University of Southern California Dissertations and Theses

Application of data-driven modeling in basin-wide analysis of unconventional resources, including domain expertise

(USC Thesis Other)

Application of data-driven modeling in basin-wide analysis of unconventional resources, including domain expertise

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content APPLICATION OF DATA-DRIVEN MODELING IN BASIN-WIDE
ANALYSIS OF UNCONVENTIONAL RESOURCES, INCLUDING
DOMAIN EXPERTISE
by
Cyrus Ashayeri

A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirement for the Degree
DOCTOR OF PHILOSOPHY
(PETROLEUM ENGINEERING)

August 2021
Copyright 2021 Cyrus Ashayeri

ii
Dedication

I dedicate this to the memory of my mother.

iii
Acknowledgments
My advisor, a visionary man, Professor Donald Paul, has continually motivated me to
explore various aspects of the energy industry. My co-advisor, Professor Iraj Ershaghi, has
been a major support and source of inspiration for my educational career in the past ten
years. Professor Birendra Jha has been an invaluable source of support during the last
stages of my Ph.D. and helped me develop a science-oriented mentality towards my work.
I would like to thank my Father, Dr. Fazlali Ashayeri, who encouraged me to never stop
learning.
During my years as a Ph.D. student, I was privileged to work as a TA with distinguished
scholars such as Professor Jincai Chang, Professor Kristian Jessen, and Professor Faruk
Civan. I have learned a lot from these individuals.
Several other scholars have been highly supportive of my research in the past few years.
The list includes Professor Najmedin Meshkati, a friend and a life mentor, Professor
Hashem Pesaran, a world-class scientist, Professor Jeff Nugent, Dr. Hossein Pourmand, Dr.
Andrei Popa, Dr. Manouchehr Takin, Dr. Parvis Mina, Professor Behrokh Khoshnevis, Dr.
Donald Hill, Professor Fred Aminzadeh, and Dr. Hossein Alimi.
I also need to mention the names of several graduate students at USC that have helped me
in various ways to achieve my goals. Dr. Azaranag Golmohammadi, Dr. Marjan Sherafati,
Dr. Mahshad Samnejad, Yahya Taghavifar, Asal Rahimi Zeynal, Dr. Hasan Shojaei, Dr.
Babak Zareyian, Dr. Mehran Hosseini, Magdalene Ante, Dr. Atefeh Jahandideh, Dr.
Mohammad Evazi, Dr. Mohammad Javaheri, Minh Tran, Saro Meguerdijian, Dr. Nima

iv
Jabbari, Ted Hayrapetian, Dr. Amirhossein Eftekharian. I will always treasure my
friendship with these brilliant individuals.
Dr. Shahram Farhadi Nia has been a great friend and contributed significantly to my
professional goals. Dr. Mahmood Shirooyeh, Dr. Vahid Arbab, Dr. Ehsan Barjasteh, and
Dr. Arash Saif are among the most valuable relationships that I established at USC.
I would also like to mention the help of Dr. Mehdi Korjani, who generously provided his
expertise.
USC administration and staff have always made themselves available and went above and
beyond to support my education. Andy Chen, Idania Takimoto, Kelly Goulis, Juli Legat,
Sorin Marghitoiu, Tina Silva, Shokry Bastorous, and Anthony Tritto.
I would also like to extend my appreciation to Chevron Corporation. Part of my doctoral
studies was supported by the Chevron Ph.D. Fellowship in Energy Resources.
Finally, I want to express my deepest gratitude to two of the most adorable women in the
world, my sister, Shylan, and my fiancé, Ashlee.

v
Table of Contents
Dedication ............................................................................................................................................. ii
Acknowledgments ............................................................................................................................. iii
List of Tables .................................................................................................................................... viii
List of Figures ..................................................................................................................................... xi
Abstract .......................................................................................................................................... xviii
Chapter 1: Introduction ..................................................................................................................... 1
1.1 Background ................................................................................................................................... 1
1.2 Introduction ................................................................................................................................... 8
1.3 Research Focus ........................................................................................................................... 14
Chapter 2: Literature Review .......................................................................................................... 22
2.1 Background ................................................................................................................................. 22
2.2 Data-driven modeling for oil and gas ........................................................................................ 32
2.2.1 Rule-based Artificial Intelligence and Expert Systems ....................................................... 34
2.2.2 Data selection, data preparation, unsupervised, and supervised learning ........................... 38
2.2.3 Clustering and classification techniques .............................................................................. 43
2.2.4 Supervised vs. Unsupervised ................................................................................................ 46
2.2.5 Artificial Neural Network and Deep Learning ..................................................................... 46
Chapter 3: Methodology ................................................................................................................... 52
3.1 Introduction ................................................................................................................................. 52
3.2. Unconventional oil and gas data landscape .............................................................................. 54
3.3 Sample Data Description and Statistical Analysis..................................................................... 58
3.4 Data preparation and exploratory phase..................................................................................... 61
3.5 Role of Petroleum Engineer in handling pitfalls in data collection, data preparation, and data
description ......................................................................................................................................... 64
3.6 Understanding data ranges and Outliers .................................................................................... 69
3.7 Analysis of Production Decline Behavior and Curve Fitting .................................................... 75
Chapter 4: Using the Adaptive Variable Structure Regression Approach in Data Selection
and Data Preparation for Improving Machine Learning-Based Performance Prediction in
Unconventional Plays ........................................................................................................................ 88
4.1 Introduction ................................................................................................................................. 88
4.2 Problem Statement ...................................................................................................................... 92
4.3 Methodology ............................................................................................................................... 94

vi
4.4 Using VSR for data selection and preprocessing ...................................................................... 99
4.5 Examination of VSR on simulation data ................................................................................. 101
4.6 Comparison of VSR and Random Forest ................................................................................ 105
4.7 Deploying VSR on the unconventional dataset ....................................................................... 108
4.8 Conclusions ............................................................................................................................... 117
Chapter 5: Evaluation of the Role of Domain Expert in Transfer Learning Applications of
Machine Learning in the Unconventional Resources ................................................................. 119
5.1 Introduction ............................................................................................................................... 119
5.1.1 Background ......................................................................................................................... 121
5.1.2 Transfer Learning for Unconventional Resources Modeling ............................................ 124
5.2 Methodology ............................................................................................................................. 126
5.2.1 Unsupervised learning for dividing the dataset .................................................................. 127
5.2.2 Predictive Modeling and Model Selection ......................................................................... 128
5.2.3 Deploying Random Forest .................................................................................................. 129
5.3 Results ....................................................................................................................................... 130
5.3.1 Application of Random Forest on data of the entire basin ................................................ 130
5.3.2 Combining Random Forest with unsupervised clustering of the asset ............................. 134
5.4 Transfer Learning implementation ........................................................................................... 146
5.4.1 Model Transfer .................................................................................................................... 148
5.4.2 Hyperparameter fine-tuning ................................................................................................ 162
5.4.3 Adding time-dependent data in the training dataset ........................................................... 164
5.5 Chapter Discussion and conclusions ........................................................................................ 166
Chapter 6: Stochastic Investigating of Basin-wide Underlying Distribution Functions of
Decline Rate Behavior for Unconventional Resources ............................................................... 168
6.1 Introduction ............................................................................................................................... 168
6.2 Methodology ............................................................................................................................. 170
6.3 Geologic heterogeneity clustering ............................................................................................ 178
6.4 Investigating the underlying statistical distributions ............................................................... 182
6.5 Application of identification and quantification of distribution function ............................... 186
6.6 Conclusions and future work .................................................................................................... 188
Chapter 7: Conclusions and Future Work .................................................................................. 191
7.1 Concluding Remarks ................................................................................................................ 191
7.2 Future Work .............................................................................................................................. 197

vii
Glossary ............................................................................................................................................ 199
References ........................................................................................................................................ 200
Appendix .......................................................................................................................................... 207

viii
List of Tables
Table 1-1. Main consideration in different stages of data-driven modeling. .................................... 13
Table 2-1. Main difference between conventional and unconventional hydrocarbon resources in
terms of production physics. ............................................................................................................... 24
Table 3-1. Examples of raw data and interpreted data in the unconventional oil and gas projects. 55
Table 3-2. Data description for main variables compiled in the Eagle Ford dataset (Segment 1:
Well information). ............................................................................................................................... 59
Table 3-3. Train and test errors using four different predictive techniques on the pre/post-2014
datasets (OLS and Bayesian GLM). ................................................................................................... 71
Table 3-4. Various cases of decline function based on (b) values (Ilk et al., 2010). ........................ 82
Table 3-5. Summary of main categories of decline curve analysis for the unconventional resources.
.............................................................................................................................................................. 82
Table 3-6. The influence of lack of, or inaccurate, the measurement for the input parameters for
decline analysis models (from Moridis et al., 2010). ......................................................................... 83
Table 4-1. Variable importance ranking based on VSR. ................................................................. 100
Table 4-2. The results of VSR variable importance rules for simulation example (local
Interpretability). ................................................................................................................................. 104
Table 4-3. The results of VSR variable importance rules for simulation example (Global
Interpretability). ................................................................................................................................. 105
Table 4-4. The results of VSR variable importance rules for the comparison test. ........................ 106
Table 4-5. The portion of predictions within 20% error margin and train and test RMSE (per barrel
of oil) for the comparison test. .......................................................................................................... 107
Table 4-6. Ranking of variable importance for first-year cumulative production target response
reported by VSR for the comparison test. ........................................................................................ 107
Table 4-7. Median and Mean errors for train and test datasets using Random Forest and VSR
algorithms. ......................................................................................................................................... 108
Table 4-8. Summary of the average geologic, engineering and production values for the ten
clusters created by formation depth and thickness. .......................................................................... 113

ix
Table 4-9. Reduced rules extracted by VSR showing the importance of four geologic variables in
the two-year cumulative production of wells in clusters 2 and 8. ................................................... 114
Table 4-10. Summary of the average geologic, engineering and production values for the ten
clusters created by surface latitude and surface longitude of the wells. .......................................... 115
Table 4-11. Rules extracted by VSR showing the importance of four geologic variables in the two-
year cumulative production of wells in clusters 1 and 7 as defined in ten clusters by well
coordinates. The last column shows the number of instances (wells) that are within each rule. ... 116
Table 4-12. Reduced rules extracted by VSR showing the importance of four geologic variables in
the two-year cumulative production of wells in clusters 1 and 7 created by well coordinates. ...... 116
Table 5-1. Train and test RMSE errors for the Eagle Ford dataset by 70-30 split ratio on the entire
basin. .................................................................................................................................................. 131
Table 5-2. Impact of feature selection (number and combination of predictor variables) on
Random Forest performance (cases 1-5). ......................................................................................... 133
Table 5-3. Summary of the average variable values in all 9 clusters demonstrated above (number
of clusters is chosen randomly to demonstrate the workflow). ....................................................... 136
Table 5-4. Error-values for train and test datasets in all possible combinations of the total number
of clusters changing from 5 to 10 (only the instances with median test error smaller than 25% are
shown here). ...................................................................................................................................... 137
Table 5-5. Performance of predictive model developed on cluster ID 6 with other clusters as the
test data set......................................................................................................................................... 139
Table 5-6. Median RMSE error values for all cluster combinations based on surface latitude and
longitude. ........................................................................................................................................... 140
Table 5-7. Descriptive statistics for different clustering numbers based on well coordinates. ...... 141
Table 5-8. Variables used for clustering in different cases. ............................................................ 143
Table 5-9. Summary of test set error statistics for different cases. ................................................. 143
Table 5-10. Summary of cluster-wide mean values in the case of ten clusters based on TVD and
API gravity. ....................................................................................................................................... 147
Table 5-11. Summary of average values for some of the variables in the six clusters generated in
the case study. .................................................................................................................................... 156
Table 5-12. Errors associated with all possible binary non-repeat train and test combinations for
the six-cluster case............................................................................................................................. 156

x
Table 5-13. The sample size for source, target, validation, and train subsamples after adding 7% of
the total wells from Filed B to the training dataset. ......................................................................... 159
Table 6-1. Various decline function types for the unconventional wells (source: Temizel et al.,
2020). ................................................................................................................................................. 175
Table 6-2. Summary of decline power-law coefficient distribution functions and parameters. .... 186
Table 6-3. Cluster summary for average geologic and engineering parameters of horizontal wells.
............................................................................................................................................................ 188

xi
List of Figures
Figure 1-1. Production of tight oil from unconventional resources in the United States between
2010 and 2020. Breakdown by major plays (Source: EIA, 2020). ...................................................... 2
Figure 1-2. Location of leading US unconventional basins (from Usman and Meehan, 2019 - EIA
2012). ..................................................................................................................................................... 3
Figure 1-3. Frequency of SPE papers published with Data-Driven and Unconventional keywords
between 2010 and 2020 (from the OnePetro database). ...................................................................... 6
Figure 2-1. Classic reservoir engineering problems that are simulation-dependent (Jansen et al.,
2009). ................................................................................................................................................... 23
Figure 2-2. Permeability ranges for conventional and unconventional reservoirs (from Navarette et
al., 2014). ............................................................................................................................................. 25
Figure 2-3. Different fluid flow regimes based on permeability ranges (from Zhang, Liehui, 2019).
.............................................................................................................................................................. 25
Figure 2-4. 3D and cross-sectional view of the multistage fractured horizontal well (Ahmed et al.,
2017). ................................................................................................................................................... 26
Figure 2-5. Three-dimensional view of multistage hydraulic fracturing job with poor stages
identified (Ahmed et al., 2017). .......................................................................................................... 26
Figure 2-6. Two-dimensional (cross-section) view of fracture conductivity (Ahmed et al., 2017).27
Figure 2-7. Generalized workflow for unconventional reservoir simulation process (Altman,
2018). ................................................................................................................................................... 28
Figure 2-8. A summarized workflow for various inputs incorporated with uncertainty in the
assessment of EUR for unconventional wells (source: this research). .............................................. 31
Figure 2-9. Various roles in a typical data-driven modeling workflow (source: inspired by the
unknown creator from internet sources). ............................................................................................ 32
Figure 2-10. Process of petrophysics used by HESPER (Peveraro, 1988). ..................................... 35
Figure 2-11. Sandstone tree used by HESPER (Peveraro, 1988). .................................................... 36
Figure 2-12. Generic data-driven framework for predicting geological and petrophysical properties
(Xu et al., 2019). .................................................................................................................................. 41
Figure 2-13. Comparison among publications on different machine learning categories between
unconventional use cases and general applications (source: OnePetro database as of April 2021). 43

xii
Figure 2-14. Architecture of an artificial neuron and a multilayered neural network (from
Abraham, 2005). .................................................................................................................................. 47
Figure 2-15. Surrogate modeling workflow (Source: Shuai Guo, 2020). ........................................ 49
Figure 3-1. Data frequency vs. Decision frequency in the Upstream oil and gas industry (source:
this research). ....................................................................................................................................... 55
Figure 3-2. Various well log data types. ............................................................................................ 56
Figure 3-3. Mud log data. ................................................................................................................... 56
Figure 3-4. Core data and various core test data. .............................................................................. 57
Figure 3-5. Drilling dynamic data. ..................................................................................................... 57
Figure 3-7. Spatial distribution of the fields based in Eagle Ford, South Texas, USA, whose data
has been utilized in this study. Well locations have been color-coded based on API Gravity. (The
map has been generated using Tableau Desktop Professional, Version 2021). ................................ 59
Figure 3-8. Correlation matrix for the independent input variables for the unconventional dataset
(this research). ..................................................................................................................................... 63
Figure 3-9. Correlation among some of the geologic parameters by 2-D visualization. ................. 64
Figure 3-10. Average first-year quarterly production in terms of barrels of oil equivalent (BOE)
for Eagle Ford wells by On Stream date. A consistent upward trend in the seven years of the study.
.............................................................................................................................................................. 65
Figure 3-11. Variance in average proppant fluid in three-year intervals before and after
01/01/2014. .......................................................................................................................................... 66
Figure 3-14. Increase in the average mass of proppant injected per unit of lateral length of the
horizontal wells between 2010 and 2016. .......................................................................................... 68
Figure 3-15. Distribution of average fluid and proppant injected per lateral foot of the wells.
Color-coded based on the first year of cumulative oil recovery (source: this research). .................. 69
Figure 3-16. Distribution of average fluid and proppant injected per lateral foot of the wells.
Color-coded based on the first year of cumulative oil recovery (source: this research). .................. 70
Figure 3-17. Train and test errors using four different predictive techniques on the pre/post-2014
datasets. The first row below the bars shows the training dataset, and the second row shows the test
dataset (source: this research). ............................................................................................................ 72
Figure 3-18. Variables’ importance ranking (including the noise variable) by OLS ....................... 74

xiii
Figure 3-19. Variables’ importance ranking (including the noise variable) by RF. ........................ 74
Figure 3-20. Variables’ importance ranking (including the noise variable) by GBM. .................... 75
Figure3-21. Variables’ importance ranking (including the noise variable) by Bayes GLM. .......... 75
Figure 3-22. The simplified production profile of a typical conventional well or field (Jakobsson et
al., 2014). ............................................................................................................................................. 76
Figure 3-23. Typical decline behavior, for example, unconventional oil wells from the Eagle Ford
basin. Significant drop compared to initial production and peak rate within the first year of
production (source: this research). ...................................................................................................... 77
Figure 3-24. The number of months on production prior to rate drop below 10% of peak oil rate
(source: this research).......................................................................................................................... 80
Figure 3-25. The number of months on production prior to rate drop below 10% of peak gas rate
(source: this research).......................................................................................................................... 81
Figure 3-26. The month of maximum oil rate (peak production) for Eagle Ford wells. ................. 84
Figure 3-27. The month of maximum gas rate (peak production) for Eagle Ford wells. ................ 85
Figure 3-28. Power law decline for a sample well. ........................................................................... 86
Figure 3-29. Distribution of error in 5-year cumulative production of oil (BOE) by using fitted
power law. ........................................................................................................................................... 86
Figure 4-1. Piecewise-linear MFs. ..................................................................................................... 95
Figure 4-2. Variable importance ranking based on Random Forest. .............................................. 101
Figure 4-3. 3D view of the test reservoir used for data generation by CMG simulation software for
Random Forest (Academic License used for CMG Version 2019 for educational purpose). ........ 102
Figure 4-4. Distribution of percentage of error for predicted vs. actual production for Simulation
data test. The vertical axis is the 3-year cumulative oil production in barrels, and the horizontal axis
well counts. ........................................................................................................................................ 105
Figure 4-5. Distribution of percentage of error for predicted vs. actual production for the train
(left) and validation (right) datasets. The vertical axis is the cumulative production based on
million barrels of oil, and the horizontal axis demonstrates the data points sorted by cumulative
production volume............................................................................................................................. 106
Figure 4-6. Ranking of variables importance reported by RF based on IncNodePurity. This
diagram is comparable to Table 4-6. ................................................................................................ 108

xiv
Figure 4-7. K-means clustering for ten clusters on two variables (left) or the first two principal
components capturing almost 75% of the data variance (right). ..................................................... 110
Figure 4-8. The geographical location of all the wells was divided into ten clusters using TVD and
Thickness. .......................................................................................................................................... 111
Figure 4-9. Location of clusters 1 and 7 in the ten-cluster case created by well coordinates (latitude
and longitude). ................................................................................................................................... 115
Figure 5-1. Modeling requirement based on the dimensionality and scale of the field planning
problem (from Klie et al., 2020). ...................................................................................................... 123
Figure 5-2. Learning process in traditional machine learning vs. Transfer Learning approach
(inspired by Pan & Yang, 2010). Fields shown here are a simplified visualization for sub-basins of
the Eagle Ford Play used in this research. ........................................................................................ 125
Figure 5-3. Distribution of RMSE in predicting two-year cumulative BOE on the dataset used for
training. .............................................................................................................................................. 131
Figure 5-4. Distribution of RMSE in predicting two-year cumulative BOE on the test dataset. .. 131
Figure 5-5. Two-dimensional visualization of k-means clustering on well coordinates with k=9.
............................................................................................................................................................ 135
Figure 5-6. Location of 9 clusters based on well coordinates superimposed on the map of South
Texas (superimposed using Google maps). ...................................................................................... 135
Figure 5-7. Geological location of clusters 6 and 7, which exhibited the best prediction accuracy
on the test data set. ............................................................................................................................ 138
Figure 5-8. Distribution of RMSE in the Random Forest algorithm in cluster 6 as the training
dataset. ............................................................................................................................................... 138
Figure 5-9. Distribution of RMSE in predicting the two-year BOE of wells in cluster 7 as the test
dataset. ............................................................................................................................................... 139
Figure 5-10. Clusters generated by k-means for 10 clusters using TVD and API gravity. ........... 145
Figure 5-11. The spatial distribution of 10 clusters based on TVD and API gravity. .................... 145
Figure 5-12. Spatial gradient maps for Eagle Ford TVD (log scale) and API gravity. .................. 146
Figure 5-13. Schematic of different Transfer Learning scenarios created by varying the amount of
training data taken from Field B. The impact on the validation median RMSE clearly shows the
role of Transfer Learning in modulating the predictive power of the model. ................................. 150

xv
Figure 5-14. Distribution of error for train dataset using static variables in addition to first six
months of BOE production data. ...................................................................................................... 152
Figure 5-15. Distribution of error for test dataset using static variables in addition to first six
months of BOE production data. In addition, we evaluated the variable importance ranking
reported by Random Forest in Figure 5-16. It is observed that the monthly production data for the
last five months in reverse order (months 6 to 2) have the highest importance among all variables.
............................................................................................................................................................ 152
Figure 5-16. Variable importance ranking as produced by Random Forest. .................................. 153
Figure 5-17. Silhouette measure to determine the optimum number of clusters for k-means. ...... 153
Figure 5-18. 2-D visualization of k-means clustering on PC1 and PC2 for TVD, API Gravity, and
Thickness. .......................................................................................................................................... 154
Figure 5-19. Location of clusters 1 to 6 on the map of Eagle Ford superimposed on South Texas
map. ................................................................................................................................................... 155
Figure 5-20. Error distribution for the training dataset (cluster 2) in the six-cluster case. ............ 157
Figure 5-21. Error distribution for the test dataset (cluster 4) in the six-cluster case. ................... 158
Figure 5-22. Location of source field or Field A (cluster 2) and target field or Field B (cluster 4)
on the Eagle Ford map. ..................................................................................................................... 158
Figure 5-23. Error distribution for the training dataset. .................................................................. 160
Figure 5-24. Predicted vs. actual value of two-year cumulative production for the training dataset.
............................................................................................................................................................ 160
Figure 5-25. Error distribution for the validation dataset. ............................................................... 161
Figure 5-26. Predicted vs. actual value of two-year cumulative production for the training dataset.
............................................................................................................................................................ 161
Figure 5-27. Change in median absolute error for cumulative 5-year production of BOE. The
graph demonstrates the incremental change in median absolute error in predicting the 5-year
cumulative production of 500 wells based in Eagle Ford after the addition of monthly production
data (from 3 to 18 months). The cumulative BOE production is one of the predictor variables in the
training model. ................................................................................................................................... 165
Figure 6-1. Typical production behavior for Eagle Ford wells with an early peak in the second or
third month, continued by a rapid decline. ....................................................................................... 171

xvi
Figure 6-2. Exponential power-law curve fit for a sample well based on using the first year of
production of BOE. ........................................................................................................................... 172
Figure 6-3. Monthly error in the predicted exponential power-law curve fit for a sample well based
on using the first year of production of BOE. The cumulative error of the first five years is shown
at -12.6%. ........................................................................................................................................... 172
Figure 6-4. Exponential power-law curve fit for a sample well based on using the first year of
production of oil. ............................................................................................................................... 173
Figure 6-5. Monthly error in the predicted exponential power-law curve fit for a sample well based
on using the first year of production of oil. The cumulative error of the first five years is shown at -
10.9%. ................................................................................................................................................ 173
Figure 6-6. Exponential power-law curve fit for a sample well based on using the first year of
production of gas. .............................................................................................................................. 173
Figure 6-7. Monthly error in the predicted exponential power-law curve fit for a sample well based
on using the first year of production of gas. The cumulative error of the first five years is shown at
-21.2%. ............................................................................................................................................... 174
Figure 6-8. Distribution of error in 5-year cumulative production by using the power-law fitted
equation for a sample of 50 wells. .................................................................................................... 174
Figure 6-9. Distribution of best (b) value estimates in the hyperbolic function. ............................ 176
Figure 6-10. Spatial distribution of best (b) values in hyperbolic decline function for a sample of
150 wells. ........................................................................................................................................... 176
Figure 6-11. Distribution of best (b) value estimate in the same 150 well samples after extending
the test range to 0 1), boundary-dominated flow When (0 0 (Eq. 6.2)
Figure 6-17 and Figure 6-18 demonstrate the distribution of power-law leading term for all
the wells within each of the five clusters applied to principal components.

Figure 6-17. Identifying and quantifying Gamma function for the distribution of leading
term for all the wells in cluster 1.
The other parameter in the exponential power law equation is the power degree. Our results
show that the statistical distribution of the power degree is governed by Normal distribution
in almost all cluster number iterations. The two parameters that determine a Gaussian
normal probability distribution are the Mean (μ) and the standard Deviation (σ).

184
(Eq. 6.3)
Figure 6-19 and Figure 6-20 demonstrate the distribution of power-law power degree
values for all the wells within each of the same clusters.

Figure 6-18. A matched gamma distribution with quantified scale and shape for the other
four clusters.
The distribution of error term or deviation between the actual production data and the fitted
power-law curve is approximately Gaussian normal in every cluster, with a mean value
close to zero.
Table 6-2 summarizes the exponential decline power-law coefficients’ underlying
distribution by type and parameters for the five clusters.

185

Figure 6-19. Identifying and quantifying Normal function for the distribution of power
degree for all the wells in cluster 3.

Figure 6-20. Matched normal distribution with quantified mean and standard deviation for
other four clusters.

186
Table 6-2. Summary of decline power-law coefficient distribution functions and
parameters.
Cluster Leading term
(Gamma distribution)
Power degree
(Normal Distribution)
Scale Shape Mean SD
1 10431.5 4.41 - 0.99 0.17
2 4931.8 4.72 - 0.89 0.14
3 9303.3 2.71 -0.91 0.2
4 10767.7 2.74 - 0.93 0.19
5 11028 3.17 - 0.97 0.19

6.5 Application of identification and quantification of distribution function
At this point, by using this workflow, we have identified the distribution types of decline
behavior parameters within each cluster and also quantified the parameters in each of the
two distribution functions. This information allows us to obtain statistical ranges, or
probability intervals, for each of the leading or power terms for a group of oil wells. We
have earlier shown that by knowing these two parameters in a power-law decline equation,
we can match the production trend, and therefore estimate the long-term cumulative
recovery of these wells with acceptable accuracy. Figure 6-20 scatter plot shows the values
for the power-law coefficients for a larger group of wells in the same shale play with at
least two years of production. This plot is color-coded based on the first two years’
cumulative oil production. It is visible that a certain zone is located in the center and
upward, emphasizing the more significant role of leading term or α on cumulative
recovery. This zone of interest includes a group of wells with a combination of leading
term and power degree coefficients that result in higher recoveries. By using the quantified
distribution functions that determine the probability intervals for the factors that were
leading and power term values, the user can calculate the probability or the ratio of the

187
wells within each cluster that would match the desired combination of power-law
coefficients in the interest zone from Figure 6-20.

Figure 6-20. Distribution of decline power-law coefficients and two-year cumulative
production heat map illustrating a potential zone of interest with wells with higher
performance.
From a field development point of view, it is important for oil companies and service
companies to be able to rank the unconventional assets based on their geologic qualities. In
addition, it is important to determine the optimum ranges of drilling and completion
parameters such as lateral drill length, the volume of pumped fluid, and mass of injected
proppant for each individual well. Needless to say, increasing each of these parameters
beyond the optimal ranges will result in over-allocation of capital. Table 6-2 summarizes
well count, the average values for geologic and engineering parameters for wells, and the
two-year and five-year cumulative production volumes.
By analyzing the combination of geologic characteristics of each cluster and the drilling
and completion recipe for the wells in these clusters, we can evaluate the impact of
reservoir qualities on well performance. In addition, the impact of engineering parameters
in the design of drilling and completion on the productivity of wells can be observed. As an

188
example, in Table 6-3, clusters 3 and 5 have very similar depth, thickness, and clay content.
Wells in cluster 3 has a longer lateral (by almost 10%). However, wells in cluster 5 have
been fractured with significantly larger volumes of fluid and proppant per lateral length on
average. The average cumulative production in cluster 5 wells is more than 20% higher
than the wells in cluster 3.
Table 6-3. Cluster summary for average geologic and engineering parameters of horizontal
wells.
Cluster
ID
Number

Well
Count

Avg.
formation
depth (ft)

Avg.
formation
thickness
(ft)

Avg.
clay
content

Avg.
lateral
section
(ft)

Avg.
proppan
t (lb per
GPI)

Avg.
inj.
fluid
(gal
per
GPI)

Avg.
cum. 2-
year
prod.
(bbl)

Avg.
cum. 5-
year
prod.
(bbl)

1 114 11,112 151 0.28 4,610 2,127 1,756 124,455 152,474
2 54 8,299 421 0.43 7,205 1,419 1,247 78,359 99,590
3 272 7,641 120 0.26 6,650 943 760 75,223 96,524
4 332 10,693 136 0.27 5,225 959 755 86,653 116,269
5 212 8,771 120 0.25 6,049 1,621 1,438 95,055 116,676

6.6 Conclusions and future work
As mentioned earlier, the unsupervised learning methods used in clustering the data with
multiple numbers of geologic parameters result in creating more well-behaved and
identifiable distribution functions in a large-scale shale play. However, the main challenge
in such an approach is the interpretability of the causal relationship among the qualities that
distinguish the clusters and the higher performance of oil wells. Further sensitivity analysis
steps may allow a better understanding of the individual impact of these qualities on the
success of drilling and hydraulic fracturing operations. Another source of the challenge is
the lack of accuracy for Greenfield shale plays with few numbers wells and shorter

189
production history. This method relied on a minimum number of wells to create a
statistically large enough sample to create reliable distribution curves. In addition, at least
18 months of production is required to establish robust decline curve fits. Therefore, the
proposed workflow is more useful in unconventional assets that have been through the
early stages of development.
Another main challenge in this part of the research was due to limited computational power
on a personal laptop and lack of access to a powerful computer. The loops created for
estimation of the best hyperbolic decline function, which required examination of close to
200 (b) values for each well, did not allow expanding the research scope for many of the
hypotheses to more than 150 wells.
This is a repeatable stochastic approach for the identification and quantification of
underlying statistical distribution functions that determine the decline behavior of
unconventional wells. This approach allows understanding the impact of the variance in
geologic properties and drilling and completion design parameters on the productivity of a
large group of wells. Field development risk can be evaluated by analyzing the efficiency
of drilling longer laterals or increasing the fluid or proppant volumes during the hydraulic
fracturing stage in different sections of the asset. It is noteworthy to mention that the data
types used in this research are very limited and mostly from publicly available sources. The
operators and service companies have access to a more comprehensive database. However,
the proposed workflow can be modified for a wider range of data types.

190
In addition, breaking the production decline curve into multiple zones which are
representative of different decline regimes is an active area of research that a combination
of machine learning and a domain expert can help develop.

191
Chapter 7: Conclusions and Future Work
7.1 Concluding Remarks
In this final chapter, we review the highlights of our research and briefly mention the main
findings. We also discuss some of the issues related to the application of data science in the
industry and try to relate our research to the areas in which our findings can help improve
the utilization of data-driven modeling in industrial applications.
In addition, we discuss some of the areas related to this thesis that we believe require
further attention, and we hope that our work can be of help to future engineers and
researchers, and they can use our proposed workflows to explore more opportunities of
integrating data-driven modeling with domain expertise.
One of the most important findings in this research is the role of using domain-specific
knowledge in framing the problem and the solution. In other words, in a reservoir
engineering context, the processes involved in a data-driven workflow should be devised
and guided by petroleum engineers. Obviously close collaboration of the engineering team
and the data science team is an essential component of this process. However, many of the
early-stage tasks such as data selection, data cleaning, and data preparation are considered
less important tasks. Handling these stages without the involvement of petroleum
engineering expertise can severely impact the accuracy of the model. In our research we
provided examples of how lack of understanding in the time sensitivity of some of the
completion design parameters can influence the model accuracy. We quantified this impact
by using real-world data and testing the impact of technology learning curve over a six-year
period of time.

192
Another important aspect of using data-driven modeling in unconventional resource
modeling is to investigate the importance of different variables on the productivity of the
wells. Variable importance ranking for several data-driven techniques was assessed in our
work. In this effort, we utilized different approaches, including the utilization of physics-
based simulation software to create ideal (or perfect) datasets, the introduction of random
noise parameters to the feature space, and hypothetical scenario generation on real-world
field data. This exercise allowed us to identify the model types that perform better in the
representation of variable importance ranking based on a domain expert’s opinion against
simulated data and identification of noise. Our results suggested that some of the linear
models that use prior knowledge in the training datasets can produce accurate variable
importance ranking. However, these models have relatively poor performance in predictive
accuracy.
This observation encouraged us to explore non-linear models that are capable of producing
high accuracy in both variable importance ranking and predictions. We were able to
identify a non-linear regression model (developed for non-petroleum applications) that uses
fuzzy logic in order to extract overarching rules that explain the relationship among various
model inputs and the target response. We devised a series of tests to examine this algorithm
against some of the most advanced ensemble learning ML methods and successfully
showed that this algorithm can generate rules that are simply understood by non-data
scientist users. These rules are capable of explaining the variance of data in sub-groups of
the training datasets. This functionality allows an improved data selection process while
dealing with large volumes of data where the relationship among some data types with the

193
target response is not known or fully explained by the physics. We also demonstrated that
this new algorithm could achieve high accuracy on both simulation-generated data as well
as real-world data.
Another major step in the process of data-driven modeling is model selection. In this
research, we conducted a comprehensive survey of the literature and observed that the
application of advanced machine learnings such as deep learning techniques has a rapidly
increasing trend in the oil and gas industry literature. We also discussed that the main
motivation for using such methods is the success of deep learning techniques in some of the
other technology areas, such as Information Technology. It is important to understand that
many of these algorithms have been developed for non-petroleum engineering problems.
Many of the deep learning techniques are deterministic black-box models by nature. As a
result, they demonstrate outstanding accuracy in certain predictive applications while they
lack descriptive capabilities. In addition, we discussed that in order to establish a causal
relationship among many model inputs and the productivity of the unconventional oil and
gas wells, a repeatable stochastic model is required.
One of the main motivations behind the widespread use of deep learning in the oil is the
industry's focus on a limited range of success metrics to evaluate data-driven modeling.
The majority of projects define values such as a certain error percentage threshold or R-
squared value as the success metric for data-driven modeling projects. In addition, in
several projects, the point of comparison for the performance of data-driven models are the
results of simulation models that have been calibrated on a particular asset for several
years. Such success metrics put data-driven modeling at a disadvantage for several reasons.

194
One of the most important aspects of data-driven modeling is the volume and the quality of
the data used for training purposes. As a result, it is expected that the performance of a
model based on ML or AI techniques would gradually improve over time by the addition of
data collected from the subject field. This means that it is important that oil and gas
companies develop data-driven frameworks which allow the addition of data over the
lifetime of the asset and facilitate incremental improvements over long periods of time.
This point of view allows a more flexible range of success metrics. As an example, a data-
driven workflow capable of capturing learnings from previous data and continually
improve those learnings by ingestion of new data should not be evaluated based on a single
forecast but rather a wide range of forecasts made in a long period of time on multiple
datasets.
Another important aspect of data-driven modeling is developing scalable models that can
be trained on one dataset and further used to predict the performance of other datasets. In
the context of unconventional oil and gas, this aspect is even more pronounced due to
insufficient physics-based simulation models. It is important that the models trained on one
basin can be reused elsewhere to improve the prediction of well performance in another
similar unconventional play. In this research we developed a Transfer Learning framework
that allows a combination of the two above-mentioned aspects of data-driven modeling.
We combined unsupervised clustering techniques with supervised learning to evaluate the
efficiency of using models trained on one subsection of a shale basin on another subsection
of the same basin with different geologic properties. In this workflow, we were able to
show that incorporating the domain knowledge in selecting the clustering variables can

195
help improve the success of transfer learning. In addition, we demonstrated that including
production data in the predictors’ feature space can significantly improve the model
performance and reduce median error margins on validation datasets from the target fields.
In this research, we discuss the importance of domain expertise in feature selection in the
Transfer Learning context. As an example, considering the practical development process
in a relatively new (greenfield) asset requires avoiding the use of long-term production
history in model training. This is because in a new field, the operators do not benefit from
the luxury of several years of production history. Therefore, a maximum of six months has
been used in our approach due to the assumption that such data will become available after
the exploration phase in a new field. We also discuss the considerations in using the data
related to the geographical location of the wells (coordinates) in the Transfer Learning
context. In our research and by evaluating the variable importance ranking for the model
inputs, we observed that surface Latitude and surface Longitude of the unconventional
wells in a mature basin include information regarding the spatial heterogeneity of the asset
because these parameters are essentially spatial proxies of target formation geologic and
petrophysical properties. Including these features in the training dataset will introduce a
bias to the model, while the well coordinates for potential drill locations in the target field
do not contain such information.
In another part of this research, we demonstrated that using fundamental statistical concepts
coupled with domain knowledge can help to develop stochastic models to analyze the
production decline behavior of the unconventional basins. We were able to estimate and
quantify the underlying distribution functions that explain the long-term production of a

196
large group of unconventional wells with high levels of accuracy. Similar to our approach
in other parts of the research, we developed streamlined workflows that can conduct
several thousand iterations on samples with hundreds of wells. This effort can fill a gap in
the literature where the majority of decline curve and rate transient analysis rely on
individual well observations. We demonstrate that production decline data analytics is one
of the main areas that can benefit from the advantages of big data analytics and machine
learning. We were able to examine various types of hyperbolic and exponential decline
functions in a large sample. The results are a wide range of decline parameters. The statical
distribution of these parameters can further be studies to discover the potential relationship
among geologic properties and well design parameters with the long-term decline behavior
of the unconventional wells.
Our last recommendation based on observations in this research, as well as interaction with
industry experts, is that for achieving the ultimate goal of using data-driven modeling in the
context of unconventional oil and gas resources, two fundamental items should be
considered:
1) We recommend that researchers do not highly rely on many of the automated features
that are embedded in advanced machine learning techniques and libraries. It is true that
examples such as automated fine-tuning or optimal range finders can accelerate the
modeling process, however the same features can detach the domain expert from engaging
with important aspects if the predictive model training process. For better underrating
causality in these types of problems it is important that the domain experts engage with the

197
development of the model as much as possible and avoid over-using of automated
functionalities.
2) It is important not to take the knowledge of petroleum engineering as granted during the
collaboration between data science and petroleum engineering teams. Many aspects of the
reservoir engineering and drilling or completion engineering that are considered common
sense for the domain experts are completely unknown to the data science teams. It is crucial
that during the problem investigation and solution framework design, guiding rules are
defined by the petroleum engineering teams that prevent disregard of fundamental physics
of the problem.
7.2 Future Work
In this section, we introduce two main research areas that our work can be continued. The
first area is the expansion of our Transfer Learnings workflow to other types of geologies
and more rich datasets with higher levels of granularity. The second area is the expansion
of the data-driven workflow developed in this research to explore the changing production
regimes in individual wells’ production history.
As discussed in Chapter 5 of this thesis, our proposed Transfer Learning workflow is
confined to the limits of the Eagle Ford Basin due to the limitations of our compiled
dataset. Deploying the full cycle workflow presented here on similarly large plays such as
the Bakken in North Dakota or the Permian Basin in Texas can show the strengths and
weaknesses of our proposed methodology. In addition, expanding the feature space to
include more data types, especially proprietary and confidential sources of data only

198
accessed by asset owners and service companies, can enhance the performance of our
models.
In Chapter 6 we were able to evaluate the impact of changing (b) factors by small
increments and estimate the best hyperbolic function fit for several hundred wells in the
Eagle Ford Basin. One of the active research areas in the application of data science in
time-series analysis of production decline curves is a multi-segment analysis of the
production history. Identifying various stages of depletion regimes such as boundary-
dominated flow, transient flow, and exponential decline through an automated estimator
can significantly improve the predictive capabilities of petroleum engineering teams.
Development of such model requires careful guidance of domain knowledge because
taking into account the physical understanding of the reservoir and its impact on the decline
behavior of the wells is essential in devising a data-driven framework to capture the
changes in decline patterns for multiple wells. Another layer of complexity in this approach
is non-quantitative data and model inputs required for well decline behavior interpretation.
As an example, well events such shut-ins, workovers, stimulation jobs, re-frac jobs, and
well interference events need to be gathered from scattered and usually non-digital sources.
The absence of such information from the datasets used by data science teams without
adequate guidance from petroleum engineering teams will compromise the success of any
data-driven project.

199
Glossary
φ
𝑣
Non-linear regressor function
𝑅 𝑠 Optimal prediction number
𝛽 𝑣 Rules’ importance variable
𝑆 𝑣 Rules based on Min-Max Theorem
𝑣 Index of the surviving set
Ai Membership function of a term assigned to a variable
𝐹 𝑣 𝑆 Surviving casual combination max arg
µ
𝐹 Min-Max set
mtry Number of variables tried at each Random Forest node
VSR Variable Structure Regression
mD Millidarcy
psi Pound per square inch
bbl Barrel of oil
Mbbl Thousand barrel of oil
Mscf Thousand standard cubic feet
MMscf Million standard cubic feet
TOC Total Organic Carbon
API American Petroleum Institute
Vclay Clay volume
SPUD Well drilling start date
GPI Gross Perforated Interval
lb Pound
gal Gallon
Cum Cumulative
Lat Latitude
Lon Longitude
Perm Permeability
SW Water saturation

200
References
1. Mohaghegh, Shahab D. "Shale analytics." Shale Analytics. Springer, Cham, 2017. 29-
81.
2. Charpentier, Ronald R., and Troy A. Cook. "Improved USGS methodology for
assessing continuous petroleum resources." US Geological Survey Data Series 547.2
(2010): 22.
3. Map of Assessed Continuous (Unconventional) Oil Resources in the United States,
2014 By U.S. Geological Survey National Assessment of Oil and Gas Resources Team,
and Laura R.H. Biewick.
4. "Trends in US oil and natural gas upstream costs." US Energy Information
Administration (2016).
5. Ashayeri, Cyrus, and Iraj Ershaghi. "OPEC and Unconventional resources." SPE
Annual Technical Conference and Exhibition. Society of Petroleum Engineers, 2015.
6. Mishra, Srikanta, and Luan Lin. "Application of data analytics for production
optimization in unconventional reservoirs: A critical review." Unconventional
Resources Technology Conference, Austin, Texas, 24-26 July 2017. Society of
Exploration Geophysicists, American Association of Petroleum Geologists, Society of
Petroleum Engineers, 2017.
7. Hastie, Trevor, Robert Tibshirani, and Martin Wainwright. Statistical learning with
sparsity: the lasso and generalizations. CRC Press, 2015.
8. Altman, Raphael, et al. "Applied Learnings in Reservoir Simulation of Unconventional
Plays." SPE Latin American and Caribbean Petroleum Engineering Conference.
Society of Petroleum Engineers, 2020.
9. Flichy, Philippe, and Claude Baudoin. "The Industrial IoT in Oil & Gas: Use
Cases." SPE Annual Technical Conference and Exhibition. Society of Petroleum
Engineers, 2018.
10. Cuddy, Steve. "The Benefits and Dangers of Using Artificial Intelligence in
Petrophysics." SPWLA 61st Annual Logging Symposium-Online. Society of
Petrophysicists and Well-Log Analysts, 2020.
11. Peveraro, R. C. A., and J. A. Lee. "HESPER: An expert system for petrophysical
formation evaluation." European Petroleum Conference. Society of Petroleum
Engineers, 1988.
12. Xu, Chicheng, et al. "When Petrophysics Meets Big Data: What can Machine
Do?." SPE Middle East Oil and Gas Show and Conference. Society of Petroleum
Engineers, 2019.
13. Bu, Tor, and Eivind Damsleth. "Errors and uncertainties in reservoir performance
predictions." SPE Formation Evaluation 11.03 (1996): 194-200.
14. Stoneburner, Richard K. "The Eagle Ford Shale field in the Gulf Coast Basin of South
Texas, USA: A “perfect” unconventional giant oil field." (2017): 121-140.
15. Clarke, Paul R., et al. "Assessing well performance in a prolific liquids-rich shale
play—An Eagle Ford case study." (2016): 213-240.
16. Jansen, J-D., et al. "Closed-loop reservoir management." First Break 23.1 (2005).

201
17. Milad, Benmadi, et al. "Practical compartmental model for shale-gas production by
multiple-fractured horizontal wells considering transport influenced by nano-pore
proximity." SPE Unconventional Resources Conference. Society of Petroleum
Engineers, 2014.
18. Bell, Michael Z. "Why expert systems fail." Journal of the Operational Research
Society 36.7 (1985): 613-619.
19. Kravis, Simon, and Rosemary Irrgang. "A case-based system for oil and gas well
design with risk assessment." Applied Intelligence 23.1 (2005): 39-53.
20. Moridis, George J., et al. "A Self-Teaching Expert System for the Analysis, Design,
and Prediction of Gas Production From Unconventional Gas Resources." Canadian
Unconventional Resources Conference. Society of Petroleum Engineers, 2011.
21. Ball, Kathy, et al. "The rise of the machines, analytics, and the digital oilfield: Artificial
intelligence in the age of machine learning and cognitive analytics." Unconventional
Resources Technology Conference, Austin, Texas, 24-26 July 2017. Society of
Exploration Geophysicists, American Association of Petroleum Geologists, Society of
Petroleum Engineers, 2017.
22. Harris, Cole. "Potential pitfalls in exploration and production applications of machine
learning." SPE Western North American and Rocky Mountain Joint Meeting. Society
of Petroleum Engineers, 2014.
23. Ahmed, I., et al. "Cracking the code of unconventional reservoir simulation." SPE
Kingdom of Saudi Arabia Annual Technical Symposium and Exhibition. Society of
Petroleum Engineers, 2017.
24. Brownlee, Jason. "Tactics to combat imbalanced classes in your machine learning
dataset." Machine Learning Mastery 19 (2015).
25. Eshkalak, Mohammad Omidvar, Shahab D. Mohaghegh, and S. Esmaili. "Synthetic,
geomechanical logs for Marcellus shale." SPE Digital Energy Conference. Society of
Petroleum Engineers, 2013.
26. Akinnikawe, Oyewande, Stacey Lyne, and Jon Roberts. "Synthetic well log generation
using machine learning techniques." Unconventional Resources Technology
Conference, Houston, Texas, 23-25 July 2018. Society of Exploration Geophysicists,
American Association of Petroleum Geologists, Society of Petroleum Engineers, 2018.
27. Klie, Hector, Bicheng Yan, and Arturo Klie. "Transfer Learning for Scalable
Optimization of Unconventional Field Operations." SPE/AAPG/SEG Unconventional
Resources Technology Conference. Unconventional Resources Technology
Conference, 2020.
28. Jakobsson, Kristofer, et al. "Bottom-up modeling of oil production: A review of
approaches." Energy Policy 64 (2014): 113-123.
29. Surguchev, Leonid, Eduardo Manrique, and Vladimir Alvarado. "Improved oil
recovery: status and opportunities." 18th World Petroleum Congress. World Petroleum
Congress, 2005.
30. Morrow, Norman, and Jill Buckley. "Improved oil recovery by low-salinity
waterflooding." Journal of Petroleum Technology 63.05 (2011): 106-112.
31. Wang, Dongying, et al. "Upscaled Gas and Water Relative Permeability from Pore and
Core Scale Experimental Data Over Hydraulic Fracturing, Flowback and Online

202
Production." Asia Pacific Unconventional Resources Technology Conference,
Brisbane, Australia, 18-19 November 2019. Unconventional Resources Technology
Conference, 2020.’
32. Li, Q., and R. Aguilera. "Unsupervised Statistical Learning with Integrated Pattern-
Based Geostatistical Simulation." SPE Western Regional Meeting. Society of
Petroleum Engineers, 2018.
33. REAL-TIME DECISION-MAKING FOR THE SUBSURFACE REPORT, Carnegie
Mellon University, Wilton E. Scott Institute for Energy Innovation, NETL, June 2019.
34. de Lima, Rafael Pires, and Kurt J. Marfurt. "Principal component analysis and K-means
analysis of airborne gamma-ray spectrometry surveys." SEG Technical Program
Expanded Abstracts 2018. Society of Exploration Geophysicists, 2018. 2277-2281.
35. Kundu, Debasis, Rameshwar D. Gupta, and Anubhav Manglick. "Discriminating
between the log-normal and generalized exponential distributions." Journal of
Statistical Planning and Inference 127.1-2 (2005): 213-227.
36. Cho, Hye-Kyung, Kenneth P. Bowman, and Gerald R. North. "A comparison of
gamma and lognormal distributions for characterizing satellite rain rates from the
tropical rainfall measuring mission." Journal of Applied meteorology 43.11 (2004):
1586-1597.
37. Navarette, Mike, Larry Chorn, and Marko Maucec. "A Holistic Approach to the
Development Stage of Shale Gas Resources." IPTC 2014: International Petroleum
Technology Conference. European Association of Geoscientists & Engineers, 2014.
38. Zhang, Liehui. "Shale gas reservoir characteristics and microscopic flow mechanisms."
Developments in Petroleum Science. Vol. 66. Elsevier, 2019. 1-47.
39. Paluszny, Adriana, Saeed Salimzadeh, and Robert W. Zimmerman. "Finite-element
modeling of the growth and interaction of hydraulic fractures in poroelastic rock
formations." Hydraulic Fracture Modeling. Gulf Professional Publishing, 2018. 1-19.
40. Ran, Qiquan. Unconventional Tight Reservoir Simulation: Theory, Technology and
Practice. Springer Nature, 2020.
41. Oliver, Dean S., and Yan Chen. "Recent progress on reservoir history matching: a
review." Computational Geosciences 15.1 (2011): 185-221.
42. Bader, Verena, and Stephan Kaiser. "Algorithmic decision-making? The user interface
and its role for human involvement in decisions supported by artificial intelligence."
Organization 26.5 (2019): 655-672.
43. Gupta, Ishank, et al. "Use of data analytics to optimize hydraulic fracture locations
along borehole." Petrophysics 59.06 (2018): 811-825.
44. Ma, Xiaodong, and Mark D. Zoback. "Lithology variations and cross-cutting faults
affect hydraulic fracturing of Woodford Shale: A case study." SPE Hydraulic
Fracturing Technology Conference and Exhibition. Society of Petroleum Engineers,
2017.
45. Schuetter, Jared, et al. "A data-analytics tutorial: Building predictive models for oil
production in an unconventional shale reservoir." SPE Journal 23.04 (2018): 1-075.
46. Bravo, C., Saputelli, L., Rivas, F., Pérez, A. G., Nikolaou, M., Zangl, G., de Guzmán,
N., Mohaghegh, S., Nunez, G. (2013). State of the art of artificial intelligence and

203
predictive analytics in the E&P industry: A technology survey. SPE Journal, 19(04),
547-563. https://doi.org/10.2118/150314-pa
47. Chaikine, I. A., Gates, I. D. (2021). A machine learning model for predicting multi-
stage horizontal well production. Journal of Petroleum Science and Engineering, 198,
108133. https://doi.org/10.1016/j.petrol.2020.108133
48. de Lima, R.P., Marfurt, K.J., 2018. Principal component analysis and K-means analysis
of airborne gamma-ray spectrometry surveys, SEG Technical Program Expanded
Abstracts, pp. 2277-2281. https://doi.org/10.1190/segam2018-2996506.1
49. Grushka-Cockayne, Y., Jose, V. R. R., Lichtendahl, K. C. (2017). Ensembles of overfit
and overconfident forecasts. Management Science, 63(4), 1110-1130.
https://doi.org/10.1287/mnsc.2015.2389
50. Gupta, S., Sharma, A., Abubakar, A., 2018. Artificial intelligence–driven asset
optimizer, SPE Annual Technical Conference and Exhibition, Dallas, Texas, USA,
September 24-26, 2018. https://doi.org/10.2118/191551-MS
51. Hajizadeh, Y. (2019). Machine learning in oil and gas; a SWOT analysis approach.
Journal of Petroleum Science and Engineering, 176, 661-663.
https://doi.org/10.1016/j.petrol.2019.01.113
52. Iferobia, C. C., Ahmad, M., Salim, A. M., Sambo, C., Michaels, I. H. (2020). Acoustic
data driven application of principal component multivariate regression analysis in the
development of unconfined compressive strength prediction models for shale gas
reservoirs, SPE Annual Technical Conference and Exhibition, Virtual, October 27 – 29,
2020. https://doi.org/10.2118/201287-MS
53. Klie, H., Yan, B., Klie, A. (2020). Transfer Learning for scalable optimization of
unconventional field operations, Proceedings of the 8th Unconventional Resources
Technology Conference, Virtual, July 20–22, 2020. https://doi.org/10.15530/urtec-
2020-2719
54. Konaté, A. A., Pan, H., Ma, H., Cao, X., Yevenyo Ziggah, Y., Oloo, M., Khan, N.
(2015). Application of dimensionality reduction technique to improve geophysical log
data classification performance in crystalline rocks. Journal of Petroleum Science and
Engineering, 133, 633-645. https://doi.org/10.1016/j.petrol.2015.06.035
55. Li, Z., Kang, Y., Feng, D., Wang, X.-M., Lv, W., Chang, J., Zheng, W. X. (2020).
Semi-supervised learning for lithology identification using Laplacian support vector
machine. Journal of Petroleum Science and Engineering, 195, 107510.
https://doi.org/10.1016/j.petrol.2020.107510
56. Liu, W., Liu, W. D., Gu, J. (2020). Predictive model for water absorption in sublayers
using a Joint Distribution Adaption based XGBoost Transfer Learning method. Journal
of Petroleum Science and Engineering, 188, 106937.
https://doi.org/10.1016/j.petrol.2020.106937
57. Lolon, E., Hamidieh, K., Weijers, L., Mayerhofer, M., Melcher, H., and Oduba, O.,
2016. Evaluating the relationship between well parameters and production using
multivariate statistical models: A middle bakken and three forks case history, SPE
Hydraulic Fracturing Technology Conference, The Woodlands, Texas, USA, February
9-11, 2016. https://doi.org/10.2118/179171-MS

204
58. Metwalli, S. A., 2020. How to choose the right machine learning algorithm for your
application. Towards Data Science. https://towardsdatascience.com/how-to-choose-the-
right-machine-learning-algorithm-for-your-application-1e36c32400b9 (accessed March
18, 2021).
59. Mohaghegh, S. D. (2017). Shale analytics, in: Mohaghegh, S. D. (Eds.), Shale
analytics, Springer, Cham, pp 29-81. https://doi.org/10.1007/978-3-319-48753-3
60. Montantes, J., 2020. 3 reasons to use Random Forest over a neural network–comparing
machine learning versus deep learning. Towards Data Science.
https://towardsdatascience.com/3-reasons-to-use-random-forest-over-a-neural-network-
comparing-machine-learning-versus-deep-f9d65a154d89 (accessed March 18, 2021).
61. Noonan, S., 2020. SPE strong: Strengthening our core: Strengthening knowledge
transfer in unconventional resources. Journal of Petroleum Technology, 72(03), 6-8.
https://doi.org/10.2118/0320-0006-JPT
62. Pan, S. J., Yang, Q. (2010). A survey on Transfer Learning. IEEE Transactions on
Knowledge and Data Engineering, 22(10), 1345-1359.
https://doi.org/10.1109/tkde.2009.191
63. Portis, D., Clarke, P., Barzola, G., Bello, H. (2016). Assessing well performance in a
prolific liquids-rich shale play –an Eagle Ford Shale case study, International
Conference and Exhibition, Barcelona, Spain, April 3-6, 2016.
https://doi.org/10.1190/ice2016-6497045.1
64. Probst, P., Wright, M. N., Boulesteix, A. L. (2019). Hyperparameters and tuning
strategies for random forest. WIREs Data Mining and Knowledge Discovery, 9(3),
e1301. https://doi.org/10.1002/widm.1301
65. Ruder, S., 2019. Neural Transfer Learning for natural language processing (Doctoral
dissertation). National University of Ireland, Galway.
66. Rui, Z., Cui, K., Wang, X., Lu, J., Chen, G., Ling, K., Patil, S. (2018). A quantitative
framework for evaluating unconventional well development. Journal of Petroleum
Science and Engineering, 166, 900-905. https://doi.org/10.1016/j.petrol.2018.03.090
67. Weiss, K., Khoshgoftaar, T. M., Wang, D. (2016). A survey of Transfer Learning.
Journal of Big Data, 3(1). https://doi.org/10.1186/s40537-016-0043-6
68. Xue, L., Liu, Y., Xiong, Y., Liu, Y., Cui, X., Lei, G. (2021). A data-driven shale gas
production forecasting method based on the multi-objective Random Forest regression.
Journal of Petroleum Science and Engineering, 196, 107801.
https://doi.org/10.1016/j.petrol.2020.107801
69. Yang, L., Ping, C., Shou, M. T. (2017). Application of well factory drilling technology
in shale gas development in the Sichuan basin, SPE/IATMI Asia Pacific Oil & Gas
Conference and Exhibition, Jakarta, Indonesia, October 17-19, 2017.
https://doi.org/10.2118/186914-MS
70. Yogatama, D., Mann, G. (2014). Efficient Transfer Learning method for automatic
hyperparameter tuning, Proceedings of the 17th International Conference on Artificial
Intelligence and Statistics, Reykjavik, Iceland, April 22 - 25, 2014. Journal of Machine
Learning Research, 33, 1077-1085.
71. Jansen, Jan-Dirk, Roald Brouwer, and Sippe G. Douma. "Closed loop reservoir
management." SPE reservoir simulation symposium. OnePetro, 2009.

205
72. Tork, Hadi Fanaee. "Spatio-temporal clustering methods classification." Doctoral
symposium on informatics engineering. Vol. 1. No. 1. Porto, Portugal: Faculdade de
Engenharia da Universidade do Porto, 2012.
73. Nguyen, Hai-Long, Yew-Kwong Woon, and Wee-Keong Ng. "A survey on data stream
clustering and classification." Knowledge and information systems 45.3 (2015): 535-
569.
74. Nikravesh, Masoud, Lotfi Asker Zadeh, and Fred Aminzadeh, eds. Soft computing and
intelligent data analysis in oil exploration. Elsevier, 2003.
75. Popa, Andrei, and William Wood. "Application of case-based reasoning for well
fracturing planning and execution." Journal of Natural Gas Science and Engineering 3.6
(2011): 687-696.
76. Gentleman, Robert, and Vincent J. Carey. "Unsupervised machine learning."
Bioconductor case studies. Springer, New York, NY, 2008. 137-157.
77. Korjani, Mohammad M., and Jerry M. Mendel. "Non-linear Variable Structure
Regression (VSR) and its application in time-series forecasting." 2014 IEEE
International Conference on Fuzzy Systems (FUZZ-IEEE). IEEE, 2014.
78. Candanedo, Inés Sittón, et al. "Machine learning predictive model for industry 4.0."
International Conference on Knowledge Management in Organizations. Springer,
Cham, 2018.
79. Hanga, Khadijah M., and Yevgeniya Kovalchuk. "Machine learning and multi-agent
systems in oil and gas industry applications: A survey." Computer Science Review 34
(2019): 100191.
80. Koh, Pang Wei, and Percy Liang. "Understanding black-box predictions via influence
functions." arXiv preprint arXiv:1703.04730 (2017).
81. Rudin, Cynthia. "Stop explaining black-box machine learning models for high stakes
decisions and use interpretable models instead." Nature Machine Intelligence 1.5
(2019): 206-215.
82. Mendel, Jerry, and Mehdi Korjani. “On establishing nonlinear combinations of
variables from small to big data for use in later processing.” Information Sciences,
2014.
83. Mendel, Jerry, and Mehdi Korjani. “Theoretical Aspects of Fuzzy Set Qualitative
Comparative Analysis (fsQCA).” Information Sciences, 2013.
84. Korjani, Mehdi M., Jerry M. Mendel, and Iraj Ershaghi. "A Predictive Model for
Improving the Efficiency of Frac Jobs." SPE Western Regional Meeting. Society of
Petroleum Engineers, 2015.
85. Navrátil, Jiří, et al. "Accelerating physics-based simulations using end-to-end neural
network proxies: An application in oil reservoir modeling." Frontiers in Big Data 2
(2019): 33.
86. Bui, Dieu Tien, et al. "Comparing the prediction performance of a Deep Learning
Neural Network model with conventional machine learning models in landslide
susceptibility assessment." Catena 188 (2020): 104426.
87. Yan, L. C., B. Yoshua, and H. Geoffrey. "Deep learning." nature 521.7553 (2015): 436-
444.

206
88. Li, Yuanjun, Ruixiao Sun, and Roland Horne. "Deep learning for well data history
analysis." SPE Annual Technical Conference and Exhibition. Society of Petroleum
Engineers, 2019.
89. Abraham, Ajith. "Artificial neural networks." Handbook of measuring system design
(2005).
90. Guo, Shuai. “An introduction to Surrogate modeling, Part I: fundamentals” Towards
Data Science, 2020.
91. Ashayeri, Cyrus, and Donald L. Paul. "A Stochastic Method in Investigating Basin-
Wide Underlying Distribution Functions of Decline Rate Behavior for Unconventional
Resources." SPE Western Regional Meeting. OnePetro, 2021.
92. Maula, Azman, et al. "Artificial Neural Network Surrogate Reservoir Proxy Modeling
for Generating Pressure and Gas Saturation at the Grid–Block Level in Shale Gas
Reservoirs." Offshore Technology Conference Asia. Offshore Technology Conference,
2020.
93. Moridis, George J., Thomas Alwin Blasingame, and Craig Matt Freeman. "Analysis of
mechanisms of flow in fractured tight-gas and shale-gas reservoirs." SPE Latin
American and Caribbean Petroleum Engineering Conference. Society of Petroleum
Engineers, 2010.
94. Ilk, Dilhan, et al. "Production data analysis--Challenges, pitfalls, diagnostics." SPE
Reservoir Evaluation & Engineering 13.03 (2010): 538-552.
95. Temizel, Cenk, et al. "Production Forecasting in Shale Reservoirs through
Conventional DCA and Machine/Deep Learning Methods." Unconventional Resources
Technology Conference, 20 –22 July 2020. Unconventional Resources Technology
Conference (URTeC), 2020.
96. Song, Xuanyi, et al. "Time-series well performance prediction based on Long Short-
Term Memory (LSTM) neural network model." Journal of Petroleum Science and
Engineering 186 (2020): 106682
97. Pearl, Judea. Causality. Cambridge university press, 2009.
98. Ashayeri, Cyrus, and Birendra Jha. “Assessment Of Unconventional Resources
Opportunities In The Middle East Tethyan Petroleum System In A Transfer Learning
Context” ADIPEC. SPE-207723-MS OnePetro, 2021.
99. Ashayeri, Cyrus, Mehdi Korjani, and Iraj Ershaghi. “Using the Adaptive Variable
Structure Regression Approach in Data Selection and Data Preparation for Improving
Machine Learning Based Performance Prediction in Unconventional Plays” URTeC,
July 2021.
100. Han, Dongkwon, Jihun Jung, and Sunil Kwon. "Comparative study on supervised
learning models for productivity forecasting of shale reservoirs based on a data-driven
approach." Applied Sciences 10.4 (2020): 1267.
101. Shojaei, Hasan, and Casey Lipp. "Optimizing Unconventional Field Development
through an Integrated Reservoir Characterization and Simulation Approach."
Unconventional Resources Technology Conference, San Antonio, Texas, 1-3 August
2016. Society of Exploration Geophysicists, American Association of Petroleum
Geologists, Society of Petroleum Engineers, 2016.

207
Appendix
In the appendix section we have provided a few screen captures from the current user interface
created for the workflow developed for this thesis.
1) General view of the first page of the browser-based dashboard. Left-hand menu items allow
the selection of various control parameters. Users can navigate to other functions of the tool
from the tabs above the page.

2) An example of spatial visualization capabilities of the workflow. Geographical
distribution of geologic variables as well as reduction volumes is possible in both
original values and log transformation scales:

208

3) Unsupervised clustering was used in many steps of the research in our work. The
proposed workflow allows conducting k-means clustering by different combinations of
variables and different cluster numbers. The model automatically performs PCA
dimension reduction in cases with more than two parameters for clustering:

209

4) Stochastic analysis of basin-wide decline behavior is embedded in the workflow. Users
can select various production related features such as peak rate and rate of change for
visualization purposes. In addition, estimation, and antifiction of the underlying
distribution functions for large groups of wells is automated in the workflow. The
distribution function types, their coefficients, as well as the errors associated with the
estimation of multi-year production can be evaluated in this tool (see next page).

210

211
5) Evaluation of the change in long-term forecast for exponential power law decline is
developed in the workflow where the user can observe the incremental changes in
forecast error by using different number of months in production history:

6) Random Forest predictive modeling functions where the user can select training
features and train to test split ratio. Filtering features allow narrowing the train sample
to wells with certain productivity levels. Errors for train and test subsets are reported in
tables and histograms (see next page).

212

7) To facilitate the feature selection process, the variables’ importance ranking is reported
by both Random Forest and Bayesian GLM. In addition, Silhouette measure is
calculated to determine the optimal number of clusters for k-means clustering (see next
page).

213

214
8) Train and test on all possible combinations of binary nonrepeating instances are
streamlined and the reports for cluster summary and errors are accessible:

215

9) Selection of the source field and the target field for Transfer Learning purposes is
possible by both the geographical positioning of the sub-basins or by selecting certain
clusters from previous clustering steps. The user can select validation groups as well as
combinations of train sub-samples to assess the impact of the addition of small
increments of wells to the train dataset from the target field (see next page).

216

Abstract (if available)

Abstract The development of oil and gas resources is a capital-intensive and challenging task. A reservoir engineer faces several problems such as “well placement,” “well control,” and “development economics.” All of these problems have been examined using reservoir simulation. In the conventional reservoir simulation approach (forward problem), the main assumption is that the model has been developed by using established fluid flow principles, where all the complexities of the reservoir influencing the flow are known. In this case, if the observation of the production behavior does not match the simulation forecast results, the reservoir engineering team can change the static parameters related to the geological and petrophysical aspect of the reservoir and improve the history matching process. A simulation model can further help with tasks such as optimization, uncertainty quantification, and sensitivity analysis. However, this process is not effective in modeling the production of unconventional oil and gas reservoirs due to several fundamental differences in these types of assets, such as yet unknown physics that govern the unconventional resources production. ? In the early years during the development of unconventional resources, companies relied on the learning curve from drilling multiple wells and using past experiences from conventional assets. This approach is expensive and does not yield any systematic way of improvement. In the past decade, the application of data analytics from inverse modeling and statistical learning has helped to identify some trends in producing wells from unconventional resources. However, utilizing various statistical or data-driven analyses without the guidance of domain expertise may result in inadequacies in the functionality of such models. Therefore, it is essential that a methodology is developed to incorporate reservoir engineering domain expertise with data-driven workflows to reduce the gap between using primarily data science expertise. ? In this research, we have explored all stages of data-driven modeling in the context of basin-wide analysis of unconventional resources. We propose a full cycle workflow in integrating data-driven techniques with reservoir engineering knowledge guidance. The first phase of this process includes data collection, data preprocessing, and the more important stage of unsupervised exploratory analysis of the data or data mining. We have explained the ways that petroleum engineering knowledge can guide the stage of unsupervised clustering to create balanced data considering the nature of various parameters. We have explored some of the recent advancements in machine learning algorithms that utilize fuzzy systems combined with non-linear regression to extract rules from the data. We demonstrate that using such models can help with the data preparation and data selection stage in data-driven modeling. ? The second stage in a data-driven approach is developing data-driven predictive algorithms. Again, in this phase, we have combined the required petroleum and geology knowledge in all steps of training, testing, and validation of the models. Taking into consideration the spatial and temporal dimensions of the data in a heterogeneous geologic formation is the core of this stage. We have shown how large volumes of production data can be used to develop stochastic models of decline behavior in a substantial number of wells. By quantifying the underlying distribution functions of production decline parameters, engineering teams can improve long-term field development planning with more accurate risk analysis. ? Third, in the post-modeling stage, we use the developed workflows and algorithms to answer some of the main questions regarding the development of a new unconventional asset. Typical problems include ranking unconventional assets based on the reservoir quality, selecting drilling locations, and identifying the geologic and engineering parameters affecting drilling and fracturing design that influence well productivity in a green field. The proposed Transfer Learning workflow in this research highlights the necessity of integrating domain expertise with using data science in unconventional predictive modeling. We have quantified the improvements in the predictive performance of machine learning models by using basic petroleum and geology knowledge in feature selection for engineering and model modes under a Transfer Learning context. ? We also discuss the pitfalls of disregarding the reservoir engineering knowledge in using data science for the predictive modeling of unconventional assets. Through this thesis, we have provided examples from the industry where incorrect use of data-driven techniques has resulted in models that are either inaccurate or only accurate in a limited dataset. Our hope is that the findings of this research, including the extensive dataset, created, streamlined workflows, and proposed guidelines, can help future students, researchers, and engineers in more efficient use of data science in the area of energy resource exploitation. ? In this research, for the first time in the literature, we have expanded the scope of data-driven modeling to an entire basin. This task introduced a major challenge in the data acquisition cleaning and pre-processing stage. Compilation of the dataset used in this research made up a significant portion of the time and energy allocated to this work. The substantial volume of this dataset also forced us to streamline and automate many of the data analytics tasks associated with our proposed workflow. The efforts associated with developing the programming of more than two thousand lines of code required for our analysis also was one of the most challenging stages of this work. However, the produced workflow, which is packaged as an interactive user interface, became an extremely helpful component of this research. This easy-to-use workflow allowed generating multiple test scenarios rapidly and facilitate the hypotheses validation/rejection process for large statistical samples possible. We consider these accomplishments as one of our most important contributions to the research community, and we hope that future students and engineers can build upon our efforts. ? In the final chapter of this work, we provide a series of conclusions that aim at helping future researchers and students in better implementation of data-driven modeling in this domain.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Integrated reservoir characterization for unconventional reservoirs using seismic, microseismic and well log data

$Integrated workflow for characterizing and modeling fracture network in unconventional reservoirs using microseismic data$

PDF

Integrated workflow for characterizing and modeling fracture network in unconventional reservoirs using microseismic data

PDF

Deep learning architectures for characterization and forecasting of fluid flow in subsurface systems

PDF

Modeling and simulation of multicomponent mass transfer in tight dual-porosity systems (unconventional)

PDF

A coupled multiphase flow-geomechanics framework for modeling material, kinematic and contact nonlinearities in faulted reservoirs

PDF

Deep learning for subsurface characterization and forecasting

PDF

Modeling and simulation of complex recovery processes

PDF

Reactivation of multiple faults in oilfields with injection and production

PDF

Integration of multi-physics data into dynamic reservoir model with ensemble Kalman filter

PDF

Dynamics of CO₂ leakage through faults

PDF

Stochastic oilfield optimization under uncertainty in future development plans

PDF

Probabilistic data-driven predictive models for energy applications

PDF

Geothermal resource assessment and reservoir modeling with an application to Kavaklidere geothermal field, Turkey

PDF

Real-time reservoir characterization and optimization during immiscible displacement processes

PDF

Modeling and recognition of events from temporal sensor data for energy applications

PDF

Data-driven methods for increasing real-time observability in smart distribution grids

PDF

Waterflood induced formation particle transport and evolution of thief zones in unconsolidated geologic layers

PDF

A method for characterizing reservoir heterogeneity using continuous measurement of tracer data

PDF

A study of diffusive mass transfer in tight dual-porosity systems (unconventional)

PDF

Assessing induced seismicity rate increase based on deterministic and probabilistic modeling

Asset Metadata

Creator Ashayeri, Cyrus (author)

Core Title Application of data-driven modeling in basin-wide analysis of unconventional resources, including domain expertise

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Petroleum Engineering

Degree Conferral Date 2021-08

Publication Date 07/26/2021

Defense Date 06/10/2021

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag artificial intelligence,data science,domain expert,Energy,hydrocarbon,machine learning,natural gas,OAI-PMH Harvest,oil,Petroleum engineering,resources,shale,unconventional

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Ershaghi, Iraj (committee chair), Jha, Birendra (committee member), Meshkati, Najmedin (committee member), Paul, Donald L. (committee member)

Creator Email ashayeri@usc.edu,cyrus.ashayeri@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC15623323

Unique identifier UC15623323

Legacy Identifier etd-AshayeriCy-9895

Document Type Dissertation

Format application/pdf (imt)

Rights Ashayeri, Cyrus

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu