APPLICATION OF MACHINE LEARNING IN WELL PERFORMANCE

PREDICTION, DESIGN OPTIMIZATION AND HISTORY MATCHING

A Dissertation

by

ADITYA VYAS

Submitted to the Office of Graduate and Professional Studies of

Texas A&M University
in partial fulfillment of the requirements for the degree of

DOCTOR OF PHILOSOPHY

Chair of Committee, Akhil Datta-Gupta

Committee Members, Michael J. King
Bani K. Mallick
Duane A. McVay
Head of Department, A. Daniel Hill

August 2017

Major Subject: Petroleum Engineering

Copyright 2017 Aditya Vyas

 ABSTRACT

Finite difference based reservoir simulation is commonly used to predict well rates

in these reservoirs. Such detailed simulation requires an accurate knowledge of reservoir

geology. Also, these reservoir simulations may be very costly in terms of computational

time. Recently, some studies have used the concept of machine learning to predict mean

or maximum production rates for new wells by utilizing available well production and

completion data in a given field. However, these studies cannot predict well rates as a

function of time. This dissertation tries to fill this gap by successfully applying various

machine learning algorithms to predict well decline rates as a function of time. This is

achieved by utilizing available multiple well data (well production, completion and

location data) to build machine learning models for making rate decline predictions for

the new wells. It is concluded from this study that well completion and location variables

can be successfully correlated to decline curve model parameters and Estimated Ultimate

Recovery (EUR) with a reasonable accuracy. Among the various machine learning models

studied, the Support Vector Machine (SVM) algorithm in conjunction with the Stretched

Exponential Decline Model (SEDM) was concluded to be the best predictor for well rate

decline. This machine learning method is very fast compared to reservoir simulation and

does not require a detailed reservoir information. Also, this method can be used to fast

predict rate declines for more than one well at the same time.

This dissertation also investigates the problem of hydraulic fracture design

optimization in unconventional reservoirs. Previous studies have concentrated mainly on

ii
optimizing hydraulic fractures in a given permeability field which may not be accurately

known. Also, these studies do not take into account the trade-off between the revenue

generated from a given fracture design and the cost involved in having that design. This

dissertation study fills these gaps by utilizing a Genetic Algorithm (GA) based workflow

which can find the most suitable fracturing design (fracture locations, half-lengths and

widths) for a given unconventional reservoir by maximizing the Net Present Value (NPV).

It is concluded that this method can optimize hydraulic fracture placement in the presence

of natural fracture/permeability uncertainty. It is also concluded that this method results

in a much higher NPV compared to an equally spaced hydraulic fractures with uniform

fracture dimensions.

Another problem under investigation in this dissertation is that of field scale

history matching in unconventional shale oil reservoirs. Stochastic optimization methods

are commonly used in history matching problems requiring a large number of forward

simulations due to the presence of a number of uncertain variables with unrefined variable

ranges. Previous studies commonly used a single stage history matching. This study

presents a method utilizing multiple stages of GA. Most significant variables are separated

out from the rest of the variables in the first GA stage. Next, best models with refined

variable ranges are utilized with previously eliminated variables to conduct GA for next

stage. This method results in faster convergence of the problem.

iii
 DEDICATION

I dedicate this dissertation to my parents, my wife, my brother and my friends for

their support during my studies at Texas A&M University.

iv
 ACKNOWLEDGEMENTS

I would like to express my sincere gratitude to my advisor, Dr. Akhil Datta-Gupta for

his continued guidance during entire period of my PhD study. His expanse of knowledge

and readiness to listen to my problems made it possible for me to study in this department

of petroleum engineering without any bottlenecks. I would also like to thank him for

continued financial support during my entire PhD studies.

I would like to thank Dr. Michael King and Dr. Bani K. Mallick for their continued

interest in my research studies. Their immense knowledge and invaluable comments

during my presentations guided me to the right direction and also helped me to continue

my PhD without any bottlenecks. I would like to thank Dr. Srikanta Mishra from Battelle

for his invaluable suggestions regarding Machine Learning study included in this

dissertation. His immense knowledge and guidance always helped me when I needed

them. I would also like to thank Dr. Duane A. McVay for being a member in my

committee.

I would also like to thank Phaedra Hopcus, Barbi Miller and Eleanor Schuler for their

help during various occasions particularly with the paperwork involved during this

graduate program.

I would also like to thank my colleagues in my research group at the department of

Petroleum Engineering, Texas A&M University – Jixiang Huang, Kenta Nakajuma,

Hyunmin Kim, Hye Young Jung, Changdong Yang, Atsushi Iino, Tsubasa Onishi,

v
Hongquan Chen, Feyisayo Olalotiti-Lawal, Xue Xu, Rongqiang Chen and Gill Hetz - for

their invaluable suggestions.

I would also like to alumni of this research group – Xia Xiaoyang, Yanbin Zhang,

Peerapong Ekkawong, Jichao Han, Kam Dongjae, Muhammed Al-Rukabi, Jeongmin

Kim, Neha Bansal, Shingo Watanabe, Shusei Tanaka and Zheng Zhang - for their

invaluable suggestions.

Finally, I would like to thank my professors in University of Oklahoma (where I did

my Masters studies) who recommended me to this PhD program – Dr. Deepak

Devegowda, Dr. Ramadan Ahmed and Dr. Jeffrey G. Callard.

vi
 CONTRIBUTORS AND FUNDING SOURCES

Contributors

This PhD dissertation work was supervised by Dr. Akhil Datta-Gupta (Committee

Chair) and other three committee members - Dr. Michael J. King, Dr. Bani K. Mallick and

Dr. Duane A. McVay.

Chapter II of this dissertation study involving Machine Learning based study

includes various suggestions made by Dr. Srikanta Mishra from Battelle. This work has

been accepted for presentation in one of the SPE conferences before the end of year 2017.

Chapter III of this dissertation involving hydraulic fracture optimization study was

done in collaboration with Changdong Yang. Changdong Yang provided the upscaling

code (Oda Method) and Fast Marching Method (FMM) based forward simulator for this

study. This work has been published in Journal of Petroleum Science and Engineering

(2017) with a modified workflow.

Chapter IV of this dissertation involving History Matching in Shale Oil reservoirs

has been done in collaboration with Atsushi Iino. Atsushi Iino provided the Fast Marching

Method (FMM) based reservoir simulator used for this study. This work has been

presented in SPE conference (SPE 185719-MS) with a modified workflow and would also

be presented in an upcoming URTeC conference (URTeC: 2693139) with a modified

workflow.

All remaining work in this dissertation has been done independently by Aditya

Vyas.

vii
Funding Sources

This work was made possible by the financial support of the member companies

of Model Calibration and Efficient Reservoir Imaging (MCERI) consortium.

viii
 TABLE OF CONTENTS

Page

ABSTRACT .......................................................................................................................ii

DEDICATION ..................................................................................................................iv

ACKNOWLEDGEMENTS ............................................................................................... v

CONTRIBUTORS AND FUNDING SOURCES ............................................................vii

TABLE OF CONTENTS ..................................................................................................ix

LIST OF FIGURES ........................................................................................................ xiii

LIST OF TABLES ........................................................................................................xxxi

CHAPTER I INTRODUCTION AND OBJECTIVES ..................................................... 1

1.1 Introduction .............................................................................................................. 1

1.2 Dissertation Outline.................................................................................................. 3

CHAPTER II MACHINE LEARNING BASED INSIGHTS ON WELL

PERFORMANCE IN EAGLE FORD WELLS ......................................... 5

2.1 Introduction and Literature Review ......................................................................... 5

2.2 Methodology .......................................................................................................... 11

2.2.1 Rate Decline Models ....................................................................................... 11

2.2.1.1 Arp’s Decline Model ................................................................................ 11

ix
 2.2.1.2 Stretched Exponential Decline Model (SEDM) ....................................... 13

2.2.1.3 Duong Model............................................................................................ 15

2.2.1.4 Weibull Model.......................................................................................... 15

2.2.2 Machine Learning Algorithms ........................................................................ 17

2.2.2.1 Random Forests (RF) ............................................................................... 18

2.2.2.2 Gradient Boosted Machine (GBM) Regression ....................................... 23

2.2.2.3 Support Vector Machines (SVM) Regression or Support Vector

Regression (SVR) ..................................................................................... 25

2.2.2.4 Multivariate Adaptive Regression Splines (MARS) ................................ 26

2.2.3 Model Averaging............................................................................................. 29

2.2.3.1 Generalized Likelihood Uncertainty Estimation (GLUE)........................ 31

2.2.4 Relative Influence of Predictor Variables ....................................................... 33

2.3 Eagle Ford Field Case Study .................................................................................. 35

2.4 Summary ................................................................................................................ 68

CHAPTER III HYDRAULIC FRACTURE DESIGN AND OPTIMIZATION IN

UNCONVENTIONAL SINGLE PHASE GAS RESERVOIR

USING GENETIC ALGORITHM ........................................................... 69

3.1 Introduction and Literature Review ....................................................................... 69

3.2 Methodology .......................................................................................................... 78

x
 3.2.1 Fast Marching Method .................................................................................... 78

3.2.2 DFN Upscaling (Oda’s Method) ..................................................................... 82

3.2.3 Hydraulic Fracturing Design ........................................................................... 85

3.2.4 Genetic Algorithm and Workflow................................................................... 87

3.3 Results and Discussion ........................................................................................... 91

3.4 Summary .............................................................................................................. 107

CHAPTER IV A MULTISTAGE GENETIC ALGORITHM FOR HISTORY

MATCHING OF SHALE OIL RESERVOIRS: FIELD CASE

STUDY................................................................................................... 108

4.1 Background and Introduction ............................................................................... 108

4.2 Methodology ........................................................................................................ 109

4.3 Results and Discussion ......................................................................................... 113

4.3.1 History matching results based on GA and three phase FMM...................... 116

4.3.2 History matching results based on GA and compositional FMM ................. 160

4.4 Summary............................................................................................................... 192

CHAPTER V CONCLUSIONS AND RECOMMENDATIONS ................................. 193

5.1 Summary and Conclusions ................................................................................... 193

5.2 Recommendations ................................................................................................ 194

NOMENCLATURE ....................................................................................................... 195

xi
SUBSCRIPTS ................................................................................................................ 198

REFERENCES ............................................................................................................... 199

APPENDIX A ................................................................................................................ 212

xii
 LIST OF FIGURES

Page

Figure 2.1 An example well prediction made by Arp’s decline model ............................ 13

Figure 2.2 Comparison of Arp’s and SEDM decline models .......................................... 14

Figure 2.3 (a) Classification Tree example (b) Equivalent partition for a two

variable case .................................................................................................... 20

Figure 2.4 An example Regression Tree from Eagle Ford data predicting maximum

oil production................................................................................................... 20

Figure 2.5 Cost complexity and size of a regression tree against misfit error using

Eagle Ford data ................................................................................................ 22

Figure 2.6 Approximate representation of a Gradient Boosted Tree Model .................... 24

Figure 2.7 An example of GCV plot using Eagle Ford data ............................................ 29

Figure 2.8 Workflow steps for model training and prediction ......................................... 31

Figure 2.9 Pairwise scatterplots of various predictor variables in Eagle Ford data ......... 37

Figure 2.10 Regression Tree fitted on EUR calculated from Arp’s Decline Model ........ 38

Figure 2.11 Regression Tree fitted on EUR calculated from SEDM Decline Model ...... 38

Figure 2.12 Regression Tree fitted on EUR calculated from Duong’s Decline

Model ............................................................................................................ 39

xiii
Figure 2.13 Regression Tree fitted on EUR calculated from Weibull’s Decline

Model ............................................................................................................ 39

Figure 2.14 Classification Tree fitted on EUR clusters derived from Arp’s Decline

Model ............................................................................................................ 40

Figure 2.15 Classification Tree fitted on EUR clusters derived from SEDM Decline

Model ............................................................................................................ 40

Figure 2.16 Classification Tree fitted on EUR clusters derived from Duong’s

Decline Model ............................................................................................... 41

Figure 2.17 Classification Tree fitted on EUR clusters derived from Weibull’s

Decline Model ............................................................................................... 41

Figure 2.18 Well clusters based on Initial Flow Rate, qi.................................................. 42

Figure 2.19 Predictor variable distribution in clusters derived from Initial Flow

Rate, qi .......................................................................................................... 43

Figure 2.20 Study wells on Texas map color coded by cluster number ........................... 44

Figure 2.21 Correlation between cluster type and different variables ............................. 45

Figure 2.22 Error metric comparison for different machine learning algorithms

taken into consideration for Arp’s model...................................................... 47

Figure 2.23 Scatterplots showing predicted vs actual values of Arp’s decline model

parameters and EUR ..................................................................................... 48

Figure 2.24 Prediction of Arp’s decline curves using GBM ............................................ 49

xiv
Figure 2.25 Error metric comparison for different machine learning algorithms taken

into consideration for SEDM model ............................................................. 50

Figure 2.26 Scatterplots showing predicted vs actual values of SEDM decline model

parameters and EUR ..................................................................................... 50

Figure 2.27 Prediction of SEDM decline curves using SVM .......................................... 51

Figure 2.28 Error metric comparison for different machine learning algorithms

taken into consideration for Duong’s model ................................................. 52

Figure 2.29 Scatterplots showing predicted vs actual values of Duong’s decline

model parameters and EUR .......................................................................... 52

Figure 2.30 Prediction of Duong’s decline curves using GBM ....................................... 53

Figure 2.31 Error metric comparison for different machine learning algorithms

taken into consideration for Weibull model .................................................. 54

Figure 2.32 Scatterplots showing predicted vs actual values of Weibull’s decline

model parameters and EUR .......................................................................... 55

Figure 2.33 Prediction of Weibull’s decline curves using SVM ...................................... 56

Figure 2.34 Comparison of predictions made by ARP’S - GBM, SEDM - SVM,

DUONG – GBM and WEIBULL - SVM ..................................................... 57

Figure 2.35 EUR prediction comparison among best candidates for each decline

model ............................................................................................................. 58

Figure 2.36 RMSE based variable ranking distribution ................................................... 60

xv
Figure 2.37 RMSE based variable ranking frequency distribution .................................. 61

Figure 2.38 RMSE based variable average rank vs rank variance ................................... 61

Figure 2.39 AAE based Variable Ranking distribution ................................................... 62

Figure 2.40 AAE based variable ranking frequency distribution ..................................... 63

Figure 2.41 AAE based variable average rank vs rank variance...................................... 63

Figure 2.42 R2 based variable ranking distribution .......................................................... 64

Figure 2.43 R2 based variable ranking frequency distribution ......................................... 65

Figure 2.44 R2 based variable average rank vs rank variance .......................................... 65

Figure 2.45 Median-Sigma ratio based variable ranking distribution .............................. 66

Figure 2.46 Median-Sigma ratio based variable ranking frequency distribution ............. 67

Figure 2.47 Median-Sigma ratio based variable average rank vs rank variance .............. 67

Figure 3.1 Natural Fracture distribution in the base model (Yang et al., 2017)............... 83

Figure 3.2 General workflow for genetic algorithm (Yang et al., 2017) ......................... 89

Figure 3.3 Workflow of objective function evaluation for each model

(Yang et al., 2017) ........................................................................................... 91

Figure 3.4 (a) Natural fracture distribution (b) Upscaled reservoir permeability field

(Yang et al., 2017) ......................................................................................... 92

Figure 3.5 FMM versus Eclipse simulated gas production for the base model

(Yang et al., 2017) ........................................................................................... 93

xvi
Figure 3.6 Effect of changing minimum matrix permeability during Oda’s

upscaling .......................................................................................................... 94

Figure 3.7 a) Gas Rates for various number of fracture stages b) Cumulative Gas

Production for different numbers of fracture stages ........................................ 96

Figure 3.8 Cost and NPV comparison for various cases of number of fracture

stages ............................................................................................................... 96

Figure 3.9 Sensitivity analysis of various variables on NPV ........................................... 98

Figure 3.10 NPV distribution in Genetic Algorithm based optimization approach ......... 99

Figure 3.11 Distribution of fracture stages and average widths in generation 1 and

generation 25 ................................................................................................. 99

Figure 3.12 Distribution of fracture stages in generation 1 and generation 25 .............. 100

Figure 3.13 NPV from Uniform spaced fractures .......................................................... 101

Figure 3.14 Hydraulic fracture placement in optimal design using genetic

algorithm ..................................................................................................... 102

Figure 3.15 Six possible realizations vs true model/base model in case of

uncertainty in natural fracture distribution .................................................. 104

Figure 3.16 Results of genetic algorithm for multiple realization based

optimization ................................................................................................ 105

Figure 3.17 Variable distribution in the first generation vs last generation ................... 105

xvii
Figure 3.18 Hydraulic fracture placement in optimal design based on multiple

realizations .................................................................................................. 106

Figure 4.1 General workflow for genetic algorithm (GA) ............................................. 113

Figure 4.2 Three regions in the field case reservoir model ............................................ 114

Figure 4.3 Well constraint Tubing Head Pressure during well production period ........ 116

Figure 4.4 Cumulative Oil Production of FMM and Eclipse as compared to History

data with base case variables (three phase FMM) ...................................... 117

Figure 4.5 Oil Rate Production of FMM and Eclipse as compared to History data

with base case variables (three phase FMM) .............................................. 117

Figure 4.6 Cumulative Water Production of FMM and Eclipse as compared to

History data with base case variables (three phase FMM) ......................... 118

Figure 4.7 Water Rate Production of FMM and Eclipse as compared to History

data with base case variables (three phase FMM) ...................................... 118

Figure 4.8 Cumulative Gas Production of FMM and Eclipse as compared to

History data with base case variables (three phase FMM) ......................... 119

Figure 4.9 Gas Rate Production of FMM and Eclipse as compared to History data

with base case variables (three phase FMM) .............................................. 119

Figure 4.10 Sensitivity analysis at the beginning of Stage 1 (three phase FMM) ......... 120

Figure 4.11 GA results for Stage 1 (three phase FMM) ................................................. 121

xviii
Figure 4.12 Uncertainty reduction in hydraulic fracture permeability during GA -

Stage 1 (three phase FMM) ......................................................................... 122

Figure 4.13 Uncertainty reduction in hydraulic fracture initial water saturation

during GA - Stage 1 (three phase FMM) .................................................... 122

Figure 4.14 Uncertainty reduction in hydraulic fracture shape factor during GA -

Stage 1 (three phase FMM) ......................................................................... 123

Figure 4.15 Uncertainty reduction in SRV porosity during GA - Stage 1

(three phase FMM) ...................................................................................... 123

Figure 4.16 Uncertainty reduction in SRV permeability during GA - Stage 1

(three phase FMM) ...................................................................................... 124

Figure 4.17 Uncertainty reduction in SRV initial water saturation during GA -

Stage 1 (three phase FMM) ......................................................................... 124

Figure 4.18 Uncertainty reduction in SRV shape factor during GA - Stage 1

(three phase FMM) ...................................................................................... 125

Figure 4.19 Variable distribution of hydraulic fracture permeability in the first

generation of GA - Stage 1 (three phase FMM) ......................................... 125

Figure 4.20 Variable distribution of hydraulic fracture initial water saturation in

the first generation of GA - Stage 1 (three phase FMM) ............................ 126

Figure 4.21 Variable distribution of hydraulic fracture shape factor in the first

generation of GA - Stage 1 (three phase FMM) ......................................... 126

xix
Figure 4.22 Variable distribution of SRV porosity in the first generation of GA -

Stage 1 (three phase FMM) ......................................................................... 127

Figure 4.23 Variable distribution of SRV permeability in the first generation of

GA - Stage 1 (three phase FMM) ................................................................ 127

Figure 4.24 Variable distribution of SRV initial water saturation in the first

generation of GA - Stage 1 (three phase FMM) ......................................... 128

Figure 4.25 Variable distribution of SRV shape factor in the first generation of

GA - Stage 1 (three phase FMM) ................................................................ 128

Figure 4.26 Variable distribution of hydraulic fracture permeability in the best

selected models of GA - Stage 1 (three phase FMM) ................................. 129

Figure 4.27 Variable distribution of hydraulic fracture initial water saturation in

the best selected models of GA - Stage 1 (three phase FMM).................... 129

Figure 4.28 Variable distribution of hydraulic fracture shape factor in the best

selected models of GA - Stage 1 (three phase FMM) ................................. 130

Figure 4.29 Variable distribution of SRV porosity in the best selected models of

GA - Stage 1 (three phase FMM) ................................................................ 130

Figure 4.30 Variable distribution of SRV permeability in the best selected models

of GA - Stage 1 (three phase FMM) ........................................................... 131

Figure 4.31 Variable distribution of SRV initial water saturation in the best

selected models of GA - Stage 1 (three phase FMM) ................................. 131

xx
Figure 4.32 Variable distribution of SRV shape factor in the best selected models

of GA - Stage 1 (three phase FMM) ........................................................... 132

Figure 4.33 Sensitivity analysis at the beginning of Stage 2 (three phase FMM) ......... 133

Figure 4.34 GA results for Stage 2 (three phase FMM) ................................................. 134

Figure 4.35 Uncertainty reduction in hydraulic fracture porosity during GA -

Stage 2 (three phase FMM) ......................................................................... 135

Figure 4.36 Uncertainty reduction in hydraulic fracture permeability during

GA - Stage 2 (three phase FMM) ................................................................ 135

Figure 4.37 Uncertainty reduction in hydraulic fracture initial water saturation

during GA - Stage 2 (three phase FMM) .................................................... 136

Figure 4.38 Uncertainty reduction in hydraulic fracture shape factor during

GA - Stage 2 (three phase FMM) ................................................................ 136

Figure 4.39 Uncertainty reduction in SRV porosity during GA - Stage 2

(three phase FMM) ...................................................................................... 137

Figure 4.40 Uncertainty reduction in SRV permeability during GA - Stage 2

(three phase FMM) ...................................................................................... 137

Figure 4.41 Uncertainty reduction in SRV initial water saturation during GA -

Stage 2 (three phase FMM) ......................................................................... 138

Figure 4.42 Uncertainty reduction in SRV shape factor during GA - Stage 2

(three phase FMM) ...................................................................................... 138

xxi
Figure 4.43 Variable distribution of hydraulic fracture porosity in the best selected

models of GA - Stage 2 (three phase FMM) ............................................... 139

Figure 4.44 Variable distribution of hydraulic fracture permeability in the best

selected models of GA - Stage 2 (three phase FMM) ................................. 139

Figure 4.45 Variable distribution of hydraulic fracture initial water saturation in

the best selected models of GA - Stage 2 (three phase FMM).................... 140

Figure 4.46 Variable distribution of hydraulic fracture shape factor in the

best selected models of GA - Stage 2 (three phase FMM) ......................... 140

Figure 4.47 Variable distribution of SRV porosity in the best selected models of

GA - Stage 2 (three phase FMM) ................................................................ 141

Figure 4.48 Variable distribution of SRV permeability in the best selected models

of GA - Stage 2 (three phase FMM) ........................................................... 141

Figure 4.49 Variable distribution of SRV initial water saturation in the best

selected models of GA - Stage 2 (three phase FMM) ................................. 142

Figure 4.50 Variable distribution of SRV shape factor in the best selected models

of GA - Stage 2 (three phase FMM) ........................................................... 142

Figure 4.51 Sensitivity analysis at the beginning of Stage 3 (three phase FMM) ......... 143

Figure 4.52 GA results for Stage 3 (three phase FMM) ................................................. 144

Figure 4.53 Uncertainty reduction in hydraulic fracture porosity during GA -

Stage 3 (three phase FMM) ......................................................................... 145

xxii
Figure 4.54 Uncertainty reduction in hydraulic fracture permeability during GA -

Stage 3 (three phase FMM) ......................................................................... 145

Figure 4.55 Uncertainty reduction in hydraulic fracture initial water saturation

during GA - Stage 3 (three phase FMM) .................................................... 146

Figure 4.56 Uncertainty reduction in hydraulic fracture shape factor during GA -

Stage 3 (three phase FMM) ......................................................................... 146

Figure 4.57 Uncertainty reduction in SRV porosity during GA - Stage 3 (three

phase FMM) ................................................................................................ 147

Figure 4.58 Uncertainty reduction in SRV permeability during GA - Stage 3

(three phase FMM) ...................................................................................... 147

Figure 4.59 Uncertainty reduction in SRV initial water saturation during GA -

Stage 3 (three phase FMM) ......................................................................... 148

Figure 4.60 Uncertainty reduction in SRV shape factor during GA - Stage 3

(three phase FMM) ...................................................................................... 148

Figure 4.61 Variable distribution of hydraulic fracture porosity in the best selected

models of GA - Stage 3 (three phase FMM) ............................................... 149

Figure 4.62 Variable distribution of hydraulic fracture permeability in the best

selected models of GA - Stage 3 (three phase FMM) ................................. 149

Figure 4.63 Variable distribution of hydraulic fracture initial water saturation in

the best selected models of GA - Stage 3 (three phase FMM).................... 150

xxiii
Figure 4.64 Variable distribution of hydraulic fracture shape factor in the best

selected models of GA - Stage 3 (three phase FMM) ................................. 150

Figure 4.65 Variable distribution of SRV porosity in the best selected models of

GA - Stage 3 (three phase FMM) ................................................................ 151

Figure 4.66 Variable distribution of SRV permeability in the best selected models

of GA - Stage 3 (three phase FMM) ........................................................... 151

Figure 4.67 Variable distribution of SRV initial water saturation in the best

selected models of GA - Stage 3 (three phase FMM) ................................. 152

Figure 4.68 Combined GA results for all stages (three phase FMM) ............................ 153

Figure 4.69 Cumulative oil history production data vs simulated production data

(a) in the first stage first generation and (b) including only the best

selected models from the last stage (three phase FMM) ............................. 154

Figure 4.70 Cumulative water history production data vs simulated production

data (a) in the first stage first generation and (b) including only the

best selected models from the last stage (three phase FMM) ..................... 155

Figure 4.71 Cumulative gas history production data vs simulated production data

(a) in the first stage first generation and (b) including only the best

selected models from the last stage (three phase FMM) ............................. 156

Figure 4.72 Oil rate history production data vs simulated production data (a) in the

first stage first generation and (b) including only the best selected

models from the last stage (three phase FMM) ........................................... 157
xxiv
Figure 4.73 Water rate history production data vs simulated production data (a) in

the first stage first generation and (b) including only the best selected

models from the last stage (three phase FMM) ........................................... 158

Figure 4.74 Gas rate history production data vs simulated production data (a) in

the first stage first generation and (b) including only the best selected

models from the last stage (three phase FMM) ........................................... 159

Figure 4.75 Cumulative Oil Production of FMM vs Eclipse as compared to History

data with base case variables (compositional FMM) .................................. 161

Figure 4.76 Oil Rate Production of FMM vs Eclipse as compared to History data

with base case variables (compositional FMM).......................................... 161

Figure 4.77 Cumulative Water Production of FMM vs Eclipse as compared to

History data with base case variables (compositional FMM) ..................... 162

Figure 4.78 Water Rate Production of FMM vs Eclipse as compared to History

data with base case variables (compositional FMM) .................................. 162

Figure 4.79 Cumulative Gas Production of FMM vs Eclipse as compared to

History data with base case variables (compositional FMM) ..................... 163

Figure 4.80 Gas Rate Production of FMM vs Eclipse as compared to History data

with base case variables (compositional FMM).......................................... 163

Figure 4.81 Sensitivity analysis at the beginning of Stage 1 (compositional FMM) ..... 164

Figure 4.82 GA results for Stage 1 (compositional FMM) ............................................ 165

xxv
Figure 4.83 Uncertainty reduction in hydraulic fracture porosity during GA -

Stage 1 (compositional FMM) .................................................................... 166

Figure 4.84 Uncertainty reduction in hydraulic fracture initial water saturation

during GA - Stage 1 (compositional FMM)................................................ 166

Figure 4.85 Uncertainty reduction in hydraulic fracture shape factor during GA -

Stage 1 (compositional FMM) .................................................................... 167

Figure 4.86 Uncertainty reduction in SRV porosity during GA - Stage 1

(compositional FMM) ................................................................................. 167

Figure 4.87 Uncertainty reduction in SRV permeability during GA - Stage 1

(compositional FMM) ................................................................................. 168

Figure 4.88 Uncertainty reduction in SRV shape factor during GA - Stage 1

(compositional FMM) ................................................................................. 168

Figure 4.89 Variable distribution of hydraulic fracture porosity in the first

generation of GA - Stage 1 (compositional FMM) ..................................... 169

Figure 4.90 Variable distribution of hydraulic fracture initial water saturation in

the first generation of GA - Stage 1 (compositional FMM)........................ 169

Figure 4.91 Variable distribution of hydraulic fracture shape factor in the first

generation of GA - Stage 1 (compositional FMM) ..................................... 170

Figure 4.92 Variable distribution of SRV porosity in the first generation of GA -

Stage 1 (compositional FMM) .................................................................... 170

xxvi
Figure 4.93 Variable distribution of SRV permeability in the first generation of

GA - Stage 1 (compositional FMM) ........................................................... 171

Figure 4.94 Variable distribution of SRV shape factor in the first generation of

GA - Stage 1 (compositional FMM) ........................................................... 171

Figure 4.95 Variable distribution of hydraulic fracture porosity in the best selected

models of GA - Stage 1 (compositional FMM) .......................................... 172

Figure 4.96 Variable distribution of hydraulic fracture initial water saturation in

the best selected models of GA - Stage 1 (compositional FMM) ............... 172

Figure 4.97 Variable distribution of hydraulic fracture shape factor in the best

selected models of GA - Stage 1 (compositional FMM) ............................ 173

Figure 4.98 Variable distribution of SRV porosity in the best selected models of

GA - Stage 1 (compositional FMM) ........................................................... 173

Figure 4.99 Variable distribution of SRV permeability in the best selected models

of GA - Stage 1 (compositional FMM) ....................................................... 174

Figure 4.100 Variable distribution of SRV shape factor in the best selected models

of GA - Stage 1 (compositional FMM) .................................................... 174

Figure 4.101 Sensitivity analysis at the beginning of Stage 2 (compositional

FMM) ........................................................................................................ 175

Figure 4.102 GA results for Stage 2 (compositional FMM) .......................................... 176

xxvii
Figure 4.103 Uncertainty reduction in hydraulic fracture porosity during GA -

Stage 2 (compositional FMM) .................................................................. 177

Figure 4.104 Uncertainty reduction in hydraulic fracture permeability during

GA - Stage 2 (compositional FMM) ......................................................... 177

Figure 4.105 Uncertainty reduction in hydraulic fracture initial water saturation

during GA - Stage 2 (compositional FMM) ............................................. 178

Figure 4.106 Uncertainty reduction in hydraulic fracture shape factor during GA -

Stage 2 (compositional FMM) .................................................................. 178

Figure 4.107 Uncertainty reduction in SRV porosity during GA - Stage 2

(compositional FMM) ............................................................................... 179

Figure 4.108 Uncertainty reduction in SRV permeability during GA - Stage 2

(compositional FMM) ............................................................................... 179

Figure 4.109 Uncertainty reduction in SRV initial water saturation during GA -

Stage 2 (compositional FMM) .................................................................. 180

Figure 4.110 Uncertainty reduction in SRV shape factor during GA - Stage 2

(compositional FMM) ............................................................................... 180

Figure 4.111 Variable distribution of hydraulic fracture porosity in the best

selected models of GA - Stage 2 (compositional FMM) .......................... 181

Figure 4.112 Variable distribution of hydraulic fracture permeability in the best

selected models of GA - Stage 2 (compositional FMM) .......................... 181

xxviii
Figure 4.113 Variable distribution of hydraulic fracture initial water saturation in

the best selected models of GA - Stage 2 (compositional FMM)............. 182

Figure 4.114 Variable distribution of hydraulic fracture shape factor in the best

selected models of GA - Stage 2 (compositional FMM) .......................... 182

Figure 4.115 Variable distribution of SRV porosity in the best selected models of

GA - Stage 2 (compositional FMM) ......................................................... 183

Figure 4.116 Variable distribution of SRV permeability in the best selected models

of GA - Stage 2 (compositional FMM) .................................................... 183

Figure 4.117 Variable distribution of SRV initial water saturation in the best

selected models of GA - Stage 2 (compositional FMM) .......................... 184

Figure 4.118 Variable distribution of SRV shape factor in the best selected models

of GA - Stage 2 (compositional FMM) .................................................... 184

Figure 4.119 Combined GA results of all stages (compositional FMM) ....................... 185

Figure 4.120 Cumulative oil history production data vs simulated production data

(a) in the first stage first generation and (b) including only the best

selected models from the last stage (compositional FMM) ...................... 186

Figure 4.121 Cumulative Water history production data vs simulated production

data (a) in the first stage first generation and (b) including only the

best selected models from the last stage (compositional FMM) .............. 187

xxix
Figure 4.122 Cumulative Gas history production data vs simulated production data

(a) in the first stage first generation and (b) including only the best

selected models from the last stage (compositional FMM) ...................... 188

Figure 4.123 Oil rate history production data vs simulated production data (a) in

the first stage first generation and (b) including only the best selected

models from the last stage (compositional FMM) .................................... 189

Figure 4.124 Water rate history production data vs simulated production data (a)

in the first stage first generation and (b) including only the best

selected models from the last stage (compositional FMM) ...................... 190

Figure 4.125 Gas rate history production data vs simulated production data (a) in

the first stage first generation and (b) including only the best selected

models from the last stage (compositional FMM) .................................... 191

Figure A.1 Input parameters in ML_Algorithms.R script – Part 1 ................................ 214

Figure A.2 Input parameters in ML_Algorithms.R script – Part 2 ................................ 215

xxx
 LIST OF TABLES

Page

Table 2.1: Exponent ‘b’ in Arp’s decline curves ............................................................. 12

Table 2.2 Response variables of decline models for Machine Learning.......................... 18

Table 2.3 Most suitable Machine Learning algorithm for each decline model ................ 46

Table 3.1 NPV variation with minimum matrix permeability used ................................. 94

Table 3.2 Economic Parameters for NPV calculations .................................................... 95

Table 3.3 Hydraulic fracture optimization variable ranges .............................................. 97

Table 3.4 NPV values correponding to various realizations vs base model or true

model ............................................................................................................. 106

Table 4.1 Uncertainty in Model parameters and their base values for Sensitivity

Analysis (Iino et al., 2017) ............................................................................ 115

Table A.1 Axis scale values used for Eagle Ford plots .................................................. 217

xxxi
 CHAPTER I

INTRODUCTION AND OBJECTIVES

1.1 Introduction

Reservoir Simulations in large and complex reservoirs can be very costly.

Specifically, in unconventional reservoirs, where reservoir models are usually represented

by millions of grid cells, oil and gas production forecasts can take a lot of time. Many

times, an engineer wants to get a quick idea about how a given well will deplete in future

so as to calculate the revenues that will be generated later on. Also, this may be needed

even before a detailed geologic information about a new well is provided. Previously,

studies have been done to predict maximum/mean oil production in a field using machine

learning approaches (LaFollette et. al, 2012 and 2013; Zhong et al., 2015). However, these

studies could not predict rate decline with time. The method presented in this chapter can

predict decline curve model parameters and predict rate decline for a new well based on

data collected from the field. This method is very fast after the needed data has been

gathered and properly cleaned/tabulated. In this chapter, this method has been applied to

calculate rate decline parameters of four commonly used decline models and also to

predict Estimated Ultimate Recovery (EUR) for a new well. This may provide an early

estimate of well production for a new well. Also, previous studies involved utilizing a

single model based predictions which is not a robust method since it would bias the model

towards the training data/machine learning tuning parameters. This chapter takes

1
advantage of a model averaging technique to make predictions based on weighted average

of multiple models built using more than one set of data/tuning parameters.

Another problem under investigation is of finding an optimum hydraulic fracturing

design in unconventional reservoirs. Previous studies in the literature involved application

of analytical models (e.g., PKN model) to predict well production. However, these models

are built for conventional reservoirs and are not suitable to be used in unconventional

reservoirs. Also optimization of hydraulic fractures in a given permeability field has been

presented earlier (Ma et. al, 2013). However, their study did not take into account the

uncertainty in the permeability field. The workflow presented in this chapter can be used

to optimize hydraulic fracture design for a given reservoir provided with some uncertainty

in the geologic data. This study also discusses uncertainty in the natural fracture

distribution and its effects on the Net Present Value (NPV). A synthetic reservoir model

has been used for this study and optimization problem is solved for maximizing the NPV.

This study also deals with a field-scale case history matching problem in which a

base model and parameters with their uncertainty are provided and a genetic algorithm

based history matching approach is utilized. Previous studies related to this work involved

history matching using a single set of uncertain parameters with a wide range of

uncertainty ranges. This chapter study utilizes a multi-stage GA approach that can be used

to identify key parameters (heavy-hitters) before proceeding to history matching. First

stage of this workflow involves using only the key parameters and matching observed

data. In subsequent stages, the refined variables achieved from the first stage are utilized

with reduced uncertainty ranges in them. The variables not included in the first stage are

2
also included in the subsequent stages. This method accelerates the convergence of a

stochastic history matching parameter which in this study is Genetic Algorithm (GA). This

study also integrates GA with a Fast Marching Method (FMM) based reservoir simulator

which is a faster alternative to commonly used commercial simulators. In this study,

simulated cumulative oil, water and gas production have been matched with their

corresponding observed/history data provided by the field operator. A production forecast

has also been made and corresponding production has been compared to test the accuracy

of history matching algorithm.

1.2 Dissertation Outline

This dissertation document contains several chapters each containing a different case

study. In Chapter II, Eagle Ford well data has been gathered from a publicly available

website and used with several machine learning algorithms in order to build models that

can predict rate declines for a new well. This method is very fast after the needed data has

been gathered and properly cleaned/tabulated. It can be used to calculate rate decline

parameters of commonly used decline models and also to predict Estimated Ultimate

Recovery (EUR) for a new well. This may provide an early estimate of well production

for a new well.

In Chapter III, a detailed workflow for hydraulic fracture design optimization has

been presented. This workflow based on genetic algorithm can be used to optimize

hydraulic fracture design for a given reservoir provided the geologic data including

permeability and porosity is known. This study also briefly discusses about the uncertainty

3
in the natural fracture distribution and its effects on the optimization of Net Present Value

(NPV). A synthetic reservoir model has been used for this study and optimization problem

is solved for maximizing the NPV.

In Chapter IV, a field case study has been presented in which a set of uncertain

parameters/variables with production history data are provided and objective is to match

history data by applying genetic algorithm based workflow. A multi-stage GA approach

has been used in this study to accelerate the convergence of GA. The multi-stage GA

approach utilizes heavy hitter variables in the first stage to fine tune the variables making

most impact. Subsequent stages, however include all variables with updated uncertainty

ranges. Simulated cumulative oil, water and gas production have been matched with their

corresponding observed/history data provided by the field operator. A production forecast

has also been made and corresponding actual production has been compared to test the

accuracy of history matching algorithm.

Finally, in Chapter V, conclusions from this dissertation study have been presented

and recommendations for possible extension/improvement to current work are suggested.

4
 CHAPTER II

MACHINE LEARNING BASED INSIGHTS ON WELL PERFORMANCE IN

EAGLE FORD WELLS

2.1 Introduction and Literature Review

Oil and gas wells have been in existence for a long time but it was only in recent

times when importance of large sets of well data are realized by the petroleum industry.

A large set of well data which includes well location data and well completion data are

becoming available in a format that can be easily used by data scientists. Since shale oil

and gas revolution started in USA, a large number of wells have been drilled and their data

collected. Many of these data are available in publically accessible websites on internet.

This chapter deals with a study done using well data collected from more than 100 wells

in the Eagle Ford reservoir. Well data used for this study include well location/depth

parameters including latitude, longitude and total vertical depth and well completion

parameters including number of hydraulic fractures, volume of fracturing fluid used,

amount of proppant used, and completed length. Well data has been collected from the

online database DrillingInfo. Only oil wells have been selected for this study.

Lee et al. (2002) applied classification and non-parametric regression algorithms

for electrofacies characterization and permeability prediction in complex reservoirs.

Model based clustering technique was used to identify clusters from well log responses.

For each cluster, non-parametric regression technique was utilized to build model and

predict corresponding permeability. The non-parametric regression algorithms include

5
ACE (Alternating Conditional Expectation), GAM (Generalized Additive Model) and

NNET (Neural Networks). ACE based regression algorithm outperformed the other two

regression methods in this study.

Perez et al. (2005) applied classification trees with well log response to predict

electrofacies, lithofacies and hydraulic flow units in uncored wells. This study also

reported the predictor variables that have most influence in classification tree based

prediction. It was also reported that larger trees may be too sensitive to the statistical noise

present in the data and therefore smaller (pruned) trees should be used for such kind of

study.

Mishra (2012) reported a method to make predictions based on multiple models

instead of single one. The final prediction is based on weighted average of predictions

from all models. It was shown that more than one decline model can be fitted to a data

with acceptable accuracy. However, their future predictions may vary a lot. To overcome

this problem, the final predicted response variable, Estimated Ultimate Recovery (EUR)

was predicted using multiple models aggregated together by Generalized Likelihood

Uncertainty Estimation or GLUE (Beven and Binley, 1992; Neuman, 2003; Singh et al.

2010) methodology.

LaFollette and Holcomb (2011) presented data analytic results using Barnett shale

horizontal wells. It was found that wells more than 3,500 – 4,500 ft of lateral length were

less efficient in terms of production per foot. Also, it was found that, most wells are drilled

in approximately 140 and 320 degrees of azimuth. Also, the best wells were those that

were drilled near horizontal.

6
 LaFollette et al. (2012) reported results for Bakken formation of the Eastern

Williston Basin. They found production efficiency (production per foot of completed

lateral) decreases with increasing lateral length. It shows that increasing number of stages

and completed length alone did not find positive correlation with maximum monthly oil

production (calculated during first 12 month production period). However, proppant

concentration seemed to have a positive correlation with maximum monthly oil

production.

LaFollette et al. (2012) presented results of North Texas Barnett Shale wells with

emphasis on well completion and fracture stimulation. It was concluded in this paper that

traditional linear regression methods are not suitable for this kind of data: prone erroneous

data, missing data, non-linear data and data containing subtle interrelationships among

variables. It was concluded that boosted tree method is more suited for this kind of data

for regression purposes. The study also found a good correlation between maximum

monthly oil production and amount of fracturing fluid used for fracking in the wells

studied.

LaFollette (2013) presented data analytics results from Barnett shale and Bakken

Shale. In Barnett shale case, relative influence of various variables in predicting maximum

monthly gas production during first 12 month period was studied. TVD is found to be the

most influential factor in this study using boosted tree model. In Bakken shale case,

relative influence of various variables in predicting maximum monthly oil production

during first 12 month period was studied. In this case, well location coordinates were

found to be most influential in the study done using boosted tree model.

7
 LaFollette et al. (2013) reported results using well data gathered from Bakken

Light Tight Oil Play. This study was carried out using multivariate analysis of production

data. It was found that well location that can be used as a proxy for reservoir quality is one

of the most influential predictor for production forecast. It was also concluded that longer

lateral wells are less efficient in terms of production per feet of lateral length.

LaFollette et al. (2014) reported results using well data gathered from Eagle Ford

Formation in South Texas. This study carried out multivariate analysis on Eagle Ford

production data. Reservoir quality was proxied by X-Y surface location since

petrophysical data was unavailable. The completion variables used for this study included

proppant amount, volume of fracturing fluid used, number of fracturing stages, and

completed length (measured as difference between measured depths of bottom perforation

and top perforation). Other variables included dip, azimuth and GOR. The proxies for

production efficiency include maximum oil rate, barrels of oil produced per unit

completed length and barrels of oil produced per pound of proppant used. The paper also

reported trends in reservoir fluid parameters.

The study reported that GOR and well location are among the most important

variables influencing multivariate analysis. This study also reported that even though

production rates increases with increase in completed lateral length, the production per

unit completed length reduces as completed length increases. Increase in proppant amount

used for completion jobs is found to increase productivity in terms of maximum monthly

production.

8
 Holcomb et al. (2015) studied the productivity effects from spatial placement and

well architecture in Eagle Ford shale horizontal wells. This study found that wells drilled

and completed in GOR less than 5000 scf/bbl have lower maximum monthly oil

production (during first 12 month period) per foot of length but then appear to have a

lower percentage decline rate than higher GOR wells. This study could not find direct

correlation between increased proppant consumption and increased well productivity.

Zhong et al. (2015) reported their results with Wolfcamp shale. They applied

several machine learning algorithms to build models that can predict first 12 months of

cumulative oil for oil wells. Machine learning algorithms used included Ordinary Least

Squares (OLS), Support Vector Machines (SVM), Random Forests (RF) and Gradient

Boosting Model (GBM). In their results, RF modeled the data most accurately. Also, they

reported the predictor relative importance based on R2 loss. In this method, each of the

predictor variable was removed from predictor set one at a time while keeping rest of the

predictors intact and checking the change in R2, i.e., R2 loss. The predictor having more

R2 loss associated with it is considered more important. Different machine learning

algorithms had different ranking/predictor importance order in this study. In case of RF,

fracturing fluid amount used for completion job turned out to be most influential factor.

Schuetter et al. (2015) reported their machine learning study using data set

comprising wells in Wolfcamp Shale in West Texas (Delaware Basin and Central Basin).

Response variable in this study was cumulative production in the first 12 months of oil

production period. This study tried to predict first 12 month cumulative production for

new test data wells based on machine learning models developed using training wells.

9
Machine learning algorithms used here were Ordinary Least Squares, Random Forest,

Gradient Boosting Machine, Support Vector Regression (SVR) and Kriging. K-fold cross-

validation technique was utilized to avoid overfitting. It was found that although Kriging

based models fits training data perfectly, they did not perform well for test data. Also,

study includes relative importance study of various predictor variables. It was found that

TVD is most influential predictor among all predictors.

Centurion et al. (2012) presented their data analytics results using Eagle Ford well

data. It was pointed out that most of the top productive wells in Eagle Ford lie in the

counties of Dewitt and Karnes. However, the worst performing wells are not located in a

particular location. Also, the wells completed using delayed release production chemicals

have higher productivity than those which didn’t use those chemicals. In the multivariable

statistical analysis, most dominant predictors were identified and they included proppant

volume, injection rates, treatment pressure, measured depth of deepest perforation,

production chemicals combined with stimulation fluids and porosity indicator.

Centurion et al. (2013) reported their multivariate analysis results using Eagle Ford

well data. The most significant variables found in their study were proppant per ft,

pressure, cluster spacing, thickness, average porosity and perforation length.

Centurion et al. (2014) reported their data analytic results using LaSalle County

wells in Eagle Ford shale. Cumulative oil production during first 3 months was considered

as a proxy for well productivity. Multivariate analysis results showed most influential

variables in this region to be completed length and stage spacing. Proppant pumped

showed positive correlation with well productivity. Also, increased shut-in time between

10
hydraulic fracture treatment and the first day of production also had a positive effect on

well productivity. Reduction in well spacing led to lower initial productivity but increased

overall productivity of the region in a longer term.

2.2 Methodology

Eagle Ford well data has been downloaded from drillinginfo (website:

info.drillinginfo.com). More than 100 well data has been collected and analyzed using

various machine learning techniques. First, well data has been analyzed using exploratory

data analytic techniques such as scatterplot and boxplot. Next, machine learning

techniques such as Random Forest (RF), Gradient Boosted Machine (GBM), Support

Vector Machine (SVM) and Multivariate Adaptive Regression Splines (MARS) have been

utilized in order to predict rate decline in Eagle Ford wells. Since the production rate data

of these wells are mostly noisy, it is difficult to model them with smooth models. However,

a novel approach explained in this section can handle this problem using machine learning

algorithms in conjunction with decline rate models used in oil industry. The well rate data

is first fitted with one of the commonly used decline models listed below.

2.2.1 Rate Decline Models

2.2.1.1 Arp’s Decline Model

Arp’s decline equation (Arps, 1945) can be represented as follows:

𝑞𝑖
q(t) = 1 (2.1)
(1+𝑏𝐷𝑖 𝑡)𝑏

where,
11
𝑞(𝑡) = rate at time t (STB/D)

𝑞𝑖 = initial rate (STB/D)

𝐷𝑖 = initial decline rate (1/month)

𝑏 = hyperbolic decline coefficient (dimensionless)

𝑡 = time (months)

Exponent b in above equation shows type of decline in a well (Table 2.1).

Table 2.1: Exponent ‘b’ in Arp’s decline curves

b value Decline type

b=0 Exponential

0<b<1 Hyperbolic

b=1 Harmonic

Fig. 2.1 shows an example well’s predictions made by Arp’s decline model

keeping Initial flow rate, Di same but varying exponent, b. It may be seen that for higher

b values, model predicts higher production rates.

12
 Figure 2.1 An example well prediction made by Arp’s decline model

2.2.1.2 Stretched Exponential Decline Model (SEDM)

Valko and Lee (2010) presented Stretched Exponential Decline Model which is a

specialized decline model for unconventional reservoirs and predicts rate decline in

transient flow regime. Since unconventional wells produce in transient flow regimes,

SEDM is more suitable for them compared to Arp’s decline model. Eq. 2.2 shows SEDM

equation.

𝑡 𝑛
q(t) = 𝑞𝑖 𝑒𝑥𝑝 [− (𝜏) ] (2.2)

where,

𝑞(𝑡) = rate at time t (STB/D)

𝑞𝑖 = initial rate (STB/D)

𝜏 = characteristic relaxation time (month)

𝑛 = exponent parameter (dimensionless)

13
𝑡 = time (months)

Johnston (2006) explained stretched exponential decay process as a sum of

exponential decay with a “fat tailed” probability distribution of time constants. Valko and

Lee (2010) explained SEDM to be a sum of large number of individual exponential decays.

It was also reported by Valko and Lee (2010) that Arp’s may predict physically unrealistic

Estimated Ultimate Recovery (EUR) values for b ≥ 1 but SEDM will always give finite

value of EUR. Fig. 2.2 shows how Arp’s can fit early rate data really well but would over

predict production at long term period.

Figure 2.2 Comparison of Arp’s and SEDM decline models

14
 2.2.1.3 Duong Model

Duong (2011) presented following equation in the case of fracture dominated flow

characteristics. This equation (Eq. 2.3) is derived empirically for shale gas and tight gas

reservoirs.

𝑎
q(t) = 𝑞1 𝑡 −𝑚 𝑒𝑥𝑝 (1−𝑚 (𝑡1−𝑚 − 1)) (2.3)

where,

𝑞(𝑡) = rate at a time t (STB/D)

𝑞1 = flow rate on first day (STB/D)

𝑎 = intercept constant

𝑚 = slope parameter. Duong (2011) showed that for the unconventional reservoirs m > 1

𝑡 = time (months)

2.2.1.4 Weibull Model

Another way to model decline curve is through Weibull growth curve (Weibull,

1951; Mishra, 2012).This equation (Eq. 2.4) is generally used for modeling time-to-failure

in applied engineering problems.

𝑡 𝛾
𝑃(𝑡) ≡ 𝐺𝑃 = 𝑀 {1 − 𝑒𝑥𝑝 (− (𝛼) )} (2.4)

where,

𝐺𝑃 = cumulative production at time t

𝑀 = carrying capacity (Max. cumulative production)

𝛾 = shape parameter

15
𝛼 = scale parameter

𝑡 = time (months)

Differentiating Eq. 2.4 gives (Weibull, 1951; Mishra, 2012):

𝛾 𝑡 𝛾−1 𝑡 𝛾
q(t) = 𝑀 𝛼 (𝛼) 𝑒𝑥𝑝 (− (𝛼) ) (2.5)

where,

𝑞(𝑡) = rate at time t (STB/month)

M, the carrying capacity, is the maximum cumulative production set by this

equation. This means that cumulative production cannot reach unrealistic values as in the

Arp’s model in some cases. Since it is a fitting parameter like 𝛼 and 𝛾, a close approximate

value of M is needed to fit Weibull curve on a well rate decline data. For this study,

cumulative well oil production during the available well oil production period with ± 10

% margin has been assumed for best range within which M should lie. 𝛼 , the scale

parameter, is that value of time at which (1-1/e) or 63.2% of the resources have been

produced (Mishra, 2012). 𝛾, the shape factor, shows how rate of growth changes with time

and is usually less than 1 for unconventional reservoirs (Mishra, 2012).

Once the well rate data is collected for all the wells included in this study, all of

the above decline models are used to fit them with a best match and the parameters of

corresponding decline models are stored for further study. Also, the Estimated Ultimate

Recovery (EUR) for each well is calculated as a numeric integral of monthly oil

production over 30 year period (360 months):

𝐸𝑈𝑅 = ∑360
𝑖=1 𝑞𝑖 (2.6)

where,
16
𝐸𝑈𝑅 = Estimated Ultimate Recovery

𝑞𝑖 = monthly oil rate (STB/month) of 𝑖 𝑡ℎ month

2.2.2 Machine Learning Algorithms

Once well rate data is collected and fitted with the decline models discussed

previously, the data is tabulated such that each row corresponds to a well and each column

corresponds to one of the variables (predictors or responses). Table 2.2 shows the

response and predictor variables used for each of the decline curve models. As shown in

Table 2.2, predictor variable are unchanged across each of the decline models but

response variables change.

The data table is divided randomly into 80% - 20% partition so that 80% of the

rows are utilized to train machine learning model (called as training data) and remaining

20% of the rows are used for testing (called as test data) the model accuracy. In this study,

different machine learning algorithms have been applied to the data under investigation.

Following subsections briefly presents the main idea behind some of these algorithms that

provided better results than the remaining ones. The three machine learning algorithms

that produced better prediction results than others are: Random Forests (RF), Gradient

Boosted Machines (GBM) and Support Vector Machines (SVM). However, results for

Multivariate Adaptive Regression Splines (MARS) are also shown in this chapter for

comparison purposes only.

17
 Table 2.2 Response variables of decline models for Machine Learning

Arp’s SEDM Duong Weibull

Response

Variables 𝐷𝑖 , 𝑏, 𝐸𝑈𝑅 𝑡𝑎𝑢, 𝑛, 𝐸𝑈𝑅 𝑎, 𝑚, 𝐸𝑈𝑅 𝛾, 𝛼, 𝑀, 𝐸𝑈𝑅

Well Latitude and Longitude, TVD, Difference between TVDs of Heel and

Predictor Toe, Completed Length, Number of Fracture stages, Amount of fracturing

Variables fluid and Proppant used for fracking

Once a model has been trained, it can then predict the decline curve parameters of

new wells which in this case are test data wells. Oil rate decline with respect to time can

then be predicted by using decline curve parameters and corresponding decline equation.

This study also deals with finding the relative influence of various predictor

variables for building a model. This can be regarded as a variable importance or sensitivity

study in which it is possible to identify most important and least important predictor

variables to build a model.

A short description of four of the machine learning algorithms applied to Eagle

Ford study is presented in the following sections.

2.2.2.1 Random Forests (RF)

Breiman (2001) reported an ensemble based learning method based on

Classification and Regression Trees (CART) concept. A single Classification Tree

consists of a series of partition such that each partition divides data points into two

18
dissimilar groups as shown in Fig. 2.3 (a). However, in reality, a partition by linear

boundaries may not be able to partition data into pure classes. This is shown in Fig. 2.3

(b) by impurities of whites among black colored circles and impurities of blacks among

white circles. These impurities can be minimized by further partitioning the variable space.

The mathematical quantity to be minimized here is called the Gini Impurity Index

(Breiman, 1996) which is a measure of impurities present in a given

partition/compartment.

𝐺𝑖𝑛𝑖 𝐼𝑚𝑝𝑢𝑟𝑖𝑡𝑦 𝐼𝑛𝑑𝑒𝑥 = ∑𝑘𝑖=1 𝑝𝑖 (1 − 𝑝𝑖 ) = 1 − ∑𝑘𝑖=1 𝑝𝑖2 (2.7)

where,

𝑝𝑖 = probability of training dataset belonging to 𝑖 𝑡ℎ class

𝑘 = number of classes (or categorical variables)

In a pure node (consisting only of one type of class), this Gini Index should be

equal to 0. In order to partition a variable space, different possibilities are tested including

different variables and different point of partition in a given variable’s range. This is

repeated at each node until Gini’s Index is minimized or number of terminal nodes exceed

the specified set limit. The final prediction value at a terminal node is governed by

majority vote.

19
 (a) (b)

Figure 2.3 (a) Classification Tree example (b) Equivalent partition for a two variable case

Regression Trees are similar to a Classification Trees but in their case prediction

is made for a continuous variable (real number) instead of a categorical variable (class) as

shown in Fig. 2.4.

Figure 2.4 An example Regression Tree from Eagle Ford data predicting maximum oil

production

20
 The values at each node is calculated by minimizing Residual Sum of Squares

(RSS) using Eqs. 2.8 and 2.9 (Shalizi, 2006):

𝑛𝑐
𝑅𝑆𝑆 = ∑𝑛𝑐=1 ∑𝑖=1(𝑦𝑖 − 𝑚𝑐 )2 (2.8)

1
𝑛𝑐
𝑚𝑐 = 𝑛 ∑𝑖=1 𝑦𝑖 (2.9)
𝑐

where,

𝑐 = number of nodes

𝑛𝑐 = number of data points in a node

𝑦𝑖 = observed or actual response value

In order to partition a variable space, different possibilities are tested including

different variables and different point of partition in a given variable’s range. This is

repeated at each node until RSS is minimized or number of terminal nodes exceed the

specified set limit. The final prediction value at a terminal node is governed by mean

prediction value. Cost Complexity (Cp) in a regression tree (Perez et. al, 2003) is given

by:

Cp = Training Error + k × No. of terminal nodes (2.10)

where,

k = cost complexity factor. If k = 0, tree will not control no. of terminal nodes and only

error rates are involved making tree larger than needed. If k is very large, tree will be very

short with high training error and biased model

Fig. 2.5 shows Cp vs cross validation error/misfit error in Eagle Ford data. As can

be seen in this figure tree size of 2 gives minimum Cp. However, it must be noted that a

very small size of tree can bias the model for the training data. In the Random Forest
21
package in R, tree sizes are controlled by providing a range within which total number of

terminal nodes should lie. This is an indirect way of controlling Cp. The default minimum

number of nodes is 5 for regression trees in Random Forest package used in this study.

Therefore in the example shown below, the tree size of 5 would be appropriate.

Figure 2.5 Cost complexity and size of a regression tree against misfit error using Eagle

Ford data

22
 A Random Forest (Breiman, 2001) is an ensemble based machine learning

algorithm which is comprised of a large number of uncorrelated trees (Classification or

Regression Trees). Instead of fitting data with a single Classification or Regression Tree,

a random forest of multiple uncorrelated trees is constructed. Each tree is derived from a

bootstrap subsample of given data as well as a bootstrap subsample of variables from

predictor variable set leading to a different order of partitioning. During prediction process

for a new dataset (not used for training the Random Forest), final prediction is based on

majority vote (Random Forest of Classification Trees) or averaged response (Random

Forest of Regression Trees).

2.2.2.2 Gradient Boosted Machine (GBM) Regression

Gradient Boosted Machine (Friedman, 2001 and 2002) is an ensemble tree based

machine learning algorithm in which a true model is represented by a series of trees such

that each subsequent tree is fitting the error residual of the previous tree (Fig. 2.6).

Friedman (2001 and 2002) reported that “Gradient Boosting of the regression trees

produces competitive, highly robust, interpretable procedures for both regression and

classification, especially mining less than clean data”.

23
 Figure 2.6 Approximate representation of a Gradient Boosted Tree Model

(Modified from Gradient Boosted Regression Trees in scikit-learn,

https://www.slideshare.net/DataRobot/gradient-boosted-regression-trees-in-scikitlearn)

A simple mathematical formulation of gradient boosted trees is presented below

(source: scikit-learn.org website (http://scikit-learn.org/stable/modules/ensemble.html).

A general form of additive model is given by:

𝐹(𝑥) = ∑𝑀
𝑚=1 𝛾𝑚 ℎ𝑚 (𝑥) (2.11)

𝛾𝑚 = step length

ℎ𝑚 (𝑥) = basis functions

The gradient boosting additive model can be represented as:

𝐹𝑚 (𝑥) = 𝐹𝑚−1 (𝑥) + 𝛾𝑚 ℎ𝑚 (𝑥) (2.12)

where,

ℎ𝑚 (𝑥) = regression/classification tree used as a basis functions/weak learners

For each stage, ℎ𝑚 (𝑥) is chosen to minimize the loss function L for the given model 𝐹𝑚−1

and its fit 𝐹𝑚−1 (𝑥𝑖 )

𝐹𝑚 (𝑥) = 𝐹𝑚−1 (𝑥) + 𝑎𝑟𝑔 min ∑𝑛𝑖=1 𝐿(𝑦𝑖 , 𝐹𝑚−1 (𝑥𝑖 ) − ℎ(𝑥)) (2.13)
ℎ

This minimization problem is solved numerically via steepest descent method.

24
𝐹𝑚 (𝑥) = 𝐹𝑚−1 (𝑥) + 𝛾𝑚 ∑𝑛𝑖=1 ∇𝐹 𝐿(𝑦𝑖 , 𝐹𝑚−1 (𝑥𝑖 )) (2.14)

where,

𝜕𝐿(𝑦𝑖 ,𝐹𝑚−1 (𝑥𝑖 ))

𝛾𝑚 = 𝑎𝑟𝑔 min ∑𝑛𝑖=1 𝐿 (𝑦𝑖 , 𝐹𝑚−1 (𝑥𝑖 ) − 𝛾 ) (2.15)
𝛾 𝜕𝐹𝑚−1 (𝑥𝑖 )

The initial model, 𝐹0 (𝑥) is usually chosen to be the mean of target values in case of

regression problems.

2.2.2.3 Support Vector Machines (SVM) Regression or Support Vector Regression

(SVR)

Smola and Schölkopf (2004) presented Support Vector Regression (SVR) or Support

Vector Machine (SVM) Regression which has become quite successful among machine

learning algorithms. This algorithm tries to fit function, f(x), on a given training dataset

such that the maximum deviation of a data point from this function is equal to ε. However,

complexity of f(x) is controlled so that f(x) is kept as flat as possible.

Eq. 2.16 shows the term that is needed to be minimized and Eq. 2.17 shows that

constraints used while minimizing Eq. 2.16.

Objective is to find: 𝑓(𝑥⃗) = 𝑤

⃗⃗⃗. 𝑥⃗ + 𝑏, by:
1
minimizing: ‖𝑤‖2 + 𝐶 ∑𝑁 ∗
𝑖=1(𝜉𝑖 + 𝜉𝑖 ) (2.16)
2

𝑦𝑖 − (𝑤 ⃗⃗⃗. 𝑥⃗ + 𝑏) ≤ 𝜀 + 𝜉𝑖
⃗⃗⃗. 𝑥⃗ + 𝑏) ≤ −(𝜀 + 𝜉𝑖∗ )
subjected to constraints: {𝑦𝑖 − (𝑤 (2.17)
𝜉𝑖 , 𝜉𝑖∗ > 0

Eq. 2.16 also shows the slack term variables (Cortes and Vapnik, 1995, Smola and

Schölkopf, 2004) in order to avoid overfitting in the model. The second term in Eq. 2.16

25
shows the cost term containing slack variables, 𝜉𝑖 , 𝜉𝑖∗ which include points with deviations

more than 𝜀 . By controlling the constant C (where C > 0), the contribution of the second

term in Eq. 2.16 can be controlled. This is also a way to control the trade-off between the

flatness of f(x) and the limit up to which data points having deviations larger than ε are

tolerated in the machine learning model. Using Lagrange multipliers (𝛼𝑖 , 𝛼𝑖∗ ) to solve

above minimization problem, the above equations become:

𝑤 = ∑𝑙𝑖=1(𝛼𝑖 − 𝛼𝑖∗ ) 𝑥𝑖 (2.18)

𝑓(𝑥) = ∑𝑙𝑖=1(𝛼𝑖 − 𝛼𝑖∗ ) < 𝑥𝑖 , 𝑥 > + 𝑏 (2.19)

where,

𝛼𝑖 , 𝛼𝑖∗ = Lagrange multiplier

<. , . > = dot product

Aizerman et al. (1964) and Nilsson (1965) showed how to map a training data to

some feature space ℱ i.e., Φ: Χ → ℱ. This process simplifies the problem such that the

optimization problem tries to find function f(x) in the feature space and not in actual input

space.

𝑓(𝑥) = ∑𝑙𝑖=1(𝛼𝑖 − 𝛼𝑖∗ )𝑘(𝑥𝑖 , 𝑥) + 𝑏 (2.20)

Once the data is in feature space, the function f(x) to be fitted can be more flat than

fitting it in original data space.

2.2.2.4 Multivariate Adaptive Regression Splines (MARS)

Freidman (1991 and 1993) reported Multivariate Adaptive Regression Splines

(MARS). Eq. 2.21 shows the basic form of MARS:

26
𝑓̂(𝑋) = 𝑎0 + ∑𝑀
𝑚=1 𝑎𝑚 𝐵𝑚 (𝑋) (2.21)

where,

𝑎0 = constant

{𝑎𝑚 }1𝑀 are the coefficients of expansion whose values are determined by least square fit

of above equation:

{𝑎𝑚 }1𝑀 = argmin ∑𝑛𝑛=1[𝑦𝑛 − 𝑎𝑚 𝐵𝑚 (𝑋)]2 (2.22)

{𝛼𝑚 }𝑀
1

X = {𝑥1 , 𝑥2 , … , 𝑥𝑝 } = variables in training data set

𝐵𝑚 (𝑋) = basis function

A basis function can be a constant, a hinge function or a product of any

combination of one or more hinge functions. A hinge function is of following form:

𝑥 − 𝑡, 𝑖𝑓 𝑥 > 𝑡
[𝑥 − 𝑡]+ = 𝑚𝑎 𝑥(0, 𝑥 − 𝑡) = { (2.23)
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

𝑡 − 𝑥, 𝑖𝑓 𝑥 < 𝑡
[𝑡 − 𝑥]+ = 𝑚𝑎 𝑥(0, 𝑡 − 𝑥) = { (2.24)
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

In above equations, the constant t is called as a knot, which is a point at which

model function f(X) changes direction. The final form of MARS equation becomes

(Friedman 1991):
𝐾𝑚
𝑓̂(𝑋) = 𝑎0 + ∑𝑀
𝑚=1 𝑎𝑚 ∏𝑘=1[±(𝑥𝑣(𝑘,𝑚) − 𝑡𝑘𝑚 )] (2.25)
+

where,

{𝑣(𝑘, 𝑚)}1𝐾𝑚 = variable set associated with 𝑚𝑡ℎ basis function 𝐵𝑚

The training process in MARS algorithm consists of a Forward Pass and a

Backward Pass. During Forward Pass, a pair of terms are added at each step until a pre-

27
specified limit of maximum number of terms is reached. On the contrary, during the

Backward Pass, the least effective term is removed in each step (one term at a time). To

decide which term needs to be discarded, Generalized Cross-Validation is used. Eq. 2.25

gives the formula to calculate GCV. It is proportional to the data fitting error but inversely

proportional to the number of terms in the model. GCV is a trade-off between the number

of terms and the Mean Squared Error (MSE) and helps dealing with the problem of

overfitting in MARS. Generalized Cross Validation (GCV) is calculated as:

1 2
∑𝑁 ̂
𝑖=1[𝑦𝑖 −𝑓𝑀 (𝑥𝑖 )]
𝐺𝐶𝑉 = 𝑁 𝐶(𝑀) 2
(2.26)
[1− ]
𝑁

𝑦𝑖 = observed values

𝑓̂𝑀 (𝑥𝑖 ) = model predicted values

𝑁 = no. of observations/predictions

𝐶(𝑀) = cost complexity function ∝ no. of basis functions used in model

At the end of the forward pass, an over fit MARS model larger than needed terms

is trained. Backward pass or the pruning pass consists of removing terms from existing

MARS equation in steps and checking GCV. GCV should first decrease to a minimum

value before taking off again. At that point optimum number of terms are achieved. Fig.

2.7 shows a GCV plot for a MARS model with Eagle Ford data. In this figure the removal

of terms should be stopped at step number 5.

28
 Figure 2.7 An example of GCV plot using Eagle Ford data

2.2.3 Model Averaging

One of the usual practice to train a machine learning model is to use an entire

training dataset by minimizing the training data misfit. Another way is to use a k-fold cross

validation approach. This dissertation section involves k-fold validation approach for

calculation of misfit. Fig. 2.8 shows steps for training a machine learning model using this

approach. Once raw well data is collected which in current study is from Eagle Ford

database, each oil well’s rate decline is fitted with one of the four decline models – Arp’s

(Arp’s 1945), SEDM (Valko and Lee, 2010), Duong (Duong, 2011) or Weibull (Weibull,

1951 and Mishra, 2012). The corresponding parameters of these decline models are then

derived based on best fit (Table 2.2). The dataset now contains both predictor variables

and response variables. Outlier points are removed based on engineering judgement, e.g.,

wells having unrealistic proppant mass or fluid volumes are removed. This dataset is now

29
split into 80% training data and 20% test data. Test data is not used for training any of the

Machine Leaning models in this study. Training data is further split into 10-folds (k = 10).

As shown in Fig. 2.8, various combinations of training data subset and test data subset can

be derived from main training data. This training data can be used to train a machine

learning Model with different input values of tuning parameters provided in the grid form

to the training data set. Therefore, each of the training data subset set with one of the

tuning parameter combination results in a single machine learning model which is tested

against corresponding test data subset resulting in an error calculated in terms of RMSE.

A large number of such models with corresponding RMSE errors are then used to predict

the main test data (not used for training purposes). However, since each model will predict

a different value of a response variable, a model averaging technique known as

Generalized Likelihood Uncertainty Estimation or GLUE is utilized here to combine the

outputs of all trained machine learning models and result in single output prediction.

Model averaging helps dealing with problem of overfitting.

30
 Figure 2.8 Workflow steps for model training and prediction

2.2.3.1 Generalized Likelihood Uncertainty Estimation (GLUE)

Generalized Likelihood Uncertainty Estimation or GLUE is derived from

Bayesian Model Averaging. Eq. 2.27 shows Bayesian Model Averaging method. This

method calculates the weights for individual models and the final output prediction is
31
weighted average of all models. For a given model j, its weight is given by (Draper 1995,

Kass and Raftery 1995 and Hoeting et al. 1999):

𝑝(𝐷|𝑀𝑗 )𝑝(𝑀𝑗 )
𝑤𝑒𝑖𝑔ℎ𝑡𝑠, 𝑤𝑗 ∝ 𝑝(𝑀𝑗 |𝐷) = ∑ (2.27)
𝑗 𝑝(𝐷|𝑀𝑗 )𝑝(𝑀𝑗 )

where,

𝑝(𝑀𝑗 ) = prior probability of Model 𝑗

𝑝(𝐷|𝑀𝑗 ) = model likelihood given by prediction error for data D

= ∫ 𝑃(𝑑|𝜃𝑗 , 𝑀𝑗 )𝑝(𝜃𝑗 |𝑀𝑗 )𝑑𝜃𝑗

𝑃(𝑑|𝜃𝑗 , 𝑀𝑗 ) = joint probability of a model 𝑗 (function of prediction errors)

𝑝(𝜃𝑗 |𝑀𝑗 ) = prior probabilities of parameters

Since it is difficult to calculate the likelihood integral, Beven and Binley (1992)

and Beven (2000) proposed GLUE formula which simplified Eq. 2.27 with Eq. 2.28.

𝜎𝑒,𝑗 2
𝑝(𝐷|𝑀𝑗 ) ∝ 𝑒𝑥𝑝 [−𝑁 ] (2.28)
𝜎𝑜 2

where,

𝑁 = shape factor

𝑆𝑢𝑚 𝑜𝑓 𝑠𝑞𝑢𝑎𝑟𝑒𝑑 𝑒𝑟𝑟𝑜𝑟𝑠

𝜎𝑒,𝑗 = variance of the errors of model 𝑗 = 𝑛𝑜.𝑜𝑓 𝑜𝑏𝑠𝑒𝑟𝑣𝑎𝑡𝑖𝑜𝑛𝑠

𝜎𝑜 = variance in the observed data

𝑁 ≫ 1 tends to give higher weightage to models with less fitting error

𝑁 ≪ 1 tends to give similar weights to all models

Therefore, model weights are given by:

32
 𝜎𝑒,𝑗 2
𝑒𝑥𝑝[−𝑁 ]𝑝(𝑀𝑗 )
𝜎𝑜 2
𝑤𝑒𝑖𝑔ℎ𝑡𝑠, 𝑤𝑗 ∝ 𝑝(𝑀𝑗 |𝐷) = 𝜎𝑒,𝑗 2
(2.29)
∑𝑗 𝑒𝑥𝑝[−𝑁 2 ]𝑝(𝑀𝑗 )
𝜎𝑜

A modified GLUE formula has been proposed by Mishra (2012) which simplifies

Eq. 2.29 even further:

𝑁
𝜎 2
𝑝(𝐷|𝑀𝑗 ) ∝ (𝜎 𝑜 2 ) (2.30)
𝑒,𝑗

𝑁
𝜎 2
( 𝑜 2 ) 𝑝(𝑀𝑗 )
𝜎𝑒,𝑗
𝑤𝑒𝑖𝑔ℎ𝑡𝑠, 𝑤𝑗 ∝ 𝑝(𝑀𝑗 |𝐷) = 𝑁 (2.31)
𝜎 2
∑𝑗( 𝑜 2 ) 𝑝(𝑀𝑗 )
𝜎𝑒,𝑗

or,
1
𝑝(𝐷|𝑀𝑗 ) ∝ 𝑅𝑀𝑆𝐸 2 (2.32)
𝑗

where,

𝑅𝑀𝑆𝐸𝑗 = Root Mean Square Error of model j to observed data

1
𝑝(𝑀𝑗 )
𝑅𝑀𝑆𝐸𝑗 2
𝑤𝑒𝑖𝑔ℎ𝑡𝑠, 𝑤𝑗 ∝ 𝑝(𝑀𝑗 |𝐷) = 1 (2.33)
∑𝑗 𝑝(𝑀𝑗 )
𝑅𝑀𝑆𝐸𝑗 2

Finally, the final output response from multiple models can be derived from

weighted sum of individual responses from all models as:

𝑛𝑜. 𝑜𝑓 𝑚𝑜𝑑𝑒𝑙𝑠
𝑅𝑒𝑠𝑝𝑜𝑛𝑠𝑒 = ∑𝑗=1 𝑤𝑗 𝑅𝑒𝑠𝑝𝑜𝑛𝑠𝑒𝑗 (2.34)

2.2.4 Relative Influence of Predictor Variables

Relative influence of a predictor variable is calculated as the relative change in the

RMSE (Root Mean Squared Error), AAE (Average Absolute Error) or R2 (Coefficient of

33
Determination) if a given predictor is removed from the training data set and rest of the

steps remain unchanged during model training process.

Eq. 2.35 shows the formula to calculate relative influence of pth predictor using

R2. From Eq. 2.35, it can be seen that relative influence of a predictor variable is its

proportion of variance that is predictable from a model. Relative Influence of a pth

predictor is given by:

𝑅 2 𝑝 −𝑅 2 −𝑝
𝑅𝐼𝑝 = 𝑎𝑏𝑠 ( ) (2.35)
𝑅2 𝑝

where,

𝑅 2 𝑝 = 𝑅 2 of model with all predictors included

𝑅 2 −𝑝 = 𝑅 2 of model with all predictors except pth predictor are included

Eq. 2.35 can be applied to other two metrics – RMSE and AAE by replacing R2

by RMSE and AAE respectively. Eqs. 2.36 and 2.37 shows formulas to calculate RMSE

and AAE.

1
𝑅𝑀𝑆𝐸 = √𝑛 ∑𝑛𝑖=1(𝑦𝑖 − 𝑦̂𝑖 )2 (2.36)

1
𝐴𝐴𝐸 = 𝑛 ∑𝑛𝑖=1|𝑦𝑖 − 𝑦̂𝑖 | (2.37)

Eq. 2.38 (Schuetter et. al, 2015) shows how 𝑝𝑠𝑒𝑢𝑑𝑜 𝑅 2 can be calculated. This

version of 𝑅 2 indicates the proportion of variance in the response/dependent variable that

is predictable from a model.

∑𝑛 (𝑦̂ −𝑦̅)2 ∑𝑛 ̂ 𝑖 )2
𝑖=1(𝑦𝑖 −𝑦
𝑝𝑠𝑒𝑢𝑑𝑜 𝑅 2 = ∑𝑖=1 𝑖
𝑛 (𝑦 −𝑦
̅)2
=1− ∑𝑛 ̅)2
(2.38)
𝑖=1 𝑖 𝑖=1(𝑦𝑖 −𝑦

34
where,

𝑦𝑖 = observed value of 𝑖 𝑡ℎ data point

𝑦̂𝑖 = predicted value of 𝑖 𝑡ℎ data point

𝑦̅ = mean of observed values

Another metric that can be utilized here is normalized mean-standard deviation

ratio (Eq. 2.39). Instead of R2, Median to Sigma ratio is utilized to create relative influence

plots. However, this ratio has been normalized w.r.t corresponding ratio in observed

data/actual data as in the case of R2.

𝑀𝑒𝑑𝑖𝑎𝑛
( )
𝜎 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑
𝑁𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑀𝑒𝑑𝑖𝑎𝑛 − 𝑆𝑖𝑔𝑚𝑎 𝑅𝑎𝑡𝑖𝑜 = 𝑀𝑒𝑑𝑖𝑎𝑛 (2.39)
( )
𝜎 𝑂𝑏𝑠𝑒𝑟𝑣𝑒𝑑

In this study, relative influence of a predictor variable is calculated by first

calculating the quantity for model evaluation - RMSE, AAE or R2 - including all the

predictor variables in training data set (𝑅 2 𝑝 ) and then calculating it without including the

predictor p in the training data set (𝑅 2 −𝑝 ). Finally, using Eq. 2.35 will give relative

influence of that predictor.

2.3 Eagle Ford Field Case Study

The Eagle Ford data is collected for about multiple wells from the commercial

database Drillinginfo (https://info.drillinginfo.com/). The raw data is cleaned to remove

outliers. Only the wells satisfying following criteria (about 100 wells) were used:

 Well Production Period > 12 months

 Initial flow rates < 40,000 STB/month

35
  STAGES > 4

 50,000 bbl < Total Fracturing fluid < 200,000 bbl

 CLENGTH > 2000 ft

 Calculated EUR <= 300 MSTB

 Wells with too much noise in rate decline data.

Fig. 2.9 shows the pairwise scatter plots for various predictor variable data collected.

It may be observed that a few pairs of variables shown in this figure have some correlation

between them. For e.g., completed length, stages and total proppant amount seem to have

some correlation among them. However, this study uses all these predictor variables in

order to see the individual effects on regression and variable relative importance study.

The EUR value for each of the wells is calculated based on decline curve extrapolation

to 30 years of production. Each of the four decline models would result in a different EUR

for a given well. As an exploratory analysis, these EURs can be regressed by a regression

tree to identify variables making more impact than others on EUR. Figs. 2.10 through 2.13

show these regression trees. As is obvious from these figures, Initial Flow Rate, qi, is

clearly making the most impact on EUR among all decline models. Another way of doing

this analysis is dividing the EUR range in Eagle Ford data into four groups or clusters

based on quartiles. Cluster 1 contains wells with lowest EURs while cluster 4 contains the

highest values of EURs. Figs. 2.14 through 2.17 show results from the classification tree

analysis for each of the decline models. Again, qi comes out to be the most important

variable.

36
Figure 2.9 Pairwise scatterplots of various predictor variables in Eagle Ford data

37
Figure 2.10 Regression Tree fitted on EUR calculated from Arp’s Decline Model

Figure 2.11 Regression Tree fitted on EUR calculated from SEDM Decline Model

38
Figure 2.12 Regression Tree fitted on EUR calculated from Duong’s Decline Model

Figure 2.13 Regression Tree fitted on EUR calculated from Weibull’s Decline Model

39
Figure 2.14 Classification Tree fitted on EUR clusters derived from Arp’s Decline Model

Figure 2.15 Classification Tree fitted on EUR clusters derived from SEDM Decline Model

40
Figure 2.16 Classification Tree fitted on EUR clusters derived from Duong’s Decline Model

Figure 2.17 Classification Tree fitted on EUR clusters derived from Weibull’s Decline

Model

41
 Based on previous results, qi has been identified to be the best candidate for

clustering the well data for further analysis. As mentioned earlier, Fig. 2.18 shows the 4

clusters created by dividing wells into four groups based on their Initial Flow Rates, q i.

Fig. 2.19 shows the distribution of other predictors in these 4 clusters. It may be observed

that cluster 4 which contains wells with highest Initial Flow Rates (qi) also contains wells

with highest Total Vertical Depths (TVD_HEEL) and Completed Lengths (CLENGTH)

if median values of these boxplots are taken as the reference.

Figure 2.18 Well clusters based on Initial Flow Rate, qi

42
 Figure 2.19 Predictor variable distribution in clusters derived from Initial Flow Rate, qi

Fig. 2.20 shows the location of the four clusters created based on Initial Flow Rate

on the Texas map. Fig. 2.21 shows wells in worst cluster 1 and best cluster 4 on map. Also

shown in this figure is the spread of other study variables on the map. Only clusters 1 and

43
4 are included in these plots to view the difference between the highest Initial Flow Rate

wells and Lowest Initial Flow Rate wells. It may be observed from these figures that most

of the wells occurring in cluster number 4 are drilled in deepest depths. However, there

are some exceptions to this observations shown in the map. This is because TVD is not be

the only criteria to predict well production. However, only TVD_HEEL has some

reasonable trend on the map.

Figure 2.20 Study wells on Texas map color coded by cluster number

44
Figure 2.21 Correlation between cluster type and different variables

45
 Fig. 2.22, 2.25, 2.28 and 2.31 show the comparison plots of different error metrics

resulting from best fit of data using the 12 machine learning algorithms applied for this

study. Best machine learning algorithm for each decline model is identified as the one

which has lowest RMSE errors but R2 to be close to unity. Table 2.3 shows the best

machine learning algorithms determined for each of the decline models.

Table 2.3 Most suitable Machine Learning algorithm for each decline model

Decline Model Best Machine Learning Algorithm

Arp’s GBM

SEDM SVM

Duong GBM

Weibull SVM

Figs. 2.23, 2.26, 2.29 and 2.32 show the scatterplots showing predicted versus

actual values of a decline curve parameter/EUR for RF, GBM, SVM and MARS

algorithms. Figs. 2.24, 2.27, 2.30 and 2.33 show the predicted decline curves for test data

wells for each of the decline models applying the best machine learning algorithm. Fig.

2.34 shows the comparison plots of predictions made in Figs. 2.24, 2.27, 2.30 and 2.33.

Since each of the four decline models under investigation have a different set of

decline model parameters, comparing them together is not possible. However, if we

compare EURs for these decline models together, it may be easier to identify the best

combination of decline model and machine learning algorithm to predict well performance

46
in Eagle Ford wells. Fig. 2.35 shows such comparison between EURs predicted from the

four decline models. It may be recalled here that EURs are estimated based on

extrapolation of a decline curve for 30 year period. Therefore Actual EURs mentioned in

these figures are calculated by extrapolating best fit decline curves using actual rate data.

This means that a well can have a different EUR for each of the four decline models for

the same well rate data. From Fig. 2.35 it may be seen that SEDM and Weibull have better

prediction results compared to other two decline models. It may also be noted that Arps

and Duong’s models are predicting higher range of EUR for the wells compared to SEDM

and Weibull models. This may be the likely reason for inaccurate prediction of EUR at

higher values in case of Arp’s and Duong’s models. It should also be recalled here that

Weibull model would require an initial estimate of the carrying capacity to fit decline

model curve on a well data. This is however not required in case of SEDM model.

Therefore, this may be regarded as an advantage of SEDM model over Weibull model.

Figure 2.22 Error metric comparison for different machine learning algorithms taken into

consideration for Arp’s model

47
Figure 2.23 Scatterplots showing predicted vs actual values of Arp’s decline model

parameters and EUR

48
Figure 2.24 Prediction of Arp’s decline curves using GBM

49
Figure 2.25 Error metric comparison for different machine learning algorithms taken into

consideration for SEDM model

Figure 2.26 Scatterplots showing predicted vs actual values of SEDM decline model

parameters and EUR

50
Figure 2.27 Prediction of SEDM decline curves using SVM

51
Figure 2.28 Error metric comparison for different machine learning algorithms taken into

consideration for Duong’s model

Figure 2.29 Scatterplots showing predicted vs actual values of Duong’s decline model

parameters and EUR

52
Figure 2.30 Prediction of Duong’s decline curves using GBM

53
Figure 2.31 Error metric comparison for different machine learning algorithms taken into

consideration for Weibull model

54
Figure 2.32 Scatterplots showing predicted vs actual values of Weibull’s decline model

parameters and EUR

55
Figure 2.33 Prediction of Weibull’s decline curves using SVM

56
Figure 2.34 Comparison of predictions made by ARP’S - GBM, SEDM - SVM, DUONG –

GBM and WEIBULL - SVM

57
Figure 2.35 EUR prediction comparison among best candidates for each decline model

58
 Fig. 2.36 shows the distribution of variable rankings based on RMSE errors. As

described previously, variable rank is calculated based on relative change in test data error

metric if the predictor variable is removed from machine learning model. Fig. 2.36 shows

variable ranking based on change in RMSE metric. A predictor variable can have a

different rank in different decline model – machine learning combination. This relative

influence/ranking plots are generated considering 4 decline models (Arp’s, SEDM, Duong

and Weibull) and 10 machine learning algorithms (RF, SVM, GBM, MARS, ANN, KNN,

LM, RIDGE, LASSO and ENET) not including ACE and AVAS due to instability issues.

Therefore, each predictor variable has 40 possible rank values across all these

combinations. Fig. 2.37 shows frequency histograms of predictor variable rank

distributions and Fig. 2.38 shows the Average Rank versus Rank Variance corresponding

to each of the predictor variable. A variable with rank close to unity and with low rank

variance is considered to be more important that others. As can be observed from these

figures, initial flow rate, qi, is ranked at the top in all cases.

Figs. 2.39 to 2.41 show similar analysis as describe above based AAE metric and

Figs. 2.42 to 2.44 show the analysis based on R2 metric. Figs. 2.45 to 2.47 show the

analysis based on Median-Sigma ratio based metric. As can be observed here different

error metric can provide different variable ranking analysis plots. However, it may be

observed here that initial flow rate is always highly ranked among all cases. Also, since

TVD has been obseved to be a critical predictor during exploratory analysis conducted

previously and since R2 metric gives TVD high importance after initial flow rate, it may

59
be logical here to assume R2 base variable importance plots to be more accurate compared

to other metrics.

TVD_HEEL_TOE_DIFF

TVD_HEEL

STAGES

qi

PROP_TOTAL

LONGITUDE

LATITUDE

FRAC_FLUID_TOTAL

CLENGTH

Figure 2.36 RMSE based variable ranking distribution

60
Figure 2.37 RMSE based variable ranking frequency distribution

Figure 2.38 RMSE based variable average rank vs rank variance

61
TVD_HEEL_TOE_DIFF

TVD_HEEL

STAGES

qi

PROP_TOTAL

LONGITUDE

LATITUDE

FRAC_FLUID_TOTAL

CLENGTH

Figure 2.39 AAE based Variable Ranking distribution

62
Figure 2.40 AAE based variable ranking frequency distribution

Figure 2.41 AAE based variable average rank vs rank variance

63
TVD_HEEL_TOE_DIFF

TVD_HEEL

STAGES

qi

PROP_TOTAL

LONGITUDE

LATITUDE

FRAC_FLUID_TOTAL

CLENGTH

Figure 2.42 R2 based variable ranking distribution

64
Figure 2.43 R2 based variable ranking frequency distribution

Figure 2.44 R2 based variable average rank vs rank variance

65
TVD_HEEL_TOE_DIFF

TVD_HEEL

STAGES

qi

PROP_TOTAL

LONGITUDE

LATITUDE

FRAC_FLUID_TOTAL

CLENGTH

Figure 2.45 Median-Sigma ratio based variable ranking distribution

66
Figure 2.46 Median-Sigma ratio based variable ranking frequency distribution

Figure 2.47 Median-Sigma ratio based variable average rank vs rank variance

67
 2.4 Summary

1. Rate decline model parameters for Arps, SEDM, Duong and Weibull decline models

can be linked to well completion and location variables using Machine Learning.

2. Rate decline curves are predicted for each of the four decline models and compared

with observed data of test wells.

3. Most suitable Machine Learning algorithms for predicting decline curve parameters

for each of decline models have been identified in this study.

4. SEDM with SVM is found to be the most suitable combination to predict EUR.

5. Relative Variable Importance study shows that initial flow rate to be most influential

predictor followed by total vertical depth.

68
 CHAPTER III

HYDRAULIC FRACTURE DESIGN AND OPTIMIZATION IN

UNCONVENTIONAL SINGLE PHASE GAS RESERVOIR USING GENETIC

ALGORITHM*

3.1 Introduction and Literature Review

In USA, shale oil and gas production has been on the rise particularly during the

last decade. However, due to very low permeability in these reservoirs, hydraulic

fracturing becomes an essential requirement for economical production. These hydraulic

fractures are created after pumping large amount of fracturing fluid and proppant to

support fractures thus created. Once created, this process increases conductivity and

surface area for fluid flow in the reservoir which increases the well production. Well

production can be increased by increasing the number of hydraulic fractures. However, it

may not be economical to increase investments in this process beyond a certain point. This

study focuses on getting close to this ‘most economical’ point by applying a class of

evolutionary algorithm know as genetic algorithm in a synthetic unconventional reservoir.

This chapter will use this reservoir case to optimize various parameters associated with

hydraulic fracturing design.

*
Parts of the text and data reported in this chapter is reprinted with permission from Yang, C., Vyas, A.,
Datta-Gupta, A., Ley, S.B. and Biswas, P., 2017. Rapid multistage hydraulic fracture design and
optimization in unconventional reservoirs using a novel Fast Marching Method. Journal of Petroleum
Science and Engineering. Copyright 2017 Elsevier

69
 Holditch (1992) reported that there is plenty of oil and gas reserves as long as it is

possible to exploit them economically. It was also reported that horizontal wells with

multiple hydraulic fractures using waterfrac technology is key for hydrocarbon production

from shales. It was also reported that going forward the biggest technological benefits will

be found in cost cutting improvements.

Saldungaray et al. (2013) emphasized the role of fracture conductivity on well

productivity. Fracture conductivity is dependent on the type of proppant/fracturing fluid

used and type of technique used for fracturing job. It was also reported that the number of

hydraulic fractures and spacing between them are dependent on rock fabric and formation

permeability. The three parameters – the rock fabric, natural fracture distribution and the

reservoir permeability – are noted as most important while optimizing the number of

hydraulic fractures used in a well.

Rankin et al. (2010) noted that since transverse fractures in horizontal wells

provide small intersection area, multiple stages with higher conductivity proppants are

needed to improve the flow capacity of the connection between fractures and wellbore.

Superior productivity is reported using more than 10 hydraulic fractures in the Bakken

study area reported.

Morales et al. (2010) presented a modified genetic algorithm to optimize well

placement in a reservoir. It was identified that in a complex heterogeneous reservoir,

optimum location of a well based on intuition is difficult to achieve.

Kennedy et al. (2012) presented well placement optimization process and

identified required combination of petrophysical, geochemical, and geomechanical

70
properties of a reservoir. It was reported that resource development simply based on

uniformly spaced hydraulic fractures may not be ideal for a heterogeneous reservoir. It

was reported that a naturally fractured reservoir can be drained better if a complex network

of fractures can be created during hydraulic fracturing process. However, in order for

optimization of well placement and hydraulic fracture design, a good amount of

knowledge about reservoir is needed. This study reports that tools such as include

electrical resistivity imaging LWD logs can be utilized in order to maximize the

knowledge about a reservoir. Also, techniques such as micro seismic monitoring can be

used to determine the details of hydraulic fractures created after fracking. A high definition

resistivity log can be used to identify natural fractures, induced fractures (from nearby

offset wells), faults and bedding planes.

Helgesen et al. (2005) presented a novel resistivity tool for accurate wellbore

placement. This tool is reported to have depth of investigation nearly 5 times the

conventional multiple propagation resistivity tools.

Biswas and Ley (2015) introduced a novel approach for natural fracture

interpretation using log data. This paper makes use of compressional waveforms instead

of shear waveforms allowing faster and accurate determination of natural fractures. At

least 4 (one in each sector) raw waveforms are used in this method as an input out of which

the first one is muted in time domain and filtered in frequency domain. This process is

repeated in each sector and RMS energy is calculated. A modified stacking algorithm is

used to amplify the finer perturbations in the data and to stabilize the waveforms. Since

compressional waveform data is not collected by every cross-dipole/wireline tool, this

71
paper suggests to make use of first arrival waveforms or “leaky mode waveforms” since

these waves have compressional velocities.

Sierra et al. (2013) concluded from their paper that reservoir permeability is the

main driver during decision making regarding hydraulic fracture spacing along horizontal

well. It was also concluded that fracture complexity is important only in reservoirs having

permeability lower than 100 nd. In reservoirs having permeability more than that,

optimally placed planar fractures should be sufficient to maximize gas recovery factor.

Also, proppant settling effects which are frequently observed in waterfracs, influence the

fracture spacing. It was also concluded that in case of stress dependent permeability and/or

porosity, smaller fracture spacing should be used. However, if the hydraulic fractures are

not properly propped, smaller fracture spacing cannot compensate. It was concluded that

knowledge of stress dependency of reservoir permeability and porosity is needed in

deciding fracture spacing. Also, type of proppant used can alter the fracture conductivity

and therefore put an effect on optimal hydraulic fracture spacing.

Ma et al. (2013) reported their hydraulic fracture placement optimization results.

This study uses both derivative free genetic algorithm based optimization and finite

difference based optimization of NPV. However, this study is not optimizing the fracture

half-length and proppant/fracturing fluid amount. It was found in their study that in a

heterogeneous reservoir, fracture spacing in high permeability region is lower than in low

permeability region. Also, in case of the finite difference based optimization method, the

optimum model showed near uniform spacing in low permeability region of the reservoir.

72
 Yang et al. (2012) reported a hydraulic fracture optimization method using a

pseudo 3D hydraulic fracturing model for a multilayered formation. Their approach

integrated Linear Elastic Fracture Mechanics (LEFM), Unified Fracture Design (UFD)

and 2D PKN model. This paper presented an algorithm that can help in determining what

treating pressure and other treatment parameters are needed to achieve optimum

placement of a given amount of proppant of specified quality. This method also informs

about the layers which act as containment barriers for vertical fracture propagation at a

specified treating pressure level.

Pitakbunkate et al. (2011) reported that fracture optimization based on Unified

Fracture Design (UFD) results in optimum fracture geometry. It was also reported that

fracture height growth depends on inter layer stress differential and not on individual stress

values. In low permeability reservoirs, large fracture height is accompanies by larger

fracture half lengths. It was also reported that there is a need to study fracture height

migration to prevent fracture migrations into water zones.

Warpinski et al. (1998 and 2005) reported how hydraulic fracture growth and

geometry can be detected using microseismic data. During hydraulic fracturing treatment,

changes in pore pressure affect planes of weakness (natural fractures and bedding planes)

adjacent to the hydraulic fracture and allow them to undergo shear slippage. These shear

slippages are like small earthquakes (and hence called “microseisms” or micro

earthquakes). These microseisms emit elastic wave signals that can be detected by

transducers located for analysis.

73
 Maxwell et al. (2002) concluded from Barnett shale studies that real time

microseismic images can be utilized to fracture geometry in currently uneconomic regions

of Barnett shale in order to make them economic. Fisher et al. (2005) also reported results

from Barnett shale. The paper reports that there can be three types of fractures – simple,

complex and very complex. In a shale reservoir having a presence of natural fractures, the

fracture complex is more likely to be “very complex”. A very complex network of

fractures allows a fracture fairway to be created with many fractures in multiple

orientations resulting in large contact area between well and reservoir. This paper reported

various technologies available that can be utilized to gather information regarding fracture

parameters such as height, length and azimuth. These technologies include Surface

Tiltmapping, Downhole Tiltmapping and Microseismic Mapping. This paper reports that

in Barnett shale wells, it is the cumulative fracture-network length (combining both

hydraulic fractures and natural fractures) that controls the reservoir connectivity and not

the conventional fracture half lengths. The paper reports ways to estimate fracture growth

by history matching recorded fracture data.

Cipolla et al. (2009) used dual permeability based reservoir model to simulate

creation of Stimulated Reservoir Volume (SRV). The paper concludes that with the

availability of reservoir geologic data such as core data and microseismic data can be used

to history match the simulated data. Gas recovery can be increased by increasing the

complexity of fracture network. The paper also reports that in low Young’s modulus

formations, effect of stress dependent network fracture conductivity becomes dominant

resulting in lower recovery. This effect is usually observed after 1-2 years of production.

74
 Savitski et al. (2013) reported from their studies that even though the aperture of a

hydraulic fracture is greater than natural fractures, the total area of activated (pressurized)

natural fractures can be significant which makes them relevant to production. Another

conclusion made by this study is that DFN connectivity does not cause a characteristic

response that would allow one to determine DFN connectivity from stimulation data. It

was also concluded that stress perturbation is not sufficient to stimulated non-conductive

natural fractures and that initial natural fracture conductivity is critically important. It was

also concluded from their study that lower injection rate will results in larger stimulated

reservoir volume in the presence of conductive natural fracture, though it will also result

in hydraulic fractures of lower width that may be susceptible to premature screen-out.

Riahi and Damjanac (2013) conducted numerical simulations to study interaction

between hydraulic fractures and natural fractures. This study concluded that for a given

injected volume, lower injection rates result in greater proportion of DFN being affected

during hydraulic fracturing propagation. It was also concluded that DFN properties such

as density, length distribution and fracture orientation are critical to the overall response

of the formation during hydraulic fracturing.

Dershowitz et al. (2000) integrated DFN methods with conventional dual porosity

reservoir simulators. It was reported that permeability of the natural fracture system

depends on the fracture intensity, the connectivity of the natural fracture system and the

distribution of the natural fracture transmissivities. This study made use of the tensor

approach of Oda (1985). Using this approach, equivalent permeability of each grid block

containing natural fractures can be generated and then further simulations can be carried

75
out. However, Oda method is suitable only in well-connected natural fractures only since

it does not take fracture connectivity into account.

Various authors have reported their methods for long term reservoir performance

forecasting. Arps (1945), Fetkovich (1980) and Valko and Lee (2010) proposed decline

curve based production predictions. Ilk et al. (2010) and Song and Ehlig-Economides

(2011) proposed their methods for reserve estimation and production forecast using

pressure/rate transient analysis. These analytic methods are fast but not as accurate as

numerical simulator available commercially due to their inadequacies to incorporate

complex heterogeneities in field. Fan et al. (2010) used a numerical simulator to predict

shale gas production in Haynesville shale. Shale gas log data is used to gather information

about reservoir porosity, permeability, TOC, saturations, etc. History matching the early

production data is then done to calibrate the reservoir properties. Microseismic data can

give idea of fractures created during hydraulic fracturing process. It was reported in this

paper that difference in stress contrast can lead to different complexities of fracture

network created during hydraulic fracturing treatment. This study shows two types of

complexities due to difference in stress anisotropies. Other factors affecting fracture

network include rock fabric, preexisting natural fractures and layering. Once a model is

calibrated using available production data, microseismic data, core data, etc., a reasonable

forecast can be made for future.

The use of commercial reservoir simulators can give a very accurate production

forecast but this method is costly and time consuming process. Lee (1982) proposed the

concept of radius of investigation in homogeneous reservoirs. It is defined as the

76
propagation distance of a “peak pressure” disturbance for an impulse source or sink (Lee

1982). Datta-Gupta et al. (2011) extended this concept to heterogeneous reservoirs with

arbitrary well conditions and the diffusive equation then turns out to be the Eikonal

equation which can be solved very efficiently by a class of front tracking methods known

as Fast Marching Methods (FMM) presented earlier by Sethian (1996 and 1999).

Sehbi et al. (2011) used the concept of drainage volume for optimizing hydraulic

fracture stages in Tight Gas Reservoirs. Their study used a high frequency asymptotic

solution of the diffusivity equation to generalize the concept of radius of drainage (Lee,

1982) to horizontal wells. In this study done in cotton valley formation well, ten hydraulic

fractures with 500 ft of half-length came out to be most optimum. Increasing number of

stages beyond that would yield diminishing returns. Besides application in optimization

problem, drainage volume calculations gave an additional advantage of flow visualization

with no additional simulations.

Xie et al. (2015a) revisited FMM and proposed a geometric pressure solution based

on depth of investigation to estimate transient pressure behavior in unconventional wells

with multistage hydraulic fractures. Well diagnostic plot was generated from pressure

depletion behavior that could be used to identify various flow regimes. The advantage of

using this technique is that transient pressure response for a multimillion grid cell based

reservoir model can be obtained within in seconds. Xie et al. (2015b) integrated shale gas

production data and microseismic data using FMM to obtain reservoir and hydraulic

fracture properties. Fracture parameters such as fracture half lengths and fracture

permeability and reservoir parameters such as matrix permeability and SRV permeability

77
were determined using a history matching process based on Genetic Algorithm (GA).

Since FMM combined with geometric approximation is computationally very efficient

compared to commercially available forward simulators, this history matching problem

could be completed very fast.

Zhang et al. (2013) extended the concept of FMM based reservoir simulation to

complex flow geometry and anisotropic properties. This study derived the FMM

formulation in corner point grids. Zhang et al. (2014 and 2016) derived a new formulation

of the diffusivity equation using diffusive time of flight as a spatial variable transforming

three dimensional simulation problem to a one dimensional one. The diffusive time of

flight (DTOF) embeds the information regarding reservoir heterogeneity. A one

dimensional problem is then solved using finite difference method rapidly.

Fujita et al. (2016) extended the DTOF formulation to triple-continuum modeling

for modeling shale gas reservoirs. Physical mechanisms like Knudsen diffusion and

slippage effects, adsorption/diffusion in nanopore surfaces, rock compaction in fractures

due to geomechanical effects and gas diffusion due to Kerogen content we included in the

FMM based unconventional shale gas simulator.

3.2 Methodology

3.2.1 Fast Marching Method

This study uses a dual porosity unconventional shale gas model for optimizing

hydraulic fracture design. The forward model to calculate gas rate production is based on

a Fast Marching Method based reservoir simulator (Zhang et. al, 2014 and 2016). A short

78
description of this method with various equations is provided in this part of dissertation.

However, a more detailed explanation can be found in the reference provided in this

section.

Description of FMM method starts with the concept of radius of investigation

proposed by Lee (1982). Radius of investigation can be defined radius of investigation in

homogeneous reservoirs as the propagation distance of a “peak pressure” disturbance for

an impulse source or sink (Lee 1982). Datta-Gupta et al. (2011) extended this concept to

heterogeneous unconventional reservoirs with horizontal wells with multistage hydraulic

fracturing. Propagation equation of peak pressure front can be derived by using asymptotic

ray theory widely used in electromagnetic and seismic wave propagation (Virieux et. al,

1994). Vasco et al. (2000), Kulkarni et al. (2000) and Datta-Gupta and King (2007) used

a high frequency asymptotic solution of the diffusivity equation to derive Eikonal equation

(Eq. 3.2) for propagating pressure front for impulse source. The general diffusivity

equation is given by:

𝑘 𝜕𝑝
⃗⃗ 𝑝) = 𝜙𝑐𝑡
∇. (𝜇 ∇ (3.1)
𝜕𝑡

The Eikonal equation is given by:

√𝛼|∇𝜏(𝑥⃗)| = 1 (3.2)

where,

𝜏 = diffusive time of flight (DTOF) or the propagation time of the pressure front

𝛼 = diffusivity = 𝑘/(𝜑𝜇𝑐𝑡 )

79
𝑘 = permeability

𝜙 = porosity

𝜇 = fluid viscosity

𝑐𝑡 = total compressibility

The diffusive time of flight, DTOF, has a unit of square root of time and shows

that pressure front propagates in the reservoir with a velocity given by the square root of

diffusivity (Datta-Gupta et. al, 2011). It is dependent on reservoir properties but

independent of flow rate (Datta-Gupta et. al, 2011). Eq. 3.2 can be solved by a class of

front tracking algorithm known as Fast Marching Method or FMM (Sethian, 1996 and

1999; Zhang et al., 2013, Xie et al., 2015a, 2015b). Using FMM, diffusive time of flight

can be calculated for each grid block of a reservoir model. In a homogeneous reservoir,

the contours of τ are related to the propagation time t of the pressure front through the

following equation (Vasco et al., 2000; Kim et al., 2009):

𝜏 = √𝛽𝑡 (3.3)

where, 𝛽 is 2, 4, and 6 for 1D linear, 2D radial, and 3D spherical flow patterns

respectively. Due to irregular flow pattern, above values of 𝛽 cannot be applied in

heterogeneous reservoirs. However, diffusive time of flight can still help in visualizing

pressure front in heterogeneous reservoirs.

The next step is to calculate well production rates based using diffusive time of

flight. Once diffusive time of flight values for each grid block in reservoir model is

calculated, different diffusive time of flight contours can be generated. The drainage pore

80
volume, 𝑉𝑝 , inside a contour can be calculated by approximating it with the total drainage

volume at cut-off. Therefore, FMM solver can generate the drainage pore volume as a

function of the diffusive time of flight, 𝑉𝑝 (𝜏). Zhang et al (2014 and 2016) derived a new

formulation of the diffusivity equation using τ as a spatial variable. Instead of writing

equation in physical coordinates, this paper presented a new equation in terms of diffusive

time of flight (Zhang et al, 2014 and 2016):

1 𝜕 𝜕𝑝 𝜕𝑝
(𝑤(𝜏) 𝜕𝜏 ) = (3.4)
𝑤(𝜏) 𝜕𝜏 𝜕𝑡

where,
𝑑𝑉𝑝 (𝜏)
𝑤(𝜏) = (3.5)
𝑑𝜏

𝑤(𝜏) gives the propagating speed of drainage surface.

Zhang et al (2016) showed the analogy between the diffusivity equation in radial

coordinate and in τ coordinate. Therefore, solving the 1-D equation in 𝜏 coordinate will

generate pressure w.r.t time. Here, 𝜏 is embedding all the heterogeneities in the reservoir.

In case of dual porosity reservoir model, fluid flow occurs only between fracture to

fracture or between matrix to fracture. Fluid flow within matrix is negligible and can be

ignored.

In a dual porosity model, Eqs. 3.6 and 3.7 are solved separately to model fluid

flow. Mass balance equation in fracture-fracture flow (Yang et. al, 2017):

𝜕(𝜌𝜙𝑓 ) 𝜌 𝑘
− ∇. (𝜇 𝑘𝑓 ∇𝑝𝑓 ) = −𝜌𝑢𝑝 𝜎 𝜇 𝑚 (𝑝𝑓 − 𝑝𝑚 ) (3.7)
𝜕𝑡 𝑢𝑝

81
Mass balance in matrix-fracture flow (Yang et. al, 2017):

𝜕(𝜌𝜙𝑚 ) 𝑘
= 𝜌𝑢𝑝 𝜎 𝜇 𝑚 (𝑝𝑓 − 𝑝𝑚 ) (3.8)
𝜕𝑡 𝑢𝑝

Since in the dual porosity model, FMM is used to solve pressure propagation in

fracture system only. The generated diffusive time of flight contours are then used to

calculate drainage pore volume. The mass balance fluid flow equations Eqs. 3.7 and 3.8

are transformed to 1-D 𝜏 coordinate. During this transformation, the mass balance

equation in matrix-fracture fluid flow keeps the same form as single porosity model but

the mass balance equation in fracture-fracture fluid flow takes the following form (Zhang

et al, 2014 and 2016):

𝑝𝑓 𝑐̃𝑡 𝜕𝑝𝑓 1 𝜕 𝑝 𝜕𝑝𝑓 1 𝑝

− 𝑤(𝜏) 𝜕𝜏 (𝑤(𝜏) 𝜇̃𝑍𝑓 ) = − 𝜙𝑐 (𝜇𝑍) 𝜎𝑘𝑚 (𝑝𝑓 − 𝑝𝑚 ) (3.9)
𝑍 𝜕𝑡 𝜕𝜏 𝑡𝑖 𝑢𝑝

where, 𝜇̃ and 𝑐̃𝑡 are dimensionless viscosity and total compressibility (Zhang et al., 2014

and 2016).

3.2.2 DFN Upscaling (Oda’s Method)

This study uses a synthetic unconventional dual porosity gas reservoir. This model

has been designed using a several clusters randomly distributed in the reservoir map. Two

extra ellipsoidal clusters of natural fractures are also put in model to create extra natural

fracture density (Fig. 3.1).

82
 1200
1000

X direction
800
600
400
200
0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Y direction

Figure 3.1 Natural Fracture distribution in the base model (Yang et al., 2017)

However, this model needs to be upscaled to corresponding permeability

distribution before simulating gas production using FMM based forward simulator. In

order to do that, Oda’s method (Oda, 1985) was utilized in this study because of its

simplicity and speed. Oda (1985) presented the following equation (Eq. 3.10) to calculate

permeability tensor for a dual permeability dual porosity reservoir model. Equation for

calculating permeability tensor in natural fractures is given by:

(𝑐)
𝑘𝑖𝑗 = 𝜆(𝑃𝑘𝑘 𝛿𝑖𝑗 − 𝑃𝑖𝑗 ) + 𝑎𝑖𝑗 (3.10)

where,

𝜆 = dimensionless constant (0< 𝜆 ≤1/12)

𝑎𝑖𝑗 = correction term

0, 𝑖𝑓 𝑖 ≠ 𝑗
𝛿𝑖𝑗 = Kronecker delta = = {
1, 𝑖𝑓 𝑖 = 𝑗

𝑃𝑘𝑘 = 𝑃11 + 𝑃22 + 𝑃33 = summation of three principal component of the crack tensor Pij

The crack tensor can be derived as,

𝜋𝜌 ∞ ∞
𝑃𝑖𝑗 = 4
∫0 ∫0 ∫Ω 𝑟 2 𝑡 3 𝑛𝑖 𝑛𝑗 𝐸(𝑛, 𝑟, 𝑡)𝑑Ω𝑑𝑟𝑑𝑡 (3.11)

83
where,

𝑟 = diameter of natural fracture

𝑡 = aperture of natural fractures

𝑛𝑖 , 𝑛𝑗 = the components of a unit normal to the fracture

𝐸(𝑛, 𝑟, 𝑡) = probability density function that describes the number of fractures whose unit

vectors n are oriented within a small solid angle 𝑑Ω

Ω = entire solid angle corresponding to the surface of a unit sphere

In a naturally fractured reservoir, each natural fracture has two opposing unit

normal vectors n(+) and n(-). Dershowitz et al. (2000) presented a simpler way of using

Oda’s equations. The total number of natural fractures in a grid cell, 𝑁 is given by:

𝑁 = ∫Ω 𝑛𝑖 𝑛𝑗 𝐸(𝑛)𝑑Ω (3.12)

Plains of permeability are given by,

1
𝑘𝑖𝑗 = 12 (𝐹𝑘𝑘 𝛿𝑖𝑗 − 𝐹𝑖𝑗 ) (3.13)

where,
1
𝐹𝑖𝑗 = fracture tensor = 𝑉 ∑𝑁
𝑘=1 𝐴𝑘 𝑇𝑘 𝑛𝑖𝑘 𝑛𝑗𝑘 (3.14)

𝑉 = grid cell volume

𝐴𝑘 = fracture area of kth natural fracture in a grid cell

𝑇𝑘 = transmissivity in kth natural fracture in a grid cell

𝑛𝑖𝑘 , 𝑛𝑗𝑘 = the components of a unit normal to the kth fracture

The fracture system porosity, 𝜙𝐹 , is given by:

84
 𝑉𝐹 ∑𝑁
𝑘=1 𝐴𝑘 .𝑒
𝜙𝐹 = 𝑉 = (3.15)
𝑐𝑒𝑙𝑙 𝑉𝑐𝑒𝑙𝑙

where,

𝑉𝐹 = fracture system volume

𝑉𝑐𝑒𝑙𝑙 = grid cell volume

𝑁 = number of fractures in a grid cell

𝐴𝑘 = fracture are of kth fracture

𝑒 = fracture storage aperture

3.2.3 Hydraulic Fracturing Design

The unconventional shale gas model that is used in this study has a non-uniform

permeability distribution due to non-uniform natural fracture density. The objective of this

chapter is to optimize the hydraulic fracture design parameters for this reservoir model

including location and number of hydraulic fractures, hydraulic fracture half lengths and

widths. Economides et al. (2002 and 2012) and Daal and Economides (2006) reported the

Unified Fracture Design algorithm to estimate the optimum hydraulic fracture dimensions

for a given amount of hydraulic fracture treatment variables such as proppant amount.

Propped volume in a single hydraulic fracture, 𝑉𝑝 is given by (Economides et al., 2002 and

2012):

𝑉𝑝 = 2𝑥𝑓 𝑤𝑓 ℎ𝑓 (3.16)

where,

𝑥𝑓 = fracture half-length

85
𝑤𝑓 = fracture average width respectively

ℎ𝑓 = fracture height

Mass of proppant used per stage, 𝑀𝑝 is given by (Economides et al., 2002 and 2012):

𝑀𝑝 = 𝑉𝑝 (1 − 𝜙𝑝 )𝜌𝑝 (3.17)

where,

𝑉𝑝 = propped volume per hydraulic fracture stage

𝜌𝑝 = proppant density

𝜙𝑝 = porosity of proppant fracture

For a given fracturing fluid injection flow rate and corresponding pumping time,

following equation can be derived keeping in consideration all the fluid losses occurring

during fracture propagation (Economides et al., 2002 and 2012):

𝑞𝑖 𝑡𝑒 − 𝜅(2ℎ𝑓 𝑥𝑓 )𝐶𝐿 √𝑡𝑒 − (2ℎ𝑓 𝑥𝑓 )𝑆𝑝 − 𝑥𝑓 𝑤𝑓 ℎ𝑓 = 0 (3.18)

where,

𝑞𝑖 = injection rate per half fracture of a bi-winged fracture

𝑡𝑒 = injection time

𝜅 = the opening time distribution factor

𝐶𝐿 = the fluid leak-off coefficient for the formation

𝑆𝑝 = spurt loss coefficient

The total proppant laden slurry volume per stage can be calculated as by (Economides et

al., 2002 and 2012):

86
𝑉𝑠𝑙𝑢𝑟𝑟𝑦 = 2𝑞𝑖 𝑡𝑒 (3.19)

Lastly, the total fracturing fluid volume per fracture stage can be calculated as by

(Economides et al., 2002 and 2012):

𝑀𝑝
𝑉𝑓𝑙𝑢𝑖𝑑 = 2𝑞𝑖 𝑡𝑒 − (3.20)
𝜌𝑝

3.2.4 Genetic Algorithm and Workflow

Since the main objective of this chapter is to optimize the hydraulic fracture design

parameters in a given reservoir model, an optimization algorithm is needed to accomplish

that. For this study, a class of evolutionary algorithms known as Genetic Algorithms

(Holland 1992 and Mitchell 1999) is utilized. A Genetic Algorithm or GA is a derivative

free optimization method based on natural selection process that mimics biological

evolution. In this algorithm, population members of current generation are evaluated for

their objective values and the population members of the next generation is reproduced

based on parents from previous generation taking into consideration their corresponding

objective values. Cheng et al. (2008) and Yin et al. (2010 and 2011) used GA to solve

optimization problems very efficiently. This study follows the same GA algorithm used

by Yin et al. (2010 and 2011). Fig. 3.2 shows the GA approach used by them. A set of

parameters are first identified with their minimum, maximum and base values. These

parameters are needed to be calibrated in order to optimize the objective function value.

Sensitivity analysis is first carried out for each of the parameters that need to be calibrated.

Some parameters can then be removed in case the model is not affected much by changing

their values. Next, an initial population of preset number of population member size is

then created using Latin Hypercube Sampling (LHS) based Design of Experiment (DOE).
87
This method takes into account the full coverage of parameter ranges provided. Each

initial population member is then used to update reservoir model used for this study and

FMM based forward simulator is used to generate production profile. The optimization

process in this study maximizes the Net Present Value (NPV) of the horizontal well with

multiple hydraulic fractures created through it into the reservoir. Therefore, after each

model simulation using FMM, NPV is calculated and stored as objective function value

for corresponding population member. The GA continues to update by creating new

population based on NPV (objective function value) values of previous generation. To

create a new generation, fittest members of the previous generation are used for crossover

or mutation so as to increase chances of creating better children. The fittest members are

chosen based on the corresponding NPV values. Newer generations evolve from previous

generations and try to reach optimum value after sufficient generations are reached or if

the maximum limit of number of generations are reached as set before optimization

process starts.

88
 Sensitivity Analysis
Design of Experiments (LHS)

Initialize/Update current
generation’s population

Run Simulation &

Evaluate Obj. Function
Select
Crossover
Mutate
Y
Stop Criteria?

N Stop

GA: accepted by fitness

Figure 3.2 General workflow for genetic algorithm (Yang et al., 2017)

Fig. 3.3 shows the steps to calculate NPV in detail. The parameters needed to be

optimized in this study are total number of hydraulic fracture, distances between hydraulic

fractures, fracture half-length and their widths. Each model in GA is updated using new

hydraulic fracture design parameters and corresponding permeability field is generated.

Amount of proppant and fracturing fluid required for creating this hydraulic fracturing

design can be calculated using Eqs. 3.16 to 3.20. Additional costs of equipment rent and

horizontal well drilling can be added to fracturing cost to get cost of entire well. The

revenue generated from well production can also be calculated based on gas prices and

cumulative gas production generated by FMM simulator. Net Present Value, NPV can

89
then be calculated as the difference between the revenue generated by the well and the

cost of well.

𝑅𝑒𝑣𝑒𝑛𝑢𝑒 = ∑𝑇𝑖=1 𝑅𝑒𝑣𝑒𝑛𝑢𝑒𝑖 (3.21)

𝑟 𝑡𝑖 /365
𝑅𝑒𝑣𝑒𝑛𝑢𝑒𝑖 = 𝑅𝑒𝑣𝑒𝑛𝑢𝑒𝑖−1 + (𝑃𝑖 − 𝑃𝑖−1 ) × 𝐺𝑎𝑠 𝑃𝑟𝑖𝑐𝑒 × (1 − 100) (3.22)

where,

𝑇 = total time of production

𝑃𝑖 = cumulative production at ith time step

𝑟 = interest rate

𝐶𝑜𝑠𝑡 = 𝐻𝑜𝑟𝑖𝑧𝑜𝑛𝑡𝑎𝑙 𝑊𝑒𝑙𝑙 𝑐𝑜𝑠𝑡 + 𝑃𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑜𝑛 𝑇𝑖𝑚𝑒 × 𝐸𝑞𝑢𝑖𝑝𝑚𝑒𝑛𝑡 𝑅𝑒𝑛𝑡 𝑐𝑜𝑠𝑡 +

𝑃𝑟𝑜𝑝. 𝐴𝑚𝑜𝑢𝑛𝑡 × 𝑃𝑟𝑜𝑝. 𝑃𝑟𝑖𝑐𝑒 + 𝐹𝑟𝑎𝑐. 𝐹𝑙𝑢𝑖𝑑 𝐴𝑚𝑜𝑢𝑛𝑡 × 𝐹𝑟𝑎𝑐. 𝐹𝑙𝑢𝑖𝑑 𝑃𝑟𝑖𝑐𝑒 (3.33)

𝑁𝑃𝑉 = 𝑅𝑒𝑣𝑒𝑛𝑢𝑒 − 𝐶𝑜𝑠𝑡 (3.34)

90
 Calculate Proppant No. of Hydraulic fractures Reservoir Model
schedule and Fracture spacing, half lengths and with DFN Network
Fracturing Fluid widths
Requirement

Generate Grid
Calculate Cost
Upscale Properties (Oda’s
Method)

Fast Marching Method

(Dual Porosity)

Gas Production and Revenue

Calculate NPV

Figure 3.3 Workflow of objective function evaluation for each model (Yang et al., 2017)

3.3 Results and Discussion

The first objective is to match the FMM prediction results with a commercial

simulator Eclipse for the reservoir model. Fig. 3.4 shows the upscaled permeability field

derived from Oda method. Since the optimum values of hydraulic fracture design

parameters are unknown, 15 hydraulic fractures with uniform spacing and half lengths are

assumed. Fig. 3.5 shows the comparison between the simulation results from FMM and

Eclipse simulators. It may be observed from this figure that FMM is predicting gas rate

very close to Eclipse results. However, the main advantage of FMM comes in terms of
91
time consumed for simulation. In this case, FMM was about 20 times faster that Eclipse

making it a more suitable candidate for this optimization study which requires large

number of simulations.

1200
1000
X direction

800
600
400
200
0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Y direction

(a)

(b)

Figure 3.4 (a) Natural fracture distribution (b) Upscaled reservoir permeability field (Yang

et al., 2017)

92
 7000

FMM
6000
Eclipse
Gas Rate (Mscf/d) 5000

4000

3000

2000

1000

0
0 4 8 12 16 20
Time (year)

Figure 3.5 FMM versus Eclipse simulated gas production for the base model (Yang et al.,

2017)

It should be noticed here that during application of the Oda’s method presented in

this study, a minimum matrix permeability is assumed to be approximately 10 nd. Fig. 3.6

shows how the cumulative gas production changes with perturbation of this assumed cut-

off value. Table 3.1 shows the variation in NPV due to changing this minimum matrix

permeability cut-off. Since the focus of this study is on the workflow for optimization and

not studying the effect of variation of this matrix permeability, this study assumes the base

value of 10 nd for this purpose.

93
 Figure 3.6 Effect of changing minimum matrix permeability during Oda’s upscaling

Table 3.1 NPV variation with minimum matrix permeability used

Minimum Matrix Permeability NPV

10 nd (Base Value) 8.06

10 x Base Value 8.87

0.1 x Base Value 7.07

2 x Base Value 8.33

0.5 x Base Value 7.84

Table 3.2 shows the economic parameters assumed in this study to calculate the

Net Present Value (NPV) for a given hydraulic fracturing design. NPV is calculated as the

94
difference between the cumulative revenue generated during a specified period of well

production time and the cost of well.

Table 3.2 Economic Parameters for NPV calculations

Properties Value

Proppant Cost (USD/ton) 550

Fracturing Fluid Cost (USD/gal) 0.4

Horizontal Well Cost (USD/well) 1.2×106

Equipment Rent Cost (USD/min) 550

Interest Rate (1/year) 10%

Gas Price (USD/Mscf) 3.6

Fig. 3.7 shows the effect of changing the number of uniformly spaced hydraulic

fractures on the well’s production. It may be observed from this figure that cumulative

production increases with increasing the number of hydraulic fracture stages in the

reservoir. Fig. 3.8 shows the effect of increasing number of hydraulic fracture stages on

NPV. As can be observed from this figure, NPV increases at the beginning but then

decreases with increasing fracture stages further. This is due to the fact that cumulative

production does not improve significantly after certain number of stages. However the

cost of fracturing increases due to larger amounts of proppant and fracturing fluid utilized

for fracking job. Therefore, NPV starts to decline after a certain number of stages. Since

a fracturing design problem such as this one involves more than one variables, there is a

95
need to come up with an optimization workflow which can provide best combination(s)

of these variables for maximizing NPV. This study takes advantages of genetic algorithm

to present such workflow.

(a) (b)

Figure 3.7 a) Gas Rates for various number of fracture stages b) Cumulative Gas

Production for different numbers of fracture stages

Stages = 18

Stages = 15

Stages = 10

Stages = 8

Stages = 5

Cost NPV

Figure 3.8 Cost and NPV comparison for various cases of number of fracture stages

96
 Table 3.3 presents variable ranges used in this optimization study. For e.g., the

number of stages can be between 8 and 18 including the boundaries. This range is decided

based on previous results that resulted in maximum NPV within this range. Fracture width

range is derived from the assumption that each hydraulic fracture is made up of a collection

of 6 cracks on either side of the well and the width of each crack is of the order of thrice

the diameter of a commonly used proppant.

Table 3.3 Hydraulic fracture optimization variable ranges

Variable Min Value Base Value Max Value

Stages No. 8 12 18

Average Width (ft) 0.02 0.05 0.08

Fracture half-length (XF1 to XF4) (ft) 150 350 550

Fracture Spacing (DIS2 to DIS25) (ft) 100 250 400

Fig. 3.9 shows the sensitivity analysis results for this study. Each variable is

perturbed to its maximum and minimum values as per the variable ranges presented

previously and corresponding fractional change in NPV was calculated compared to the

base NPV value (NPV resulting from keeping all variables at their base values). It may be

observed from this figure that NPV is most sensitive to the average width in the current

set of variable ranges. Although fracture spacing is not so sensitive in this figure, they can

be more dominant if more than one fracture to fracture spacing is changed during

97
optimization study. Therefore, current study has kept all the variables for optimization

process.

Stages
Average Width

Fracture Half-Lengths

Width

Fracture Spacing

Figure 3.9 Sensitivity analysis of various variables on NPV

Fig. 3.10 shows the results from genetic algorithm based optimization of NPV. As

explained in previous section of this chapter, genetic algorithm based optimization

consists of updating generations based on previous generations based on cross over and

mutation. As can be observed from this figure, subsequent generations tend to be better in

terms of objective function NPV. Fig. 3.11 shows variable distributions in the first

generation and the last generation. It may be observed that the first generation consists of

all possible values of this variable as provided in Table 3.3. However, as we move from

first generation to last generation, this variable ranges shrinks. This shows that this

algorithm is reaching an optimum set of variable values.

98
 Figure 3.10 NPV distribution in Genetic Algorithm based optimization approach

Figure 3.11 Distribution of fracture stages and average widths in generation 1 and

generation 25

99
 Fig. 3.12 shows the distribution of stage numbers in the first and the last

generations. As can be observed from this figure, two optimum number of stage numbers

are available in this problem – 13 and 18.

Figure 3.12 Distribution of fracture stages in generation 1 and generation 25

Figs. 3.13 and 3.14 show uniformly placed and optimally placed hydraulic fracture

designs. Optimum designs corresponding to both 13 and 18 number of stages are

compared in these figures. Comparing NPV values provided in Figs. 3.13 and 3.14 shows

that a reasonable improvement in NPV can be achieved by using the workflow utilized in

this study.

100
 Uniform Design Uniform Design

(NPV = $ 7.97 million) (NPV = $ 7.46 million)

(13 fracs) (18 fracs)

Figure 3.13 NPV from Uniform spaced fractures

101
 Optimum Design - 1 Optimum Design - 2

(NPV = $ 9.5 million) (NPV = $ 9.5 million)

(13 fracs) (18 fracs)

Figure 3.14 Hydraulic fracture placement in optimal design using genetic algorithm

Previous discussion assumed having a good knowledge about natural fracture

distribution in the reservoir model. However, if there is some uncertainty present in natural

fracture distribution, NPV based on multiple possible realizations can be chosen to be the

objective needed to be maximized. Fig. 3.15 shows possible realizations different from

original base model presented before. In this case NPV can be integrated using:
1
𝑁𝑃𝑉 = 𝑁 ∑𝑁
𝑖=1 𝑤𝑖 . 𝑁𝑃𝑉𝑖 (3.35)

where,

102
𝑁𝑃𝑉𝑖 = 𝑁𝑃𝑉 𝑜𝑓 𝑡ℎ𝑒 𝑖 𝑡ℎ 𝑟𝑒𝑎𝑙𝑖𝑧𝑎𝑡𝑖𝑜𝑛

𝑤𝑖 = 𝑤𝑒𝑖𝑔ℎ𝑡𝑠 𝑎𝑠𝑠𝑖𝑔𝑛𝑒𝑑 𝑡𝑜 𝑖 𝑡ℎ 𝑁𝑃𝑉

Fig. 3.16 shows the results from genetic algorithm based maximization of NPV

calculated using Eq. 3.35. For this study equal weights have been assigned to all reservoir

models. It can be observed here that genetic algorithm can successfully converge to a set

of models having low variance in NPVs compared to initial set of population. Fig 3.17

shows the variable distribution in the first and the last generations. It is clear from this

figure that variable ranges in the last generation has shrunk compared to the first

generation reducing uncertainty in those variables.

Fig. 3.18 shows the most optimum hydraulic fracture design based on multiple

realizations when applied to the true model/base model. It can be seen here that the NPV

has reduced from $ 9.5 million to 9.48 million when using the six realizations intead of

the actual model for optimization problem. This small loss of NPV shows robustness of

this algorithm using six realization. Table. 3.4 shows the variation in NPV if the true

model is one of the siz realizations or the base model presented earlier. It may be observed

from the numbers provided in this table that moderate uncertainty in true model can have

some effect on true NPV, i.e., it may be slightly higher or lower then the expected value.

This minor change is however, insignificant compared to the difference between the

optimum NPV and the NPV resulting from uniformly placed hydraulic fractures with base

variable values presented earlier.

103
Figure 3.15 Six possible realizations vs true model/base model in case of uncertainty in

natural fracture distribution

104
Figure 3.16 Results of genetic algorithm for multiple realization based optimization

Figure 3.17 Variable distribution in the first generation vs last generation

105
Figure 3.18 Hydraulic fracture placement in optimal design based on multiple realizations

Table 3.4 NPV values correponding to various realizations vs base model or true model

Realization NPV

Base Model 9.48

Realization 1 9.48

Realization 2 9.44

Realization 3 9.52

Realization 4 9.54

Realization 5 9.61

Realization 6 9.56

106
 3.4 Summary

1. For a given model with known natural fracture distribution, increasing number of

hydraulic fractures would increase cumulative production but the corresponding

cost of hydraulic fracturing would also increase. There is an optimum number of

hydraulic fractures for a given reservoir model.

2. Genetic Algorithm based hydraulic fracture optimization workflow presented in

this chapter can be utilized to maximize NPV by optimizing multiple hydraulic

fracture variables such as number of hydraulic fractures, widths of hydraulic

fractures, fracture half lengths and spacing between hydraulic fractures..

3. This chapter also presents how to deal with uncertainty in natural fracture

distribution and presents the modified workflow for such cases. Variance in NPV

due uncertainty in true model uncertainty has been presented for example case.

Moderate uncertainty in true model can lead to small variation in expected NPV.

4. FMM based simulator has been proven to be an accurate and faster alternative to

commercial simulator for an optimization study requiring large number of forward

simulations.

107
 CHAPTER IV

A MULTISTAGE GENETIC ALGORITHM FOR HISTORY MATCHING OF

SHALE OIL RESERVOIRS: FIELD CASE STUDY*

4.1 Background and Introduction

This chapter deals with application of FMM based reservoir simulator in field case

reservoir models. Since FMM has already been described earlier in Chapter 2, only the

improvements in FMM simulator associated with upgrading it to a three phase and

compositional simulator have been presented in this Chapter.

Zhang et al. (2014 and 2016) presented a genetic algorithm (GA) based history

matching study in a field case using FMM based reservoir simulator. In their study, the

reservoir model was divided into three groups – hydraulic fracture region, Stimulated

Reservoir Region (SRV) and outer region. The SRV region is box shaped whose

dimensions are needed to be calibrated during history matching. Hydraulic fractures are

in transverse direction to the horizontal well and changed in vertical direction only. These

hydraulic fractures are divided into several groups such that each group has hydraulic

*
Parts of the text and data reported in this chapter is reprinted with permission from:
 Iino, A., Vyas, A., Huang, J., Datta-Gupta, A., Fujita, Y., Bansal, N. and Sankaran, S., April, 2017.
Efficient Modeling and History Matching of Shale Oil Reservoirs Using the Fast Marching Method:
Field Application and Validation. SPE Western Regional Meeting held in Bakersfield, California,
USA. Copyright 2017 Society of Petroleum Engineers (SPE)
 Iino, A., Vyas, A., Huang, J., Datta-Gupta, A., Fujita, Y. and Sankaran, S., July, 2017. Rapid
Compositional Simulation and History Matching of Shale Oil Reservoirs Using the Fast Marching
Method. Unconventional Resources Technology Conference held in Austin, Texas, USA. Copyright
2017 Unconventional Resources Technology Conference (URTeC)

108
fractures with similar history matching parameters. This is done to reduce the number of

parameters needed to be calibrated during history matching. This chapter study follows a

similar approach of dividing current field case model into various regions before applying

genetic algorithm based history matching.

4.2 Methodology

The methodology followed here is similar to Chapter 3 of this dissertation using

GA. However different versions of FMM based reservoir simulators are applied in this

studied incorporating both three phase and compositional field case models. A short

description of dual porosity based two phase FMM simulator is provided in Chapter 3 of

this dissertation. This study involves extending application of FMM based simulator to

field case scenario for history matching purpose. Necessary updates in FMM based

simulator have been incorporated (Iino et al. (2017)) and the newer versions of these

simulators are applied in the field case study.

In the three phase FMM algorithm, single phase diffusivity is replaced by

multiphase diffusivity (Iino et al. 2017):

𝜆 𝑘
𝛼𝑚𝑝 = 𝜙𝑐𝑡 (4.1)
𝑡

where,

𝜆𝑡 = total mobility

𝑐𝑡 = total compressibility

Kazemi et. al. (1976) and Gilman and Kazemi (1983) reported following equations

for mass balance in dual porosity model:

109
Mass balance equation for oil phase:

𝜕 𝑆𝑜𝑓 𝑘 𝑞̃ Γ
(𝜙𝑓 ) = ∇. (𝒌𝑓 𝐵𝑟𝑜𝑓 ∇𝑝𝑓 ) + 𝐵𝑜 − 𝐵𝑜 (4.2)
𝜕𝑡 𝐵𝑜 𝜇 𝑜 𝑜 𝑜 𝑜

Mass balance equation for water phase:

𝜕 𝑆𝑤𝑓 𝑘 𝑞̃ Γ
(𝜙𝑓 ) = ∇. (𝒌𝑓 𝐵 𝑟𝑤𝑓 ∇𝑝𝑓 ) + 𝐵𝑤 − 𝐵𝑤 (4.3)
𝜕𝑡 𝐵𝑤 𝜇 𝑤 𝑤 𝑤 𝑤

Mass balance equation for gas phase:

𝜕 𝑆𝑔𝑓 𝑆𝑜𝑓 𝑘𝑟𝑔𝑓 𝑘 𝑞̃𝑔 𝑞̃ Γ𝑔 Γ

[𝜙𝑓 ( 𝐵 + 𝑅𝑠 )] = ∇. [𝒌 (𝐵 + 𝑅𝑠 𝐵𝑟𝑜𝑓 ) ∇𝑝𝑓 ] + (𝐵 + 𝑅𝑠 𝐵𝑜 ) − (𝐵 + 𝑅𝑠 𝐵𝑜 )
𝜕𝑡 𝑔 𝐵𝑜 𝑔 𝜇𝑔 𝜇 𝑜 𝑜 𝑔 𝑜 𝑔 𝑜

(4.4)

where,:

𝑘
Γ𝑗 = 𝑓𝑙𝑢𝑖𝑑 𝑡𝑟𝑎𝑛𝑠𝑓𝑒𝑟 𝑡𝑒𝑟𝑚 = 𝜎𝑘𝑚 ( 𝜇𝑟𝑗) (𝑝𝑓 − 𝑝𝑚 ) (4.5)
𝑗

𝜎 = shape factor that depends on connectivity between matrix and surrounding fractures

𝑗 = phase type: oil/water/gas

To transform coordinate system from physical coordinates to τ coordinate, Eq. 4.6

is used (Iino et al, 2017):

𝑘 𝜙𝑓,𝑟𝑒𝑓 𝜕 𝑐 𝑘𝑟𝑗 𝜕𝑝
∇. (𝒌𝑓 𝐵 𝑟𝑗 ∇𝑝𝑓 ) ≡ − [𝑤(𝜏) (𝜆𝑡 ) ] (4.6)
𝜇 𝑗 𝑗 𝑤(𝜏) 𝜕𝜏 𝑡 𝑟𝑒𝑓 𝐵𝑗 𝜇𝑗 𝜕𝜏

The new mass balance equations for oil, water and gas phases then become (Iino et al.,

2017):

Mass balance equation for oil phase:

𝜕 𝑆𝑜𝑓 𝜙𝑓,𝑟𝑒𝑓 𝜕 𝑐 𝑘𝑟𝑜 𝜕𝑝𝑓 𝑞̃ Γ

(𝜙𝑓 )= (𝑤(𝜏) (𝜆𝑡 ) ) + 𝐵𝑜 𝛿(𝜏𝑤𝑏 ) − 𝐵𝑜 (4.7)
𝜕𝑡 𝐵𝑜 𝑤(𝜏) 𝜕𝜏 𝑡 𝑟𝑒𝑓 𝐵𝑜 𝜇𝑜 𝜕𝜏 𝑜 𝑜

Mass balance equation for water phase:

110
𝜕 𝑆𝑤𝑓 𝜙𝑓,𝑟𝑒𝑓 𝜕 𝑐 𝑘𝑟𝑤𝑓 𝜕𝑝𝑓 𝑞̃ Γ
(𝜙𝑓 )= (𝑤(𝜏) (𝜆𝑡 ) ) + 𝐵𝑤 𝛿(𝜏𝑤𝑏 ) − 𝐵𝑤 (4.8)
𝜕𝑡 𝐵𝑤 𝑤(𝜏) 𝜕𝜏 𝑡 𝑟𝑒𝑓 𝐵𝑤 𝜇𝑤 𝜕𝜏 𝑤 𝑤

Mass balance equation for gas phase:

𝜕 𝑆𝑔𝑓 𝑆𝑜𝑓 𝜙𝑓,𝑟𝑒𝑓 𝜕 𝑐 𝑘𝑟𝑔𝑓 𝑘 𝜕𝑝𝑓 𝑞̃𝑔

[𝜙𝑓 ( 𝐵 + 𝑅𝑠 )] = [𝑤(𝜏) (𝜆𝑡 ) (𝐵 + 𝑅𝑠 𝐵𝑟𝑜𝑓 ) ] + (𝐵 +
𝜕𝑡 𝑔 𝐵𝑜 𝑤(𝜏) 𝜕𝜏 𝑡 𝑟𝑒𝑓 𝑔 𝜇𝑔 𝜇
𝑜 𝑜 𝜕𝜏 𝑔

𝑞̃ Γ𝑔 Γ
𝑅𝑠 𝐵𝑜 ) 𝛿(𝜏𝑤𝑏 ) − (𝐵 + 𝑅𝑠 𝐵𝑜 ) (4.9)
𝑜 𝑔 𝑜

Eqs. 4.7 to 4.9 show that mass balance equations can be solved w.r.t 1-D τ

coordinate system. These equations can be solved using a finite difference method to

calculate oil, water and gas rates. A detailed description of this FMM based reservoir

simulator is provided in Iino et al. (2017). The compositional FMM version follows

similar concept except that it incorporates compositional effects (Iino et. al, 2017)

The field case under investigation in this chapter is used to match history data and

to forecast future production. The history matching problem in this chapter is based on

Genetic Algorithm (GA). However instead of maximizing the objective function (NPV)

as in the case of Chapter 2, the objective function in this study (mismatch error) is to be

minimized in this chapter. The objective function or error function,𝑓(𝑚), to be minimized

in this study is given by:

𝑓(𝑚) = 𝑙𝑛|∆𝐶𝑢𝑚_𝑂𝑖𝑙| + 𝑙𝑛|∆𝐶𝑢𝑚_𝑊𝑎𝑡𝑒𝑟| + 𝑙𝑛|∆𝐶𝑢𝑚_𝐺𝑎𝑠| (4.1)

where,

∆𝐶𝑢𝑚_𝑂𝑖𝑙, ∆𝐶𝑢𝑚_𝑊𝑎𝑡𝑒𝑟 and ∆𝐶𝑢𝑚_𝐺𝑎𝑠 = root mean squared errors of observed

cumulative production and simulated cumulative production for corresponding phases:

oil/gas/water

111
 Iino et al. (2017) presented history matching results using three phase FMM and

compositional FMM. This study uses the same reservoir model but applies a slightly

different approach of GA based workflow. Fig. 4.1 shows various steps in history

matching using this modified GA consisting of various GA stages. First, the objective

function is tested for sensitivity w.r.t various reservoir model parameters needed to be

calibrated for history matching. To calculate sensitivity, a parameter is perturbed to its

maximum and minimum values keeping all other parameters at their base values. The

relative change in the objective function compared to the base model (in which all

parameters are kept at their base values) is calculated. This is repeated for all parameters

to be calibrated one at a time and compared together in the end. Finally, an engineering

judgement is made to decide if any parameter is needed to be removed from further study.

If one or more parameters are not affecting the objective function significantly, they can

be discarded for next GA stage. Once GA results show no further significant improvement

(in terms of variable ranges and objective error values), the GA is stopped and a collection

of best models are selected. Next, the updated variable ranges for the variables included

in the previous GA stage is utilized for next GA stage. Also, the variables that were

discarded in previous stage are also incorporated. Similar process is repeated in the next

GA until reasonably good history matching results are observed.

112
 Sensitivity Analysis
Selection of Significant Variables

Initialize/Update current Update Variables

generation’s population Update Variable Ranges

Run Simulation &

Evaluate Obj. Function
Select
Crossover N
Mutate
Y
Stop Criteria? HM err < Ɛ

N Y

GA: accepted by fitness STOP

Figure 4.1 General workflow for genetic algorithm (GA)

4.3 Results and Discussion

The field case dual porosity model studied here is dimensioned 7,100 ft × 2,500 ft

× 180 ft. The reservoir model has 71 × 25 × 13 (= 23,075) grid blocks. Initial reservoir

pressure is 3,953 psi with bubble point pressure of 2,930 psi and therefore the reservoir is

initially under saturated. The model has a single horizontal well with ten stages of

hydraulic fractures. The model is divided mainly in three regions - Hydraulic Fractures,

Stimulated Reservoir Volume (SRV) and non-SRV region (outer region) (Fig. 4.2). Table

4.1 lists various variables where the uncertainty exists with corresponding minimum and

maximum values. The base values are the best estimate of a given variable. These variable

ranges are determined with active discussions with the operator of this field.
113
Figure 4.2 Three regions in the field case reservoir model

114
Table 4.1 Uncertainty in Model parameters and their base values for Sensitivity Analysis (Iino

et al., 2017)

Region Uncertain Parameters Low High Base

Porosity
0.005 0.02 0.01
(HF_poro1, HF_poro2, HF_poro3)
Permeability (mD)
0.2 3.0 0.55
(HF_perm1, HF_perm2, HF_perm3)
Water saturation (HF_Swi) 0.75 0.95 0.85
Compaction table (HF_comp) 2 12 2
Hydraulic Shape factor (ft-2)
0.0025 0.5 0.005
(HF_sigma1, HF_sigma2, HF_sigma3)
Fracture Fracture half length (ft)
50 150 50
(HF_Xf1, HF_Xf2, HF_Xf3 )
Fracture height (ft)
40 100 60
(HF_h1, HF_h2, HF_h3)
Stage length (ft)
300-400 500-600 500-600
(HF_len1, HF_len2, HF_len3)
Porosity
0.005 0.012 0.01
(SRV_poro1, SRV_poro2, SRV_poro3)
Permeability (mD)
0.01 0.2 0.1
(SRV_perm1, SRV_perm2, SRV_perm3)
Water saturation
0.175 0.7 0.35
(SRV_ Swi1, SRV_ Swi2, SRV_ Swi3 )
SRV Compaction table (SRV_ comp ) 2 12 2
Shape factor (ft-2)
1.25×10-4 0.02 1.25×10-3
(SRV_ sigma1, SRV_ sigma2, SRV_ sigma3)
SRV_Width (ft)
300 900 500
(SRV_W1, SRV_W2, SRV_W3)
Porosity (Mat_poro) 0.059 0.094 0.08
Permeability (Mat_perm), mD 2.3×10-7 1.3×10-4 2.7×10-5
Matrix Water saturation (Mat_Swi) 0.3 0.77 0.41
Connate water saturation (Mat_Swc) 0.5*Swi 1.0*Swi 1.0*Swi

115
 4.3.1 History matching results based on GA and three phase FMM

Iino et al. (2017) presented a FMM based three phase unconventional reservoir

simulator that is multiple times faster than a commercially available finite difference based

reservoir simulator. This study applied FMM as a suitable candidate for history matching

problem involving large number of simulations. Current study also utilizes the advantages

of FMM for history matching. To test accuracy of FMM relative to Eclipse, simulations

have been conducted for both FMM based simulator and Eclipse for the field case model

under investigation using the base values of each variable. Fig. 4.3 shows the well

constraint utilized here which is tubing head pressure. Figs. 4.4 to 4.9 present the

comparison plots of the simulation results using three phase FMM simulator and Eclipse

100 simulator. It is clear from these figures that FMM and Eclipse are reasonably close to

each other and therefore, FMM can be a good candidate for further history matching

simulations due to faster simulations (Iino et. al, 2017).

Figure 4.3 Well constraint Tubing Head Pressure during well production period

116
 Figure 4.4 Cumulative Oil Production of FMM and Eclipse as compared to History data

with base case variables (three phase FMM)

Figure 4.5 Oil Rate Production of FMM and Eclipse as compared to History data with base

case variables (three phase FMM)

117
Figure 4.6 Cumulative Water Production of FMM and Eclipse as compared to History data

with base case variables (three phase FMM)

Figure 4.7 Water Rate Production of FMM and Eclipse as compared to History data with

base case variables (three phase FMM)

118
Figure 4.8 Cumulative Gas Production of FMM and Eclipse as compared to History data

with base case variables (three phase FMM)

Figure 4.9 Gas Rate Production of FMM and Eclipse as compared to History data with

base case variables (three phase FMM)

119
 As presented in the previous section of this chapter, a multi-stage GA approach

has been utilized for this study. In stage 1, sensitivity analysis is done and relative

importance of variaous variables are checked. Heavy hitter variables or the variables

making relatively larger impact on the objective error functions are identified and rest of

the variables are discarded for this stage. Fig 4.10 shows the results of sensitivity analysis.

Parameters not included for this stage GA are shown in green boxes.

HF Porosity

HF Perm

HF Swi
HF Compaction
HF Sigma

HF Half-Lengths

HF Height

HF Stage Length
SRV Porosity

SRV Perm

SRV Swi
SRV Compaction
SRV Sigma
SRV Width
Matrix Porosity
Matrix Perm
Swi
Swc Multiplier

Figure 4.10 Sensitivity analysis at the beginning of Stage 1 (three phase FMM)
120
 Fig. 4.11 shows the results of GA in stage 1. As can be observed from this figure,

after multiple generations, improvement in objective error function reduces. Also, since

variables in this GA operation show large shrinkage in their ranges from generation 1 to

generation 12 (Figs. 4.12 to 4.18), GA was stopped at this point and a collection of best

models was selected (Fig. 4.11). These best models are chosen to derive new variable

ranges of the variables included for the next GA stage. Figs. 4.19 to 4.25 show the variable

distribution in generation 1 of this stage while Figs. 4.26 to 4.32 show the variable ranges

in the best models selected at the end of this GA stage. It may be observed that a relatively

uniform variable distribution transforms into a narrower and close to normal distribution.

Figure 4.11 GA results for Stage 1 (three phase FMM)

121
Figure 4.12 Uncertainty reduction in hydraulic fracture permeability during GA - Stage 1

(three phase FMM)

Figure 4.13 Uncertainty reduction in hydraulic fracture initial water saturation during GA -

Stage 1 (three phase FMM)

122
Figure 4.14 Uncertainty reduction in hydraulic fracture shape factor during GA - Stage 1

(three phase FMM)

Figure 4.15 Uncertainty reduction in SRV porosity during GA - Stage 1 (three phase FMM)

123
Figure 4.16 Uncertainty reduction in SRV permeability during GA - Stage 1 (three phase

FMM)

Figure 4.17 Uncertainty reduction in SRV initial water saturation during GA - Stage 1

(three phase FMM)

124
Figure 4.18 Uncertainty reduction in SRV shape factor during GA - Stage 1 (three phase

FMM)

Figure 4.19 Variable distribution of hydraulic fracture permeability in the first generation

of GA - Stage 1 (three phase FMM)

125
Figure 4.20 Variable distribution of hydraulic fracture initial water saturation in the first

generation of GA - Stage 1 (three phase FMM)

Figure 4.21 Variable distribution of hydraulic fracture shape factor in the first generation

of GA - Stage 1 (three phase FMM)

126
 Figure 4.22 Variable distribution of SRV porosity in the first generation of GA - Stage 1

(three phase FMM)

Figure 4.23 Variable distribution of SRV permeability in the first generation of GA - Stage

1 (three phase FMM)

127
Figure 4.24 Variable distribution of SRV initial water saturation in the first generation of

GA - Stage 1 (three phase FMM)

Figure 4.25 Variable distribution of SRV shape factor in the first generation of GA - Stage

1 (three phase FMM)

128
Figure 4.26 Variable distribution of hydraulic fracture permeability in the best selected

models of GA - Stage 1 (three phase FMM)

Figure 4.27 Variable distribution of hydraulic fracture initial water saturation in the best

selected models of GA - Stage 1 (three phase FMM)

129
Figure 4.28 Variable distribution of hydraulic fracture shape factor in the best selected

models of GA - Stage 1 (three phase FMM)

Figure 4.29 Variable distribution of SRV porosity in the best selected models of GA -

Stage 1 (three phase FMM)

130
Figure 4.30 Variable distribution of SRV permeability in the best selected models of GA -

Stage 1 (three phase FMM)

Figure 4.31 Variable distribution of SRV initial water saturation in the best selected

models of GA - Stage 1 (three phase FMM)

131
Figure 4.32 Variable distribution of SRV shape factor in the best selected models of GA -

Stage 1 (three phase FMM)

132
 In the next GA stage, the variables of stage 1 are kept with updated ranges based

on best models selected previously and the previously discarded variables are also

included. Fig. 4.33 shows the new sensitivity plot. It can be observed that this time, more

uniformity is seen in terms of variable importance.

HF Porosity

HF Perm
HF Swi
HF Compaction
HF Sigma
HF Half-Lengths
HF Height

HF Stage Length

SRV Porosity
SRV Perm
SRV Swi
SRV Compaction
SRV Sigma

SRV Width
Matrix Porosity
Matrix Perm
Swi Swc Multiplier

Figure 4.33 Sensitivity analysis at the beginning of Stage 2 (three phase FMM)

133
 Fig. 4.34 shows the results of GA in stage 2. As can be observed from this figure,

after multiple generations, improvement in objective error function reduces. Also, since

variables in this GA operation show large shrinkage in their ranges from generation 1 to

generation 12 (Figs. 4.35 to 4.42), GA was stopped at this point and a collection of best

models was selected (Fig. 4.34). These best models are chosen to derive new variable

ranges of the variables included for this GA stage. Figs. 4.43 to 4.50 show the variable

ranges in the best models selected at the end of this GA stage. It may be observed that

distributions of the variables common with previous stage have become narrower showing

further reduction in uncertainty.

Figure 4.34 GA results for Stage 2 (three phase FMM)

134
Figure 4.35 Uncertainty reduction in hydraulic fracture porosity during GA - Stage 2 (three

phase FMM)

Figure 4.36 Uncertainty reduction in hydraulic fracture permeability during GA - Stage 2

(three phase FMM)

135
Figure 4.37 Uncertainty reduction in hydraulic fracture initial water saturation during GA -

Stage 2 (three phase FMM)

Figure 4.38 Uncertainty reduction in hydraulic fracture shape factor during GA - Stage 2

(three phase FMM)

136
Figure 4.39 Uncertainty reduction in SRV porosity during GA - Stage 2 (three phase FMM)

Figure 4.40 Uncertainty reduction in SRV permeability during GA - Stage 2 (three phase

FMM)

137
 Figure 4.41 Uncertainty reduction in SRV initial water saturation during GA - Stage 2

(three phase FMM)

Figure 4.42 Uncertainty reduction in SRV shape factor during GA - Stage 2 (three phase

FMM)

138
Figure 4.43 Variable distribution of hydraulic fracture porosity in the best selected models

of GA - Stage 2 (three phase FMM)

Figure 4.44 Variable distribution of hydraulic fracture permeability in the best selected

models of GA - Stage 2 (three phase FMM)

139
Figure 4.45 Variable distribution of hydraulic fracture initial water saturation in the best

selected models of GA - Stage 2 (three phase FMM)

Figure 4.46 Variable distribution of hydraulic fracture shape factor in the best selected

models of GA - Stage 2 (three phase FMM)

140
 Figure 4.47 Variable distribution of SRV porosity in the best selected models of GA -

Stage 2 (three phase FMM)

Figure 4.48 Variable distribution of SRV permeability in the best selected models of GA -

Stage 2 (three phase FMM)

141
 Figure 4.49 Variable distribution of SRV initial water saturation in the best selected

models of GA - Stage 2 (three phase FMM)

Figure 4.50 Variable distribution of SRV shape factor in the best selected models of GA -

Stage 2 (three phase FMM)

142
 In the next GA stage, the variables of the previous stage are kept with updated

ranges based on best models selected previously. Fig. 4.51 shows the new sensitivity plot.

It can be observed that this time, some of the variables are not making big impact due to

shrinkage of their ranges in the previous GA stages. However, all the variables are

included in this GA stage.

HF Porosity

HF Perm
HF Swi
HF Compaction
HF Sigma

HF Half-Lengths

HF Height

HF Stage Length

SRV Porosity
SRV Perm

SRV Swi
SRV Compaction
SRV Sigma

SRV Width
Matrix Porosity
Matrix Perm
Swi
Swc Multiplier

Figure 4.51 Sensitivity analysis at the beginning of Stage 3 (three phase FMM)
143
 Fig. 4.52 shows the results of GA in stage 3. As can be observed from this figure,

after multiple generations, improvement in objective error function reduces. Also, since

variables in this GA operation show large shrinkage in their ranges from generation 1 to

generation 12 (Figs. 4.53 to 4.60), GA was stopped at this point and a collection of best

models was selected (Fig. 4.52). These best models are chosen to derive new variable

ranges of the variables included for this GA stage. Figs. 4.61 to 4.67 show the variable

ranges in the best models selected at the end of this GA stage. It may be observed that

distributions of the variables common with previous stage have become narrower showing

further reduction in uncertainty.

Figure 4.52 GA results for Stage 3 (three phase FMM)

144
Figure 4.53 Uncertainty reduction in hydraulic fracture porosity during GA - Stage 3 (three

phase FMM)

Figure 4.54 Uncertainty reduction in hydraulic fracture permeability during GA - Stage 3

(three phase FMM)

145
Figure 4.55 Uncertainty reduction in hydraulic fracture initial water saturation during GA -

Stage 3 (three phase FMM)

Figure 4.56 Uncertainty reduction in hydraulic fracture shape factor during GA - Stage 3

(three phase FMM)

146
Figure 4.57 Uncertainty reduction in SRV porosity during GA - Stage 3 (three phase FMM)

Figure 4.58 Uncertainty reduction in SRV permeability during GA - Stage 3 (three phase

FMM)

147
 Figure 4.59 Uncertainty reduction in SRV initial water saturation during GA - Stage 3

(three phase FMM)

Figure 4.60 Uncertainty reduction in SRV shape factor during GA - Stage 3 (three phase

FMM)

148
Figure 4.61 Variable distribution of hydraulic fracture porosity in the best selected models

of GA - Stage 3 (three phase FMM)

Figure 4.62 Variable distribution of hydraulic fracture permeability in the best selected

models of GA - Stage 3 (three phase FMM)

149
Figure 4.63 Variable distribution of hydraulic fracture initial water saturation in the best

selected models of GA - Stage 3 (three phase FMM)

Figure 4.64 Variable distribution of hydraulic fracture shape factor in the best selected

models of GA - Stage 3 (three phase FMM)

150
 Figure 4.65 Variable distribution of SRV porosity in the best selected models of GA -

Stage 3 (three phase FMM)

Figure 4.66 Variable distribution of SRV permeability in the best selected models of GA -

Stage 3 (three phase FMM)

151
 Figure 4.67 Variable distribution of SRV initial water saturation in the best selected

models of GA - Stage 3 (three phase FMM)

Fig. 4.68 shows the combined plot showing all GA stages. It may be observed that

there is significant improvement from one GA stage to the next one. At this point the best

models are selected as mentioned previously and plotted against history data (Figs. 4.69

to 4.74).

152
Figure 4.68 Combined GA results for all stages (three phase FMM)

153
 (a)

History Matching

Forecast

(b)

Figure 4.69 Cumulative oil history production data vs simulated production data (a) in the

first stage first generation and (b) including only the best selected models from the last

stage (three phase FMM)

154
 (a)

History Matching

Forecast

(b)

Figure 4.70 Cumulative water history production data vs simulated production data (a) in

the first stage first generation and (b) including only the best selected models from the

last stage (three phase FMM)

155
 (a)

History Matching

Forecast

(b)

Figure 4.71 Cumulative gas history production data vs simulated production data (a) in

the first stage first generation and (b) including only the best selected models from the

last stage (three phase FMM)

156
 (a)

History Matching

Forecast

(b)

Figure 4.72 Oil rate history production data vs simulated production data (a) in the first

stage first generation and (b) including only the best selected models from the last stage

(three phase FMM)

157
 (a)

History Matching

Forecast

(b)

Figure 4.73 Water rate history production data vs simulated production data (a) in the first

stage first generation and (b) including only the best selected models from the last stage

(three phase FMM)

158
 (a)

History Matching

Forecast

(b)

Figure 4.74 Gas rate history production data vs simulated production data (a) in the first

stage first generation and (b) including only the best selected models from the last stage

(three phase FMM)

159
 4.3.2 History matching results based on GA and compositional FMM

Iino et al. (2017) presented a FMM based compositional unconventional reservoir

simulator that is multiple times faster than a commercially available finite difference based

reservoir simulator. Their study applied compositional FMM as a suitable candidate for

history matching problem involving large number of simulations. Current dissertation

study also utilizes the advantages of compositional FMM for history matching. To test

accuracy of FMM relative to Eclipse, simulations have been conducted for both FMM

based simulator and Eclipse for the field case model under investigation using the base

values of each variable. Figs. 4.75 to 4.80 present the comparison plots of the simulation

results using compositional FMM simulator and Eclipse 300 simulator. It is clear from

these figures that FMM and Eclipse are reasonably close to each other and therefore, FMM

can be a good candidate for further history matching simulations due to faster simulations

(Iino et. al, 2017).

160
 Figure 4.75 Cumulative Oil Production of FMM vs Eclipse as compared to History data

with base case variables (compositional FMM)

Figure 4.76 Oil Rate Production of FMM vs Eclipse as compared to History data with base

case variables (compositional FMM)

161
Figure 4.77 Cumulative Water Production of FMM vs Eclipse as compared to History data

with base case variables (compositional FMM)

Figure 4.78 Water Rate Production of FMM vs Eclipse as compared to History data with

base case variables (compositional FMM)

162
Figure 4.79 Cumulative Gas Production of FMM vs Eclipse as compared to History data

with base case variables (compositional FMM)

Figure 4.80 Gas Rate Production of FMM vs Eclipse as compared to History data with

base case variables (compositional FMM)

163
 As presented in the previous section of this chapter, a multi-stage GA approach

has been utilized for this study. In stage 1, sensitivity analysis is done and relative

importance of variaous variables are checked. Heavy hitter variables or the variables

making relatively larger impact on the objective error functions are identified and rest of

the variables are discarded for this stage. Fig 4.81 shows the results of sensitivity analysis.

Parameters not included for this stage GA are shown in green boxes.

HF Porosity

HF Perm

HF Swi
HF Compaction
HF Sigma

HF Half-Lengths

HF Height

HF Stage Length
SRV Porosity

SRV Perm

SRV SRV Swi

Compaction
SRV Sigma
SRV Width
Matrix Porosity
Matrix Perm
Swi
Swc Multiplier

Figure 4.81 Sensitivity analysis at the beginning of Stage 1 (compositional FMM)

164
 Fig. 4.82 shows the results of GA in stage 1. As can be observed from this figure,

after multiple generations, improvement in objective error function reduces. Also, since

variables in this GA operation show large shrinkage in their ranges from generation 1 to

generation 12 (Figs. 4.83 to 4.88), GA was stopped at this point and a collection of best

models was selected (Fig. 4.82). These best models are chosen to derive new variable

ranges of the variables included for this GA stage. Figs. 4.89 to 4.94 show the variable

distribution in generation 1 of this stage while Figs. 4.95 to 4.100 show the variable ranges

in the best models selected at the end of this GA stage. It may be observed that a relatively

uniform variable distribution transforms into a narrower and close to normal distribution.

Figure 4.82 GA results for Stage 1 (compositional FMM)

165
 Figure 4.83 Uncertainty reduction in hydraulic fracture porosity during GA - Stage 1

(compositional FMM)

Figure 4.84 Uncertainty reduction in hydraulic fracture initial water saturation during GA -

Stage 1 (compositional FMM)

166
Figure 4.85 Uncertainty reduction in hydraulic fracture shape factor during GA - Stage 1

(compositional FMM)

Figure 4.86 Uncertainty reduction in SRV porosity during GA - Stage 1 (compositional

FMM)

167
Figure 4.87 Uncertainty reduction in SRV permeability during GA - Stage 1 (compositional

FMM)

Figure 4.88 Uncertainty reduction in SRV shape factor during GA - Stage 1 (compositional

FMM)

168
Figure 4.89 Variable distribution of hydraulic fracture porosity in the first generation of

GA - Stage 1 (compositional FMM)

Figure 4.90 Variable distribution of hydraulic fracture initial water saturation in the first

generation of GA - Stage 1 (compositional FMM)

169
Figure 4.91 Variable distribution of hydraulic fracture shape factor in the first generation

of GA - Stage 1 (compositional FMM)

Figure 4.92 Variable distribution of SRV porosity in the first generation of GA - Stage 1

(compositional FMM)

170
Figure 4.93 Variable distribution of SRV permeability in the first generation of GA - Stage

1 (compositional FMM)

Figure 4.94 Variable distribution of SRV shape factor in the first generation of GA - Stage

1 (compositional FMM)

171
Figure 4.95 Variable distribution of hydraulic fracture porosity in the best selected models

of GA - Stage 1 (compositional FMM)

Figure 4.96 Variable distribution of hydraulic fracture initial water saturation in the best

selected models of GA - Stage 1 (compositional FMM)

172
Figure 4.97 Variable distribution of hydraulic fracture shape factor in the best selected

models of GA - Stage 1 (compositional FMM)

Figure 4.98 Variable distribution of SRV porosity in the best selected models of GA -

Stage 1 (compositional FMM)

173
Figure 4.99 Variable distribution of SRV permeability in the best selected models of GA -

Stage 1 (compositional FMM)

Figure 4.100 Variable distribution of SRV shape factor in the best selected models of GA -

Stage 1 (compositional FMM)

In the next GA stage, the variables of the previous stage are kept with updated

ranges based on best models selected previously. Fig. 4.101 shows the new sensitivity

plot. It can be observed that this time, some of the variables are not making big impact

174
due to shrinkage of their ranges in the previous GA stages. However, all the variables are

included in this GA stage.

HF Porosity

HF Perm

HF Swi
HF Compaction
HF Sigma
HF Half-Lengths

HF Height

HF Stage Length
SRV
Porosity
SRV Perm

SRV SRV Swi

Compaction
SRV Sigma

SRV Width

Matrix Porosity
Matrix Perm
Swi
Swc Multiplier

Figure 4.101 Sensitivity analysis at the beginning of Stage 2 (compositional FMM)

Fig. 4.102 shows the results of GA in stage 2. As can be observed from this figure,

after multiple generations, improvement in objective error function reduces. Also, since
175
variables in this GA operation show large shrinkage in their ranges from generation 1 to

generation 10 (Figs. 4.103 to 4.110), GA was stopped at this point and a collection of best

models was selected (Fig. 4.102). These best models are chosen to derive new variable

ranges of the variables included for this GA stage. Figs. 4.111 to 4.118 show the variable

ranges in the best models selected at the end of this GA stage. It may be observed that

distributions of the variables common with previous stage have become narrower showing

further reduction in uncertainty.

Figure 4.102 GA results for Stage 2 (compositional FMM)

176
 Figure 4.103 Uncertainty reduction in hydraulic fracture porosity during GA - Stage 2

(compositional FMM)

Figure 4.104 Uncertainty reduction in hydraulic fracture permeability during GA - Stage 2

(compositional FMM)

177
Figure 4.105 Uncertainty reduction in hydraulic fracture initial water saturation during GA

- Stage 2 (compositional FMM)

Figure 4.106 Uncertainty reduction in hydraulic fracture shape factor during GA - Stage 2

(compositional FMM)

178
Figure 4.107 Uncertainty reduction in SRV porosity during GA - Stage 2 (compositional

FMM)

Figure 4.108 Uncertainty reduction in SRV permeability during GA - Stage 2

(compositional FMM)

179
Figure 4.109 Uncertainty reduction in SRV initial water saturation during GA - Stage 2

(compositional FMM)

Figure 4.110 Uncertainty reduction in SRV shape factor during GA - Stage 2

(compositional FMM)

180
 Figure 4.111 Variable distribution of hydraulic fracture porosity in the best selected

models of GA - Stage 2 (compositional FMM)

Figure 4.112 Variable distribution of hydraulic fracture permeability in the best selected

models of GA - Stage 2 (compositional FMM)

181
Figure 4.113 Variable distribution of hydraulic fracture initial water saturation in the best

selected models of GA - Stage 2 (compositional FMM)

Figure 4.114 Variable distribution of hydraulic fracture shape factor in the best selected

models of GA - Stage 2 (compositional FMM)

182
 Figure 4.115 Variable distribution of SRV porosity in the best selected models of GA -

Stage 2 (compositional FMM)

Figure 4.116 Variable distribution of SRV permeability in the best selected models of GA -

Stage 2 (compositional FMM)

183
 Figure 4.117 Variable distribution of SRV initial water saturation in the best selected

models of GA - Stage 2 (compositional FMM)

Figure 4.118 Variable distribution of SRV shape factor in the best selected models of GA -

Stage 2 (compositional FMM)

184
 Fig. 4.119 shows the combined plot showing all GA stages. It may be observed

that there is significant improvement from one GA stage to the next one. At this point the

best models are selected as mentioned previously and plotted against history data (Figs.

4.120 to 4.125).

Figure 4.119 Combined GA results of all stages (compositional FMM)

185
 (a)

History Matching

Forecast

(b)

Figure 4.120 Cumulative oil history production data vs simulated production data (a) in

the first stage first generation and (b) including only the best selected models from the

last stage (compositional FMM)

186
 (a)

History Matching

Forecast

(b)

Figure 4.121 Cumulative Water history production data vs simulated production data (a)

in the first stage first generation and (b) including only the best selected models from the

last stage (compositional FMM)

187
 (a)

History Matching

Forecast

(b)

Figure 4.122 Cumulative Gas history production data vs simulated production data (a) in

the first stage first generation and (b) including only the best selected models from the

last stage (compositional FMM)

188
 (a)

History Matching

Forecast

(b)

Figure 4.123 Oil rate history production data vs simulated production data (a) in the first

stage first generation and (b) including only the best selected models from the last stage

(compositional FMM)
189
 (a)

History Matching

Forecast

(b)

Figure 4.124 Water rate history production data vs simulated production data (a) in the

first stage first generation and (b) including only the best selected models from the last

stage (compositional FMM)

190
 (a)

History Matching

Forecast

(b)

Figure 4.125 Gas rate history production data vs simulated production data (a) in the first

stage first generation and (b) including only the best selected models from the last stage

(compositional FMM)

191
 4.4 Summary

1. History matching using GA can be an effective tool in reducing model variable

uncertainty. Results show that variable uncertainty can be significantly reduced from

first generation to the final generation.

2. In a scenario with unknown variable sensitivities and ranges, taking a larger initial

variable range is common. This study shows how heavy-hitter variables can be

separated out from other variables and GA can then be conducted only using heavy-

hitter variables. GA can be repeated in the next stage(s) by including previously

eliminated variables and refined ranges of heavy hitters.

3. Best models can be selected from a GA stage to repeat workflow for the new stage

thus converging to the solution faster. Variable distribution plots presented for best

selected models explain how a uniform distribution in the beginning of stage 1 can

be reduced to a smaller range of normal distribution in a stage. This refined variable

range can then be carried over to the next GA stage.

4. History matching and forecast results for the field case has been presented using a

multi-stage GA approach. It has been shown that multi-stage GA can be a faster

alternative to single stage GA to get reasonable history matching results.

5. FMM based simulator has been proven to be an accurate and faster alternative to

commercial simulator for an optimization study requiring large number of forward

simulations. However, this study can be repeated using any commercial finite

difference simulator.

192
 CHAPTER V

CONCLUSIONS AND RECOMMENDATIONS

5.1 Summary and Conclusions

This dissertation study has presented different applications of machine learning

algorithms including GA. Following conclusions can be drawn from this dissertation:

1. In the second chapter, Eagle Ford well data was collected from public website and

fitted with various decline curve models to get best fit decline curve parameters and

expected EUR for each well. Several machine learning algorithms such as Random

Forest, Support Vector Machine and Gradient Boosting Machines are then applied to

correlate well decline curve parameters and EUR to well completion and well location

variables. The models thus developed have been utilized to predict well rate

production as a function of time and also well EUR with reasonable accuracy. Also,

variables making most impact on the EUR have been identified in this study.

2. In the third chapter, Genetic Algorithm (GA) based workflow has been presented to

optimize the Net Present Value (NPV) during well production period. It has been

found in this chapter that NPV cannot be optimized simply by increasing the number

of stages in a horizontal well. A GA based workflow which involves various

fracturing variables such as proppant amount and fracturing fluid amount has been

presented and applied to a synthetic unconventional shale gas reservoir model. The

most optimum design variable set has been compared to the uniformly spaced design

to compare the difference between the two cases. Also, this chapter presents the

effects of uncertainty in reservoir permeability on NPV if the presented workflow is

193
 used to optimize the hydraulic fracture design.

3. In the fourth chapter, a multistage GA approach has been presented to match history

data in a shale oil field case. In this method only the most significant history matching

variables are utilized in the first stage of GA. Once first stage converges based on

criteria mentioned in this chapter, next stage including updated variables and their

ranges are utilized. The updated variable ranges are based upon the best models in

the previous stage. This method can further fine tune variable ranges with better

history matching error as compared to single stage GA.

5.2 Recommendations

Following points are recommended as an extension/improvement to current

dissertation work:

1. In the second chapter study, more variables can be included that impact well rates such

as well head pressure/bottom hole pressure. Also, in case of major changes in the well

constraint variables, fitting a single decline curve may not be suitable for a given well.

In that case multiple decline curves may be fitted and predicted.

2. In the third chapter, ways to predict natural fracture distribution in larger uncertainty

is needed in case this workflow is applied to reservoirs with little or no knowledge of

natural fracture density distribution.

194
 NOMENCLATURE

a = Intercept Constant (Duong Model)

α or alpha = Scale parameter (Weibull Model)

b = Decline coefficient (Arps)

ACE = Alternating Conditional Expectation

BMA = Bayesian Model Averaging

BHP = Bottom Hole Pressure

CART = Classification and Regression Trees

CLENGTH = Completed Length

DCA = Decline Curve Analysis

DFN = Discrete Fracture Network

DOE = Design of Experiments

Di = Initial Decline Rate (Arps)

DTOF = Diffusive Time of Flight

EUR = Estimated Ultimate Recovery

FRAC_FLUID_TOTAL = Total Fracturing Fluid used for a well

FMM = Fast Marching Method

GA = Genetic Algorithm

γ or gamma = Shape parameter (Weibull Model)

GAM = Generalized Additive Model

GBM = Gradient Boosting Machine

195
GCV = Generalized Cross-Validation

GLUE = Generalized Likelihood Uncertainty

Estimation

GOR = Gas-Oil ratio

LATITUDE = Latitude of a well’s location

LHS = Latin Hypercube Sampling

LONGITUDE = Longitude of a well’s location

LEFM = Linear Elastic Fracture Mechanics

m = Slope parameter (Duong Model)

M = Carrying capacity (Weibull)

MARS = Multivariate Adaptive Regression Splines

MD = Measured Depth

MSE = Mean Squared Error

n = Exponent parameter (SEDM)

NNET = Neural Networks

NPV = Net Present Value

OLS = Ordinary Least Squares

PKN = Perkins-Kern-Nordgren

PROP_TOTAL = Total proppant amount used for a well

qi = Initial flow rate or Maximum Flow Rate

q1 = Flow rate during first month (Duong Model)

𝑅2 = Coefficient of Determination

196
𝑅𝐼𝑝 = Relative Variable Importance

𝑅2𝑝 = 𝑅 2 of a model using all predictors

𝑅 2 −𝑝 = 𝑅 2 of a model using all predictors except 𝑝𝑡ℎ

predictor

RF = Random Forest

RMSE = Root Mean Squared Error

RSS = Residual Sum of Squares

SEDM = Stretched Exponential Decline Model

SVM = Support Vector Machine

SVR = Support Vector Regression

SRV = Stimulated Reservoir Volume

STAGE = Number of hydraulic fracture stages in a well

t = Time elapsed during well production

τ = Characteristic time (SEDM)

TOC = Total Organic Content

TVD = Total Vertical Depth

TVD_HEEL = Total Vertical Depth of horizontal well heel

TVD_HEEL_TOE_DIFF = Difference between TVDs of Heel and Toe

UFD = Unified Fracture Design

197
 SUBSCRIPTS

𝑓 = fracture

𝑖 = initial condition

𝑚 = matrix

𝑝 = proppant

𝑢𝑝 = upstream

198
 REFERENCES

Aizerman M.A., Braverman E.M., and Rozonoer L.I. 1964. Theoretical foundations of the
potential function method in pattern recognition learning. Automation and Remote
Control 25: 821–837.

Arps, J.J. 1945. Analysis of Decline Curves. Trans. AIME: 160: 228-247

Beven, K.J., and A. Binley. 1992. The future of distributed models: Model calibration and
uncertainty prediction. Hydrological Processes 6, 279–298

Biswas, P., & Ley, S. B. (2015). Seismic Methodologies Adapted For Use In Acoustic
Logging. Society of Petroleum Engineers. doi:10.2118/175995-MS

Breiman, L., 1996. Technical note: Some properties of splitting criteria. Machine
Learning, 24(1), pp.41-47.

Breiman, L. 2001. "Random forests," Machine Learning, vol. 45, no. 1, pp. 5-32

Centurion, S.M., Cade, R. and Luo, X.L., 2012, January. Eagle Ford Shale: Hydraulic
Fracturing, Completion, and Production Trends: Part II. In SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers.

Centurion, S., Cade, R., Luo, X.L. and Junca-Laplace, J.P. 2013, September. Eagle Ford
Shale: Hydraulic Fracturing, Completion and Production Trends, Part III. In SPE Annual
Technical Conference and Exhibition.

199
Centurion, S., Junca-Laplace, J.P., Cade, R. and Presley, G., 2014, January. Lessons
Learned From an Eagle Ford Shale Completion Evaluation. In SPE Annual Technical
Conference and Exhibition. Society of Petroleum Engineers.

Cheng, H., Dehghani, K., & Billiter, T. C. (2008). A Structured Approach for
Probabilistic-Assisted History Matching Using Evolutionary Algorithms: Tengiz Field
Applications. Society of Petroleum Engineers. doi:10.2118/116212-MS

Cipolla, C. L., Lolon, E., Erdle, J., & Tathed, V. S. (2009). Modeling Well Performance
in Shale-Gas Reservoirs. Society of Petroleum Engineers. doi:10.2118/125532-MS

Cortes, C. and Vapnik, V. 1995. Support vector networks. Machine Learning 20: 273–
297.

Cosma Shalizi. 2006. Statistics 36-350: Data Mining, Fall 2006 online lecture notes.

Daal, J. A., & Economides, M. J. (2006). Optimization of Hydraulically Fractured Wells

in Irregularly Shaped Drainage Areas. Society of Petroleum Engineers.
doi:10.2118/98047-MS

Datta-Gupta, A. and King, M. J., Streamline Simulation: Theory and Practice, Textbook
Series #11, Society of Petroleum Engineers, Richardson, TX, ISBN 978-1-55563-111-6
(2007)

Datta-Gupta, A., Xie, J., Gupta, N. et al. 2011. Radius of Investigation and its
Generalization to Unconventional Reservoirs. Journal of Petroleum Technology 63 (7):
52-55.

200
Dershowitz, B., LaPointe, P., Eiben, T., Wei, L. 2000. Integration of Discrete Feature
Network Methods with Conventional Simulator Approaches. SPE Reservoir Eval. & Eng.,
3 (2).

Draper, D. 1995. Assessment and propagation of model uncertainty. Journal of the Royal
Statistical Society: Series B 57, no. 1: 45–97.

Duong, A. N. 2011. "Rate-Decline Analysis for Fracture-Dominated Shale Reservoirs."

SPEREE 14 (3): 377-387. http://dx.doi.org/10.2118/137748-PA.

Economides, M.J., Oligney, R.E. and Valko, P.P. “Unified Fracture Design”. Orsa Press,
Alvin TX, May 2002.

Economides, M.J., Hill, A.D., Ehlig-Economides, C. and Zhu, D. “Petroleum Production

Systems”. Second Edition. Prentice Hall, 2012.

Fan, L., Thompson, J. W., & Robinson, J. R. (2010). Understanding Gas Production
Mechanism and Effectiveness of Well Stimulation in the Haynesville Shale through
Reservoir Simulation. Society of Petroleum Engineers. doi:10.2118/136696-MS

Fetkovich, M.J. 1980. Decline Curve Analysis Using Type Curves. J PetTechnol 32 (6):
1065–1077.

Fisher, M.K., Wright, C.A., Davidson, B.M., Goodwin, A.K., Fielder, E.O., Buckler, W.S.
and Steinsberger, N.P., 2005, January. Integrating fracture mapping technologies to
improve stimulations in the Barnett Shale. SPE Productions and Facilities 20 (2): 85-93.
doi: 10.2118/77441-PA

201
Friedman, J. H. 1991. Multivariate Adaptive Regression Splines. The Annals of Statistics.
Vol. 19. No. 1: 1-141.

Friedman, J. H. 1993. Fast MARS Stanford University Department of Statistics, Technical

Report 110.

Friedman, J.H., 2001. Greedy function approximation: a gradient boosting

machine. Annals of statistics, pp.1189-1232.

Friedman, J.H., 2002. Stochastic gradient boosting. Computational Statistics & Data
Analysis, 38(4), pp.367-378.

Fujita, Y., Datta-Gupta, A. and King, M., 2016. A Comprehensive Reservoir Simulator
for Unconventional Reservoirs That Is Based on the Fast-Marching Method and Diffusive
Time of Flight. SPE Journal.

Hartigan, J.A. and Wong, M.A., 1979. Algorithm AS 136: A k-means clustering
algorithm. Journal of the Royal Statistical Society. Series C (Applied Statistics), 28(1),
pp.100-108.

Helgesen, T. B., Fulda, C., Meyer, W. H., Thorsen, A. K., Baule, A., Ronning, K. J., &
Iversen, M. (2005). Accurate Wellbore Placement using a Novel Extra Deep Resistivity
Service. Society of Petroleum Engineers. doi:10.2118/94378-MS

Hoeting, J.A., Madigan, D., Raftery, A.E. and Volinsky, C.T., 1999. Bayesian model
averaging: a tutorial. Statistical science, pp.382-401.

Holcomb, W.D., Lafollette, R.F. and Zhong, M., 2015, February. The Third Dimension:
Productivity Effects From Spatial Placement and Well Architecture in Eagle Ford Shale

202
Horizontal Wells. In SPE Hydraulic Fracturing Technology Conference. Society of
Petroleum Engineers.

Holditch, S. A. 2010. Shale Gas Holds Global Opportunities. The American Oil & Gas
Reporter, August 2010 Editor’s Choice.

Holland, J.H. 1992. Genetic Algorithms. Scientific American July 1992: 66-72.

Iino, A., Vyas, A., Huang, J., Datta-Gupta, A., Fujita, Y., Bansal, N. and Sankaran, S.,
April, 2017. Efficient Modeling and History Matching of Shale Oil Reservoirs Using the
Fast Marching Method: Field Application and Validation. SPE Western Regional Meeting
held in Bakersfield, California, USA

Iino, A., Vyas, A., Huang, J., Datta-Gupta, A., Fujita, Y. and Sankaran, S., July, 2017.
Rapid Compositional Simulation and History Matching of Shale Oil Reservoirs Using the
Fast Marching Method. Unconventional Resources Technology Conference held in
Austin, Texas, USA

Ilk, D., Anderson, D. M., Stotts, G. W. J., Mattar, L., & Blasingame, T. (2010). Production
Data Analysis--Challenges, Pitfalls, Diagnostics. Society of Petroleum Engineers.
doi:10.2118/102048-PA

Johnston, D.C. 2006. Stretched Exponential Relaxation Arising From a Continuous Sum
of Exponential Decays. Phys. Rev. B 74: 184430

Kanungo, T., Mount, D.M., Netanyahu, N.S., Piatko, C.D., Silverman, R. and Wu, A.Y.,
2002. An efficient k-means clustering algorithm: Analysis and implementation. IEEE
transactions on pattern analysis and machine intelligence, 24(7), pp.881-892.

203
Kaplan, S., 1981. On the method of discrete probability distributions in risk and reliability
calculations–application to seismic risk assessment. Risk Analysis, 1(3), pp.189-196.

Kass, R.E., and A.E. Raftery. 1995. Bayes factors. Journal of the American Statistical
Association 90, 773–795.

Kennedy, R. L., Gupta, R., Kotov, S. V., Burton, W. A., Knecht, W. N., & Ahmed, U.
(2012). Optimized Shale Resource Development: Proper Placement of Wells and
Hydraulic Fracture Stages. Society of Petroleum Engineers. doi:10.2118/162534-MS

Kim, J. U., Datta-Gupta, A., Brouwer, R., & Haynes, B. (2009). Calibration of High-
Resolution Reservoir Models Using Transient Pressure Data. Society of Petroleum
Engineers. doi:10.2118/124834-MS

Kulkarni, K. N., Datta-Gupta, A., & Vasco, D. W. (2000). A Streamline Approach for
Integrating Transient Pressure Data into High Resolution Reservoir Models. Society of
Petroleum Engineers. doi:10.2118/65120-MS

LaFollette, R.F. and Holcomb, W.D., 2011, January. Practical Data Mining: Lessons-
Learned From the Barnett Shale of North Texas. Paper SPE 140524 presented at the
Hydraulic Fracturing Technology Conference and Exhibition held in the Woodlands,
Texas, USA, 24-26 January.

Lafollette, R., Holcomb, W.D. and Aragon, J., 2012, January. Impact of completion
system, staging, and hydraulic fracturing trends in the Bakken Formation of the Eastern
Williston Basin. In SPE Hydraulic Fracturing Technology Conference. Society of
Petroleum Engineers.

204
Lafollette, R., Holcomb, W.D. and Aragon, J., 2012. Practical Data Mining: Analysis of
Barnett Shale Production Results with Emphasis on Well Completion and Fracture
Stimulation. Paper SPE 152531 presented at the SPE Hydraulic Fracturing Technology
Conference, The Woodlands, Texas, USA, 6–8 February.

LaFollette, R.F. 2013. Shale Gas and Light Tight Oil Reservoir Production Results: What
Matters?. Proceedings of the Twenty-third (2013) International Offshore and Polar
Engineering Conference. International Society of Offshore and Polar Engineers,
Anchorage, Alaska, USA, June 30 – July 5.

LaFollette, R.F., Izadi, G. and Zhong, M. 2013, February. Application of Multivariate

Analysis and Geographic Information Systems Pattern-Recognition Analysis to
Production Results in the Bakken Light Tight Oil Play. In SPE Hydraulic Fracturing
Technology Conference. Society of Petroleum Engineers.

LaFollette, R.F., Izadi, G. and Zhong, M., 2014, February. Application of Multivariate
Statistical Modeling and Geographic Information Systems Pattern-Recognition Analysis
to Production Results in the Eagle Ford Formation of South Texas. In SPE Hydraulic
Fracturing Technology Conference. Society of Petroleum Engineers.

Lee S.H., Kharghoria, A. and Datta-Gupta, A. 2002. Electrofacies Characterization and

Permeability Predictions in Complex Reservoirs. SPE Reservoir Evaluation and
Engineering, 5 (03), pp. 237-248.

Lee, W. J. Well Testing, Society of Petroleum Engineers, Richardson, TX (1982)

Ma, X., Plaksina, T., & Gildin, E. (2013). Optimization of Placement of Hydraulic
Fracture Stages in Horizontal Wells Drilled in Shale Gas Reservoirs. Society of Petroleum
Engineers. doi:10.1190/URTEC2013-151

205
Maxwell, S. C., Urbancic, T. I., Steinsberger, N., & Zinno, R. (2002). Microseismic
Imaging of Hydraulic Fracture Complexity in the Barnett Shale. Society of Petroleum
Engineers. doi:10.2118/77440-MS

Mishra, S. A New Approach to Reserves Estimation in Shale Gas Reservoirs Using

Multiple Decline Curve Analysis Models. Paper SPE 161092 presented at the SPE Eastern
Regional Meeting held in Lexington, Kentucky, USA, 3-5 October 2012.

Mishra, S., Choudhary, M.K. and Datta-Gupta, A., 2002. A novel approach for reservoir
forecasting under uncertainty. SPE Reservoir Evaluation & Engineering, 5(01), pp.42-48.
Mitchell, M. 1999.

An Introduction to Genetic Algorithms. The MIT Press, Cambridge, Massachusetts.

Morales, A. N., Nasrabadi, H., & Zhu, D. (2010). A Modified Genetic Algorithm for
Horizontal Well Placement Optimization in Gas Condensate Reservoirs. Society of
Petroleum Engineers. doi:10.2118/135182-MS

Neuman, S.P., 2003. Maximum likelihood Bayesian averaging of uncertain model

predictions. Stochastic Environmental Research and Risk Assessment, 17(5), pp.291-305.

Nilsson, N.J. 1965. Learning machines: Foundations of Trainable Pattern Classifying

Systems. McGraw-Hill.

Oda, M. 1985. Permeability Tensor for Discontinuous Rock Masses. Geotechnique, 35

(4), pp. 483-495.

Perez, H.H., Datta-Gupta, A. and Mishra, S., 2005. The Role of Electrofacies, Lithofacies,
and Hydraulic Flow Units in Permeability Prediction from Well Logs: A Comparative

206
Analysis Using Classification Trees. SPE Reservoir Evaluation & Engineering, 8(02),
pp.143-155

Pitakbunkate, T., Yang, M., Valko, P. P., & Economides, M. J. (2011). Hydraulic Fracture
Optimization with a p-3D Model. Society of Petroleum Engineers. doi:10.2118/142303-
MS

Rankin, R.R., Thibodeau, M., Vincent, M.C. and Palisch, T., 2010, January. Improved
production and profitability achieved with superior completions in horizontal wells: a
bakken/three forks case history. In SPE Annual Technical Conference and Exhibition.
Society of Petroleum Engineers.

Riahi, A and Damjanac, B. (2013). Numerical Study of Interaction Between Hydraulic

Fracture and Discrete Fracture Network. Proceedings of the International Conference for
Effective and Sustainable Hydraulic Fracturing, Brisbane, Australia.

Saldungaray, P. M., Palisch, T., & Shelley, R. (2013). Hydraulic Fracturing Critical
Design Parameters in Unconventional Reservoirs. Society of Petroleum Engineers.
doi:10.2118/164043-MS

Savitski, A. A., Lin, M., Riahi, A., Damjanac, B., & Nagel, N. B. (2013). Explicit
Modeling of Hydraulic Fracture Propagation in Fractured Shales. International Petroleum
Technology Conference. doi:10.2523/17073-MS

Schuetter J., Mishra S., Zhong M. and LaFollette R. 2015. Data Analytics for Production
Optimization in Unconventional Reservoirs. Paper SPE 178653-MS/URTeC:2167005
presented at the Unconventional Resources Technology Conference held in San Antonio,
Texas. USA, 20-22 July.

207
Sehbi, B. S., Kang, S., Datta-Gupta, A., & Lee, W. J. (2011). Optimizing Fracture Stages
and Completions in Horizontal Wells in Tight Gas Reservoirs Using Drainage Volume
Calculations. Society of Petroleum Engineers. doi:10.2118/144365-MS

Sethian, J. A. 1996. A Fast Marching Level Set Method for Monotonically Advancing
Fronts. Proceedings of the National Academy of Science 93:1591-1595.

Sethian, J. A., Level Set Methods and Fast Marching Methods, Cambridge University
Press, New York City (1999).

Sierra, L., Mayerhofer, M., & Jin, C. J. (2013). Production Forecasting of Hydraulically
Fractured Conventional Low-Permeability and Unconventional Reservoirs Linking the
More Detailed Fracture and Reservoir Parameters. Society of Petroleum Engineers.
doi:10.2118/163833-MS

Singh, A., Mishra, S. and Ruskauff, G., 2010. Model averaging techniques for quantifying
conceptual model uncertainty. Ground Water, 48(5), pp.701-715.

Smola, A., J. and Schölkopf, B. 2004. “A tutorial on support vector regression.” Statistics
and Computing, vol.14, no. 3, pp. 199-222.

Song, B., & Ehlig-Economides, C. A. (2011). Rate-Normalized Pressure Analysis for

Determination of Shale Gas Well Performance. Society of Petroleum Engineers.
doi:10.2118/144031-MS

Valko, P.P. and Lee, J.W. 2010. A Better Way To Forecast Production From
Unconventional Gas Wells. Paper SPE 134231 presented at the SPE Annual Technical
Conference and Exhibition, Florence, Italy, 19-22 September.

208
Virieux, J., Flores-Luna, C. and Gibert, D., 1994. Asymptotic theory for diffusive
electromagnetic imaging. Geophysical Journal International, 119(3), pp.857-868.

Vasco, D. W., Keers, H., and Karasaki, K. 2000. Estimation of Reservoir Properties Using
Transient Pressure Data: An Asymptotic Approach. Water Resources Research 36 (12):
3447-3465.

Warpinski, N.R., Branagan, P.T., Peterson, R.E., Wolhart, S.L. and Uhl, J.E., 1998,
January. Mapping hydraulic fracture growth and geometry using microseismic events
detected by a wireline retrievable accelerometer array. In SPE Gas Technology
Symposium. Society of Petroleum Engineers.

Warpinski, N.R., Kramm, R.C., Heinze, J.R. and Waltman, C.K., 2005. Comparison of
Single-and Dual-Array Microseismic Mapping Techniques in the Barnett Shale. Paper
SPE 95568 presented at the SPE Annual Technology Conference and Exhibition, Dallas,
9–12 October.

Weibull, W. 1951. A Statistical Distribution Function of Wide Applicability. J. Appl.

Mech. 18: 293-297.

Xie, J., Yang, C., Gupta, N., King, M. J., & Datta-Gupta, A. (2015a). Depth of
Investigation and Depletion in Unconventional Reservoirs With Fast-Marching Methods.
Society of Petroleum Engineers. doi:10.2118/154532-PA

Xie, J., Yang, C., Gupta, N., King, M. J., Datta-Gupta, A. (2015b). Integration of Shale-
Gas-Production Data and Microseismic for Fracture and Reservoir Properties With the
Fast Marching Method. Society of Petroleum Engineers. doi:10.2118/161357-PA

209
Yang, C., Vyas, A., Datta-Gupta, A., Ley, S.B. and Biswas, P., 2017. Rapid multistage
hydraulic fracture design and optimization in unconventional reservoirs using a novel Fast
Marching Method. Journal of Petroleum Science and Engineering.

Yang, M., Valko, P.P. and Economides, M.J., 2012, March. Hydraulic Fracture Production
Optimization with a Pseudo-3D Model in Multi-layered Lithology. In SPE/EAGE
European Unconventional Resources Conference & Exhibition

Yin, J., Park, H., Datta-Gupta, A., & Choudhary, M. K. (2010). A Hierarchical Streamline-
Assisted History Matching Approach With Global and Local Parameter Updates. Society
of Petroleum Engineers. doi:10.2118/132642-MS

Yin, J., Xie, J., Datta-Gupta, A., & Hill, A. D. (2011). Improved Characterization and
Performance Assessment of Shale Gas Wells by Integrating Stimulated Reservoir Volume
and Production Data. Society of Petroleum Engineers. doi:10.2118/148969-MS

Zhang, Y., Yang, C., TETKing, M. J., & Datta-Gupta, A. (2013). Fast-Marching Methods
for Complex Grids and Anisotropic Permeabilities: Application to Unconventional
Reservoirs. Society of Petroleum Engineers. doi:10.2118/163637-MS

Zhang, Y., Bansal, N., Fujita, Y., Datta-gupta, A., King, M. J., & Sankaran, S. (2014).
From Streamlines to Fast Marching: Rapid Simulation and Performance Assessment of
Shale Gas Reservoirs Using Diffusive Time of Flight as a Spatial Coordinate. Society of
Petroleum Engineers. doi:10.2118/168997-MS

Zhang, Y., Neha., B., Fujita, Y., Datta-Gupta, A., King., M. and Sankaran, S. 2016. "From
Streamlines to Fast Marching: Rapid Simulation and Performance Assessment of Shale-
Gas Reservoirs by Use of Diffusivity Time of Flight as a Spatial Coordinate." SPEJ 21
(5): 1-16. http://dx.doi.org/10.2118/168997-PA.

210
Zhong M., Schuetter J., Mishra and S. LaFollette. 2015. Do Data Mining Methods Matter?
: A Wolfcamp “Shale” Case Study. Paper SPE 173334-MS presented at the SPE Hydraulic
Fracturing Technology Conference held in The Woodlands, Texas, USA, 3-5 February.

211
 APPENDIX A

This appendix describes how to regenerate figures and results presented in Chapter 2. This
is a standalone R application code. A new user needs to copy the R code folder named as
‘ML’ in C drive keeping the names of this folder and all the subfolders unchanged.

Prerequisites: As a prerequisite R needs to be installed on the user computer. R Studio

should be installed in order to edit code if needed. Also, some of the libraries needs to be
installed before running code.
Following list of libraries need to be downloaded/installed: ‘xlsx’, ‘GA’, ‘Metrics’,
‘randomForest’, ‘earth’, ‘e1071’, ‘MASS’, ‘glmnet’, ‘gbm’, ‘acepack’, ‘ggplot2’,
‘cvTools’, ‘neuralnet’, ‘class’, ‘maps’, ‘devtools’, ‘rpart.plot’, ‘FNN’, ‘reshape2’.
In order to install a library, go to R Studio menu bar and press Tools  Install
Packages. A window should be opened up where the needed library can be installed.
Another way to install more than one package is through R commands. An R script
file named as Install_Packages.R is provided with other R files. This file can be run in
order to install all packages needed.
In case a library is needed but not installed, R Studio should generate error in
console.

The contents of ML folder and their main job are:

1. DCA_Well_Data: This folder contains several excel sheets e.g., ‘DCA_100.xlsx’.
Each excel sheet belongs to a well and contains monthly rate data. The corresponding
well API number is also provided in each file. This folder also contains well
completion data of all wells in H_VAR_EXPORT_DCA.xlsx.

2. Output_Files: Output files of all the R script files are saved in this folder.

3. DCA_FIT_ARPS.R: This R script file reads the monthly rate data and completion

212
 data for each of the study wells in DCA_Well_Data folder and fits Arp’s decline
curves. It fits the best decline model parameters (‘Di’ and ‘b’) and predicts the
Estimated Ultimate Recovery (EUR) based on them. EUR is calculated for each well
based on 30 years of production using decline curve extrapolation. Each well’s initial
flow rate (taken as maximum flow rate) is also identified for monthly rate data and
referred to as ‘qi’ or initial flow rate in this study. Finally, the fitted decline model
parameters and the corresponding completion data e.g., no. of stages, proppant
amount, etc. (pulled from H_VAR_EXPORT_DCA.xlsx) for each well are stored in
an excel sheet named as ‘Model_data_ARPS.xlsx’. In this excel sheet, each row
corresponds to a well identified by a serial number. Wells are identified by their unique
serial number or well number. If needed, API number corresponding to a well serial
number can be retrieved from a well’s corresponding excel file in DCA_Well_Data
folder. It should be noted here that those wells with less than 12 months of production
history are not included in this study.

4. DCA_FIT_SEDM.R: This R script file has similar job to do as DCA_FIT_ARPS.R

except that it is trying to fit SEDM parameters (‘tau’ and ‘n’) instead of Arp’s
parameters. EUR is also calculated based on extrapolated SEDM curve.

5. DCA_FIT_DUONG.R: This R script file has similar job to do as DCA_FIT_ARPS.R

except that it is trying to fit Duong’s model parameters (‘a’ and ‘m’) instead of Arp’s
parameters. EUR is also calculated based on extrapolated DUONG curve.

6. DCA_FIT_WEIBULL.R: This R script file has similar job to do as

DCA_FIT_ARPS.R except that it is trying to fit Weibull’s model parameters
(‘gamma’, ‘alpha’ and ‘M’) instead of Arp’s parameters. EUR is also calculated based
on extrapolated WEIBULL curve.

7. DCA_Data_Clean.R: This R script file combines the output files of

DCA_FIT_ARPS.R, DCA_FIT_SEDM.R, DCA_FIT_DUONG.R and
213
 DCA_FIT_WEIBULL.R and generates a single file (Model_data.xlsx) which contains
decline curve parameters for each well. This file also generates boxplots for
distribution of various predictor variables in each of the 4 clusters clustered with
respect to Initial flow rate, qi. This file also generates bubble plots for various predictor
variables on Texas map. This file is also used to filter out outlier wells which have
unrealistic predictor/response values.

8. ML_Algorithms.R: This R script file fits one or more machine learning algorithms
selected by the user and builds models to predict decline model parameters. A user can
change some of the parameters as discussed below:

Figs. A.1 and A.2 shows the snapshots from R script file ML_Algorithms.R. These
snapshots show where exactly a user can change inputs.

Figure A.1 Input parameters in ML_Algorithms.R script – Part 1

214
 Figure A.2 Input parameters in ML_Algorithms.R script – Part 2

The explanation of various variables and their possible values are provided below:

DATA_FILE_PATH
This variable assigns the path of excel sheet containing all predictors and responses that
are needed various for machine learning algorithms. For e.g., for the current settings, it is
set to “C:\\ML\\Output_Files\\Model_data.xlsx”.

ML_ALGORITHMS
This variable assigns type of machine learning algorithm used. One or more algorithms
can be run at a time. E.g., c(“RF”, “SVM”, “MARS”) would run code for RF, SVM and
MARS in that order. In total, 12 machine learning algorithms are allowed.
Suggested Values: one or more of “RF”, “SVM”, “MARS”, “GBM”, “ACE”, “AVAS”,
“RIDGE”, “LASSO”, “ENET”, “KNN”, “ANN”, “LM”

Above acronyms stand for following machine learning algorithms:

RF: Random Forest

215
SVM: Support Vector Machine
MARS: Multivariate Adaptive Regression Splines
GBM: Gradient Boosting Machine
ACE: Alternative Conditional Expectations
AVAS: Additivity Variance Stabilization
RIDGE: Ridge Regression
LASSO: Least Absolute Shrinkage and Selection Operator
ENET: Elastic Net regression
KNN: K-Nearest Neighbors
ANN: Artificial Neural Network
LM: Linear Model

PREDICTORS_ALL
This variable assigns the list of predictor variables. These variables must be present in the
data file – “Model_data.xlsx”.
Suggested Values: For Chapter 1 study it is set to c(“PROP_TOTAL”,
“FRAC_FLUID_TOTAL”, “CLENGTH”, “STAGES”, “TVD_HEEL”,
“TVD_HEEL_TOE_DIFF”, “LONGITUDE”, “LATITUDE”, “qi”)

RESPONSES
Response variable to be predicted. Can be one or more variables.
Suggested values:
For ARPS, it can be set to “ARPS_Di”, “ARPS_b” or “ARPS_EUR”. Multiple response
variables can be predicted as in c(“ARPS_Di”, “ARPS_b”, “ARPS_EUR”)

For SEDM, it can be set to “SEDM_tau”, “SEDM_n” or “SEDM_EUR”. Multiple

responses can be predicted as in c(“SEDM_tau”, “SEDM_n”, “SEDM_EUR”)

216
For DUONG, it can be set to “DUONG_a”, “DUONG_m” or “DUONG_EUR”. Multiple
responses can be predicted as in c(“DUONG_a”, “DUONG_m”, “DUONG_EUR”)

For WEIBULL, it can be set to “WEIBULL_gamma”, “WEIBULL_alpha”,

“WEIBULL_M” or “WEIBULL_EUR”. Multiple responses can be predicted as in
c(“WEIBULL_gamma”, “WEIBULL_alpha”, “WEIBULL_M”, “WEIBULL_EUR”)

SCATTER_PLOT_AXIS_LIMIT
This variable sets minimum and maximum limits for the response variable for which
model training is being done.
Suggested Values: In the Eagle Ford study case, following values for this variable has
been used.

Table A.1 Axis scale values used for Eagle Ford plots

DCA Model Response SCATTER_PLOT_AXIS_LIMIT

ARPS_Di c(0,1)
ARPS ARPS_b c(0,1)
ARPS_EUR c(0,300)
SEDM_tau c(0,20)
SEDM SEDM_n c(0,2)
SEDM_EUR c(0,300)
DUONG_a c(0,3)
DUONG DUONG_m c(1,3)
DUONG_EUR c(0,300)
WEIBULL_gamma c(0,1)
WEIBULL WEIBULL_alpha c(0,20)
WEIBULL_M c(0,2e+5)
WEIBULL_EUR c(0,300)

217
IS_RI
If this script needs to be run for calculation of relative importance of various predictor
variables, this variable is set to “Y” otherwise “N”. In case the current run is for Relative
Influence calculations, ACE and AVAS algorithms need to be removed from the set of
ML_ALGORITHMS in input sections. For ML_ALGORITHMS, use one or more of
“RF”, “SVM”, “MARS”, “GBM”, “RIDGE”, “LASSO”, “ENET”, “KNN”, “ANN”,
“LM”

TRAIN_FRAC
The fraction of data points used for training purpose. The rest will be used for testing the
machine learning model.
Suggested values: 0.8

IS_NORM
This variable decides whether the data needs to be normalized for learning or not. The
final predictions are stored after de-normalizing the data.
Suggested Values: Choose “Y” if data needs to be normalized and choose “N” otherwise.

Note: If using “ANN” as the machine leaning algorithm, normalizing the data is
necessary. Therefore, IS_NORM should be set to “Y” if one of the machine learning
algorithms is “ANN”.

AVG_METHOD
This variable assigns the type of averaging algorithm used
Suggested values: “GLUE”, “MLBMA”, “AICMA” or “ARITHMETIC”

Each of these averaging keywords stand for different ways of assigning model weights to
be used for model averaging:

218
GLUE: Generalized Likelihood Uncertainty Estimation
MLBMA: Maximum Likelihood Bayesian Model Averaging
AICMA: Akaike Information Criterion Model Averaging
ARITHMETIC: All models are assigned equal weights. The averaging is based on
arithmetic average of all models.

NO_SEEDS
This variable assigns the number of seeds used to reshuffle the given training dataset.
Reshuffling the data would give different data points in k-folds that are generated during
model building and will generate extra set of models in model pool.

FOLDS_NO
This variable assigns the number of folds into which the training data is split. If this
number is high, smaller sets of data would lie in each fold. On the other hand, smaller
values would split training data into bigger sets of data in each fold.
Suggested Value: This is set to 5 or 10 most commonly. In the current settings, it is set to
10.

IS_SINGLE_MODEL
This variable indicates whether model averaging needs to be done or not. If it is set to “Y”,
then only the best model is used for final prediction for test data. If it is set to “N”, then
model averaging is done with corresponding weights of each model.

IS_CLUSTER
This variable decides if a machine learning is to be done for a particular cluster or not. If
it is set to “Y” data is divided into 4 clusters based on the variable name specified.
Suggested Values: “Y” or “N”.

219
CLUSTER_VARIABLE
The variable to be used to partition data into 4 clusters based on quartiles. This is useful
only if IS_CLUSTER is set to “Y”.
Suggested Values: This is assigned to one of the predictor variables, for e.g., “qi”.

CLUSTER_NO
This variable assigns the cluster number to be used for machine learning. This variable is
useful only if IS_CLUSTER is set to “Y” otherwise it is ignored and entire dataset is used.
Suggested Values: 1, 2, 3 or 4

NTREE
This variable is a tuning parameter for Random Forest model which is equal to the number
of trees used.
Suggested values: Usually a large number will help dealing with overfitting. For Eagle
Ford data, NTREE = 300 has been used.

MTRY_SEQ_VALUES
This variable is a tuning parameter for Random Forest and gives sequence of options for
number of predictor variables to be considered to partition data at each node of a tree in
Random Forest.
Suggested values: It is suggested to use all possible subsets of predictor variables. In
Eagle Ford data, since there are 9 predictors, MTRY_SEQ_VALUES is set to seq(from =
1,to = 9,by = 1) giving all possible subsets of predictor variables to be used at each node.

KERNEL_TYPES
This variable is a tuning parameter for SVM and assigns the kernel type(s) to be used for
SVM learning.
Suggested values: One or more of “linear”, “radial” and “polynomial” kernels are
suggested. In current settings all of them are assigned as a sequence - c(“linear”, “radial”,

220
“polynomial”). Multiple kernel types can be used for building multiple models for model
averaging.

COST_VALUES
This is a tuning parameter for SVM. It assigns the cost parameter for SVM. Changing cost
value can reduce overfitting.
Suggested values: In current settings, a sequence of cost values are provided as
seq(0.1,3,0.1) ranging between 0.1 and 3.0 in steps of 0.1.

DEGREE_VALUES
This variable is a tuning parameter for MARS. It sets possible degree values for MARS
model. Degree in MARS model controls the maximum degree of interaction. If degree is
set to 1, no interaction terms are included, i.e., an additive model is built.
Suggested values: In the current settings a range of degree values are given - seq(from =
1,to = 3,by = 1). Therefore degree can be 1, 2 or 3.

LAMBDA_VALUES
This is a tuning parameter for Ridge and LASSO regression models. This variable assigns
values for lambda which controls model regularization term.
Suggested values: In the current settings, it is in the sequence from 0 to seq(from = 0,to
= 0.01,by = 0.0001)

ALPHA_VALUES
This variable is a tuning parameter for Elastic Net (ENET) regression and assigns one or
more values for the Elastic Net mixing parameter, alpha.
Suggested values: In case of Elastic Net regression, alpha should lie between 0 and 1. In
current settings, it is within a range of 0.1 and 0.9 in the steps of 0.1, i.e., seq(from = 0.1,to
= 0.9,by = 0.1). In case alpha is set to 0, the model becomes Ridge regression and if alpha

221
is set to 1, model becomes LASSO regression. In case of Ridge or LASSO regression, the
corresponding alpha values are automatically used by the code and this variable is ignored.

NTREES_VALUES
This is a tuning parameter for GBM and it assigns the number of trees to fit. A single value
or a sequence of values may be provided.
Suggested values: In the current settings, this variable is assigned to a range of values
from 10000 to 30000 in steps of 10000, i.e., seq(from = 10000,to = 30000,by = 10000).

HIDDEN_VALUES
This is a tuning parameter for ANN model. This variable assigns number of neurons in a
hidden layer. It may be a single value or a sequence of possible values.
Suggested Values: In the current settings, this variable is set to have a sequence of
possible values ranging from 9 to 30 in steps of 3, i.e., seq(from = 9, to = 20, by = 3). A
large number of neurons may lead to over fitting.

HIDDEN_LAYERS_VALUES
This variable assigns the number of hidden layers in ANN network. In current code
settings, each hidden layer is set to be of equal number of neurons.
Suggested Values: In the current settings, it is set to a sequence from 1 to 3 in steps of 1s,
i.e., seq(from = 1, to = 3, by = 1). Larger number of layers may lead to over fitting.

THRESHOLD_VALUES
This variable assigns the threshold value for the partial derivatives of the error function as
stopping criteria. A small value may over fit model.
Suggested Values: In current settings, threshold is set to a range of values ranging from
0.1 to 10 in steps of 1.

222
KNN_VALUES
This is a tuning parameter for KNN regression. This variable assigns the number of nearest
neighbors considered.
Suggested Values: In current settings, a range of values from 1 to 10 in steps of 1 are
used, i.e., seq(from = 1, to = 10, by = 1).

MAX_TERMS_VALUES
This is a tuning parameter for LM (linear model) fitting. This variable sets maximum
number of terms in a linear model including interaction terms. In current code, up to three
way interactions are considered.
Suggested Values: More number of terms are likely to over fit model. In the current code
settings, it is set to a range of values from 20 to 30 in steps of 1, i.e., seq(from = 20, to =
30, by = 1).

9. DCA_Decline_Curves.R: This R script file plots the test data well decline curves
against actual rate data. In the input section of this R script file, user needs to specify
values for following variables:

DCA_METHOD
This variable assigns the decline model for which plots need to be generated.
Suggested Values: One of the decline models - “ARPS”, “SEDM”, “DUONG” or
“WEIBULL”

ML_ALGORITHM
The machine learning algorithm for which the decline model predictions have to be
plotted.
Suggested Values: One of the following algorithms:
“RF”, “SVM”, “MARS”, “GBM”, “ACE”, “AVAS”, “RIDGE”, “LASSO”, “ENET”,
“KNN”, “ANN”, “LM”

223
IS_CLUSTER
If the learning was done for each cluster, decline models would be plotted for each cluster
separately.
Suggested Values: “Y” or “N”

10. ERR_PLOTS_RELATIVE.R: This R script file plots the error bar plots for training
and test data predictions. Error plots are based on normalized RMSE, AAE or R2 errors
relative to the maximum value among all algorithms under investigation. Following
input variables need to be set before running this script.

ML_ALGORITHMS
This variable assigns the list of machine learning algorithms that need to be included in
error bar plots. Corresponding machine leaning algorithms need to be run before including
them in this list.
Suggested Values: One or more of following machine learning algorithms:
“RF”, “SVM”, “MARS”, “GBM”, “ACE”, “AVAS”, “RIDGE”, “LASSO”, “ENET”,
“KNN”, “ANN”, “LM”

RESPONSE
This variable needs to be assigned to the response variable for which error bars need to be
compared for different machine learning algorithms.
Suggested Values: E.g., “SEDM_EUR”, “ARPS_EUR”, etc.

11. ERR_PLOTS.R: This file does the same job as ERR_PLOTS_RELATIVE.R except
that it creates bar plots based on un-normalized errors.

12. RI_PLOTS: This R script file needs to be executed in order to generate relative
influence plots for the current study. Following variables need to be set before running
this file.

224
ML_ALGORITHMS
This variable needs to be set to a list of machine learning algorithms which need to be
included in relative influence.
Suggested Values: One or more of following machine learning algorithms:
“RF”, “SVM”, “MARS”, “GBM”, “ACE”, “AVAS”, “RIDGE”, “LASSO”, “ENET”,
“KNN”, “ANN”, “LM”

RESPONSES
This variable is assigned to the list of variables that need to be included in relative
influence plots.
Suggested Values: For. e.g., c(“ARPS_EUR”, “SEDM_EUR” ,”DUONG_EUR”)

RANKING_POLICY
This variable is assigned to the metric type used to calculate relative influence of a
variable.
Suggested Values: One of “RMSE_Test”, “AAE_Test” or “R2_Test”.

225