is by far the most prolific in the United
States. Only considering the portions of Marcellus Shale in Pennsylvania and West Virginia, the EIA has
reported the Marcellus Shale play as having the largest amount of natural gas proved reserves (84.1
trillion cubic feet) in 2016 (U.S. Energy Information Administration 2018). After the Marcellus Shale, the
EIA reported the Eagle Ford in Texas to be the second largest reserve in the United States, containing
22.7 trillion cubic feet of gas (U.S. Energy Information Administration 2018).
The Marcellus Shale is a Devonian-age organic-rich formation in the Appalachian basin that covers
a subsurface area of over 95,000 square miles, stretching from New York in the north to Kentucky and
Tennessee in the south (U.S. Energy Information Administration 2017). The formation of Marcellus Shale
includes two organic-rich intervals, the Upper Marcellus and Lower Marcellus. The Lower Marcellus
with higher Total Organic Carbon (TOC) is usually the target of horizontal well landing. Figs. 1 and 2
show the structural and thickness maps of the Marcellus Shale. The top of Marcellus Shale formation
varies from 100 feet to 9,900 feet below the surface (Fig. 1). The thickness ranges from ten feet (south of
Marcellus Shale play) to nearly one thousand feet in the northeastern region (Fig. 2) (U.S. Energy
Information Administration 2017). In Pennsylvania, the portion of the Marcellus Shale located in the
northeastern and southeastern regions are extremely prolific and popular among operators, as shown in
Fig. 3. These two regions include thick and high TOC intervals of Marcellus Shale. In the Southwestern
Pennsylvania, the average thickness of Marcellus interval ranges from 150-200 feet.
Due to natural gas commodity pricing, shale asset operations are high-risk endeavors. Exploration and
production companies aim to reduce the uncertainties involved in their operations by predicting
production performance and optimizing the costly elements of operations involved in horizontal drilling
and completion design. Factors such as inter-lateral well spacing, hydraulic fracturing design, and well
placement play important roles in the economic success of shale operations. Understanding the geologic
characteristics of the formation is also a critical element. Geological modeling and numerical reservoir
simulation have been the standard tools utilized in the industry to integrate these controlling factors
(Shahkarami, Wang and Belyadi 2017). When it comes to shale reservoirs, due to the ultra low
permeability of rock, the physics governing the fluid flow is more complex than the conventional
reservoirs (Javadpour, Fisher and Unsworth 2007). Consequently, the numerical reservoir simulations for
shale reservoirs are less reliable, biased toward the conventional reservoirs, computationally expensive,
and unable to perform a full-field analysis. If enough data is available, the data-driven techniques are the
appropriate tools to address the shortcomings of numerical reservoir simulations. These techniques are
based on the pattern recognition capabilities of artificial intelligence, data mining, and machine learning
algorithms. The application of data analytics techniques in petroleum engineering is not new (Boomer
1995, Mohaghegh 1995), but in the recent years they have been widespread (Alabboodi and Mohaghegh
2016, Gaurav 2017, Suhag, Ranjith and Aminzadeh, Comparison of Shale Oil Production Forecasting
using Empirical Methods and Artificial Neural Networks 2017, Pankaj, et al. 2018).
Figure 1—Structural map of the Marcellus Shale formation (U.S. Energy Information Administration
2017).

Figure 2—Thickness map of the Marcellus Shale formation (U.S. Energy Information Administration
2017).
 Figure 3—Wells drilled in the Marcellus Shale and Utica/Point-Pleasant Shale
formation through April 2017 (U.S. Energy Information Administration 2017).

The focus of this study is on the Southwestern Pennsylvania area of the Marcellus Shale. The database
is compiled from public data and includes drilling, completion, and production history data from wells in
Allegheny, Greene, Fayette, Washington, and Westmoreland counties. Public data is not always reliable
and extensive preprocessing and screening should be done before conducting any modeling. This study
utilized various machine learning techniques, including Linear Regression, Support Vector Machines,
Artificial Neural Networks, and Gaussian Processes, to characterize the production performance of the
wells in the database. The objective was to construct predictive models for the area covered by the
database. The predictive models were validated using the cross-validation technique. The validated
models could be used in reservoir management decision-making workflows to answer questions about the
best infill drilling scenarios, optimum hydraulic fracturing design, estimated initial production rate, and
normalized estimated ultimate recovery.

Database
The final, filtered database includes 820 horizontal wells drilled and completed in the Marcellus Shale
formation in the Southwest of Pennsylvania. The data used in this study was compiled and provided for
use by James Knobloch Petroleum Consultants Inc. (James Knobloch Petroleum Consultants Inc. 2018)
as a part of their Appalachian Unconventional Well database. The source is deemed as public information
and supplied by the Pennsylvania Department of Environmental Protection (PA DEP) via the EDWIN
data portal (The Bureau of Topographic & Geologic Survey 2018). Data accuracy is limited to the quality
of the data provided by the operators to the PA DEP. Myers et al. (Myers, et al. 2017) also drew from this
database in their study and offered a comprehensive discussion on these public sources, their quality, and
the ways to interpret the errors. The final information and parameters in the database are detailed in Fig.
4.
 Figure 4—Database used in this study.

The original dataset contained more than 2000 wells. However, more than half of these wells had key
information missing and/or obvious data quality issues. The locations of the wells are displayed in Fig. 5.
Fig. 6 displays the color-coded study wells to indicate the original operators. Figs. 7 and 8 map the wells
based on their relative locations (Inner/Outer/Loner) and development history (Parent/Child well). More
than 99% of these wells are developed between 2011 and 2015 (Fig. 9).

Figure 5—Green circles show the location of 820 horizontal wells in the database. The public
sources used in the database are discussed thoroughly in Myers et al. (Myers, et al. 2017).
 Figure 6—Wells are color-coded to indicate the original operators

Figure 7—Relative locations (Inner/Outer/Loner) of wells in the database

 Figure 8—Demonstration of well generation (Parent/Child well) in the database

Figure 9—More than 99% of wells in the database have been

developed (meaning hydraulically fractured) between 2011 and
2015.

To represent the reservoir quality in the database, average thickness values of the Marcellus Shale
were included. Additionally, the database included the thickness values of the Purcell Limestone layer
(also referred to as the Cherry Valley formation). An organic-rich mud rock, the Marcellus Shale contains
several limestone interlayers (Wang, Ju, et al. 2014, Wang, Shakarmi and Bruno 2016). Most of these
limestone interlayers do not cover a large area. Because of the different mechanical properties, these
limestone interlayers could impact the success of a completion job in the Marcellus Shale. In
Southwestern
Pennsylvania, the thickest limestone interlayer is the Purcell Limestone interval that separates the Upper
Marcellus Shale and the Lower Marcellus Shale layers. 228 well logs were used to interpret the Marcellus
Shale and Purcell Limestone thickness values (Fig. 10). The gamma ray and density values were the main
features used to estimate the thickness values. Fig. 10 shows the well log database location and Figs. 11
and 12 display the thickness of the Marcellus Shale and the Purcell Limestone layers in the studied area.

Figure 10—Marcellus Shale and Purcell Limestone thickness values were estimated using 228 well logs.

Figure 11—Average Marcellus Shale gross thickness for the study area
 Figure 12—Purcell Limestone thickness for the study area

Methodology
Fig. 13 demonstrates the general workflow of a data-driven project. In this study, the "problem" to
address was the estimation of EUR. In this workflow, understanding the problem and data are very critical
to developing an efficient model. The understanding of previous steps impacts the database preparation
procedures. For instance, the EUR value for a well is known to be directly impacted by the reservoir
quality, drilling and completion design and operation, history and relative location of the well, and initial
well performance. Database preparation involves preprocessing the raw data before the application of a
machine learning algorithm or modeling step. For example, because of the public nature of the data,
considerable time was spent on cleaning and removing the outliers. This step may also require modeling
procedures. The EUR values in the database were estimated using Decline Curve Analysis (DCA) using
"b" value of 1.2 and 6% terminal decline over 50 years. The EUR estimations are courtesy of James
Knobloch Petroleum Consultants (James Knobloch Petroleum Consultants Inc. 2018). Another example
of modeling in the data processing step is calculating the thickness values for the Marcellus Shale and the
Purcell Limestone intervals, as discussed above.
 Figure 13—General workflow for a data-driven project

The modeling step in the general workflow of data-driven projects involves utilizing a machine
learning algorithm to develop a predictive model. The predictive model relies heavily on data quality
and preprocessing procedures during the database preparation. It should be understood that the modeling
and database preparation is a back-and-forth process. It is common to go back to the database and apply
a new preprocessing technique based on the insights gained during the modeling stage. Once a model
is developed or trained, its predictability is tested through the "validation" process. The validation (or
verification/evaluation/testing) step involves using a set of data that has not been used (seen) during the
model development (training). Therefore, a portion of the database should always be saved for validation
proposes. If the training data set is not a good representative of the database, the model will not perform
well on the validation set. To mitigate this problem, it is common to repeat the whole process of training
and validation several times using random samples. Each time a portion of the database is chosen
randomly to train the model, the remainder is used for validation. The reported error is an average error
over the number of training/validation processes. This process is known as "k-fold cross validation," with
"k" indicating the number of portions that our database is split into. When data is limited, cross validation
is a common practice used to test the prediction capabilities of machine learning algorithms. In addition,
cross validation is a standard way of comparing the limits of different algorithms. Another benefit of
cross validation is that all data points are used for both training and validation, and each data point is used
for validation once. This study used a 10-fold cross validation. A strong validation performance provides
more trust for the "implementation" step. A poor validation performance requires going back to the
problem, understanding every step, and studying the objectives of the project.
This study utilized four different machine learning algorithms: Simple Linear Regression (SLR),
Support Vector Machines (SVMs), Artificial Neural Networks (ANNs), and Gaussian Process (GP). The
following paragraphs give an overview of these algorithms using surveyed studies that apply them on a
Marcellus Shale database.
Simple Linear Regression is a statistical technique that summarizes the relationships between two sets
of variables as a linear combination. Cunningham et al. (Cunningham, et al. 2012) used a linear
regression
algorithm to characterize the EUR for the Marcellus Shale assets owned by EQT. Their database had
170 horizontal wells in Pennsylvania and West Virginia. Izadi et al. (Izadi, et al. 2014) presented another
Marcellus Shale study integrating parameters such as completion design, geo-mechanical properties, and
production profile to a regression technique. They aimed to link the efficiency of hydraulic fracturing jobs
and production performance to rock properties.
A disadvantage of simple regression models is that they only work well using data with linear
behavior. This characteristic makes regression models too simple for many practical applications. On the
other hand, machine learning algorithms such as Support Vector Machines and Artificial Neural Networks
perform well on data with non-linear relationships among variables. In SVMs, the inputs are transferred
into a new dimension and perform linear regression in the new dimension. The linear behavior in the new
space can represent non-linear relationships in the original space. 1ANNs were motivated by the goal of
having machines that are able to mimic the behavior of a human brain. ANNs include interconnected
artificial neurons, an idea originally adopted from the anatomy of the human brain. Several applications
of ANNs have been discussed thoroughly in (Shahkarami, Mohaghegh and Hajizadeh 2015, A.
Shahkarami, S. D. Mohaghegh and V. Gholami, et al. 2015). Wang et al. (Wang, Ju, et al. 2014) and
include a comprehensive overview of SVMs and ANNs for predicting the Marcellus Shale litho-facies. In
another study, Wang et al. (Wang, Shakarmi and Bruno 2016) use SVMs and ANNs to predict the total
organic content in the Utica- Point Pleasant formations in the Appalachian Basin. Shahkarami et al.
(Shahkarami, Wang and Belyadi 2017) used ANNs to create a proxy model for fast track analysis of well
spacing optimization in the Utica- Point Pleasant formations. Suhag et al. (Suhag, Ranjith and Aminzadeh
2017) performed a comparison study of oil shale production using non-linear regression and ANNs.
Applications of Gaussian Processes for predictive modeling are not new. Kriging, the popular geo-
statistical technique, is based on Gaussian Processes (Rasmussen and Williams 2006). Known to be a less
parametric algorithm, the Gaussian Process has a tendency to be more generalized, suitable for small data
sets, and adaptive to non-linear behaviors (Dudley 2010). A good source on the theory and applications of
Gaussian Processes is Rasmussen and Williams’ book (Rasmussen and Williams 2006). Yu et al. (2016)
utilized Gaussian Processes to estimate total organic carbon from well logs in a shale reservoir. Hamdi et
al. (Hamdi, Hajizadeh and Sousa 2015) used the Gaussian Process for improving the efficiency of
assisted history matching workflow.

Results
As described in Fig. 4 above, 25 attributes were selected as input parameters for the modeling. These
parameters represent the geographical location, geologic conditions, drilling and completion designs, and
initial production results while keeping the model as streamlined and efficient as possible. As discussed
above, the well data set was reduced to 820 instances from more than 2000 wells to ensure the utilization
of the highest quality of data. Fig. 14 below shows the results of the testing trellised by the various
machine learning algorithms. Different statistical means were used to quantify the accuracy of the
modeling results. These statistical analysis results are found below in Table 1.
 Table 1—Machine Learning algorithm results

Figure 14—Results of Algorithms. Actual EUR vs Predicted EUR (BCF) trellised by test.

Considering the known issues and lack of granularity associated with public data, the algorithms
produced a rather high correlation coefficient. The SVM produced the highest correlation coefficient,
although GP and SLR were really close. The ANNs model had the lowest correlation coefficient. In
general all four algorithms tend to underestimate the EUR values.
With the model's reliability now established, it is possible to predict EURs under a set of conditions
within the bounds of the study. For example, given township (proxy for geologic conditions), lateral
length, and basic stimulation parameters, the model is able to predict EUR with an acceptable level of
certainty. However, in areas of Southwestern Pennsylvania, the range of parameters (such as stage length
and stimulation size) is limited to an older generation of completions. This obstacle may be overcome by
establishing an analogy set of wells in an area with similar conditions. These conditions require wells
with modern completions, analysis of the data to understand the "completion uplift" factor, and
application of the factor to the area with older completions. Another solution is normalizing the output.
For instance we can do the same analysis using the EUR per lateral length.
Conclusion
It has been shown that public data may be used to predict dry gas reservoir performance with a reasonable
degree of certainty in southwest Pennsylvania by using four different machine learning algorithms:
Simple Linear Regression (SLR), Support Vector Machines (SVMs), Artificial Neural Networks (ANNs),
and Gaussian Process (GP). SVM was found to be the most reliable model and may be used to predict
EURs for future wells.
The authors plan to continue to fine tune and refine the models discussed in this paper to improve
accuracy. Along with the continued refinement of the data set, the authors would like to incorporate other
data streams, such as inter-well spacing and sand loading, in future work. Additionally, these models
may be further refined for specific areas of interest for the use of type curve generation. As operators
continue to develop the Marcellus Shale in Southwestern Pennsylvania, the additional well results and
improved completions designs will lead to further efficiency and accuracy gains in both the prediction
and the production of natural gas reserves.

Acknowledgments
We would like to extend our appreciation to James Knobloch Petroleum Consultant Inc. for providing the
dataset used in this study. We also acknowledge the University of Waikato, New Zealand for developing
the software used in our work (Waikato Environment for Knowledge Analysis).

References
Alabboodi, M. J., & Mohaghegh, S. D. (2016). Conditioning the Estimating Ultimate Recovery of Shale Wells to
Reservoir and Completion Parameters. SPE Eastern Regional Meeting. Canton, OH: Society of Petroleum Engineers.
Beckwith, R. (2013, June). The Marcellus Shale Gas Boom Evolves. Journal of Petroleum Technology.
Boomer, R. J. (1995). Predicting Production Using a Neural Network (Artificial Intelligence Beats Human Intelligence).
Petroleum Computer Conference. Houston, TX: Society of Petroleum Engineers.
Cunningham, C. F., Cooley, L., Wozniak, G., & Pancake, J. (2012). Using Multiple Linear Regression To Model EURs of
Horizontal Marcellus Wells. 2012 SPE Eastern Regional Meeting. Lexington, KY: Society of Petroleum Engineers.
Dudley, R. M. (2010). Selected Works of R.M. Dudley. (Giné,E. Koltchinskii,V. & Norvaiša,R. Eds.) Verlag, New York:
Springer.
Alabboodi, M. J., & Mohaghegh, S. D. (2016). Conditioning the Estimating Ultimate Recovery of Shale Wells to
Reservoir and Completion Parameters. SPE Eastern Regional Meeting. Canton, OH: Society of Petroleum Engineers.
Beckwith, R. (2013, June). The Marcellus Shale Gas Boom Evolves. Journal of Petroleum Technology.
Boomer, R. J. (1995). Predicting Production Using a Neural Network (Artificial Intelligence Beats Human Intelligence).
Petroleum Computer Conference. Houston, TX: Society of Petroleum Engineers.
Cunningham, C. F., Cooley, L., Wozniak, G., & Pancake, J. (2012). Using Multiple Linear Regression To Model EURs
of Horizontal Marcellus Wells. 2012 SPE Eastern Regional Meeting. Lexington, KY: Society of Petroleum
Engineers.
Dudley, R. M. (2010). Selected Works of R.M. Dudley. (Giné,E. Koltchinskii,V. & Norvaiša,R. Eds.) Verlag, New York:
Springer.
Gaurav, A. (2017). Horizontal Shale Well EUR Determination Integrating Geology, Machine Learning, Pattern
Recognition and MultiVariate Statistics Focused on the Permian Basin. SPE Liquids-Rich Basins Conference -
North America. Midland, TX: Society of Petroleum Engineers.
Hamdi, H., Hajizadeh, Y., & Sousa, M. C. (2015). Gaussian Process for Uncertainty Quantification of Reservoir
Models. SPE/IATMI Asia Pacific Oil & Gas Conference and Exhibition. Nusa Dua, Bali, Indonesia: Society of
Petroleum Engineers.
Izadi, G., Junca, J.-P., Cade, R., & Rowan, T. (2014). Production Performance in Marcellus Shale: Multidisciplinary
Study of Hydraulic Fracturing. Abu Dhabi International Petroleum Exhibition and Conference. Abu Dhabi, UAE:
Society of Petroleum Engineers.
James Knobloch Petroleum Consultants Inc. (2017, August 1). Appalachian Basin Horizontal Well Database. Marietta,
Ohio, USA.
James Knobloch Petroleum Consultants Inc. (2018, August 1). Appalachian Basin Horizontal Well Database. Marietta,
Ohio, USA.
 Javadpour, F., Fisher, D., & Unsworth, M. (2007). Nanoscale Gas Flow in Shale Gas Sediments. Journal of Canadian
Petroleum Technology, 55–61.
Mohaghegh, S. (1995). Neural Network: What It Can Do For Petroleum Engineers. Journal of Petroleum Technology, 42.
Myers, R. R., Knobloch, T. S., Jacot, R. H., & Ayers, K. L. (2017). Production Analysis of Marcellus Shale Well
Populations Using Public Data: What Can We Learn from The Past? Eastern Regional Meeting. Lexington, KY: Society
of Petroleum Engineers.
Nine Energy Service, I. (2016, May 17). PR News. Retrieved 7 15, 2017, from http://www.prnewswire.com/news-
releases/ nine-energy-completes-124-stage-well-in-utica-shale-579799881.html
Pankaj, P., Geetan, S., MacDonald, R., Shukla, P., Sharma, A., Menasria, S., et al. (2018). Need for Speed: Data
Analytics Coupled to Reservoir Characterization Fast Tracks Well Completion Optimization. SPE Canada
Unconventional Resources Conference. Alberta, Canada: Society of Petroleum Engineers.
Rasmussen, C. E., & Williams, C. K. (2006). Gaussian Processes for Machine Learning. Cambridge, MA: MIT Press.
Shahkarami, A., Mohaghegh, S. D., & Hajizadeh, Y. (2015). Assisted History Matching Using Pattern Recognition
Technology. SPE Digital Energy Conference and Exhibition. he Woodlands, Texas: Society of Petroleum Engineers.
Shahkarami, A., Mohaghegh, S. D., Gholami, V., & Bromhal, G. (2015). Application of Artificial Intelligence and Data
Mining Techniques for Fast Track Modeling of Geologic Sequestration of Carbon Dioxide a Case Study of SACROC
Unit. SPE Digital Energy Conference and Exhibition. he Woodlands, Texas: Society of Petroleum Engineers.
Shahkarami, A., Mohaghegh, S., & Hajizadeh, Y. (2018). Assisted History Matching Using Pattern Recognition
Technology. Int. J. of Oil, Gas and Coal Technology, 412 – 442.
Shahkarami, A., Wang, G., & Belyadi, H. (2017). Horizontal Well Spacing and Hydraulic Fracturing Design
Optimization: A Case Study on Utica-Point Pleasant Shale Play. Journal of Sustainable Energy Engineering, 148–
162.
Suhag, A., Ranjith, R., & Aminzadeh, F. (2017). Comparison of Shale Oil Production Forecasting using Empirical
Methods and Artificial Neural Networks. SPE Annual Technical Conference and Exhibition. San Antonio, TX:
Society of Petroleum Engineers.
Suhag, A., Ranjith, R., & Aminzadeh, F. (2017). Comparison of Shale Oil Production Forecasting using Empirical
Methods and Artificial Neural Networks. SPE Annual Technical Conference and Exhibition. San Antonio, Texas:
Society of Petroleum Engineers.
The Bureau of Topographic & Geologic Survey. (2018). Exploration and Development Well Information Network.
Retrieved 2018, from http://docs.dcnr.pa.gov/topogeo/econresource/oilandgas/EDWIN_home/index.htm
U.S. Energy Information Administration. (2017). Marcellus Shale Play Geology review. Washington DC: U.S. Energy
Information Administration (EIA).
U.S. Energy Information Administration. (2018). U.S. Crude Oil and Natural Gas Proved Reserves, Year-end 2016.
Washington DC: U.S. Energy Information Administration.
Vapnik, V., Golowich, S., & and Smola, A. (1997). Support vector method for function approximation, regression
estimation, and signal processing. Advances in Neural Information Processing Systems9, 281–287.
Ventura, J., WalkerR.,Jr., Zagorski, W., Davis, G., Applegath, J., Curry, M., et al. (2013). The Discovery of the
Marcellus Shale Play, An Operator's Experience. 2013 Unconventional Resources Technology Conference,. Denver,
CO: Unconventional Resources Technology Conference.
Wang, G., Ju, Y., Carr, T. R., Li, C., & Cheng, G. (2014). Application of Artificial Intelligence on Black Shale
Lithofacies Prediction in Marcellus Shale, Appalachian Basin. 2014 Unconventional Resources Technology
Conference. Denver, CO: Society of Petroleum Engineers.
Wang, G., Shahkarami, A., & He, Q. (2018). Characterization of Limestone Interlayers in Marcellus Shale and its Effect
on Production in Southwestern PA. ACE 2018 Annual Convention & Exhibition. Salt Lake City, Utah: American
Association of Petroleum Geologists.
Wang, G., Shakarmi, A., & Bruno, J. (2016). TOC Content Distribution Features in Utica-Point Pleasant Formations,
Appalachian Basin. 2016 Unconventional Resources Technology Conference. San Antonio, Texas: Society of
Petroleum Engineers.
Wang, G., Shakarmi, A., & Bruno, J. (2016). TOC Content Distribution Features in Utica-Point Pleasant Formations,
Appalachian Basin. 2016 Unconventional Resources Technology Conference. San Antonio, Texas: Society of
Petroleum Engineers.
Yu, H., Wang, Z., Rezaee, R., Zhang, Y., Xiao, L., Luo, X., et al. (2016). The Gaussian Process Regression for TOC
Estimation Using Wireline Logs in Shale Gas Reservoirs. International Petroleum Technology Conference.
Bangkok, Thailand: International Petroleum Technology Conference.