THE UNIVERSITY OF TULSA

THE GRADUATE SCHOOL

OIL FIELD DEVELOPMENT OPTIMIZATION

UNDER GEOLOGICAL UNCERTAINTY

by
Ranran Lu

A thesis submitted in partial fulfillment of

the requirements for the degree of Doctor of Philosophy
in the Discipline of Petroleum Engineering

The Graduate School

The University of Tulsa

2020
THE UNIVERSITY OF TULSA
THE GRADUATE SCHOOL

OIL FIELD DEVELOPMENT OPTIMIZATION

UNDER GEOLOGICAL UNCERTAINTY

by
Ranran Lu

A THESIS
APPROVED FOR THE DISCIPLINE OF
PETROLEUM ENGINEERING

By Thesis Committee

Albert C. Reynolds Jr., Chair

Rami Mustafa Younis
Mustafa Onur
Siamack A. Shirazi

ii
 COPYRIGHT STATEMENT

Copyright c 2020 by Ranran Lu

All rights reserved. No part of this publication may be reproduced, stored in a

retrieval system, or transmitted, in any form or by any means, electronic, mechanical, pho-
tocopying, recording, or otherwise, without the prior written permission of the author.

iii
 ABSTRACT

Ranran Lu (Doctor of Philosophy in Petroleum Engineering)

Oil Field Development Optimization Under Geological Uncertainty
Directed by Albert C. Reynolds Jr.
227 pp., Chapter 6: Discussions and Conclusions

(451 words)

Field development optimization is a critical part of the decision making process for
reservoir management. Many researchers have studied individual parts of this optimiza-
tion process, where the two most typical problems are the well control optimization and
well placement optimization where the number, types and drilling orders of the wells to be
drilled are assumed to be known which is not the case most of the time; hence, there is a
strong need to do research on the full field development optimization problem where the
number of wells, their types, drilling sequences, locations and control settings are considered
as optimization (design) variables. The full field development optimization problem is a
mixed integer problem. Common mathematical optimization algorithms, e.g. branch and
bound, for solving mixed integer problems are too computationally intensive to apply to field
optimization because each iteration of an optimization requires at least one run of a reservoir
simulation to evaluate the cost function, which in this work is represented by the net present
value (NPV) of production over the life of the reservoir. The full field development prob-
lem becomes even more computationally intensive when one considers geological uncertainty
where one generally performs robust optimization to maximize the average NPV of life-cycle
production over a set of Ne plausible reservoir models representing geological uncertainty.
For robust optimization, one needs to perform at least Ne reservoir simulation runs at each

iv
iteration of an optimization algorithm. Moreover, when geological uncertainty is considered,
it may be desirable to conduct bi-objective optimization to minimize the downside risk.
This research focuses on proposing parameterizations for field optimization, and de-
veloping efficient methodologies to solve the field development optimization problem. The
focus is a sequential optimization as an alternative to simultaneous optimization. The prob-
lems considered include (i) selecting a fixed number of wells from a given set of potential
drilling paths and determining the well types (injectors or producers), drilling order and con-
trol settings, (ii) optimizing number of wells, types and locations followed by optimizing well
controls, (iii) robust optimization of the second problem where a set of realizations is used
to characterize the geological uncertainty, and (iv) bi-objective well placement optimization
where the first objective is to maximize the average NPV and the second is to maximize the
minimum NPV of a given set of realizations. The methodologies focus on hybridizing the
Genetic Algorithm (GA), the Stochastic Simplex Approximate Gradient method (StoSAG)
and the General Pattern Search (GPS) to handle different types of optimization variables
and to combine global search ability and local search ability. The validity of the proposed
methods are tested mainly on waterflooding examples, considering vertical wells, slanted
wells and multi-segment wells subjected to the nonlinear well spacing constraints.

v
 ACKNOWLEDGEMENTS

I would like to express my deepest appreciation to my advisor, Dr. Albert C.

Reynolds, for his guidance, support and patience during my Ph.D. study. His integrity,
hard work and dedication will always be an example in my life. I would also like to extend
my appreciation to my committee members: Dr. Mustafa Onur, Dr. Rami Younis and
Dr. Siamack Shirazi for their advice and comments on my research. I am very grateful to
all the member companies of the TUPREP, Graduate School and the McDougall School of
Petroleum Engineering for all kinds of support provided during my Ph.D. study. I also would
like to show my thanks and appreciations to Judy, Ronda and Loreta for the administrative
help.
I deeply appreciate all the help, support and encouragement from my friends and
colleagues: Cintia, Sepide, Xin Li, Yuanshan, Bailian, Emilio, Guotong, Gosha, Javad, Reza,
Shuaiwei, Wen, Xin Liu, Ying, Yu, Yang, Zhe and Zhenyu, et al..
I would like to show my greatest gratitude to my parents, my siblings and my husband,
to whom I dedicate this dissertation, for their unconditional support and love. I am truly
thankful for having them in my life.

vi
 TABLE OF CONTENTS

COPYRIGHT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv

ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

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

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xxix

CHAPTER 1: INTRODUCTION 1
1.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Research Objectives and Contextures . . . . . . . . . . . . . . . . . . . 21
1.3 Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

CHAPTER 2: COMPUTATIONAL ALGORITHMS FOR FIELD DEVEL-

OPMENT OPTIMIZATION 26
2.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.2 Basic Methodology Considered . . . . . . . . . . . . . . . . . . . . . . . 32
2.2.1 Genetic Algorithm (GA) . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2.2 Stochastic Simplex Approximate Gradient (StoSAG) . . . . . . . . . . 35
2.2.3 General Pattern Search (GPS) . . . . . . . . . . . . . . . . . . . . . . 39

CHAPTER 3: SELECTING WELLS FROM A GIVEN SET OF POTEN-

TIAL PATHS 43
3.1 Problem Description and Methodology . . . . . . . . . . . . . . . . . . 43
3.1.1 Optimization of well locations, types and drilling order . . . . . . . . 47
3.1.2 Optimization of well controls u . . . . . . . . . . . . . . . . . . . . . 54
3.2 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2.1 Example 1: channelized reservoir . . . . . . . . . . . . . . . . . . . . 56
3.2.2 Example 2: Brugge model . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2.3 Investigation on drilling order effect . . . . . . . . . . . . . . . . . . . 83
3.2.4 Investigation on population size . . . . . . . . . . . . . . . . . . . . . 86
3.3 Comparison with StoSAG Using Priority Parametrization . . . . . . 89
3.3.1 StoSAG using priority parametrization . . . . . . . . . . . . . . . . . 89

vii
 3.3.2 Numerical examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.4 Comparison with a Gradient-based Algorithm . . . . . . . . . . . . . 94
3.4.1 A gradient-based algorithm . . . . . . . . . . . . . . . . . . . . . . . . 94
3.4.2 Optimization of (x, T , u) using GA . . . . . . . . . . . . . . . . . . . 97
3.4.3 Numerical example: channelized reservoir . . . . . . . . . . . . . . . . 100
3.5 Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

CHAPTER 4: JOINT OPTIMIZATION OF WELL NUMBER, TYPES AND

LOCATIONS 104
4.1 Problem Description and Methodology . . . . . . . . . . . . . . . . . . 104
4.1.1 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.1.2 Constraints handling . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.1.3 Joint optimization of well status, types and locations . . . . . . . . . 108
4.2 Numerical Examples of Deterministic Optimization . . . . . . . . . . 115
4.2.1 Example 1: PUNQ-S3 model . . . . . . . . . . . . . . . . . . . . . . . 116
4.2.2 Example 2: Brugge model . . . . . . . . . . . . . . . . . . . . . . . . 128
4.3 Field Development Optimization Considering Geological
Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.4 Numerical Examples of Robust Optimization . . . . . . . . . . . . . . 142
4.4.1 Example 1: PUNQ model . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.4.2 Example 2: Brugge model . . . . . . . . . . . . . . . . . . . . . . . . 151
4.5 Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

CHAPTER 5: WELL PLACEMENT OPTIMIZATION WITH COMPLI-

CATED WELL TRAJECTORIES 161
5.1 Well Trajectory Parameterization . . . . . . . . . . . . . . . . . . . . . 161
5.1.1 Slanted well trajectory parameterization . . . . . . . . . . . . . . . . . 161
5.1.2 Multi-segmented well trajectory parameterization . . . . . . . . . . . . 163
5.2 Constrained Optimization using Augmented Lagrangian Method . 168
5.3 Bi-objective Well Placement Optimization . . . . . . . . . . . . . . . . 172
5.4 Numerical Example: Channelized Reservoir . . . . . . . . . . . . . . . 177
5.4.1 Deterministic optimization using realization #1 . . . . . . . . . . . . 180
5.4.2 Bi-objective optimization using 10 realizations . . . . . . . . . . . . . 187
5.5 Numerical Example: Oseberg Reservoir . . . . . . . . . . . . . . . . . 194
5.5.1 Deterministic optimization using realization #1 . . . . . . . . . . . . 197
5.5.2 Bi-objective optimization using 10 realizations . . . . . . . . . . . . . 199
5.6 Summary and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 202

CHAPTER 6: DISCUSSION AND CONCLUSIONS 203

NOMENCLATURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

APPENDIX A: MINIMUM WELL SPACING CONSTRAINT . . . . . . . . 216

APPENDIX B: MULTI-SEGMENTED WELL TRAJECTORY CALCULA-
TION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219

viii
APPENDIX C: EQUIVALENT WELL INDEX CALCULATION . . . . . . . 221
C.1 Calculation of the Inputs for CMG . . . . . . . . . . . . . . . . . . . . 221
C.2 Two-segmented Perforation within a Gridblock . . . . . . . . . . . . . 225

ix
 LIST OF TABLES

3.1 Petrophysical parameters for each facies, Example 1, i.e., selecting Nw = 8

wells given Kmax = 30 potential locations in a channelized reservoir. . . . . . 59
3.2 Perforated layers of the horizontal wells, Example 1. . . . . . . . . . . . . . . 59
3.3 Fixed BHP’s used in the optimization of (x, T , O) and initial guesses used in
the subsequent control optimization, Bound A, Bound A→B and Bound B for
Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4 Optimal solutions obtained at different number of generations for the joint
optimization of well locations and types using seed 2, Bound A, Example 1. . 62
3.5 Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Bound A, Example 1. For seed 1 and seed 3 results, wells that are
different from the solution of seed 2 are marked in red. . . . . . . . . . . . . . 65
3.6 Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Bound B, Example 1. Injectors are in bold. . . . . . . . . . . . . 65
3.7 Effects of initial guesses and bounds on sequential optimization of well locations
and types, drilling order (i.e., (x, T ) → O) and controls using three seeds,
Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.8 Effects of initial guesses and bounds on simultaneous optimization of well loca-
tions and types, drilling order (i.e., (x, T , O)) and controls using three seeds,
Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.9 Optimal solutions obtained at different number of generations for the opti-
mization of well locations and types using seed 3, Example 2-A. . . . . . . . . 73

x
3.10 Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Example 2-A. Wells that are different from the solution of seed 3
are in red. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.11 Comparison of the optimal NPV’s obtained with optimization of (x, T ) → O
and (x, T , O) followed by control optimization, Example 2-A. . . . . . . . . . 75
3.12 Comparison of the optimal NPV’s obtained with sequential optimization of
(x, T ) → (O) followed by control optimization, i.e., (x, T ) → (O) → (u),
Example 2-A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.13 Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Bound A, Example 2-A. Wells that are different from the solution
of seed 3 in Table 3.10 are marked in red. . . . . . . . . . . . . . . . . . . . . 76
3.14 Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Example 2-B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.15 Comparison of the optimal NPV’s obtained with optimization of (x, T ) → O
and (x, T , O) followed by control optimization, Example 2-B. . . . . . . . . . 80
3.16 Comparison of the optimal NPV’s obtained with simultaneous optimization of
(x, T , O) followed by control optimization, i.e., (x, T , O) → (u), Example 2-C. 82
3.17 Optimal drilling order within different reservoir lives using the optimal well
locations and types (x, T ) obtained with seed 1, Example 2-B. . . . . . . . . 84
3.18 Summary of NPV’s obtained with drilling order optimization within different
reservoir lives using the optimal well locations and types (x, T ) obtained with
seed 1, Example 2-B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.19 Summary of NPV’s obtained with drilling order optimization within different
drilling time Td using the optimal well locations and types (x, T ) obtained
with seed 3, Example 2-A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.20 Optimal drilling order with different drilling time Td using the optimal well
locations and types (x, T ) obtained with seed 3, Example 2-A. . . . . . . . . 86

xi
3.21 Best NPV’s obtained by optimizing well locations and types with about 2000
reservoir simulation runs using different population sizes, Example 2-A. . . . 88
4.1 Comparison of the optimal NPV’s obtained with three different algorithms
using three seeds, PUNQ model. . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.2 Comparison of simulation cost obtained with three different algorithms using
three seeds, PUNQ model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.3 Comparison of optimal well numbers obtained with three different algorithms
using three seeds, PUNQ model. . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.4 Comparison of optimal NPV’s obtained with three different algorithms using
three seeds, Brugge model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.5 Comparison of the simulation cost for the three different algorithms using three
seeds, Brugge model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.6 Comparison of optimal well numbers obtained with three different algorithms
using three seeds, Brugge model. . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.7 Summary of the optimal solutions and computational cost obtained with the
three robust optimization procedures using three seeds, robust optimization of
PUNQ model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.8 Summary of the optimal solutions and computational cost obtained with the
three robust optimization procedures with GA and GPS, robust optimization
of PUNQ model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.9 Summary of the optimal solutions and computational cost obtained with three
robust optimization procedures, robust optimization of the Brugge model. . . 155
4.10 Summary of the optimal solutions and computational cost obtained with the
three robust optimization procedures using the iterative sequential method
with GA and GPS search steps, robust optimization of Brugge model. . . . . 159

xii
5.1 Summary of the trajectories of all 8 wells obtained with the initial and optimal
well placements after slanted well trajectory optimization followed by multi-
segmented trajectory optimization using seed 3, deterministic optimization
using realization #1, channelized model. . . . . . . . . . . . . . . . . . . . . . 184
5.2 Summary of the trajectories of all 8 wells obtained for multi-segmented tra-
jectory optimization using seed 1, deterministic optimization using realization
#1, channelized model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
5.3 Petrophysical property parameters, Oseberg model. . . . . . . . . . . . . . . 195

xiii
 LIST OF FIGURES

1.1 Schematic diagram of closed-loop field development management. . . . . . . . 2

2.1 Schematic of GA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 Illustration of the StoSAG search direction. . . . . . . . . . . . . . . . . . . . 39
2.3 Illustration of the General Pattern Search (GPS) search direction. . . . . . . 42
3.1 SUS selection for the GA algorithm where A, B, C, D, E, F, G, H represent the
ranked individuals in a population with their lengths proportional to their
ranks in descending order; Fp represents the sum of ranks for all individuals
and N represents the number of parents to keep in the parent pool. . . . . . 49
3.2 Illustration of the crossover operation for the mixed encoding chromosomes
(x, T , O). The red and blue colors represent (x, T ) of the two parent chro-
mosomes respectively. The green and orange colors represent (O) of the two
parent chromosomes, respectively. . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3 Illustration of one-point crossover. The red and blue colors represent x of two
parent chromosomes respectively. . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.4 Illustration of the original one-point crossover and the modified one-point
crossover. The red and blue colors represent x of the two parent chromo-
somes respectively. The yellow color corresponds to a specific gene unit xbj , j =
1, 2, · · · , Nw . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 The full set of potential well locations and trajectories projected to the hori-
zontal permeability field, Example 1. . . . . . . . . . . . . . . . . . . . . . . . 59
3.6 NPV versus generations obtained with joint optimization of well locations and
types using three seeds, Bound A, Example 1. . . . . . . . . . . . . . . . . . 62

xiv
3.7 NPV versus generations obtained with the optimization of drilling order using
three seeds, Bound A, Example 1. . . . . . . . . . . . . . . . . . . . . . . . . 62
3.8 Optimal solutions obtained with sequential optimization of (x, T ) → O, using
three seeds, plotted on the horizontal permeability field (upper row) and the
oil saturation field (bottom row) of layer 7 at end of 4,000 days of production,
Bound A, Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.9 Optimal solutions obtained with sequential optimization of (x, T ) → O, using
three seeds, plotted on the horizontal permeability field (upper row) and the
oil saturation field (bottom row) of layer 7 at end of 4,000 days of production,
Bound B, Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.10 Optimal controls for the well locations, types and drilling order obtained with
sequential optimization of (x, T ) → O using seed 1, Bound A, Example 1. . . 66
3.11 Optimal controls for the well locations, types and drilling order obtained with
sequential optimization of (x, T ) → O using seed 2, Bound A→B, Example 1. 67
3.12 Optimal controls for the well locations, types and drilling order obtained with
sequential optimization of (x, T ) → O using seed 2, Bound B, Example 1. . . 67
3.13 Top structure of the Brugge model with the original 30 wells. Left: 3D struc-
ture plotted on the initial oil saturation field. Right: 2D depth contour of layer
1 where wells with label 21 to 30 are the locations of the original injectors and
wells with label 1 to 20 are the locations of the original producers, Example 2. 69
3.14 The log-horizontal permeability (ln kx ) field and porosity (φ) field of the Brugge
model, Example 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.15 NPV versus generations obtained with the optimization of well locations and
types, i.e., (x, T ), using three seeds, Example 2-A, i.e., selecting Nw = 12 wells
given Kmax = 30 potential locations. . . . . . . . . . . . . . . . . . . . . . . . 73
3.16 NPV versus generations obtained with drilling order optimization using three
seeds, Example 2-A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

xv
3.17 Well configurations and oil saturation fields of layer 1. (a) shows the set of
initial potential locations plotted on the oil saturation field; (b)-(d) show the
optimal well configurations obtained with sequential optimization of (x, T ) →
O for three seeds, with the corresponding oil saturation fields after 20 years
of production on the top layer, Example 2-A. ⊗ represents an injector and o
represent a producer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.18 NPV versus generations obtained with the optimization of well locations and
types using three seeds, Example 2-B, i.e., selecting Nw = 12 wells given
Kmax = 64 potential locations in a channelized reservoir. . . . . . . . . . . . . 79
3.19 NPV versus generations obtained with drilling order optimization using three
seeds, Example 2-B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.20 Well configurations and oil saturation fields of layer 1. (a) shows the set of
potential locations plotted on the initial oil saturation field; (b)-(d) show the
optimal well configurations obtained with sequential optimization of (x, T ) →
O for three seeds, with the corresponding oil saturation fields after 20 years
of production on the top layer, Example 2-B. ⊗ represents an injector and o
represent a producer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.21 Well configurations and oil saturation fields of layer 1. (a) shows the set
of potential locations plotted on the initial oil saturation field; (b)-(d) show
the optimal well configurations obtained with simultaneous optimization of
(x, T , O) for three seeds, with the corresponding oil saturation fields after 20
years of production on the top layer, Example 2-C. ⊗ represents an injector
and o represent a producer. The first 12 wells drilled are marked in black color
and the 8 infill wells are marked in blue color. . . . . . . . . . . . . . . . . . 82
3.22 NPV versus generations obtained with drilling order optimization within dif-
ferent reservoir lives using the optimal well locations and types (x, T ) obtained
with seed 1, Example 2-B. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

xvi
3.23 NPV versus generations obtained with drilling order optimization with differ-
ent drilling time Td using the optimal well locations and types (x, T ) obtained
with seed 3, Example 2-A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.24 Effect of population size on GA performance from Lobo et al. [51]. . . . . . . 87
3.25 NPV versus generations obtained for the optimization of well types and loca-
tions using three seeds, Np =24, Example 2-A. . . . . . . . . . . . . . . . . . . 87
3.26 NPV versus generations obtained for the optimization of well types and loca-
tions using three seeds, Np =72, Example 2-A. . . . . . . . . . . . . . . . . . . 88
3.27 NPV versus generations obtained for the optimization of well types and loca-
tions using three seeds, Np =108, Example 2-A. . . . . . . . . . . . . . . . . . 88
3.28 NPV versus iterations for the drilling order optimization process starting from
four different initial guesses (Fig. 3 in [44]). . . . . . . . . . . . . . . . . . . . 90
3.29 NPV versus iterations for the drilling order optimization of 4 injectors and
8 producers (Fig. 6 in [44].) Trial update using best performing perturbed
priority sample is shown in red line where individual samples are shown as gray
dots. Trial updates obtained with line search using StoSAG gradient estimated
from normal distributed perturbations is shown in blue line. Trial updates
obtained with line search using StoSAG gradient estimated from Bernoulli
perturbations is shown in green line. . . . . . . . . . . . . . . . . . . . . . . 90
3.30 NPV versus generations obtained with simultaneous optimization of (x, T , O)
using three seeds, Bound A, Example 1. . . . . . . . . . . . . . . . . . . . . . 92
3.31 NPV versus iterations obtained with simultaneous optimization of (x, T , O)
following the methodology of [44] using three seeds, Example 1. The green
square represents a successful update of the best performing perturbation, the
blue star represents a successful update along the StoSAG search direction,
and the red circle represents a iteration where no successful updates is obtained. 93
3.32 NPV versus generations obtained with simultaneous optimization of (x, T , O)
using three seeds, Example 2-A. . . . . . . . . . . . . . . . . . . . . . . . . . 93

xvii
3.33 NPV versus iterations obtained with simultaneous optimization of (x, T , O)
following the methodology of [44] using three seeds, Example 2-A. The green
square represents a successful update of the best performing perturbation, the
blue star represents a successful update along the StoSAG search direction,
and the red circle represents a iteration where no successful updates is obtained. 93
3.34 Illustration of the logistic function and its derivative in Eq. 3.14. . . . . . . . 96
3.35 NPV and number of wells versus simulation runs for the joint optimization
of well locations, types and rates obtained with the gradient-based method,
Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.36 NPV versus generations for the joint optimization of well locations, types and
rates obtained with GA, Example 1. . . . . . . . . . . . . . . . . . . . . . . . 101
3.37 Optimal well configurations obtained with the gradient-based method and GA
method respectively, plotted on the oil saturation fields of layer 7 at the end
of 4,000 days, Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.38 Optimal well injection/production rates obtained with the gradient-based method
and GA, Example 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.1 Constraint handling when a well is placed in an inactive gridblock. . . . . . . 107
4.2 The top layer of PUNQ model. Left: the depth of the top layer. Right:
the initial pressure of the top layer. The black curve delineates the water-oil
contact and the red curve delineates the gas-oil contact. . . . . . . . . . . . . 117
4.3 Locations of 20 wells in the reference scenario (left) and their locations after
well placement optimization (right), plotted on the top layer of the PUNQ
model. Injector locations are marked in brown color and labeled with ⊗;
producer locations are marked in yellow color and labeled with o; the gas cap
is marked in green color; the aquifer is marked in dark blue and the oil zone
is marked in light blue color. . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

xviii
4.4 NPV versus the number of simulation runs in the well placement optimization
of the 20 wells in the reference scenario of the PUNQ model. o’s represent the
StoSAG iterations and ∗’s represent the GPS iterations. . . . . . . . . . . . . 118
4.5 NPV versus number of simulations for the well placement optimization of the
reference scenario, using GPS with three different initial mesh sizes, PUNQ
model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.6 NPV versus number of simulation runs during the well control optimization
using the optimal well locations of the reference scenario, PUNQ model. The
red dashed line represents the optimal NPV’s obtained after the well placement
optimization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.7 Optimal controls obtained for the reference scenario with optimal well loca-
tions, PUNQ model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.8 The set of potential locations (marked in yellow and red colors) for the initial
population, plotted on the top layer of the PUNQ model. . . . . . . . . . . . 123
4.9 Comparison of the optimal well numbers obtained with three different algo-
rithms using three seeds, PUNQ model. “W1-(s,T,x) and “W1-x” respec-
tively represent the first-stage and the second-stage optimization of the itera-
tive sequential algorithm, “W2-(s,T,x) and “W2-x” respectively represent the
first-stage and the second-stage optimization of the mixed-encoded GA, “W3-
(s,T,x) and “W3-x” respectively represent the first-stage and the second-stage
optimization of the Genocop III method, and the black dashed line represents
the NPV of the reference scenario. . . . . . . . . . . . . . . . . . . . . . . . . 125
4.10 NPV’s versus the number of simulation runs obtained with the best seed of
each of the three algorithms, PUNQ model. In the iterative sequential algo-
rithm, ∗’s represent the NPV’s of current population re-evaluated with the
updated locations after a StoSAG search step, ∗’s represent the NPV’s of cur-
rent population re-evaluated with the updated locations after a GPS search
step. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

xix
4.11 Comparison of optimal well numbers obtained with three different algorithms
using three seeds, PUNQ model. “W1” represents the iterative sequential
algorithm, “W2” represents the mixed-encoded GA, and “W3” represents the
Genocop III method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.12 Optimal well locations obtained with the three algorithms using the best seed,
plotted on the top layer of the PUNQ model. . . . . . . . . . . . . . . . . . . 127
4.13 Optimal NPV’s obtained with different drilling cost using the mixed-encoded
GA algorithm, PUNQ model. . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.14 Optimal number of wells obtained with different drilling cost using the mixed-
encoded GA algorithm, PUNQ model. . . . . . . . . . . . . . . . . . . . . . . 128
4.15 Locations of the 30 wells in the reference scenario (left) and their locations after
well placement optimization (right), plotted on the top layer of the Brugge
model. ⊗ represents an injector; o represents a producer. . . . . . . . . . . . 130
4.16 NPV versus the number of simulation runs in the well placement optimization
of the 30 wells in the reference scenario of the Brugge model. o’s represent the
StoSAG iterations and ∗’s represent the GPS iterations. . . . . . . . . . . . . 130
4.17 Comparison of optimal well numbers obtained with three different algorithms
using three seeds, Brugge model. “W1-(s,T,x) and “W1-x” respectively repre-
sent the first-stage and the second-stage optimization of the iterative sequen-
tial algorithm, “W2-(s,T,x) and “W2-x” respectively represent the first-stage
and the second-stage optimization of the mixed-encoded GA, “W3-(s,T,x) and
“W3-x” respectively represent the first-stage and the second-stage optimiza-
tion of the Genocop III method, and the black dashed line represent the NPV
of the reference scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.18 NPV’s versus the number of simulations obtained with the best seed of each
of the three algorithms, Brugge model. In the iterative sequential algorithm,
∗’s and ∗’s respectively represent the NPV’s of current population re-evaluated
with the updated locations after a StoSAG search step and a GPS search step. 133

xx
4.19 Comparison of the optimal well numbers obtained with three different algo-
rithms using three seeds, Brugge model. “W1” represents the iterative sequen-
tial algorithm, “W2” represents the mixed-encoded GA, and “W3” represents
the Genocop III method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.20 Optimal well locations obtained with the three algorithms using the best seed,
plotted on the oil saturation field of the top layer after 20 years of production,
Brugge model. ⊗ represents an injector; o represents a producer. . . . . . . . 134
4.21 Cdf’s of NPV obtained with the reference scenario (black curve) and the op-
timal well placements for the three robust procedures using three seeds (blue
curve for seed 1, red curve for seed 2 and green curve for seed 3), robust op-
timization of the PUNQ model. The green, blue and red circles respectively
represent the P2, P50 and P100 of the cdf of NPV obtained with the reference
scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.22 Realizations of the log-horizontal permeability fields (upper row) and the log-
vertical permeability fields (bottom row), corresponding to P2, P50 and P100
of the cdf of NPV obtained with the reference well placement scenario, PUNQ
model. kh and kv represent the horizontal and vertical permeability respec-
tively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.23 Expected NPV of the susbet ensemble versus number of simulation runs ob-
tained with Robust-1 using three seeds (blue curve for seed 1, red curve for seed
2 and green curve for seed 3), robust optimization of the PUNQ model. The
solid lines represent the expected NPV’s of the subset ensemble; the dashed
lines in color represent the optimal expected NPV’s of the full ensemble; the
black dashed line is the expected NPV of the full ensemble obtained with the
reference scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

xxi
4.24 Expected NPV of the full ensemble versus number of simulations obtained for
Robust-2 using three seeds, robust optimization of PUNQ model. o represents
an update by GA, ∗ represents an update by StoSAG, and  represents an
update by GPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.25 Expected NPV versus number of simulations obtained with Robust-2 using
seed 2, robust optimization of PUNQ model. o, ∗, and  respectively represent
the expected NPV of the full ensemble obtained using GA, StoSAG and GPS; .
and . respectively represent the expected NPV of the subset ensemble obtained
using StoSAG and GPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
4.26 Expected NPV of the full ensemble versus number of simulations obtained for
Robust-3 using three seeds, robust optimization of PUNQ model. o represents
an update by GA, ∗ represents an update by StoSAG, and  represents an
update by GPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.27 Optimal types and locations of drilled wells obtained with Robust-1 using three
seeds, robust optimization of the PUNQ model. Injector locations are marked
in brown color and labeled with ⊗; producer locations are marked in yellow
color and labeled with o; the gas cap is marked in green color; the aquifer is
marked in dark blue and the oil zone is marked in light blue color. . . . . . . 148
4.28 Optimal types and locations of drilled wells obtained for Robust-2 using three
seeds, robust optimization of PUNQ model. Injector locations are marked in
brown color and labeled with ⊗; producer locations are marked in yellow color
and labeled with o; the gas cap is marked in green color; the aquifer is marked
in dark blue and the oil zone is marked in light blue color. . . . . . . . . . . . 148
4.29 Optimal types and locations of drilled wells obtained for Robust-3 using three
seeds, robust optimization of PUNQ model. Injector locations are marked in
brown color and labeled with ⊗; producer locations are marked in yellow color
and labeled with o; the gas cap is marked in green color; the aquifer is marked
in dark blue and the oil zone is marked in light blue color. . . . . . . . . . . . 149

xxii
4.30 Expected NPV of the full ensemble versus number of simulations obtained
for Robust-2 using the iterative sequential method with GA and GPS steps,
robust optimization of the PUNQ model. o’s and ’s respectively represent
updates by GA and GPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.31 Expected NPV of the full-ensemble versus number of simulations obtained for
Robust-3 using the iterative sequential method with GA and GPS steps, robust
optimization of PUNQ model. o’s and ’s respectively represent updates by
GA and GPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.32 Cdf’s of NPV’s obtained with the reference scenario (black curve) and the
optimal well placement for three robust optimization procedures, robust op-
timization of Brugge model. The green, blue and red circles, respectively,
represent the P2, P50 and P100 of the cdf of NPV obtained with the reference
scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
4.33 Petrophysical parameters for the top layer of the realizations corresponding
to P2, P50 and P100 of the cdf of NPV obtained with the reference well
placement scenario, Brugge model. kh and kv represents the horizontal and
vertical permeability respectively, φ is the porosity, N T G is the net to gross
ratio and Swcon is the connate water saturation. . . . . . . . . . . . . . . . . . 153
4.34 Expected NPV of the subset ensemble obtained with the best candidate solu-
tion of a population versus number of simulation runs obtained with Robust-1,
robust optimization of Brugge model. The solid line in blue represents the ex-
pected NPV’s of the subset ensemble, the blue dashed line represents optimal
expected NPV of the full ensemble, and the black dashed line represents the
expected NPV of the full ensemble obtained with the reference scenario. . . 155
4.35 Expected NPV versus number of simulations obtained with Robust-2, robust
optimization of the Brugge model. o, ∗ respectively represent the expected
NPV of the full ensemble obtained using GA and StoSAG; . represents the
expected NPV of the subset ensemble obtained using StoSAG. . . . . . . . . 156

xxiii
4.36 Expected NPV of the full ensemble versus number of simulations obtained for
Robust-3, robust optimization of the Brugge model. o represents an update
by GA, ∗ represents an update by StoSAG, and  represents an update by GPS.156
4.37 Optimal types and locations of drilled wells obtained for Robust-1, plotted
on the oil saturation field at the top layer of realization #14 after 20 years of
production, robust optimization of the Brugge model. ⊗ represents an injector
and o represents a producer. . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
4.38 Optimal types and locations of drilled wells obtained for Robust-2, plotted
on the oil saturation field at the top layer of realization #14 after 20 years of
production, robust optimization of the Brugge model. ⊗ represents an injector
and o represents a producer. . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
4.39 Optimal types and locations of drilled wells obtained for Robust-3, plotted
on the oil saturation field at the top layer of realization #14 after 20 years of
production, robust optimization of the Brugge model. ⊗ represents an injector
and o represents a producer. . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.40 Cdf’s of NPV’s obtained with the reference scenario (black curve) and optimal
well placement for the three robust optimization procedures with GA and GPS
search steps, robust optimization of Brugge model. The green, blue and red
circles respectively represent the P2, P50 and P100 of the cdf of NPV obtained
with the reference scenario. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.41 The expected NPV of the full ensemble versus number of simulation runs
obtained for Robust-2 using the iterative sequential method with GA and GPS
steps, robust optimization of Brugge model. o’s and ’s respectively represent
updates by GA and GPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
4.42 The expected NPV of the full ensemble versus number of simulation runs
obtained for Robust-3 using the iterative sequential method with GA and GPS
steps, robust optimization of Brugge model. o’s and ’s respectively represent
updates by GA and GPS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

xxiv
5.1 Illustration of well trajectory parameterization for a slanted and a multi-
segmented well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.2 Illustration of the smoothed trajectory of two consecutive well segments for a
multi-segmented well. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
5.3 Projections of the initial well trajectories on the horizontal permeability field
(upper row) and the porosity field (bottom row) of layer 7 for three different
realizations, channelized model. . . . . . . . . . . . . . . . . . . . . . . . . . . 179
5.4 NPV versus number of simulation runs obtained for slanted well trajectory op-
timization followed by multi-segmented trajectory optimization, deterministic
optimization using realization #1, channelized model. The solid lines repre-
sent the slanted well trajectory optimization (StoSAG in blue and GPS in red)
and the dashed lines represent the subsequent multi-segmented well trajectory
optimization (StoSAG in blue and GPS in red). . . . . . . . . . . . . . . . . . 182
5.5 Cumulative oil production, water production and water injection obtained with
the initial and optimal well placements after slanted well trajectory optimiza-
tion followed by multi-segmented trajectory optimization using seed 3, deter-
ministic optimization using realization #1, channelized model. . . . . . . . . 183
5.6 Projections of the initial and optimal well placements after slanted well tra-
jectory optimization followed by multi-segmented trajectory optimization on
layer 7 using seed 3, deterministic optimization using realization #1, channel-
ized model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
5.7 Oil saturation after 3,000 days of production plotted on layer 7, obtained
with the initial and optimal well placements after slanted well trajectory op-
timization followed by multi-segmented trajectory optimization using seed 3,
deterministic optimization using realization #1, channelized model. . . . . . . 185

xxv
5.8 Projections of the initial and optimal injector perforations obtained with slanted
followed by multi-segmented trajectory optimization using seed 3, onto the
x − z plane of the horizontal permeability field, deterministic optimization
using realization #1, channelized model. . . . . . . . . . . . . . . . . . . . . . 186
5.9 NPV versus number of simulations obtained for multi-segmented trajectory op-
timization, deterministic optimization using realization #1, channelized reser-
voir. The blue lines represent the StoSAG algorithm and the red lines represent
the GPS algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
5.10 Cumulative oil production, water production and water injection obtained for
the initial and optimal well placements after multi-segmented well trajectory
optimization using seed 1, deterministic optimization using realization #1,
channelized model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
5.11 Projections of the initial and optimal well placement after multi-segmented
trajectory optimization onto layer 7 using seed 1, deterministic optimization
using realization #1, channelized model. . . . . . . . . . . . . . . . . . . . . . 188
5.12 Oil saturation after 3,000 days of production plotted on layer 7, obtained
for the initial and optimal well placement after multi-segmented trajectory
optimization using seed 1, deterministic optimization using Realization #1,
channelized model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
5.13 Projections of the initial and optimal injector perforations obtained with multi-
segmented well trajectory optimization using seed 1, onto the x − z plane of
the horizontal permeability field, deterministic optimization using Realization
#1, channelized model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189

xxvi
5.14 Bi-objective well placement optimization using 10 realizations, channelized
model. (a) Average NPV over 10 realizations versus number of simulation
runs in primary optimization using StoSAG (blue) and GPS (red); (b) the
augmented Lagrangian function, LA , minimum NPV, JR , and average NPV,
JE , versus number of simulation runs in secondary optimization using StoSAG
(blue) and GPS (red) where LA is represented by marker x, JR is represented
by the solid line, JE is represented by the dashed line and JE (vE∗ ) is represented
by the dotted line; (c) Cdf’s of NPV’s for the initial well placement and the
optimal well placement after primary and secondary optimization. . . . . . . 191
5.15 Number of violated constraints and indicies of the risk realization during the
secondary optimization of the bi-objective optimization of slanted well trajec-
tories using 10 realizations, channelized model. . . . . . . . . . . . . . . . . . 191
5.16 Cumulative oil production, water production and water injection obtained af-
ter the primary and secondary optimization of the bi-objective optimization,
channelized model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
5.17 NPV versus number of simulation runs for the deterministic optimization us-
ing realization #8, channelized model. The blue line represents the StoSAG
algorithm and the red line represents the GPS algorithm. . . . . . . . . . . . 193
5.18 Bi-objective well placement optimization using 10 realizations using a different
seed, channelized model. (a) Average NPV over 10 realizations versus number
of simulation runs in primary optimization using StoSAG (blue) and GPS
(red); (b) the augmented Lagrangian function, LA , minimum NPV, JR , and
average NPV, JE , versus number of simulation runs in secondary optimization
using StoSAG (blue) and GPS (red) where LA is represented by marker x, JR is
represented by the solid line, JE is represented by the dashed line and JE (vE∗ ) is
represented by the dotted line; (c) Cdf’s of NPV for the initial well placement
and the optimal well placement after primary and secondary optimization. . 193

xxvii
5.19 Top structure and initial oil saturation of the Oseberg model. The tar mat is
highlighted with a blue line. . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
5.20 Horizontal permeability (mD, upper row) and porosity field (bottom row) for
realization #1, Oseberg. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
5.21 Horizontal permeability (mD, upper row) and porosity field (bottom row) for
realization #8, Oseberg. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
5.22 NPV versus number of simulations obtained for deterministic optimization of
slanted well trajectories using realization #1, Oseberg. The blue line represents
the StoSAG algorithm and the red line represents the GPS algorithm. . . . . 198
5.23 Oil saturation after 3,900 days of production obtained with initial and optimal
well locations, deterministic optimization using realization #1, Oseberg. . . . 199
5.24 Bi-objective well placement optimization using 10 realizations, Oseberg reser-
voir. (a) Average NPV over 10 realizations versus number of simulation runs in
primary optimization using StoSAG (blue) and GPS (red); (b) the augmented
Lagrangian function, LA , minimum NPV, JR , and average NPV, JE , versus
number of simulation runs in secondary optimization using StoSAG (blue) and
GPS (red) where LA is represented by marker x, JR is represented by the solid
line, JE is represented by the dashed line and JE (vE∗ ) is represented by the
dotted line; (c) Cdf’s of NPV for the initial well placements and the optimal
well placement after primary and secondary optimization. . . . . . . . . . . 201
5.25 Number of violated constraints and indicies of the risk realization during sec-
ondary optimization of the bi-objective optimization of slanted well trajecto-
ries, Oseberg. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
5.26 Oil saturation in layer 10 after 3,900 days of production obtained with initial
well locations, and optimal well locations after robust optimization and bi-
objective optimization, for realization #1 and #7, Oseberg. . . . . . . . . . . 202
C.1 Numbering of the Cartesian grids. Numbers in parenthesis on the top surface
of each gridblock indicate the (i, j, k) gridblock indicies. . . . . . . . . . . . . 223

xxviii
C.2 Two-segmented perforation within one simulation gridblock. . . . . . . . . . . 225

xxix
 CHAPTER 1

INTRODUCTION

Closed-loop reservoir management (CLRM) is a general reservoir management frame-

work that consists of a loop of three sequentially performed inter-related tasks during the life
of a reservoir: data collection, history matching and optimization [36]. Fig. 1.1 presents the
schematic diagram of the CLRM framework. Relevant data, e.g., geological, seismic, well
logs and production data can be gathered via different surveillance operations. These data
can be assimilated to obtain a posterior reservoir model that better matches the data. The
importance of this step is that the new model is expected to have more accurate predictive
capability as well as reduce the uncertainty about the reservoir. Based on the calibrated
reservoir model, the most beneficial operating strategies can be devised using an optimiza-
tion algorithm. Typically, in CLRM, only the well control parameters of existing wells are
optimized, such as bottomhole pressures, production rates and valve settings.
Our focus in this work is on the optimization step of CLRM where the cost (objective)
function is the net present value (NPV) of production over the remaining life of the reservoir
(life-cycle optimization) and the design (optimization) variables may include well controls
(operating pressure or rates on specified time intervals or control step), well paths, well
drilling order, well type (injector or producer) and the number of wells. When the optimiza-
tion process includes some of the aforementioned key decision variables, it is referred to as
field development optimization, rather than simply production optimization. This problem
is essentially a mixed-integer problem determining high-dimensional variables which can be
either numerical or categorical, continuous or discrete. This research focuses on developing
suitable algorithms to solve the field development optimization and its sub-problems under
geological uncertainty.

1
 Figure 1.1: Schematic diagram of closed-loop field development management.

When geological uncertainty is considered, instead of a single reservoir model, a set

of models that are consistent with observed data, geological information and seismic inter-
pretation are obtained in the data assimilation process. Correspondingly, the optimization
process has to be adapted to improve the average reservoir performance such as the life-cycle
NPV over a set of reservoir models selected to represent the geological uncertainty, which is
commonly referred to as robust optimization. Note the average NPV of production over the
reservoir life is the standard estimation of expected value and we sometimes see the average
NPV referred to as the expected NPV and we may occasionally do so here. Although max-
imizing the average NPV of life-cycle production is important, the variance in the resulting
NPV’s of individual models at the computed optimal values of the design variables may be
large and in this case, there is considerable downside risk. To mitigate downside risk, one can
consider bi-objective optimization [22, 86, 20, 49, 50], or even multi-objective optimization
[48]. One common bi-objective optimization formulation is to consider the two conflicting
objectives of maximizing the expected value of the life-cycle NPV of production and max-
imizing the minimum value of the ensemble NPV’s [49, 50], and this is the bi-objective
optimization considered here.

1.1 Literature Review

Field development optimization is a critical part of the decision making process for a
reservoir. Many researchers have studied several individual parts, where the two most typical

2
problems are the well control optimization in which the injection/production rate and/or the
bottomhole pressure of wells are optimized and well location optimization in which the well
locations and trajectories are optimized. Recently, there has been a growing interest in
solving the joint optimization problem of well locations and controls with both sequential
and simultaneous approaches. Some researchers also investigated on the field development
optimization problem using global search algorithms. Previous procedures in the literature
are discussed in this section according to their design parameters and algorithms applied.
The well control optimization problem has been studied extensively in previous liter-
ature, using both adjoint gradient-based method [42, 73] and stochastic methods, e.g. PSO
[84, 80], CMA-ES [25, 77], EnOpt [13], StoSAG [19, 24], SPSA [47], etc. The adjoint method
provides the most accurate gradient and is computationally the most efficient method when
the cost function is a differential function of the optimization variables which is the case
when the optimization variables are well controls, but is not the case when the optimization
variables are well types and number of wells. In addition to lack of differentiability issue,
the use of a gradient-based optimization algorithm is precluded if the reservoir simulator
used lacks the capability to compute all the gradient information needed which is often the
case. When the simulator can compute the needed gradient, one should use adjoint gradient
for well control optimization and to some extend for well placement optimization. Of these
two types of design variables, the use of adjoint gradient of the NPV with respect to the
well placement variables is more tenuous for two reasons. Firstly, unlike the well control
variables which explicitly affect the source/sink terms in reservoir simulation equations, the
well locations exert their effects on the well index which indirectly controls the value of
the source/sink terms. Secondly, the NPV function is rough and becomes non-differentiable
when part of a well path crosses a gridblock.
Despite the aforementioned obstacles, there has been attempts to use the adjoint
based techniques in the well placement optimization. Sarma et al. [74] considered to op-
timizing the well locations for fully-penetrating vertical wells under bottomhole pressure
control using the adjoint formulation. In their work, the discrete parameters ((i, j) well lo-

3
cation indices) are replaced with their continuous counterparts in the spatial domain ((x, y)
well coordinates) and the discontinuous Dirac-delta function (defining wells as point sources)
in the underlying governing PDEs (reservoir simulation equations) is replaced with a continu-
ous function (a bivariate Gaussian function) of the defined well coordinates (x, y). With this
modification, the source/sink term becomes a continuous and differentiable function with re-
spect to the continuous well location variables which makes the adjoint formulation possible.
The discretization of the modified PDEs yields (additional) non-zero source/sink beyond the
original well blocks. The extent and total number of these additional terms are determined
by the size of the base of the approximating Gaussian function. In practice, the base is made
small enough so that the non-zero source/sink term (pseudo-wells) for gridblocks other than
the nearest ones are discarded. In their work, the NPV function is defined based on the
well terms for both the original and the additional perforated gridblocks. However, there is
an inconsistency when mapping the continuous source/sink back to point source/sink term
since they define similar but different dynamics and only keeping the nearest pseudo-wells
in the continuous source/sink term may underestimate the point source/sink term.
Forouzanfar and Reynolds [27] used an idea similar to that of Sarma et al. [74]
to smooth the objective functions for the placement of vertical, horizontal and deviated
wells operating at specified rates or bottomhole pressures. In their work, they perforate the
surrounding gridblocks of the well trajectory as well as the gridblocks actually penetrated
by the well. However, their representation is only an approximation since they assumed the
surrounding gridblocks and the actually penetrated gridblocks have (i) same well indices
(ii) same phase mobilities and (iii) same drawdown pressures. These assumptions are not
valid if there are abrupt changes in permeabilities, pressures and saturations between the
surrounding gridblocks and the actual penetrated gridblocks. They found that the NPV
obtained with wells under rate control is better estimated than those under bottomhole
pressure control. For rate-controlled wells, the total production rate remains constant while
for the bottomhole pressure controlled wells, the total production rate may be overestimated
due to the extra perforations in surrounding gridblocks. In their work, the well trajectories

4
are parametrized as the coordinates of the perforation center (xw , yw , zw ), the perforation
length (lw ), the inclination angle (ϕ) and the azimuth angle (θ). Instead of using the adjoint
framework, they carried out the optimization using a derivative free method, i.e., Bound
Optimization by Quadratic Algorithm (BOBYQA) which does not require the NPV function
to be differentiable everywhere. One advantage of the BOBYQA is that it uses the reservoir
simulator as a “black box,” and hence can handle the economic and field operating constraints
(e.g., maximum water cut constraint) as reactive constraints in the reservoir simulator.
Volkov and Bellout [82] proposed an efficient approach to approximate the gradient
of the NPV term with respect to the well placement design parameters through the adjoint
framework using one forward run where the reservoir simulation is performed and the state
variables are saved and one backward run where the adjoint system is solved. At each step
of the backward simulation run, the adjoint system is formulated with the state variables
and solved with respect to the corresponding adjoint variables λ. The adjoint gradient for
the control variables are solved directly using Eq. 1.1.

∂J ∂g
∇u J = + ( )T λ, (1.1)
∂u ∂u

where u is the vector of well control variables, J is the objective function (e.g. the cumulative
oil production or the net present value of life-cycle production); g represents the nonlinear
residual equation at each gridblock of the reservoir; λ is the adjoint vector. Denote the vector
of well location variables as x. The gradient components corresponding to the well config-
uration variables are obtained by a finite difference scheme by perturbing each element xj
individually and evaluating the residual and NPV terms of the perturbed xj . Unlike analyt-
ical derivatives, finite difference may be able to capture the coarse-scale slope of the function
in the case of function roughness or discontinuities. However, the suitable perturbation size
is case-dependent, which means that parameter tuning has to be performed for each differ-
ent reservoir model considered and that the obtained perturbation ranges are assumed valid
for all feasible well configurations explored by the optimization procedure. In their work,

5
they considered optimization of the placement of both vertical wells and slanted wells. Since
the (approximated) adjoint gradient can be obtained for both the control variables and well
configuration variables, this approach can handle simultaneous optimization of both terms.
They tested the simultaneous optimization on a 2D reservoir with 15 vertical wells under
rate control. They did not show the comparison between simultaneous optimization and se-
quential optimization using the adjoint framework. The well spacing constraints and signed
well-to-boundary distance constraints are considered where the derivative of the constraints
with respect to the well configuration variables are also obtained by finite difference.
Zandvliet et al. [87] placed eight “pseudo-wells” in the neighboring gridblocks around
each actual well. A pseudo-well is a well with a very low rate with negligible effect on the
reservoir’s overall performance. They used an adjoint method to calculate the gradients
of NPV with respect to the small flow rates prescribed for the pseudo-wells. The pseudo-
well that gives the largest gradient determines the location of the well in next iteration.
Vlemmix et al. [81] applied a similar concept to find the optimum well trajectory for a single
horizontal/deviated well. Their method includes creating “pseudo-side-tracks” from each
trajectory point. Each pseudo-side-track is assigned a very small perforation length that
makes their effect on the overall reservoir performance negligible. These side tracks are then
used to find the approximate gradient direction for each trajectory coordinate. A smoothing
step is needed after every update of the trajectory coordinates, to ensure the drillability of
the well (i.e., the curvature of the well trajectory has to stay below a maximum value allowed
by the directional drilling technology and equipment). However, this search direction is not
guaranteed to be uphill due to the following reasons: (i) the largest gradient of the NPV with
respect to the control settings of the pseudo-wells (pseudo-tracks) is not necessarily positive
(ii) the interactions between wells may counter the positive effect expected by increasing the
flow rates of each pseudo-well (pseudo-track) individually (iii) the decrease in the flow rate
of a main well may have a negative effect on the performance countering the positive effect
by increasing the flow rate of its best pseudo-well (pseudo-track). In both of their works,
only the well placement variables are optimized.

6
 Some researchers rephrased the well placement optimization problem as the well con-
trol optimization problem. Wang et al. [83] proposed to put a well in every gridblock and
optimize a modified objective function (NPV minus an approximation of the drilling cost
as a function of the injection rate) using the adjoint gradient in order to find the optimal
number of injectors, their locations and rates. In their work, for each well, when the injection
rate is zero, the approximate drilling cost becomes zero, and this well is eliminated; when
the injection rate is large enough, the approximated drilling cost equals the true drilling
cost and this well needs to be drilled. They also imposed a total injection rate constraint
which helps to eliminate unnecessary wells. However, it is challenging to find a suitable
approximation function for the drilling cost and to decide the total injection rate. Zhang et
al. [89] speeded up the algorithm proposed in [83] by eliminating more than one injectors
at each iteration. Forouzanfar and Reynolds [26] experimented with different approximation
functions of the drilling cost, proposed an heuristic initialization procedure to determine the
total injection rate and total production rate, and extended the method to the placement of
both injectors and producers. However, there is an inconsistency in their objective function
and the gradient calculation. The method used in [83], [89] and [26], is only applicable to
wells under rate control. Issues with the methodology are that (i) wells eliminated at the
current iteration may not be brought back in future iterations; (ii) an unnecessarily large
amount of wells may be obtained in the final solution which means the solution may be
somewhat suboptimal. Compared to the previously mentioned algorithms which also use
the adjoint framework, this method can optimize the number, type and control settings for
wells under total liquid rate control, as well as the well locations. However, their method
has only been applied to the placement of vertical wells with only one control step and it is
difficult to include the well trajectory parameters in the optimization process.
As the adjoint gradient is not generally available, some researchers used the stochastic
gradient to approximate the true gradient and thus calculate a search direction. Calculating
stochastic gradients does not require the objective function to be differentiable. Hence,
there is no need to smooth the objective function or to change the optimization variables

7
to well control variables. However, the advantage of the stochastic gradient procedures
comes with the disadvantage that an approximate gradient generally reduces the overall
performance of the optimization algorithm. Common stochastic gradient-based methods
applied in well placement optimization include SPSA, EnOpt and StoSAG. Bangerth et
al. [5] applied an integer variant of the SPSA method to well placement optimization by
simply rounding the perturbation size and the step size to the nearest integers. In their
work, they considered the placement of 7 vertical wells for a 2D reservoir where the well
locations are represented by the (i, j) gridblock indices and compared the optimum results
obtained with SPSA, finite different method (FDM) and the very fast simulated annealing
(VFSA). They concluded that SPSA was more computationally efficient in finding good
locations than VFSA, while VFSA obtained good locations more reliably but with more
function evaluations than SPSA; the FDM is less reliable and requires more simulation runs
than both SPSA and FDM. Li et al. [47] combined the optimization of the placement of
vertical wells and their corresponding rates into a joint problem using a modified SPSA
algorithm. The total injection rate and total production rate are constrained to be equal
to 1 pore volume per year (linear constraints). In the modified SPSA, the perturbations to
the well location variables and well rates are randomly chosen from the full set of feasible
directions. For the well placement variables, at each iteration, a random neighbor of the
grid block is chosen and the NPV is evaluated for locating a vertical well in this gridblock
as well as in the gridblock in the opposite direction, and the well is then moved to the
grid block with the higher NPV. In this SPSA variant, the step size of the line search is
limited to a single grid block. Li et al. investigated the two common approaches for the joint
optimization of well locations and controls, which are namely the simultaneous and sequential
approaches. In the simultaneous approach, the vector of optimization variables includes
both the control variables and the well trajectory variables. In the sequential approach,
the well placement optimization is carried out first with the controls fixed at their initial
values, and then based on the optimal well locations obtained, the well control optimization is
carried out. According to their paper, simultaneous optimization for vertical wells under rate

8
control outperformed sequential optimization for both deterministic and robust optimization
regarding the number of iterations and the optimum NPV’s obtained. They showed that
in robust optimization, using only a small random subset of model realizations resulted in
optimization results similar to those obtained when the entire ensemble of models is used.
Jesmani et al. [37] considered two well placement optimization problems where the location
of a vertical well is optimized in one problem and the trajectory for a horizontal well is
optimized in the other problem, using the original SPSA algorithm. They also investigated
an approach to handle geological uncertainty in [47]. Their results suggested that the robust
well placement optimization using randomly selected subset of realizations saves significant
computational time, while the optimum of the expected NPV remains almost the same as
in the case where all realizations are used. They also observed that the SPSA algorithm
is a local search method in which the performance is highly dependent on the initial well
configuration. Besides that, Jesmani et al. [37] considered well length constraints and the
inter-well distance constraint by combining the SPSA algorithm with the decoder method
and the projection method respectively; however, no conclusion of the performance of these
constraint handling methods have been drawn due to the limited trials.
Chen et al. [12] proposed the ensemble based optimization (EnOpt) for the robust
control optimization problem. In EnOpt, at each iteration, (i) for each realization, one per-
turbation vector is generated and the corresponding NPV is calculated for the perturbed
control vector, and (ii) the cross-covariance between the perturbed controls and NPV’s is
used as the search direction. Fonseca et al. [24] proposed a variant for robust control
optimization, i.e., StoSAG, in which the search direction is estimated as a stochastic approx-
imation of the simplex gradient. Fonseca et al. [24] theoretically show that when there is
a large variation in the ensemble of reservoir models, StoSAG provides a better approxima-
tion of the true gradient than EnOpt. They provide examples to illustrate this theoretical
result, i.e., they show, via computational experiments, that StoSAG generally yields a sig-
nificantly higher value of the life-cycle NPV for robust optimization than is achieved with
the standard EnOpt algorithm. Though both StoSAG and EnOpt are designed for robust

9
optimization, i.e., optimization over an ensemble of reservoir realizations characterizing the
geological uncertainty, they can be readily adapted to deterministic optimization, see Do
and Reynolds [19]. Leeuwenburgh et al. [45] applied EnOpt to deterministic optimization
of vertical well placement in a 2D reservoir for up to 9 wells whose locations are represented
by the (x, y) coordinates. They concluded that EnOpt works well for the tested determin-
istic well placement problems. Zhang et al. [90] applied StoSAG to the deterministic well
placement and control optimization problem for a 3D channelized reservoir with 13 slanted
wells with water-alternating-gas/surfactant-alternating-gas flooding process. The well loca-
tions are represented as the (x, y, z) coordinates of the heel point and toe point of a well
in the reservoir. He concluded that StoSAG works well for both problems, however, simul-
taneous optimization did not outperform sequential optimization. Lu et al. [52] improved
the StoSAG search direction and proposed updating the well placement variables and the
control variables alternatively iteration by iteration, referred to as an iterative simultane-
ous procedure. Their results showed that the iterative simultaneous procedure outperforms
the sequential procedure for deterministic optimization, but these two procedures perform
similarly for robust optimization where 10 realizations are used to represent the geological
uncertainty. They also developed an efficient StoSAG gradient estimation for bi-objective
optimization where the two objectives are to maximize the expected NPV and minimize the
minimum NPV among all the realizations representing the geological uncertainty.
Some researchers ([31, 44]) reparameterized the categorical variables (to-drill or not-
to-drill, injector or producer, drilling order of wells) as continuous variables and carried out
optimization using a search direction estimated by stochastic gradients. Hanea et al. [31]
determined the well drilling order, types and times for converting a producer to an injector
under geological uncertainty using the Stochastic Simplex Approximate Gradient (StoSAG)
in a Quasi-Newton method. For the case of determining drilling order and well type for a
given set of well paths, each well is assigned one continuous priority variable (for drilling
order) and one continuous type variable. The larger the priority variable, the earlier the well
is drilled. The well is an injector if the type variable is greater than 0.5 and otherwise, the

10
well is a producer. However, this procedure can be problematic as perturbing a continuous
priority or type variables may not change the actual drilling order or well type which then
leads to a low quality stochastic gradient, i.e., a poor search direction. In the extreme case
where the perturbation produces no change in drilling order and does not change the type
of even one well, the stochastic gradient would be zero even when the current estimate of
the optimization variables is far from optimum.
Evolutionary methods have also been tried for optimization problems like those con-
sidered in this work. The Genetic algorithm (GA) is one of the most popular methods in
well placement optimization. GA can be thought of as a self-adapting system. Bittencourt
and Horne [8] developed a hybrid GA and applied it to determine the best well locations
and trajectories for a real case of an offshore oil reservoir development project. In their
project, geological uncertainty is not considered; the total number and types of wells are
pre-specified; wells trajectories can only be vertical or horizontal and if horizontal, wells
can take any orientation in the oil zone; no restrictions concerning well spacing and injec-
tor/producer patterns are applied. Their work focuses on well placement optimization where
the control settings are pre-fixed. The chromosome is formulated as a binary string of the
head point coordinates (gridblock indicies), the trajectory type (vertical or horizontal) and
the orientation (8 directions) for 33 wells. The length of the well is not optimized either.
In the final solution, they found that about five wells were placed very close to injectors,
faults or even to other producers which can provide heuristic guidance for reducing the num-
ber of wells. Montes et al. [59] studied the effect of several GA parameters such as initial
population, population size, and mutation rate on optimizing the well locations for vertical
wells. These are important GA parameters that can directly affect the efficiency and relia-
bility of GA convergence to a result close to a global optimum. If these parameters are not
selected carefully, GA may require an unnecessary large number of reservoir simulations (for
fitness evaluation) or can get trapped around a local optimum. Montes et al. [59] tested
a deterministic layercake reservoir with 5 vertical wells where the (i, j) indicies are used
to describe well locations. Their results show that the initial population does not have a

11
significant effect on the results. They also conclude that the general rule in GA literature
for the optimum population size (twice the number of bits used in a chromosome) applies
for their problem too. However, it is not clear that this conclusion applies when the number
of variables becomes very large. Their results are not conclusive about the mutation rate.
Yeten et al. [86] determined the optimal location and trajectory of a non-conventional well
(heel and toe point of the main bore, number of junctions and number of laterals emanating
from the same junction, lateral trajectory, the wellbore diameter and the total liquid produc-
tion rate) using the genetic algorithm with a binary encoding. In their work, the trajectory
of both the main bore and the laterals are described by the coordinates of the entry point
(hx , hy , hz ), the length of the trajectory projected onto the x-y plane lxy , the orientation of
the well in the x-y plane θ, and the depth to the trajectory endpoint tz . Both deterministic
optimization and robust optimization are tested. In robust optimization, 10 realizations are
used represent the geological uncertainty and the objective function is defined as a weighted
sum of the average NPV and the standard deviation. In all their test cases, the maximum
number of junctions on the main bore is set to be 4 and maximum number of laterals ema-
nating from the same junction is set to 1. Emerick et al. [21] used the Genetic Algorithm for
Numerical Optimization of Constrained Problems (Genocop III) to determine the number,
types, locations and trajectories of slanted wells subject to various constraints for a deter-
ministic reservoir. In their work, the chromosome is composed of both binary variables (well
type, i.e., injector or producer, and well status, i.e., open or shut-in) and integer variables
((i, j, k) gridblock indices of the heels and toes for each well), but treated as a string of real
numbers. In their work, an arithmetic crossover is used to make linear combinations of a
randomly selected parent pair and the genes of the produced offsprings are truncated to
the nearest integers. They also mutate the locations variables, the type variables and the
status variables in different ways and at different mutation probabilities. Emerick et al. [21]
considered both linear and nonlinear constraints where any individual in the population has
to be repaired before its evaluation if it is infeasible. These constraints include maximum
well length, minimum distance between wells and user-defined undesirable regions for well

12
placement. In their work, they are able to test different cases with up to a maximum number
of 16 wells within 5000 simulations. Based on three full-field reservoir models, (i) they have
shown significant increases in the NPV and cumulative oil production using Genocop III over
the well placement scenarios proposed by reservoir engineers; (ii) the optimization defining
the whole initial population randomly achieved a value of NPV lower than the results using
the engineer’s suggested case in the initial population. Lee et al. [43] used GA to optimize
the number, location and trajectory of horizontal wells using a 2D node-based configuration.
In their method, a horizontal well can have multiple kick-off points. However, they only
tested their methodology on two small 2D synthetic reservoirs (one with 10 × 10 gridblocks
and one with 20 × 20 gridblocks) and considered the placement of three horizontal wells at
maximum. Though their estimated optimal NPV’s are larger compared to placing three ver-
tical wells, increasing the number of kick-off points gives negligible increase in the estimated
optimal NPV’s. Morales et al. [60] used GA to optimize the heel and toe coordinates of a
horizontal well in a reservoir with 30 × 30 × 11 grids considering 5 different realizations of
the reservoir. In their method, the fitness value is calculated as the weighted sum of the
NPV’s of realizations whose NPV’s are among the top ps percent and they showed that the
trajectories of the horizontal well obtained with different ps values have significant changes.
Salam et al. [72] proposed to calculate the fitness of each candidate solution using a proxy
trained by artificial neural networks (ANN) and to evolve to next generation following the
GA operators. They optimized the placement of 10 vertical producers and their conversion
time from production into injection respectively and showed that the hybrid algorithm of GA
and ANN only requires a few hundred simulation runs for the two problems they considered.
Besides GA, some other methods have also been tried, PSO [65, 66], simulated annealing
[6], CMA-ES [18, 25], and pattern search algorithms [7, 33, 32].
Bellout et al. [7] proposed a nested joint optimization algorithm where the upper level
optimization problem is the well placement optimization using Pattern Search Algorithms,
i.e., Generalized Pattern search (GPS), Hooke-Jeeves Direct Search (HJDS) or a hybrid
optimization parallel search package (HOPSPACK, a distributed computing implementation

13
of GPS, and the lower level is the well control optimization using the adjoint gradient.
In the sequential optimization (well placement optimization with fixed controls followed
by well control optimization using the optimal locations obtained), two control strategies
are considered in the well placement optimization: a fixed control strategy where injectors
operate at their upper bounds of the bottomhole pressure and producers operate at their
lower bounds of the bottomhole pressure, and a “reactive” control strategy where injectors
operate at their upper bounds of the bottomhole pressure and producers operate at their
lower bounds of the bottomhole pressure until a prescribed limit to the maximum water cut
is reached. In their work, they consider a 2D reservoir with 5 vertical wells and 10 control
intervals, based on which the nested joint optimization algorithm outperformed sequential
optimization with both aforementioned control strategies. From their results, GPS and
HOPSPACK are slightly faster than HJDS, but they yield lower NPV values than HJDS.
However, GPS and HOPSPACK can be accelerated, in terms of clock time, if a cluster is
available. Li and Jafarpour [46] used an iterative procedure in which they alternated between
a full well placement optimization and a full well control optimization. Again, different
optimization approaches were used for the two problems, where a coordinate descent random
search algorithm is used for the well placement optimization and the adjoint gradient-based
method is used for the well control optimization. Their work pertains to vertical wells under
total liquid rate control. Though the number of control steps is only provided for one case
with five wells, they tested cases with up to 15 wells. For all cases, they were able to achieve
significant improvements over a single well placement followed by well control optimization.
As we mentioned previously, Lu et al. [52] used a modified iterative procedure based on
StoSAG to optimize well trajectories and controls for 13 slanted wells with 12 control steps,
and found that the iterative procedure also outperformed the procedure of a single well
placement followed by well control optimization. The above mentioned works essentially all
decompose the joint problem into two smaller subproblems. This enables the application of
a specialized optimization method for each of the two subproblems.

14
 Li et al. [47] combined the optimization of vertical well placements and their cor-
responding rates into a joint problem using a modified SPSA algorithm where the total
injection rate and total production rate are constrained to be 1 pore volume per year (lin-
ear constraints). The results they obtained for the PUNQ model showed that simultaneous
optimization outperformed sequential optimization. Isebor et al. [34] applied a hybrid tech-
nique that combined the PSO and the mesh adaptive direct search (MADS) for the joint
optimization problem of the example considered by Bellout et al. [7], but with only 5 con-
trol steps. Besides that, they also developed a consistent filter-based treatment ([23]) of
nonlinear constraints in both stand-alone PSO and the PSO-MADS hybrid. In their work,
they handled three types of nonlinear constraints (minimum distance between wells, max-
imum field rate, and maximum well water cut) whereas in Bellout et al. [7] only bound
constraints are considered. Bellout et al. [7] required about 4,000 simulations for each of
their optimization runs for 10 location variables with 50 control variables, whereas Isebor et
al. [34] used around 30,000 simulations for each run where there are 10 location variables
and 25 optimization variables. Forouzanfar et al. [25] implemented the CMA-ES optimiza-
tion algorithm to solve both the simultaneous and sequential joint optimization problem of
slanted wells where they kept the parametrization of well trajectories and introduced a new
parametrization strategy for the control variables. Their numerical results for PUNQ model
showed that simultaneous optimization algorithm gave slightly better NPV’s than sequen-
tial optimization framework. In summary, Li et al. [47], Isebor et al. [34] and Forouzanfar
et al. [25] all show that the simultaneous approach outperforms the sequential approach.
However, the alternate conclusion has been reached by others. Specifically, Humphries et
al. [33] used a hybrid algorithm that combines PSO and GPS for simultaneous optimiza-
tion, which is an approach similar to the one used by Isebor et al. [34]. In their sequential
optimization implementation, PSO is used for the well placement optimization and GPS is
used for the well control optimization. Humphries et al. [33] did not find the simultaneous
approach to consistently provide better solutions than the sequential procedure. Humphries
et al. [32] later applied their method to joint optimization for non-conventional wells with

15
different complexity of trajectories and suggested that the sequential approach is better able
to deal with increasingly complex well parameterizations than the simultaneous approach,
and for their complex problems where there are 24 trajectory variables (4 slanted wells) and
20 control variables (5 control intervals), compared to simultaneous optimization, sequential
optimization gives 5% higher NPV’s under bound constraints and 9% higher NPV’s under
nonlinear constraints.
Most of the well control and placement optimization problems mentioned above as-
sume the number, types and drilling orders of the wells to be drilled are known which is not
the case most of the time. Hence, there is a strong need to research on the full field de-
velopment optimization problem where the number of wells, their types, drilling sequences,
locations and control settings are optimized altogether. The full field development optimiza-
tion problem determines both categorical variables (drill-or-not, injector or producer, and
well drilling sequence) and numerical variables (well locations and control settings). Hence,
as a mixed-integer problem, it is far more complicated than the joint optimization of well
locations and controls. Recall that Wang et al. [83], Zhang et al. [89], and Forouzanfar
and Reynolds [26] rephrased the field development optimization problem as a well control
optimization problem where the drilling cost is defined as a smooth function of the injec-
tion/production rate and optimize the number, type, locations and control settings for wells
under total liquid rate control. Ciaurri et al. [14] proposed to solve the field development
optimization problem by well pattern optimization followed by a second-stage well placement
optimization which provides a minor change of the well locations within the patterns using
PSO method. Although this method is more computationally tractable, it suffers from three
drawbacks: firstly, it is only applicable to newly discovered reservoirs where no wells has
been drilled; secondly, the control setting of the wells has to be pre-specified; thirdly, this
method theoretically will lead to suboptimal results since wells are constrained to repeated
patterns. Awotunde [2] optimized the well number, types, locations and controls using dif-
ferential evolution (DE). He proposed to divide the search space of the control variables
at the first control step (or the coefficient of a polynomial/trigonometric parametrization

16
of the control variables) into three intervals (indicating an injector, no-well and a producer
respectively), referred to as well control zonation method. Hence, the “drill-or-not” status
and “injector/producer” type of a well can be determined from the control variables. The
settings of the no-well zone has a large effect on the algorithm performance and its optimal
settings are case dependent. Based on the well control zonation method, Awotunde [1] in-
troduced two additional design variables, the project life and the fractional times at which
the wells are brought online, and he is able to simultaneously optimize well types, locations,
controls and well open time. However, the well controls over the reservoir life were assumed
to follow a Cosine function of time and wells were brought online without considering the rig
availability. Awotunde [3] compared the well control zonation method and the mixed-integer
non-linear programming (MINLP) procedure using DE and his results showed the former
method outperforms the latter in most cases provided that the no-well zone is set as large
as or larger than the combined width of the two other zones. Khan et al. [41] extended the
well control zonation method to include the vertical/horizontal well type in addition to the
injector/producer type and used PSO as the optimizer. They found that the well control
zonation method outperforms the MINLP method, which is consistent with the observations
of Awotunde [3].
Isebor et al. [35] introduced a ternary categorical variable (where the three levels rep-
resents injector, do-not-drill and producer) in addition to the locations and control settings
and then optimized these variables using PSO-MADS. To handle the integer and ternary
variables, they simply round the variables to the nearest integer value before evaluation in
the PSO search step and use a discrete mesh in the MADS poll step. They noted that both
PSO and MADS perform satisfactorily for continuous problems, but with the inclusion of
integer variables, they may converge to relatively poor local optima. What is more, the
convergence of PSO-MADS with multi-level categorical variables (the ternary variable) is
not clearly understood. Shirangi and Durlofsky [76] extended the work of [35] to closed
loop field development optimization under geological uncertainty where the history match-
ing process is solved with the Random Maximum Likelihood (RML, [64]) method using the

17
adjoint gradient and the optimization process is carried out using the PSO-MADS method.
In the work of Shirangi and Durlofsky [76], the well locations are represented by the (i, j, k)
gridblock indices of their two end points and the number of control steps per well is set
equal to the number of wells, Nw , with the length of the first Nw − 1 control steps equal to
the specified constant drilling time. Based on a deterministic 2D reservoir with four vertical
wells where each new well is drilled every 210 days, they demonstrated that simultaneous
optimization (optimizing the well types, locations, time-varying controls of these four wells
simultaneously) obtained 16% higher NPV than well-by-well optimization where the location
and type of the first well is optimized first and given the optimal location and type of the
first well, the location and type of the second well (drilled at 210 days) are optimized, and
so on. This simple case took 200 equivalent simulation runs (number of times a batch of
simulation runs is submitted to the compute cluster with a maximum of 400 available cores),
which equals 80,000 actual simulation runs. They solved the full problem of closed-loop field
development for a 2D reservoir with a maximum number of 8 vertical wells and 8 control
steps with 50 realizations used to represent the geological uncertainty. However, due to
the heavy computational cost of PSO-MADS, they proposed a sample validation procedure
(optimizing over an representative subset, e.g. 10 realizations) in the field development op-
timization. This case took 1700 equivalent simulation runs using a cluster with a maximum
of 400 cores, i.e., approximately 350,000 actual reservoir simulation runs. A 3D example
with 6 wells are also presented where they fix the number, type and drilling sequences of
wells and pre-specify whether a well is vertical or horizontal. Although the procedure using
PSO-MADS may be computationally tractable for realistic field cases using parallelization
with a large number of cores, there is a concern about whether simultaneous optimization
will still outperform the well-by-well procedure with an increase of the problem complexity
(e.g. considering geological uncertainty, optimizing number of wells and trajectories for com-
plicated wells) and number of variables (optimizing over more wells and using more control
steps), according to the findings in [32].

18
 Yeten et al. [86], Essen et al. [22], Durlofsky et al. [20], Liu and Reynolds [49],
Lu et al. [52] have all considered the optimization process using a set of model realizations
representing the geological uncertainty. However, only a small ensemble is used in robust
optimization, which may not be able to properly capture the subsurface uncertainty. The
computational cost for a large ensemble of realizations may be prohibitive. Thus, it is
significant to develop strategies to improve computational efficiency for large ensemble size.
Yang et al. [85] proposed conducting robust optimization on a small subset of repre-
sentative realizations which is selected based on a reliable ranking of the calculated perfor-
mance of all the realizations under a reference operating condition. However, the ranking of
reservoir performance changes if the well operational controls are updated in the optimiza-
tion process and in fact can change significantly so that the chosen subset no longer gives an
adequate representation of the full ensemble [53]. To further reduce the computational cost,
Yang et al. also replaced the reservoir simulation process with a trained response surface
model. Li et al. [47] evaluated the objective function with a subset of randomly selected
realizations at each iteration of the SPSA algorithm in order to achieve the desired com-
putational efficiency. Although SPSA is able to handle a noisy objective function, there is
no guarantee that selected subsets can represent the full set. Shirangi and Durlofsky [76]
proposed reselecting the representative realizations based on the solutions obtained with
a previous complete subset optimization, and repeat the subset optimization until a pre-
specified relative improvement ratio is reached, i.e., 0.5 in their work. In their work, the
relative improvement ratio is defined as the increase in the expected NPV for the subset
of realizations (based on the initial expected NPV for the subset) over the increase in the
expected NPV for the full set of realizations (compared to the initial expected NPV in the
full set). However solving a complete optimization problem for each subset is expensive, and
the obtained solution is not guaranteed to be the optimal solution of the full-set.
Many researchers have worked on multi-objective optimizations of the well controls.
For both deterministic and roust optimization, it is desirable to maximize the long-term NPV
and the short-term NPV at the same time so that the invested money can be payed back

19
sooner and hence can be invested on other projects without significantly harming the long-
term benefit. For cases considering geological uncertainty, it is desirable to maximize the
expected NPV and to minimize the risk (i.e., minimize the standard deviation of the NPV’s of
all realizations or maximize the minimum NPV). Essen et al. [22] proposed an adjoint-based
hierarchical production optimization method in which the secondary objective (maximization
of the short-term NPV) is optimized based on the null space of the Hessian matrix of the
primary objective (maximization of the long-term NPV). They observed a significant increase
in short-term objectives without significantly decreasing the life-cycle NPV. However, the
authors point out that the convergence of the method is slow, due to infeasible solution steps.
They also introduced an alternative method, referred to as switching method in which the
long-term NPV is not treated as a strict constraint. Chen and Reynolds [11] carried out a
robust short term control optimization after long-term control optimization subject to the
constraint that the average long-term NPV should not be decreased using the Augmented
Lagrangian method for waterflooding examples and showed that robust long-term alternating
with short-term optimization is able to increase the short-term NPV without compromising
the life-cycle NPV. Liu and Reynolds [49] built the Pareto front of maximizing the long-term
NPV and maximizing the short-term NPV using an adjusted weighted sum method and
an normal boundary intersection method with adjoint gradient for an optimal well control
problem of a waterflood example. Liu and Reynolds [49] found that both methods gave
similar estimates of the Pareto front where the solution points distribute fairly evenly. Yeten
et al. [86] and Durlofsky et al. [20] used a simple treatment of uncertainty by incorporating
the standard deviation of the NPV’s for all realizations into the objective function with
the framework of GA to maximize the expected NPV and to minimize the risk over an
ensemble of realizations. However, Liu and Reynolds [49] observed the undesired feature
that the reduction in the standard deviation of the NPV’s is achieved mostly by reducing
the highest plausible NPV’s. Thus, they propose to maximize the minimum NPV instead of
minimizing the standard deviation in bi-objective optimization where the primary objective
is to maximize the average NPV. However, the computational cost of finding the Pareto

20
front may be expensive. What is more, for a problem with rough objective function surface
and multiple local solutions, e.g. the joint optimization of well control settings and well
placement, our experience suggests that it may be difficult to obtain a good approximation
of the Pareto front. Hence, it is much cheaper and practical to simply find a reasonable
trade-off instead of the whole front. Liu and Reynolds [50] combined lexicographic method
with adjoint based gradient to solve a bi-objective well control optimization problem where
the two objectives are optimized in a sequence according to the order of importance in order
to find a reasonable trade-off solution.

1.2 Research Objectives and Contextures

Formulating field development planning as an optimization problem is slowly be-

coming an acceptable approach in the petroleum industry. Many researchers have studied
several individual parts, where the two most typical problems are the well control optimiza-
tion which has been extensively studied in the literature and well placement optimization
which is continuing to gain more and more attention.
As limited adjoint capability is available in commercial simulators, we focus our atten-
tion on a gradient-based algorithm where the true gradient is approximated by a stochastic
gradient when optimizing well controls and well trajectories. Specifically we use the StoSAG
(stochastic simplex approximate gradient) to replace the true gradient in a steepest ascent
(or descent) algorithm. We consider both deterministic optimization based on a single reser-
voir model and robust optimization when we maximize the average net present value of
productions over a set of reservoir models selected to represent the geological uncertainty in
reservoir model parameters.
For general field development optimization, we will also need to incorporate some
combinations of the following variables: the number of wells that will be drilled, the well types
(injector or producer) and the drilling order. Generally, there is a strong need to conduct
research on the complete field development problem where the number of wells, their types,
drilling order, locations and control settings are optimized altogether. Since the complete

21
field development optimization is in fact a mixed-integer programming problem, different
algorithms (e.g. evolutionary algorithms), other than (stochastic) gradient-based algorithms,
will likely need to be applied. Previous works solve the field development optimization
problem in two ways. The first way re-parameterizes the categorical design variables as
continuous variables (i.e., controls [83, 89, 26, 3]) or priorities [31, 44]), uses optimizers
designed for continuous variables to estimate their optimal value and then map the estimated
continuous variables back to the categorical variables. According to Lu and Reynolds [54],
this type of methods not only enlarge the search space, but also give sub-optimal solutions.
The second way parameterizes the categorical variables as they are and uses a hybrid of
global and local search optimizers, i.e., PSO-MADS, where at each iteration, the categorical
variables are again treated continuous and then truncated back to their nearest discrete
values after an update, which does not seem ideal. As part of our research, we will attempt
to develop suitable encodings for variables of different types and use the Genetic algorithm
(GA), which handles the categorical variables more naturally, as the optimizer. Since it is
well known that GA is computationally expensive, we also attempt to improve its efficiency
by either decoupling the problem or combining GA with stochastic gradient methods.
Geological uncertainty, which leads to the uncertainty in the performance prediction,
has a relevant effect on the decision-making process regarding reservoir development. The
common method is to use multiple plausible geostatistical realizations when evaluating the
objective function. Given a large ensemble of realizations which reasonably covers the un-
certainty of the model parameters, the primary objective function of robust optimization
will be to maximize the average NPV over all realizations efficiently. However, when the
number of realizations of the reservoir model is large, computational feasibility requires that
one reduces the number of models on which robust optimization will be preformed, i.e., to
determine a representative subset of the the whole ensemble so that robust optimization on
the subset will produce an expected value of NPV close to the one obtained via robust op-
timization on the full ensemble. Previous work on well placement and control optimization
problems uses a subset of representative models equally spaced on the cumulative distribu-

22
tion function of NPV’s of all realizations evaluated at a reference operating condition where
the subset size varies from 3 [85] to 50 [16] depending on the user preference. Most of the
previous work on the field development optimization problem, due to its complexity, uses
only one reservoir model [83, 89, 26, 14, 2, 1, 3, 41, 35], except Shirangi and Durlofsky who
used a subset of 10 realizations of a 2D synthetic reservoir with 56 optimization variables at
the cost of around 350,000 simulations which is hardly affordable without parallel comput-
ing. However, it is difficult to specify the reference operating condition and it seems obvious
that it is highly unlikely that the selected realizations will remain representative during the
whole optimization process. As part of our research, we will attempt to develop a procedure
to select an initial representative subset of reservoir models from the full set that represents
geological uncertainty and to update this subset during the optimization so that at the end
of robust optimization, the optimal parameter for the subset will also be at least weakly op-
timal for the complete set of reservoir models. Robust optimization maximizes the average
NPV of production over a plausible set of reservoir models. However, the set of optimal
design variables obtained by maximizing average NPV may produce widely varying values
of NPV when their optimization variables are applied to each reservoir model. Hence, we
also consider to achieve a secondary objective to reduce the risk which is to maximize the
minimum NPV or minimize the standard deviation of the NPV’s over the given ensemble.
The objectives of this research and our associated contributions are as follows:

• Our first objective was to explore the applicability of StoSAG algorithm in well place-
ment optimization of slanted wells and multi-segmented wells considering minimum
well spacing, maximum well length and drillability constraints. We firstly proposed
suitable parameterizations for slanted wells and multi-segmented wells and then cou-
pled the augmented Lagrangian method with StoSAG to solve the constrained well
placement optimization problem. Though StoSAG has been shown to be efficient for
well placement optimization problems [52, 30], we found the solutions obtained by
StoSAG are only sub-optimal and added a subsequent General Pattern Search (GPS)
method to fine tune the well locations and trajectories.

23
• A second objective was to develop improved methods for the field development op-
timization problem. Here, we consider two problems. The first one has not been
considered previously in any journal publications where we want to select a fixed num-
ber of wells from a given set of potential drilling paths and to determine their types,
drilling order and control settings. The second is a more general problem where we
optimize well number, types, locations and control settings. We decouple the control
optimization from both problems and optimize well control settings using StoSAG at
the end to reduce problem complexity and to be computationally feasible. For the
first problem, we developed different encodings for well locations, types and drilling
order, and used GA with mixed operators which is shown to outperform the methods
of [83, 89, 26] and [31, 44] in this work. For the second problem, we develop two dif-
ferent hybrids of GA, StoSAG and GPS to efficiently optimize well number, types and
locations, which are able to consider a much larger problem compared to Isebor et al.
[35] and are shown to significantly outperform the Genocop III method by [21].

• A third objective was to improve computational efficiency for robust optimization

when a large number of realizations are used to characterize the uncertainty. Here, we
developed a robust procedure to adaptively choose representative subsets from the full
ensemble of reservoir models during the optimization process. The adaptive procedure
outperforms the case of using a fixed subset throughout the optimization process,
and only requires a few thousand simulations in contrast to a few hundred thousand
simulations required for Shirangi and Durlofsky [76].

• In addition, we consider bi-objective optimization within the context of robust op-

timization. Here, we consider the bi-objective well placement optimization problem
where the two objectives are to maximize the average NPV and maximize the mini-
mum NPV. Following the lexicographic method, the two objectives are optimized in
a sequence where the average NPV is optimized first, and starting from the optimal
solution obtained, the minimum NPV is optimized subject to an additional constraint

24
 that the average NPV is not decreased by 1%. The constrained optimization prob-
lem is solved using augmented Lagrangian method where an efficient implementation
of StoSAG and GPS are developed to solve the inner loop problem. Compared to
previous work [86, 22, 11, 49, 50], this piece of work extends the bi-objective optimiza-
tion from the optimal control problem to the optimal well placement problem whose
objective function surface is rougher.

• Our final contribution was to embed the optimization methodology in usable software
so that it can be applied to field problems.

1.3 Dissertation Outline

There are six chapters in this dissertation. Chapter 1 gives a brief review of the
literature pertaining to various optimization problems in oil field development, our initial
research objectives and our research contribution. In Chapter 2, the generalized field devel-
opment optimization problem is described and the methodologies considered in this disser-
tation are discussed, including Genetic Algorithm (GA), Stochastic Simplex Approximate
Gradient method (StoSAG) and Generalized Pattern Search Algorithm (GPS). In Chapter
3, we present the first major contribution of our work, a viable solution to select a fixed
number of wells from a given set of potential drilling paths and to determine the well types,
drilling order and control settings based on the Brugge model and a channelized reservoir
model. In Chapter 4, we present the second major contribution of our work, a viable solu-
tion to simultaneous optimization of well number, types and locations under assumed well
controls for both the Brugge model and the PUNQ model. Here, we also present an effi-
cient robust optimization procedure given a large number of realizations to represent the
geological uncertainty. In Chapter 5, we present an advanced parameterization of slanted
and multi-segmented well trajectories and a bi-objective well placement optimization process
to maximize the average NPV and minimize risk given multiple realizations of the reservoir
using lexicographic method. Chapter 6 summarizes our work and presents conclusions.

25
 CHAPTER 2

COMPUTATIONAL ALGORITHMS FOR FIELD DEVELOPMENT

OPTIMIZATION

In this chapter, we firstly present the mathematical description of a general field

development optimization problem, including the definition of the objective function, the
parameterization of the design variables and the related constraints. In this dissertation, the
field development optimization is solved using hybrids of three base algorithms, i.e., Genetic
Algorithm (GA), Stochastic Simplex Approximate Gradient (StoSAG) and General Pattern
Search (GPS). GA is introduced to handle the categorical variables naturally, StoSAG is
mainly used to optimize well locations and controls, and GPS is mainly used to fine tune
the well locations obtained by StoSAG. The three computational algorithms are described
respectively in the second part of this chapter while their hybrids are presented in the
following three chapters according to the problems considered.

2.1 Problem Description

The field development optimization considers maximizing the net present value (NPV)
over the presumed life of the reservoir to find the optimal number of wells, the corresponding
drilling order, types, locations and control settings. In this section, we present a general field
development optimization problem but this dissertation focuses on solving various subsets
of this general problem. Here, the general field development optimization problem is defined
by

max J (s, T , O, x, u) (2.1a)

s.t. c(s, T , O, x, u) ≤= 0, (2.1b)

26
 sj ∈ {0, 1} ∀ j = 1, · · · , Kmax , (2.1c)

Tj ∈ {0, 1} ∀ j = 1, · · · , Kmax , (2.1d)

x ∈ Ω, (2.1e)

ulow ≤ u ≤ uup , (2.1f)

where Kmax defines the maximum number of wells to be drilled; s is a Kmax -dimensional
categorical variable vector where the jth element, sj , is a binary variable defining whether
well j is to be drilled or not, i.e., sj = 0 corresponds to “not-to-drill” and sj = 1 corresponds
to “to-drill”; T is a Kmax -dimensional categorical variable vector where its jth element, Tj ,
is a binary variable defining the type of well j, i.e., Tj = 0 corresponds to “injector” and
Tj = 1 corresponds to “producer”; O is a sequence of ordinal number, i.e., a permutation of
{1, 2, · · · , Kmax }, where the jth element, Oj , represents well j is drilled as the Oj th well; x
is a bpl ∗ Kmax -dimensional vector of the well location parameters, where bpl is the number of
parameters representing the location of each well; Ω represents the potential set of locations
and/or paths that wells can take; u is a Nc ∗ Kmax -dimensional vector of the well control
parameters where Nc is the number of control steps (here Nc is assumed to be the same
for all wells), where u includes the injection/production rates of wells under rate control or
bottom-hole pressure of wells under pressure control at each control step; ulow and uup are
the lower and upper bounds of the control settings respectively; c(s, T , O, x, u) represents
the linear/non-linear constraints, where the notation ≤= means the constraints could be
either of the form c ≤ 0 or c = 0. Possible constraints include bounds on field production or
injection rates, balanced injection and production constraint, maximum water cut constraint,
or minimum well spacing constraints, etc.
The NPV for a given vector m of the reservoir model parameter is denoted by
J (m, s, T , O, x, u) and defined by,

27
   
Nt 
PNP n n n n n n

 Kmax
∆tn j=1 ro ·q o,j −cw ·q w,j −cg ·q g,j
X   X
J (m, s, T , O, x, u) =
 − sj Cw,j ,
 
tn 
n=1  (1+b) − N n n n n
P I
k=1 cwi ·q wi,k +cgi ·q gi,k
 365  j=1

(2.2)
where n denotes the nth time step of the reservoir simulation; Nt is the total number of time
steps; tn is the simulation time at the end of the nth time step; ∆tn is the nth time step size;
NP and NI denote the number of producers and injectors, respectively, which depend on
the “drill or not” variables represented by the components of s and the components of the
type indicator vector T ; q no,j , q nw,j , q ng,j , respectively, denote the average oil production rate
(STB/day), the average water production rate (STB/day) and the average gas production
rate (Mscf/day) at the jth producer over the nth simulation time step; q nwi,k and q ngi,k denote
the average water injection rate (STB/day) and the average gas injection rate (Mscf/day)
over the kth injector for the nth time step; ron is the oil price ($/STB); cnw is the water
disposal cost ($/STB); cng is the gas disposal cost ($/Mscf) (negative if gas is sold); cnwi and
cngi are the water injection cost ($/STB) and gas injection cost ($/Mscf), respectively; Cw,j
denotes the drilling cost for the jth well which depends on the trajectory variables defining
this well; and b is the annual discount rate. In this work, we use a commercial reservoir
simulator to simulate the dynamic performance of the flooding process. The NPV function
of Eq. 2.2 is computed using the output of a reservoir simulation run based on (s, T , O, x, u)
and a specific plausible reservoir model m.
The problem described in Eq. 2.1 can be extremely complicated when (s, T , O, x, u)
are all considered as design variables even for the deterministic case where the reservoir
model m is assumed known. Previous work in the literature focuses on solving subsets of
Eq. 2.1 where some design variables are assumed known. Only a few works (e.g., Hanea et
al. [31] and Leeuwenburgh et al. [44]) have considered drilling order optimization. When
(s, T , O, x) is known, Eq. 2.1 describes an optimal control problem with a smooth objec-
tive function and continuous optimization variables which can be solved efficiently using

28
gradient-based algorithms, given the availability of a reservoir simulator that provides the
relevant adjoint solutions needed to calculate the gradient of J(m, u) efficiently [42, 73, 10];
otherwise, one can use methods based on stochastic gradients, e.g. ensemble optimization
(EnOpt) [13], stochastic simplex approximate gradient (StoSAG) method [24], simultane-
ous perturbation stochastic approximation (SPSA) [47]. Another possibility is to use a
derivative free algorithm such as particle swarm optimization (PSO) [84, 80], covariance ma-
trix adaptation-evolutionary strategy (CMA-ES) [25, 77] and Genetic Algorithm (GA) [86].
When (s, T , O, u) is known, Eq. 2.1 describes an optimal well placement problem where
the objective function may not be differentiable for all value of x; in fact, the optimization
variables can be either continuous or discrete depending on how the well trajectory is pa-
rameterized. For vertical wells, the well location is usually defined by the discrete (i, j, k)
indices of the gridblocks penetrated by a well [5, 7, 47] where in most papers the well is
assumed to be completely vertically penetrating. Emerick et al. [21] used the (i, j, k) grid-
block indices of the starting point and end point to describe slanted well trajectories. In this
case, the design parameters are again discrete variables since gridblock indices can only take
integer values. However, for slanted wells, the well trajectory can also be parametrized as
continuous variables, for example using the (x, y, z) coordinates of the center point of each
well, the perforation length l, the inclination angle ϕ and the azimuth angle θ [66, 27]. How-
ever, this parametrization requires a special scaling treatment for the optimization process
if the (stochastic) gradient-based method is used (Forouzanfar and Reynolds [27]). Vlem-
mix et al. [81] parameterized a single multi-segmented well as the (x, y, z) coordinates for
a set of trajectory points where the number of trajectory points is equal to the number of
gridblocks through which the trajectory passes. However, a smoothing step is needed after
every update of the coordinates of the trajectory points, to ensure the drillability of the
well. Yeten et al. [86] optimized the well trajectory for a multilateral unconventional well
where the main bore is described by coordinates of the entry point (hx , hy , hz ), the length
of the trajectory projected onto the x-y plane lxy , the azimuth angle θ, and the depth to
the trajectory endpoint tz ; and the lateral is described by the coordinates of the junction

29
point (there can be multiple laterals from one junction, the number of junctions and the
number of laterals emanating from each junction are to be determined). In this case, the
optimization variables are mixed discrete and continuous variables. There currently exists
no adjoint code to calculate the gradient of the NPV function with respect to the well lo-
cation variables. Gradient-based algorithms using stochastic gradient and derivative free
algorithms are widely used. When (s, T , O) is known, Eq. 2.1 reduces to the problem of
jointly optimizing well controls and well locations. The problem of optimizing well locations
leads to a rougher objective function surface (Forouzanfar and Reynolds [28]) than the one
for the problem of optimizing only well controls. For problems involving rough objective
functions, gradient-based algorithms generally work poorly although a gradient-based algo-
rithm tends to be highly efficient for the pure well control optimization problem. Thus, it
may be advantageous to use a sequential algorithm to jointly optimize well controls and
trajectories where a gradient-based algorithm is used for the well controls and an algorithm
with a better capability for finding global optima of a noisy objective function is employed
for the well trajectory optimization. Another reason for decoupling the joint optimization
problem into two subproblems, i.e., well control optimization and well placement optimiza-
tion is that, in practice, well control optimization is not conducted until all wells have been
drilled. A third reason is that when all the optimization variables (s, T , O, x, u) need to be
determined, Eq. 2.1 is in fact a mixed-integer problem with both continuous and categorical
variables, referred to as the full field development optimization problem. In this case, the
gradient (or stochastic gradient) based method is no longer suitable since the gradient with
respect to the categorical variables is not defined. As discussed below, some authors have
tried to replace the categorical variables as continuous variables, but it is not clear that such
a procedure can yield optimal results.
Only a few papers have included well status, types and drilling order in addition to
well paths as design variables [35, 76, 2, 31, 44]. The first two papers (Isebor et al. [35] and
Shirangi et al. [76]) assign each of the Kmax wells a label, preset their drilling order and
opening time, and optimize the well status and type (-1 represents a producer, 0 represents

30
do-not-drill and 1 represents an injector), locations and controls for these Kmax wells. Even
though the well status and types are considered as ternary variables, at each iteration of the
Particle Swarm Optimization (PSO), the ternary variable are treated as continuous variables
and then truncated back to the nearest integer within the set {−1, 0, 1}. What is more, to
switch the drilling order of well i and well j (i 6= j), the locations and controls for well i
and j have to be switched which is quite difficult for PSO or MADS when well i and well
j are far away from each other. Moreover, even if well i is not drilled (i.e., well status
of well i corresponds to do-not-drill), it still holds (reserves) Td days of drilling time (i.e.,
the drilling rig does not work for the next Td days) and because of the “rig down time,” the
solution obtained must be at least somewhat suboptimal. In the last two of these papers, well
status is not optimized while well type and drilling order were determined using continuous
variables which they refer to as priority variables (see section 1.1 for an explanation of
priority variables). As well status, types and drilling order are discrete variables, treating
them as continuous variables and then truncating back to discrete values at each iteration
of an optimization algorithm, does not seem ideal. Instead, we believe that it makes more
sense to treat them directly as discrete categorical variables which enables the use of the
genetic algorithm.
Probably one of the most important concerns in the decision-making process based on
the optimization results would be the uncertainty in the reservoir geological description. The
most common approach for handling geological uncertainty is to generate multiple reservoir
descriptions and do robust optimization in which, the expected value of NPV is optimized.
As noted previously, we use the terminology expected value whereas the expected value is
actually replaced by its standard estimator, i.e., the average of the NPV where the average is
over the set of reservoir models (vectors of reservoir model parameters) mk , k = 1, 2, · · · , Ne
which represent the uncertainty. Thus, the expectation of NPV is defined by

Ne
1 X
JE (s, T , O, x, u) = J(mk , s, T , O, x, u) (2.3)
Ne k=1

31
 Robust optimization is much more expensive than deterministic optimization in Eq.
2.1 since each objective function evaluation requires Ne simulation runs where Ne generally
ranges from ten to one, or two hundreds. There are at least three major concerns with robust
optimization, one is that maximizing the expected objective function over the given ensemble
still may incur considerable downside risk [49], and a second is that optimization over a
large set of realizations of m is generally not computationally feasible for field development
optimization problems and a third concern is that life-cyle robust optimization may yield
unacceptably low short-term production. The last issue is not dealt with here, but is generally
addressed by bi-objective optimization (Essen et al. [22], Chen and Reynolds [11] and Liu
and Reynolds [49]).

2.2 Basic Methodology Considered

The general problem stated here is a mixed-integer problem with a high-dimensional

search space which includes binary variables (“drill-or-not” status and types), ordinal vari-
ables (drilling order), numerical discrete variables (locations for vertical wells) and numerical
continuous variables (well control settings). Considering that (i) GA handles categorical vari-
ables naturally; (ii) the Stochastic Simplex Approximate Gradient (StoSAG) based method
has been proven to be efficient in optimizing well locations and controls, but searches rel-
atively poorly when close to a local optimum due to gradient estimation error, and (iii)
General Pattern search method (GPS) is a good local search method once we are close to an
optimum, but is computationally expensive for high-dimensional optimization problems, we
focus on developing hybrids of these three main algorithms when solving different aspects of
the optimization problem defined in Eq. 2.1.
The performance of an algorithm (e.g. steepest descent and general pattern search)
may depend crucially on scaling. For example, if the producers are under bottomhole pressure
control and the injectors are under rate control, changes to BHP may produce much larger
variations in the NPV than do similar changes to injection rate; same thing may happen to
the coordinates of the well end points along the (x, y) directions and the z direction. Hence,

32
if the optimization problem is poorly scaled, we introduce a new variable vector which has
similar magnitude in all its components and then solve the optimization problem in terms of
the new variable vector. In this dissertation, for all the optimal well control problem and the
well placement optimization of slanted wells and multi-segmented wells, scaling is adopted;
see Chapter 5 for detailed explanation.

2.2.1 Genetic Algorithm (GA)

The genetic algorithm (GA) is a metaheuristic algorithm inspired by the process of

natural selection which is often used to solve global optimization problems. GA evolves
iteratively starting from a population of randomly generated individuals through a set of
operators: selection, crossover, mutation and replacement which are designed to move the
population away from a local optima where a traditional hill climbing algorithm might get
stuck. Each individual in the population corresponds to a set of decision variables (or scaled
variables), which in GA terminology is referred to as a chromosome or genotype, denoted as
c. The parameterization of the decision variables is referred to as encoding. Traditionally,
solutions are represented as strings of binary variables, but other encodings are possible, i.e.,
permutation encoding and real value encoding. Different encodings, crossover and mutation
strategies are used to account for different types of variables (e.g. well “on/off status,” well
types, drilling order, locations and controls), depending on the specific problem considered.
In each generation, the fitness of every individual in the population is evaluated and
a parent pool is selected from the current population by a stochastic procedure in which the
more fit individuals are likely to be selected. Next, a new generation is formed based on re-
combining and mutating chromosomes of the individuals from the parent pool. Furthermore,
the best chromosome will randomly replace one individual in this new generation. The new
generation of candidate solutions forms the population from which the parent pool will be
selected at the next iteration of the algorithm. This generational process is repeated until a
termination condition is reached. Common termination conditions include specifying a fixed
number of generations based on available computational resources or terminating GA when

33
iterations (generations) no longer produce better results. A flow chart of GA is shown in
Fig. 2.1 and a general description of GA using binary encoding as an example is provided
in Algorithm 1.

Figure 2.1: Schematic of GA.

Algorithm 1 Algorithm outline for GA with binary encoding

1. Preset population size Np , maximum number of generations Ng , crossover rate pc and

mutation rate pm .

2. Initialization: generate Np feasible individuals with chromosome c randomly sampled

from {0, 1}.

3. Evaluation: evaluate the objective function of Eq. 2.2 for each individual in the current
population.

4. Selection: select Np individuals from current population to form a parent pool.

5. Crossover: recombine the candidate solutions from the parents pool with a probability
of pc using a modified one-point crossover to produce offsprings.

6. Mutation: the produced offsprings invert bits at positions randomly chosen with proba-
bility pm .

7. Termination criteria check. If number of generations exceeds Ng , terminate; otherwise,

go to step 3.

Usually in Genetic Algorithms (GA), the crossover probability is set fairly high, 0.6
to 0.9, and mutation probability of a gene is kept far smaller, 0.001 to 0.01. A higher
crossover probability helps to explore the solution space more globally while a small crossover
probability may slow down the evolution process. A very small mutation rate may lead to
genetic drift (which is non-ergodic in nature) while a high mutation rate may lead to a loss of
good solutions, unless elitist selection is employed. Finding an appropriate population size is

34
a difficult task. If it is too small, GA may not be able to reach high quality solutions. If it is
too large, GA converges so slowly that it mat require an infeasible number of forward model
runs (reservoir simulation runs in this work) in order to obtained good optimum. It has
been shown, both theoretically and empirically, that the optimal size differs from problem
to problem. A somewhat widely-accepted rule of thumb on selecting the population size is
that it should be set proportionally to the problem’s size and difficulty. However, problem
difficulty is very hard to estimate for real-world problems. The general rule in GA literature
for the optimum population size is to set it equal to two times the number of bits used in
a chromosome [59]. However, this rule of thumb is computationally expensive for a high-
dimensional optimization problem, e.g. problems considered in this work, and we find that
good results can be generated with a considerably smaller population size.

2.2.2 Stochastic Simplex Approximate Gradient (StoSAG)

In this section, we simply use u to denote the optimization variables. For the well
control optimization problem, u represents the (scaled) well control settings of all wells at
all control steps (e.g well bottomhole pressure, well injection/production rate); for the well
placement optimization problem, u refers to the (scaled) well trajectory parameters for all
wells; for the simultaneous optimization problem, u includes both the (scaled) well control
settings and the (scaled) well placement variables. Following the formulation of the gradient
of Eq. 2.3, the gradient of JE is equal to the sum of the gradient of NPV calculated for each
model, i.e.,
Ne
1 X
∇u JE (u) = ∇u J(u, mk ). (2.4)
Ne k=1

In Eq. 2.4, the ∇u is used to represent the true gradient and ∇u JE (u) represents an uphill
direction, i.e., if u` is the estimate of the optimal value u∗ obtained at iteration `, JE (u` )
increases in the direction ∇u JE (u` ). Thus the steepest ascent algorithm is given by

ul+1 = u` + α` ∇u JE (u` ), for ` = 0, 1, · · · , (2.5)

35
where u0 is the initial guess and α` is the step size which can be chosen by a line search
procedure. In the StoSAG algorithm, ∇u J(u` , mk ) is approximated by a stochastic gradient
denoted by d`k and ∇u JE (u` ) in Eq. 2.5 is replaced by

Ne
` 1 X
d = d` , (2.6)
Ne k=1 k

so the steepest ascent algorithm of Eq. 2.5 is replaced by

d`
ul+1 = u` + α` , (2.7)
kd` k∞

for ` = 0, 1, · · · , until convergence. α` is commonly chosen by backtracking where a max-

imum allowable step size, α0 , is usually set between 0.1[uup − ulow ] and 0.2[uup − ulow ] to
enable a large initial trial update, and a maximum allowable number of step size cuts is
usually set as 5.
To generate a simplex gradient of J(u` , mk ), we need to generate different perturba-
tions of u` for each mk . We define the perturbed vector by û`j,k , j = 1, 2, · · · , Npert where
the hat denotes a perturbed vector and the subscript j denotes the jth perturbation vector
for the realization mk . Next, we define the Nu ×Npert matrix ∆Uk` and the Npert -dimensional
column vector ∆Jk` respectively by

∆Uk` = [û`1,k − u` , û`2,k − u` , · · · , û`Np ,k − u` ],

and

∆Jk` = [J(û`1,k , mk ) − J(u` , mk ), J(û`2,k , mk ) − J(u` , mk ), · · · , J(û`Np ,k , mk ) − J(u` , mk )]T .

Then the approximate simplex gradient, d`k , in Eq. 2.6 is given by

d`k = (∆Uk`,+ )T ∆Jk` , (2.8)

36
where throughout the superscript + denotes the Moore Penrose pseudo-inverse which can
be obtained using singular value decomposition, Npert represents the total number of per-
turbations for each realization mk , and Nu represents the dimension of the design variables
vector. The perturbations can be generated by any method, for example, we can use

1/2
û`j,k = u` + LZj,k
`
= u` + CU Zj,k
`
, (2.9)

`
where, Zj,k is a Nu -dimensional vector of independent standard random normal deviates
and L is the lower triangular matrix of the Cholesky decomposition of a specified covariance
matrix CU which controls the perturbation sizes and correlations between the components
of the perturbed vectors [9]. CU is a block diagonal matrix with Nw blocks and each block,
CUw , corresponds to the correlation matrix over the Nc control steps of well w where each
element, CUi,j,w is given by

"    3 #
3 |ti − tj | 1 |ti − tj |
CUi,j,w = σ 2 1− + , (2.10)
2 Ns 2 Ns

where, σ is the standard deviation of the perturbations, e.g., 0.03 for the normalized vari-
ables in the scale of [0,1], i and j represent the ith and jth control steps respectively,
i = 1, 2, · · · , Nc , j = 1, 2, · · · , Nc , ti and tj represent the times corresponding to the middle
of ith and jth control step interval respectively, and Ns is the length of time over which we
wish the controls of well w to be temporally correlated.
When Ne = 1, the optimization process is referred to as deterministic optimization
where Npert is usually set to a number ranges from 5 to 50. Do and Reynolds [19] used only 5
to 10 for the deterministic case where there is only a single reservoir model, but Fonseca et al.
[24] found that we may need far more than 50 perturbations in order to obtain a stochastic
gradient which has a direction fairly close to the true gradient. When we consider robust
optimization where the number, Ne , is such that Ne ≥ 10, following Fonseca et al. [24], we
use only one perturbation per reservoir model. Because the stochastic gradient of JE (u` )

37
is only an approximation of the true gradient, the search direction dl may not point uphill
from u` , and if not, it may not be possible to obtain a higher value of JE (u` ) by searching
in the direction dl . When this occurs, we may try to find an uphill direction by generating
new perturbations and computing new stochastic gradient, dl . In our implementation, if we
cannot improve the expected NPV after five trials, the StoSAG algorithm is terminated. The
StoSAG algorithm is summarized in Algorithm 2 and illustrated in Fig. 2.2. In Fig. 2.2, we
suppose there are only two wells (one injector under rate control and one producer under
bottomhole pressure control) and only one control step, the x−axis represents the injection
rate and the y−axis represents the production BHP, the 2D contour map represents the NPV
(for illustration only). In Fig. 2.2, lines in red and blue respectively represent the gradient
direction and the StoSAG direction at different estimates of the design variables, where in
this schematic, the angle between the gradient direction and the StoSAG search direction is
less than 90◦ , i.e., the StoSAG direction is in an uphill search direction.

Figure 2.2: Illustration of the StoSAG search direction.

For the case of a single reservoir model, the expectation of the StoSAG search direction
 
` i 2 i=Npert ∗Ne
d is the true gradient ∇u J with an error term O max{kδu k2 }i=1 [24]. Hence, the
larger is the perturbation size, the larger is the error term. When optimizing the well place-

38
Algorithm 2 Outline for the StoSAG Algorithm

Initialize the parameters: nsc , ncuts = 0, Nres , is = 0, α` = α0 , where nsc is the maximum
allowed number of step size cuts, ncuts is the number of step size cuts, Nres is the
maximum number of the consecutive search direction re-computation, is is the number
of search direction re-computation, α0 is the maximum step size.

1. Generate Npert ∗ Ne samples of perturbation at current estimate u` and compute the

search direction with Eq. 2.6. Set is = is + 1.

2. Compute a trial update utrial,`+1 using Eq. 2.5 (truncate each element of utrial,`+1 to its
nearest integer for discrete variables).

3. Check if the trial update is acceptable.

• If JE (utrial,`+1 ) > JE (u` ),

– set u`+1 = utrial,`+1 , ` = ` + 1, is = 0, ncuts = 0, α`+1 = α0 and go to step 1.
• Otherwise,
– set ncuts = ncuts + 1,
– if ncuts > nsc and is > Nres , terminate the optimization algorithm,
– if ncuts > nsc and is < Nres , then go to step 1,
– if ncuts ≤ nsc , set α` = α` /2, then go to step 2.

4. Terminate if the maximum number of simulations or iterations are used.

ment of vertical wells where the location variables ((i, j) gridblock indicies) are discrete, the
±1 Bernoulli distribution is often used to generate random perturbations of the current esti-
 
i=N ∗N
mate of the well locations. In this circumstance, it is possible for O max{kδui k22 }i=1 pert e
to be quite large. Meanwhile, the objective function surface of the well placement problem is
quite rough compared to the well control optimization problem, and this can lead to a poor
StoSAG performance. In this sense, we may not be able to obtain a result close to optimum.

2.2.3 General Pattern Search (GPS)

General Pattern Search (GPS) [15] is a class of direct search methods, originally
introduced by Torczon [79]. Each iteration consists of an optional search step and a poll
step. Both steps evaluate the objective function at a finite number of trial points on a mesh
M` in order to find an improved mesh point. In this work, M` is given by

39
 M` = u` ± ∆` ei : i ∈ {1, 2, · · · , Nx } ,

(2.11)

where, u` is the incumbent point at the lth iteration, ∆` is the mesh size at lth iteration, Nx
is the number of elements in u` , and ei represents the unit vector where the ith element of ei
is 1. The difference between search and poll steps is in the way the trial points are selected.
Any strategy, including none, can be used in the search step, and it is often tailored to the
particular application to obtain specific efficient algorithms. Since GPS is used to fine tune
the well locations obtained by StoSAG, we only consider the poll step in this work. Our
implementation of GPS is described below.
There are Nx coordinate directions in total and each coordinate direction is associated
with two objective function evaluations, J(u` +∆` ei ) and J(u` −∆` ei ). If both J(u` +∆` ei )
and J(u` −∆` ei ) are smaller than J(u` ), then we move on to the (i+1)th coordinate direction.
If J(u` +∆` ei ) > J(u` ) and J(u` +∆` ei ) > J(u` −∆` ei ), then incumbent point u` is updated,
i.e., u`+1 = u` + ∆` ei and consecutive trial updates u`+1 = u`+1 + k∆` ei , k = 1, 2, · · · , 5 on
the mesh M` along ei direction are taken until objective function stops increasing. Similarly,
if J(u` −∆` ei ) > J(u` ) and J(u` −∆` ei ) > J(u` +∆` ei ), then incumbent point u` is updated,
i.e., u`+1 = u` − ∆` ei and consecutive trial updates u`+1 = u`+1 − k∆` ei , k = 1, 2, · · · , 5
on the mesh M` along −ei direction are taken until objective function stops increasing. One
should note, in the original GPS algorithm, ∆` is initially set large to search globally and ∆`
shrinks size to search locally whenever a local optimizer on M` is found, i.e., when no trial
update that improves the objective function is found after searching along Nx coordinate
directions. GPS algorithm is only considered to converge when ∆` is smaller than a user-
defined minimum allowable mesh size. The value of the initial mesh size ∆0 depends on the
actual problem to be solved. For example, in the well placement optimization of vertical
wells, when well locations are represented by the (i, j) gridblock indicies, we can start from
a small mesh size (e.g., ∆0 = 1 gridblock) if the GPS method is used to fine tune the well
locations after the StoSAG optimization or start from a larger mesh size (e.g., ∆0 = 5) if

40
GPS method is used as the standalone optimizer. The way to shrink the mesh size ∆` is
also problem dependent. For example, for the case where the mesh size has to be a positive
integer, one can use ∆` = ∆` − 1; otherwise if the mesh size is a continuous variable, one
can use ∆` = ∆` /2. For detailed implementation of GPS in this work, please see Algorithm
3. Fig. 2.3 shows an illustration of the GPS algorithm for the well control optimization
where there is one injector under liquid rate control and one producer under BHP control
considering one control step.

Algorithm 3 General Pattern Search Algorithm

Set ` = 0, idx = 0, F LGP S = 0 and ∆0 . Initialize the design variable u` , set the initial
mesh size ∆` = ∆0 and the minimum allowable mesh size ∆min . Set the maximum
allowable consecutive trial updates along a coordinate direction α0 = 5, create the
mesh M` = {u` ± ∆` ei : i ∈ {1, 2, · · · , Nx }}.

Do

• Set idx = idx + 1, set i equal to the modulus of idx divided by Nx .

• Evaluate J(u` ± ∆` ei ).
If J(u` + ∆` ei ) > J(u` ) or J(u` − ∆` ei ) > J(u` ),
– set F LGP S = 0 and k = 0.
– If J(u` + ∆` ei ) > J(u` − ∆` ei ), set u`,k = u` + ∆` ei , d` = ei ;
– otherwise, set u`,k = u` − ∆` ei , d` = −ei .
– For k = 1, · · · , α0
∗ If J(u`,k−1 + ∆` d` ) > J(u`,k−1 ), set u`,k = u`,k−1 + ∆` d` ;
∗ else, break the for-loop.
– EndFor
– Set u`+1 = u`,k and ` = ` + 1.
Otherwise,
– set F LGP S = F LGP S + 1.
• If F LGP S = Nx ,
– u` is a local optimizer of the current mesh, then reduce current mesh size ∆` ,
e.g. ∆` = ∆` /2.
– If ∆l ≥ ∆min , then re-create the mesh M` = {u` ± ∆` ei : i ∈ {1, 2, · · · , Nx }},
and set F LGP S = 0;
– otherwise, terminate the algorithm.

EndDo

41
Figure 2.3: Illustration of the General Pattern Search (GPS) search direction.

42
 CHAPTER 3

SELECTING WELLS FROM A GIVEN SET OF POTENTIAL PATHS

Reservoir engineers are often asked to select “the best” Nw well paths from a set of
Kmax potential paths where Kmax > Nw , determine whether each of the Nw wells should be
an injector or a producer and perhaps determine the optimal drilling order. The number
of wells, Nw , that will be drilled is determined by the drilling budget. Often optimizing
operating well controls is not done until the well paths, types and drilling order have been
selected. In this chapter, our focus is solely on the problem: given Kmax plausible well
paths, select the optimal Nw wells, the type of each of these wells, the drilling order and
find the optimal wells controls where the objective function we wish to maximize is the
net present value (NPV) of production over the life of the reservoir. That is to say, we
consider the optimization problem given in Eq. 2.1 with the design variables as (x, T , O).
Well controls are optimized only after (x, T , O) is optimized for a given set of well controls.
The set of Kmax potentially good well paths is determined by drilling engineers working
with geoscientists and production engineers focused on drilling sweet zones with good flow
capacity. As an aid to finding good potential well paths, one may use a reservoir quality
map generated from geological information and static data to identify high permeable regions
with significant mobile oil; see [56, 39, 40, 17].

3.1 Problem Description and Methodology

In this chapter, each path is described by its perforations where the perforation seg-
ments can be pre-specified with different shapes, i.e., vertical, horizontal, slanted or curved.
As the objective is to select the Nw “best” well paths out of Kmax possibilities rather than
to optimize the trajectories of Kmax paths, we simply map each potential path to its label.

43
Hence the set of potential paths is represented as a set of positive integers, denoted as Ω,
where the numbers are used for labeling or identification only, and they do not indicate
quantity, rank, or any other measurement. Thus,

Ω = {i|i = 1, 2, · · · , Kmax }

where, each i in Ω corresponds to a specific path. Let x = [x1 , x2 , · · · , xNw ] denote any Nw
paths out of Kmax possibilities. The entries xj , j = 1, 2, · · · , Nw must all be distinct but can
correspond to any value in Ω encoded as its binary representation. One realization of x can
Kmax !
be obtained by making Nw selections without replacement from Ω, so there are Nw !(Kmax −Nw )!

possibilities for x. In this chapter, xj is denoted as a xbj , where the superscript b is used to
emphasize it is a binary vector. It is important to emphasize that xbj is actually the binary
representation of one of the numbers in Ω where the number of labels in Ω determines the
length of xbj . If for example, there are 32 possible paths, then we need a binary string of
length 5 in order to define a unique binary string for each of the 32 (25 ) potential paths.
If there are only 30 potential paths, we still require a binary string of length 5 to uniquely
describe the 30 potential paths but in this case, there will be two coded strings which will
pertain to none of the potential paths. In general, we need a string of length bpl to uniquely
define Kmax potential paths, where bpl is the integer such that

2bpl−1 < Kmax ≤ 2bpl . (3.1)

Denote the vector T as a Nw -dimensional binary vector of the form T = [T1 , T2 , · · · , TNw ]
where each entry Tj , j = 1, 2, · · · , Nw determines whether the well placed at the potential
path corresponding to xbj is an injector or a producer. Tj is a two-level categorical variable,
where Tj = 0 indicates a producer and Tj = 1 indicates an injector. The vector T has 2Nw
possibilities. Similarly, denote O = [O1 , O2 . · · · , ONw ] as an Nw -dimensional vector where
each entry Oj determines the drilling sequence of the well with its path corresponding to
xbj . For example, Oj = 5 (when Nw ≥ 5) means the well represented by xbj is the 5th well

44
drilled. As Oj is an ordinal number which tells the drilling sequence of the well represented
by xbj and we are going to select Nw wells, O always represents a permutation of the numbers
from 1, 2,· · · , Nw . We assume there is only one rig and wells are drilled one by one which
means that all entries of O have to be distinct. Hence a realization of O can be obtained
by making Nw selections without replacement from a list of numbers from 1 to Nw , where
there are Nw ! possibilities for O. We assume each well is drilled in Td days which means
that the well with drilling sequence Oj will be drilled at time Td (Oj − 1) days. We let
u = [u(1) , u(2) , · · · , u(Nw ) ]T denote the vector of well controls for Nw wells where each entry
u(j) denotes the control settings of the well placed at the location represented by xbj . Since
(j) (j) (j)
there can be Nc control steps, u(j) = [u1 , u2 , · · · , uNc ] where Nc is the total number of
control steps, u is a Nw ∗ Nc -dimensional vector of continuous variables, determining the
injection/production rates of the wells under rate control or the BHP’s of wells under bot-
tomhole pressure control. Note before a well is drilled, its controls are immaterial and this
well is fixed to shut-in status. Therefore, the number and the length of the active control
steps for each well may be different.
The problem considered in this chapter is a mixed-integer problem with a high-
dimensional design variable vector c = (x, T , O, u). The genetic algorithm (GA) is adopted
as the main methodology to optimize (x, T , O). Even though it is possible to simultane-
ously optimize the control variables u together with the well locations, types and drilling
order (x, T , O), we consider a sequential optimization workflow where (x, T , O) is firstly
optimized with fixed u and then u is optimized based on the optimal (x, T , O) obtained,
for the reasons discussed below. Note if we choose Nw = 12 wells out of Kmax = 64 locations
as in the second numerical example presented later, then bpl = 6, dim(x) = Nw ∗ bpl = 72,
dim(T ) = 12, dim(O) = 12, which sums up to 96. If each of the 12 wells had 20 control
steps, them dim(u) ≈ 240. If u is included in the design variable, i.e., c = (x, T , O, u),
then c would contain 336 components which would require a population size of 772, based on
common rule of thumb for GA that the population size be twice the length of c. If 50 genera-
tions were produced using GA with this population size, over 30,000 reservoir simulation runs

45
would be required which would likely make GA practically infeasible. The computational
cost would of course be reduced by using only one to two control steps per well, but this
would in many cases yield a suboptimal result (Oliveira and Reynolds [63]). A second reason
for sequential optimization is that the NPV function is generally a fairly smooth function
of the controls [42, 73], so the gradient-based methods using gradients computed from the
adjoint method or from a stochastic approximation are applicable and relatively effective for
optimizing controls but is relatively inefficient for optimizing categorical variables. Hence, a
sequential optimization procedure can break a high-dimensional optimization problem into
two smaller problems, and enables us to use different algorithms suitable for each specific
sub-problem, i.e in the first stage of optimization, the optimization variables (x, T , O) are
all categorical variables, hence GA, an algorithm with global search ability, can be employed
whereas in the second stage of optimization, the vector of well controls, u, is optimized using
the Stochastic Simplex Approximate Gradient (StoSAG) [24]. A third reason for not includ-
ing u directly in c is that in industry, control optimization is often done after the wells have
been drilled. Thus our general procedure is to maximize J(x, T , O) with each component
of u set equal to its upper bound for a BHP-controlled injector and set equal to its lower
bound for a BHP-controlled producing well, and then maximize J(u) based on the optimal
(x, T , O) obtained. Note that there is no guarantee that sequential algorithms converge.
Isebor et al. [34] stated that simultaneous optimization of well locations and controls gave
a higher NPV than sequential optimization. Different than [34] who focused on optimizing
vertical wells, Humphries and Haynes [32] optimized the well placement and control settings
for non-conventional wells and found that sequential approaches gave a higher NPV than
simultaneous optimization for complex well parameterizations. After providing results of our
workflow for two examples, we comment for sequential versus simultaneous optimization.
In this chapter, we used mixed encodings to account for different types of variables
in order to fully characterize the variable set c = [x, T , O]. The location variables x and
the type variables T are encoded as binary variables, and O is parameterized as ordinal
numbers using permutation encoding. Note x is encoded with binary strings instead of

46
ordinal numbers to enable the generation of new potential locations which do not belong
to any of the parents in the crossover operation. The commonly used crossover operation
for ordinal numbers, Partially Mapped Crossover (PMX), can only recombine the existing
potential locations in the parent pair. If well locations are parameterized as ordinal variables
and PMX is applied, when Kmax >> Nw , it is likely that the initial population of candidate
solutions (randomly generated) will not include some of the potential paths and it is highly
unlikely that these paths will be brought back to the gene pool in following iterations by
mutation alone.

3.1.1 Optimization of well locations, types and drilling order

In this section, we give a detailed description of the five GA operators: initialization,

selection, crossover, mutation and replacement, adapted to solve for the specific problem
of this chapter. Note that the three parts (x, T , O) of design variable vector c have dif-
ferent encodings and each encoding requires a different crossover and mutation procedure.
Hence, GA operators suitable for the mixed encodings are designed to fit for the purpose of
simultaneous optimization of c = (x, T , O).
1. Initialization
In the implementation of GA, we need to preset the population size Np , the maximum
number of generations Ng , the crossover rate pc and the mutation rate pm . Initially, Np
individuals will be generated to form the initial population in which each entry of chromosome
c has to be sampled randomly and independently.
As both x and T are binary strings, each of their entry is sampled randomly from
{0, 1}. Note, there are two types of constraints in the parameterization of x. Firstly, when
Kmax is strictly less than 2bpl , binary encodings can be generated that do not correspond
to any of the potential locations. Secondly, we do not allow two wells to be drilled at the
same location, or two potential locations cannot be selected twice in a single chromosome,
i.e., xbi 6= xbj for i 6= j. Any infeasible trial chromosomes (chromosomes violating any of
these two constraints) are abandoned and only feasible chromosomes are accepted in the

47
initial population. Due to the ordinal feature of O, the drilling sequence is encoded as a
permutation of the sequence [1, 2, · · · , j, j + 1, · · · , Nw ] where the jth entry represents the
drilling order of the well placed at the path represented by xbj . For example, if Nw = 8, a
possible chromosome segment for the drilling order is

O = [6 4 5 3 7 8 1 2]. (3.2)

According to Eq. 3.2, well xb1 is drilled the 6th (i.e., 5Td days), well xb2 is drilled 4th (i.e.,
3Td days), well xb3 is drilled 5th (i.e., 4Td days) and so on. For each individual of the initial
population, O is sampled as a random permutation. It is worth mentioning that in the
initialization process, we force the first well drilled to be a producer as it generally would
make no sense to open an injector first. Then a random sample of (x, T ) are concatenated
with a random sample of O to form a random chromosome c. After all Np individuals are
generated, the objective function defined in Eq. 2.2 for each individual is evaluated.
2. Selection
After the current generation has been evaluated, a portion of this generation is se-
lected in order to breed a new generation where the more fit individuals are likely to be
selected. Stochastic Universal Sampling (SUS) is a development of fitness proportionate
selection where fitness can be defined as any function positively correlated to the objective
function J. To apply SUS, we first order and rank the Np chromosomes according to their
NPV in ascending order so that the chromosome with the highest NPV has a ranking of
Np , the chromosome corresponding to the second highest NPV is given a ranking of Np − 1
and so on. The ranking/ordering of two individuals of equal NPV’s is determined randomly.
The ranking is used as the fitness function. We let Fj denote the fitness of chromosome
j where the chromosomes j = 1, 2, · · · , Np are ordered according to their fitness i.e.,
F1 ≥ F 2 ≥ F3 ≥ · · · ≥ FNp . The cumulative fitness of individual j is denoted by Fp (j)
PNp
and defined by Fp (j) = jk=1 Fk , so the total fitness is Fp (Np ) = k=1
P
Fk . SUS then uses a
single random value r to sample the ranked individuals by choosing them at evenly spaced

48
intervals. This gives a weaker member of the population (according to its fitness) a chance
to be chosen and thus avoids having the candidate space saturated by the fittest members.
For illustration purpose, consider selecting 8 parents from a population of 8, i.e., Np = 8.
In Fig. 3.1, A, B, C, D, E, F, G and H represent the eight individuals with descending
fitness (values 8, 7, 6, · · · , 1, respectively). Note that the total fitness is Fp (Np ) = 36 for
 
this case. First, a random number r ∼ U 0, FpN(Np p ) is generated. For i = 1, 2, · · · , Np ,
the ith selected individual is the one with fitness Fj where j ∈ {1, 2, · · · , Np } is the index
(i−1)Fp (Np )
satisfying Fp (j − 1) < r + Np
≤ Fp (j), j ∈ {1, 2 · · · , Np } where Fp (0) = 0. The Np
selected individuals compose the parent pool to breed the next generation. The selected
Np
parents are then shuffled and divided into 2
parent pairs. For the r selected in Fig. 3.1,
the individuals selected as parents are A, A, B, C, C, D, F and G.

Figure 3.1: SUS selection for the GA algorithm where A, B, C, D, E, F, G, H represent the
ranked individuals in a population with their lengths proportional to their ranks in descend-
ing order; Fp represents the sum of ranks for all individuals and N represents the number of
parents to keep in the parent pool.

3. Crossover
The crossover operation produces offsprings by recombining a selected pair of can-
didate solutions from the parent pool at the crossover rate pc , which is the probability
of performing a crossover operation. For each pair of selected parents, a random number
ri , i = 1, 2 · · · , Np /2 is generated from the uniform distribution U[0, 1] and compared with
the crossover rate pc . If ri > pc , the pair of offsprings are kept the same as the parent
pair; otherwise, a random position is selected on chromosome c. If the randomly generated
crossover point falls in the part of c corresponding to the drilling order O as in Fig. 3.2(a),
then Partially Mapped Crossover (PMX), which is discussed later, is conducted for O. If the

49
crossover point falls in between two genes corresponding to (x, T ) as in Fig. 3.2(b), then a
modified version of one-point crossover is applied as discussed later.

(a) Crossover point in O (b) Crossover point in (x, T )

Figure 3.2: Illustration of the crossover operation for the mixed encoding chromosomes
(x, T , O). The red and blue colors represent (x, T ) of the two parent chromosomes re-
spectively. The green and orange colors represent (O) of the two parent chromosomes,
respectively.

One-point crossover is the most common crossover mechanism used in GA, and is
applied wherever the random position selected is in the (x, T ) part of the chromosome c.
In this case, the genes on the right of selected position are exchanged between the pair of
parents (see Fig. 3.3). However, if the crossover point falls in the x part, the produced
offsprings may possibly violate one of the two types of constraints mentioned previously, i.e.,
(i) xbj may correspond to a chromosome which is not one of the potential paths in Ω or (ii)
c may be a chromosome such that xbj = xbi for i 6= j. Whenever (i) is violated, we abandon
the infeasible trial offsprings and repeat the crossover operation where the crossover point
is selected in the x part of c until feasible offsprings are formed. Quite often the selected
pair of parents have wells in common at different locations in the two chromosomes, see the
yellow part of the chromosomes in Fig. 3.4(a). If the selected crossover point is in between
two positions represented by the yellow genes as is the case for the two parents of Fig. 3.4(a),
then the produced offsprings violate constraint (ii) since one potential location (represented
by the yellow genes) appears twice in one single chromosome, see Child 1 in Fig. 3.4(a).
However, abandoning the infeasible offspring and repeating the crossover operation until
feasible offsprings are obtained greatly reduces the diversity. Hence, instead, we propose to
do crossover on the remaining part of chromosomes that excludes the common wells, and
then replace the common wells back in their original place, see Fig. 3.4(b). The crossover
process is repeated until Np offsprings are generated.

50
Figure 3.3: Illustration of one-point crossover. The red and blue colors represent x of two
parent chromosomes respectively.

(a) One-point crossover (b) Modified one-point crossover

Figure 3.4: Illustration of the original one-point crossover and the modified one-point
crossover. The red and blue colors represent x of the two parent chromosomes respectively.
The yellow color corresponds to a specific gene unit xbj , j = 1, 2, · · · , Nw .

Adding more crossover sites (i.e., N-point crossover) allows the head and tail section
of a chromosome to be accepted together in the offspring and can give offsprings of a larger
diversity compared to one-point crossover. However, it can destroy good gene segments
and can sometimes reduce the performance of the genetic algorithm. We started with the
one-point crossover and did not see a loss of diversity in the offsprings. Hence, one-point
crossover are used throughout our work for the binary encoded vectors (x, T ). The classical
one-point (or N-point) crossover operator is no longer suitable for the ordinal vector O since
they may generate offspring chromosomes where two or more wells are drilled at the same
time, hence, we used PMX crossover instead.
The partially mapped crossover (PMX) was proposed by Goldberg and Lingle [29]
for ordinal numbers. Firstly, two random crossover points are chosen along the chromosome
string, then the substrings between these two crossover points are exchanged between the
selected pair of parents, and then the remaining information of the offsprings are determined
according to the mapping relationship in the selected substrings. Consider, for example, two
parents (denoted as P1 and P2 ) with one random crossover point between the 3rd and 4th
bits and the other crossover point between the 6th and 7th bits where, with the two crossover

51
points (shown below) marked with |, P1 and P2 are given by:

P1 = [8 4 3 | 5 7 6 | 1 2],
(3.3)
P2 = [4 6 5 | 3 7 8 | 1 2].

The mapping sections that are interchanged are between these two crossover points. In this
example, the mapping systems are 5 → 3, 7 → 7 and 6 → 8. Now these two mapping
sections are exchanged with each other so that the two offsprings have the following form,

C1 = [× × × |3 7 8| × ×],
(3.4)
C2 = [× × × |5 7 6| × ×],

where the entries of genes marked × are still to be determined. Then, determine the re-
maining bits using information from the original parents that produce no conflicts, i.e., do
not generate two equal bits in one chromosome, to obtain

C1 = [× 4 × | 3 7 8 | 1 2],
(3.5)
C2 = [4 × × | 5 7 6 | 1 2].

If it did not cause a conflict, the first × in the first offspring C1 would be 8 which comes
from the first parent P1 . However, 8 is already in C1 so it cannot be used again. According
to the mapping sections, 8 → 6, and the fact that 6 is absent from C1 in Eq. 3.5, 6 should
then occupy the first × in C1 . The second × in C1 is 3 according to Eq. 3.3, but 3 is already
in this offspring, C1 of Eq. 3.5. Thus, according to mapping 3 → 5, 5 should replace the
second × since 5 has no conflict with the existing elements. Thus the first offspring C1 is
given by
C1 = [6 4 5 | 3 7 8 | 1 2]. (3.6)

Analogously, the second offspring C2 is given by

C2 = [4 8 3 | 5 7 6 | 1 2]. (3.7)

52
 4. Mutation
All Np offsprings are mutated before they are evaluated. Again, we used different
mutation operators for the binary encoded variables (x, T ) and the permutation encoded
variables O.
For each gene of (x, T ), a random number ri is generated from the uniform distri-
bution U[0, 1] and compared with the mutation rate pm , which represents the probability
of mutation operation. If ri < pm , the corresponding gene will be flipped, i.e., we change
0 to 1 or vice versa. If either constraint (i) or (ii) is violated after the mutation operation,
the chromosome will be repaired by replacing the infeasible chromosome, x, with a feasible
chromosome following three steps: (i) find the potential paths that do not appear in x and
their corresponding binary representations; (ii) calculate the Euclidean distance between xbj
and each binary representation; (iii) replace xbj (the well that cause a conflict) with one of
these binary representations which has the shortest Euclidean distance to xbj . Note that
mutation of a type variables gene Tj will not cause infeasibility.
Since each element of the drilling order O is related to each other, i.e., O is a sequence,
we cannot perform the classical gene by gene mutation as we do for (x, T ). Instead, we do
mutations, chromosome by chromosome, with a larger mutation probability, p̂m , than is
used when mutating genes of binary variables (x, T ). In particular, for each individual in
the current population, we generate one random number ri ∼ U(0, 1), i = 1, 2, · · · , Np . If
ri < p̂m , then mutate the drilling order O of ith individual, otherwise, leave O of the ith
chromosome unchanged. Three mutation operations are used with equal probability. We use
O in Eq. 3.2 as an example to demonstrate the procedure used when ri < pm .

• Two-point swap. Randomly select two points in the O segment of a chromosome, and
swap the ordinal numbers at these two points. For example, select the 3rd and 6th
locations in Eq. 3.2, then the mutated drilling order is obtained as

[6 4 |8| 3 7 |5| 1 2]. (3.8)

53
 • Flipping of genes. Randomly select two points out of the O segment of a chromosome,
and flip the ordinal numbers in between these two points, i.e., reverse the order in
which they appear. For example, select the 3rd and 6th locations, then the mutated
drilling order of Eq. 3.2 is obtained as

[6 4 |8 7 3 5| 1 2]. (3.9)

• Re-combine genes. Randomly select one point out of the O segment of a chromosome,
split it into two segments and re-combine the two segments by putting the first segment
of genes after the second segment of genes. For example, select one point between the
6th and 7th locations and then the mutated drilling order of Eq. 3.2 is obtained as

[ 1 2 |6 4 5 3 7 8]. (3.10)

5. Replacement
In the replacement operation, the best candidate solution from the current generation
(with highest NPV or fitness) will randomly replace one of the newly generated offsprings
so that the best chromosome will be carried over to the next generation.
In this chapter, besides simultaneous optimization of (x, T , O), we also considered
a two-stage sequential optimization procedure. In the first stage, the locations of Nw wells
from Kmax potential locations, x, and the well types, T , are optimized where all Nw wells
operate from time zero since x and T are always represented as binary vectors. In the second
stage, the drilling order of Nw wells O is optimized based on the optimal (x, T ) obtained
in the first stage. In this way the drilling order O is considered as a second order effort.
The performance of the sequential procedure and the simultaneous procedure are compared
based on two numerical examples.

3.1.2 Optimization of well controls u

As mentioned earlier, StoSAG combined with steepest ascent is used to optimize u

54
based on the optimal well locations, types and drilling orders, (x, T , O), obtained for fixed
values of the well controls. In the implementation of StoSAG, at each iteration, based on
the current estimate of control settings u` , 10 perturbations are generated from a normal
low 2
distribution where the ith diagonal of Cu is 0.03 ∗ (uup


i − ui ) and Cu is selected to
enforce a temporal correlation of the controls of each well; see Chen and Reynolds [9]. The
maximum step size, along the normalized stochastic gradient direction (gradient divided
by its infinity norm) is set as 0.2 ∗ [uup − ulow ]. If the NPV is increased, then update u`
and move to the next iteration. If the NPV is not increased, then the step size is cutback
by 50%. The maximum number of step-size cuts is set as 5. If after 5 step-size cuts, the
NPV cannot be improved, 10 new perturbations around u` are re-sampled to estimate a new
stochastic gradient. The maximum number of re-samples is set as 5, i.e., after 5 re-samples,
the algorithm is terminated.

3.2 Numerical Examples

In this section, the proposed methodology is tested on two waterflooding examples, a

channelized reservoir with three different facies (Example 1) and the Brugge model (Example
2). In Example 1, we select 8 best wells from 30 potential paths (either vertical or horizontal).
In Example 2, we consider three cases: in case A, we try to determine the best 12 well
locations out of 30 potential locations; in case B, we still select 12 best wells, but the
number of potential paths is enlarged to 64. In case C, we consider simultaneously selecting
12 wells starting from time 0 and 8 infill wells starting from the fourth year, out of 64
potential locations. For both examples, we first determine the optimal well locations, types,
drilling order using GA with BHP’s for producers and injectors, respectively, fixed at their
minimum and maximum values, and then the optimal control settings are determined using
the StoSAG algorithm while we fix the well locations, types and drilling order at the optimal
values obtained in the first optimization. We also compare simultaneous optimization of
(x, T , O) and sequential optimization where (x, T ) are optimized first assuming all wells
start operation at 0 days and then drilling order (O) is optimized with (x, T ) fixed at the

55
optimal values obtained, referred to as (x, T ) → O.
In the implementation of GA, the crossover probability pc is set to 0.8 for both (x, T )
and O. However, for the binary encoded parameters (x, T ), the mutation probability p̂m is
set to 0.01. For the ordinal parameters O, pm is defined by

p̂m = min{0.03 ∗ ig , 0.3}, (3.11)

where ig represents the ig th generation. As discussed previously, we do mutations gene by

gene for (x, T ) and chromosome by chromosome for O. The population size Np is set to be
50; the maximum number of generations Ng are set differently for Example 1 and Example 2.
Based on our numerical experiments, we found that the rule of thumb in the GA literature
that Np should be set equal to twice the number of bits in a chromosome gives a larger value
of Np than what is needed. Np = 50 gives a reasonable trade-off between fast convergence
and the good quality of the solution.

3.2.1 Example 1: channelized reservoir

The first example considered is a channelized reservoir model defined on a 50 ×

50 × 14 reservoir simulation grid where each grid block is 200 ft × 200 ft × 10 ft. The
reservoir is composed of three geologic zones: zone 1 is composed of the first four reservoir
simulation layers, zone 2 contains the middle six simulation layers, zone 3 is composed of
the bottom four simulation layers. There are three facies, i.e., the channel facies, the levee
facies and the shale facies generated using the object-based modeling. When generating the
petrophycical parameters, the azimuth angle is set to 0◦ and the correlation length is set
to 1,000 ft in both the major and minor correlation directions. We assume no correlation
in the vertical direction. Both the horizontal permeability field and the porosity field are
generated following normal distribution where the mean and deviation of the petro-physical
parameters for each individual facies are given in Table 3.1. The vertical permeability field is
set equal to 1/10 of the horizontal permeability field. The top of the reservoir is at the depth
of 4,800 ft. The initial reservoir pressure is 3,800 psi and the initial reservoir oil saturation

56
is 0.9. Note the upper bound on the production BHP’s and the lower bound on the injection
BHP’s are set equal to the initial reservoir pressure, i.e., 3,800 psi.
Wells can only be drilled at the 30 possible locations shown in Fig. 3.5 and each
location corresponds to one prescribed well trajectory (a fully penetrating vertical well or
a horizontal well). Each potential well location corresponds to a label. The potential hori-
zontal wells are completed in different layers and the prescribed well trajectories at different
potential locations do not intersect with each other. Note that two potential well locations
can intersect or overlap with each other areally when their paths are projected. For example,
the label “24/28” in Fig. 3.5 means that path 24 and path 28 project to the same line in the
x-y plane but path 24 is a horizontal well located in layer 12 whereas path 28 is a horizontal
well located in layer 14. The square symbol represents the starting point of a well, the circle
symbol represents the end point of a well and the bold line represents the perforation in
the plotted layer. This configuration contains 11 vertical wells fully penetrating all 14 layers
(with labels from 1 to 11) and 19 horizontal wells completed in different layers (with labels
from 12 to 30), see Table 3.2. The perforation lengths of the horizontal wells are equal to
2,000 ft which are much longer than those of the vertical paths which have a perforation
length equal to the reservoir thickness, 140 ft.
The problem is to select Nw = 8 best wells from Kmax = 30 potential locations and
determine their optimal locations x, types T , drilling order O and control settings u. All
wells are under BHP control. We consider the effect of the value of the upper bound on
BHP’s at injectors and the lower bound on BHP’s at producers and the effect of the values
of fixed BHP’s when we optimize (x, T , O) with the value of BHP’s held fixed. Two sets
of bounds on the control settings are considered, where in bound A the upper bound for
injectors is 4,600 psi and the lower bound for producers is 3,000 psi and in bound B the
upper bound for injectors is 5,000 psi and the lower bound for producers is 1,000 psi. A
producer will be shut-in if its water cut goes above 0.98. The production life, which is
set to be 4,000 days, is divided into 16 control steps where the first 7 control steps are all
90 days in length, the subsequent 8 control steps are all 365 days in length and the last

57
control step is 450 days. We assume it takes three months to drill a well, Td = 90 days.
Accordingly, the first well has 16 control steps, the second well has 15 control steps, and so
on. The eighth well drilled has 9 control steps. Hence, there are 100 control variables in
total, i.e., the vector of the continuous control variables, u, is 100-dimensional. In this case,
the drilling cost of a vertical well is Cw,j = $8 × 106 and the drilling cost of each horizontal
well is Cw,j = $16 × 106 . The economic parameters for the computation of NPV are set as:
ron = $50.0/STB, cnw = $5.0/STB, cnwi = $5.0/STB, b=10.0%. Only oil and water flow in
the reservoir.
Given Kmax = 30, a bpl = 5 bit binary vector is used to describe the possibility of each
of the 30 potential locations. Hence, with Nw =8, x is a Nw ∗ bpl = 40-dimensional binary
variable vector, T is a 8-dimensional binary variable vector and O is a 8-dimensional ordinal
vector, so there are 56 categorical optimization variables. Hence, a sequential optimization
of (x, T , O) and u is adopted where (x, T , O) are determined first fixing the BHP’s of
injectors and producers at their upper bound and lower bound respectively and then u
is optimized based on the optimal (x, T , O) obtained. In simultaneous optimization of
(x, T , O) with GA using mixed encoding, the maximum number of generations is set equal
to 80. In sequential optimization of (x, T ) using binary encoding followed by optimization
of O using permutation encoding, the maximum number of generations are set as 50 and 30,
respectively, so that the computational effort of simultaneous and sequential optimization
are comparable in that 80 generations are used in both cases.
In this example, we consider the determination of optimal (x, T , O, u) under bound
A and bound B, where we denote the optimization processes as “Bound A” and “Bound
B” respectively. For both Bound A and Bound B, we set the BHP’s at the bounds when
determining the optimal (x, T , O) and then determine the optimal u using the optimal
(x, T , O) obtained. To investigate the effects of controls used to determine well locations,
types and drilling order, we also considered a case where we set the controls at bound A to
determine the optimal (x, T , O) and then optimize well controls within bound B, referred
to as “Bound A→B”. Note that the bounds in Bound A are closer to the initial reservoir

58
pressure than those of Bound B. Bound A→B is a special case of Bound B where the controls
are not set at the bounds (1,000 psi for production BHP and 5,000 psi for injection BHP)
when determining (x, T , O), but are fixed equal to the initial guesses used for the subsequent
optimization of well controls (3,000 psi for production BHP and 4,600 psi for injection BHP).

Table 3.1: Petrophysical parameters for each facies, Example 1, i.e., selecting Nw = 8 wells
given Kmax = 30 potential locations in a channelized reservoir.

Porosity Horizontal Permeability, mD

Facies
Mean Variance Mean Variance
Channel 0.25 0.03 1500 200
Levee 0.20 0.03 500 100
Shale 0.15 0.03 50 10

(a) Layer 1 (b) Layer 7 (c) Layer 12

Figure 3.5: The full set of potential well locations and trajectories projected to the horizontal
permeability field, Example 1.

Table 3.2: Perforated layers of the horizontal wells, Example 1.

Path Label 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Completed Layer 1 1 1 1 1 3 3 3 3 3 3 12 12 12 12 14 14 14 14

1. Optimal solution of (x, T , O)

In this section, we use Bound A as the base case to illustrate the performance of the
proposed methodology. In Bound A, the BHP controls at all producers are fixed at their
lower bound (3,000 psi), the BHP controls at all injectors are fixed at their upper bound
(4,600 psi) when optimizing (x, T , O). Fig. 3.6 shows the NPV versus number of generations

59
Table 3.3: Fixed BHP’s used in the optimization of (x, T , O) and initial guesses used in the
subsequent control optimization, Bound A, Bound A→B and Bound B for Example 1.

Case Bound A Case Bound A→B Case Bound B

Upper bound on injection BHP, psi 4600 5000 5000
Lower bound on production BHP, psi 3000 1000 1000
Injection BHP for (x, T , O) optimization, psi 4600 4600 5000
Production BHP for (x, T , O) optimization, psi 3000 3000 1000
Initial injection BHP for control optimization, psi 4500 4600 4800
Initial production BHP for control optimization, psi 3100 3000 1200

for the joint optimization of only well locations and types (no optimization of O) obtained
with three different initial random seeds. The NPV’s for the candidate solutions in the initial
generation range widely, approximately from $0 × 109 to $4 × 109 , i.e., the well locations
and types significantly impact the NPV. The best NPV’s obtained with seed 2 and seed 3
are somewhat higher than the best NPV obtained with the run using seed 1, but seed 1
gives only a 4% lower optimal NPV than the highest optimal NPV so the method is fairly
robust to the choice of the initial population. The best NPV among the Np = 50 candidate
solutions increases significantly during the first 20 generations but the rate of improvement
slows down dramatically during the following 30 generations and is essentially stable for the
last few generations for all three seeds. Table 3.4 shows the best candidate solution obtained
up to 10, 20, 30, 40 and 50 generations using seed 2. In Table 3.4 and similar tables presented
later, “P” corresponds to a producer and “I” corresponds to an injector, e.g., “I04” represent
an injector at the well path represented by label 4 in Fig. 3.5; Ninj , Npro , NV , NH respectively
represent the optimal number of injectors, producers, vertical wells and horizontal wells. As
we can see, the optimal well locations and types change greatly in the first 20 generations
(specifically, generations 10 and 20 have only one well in common), but generations 40 and
50 have identical optimal solutions.
Based on the optimal well locations and types, a second stage of drilling order op-
timization is carried out and the resulting NPV versus the number of generations is shown
in Fig. 3.7. Compared to the optimal NPV in Fig. 3.6, the initial NPV’s in Fig. 3.7 are
much smaller. This is because in the first optimization stage where (x, T ) is optimized,

60
all wells are assumed to begin operating at time 0 while in the second optimization stage,
wells are brought online sequentially (every Td = 90 days) which corresponds to a shorter
total operation time and hence gives smaller NPV’s. The NPV’s for the candidate solutions
in the initial generation in Fig. 3.7 have a relatively small range, about $0.2 × 109 which
indicates the well drilling order has a much smaller impact on the NPV compared to the
well locations and types. The best NPV obtained with each seed increases only slightly (on
average 0.7%) by optimizing the drilling order and stabilized at around 30 generations for all
three seeds. More precisely, if we had chosen the “optimal” drilling order based on the best
NPV of the initial population of Fig. 3.7(b), we would have obtained an “optimal” NPV
equal to $4.67 × 109 whereas as drilling order optimization increased this to $4.72 × 109 , an
absolute increase of 50 million dollars at the cost of 1500 reservoir simulation runs. How-
ever, if we simply ignored any evaluation of NPV for a set of possible drilling order, i.e.,
did not even generate the initial population of Fig. 3.7(b), we could have picked a drilling
order that resulted in an NPV of $4.44 × 109 which is about 230 million lower than the best
NPV from the initial population. Table 3.5 shows the optimal solutions including the well
locations, types and drilling order for all three seeds. Each well is represented in the form
“type-label-order,” e.g., “P05-6” represents drilling the producer with potential path labeled
5 starting from (6 − 1) ∗ Td = 300 days. “V” represent a vertical well and “H” represents
a horizontal well. If a well is horizontal, the completed layer of that well is also given, e.g.
“H-1” represents a horizontal well completed in layer 1. As we can see from Table 3.5, the
number of injectors is 3 or 4 for all 3 seeds. The variation in the optimal NPV is also small.
Comparing the optimal solutions obtained with seed 1 and seed 2, they have 6 similar wells,
including four wells in common {P05, P16, P23, I29}, one pair of wells at same areal location
but completed in different layers {I21 and I15}, and one pair of wells close to each other {I13
and I18}. Comparing the optimal solutions obtained with seed 2 and seed 3, they have 6
wells in common {I04, P05, I15, P16, P20, P23} and one pair of wells close to each other {I13
and I18}. For all three seeds, the common wells have the same relative drilling order and the
first four wells drilled follow the sequence “producer-injector-producer-injector”. Fig. 3.8

61
shows the optimal well placement obtained with three initial random seeds plotted on the
horizontal permeability field and the remaining oil saturation field of layer 7 after 4,000 days
of production. Note that the distribution of the remaining oil saturation is fairly similar for
the three estimates of optimal solutions. Hence, the optimization gives us a clear idea about
good locations, their types, drilling order, the optimal NPV and the sweep efficiency.

Best Chrom: $4.7506 # 109 Best Chrom: $4.9607 # 109 Best Chrom: $4.9032 # 109

4 4 4
NPV, 109$

NPV, 109$

NPV, 109$
2 2 2

0 0 0
0 20 40 0 20 40 0 20 40
No. of Generations No. of Generations No. of Generations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.6: NPV versus generations obtained with joint optimization of well locations and
types using three seeds, Bound A, Example 1.

Table 3.4: Optimal solutions obtained at different number of generations for the joint opti-
mization of well locations and types using seed 2, Bound A, Example 1.

Generations NPV, $ × 109 Ninj Npro NV NH Wells

10 4.49 4 4 3 5 I04 I08 P10 I13 I14 P16 P23 P28
20 4.68 4 4 2 6 I04 P09 P12 I13 I15 P16 P23 I29
30 4.75 4 4 2 6 P01 I04 I13 I15 P16 P20 P23 I29
40 4.96 4 4 2 6 I04 P05 I13 I15 P16 P20 P23 I29
50 4.96 4 4 2 6 I04 P05 I13 I15 P16 P20 P23 I29

Best Chrom: $4.5476 # 109 Best Chrom: $4.7155 # 109 Best Chrom: $4.6642 # 109
4.8 4.8 4.8
NPV, 109$

NPV, 109$

NPV, 109$

4.6 4.6 4.6

4.4 4.4 4.4

4.2 4.2 4.2

0 10 20 30 0 10 20 30 0 10 20 30
No. of Generations No. of Generations No. of Generations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.7: NPV versus generations obtained with the optimization of drilling order using
three seeds, Bound A, Example 1.

62
 (a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.8: Optimal solutions obtained with sequential optimization of (x, T ) → O, using
three seeds, plotted on the horizontal permeability field (upper row) and the oil saturation
field (bottom row) of layer 7 at end of 4,000 days of production, Bound A, Example 1.

(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.9: Optimal solutions obtained with sequential optimization of (x, T ) → O, using
three seeds, plotted on the horizontal permeability field (upper row) and the oil saturation
field (bottom row) of layer 7 at end of 4,000 days of production, Bound B, Example 1.

63
 Fig. 3.9 presents the optimal wells obtained with sequential optimization using three
seeds for Bound B, plotted on the horizontal permeability field (upper row) and the oil
saturation field (bottom row) of layer 7 at end of 4,000 days of production. In this case,
the BHP’s are fixed at their bounds (5,000 psi for injectors, 1,000 psi for the producers)
at all control steps when optimizing (x, T , O). From Fig. 3.9, with seed 1 and seed 3, we
obtain similar flooding patterns (with 6 wells in common) with two injection wells where
the injected water at I24 flows from the southwest corner to the northeast direction along
the channels and the producers are gathered at the north and east sides of the reservoir.
On the other hand, for seed 2, there is only one injector (i.e., I13) and the injected water
flows from the northeast to the southwest with most producers at the south and west sides
of the reservoir. Despite the difference in flow paths and well paths obtained for seed 1
and 3 compared to those generated with seed 2, the estimated optimal NPV’s are almost
identical for the 3 seeds as shown in Table 3.6. Table 3.6 shows a summary of the optimal
solutions obtained with sequential optimization of (x, T ) → O for Bound B. Compared to
Bound A where 3 to 4 injectors are obtained, only one (horizontal) to two injectors (one
horizontal, one vertical) are obtained in case Bound B, due to the higher injection BHP’s.
Meanwhile, for all three seeds of Bound B, (i) the first injector drilled is horizontal and is
brought online Td = 90 days later than that in Bound A; (ii) the first four wells drilled follow
the sequence “producer-producer-injector-producer”. Hence, the optimal results of Bound B
also illustrate that the proposed optimization procedure can provide a clear guidance on the
optimal well combinations and drilling sequence, indicating GA algorithm is quite robust.
Based on the optimal well locations, types and drilling order obtained with sequential
optimization of (x, T ) → O, control optimization is carried out using StoSAG. For both
Bound A and Bound B, the initial guesses of BHP’s for the control optimization are set close
to the bounds and are given in Table 3.3. For Bound A→B, the initial guesses of BHP’s
are set at the upper bound on injection BHP’s (4,600 psi) and lower bound on production
BHP’s (3,000 psi) of bound A, which are far from the bounds of case Bound B (upper bound
on injection BHP’s at 5,000 psi and lower bound on production BHP’s at 1,000 psi).

64
Table 3.5: Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Bound A, Example 1. For seed 1 and seed 3 results, wells that are different from
the solution of seed 2 are marked in red.
Seeds NPV, $ × 109 Ninj Npro NV NH Wells
P05-6 P09-8 P12-7 P16-3 I18-4 I21-2 P23-1 I29-5
Seed 1 4.55 3 5 2 6
V V H-1 H-1 H-3 H-3 H-12 H-14
I04-4 P05-8 I13-5 I15-2 P16-3 P20-7 P23-1 I29-6
Seed 2 4.72 4 4 2 6
V V H-1 H-1 H-1 H-3 H-12 H-14
I04-5 P05-6 I15-2 P16-3 P17-7 I18-4 P20-8 P23-1
Seed 3 4.66 3 5 2 6
V V H-1 H-1 H-3 H-3 H-3 H-12

Table 3.6: Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Bound B, Example 1. Injectors are in bold.
Seeds NPV, $ × 109 Ninj Npro NV NH Wells
P01-8 I02-5 P03-4 P12-1 P13-7 P16-2 P20-6 I24-3
Seed 1 6.39 2 6 3 5
V V V H-1 H-1 H-1 H-3 H-12
P05-5 P08-6 P12-1 I13-3 P14-8 P15-4 P16-2 P19-7
Seed 2 6.36 1 7 2 6
V V H-1 H-1 H-1 H-1 H-1 H-3
P01-6 I02-5 P03-8 P12-7 P16-2 P17-1 P18-4 I24-3
Seed 3 6.40 2 6 3 5
V V V H-1 H-1 H-3 H-3 H-12

2. Investigation of the effect of initial guesses and bounds

The estimated NPV’s obtained for case Bound A, Bound A→B as well as Bound
B obtained for the optimization of (x, T ) → O followed by well control optimization, are
shown in Table 3.7. Figs. 3.10 and 3.12 show the optimal controls obtained for Bound A
using seed 1 and for Bound B using seed 2. Note in both figures, wells tend to operate at
or very close to the bounds for the first 2,000 days which contributes more than 80% of the
NPV. Fig. 3.11 shows the optimal controls obtained for Bound A→B using seed 2. Note
unlike cases Bound A and Bound B where the well control optimization did not increase
NPV significantly beyond the NPV obtained by optimization of (x, T , O) with fixed BHP
controls, in case Bound A→B, optimizing well controls resulted in about a 30% increase
in NPV. However, the final optimal NPV results for Bound A→B are on the order of 5%
lower than those obtained for the Bound B case where during (x, T , O) optimization well
controls were fixed at bounds; see Table 3.3. It is important to note the bounds on BHP’s are
the same for case Bound B and case Bound A→B, the only differences are the fixed values
of BHP’s used when optimizing (x, T , O) and the initial guesses used in the subsequent

65
optimization step where we optimize only controls. Specifically, if we put the controls at the
bounds of Bound B (i.e., injecting at 5,000 psi and producing at 1,000 psi) for the optimal
(x, T , O) obtained in Bound A→B (i.e., injecting at 4,600 psi and producing at 3,000 psi),
we obtain an average NPV which is 21% lower than that in Bound A→B, and 26% lower
than that obtained in Bound B, indicating that the well controls used when optimizing
(x, T , O) strongly affect the optimal well locations, type, drilling order (see Tables 3.5 and
3.6) which in turn affect the optimal well controls (see Figs. 3.11 and 3.12). The results
presented to this point and other results not presented, indicate that with our workflow when
the optimization of (x, T , O) and the optimization of well controls are done in two separate
steps, the fixed values of BHP used in the optimization of (x, T , O) may strongly influence
the result and lead to at least a somewhat suboptimal final estimate of the optimal NPV.

Table 3.7: Effects of initial guesses and bounds on sequential optimization of well locations
and types, drilling order (i.e., (x, T ) → O) and controls using three seeds, Example 1.

Optimal NPV, $ × 109

Seeds Bound A Bound A→B Bound B
(x, T ) → O (u) (x, T ) → O (u) (x, T ) → O (u)
Seed 1 4.55 4.62 4.55 5.96 6.39 6.39
Seed 2 4.72 4.72 4.72 6.09 6.36 6.43
Seed 3 4.66 4.70 4.66 6.07 6.40 6.43
Average 4.64 4.68 4.64 6.04 6.38 6.42

3800 4600
P23
I21
3600 4400
Producer index

P16
Injector index

3400
P05 I18 4200

P12 3200
4000
I29
P09
3000
3800
1000 2000 3000 4000 1000 2000 3000 4000
Simulation Time Simulation Time
(a) Production Control (b) Injection Control

Figure 3.10: Optimal controls for the well locations, types and drilling order obtained with
sequential optimization of (x, T ) → O using seed 1, Bound A, Example 1.

66
 5000

P23 3500 I15

4800

Producer index
3000

Injector index
4600
P16 I04
2500
4400

P20 2000 I13 4200

1500 4000
P05 I29
1000 3800
1000 2000 3000 4000 1000 2000 3000 4000
Simulation Time Simulation Time
(a) Production Control (b) Injection Control

Figure 3.11: Optimal controls for the well locations, types and drilling order obtained with
sequential optimization of (x, T ) → O using seed 2, Bound A→B, Example 1.

5000
P12
3500
4800
P16
Producer index

3000 Injector index

4600
P15
2500
P05 I13 4400

P08 2000 4200

P19 1500 4000

P14
1000 3800
1000 2000 3000 4000 1000 2000 3000 4000
Simulation Time Simulation Time
(a) Production Control (b) Injection Control

Figure 3.12: Optimal controls for the well locations, types and drilling order obtained with
sequential optimization of (x, T ) → O using seed 2, Bound B, Example 1.

3. Investigation of simultaneous (x, T , O) versus sequential (x, T ) → O) optimiza-

tion
Simultaneous optimization of the well locations, types and drilling order using GA
with mixed encodings with three initial random seeds is also carried out and the optimal
NPV’s obtained for the three cases are presented in Table 3.8. Comparing Tables 3.7 and 3.8,
sequential optimization slightly outperformed simultaneous optimization with same compu-
tation cost based on the average optimal NPV’s obtained for three different runs. However,
it is impossible to argue this is a general conclusion and based purely on mathematical intu-
ition, one would guess that simultaneous optimization would generally perform as well as or
better than sequential optimization as simultaneous optimization ensures the best possible

67
solution exists somewhere in the search space. Isebor et al. [34] compare simultaneous versus
sequential optimization for only one simple problem where the reservoir model is 2D, and
the number of wells is fixed at five with three producers and two injectors. Moreover, the
only optimization variables are the well controls based on five control steps and the loca-
tions of the five vertical wells, i.e., unlike our work, well types and drilling order are not
considered. The problem they consider involves only 35 total design variables. In addition,
they only compare sequential and simultaneous optimization using PSO-MADS. For their
case, based on five distinct optimizations, simultaneous optimization gives an NPV about
4% to 8% higher than sequential optimization which is in conflict with our conclusion on this
issue. On the other hand, Humphries and Haynes [32] compare sequential and simultaneous
optimization of locations and controls for a set of problems using PSO-MADS. They find
that as the problem become more complex, e.g., as allowed well paths become more “com-
plex,” i.e., slanted well trajectories are allowed as opposed to requiring all wells be vertical,
sequential optimization results in 5% to 9% higher NPV than simultaneous optimization.
Moreover, as we consider well locations, types and drilling order while [34] and [32] consider
only well locations and controls, when comparing sequential versus simultaneous optimiza-
tion, there is no way to know a priori whether the conclusions of either paper on sequential
versus simultaneous optimization would apply to the problem we consider here. Given the
expected highly rough optimization problem where two types of categorical design variables
are used and the large search space, we are not surprised to find that, for the problems we
consider, sequential optimization performs slightly better than simultaneous optimization
which agrees with the conclusion of [32].

3.2.2 Example 2: Brugge model

The Brugge field was developed by TNO in conjunction with the SPE Applied Tech-
nology Workshop held in Brugge in June 2008 to benchmark the technology available at that
time for closed loop (real-time) reservoir management, see Peters et al. [71]. The original
model was constructed with approximately 20 million gridblocks and then upscaled to a

68
Table 3.8: Effects of initial guesses and bounds on simultaneous optimization of well locations
and types, drilling order (i.e., (x, T , O)) and controls using three seeds, Example 1.

Optimal NPV, $ × 109

Seeds Bound A Bound A→B Bound B
(x, T , O) (u) (x, T , O) (u) (x, T , O) (u)
Seed 1 4.61 4.64 4.61 5.93 6.16 6.36
Seed 2 4.53 4.54 4.53 6.12 6.33 6.39
Seed 3 4.65 4.70 4.65 5.95 6.38 6.43
Average 4.60 4.63 4.60 6.00 6.29 6.39

60,048 (139 × 48 × 9) gridblock model. The stratigraphy of the Brugge field is modeled
after a typical North Sea Brent field and is an elongated half-dome with one internal fault
and one boundary fault. Originally, 20 vertical producers are placed inside the oil zone while
10 vertical injectors are placed close to the water oil contact. The top structure map with the
original well locations is shown in Fig. 3.13. Fig. 3.14 shows the log horizontal permeability
field and the porosity field of layer 1 and layer 9 respectively.

Figure 3.13: Top structure of the Brugge model with the original 30 wells. Left: 3D structure
plotted on the initial oil saturation field. Right: 2D depth contour of layer 1 where wells
with label 21 to 30 are the locations of the original injectors and wells with label 1 to 20 are
the locations of the original producers, Example 2.

In this example, we considered three cases. In case A, we use the original 30 well
locations as 30 potential locations and select 12 wells out of these 30 potential locations.
In case B, we enlarge the number of potential locations to 64 but still select 12 wells. For
both cases, we determined the optimal locations, types, drilling order and control settings
for 12 wells. In case C, we considered simultaneously selecting 12 wells starting from time

69
 0.3 0.3

8 8
0.25 0.25

6 6
0.2 0.2

4 4
0.15 0.15

2 2
0.1 0.1

0 0 0.05 0.05

-2 -2 0 0

(a) ln kx -1st layer (b) ln kx -9th layer (c) φ-1st layer (d) φ-9th layer

Figure 3.14: The log-horizontal permeability (ln kx ) field and porosity (φ) field of the Brugge
model, Example 2.

0 and 8 infill wells starting from the fourth year, from 64 potential locations. The initial
average reservoir pressure is 2,514 psi. All wells operate under BHP control. The upper
bound on the production BHP’s and the lower bound on the injection BHP’s are set the
same as the initial reservoir pressure, i.e., 2,514 psi. The lower bound on the production
BHP’s is set equal to 725 psi and the upper bound on injection BHP’s is set equal to 2,662
psi. In this example, producers only penetrate the first 8 layers and injectors penetrate all 9
layers. Either producers or injectors can be drilled at any potential location in the oil zone.
We force a well drilled at potential locations in the aquifer to be an injector. A producer will
be shut-in if its water cut goes above 0.94. The reservoir life is set as 20 years. The economic
parameters are set as: ron = $80.0/STB, cnw = $5.0/STB, cnwi = $5.0/STB, b=10.0%. The
drilling cost for each well is set as Cw,j = $8 × 106 . We assume it takes one month to drill a
well, i.e., Td = 30 days.
1. Case A of Example 2: Nw = 12, Kmax = 30
In this case, the locations of the 30 wells used in the original Brugge model are used
as the 30 potential locations. Given Kmax = 30 and Nw = 12, a bpl = 5 bit binary vector
is used to describe the possibility of each potential location. Hence, x is a Nw ∗ bpl = 60-

70
dimensional binary variable vector, T is a 12-dimensional binary variable vector and O is a
12-dimensional ordinal vector. The reservoir life is divided into 20 control steps where each
control step contains 365 days. Since it takes 30 days to drill a well, all wells are drilled
within the first year, i.e., all wells have 20 control steps. However, the first control step is
365 days for the first well (from days 0 to 365 days), 335 days for the second well (from 30
days to 365 days), 305 days for the third well (from 60 days to 365 days), etc. Hence, u is
a Nw ∗ Nc = 240-dimensional vector composed of continuous variables.
A sequential optimization of (x, T , O) and u is adopted where (x, T , O) is deter-
mined first with the BHP’s of injectors and producers fixed at their upper bound and lower
bound respectively and then u is optimized based on the optimal (x, T , O) obtained. In
simultaneous optimization of (x, T , O) with GA using mixed encoding, the maximum num-
ber of generations are set as 60. In sequential optimization of (x, T ) using binary encoding
and O using permutation encoding, the maximum number of generations are set as 40 and
20 respectively.
Fig. 3.15 shows the NPV’s versus number of generations for the joint optimization of
well locations and types with three seeds. In this case, the NPV’s for the candidate solutions
in the initial generation range approximately from $0 × 109 to $7.4 × 109 , again indicating
that well locations and types have a big impact on the NPV. The highest optimal NPV is
obtained with seed 3, but the worst optimal NPV estimate is only 0.5% worse than the best
optimal NPV of $9.34 × 109 which indicates that GA algorithm is quite robust. The average
best NPV in the initial population is $7.32 × 109 and the average best NPV in the final
solution is $9.34 × 109 , i.e., GA algorithm improved the NPV by 21.6% on average. Table
3.9 shows the best solutions obtained at different generations using seed 3. From Table 3.9,
the optimal well locations and types change substantially during the first 20 generations
while in the following 20 generations, only two wells changed locations and types. The
difference between the maximum and the minimum of the optimal NPV’s for all three seeds
(0.025×109 $) is quite small compared to their absolute average value (9.32×109 $). Based
on the optimal well locations and types, a second-stage drilling order optimization is carried

71
out and NPV versus the number of generations is shown in Fig. 3.16. Similar to Example
1, the optimal NPV’s obtained are lower than that in Fig. 3.15 due to a shorter operational
time when drilling order is considered. The NPV’s for the candidate solutions in the initial
generation have a relatively small range, about $0.2 × 109 . If we had chosen the “optimal”
drilling order based on the best NPV of the initial population of Fig. 3.16(c), we would
have obtained an “optimal” NPV equal to $9.04 × 109 whereas as drilling order optimization
increased this to $9.08 × 109 , an absolute increase of 40 million dollars at the cost of 1000
reservoir simulation runs. However, if we simply ignored any evaluation of NPV for a set of
possible drilling order, i.e., did not even generate the initial population of Fig. 3.16(c), we
could have picked a drilling order that resulted in an NPV of $8.89 × 109 which is about 150
million lower than the best NPV from the initial population.
Table 3.10 shows the optimal solutions obtained with sequential optimization of
(x, T ) → O. From Table 3.10, all three seeds resulted in 4 injectors and 8 producers.
Comparing the optimal well configurations obtained with seed 2 and seed 3, they have 9 out
of 12 wells in common. Comparing the optimal solutions obtained with seed 1 and seed 3,
they again have 9 out of 12 wells in common. For all three seeds, the first injector drilled is
I01 and is drilled as the 6th or 7th well; the first producer drilled is P19. Fig. 3.17 shows
the optimal well configurations obtained with all three seeds and the corresponding oil sat-
uration fields of layer 1 after 20 years of production. From Fig. 3.17, the well locations and
the oil saturation field obtained for the three seeds are quite similar to each other. Note that
all three seeds obtained four injectors where two injectors are placed in the aquifer and two
injectors are placed in the oil zone. Among the two injectors placed in the oil zone, all three
seeds chose I01 as the first injector. Hence, as in Example 1 (the channelized reservoir), the
optimization gives us a clear indication about good potential locations, their corresponding
types and when to drill the injectors.
Based on the optimal well locations, types and drilling order obtained with sequential
optimization of (x, T ) → O, we carried out control optimization for the selected 12 best
wells using StoSAG starting from the initial guesses where at all control steps, the BHP’s

72
 Best Chrom: $9.2928 # 109 Best Chrom: $9.3193 # 109 Best Chrom: $9.3434 # 109
10 10 10
NPV, 109$

NPV, 109$

NPV, 109$
5 5 5

0 0 0
0 20 40 0 20 40 0 20 40
No. of Generations No. of Generations No. of Generations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.15: NPV versus generations obtained with the optimization of well locations and
types, i.e., (x, T ), using three seeds, Example 2-A, i.e., selecting Nw = 12 wells given
Kmax = 30 potential locations.

Best Chrom: $9.0329 # 109 Best Chrom: $9.064 # 109 Best Chrom: $9.0812 # 109

9 9 9
NPV, 109$

NPV, 109$

NPV, 109$
8.9 8.9 8.9

8.8 8.8 8.8

8.7 8.7 8.7

0 10 20 0 10 20 0 10 20
No. of Generations No. of Generations No. of Generations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.16: NPV versus generations obtained with drilling order optimization using three
seeds, Example 2-A.

Table 3.9: Optimal solutions obtained at different number of generations for the optimization
of well locations and types using seed 3, Example 2-A.
Generations NPV, $ × 109 Ninj Npro Wells
10 8.58 3 9 P01 P02 P07 P08 P11 P12 P13 I16 P17 P19 I21 I30
20 9.02 3 9 I01 P02 P03 P07 P08 P11 P13 P16 P17 P19 I21 I25
30 9.04 3 9 I01 P02 P03 P07 P08 P11 P13 P16 P17 P19 I21 I26
40 9.34 4 8 I01 P02 P03 P07 P08 I09 P11 P13 P16 P19 I21 I25

Table 3.10: Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Example 2-A. Wells that are different from the solution of seed 3 are in red.
Seeds NPV, $ × 109 Ninj Npro Wells
Seed 1 9.03 4 8 I01-6 P02-4 P03-3 P06-5 I07-12 P08-11 P11-8 P13-9 P16-2 P19-1 I21-7 I27-10
Seed 2 9.06 4 8 I01-6 P02-5 P03-2 P06-5 I07-12 P08-10 P11-8 P13-11 P16-3 P19-1 I21-7 I30-9
Seed 3 9.08 4 8 I01-7 P02-3 P03-2 P07-6 P08-12 I09-9 P11-5 P13-10 P16-4 P19-1 I21-11 I25-8

at producers are 870 psi (150 psi more than the fixed production BHP during the (x, T , O)
optimization) and the BHP’s at injectors are 2,640 psi (22 psi less than the fixed injection

73
 0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7

+ + + +
+ 0.6 0.6 0.6 0.6
+ o o
o o o o o o o
+ o 0.5 0.5 0.5 0.5
+ o + + + o
o o o o o
+ o o 0.4 o o 0.4 o o 0.4 o o 0.4
o o
+ o o + + + +

+ o
0.3 o 0.3 o 0.3 o 0.3
o +
oo o
o o o
+ o o o o
0.2 0.2 0.2 0.2
+ +

0.1 0.1 0.1 0.1

0 0 0 0

(a) So -initial (b) Soopt -seed 1 (c) Soopt -seed 2 (d) Soopt -seed 3

Figure 3.17: Well configurations and oil saturation fields of layer 1. (a) shows the set of
initial potential locations plotted on the oil saturation field; (b)-(d) show the optimal well
configurations obtained with sequential optimization of (x, T ) → O for three seeds, with the
corresponding oil saturation fields after 20 years of production on the top layer, Example
2-A. ⊗ represents an injector and o represent a producer.

BHP during the (x, T , O) optimization). Note different wells have different starting times
since they are drilled one by one. However, well control optimization did not increase the
value of NPV obtained by optimizing only (x, T , O) with well controls fixed at bounds, see
Table 3.11. We did observe that, the control settings of the injectors increased to their upper
bounds and those of the producers decreased to their lower bounds in a few iterations, but
caused an increase in NPV smaller than $0.01 × 109 (10 million dollar), which suggests that
the upper bound on the injection pressure and the lower bound of the producing pressure are
the optimal control settings for this example. The fact that the optimal controls found were
at bounds, for this example, is likely due to the following: we are placing a relatively small
number of wells for the lengths of reservoir life specified, compared with 30 wells specified
in the original Brugge example.
Table 3.11 presents a summary of the optimal NPV’s obtained with both sequential
optimization of (x, T ) → O and simultaneous optimization of (x, T , O). Based on the aver-
age optimal NPV, the sequential process performed as well or better than the simultaneous

74
process, which is consistent with the results of Example 1.
Table 3.11: Comparison of the optimal NPV’s obtained with optimization of (x, T ) → O
and (x, T , O) followed by control optimization, Example 2-A.

Optimal NPV, $ × 109

Seeds Sequential Simultaneous
(x, T ) (O) (u) (x, T , O) (u)
Seed 1 9.29 9.03 9.03 8.96 8.96
Seed 2 9.32 9.06 9.06 9.11 9.11
Seed 3 9.34 9.08 9.08 9.11 9.11
Average 9.32 9.06 9.06 9.06 9.06

We denote the minimum production BHP’s of 725 psi and the maximum injection
BHP’s of 2,662 psi as Bound B which is the case we just considered. Now, we also consider a
set of bounds (referred to as Bound A) where the lower bound on producer BHP’s is 1,619.5
psi and the upper bound on injection BHP’s is 2,588 psi. Note the bounds of Bound A are
midway between the value of Bound B and the initial reservoir pressure. Similar to Example
1, the determination of the optimal (x, T , O, u) using Bound A is denoted as “Bound A.”
The case where we set the controls at Bound A to determine the optimal (x, T , O) and then
optimize well controls within Bound B, is referred to as “Bound A→B”. Table 3.12 shows
the optimal NPV’s obtained for the optimization process of (x, T ) → (O) → (u) for Bound
A, Bound B and Bound A→B.
From Table 3.12, the control optimization improves the NPV’s only slightly for Bound
A, but significantly for Bound A→B, indicating that the bounds on BHP’s affect the optimal
controls and that increasing the upper bound for injection pressure and reducing the lower
bound on production BHP’s during the control optimization can further improve the NPV’s.
Compared to Bound B, slightly less NPV’s are obtained for Bound A→B where the optimal
(x, T , O) are obtained with BHP’s fixed at the middle of Bound B, and controls are optimized
within Bound B. If we obtain a global optimizer for the optimization of (x, T , O) with well
control fixed, it is optimal for the controls used. However, it may be possible to increase the
optimal NPV by well control optimization using the optimal (x, T , O) obtained. We find that
for all the examples considered, that if we fix controls at bounds (upper bound of the injection

75
BHP and lower bound for the production BHP) during (x, T , O) optimization, then the well
control optimization have a negligible effect on the optimal NPV. However, if we fix well
controls at non-bounds during the optimization of (x, T , O), then well control optimization
with the (x, T , O) fixed, tends to move pressure closer the bounds but ultimately produce
a lower final NPV than is obtained by optimization of (x, T , O) with controls fixed at the
bounds. Similar to Example 1, this fact recommends that it is preferable to operate the
injectors at their maximum BHP and producers at their minimum BHP to determine the
optimal (x, T , O). Table 3.13 shows the optimal (x, T , O) obtained for Bound A where wells
that are different from the solution of seed 3 in Table 3.10 are marked in red. Different from
Example 1 where different flooding patterns for Bound A→B and Bound B are obtained,
similar wells are obtained in this case. From Table 3.13, Bound A also gives 4 injectors and 8
producers which is consistent with the result of the base case (Bound B) given in Table 3.10.
What is more, 8 out of 12 wells selected in Table 3.13 are also selected in Table 3.10. Hence,
it is not surprising that similar optimal NPV’s after control optimization are obtained for
Bound A→B and Bound B.

Table 3.12: Comparison of the optimal NPV’s obtained with sequential optimization of
(x, T ) → (O) followed by control optimization, i.e., (x, T ) → (O) → (u), Example 2-A.

Optimal NPV, $ × 109

Seeds Bound A Bound A→B Bound B
(x, T ) → (O) (u) (x, T ) → (O) (u) (x, T ) → (O) (u)
Seed 1 6.29 6.29 6.29 8.73 9.03 9.03
Seed 2 6.40 6.29 6.29 8.96 9.06 9.06
Seed 3 6.29 6.29 6.29 8.87 9.08 9.08
Average 6.33 6.29 6.29 8.85 9.06 9.06

Table 3.13: Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Bound A, Example 2-A. Wells that are different from the solution of seed 3 in
Table 3.10 are marked in red.
Seeds NPV, $ × 109 Ninj Npro Wells
Seed 1 6.29 4 8 I01-5 P02-3 I03-12 P06-10 P08-9 P11-6 P13-11 P16-2 P17-4 P19-1 I21-8 I30-7
Seed 2 6.40 4 8 I01-7 P02-3 P06-8 I07-10 P08-12 P11-4 P13-6 P16-1 P17-11 P19-2 I21-9 I30-5
Seed 3 6.29 4 8 I01-6 P02-4 P03-1 P06-10 I07-9 P08-12 P11-5 P16-3 P18-11 P19-2 I21-8 I30-7

76
 2. Case B of Example 2: Nw = 12, Kmax = 64
In this case, the number of potential locations is increased to 64 by adding 34 addi-
tional potential paths to the 30 locations used in case A, including 4 more locations in the
aquifer near the water-oil contact and 30 more locations in the oil zone, see Fig. 3.20(a).
Given Kmax = 64 and Nw = 12, a bpl = 6 bit binary vector is used to provide a unique
representation of each of the 64 potential well locations. Hence, x is a Nw ∗ bpl = 72-
dimensional binary variable vector, T is a 12-dimensional binary variable vector and O is a
12-dimensional ordinal vector. The reservoir life is divided into 40 control steps where each
control step contains 182.5 days. As it takes 30 days to drill a well, the first 7 wells are drilled
within the 182.5 days. That is to say, the first 7 wells all have 20 control steps, but with a
different length for the first control step which is 182.5 days for the first well (from days 0
to 182.5 days), 152.5 days for the second well (from 30 days to 182.5 days), and so on. The
remaining 5 wells are drilled in the second half year, i.e., they have 39 control steps where
the first control step length is 155 days for the eighth well (from 210 days to 365 days), 125
days for the ninth well (from 240 days to 365 days), and so on. Hence, u is a 475-dimensional
vector composed of continuous variables. A sequential optimization of (x, T , O) and u is
adopted where (x, T , O) are determined by first fixing the BHP’s of injectors and producers
at their upper bound and lower bound, respectively, and then u is optimized based on the
optimal (x, T , O) obtained. In sequential optimization of (x, T ) using binary encoding and
O using permutation encoding, we still use the population size equal to 50, however, the
maximum number of generations are set as 50 and 20 respectively. Note GA is allowed to
run for 10 more generations compared to case A when optimizing (x, T ) since the length of
the chromosome is 12 bits longer. One may also choose to increase the population size Np .
Fig. 3.18 shows the NPV’s versus the number of generations for the joint optimization
of well locations and types with three seeds. Similar to case A, the NPV’s for the candidate
solutions in the initial generation range approximately from $0 × 109 to $8.0 × 109 . The
highest NPV is obtained with seed 1, but the worst optimal NPV estimate is only 2.5% less
than the best optimal NPV of $10.44 × 109 , so again the GA algorithm is quite robust. The

77
average best NPV in the initial population is $7.82 × 109 and the average best NPV in the
final population is $10.31 × 109 . Hence, the GA algorithm improved the NPV by 24% on
average. Compared to case A, the average optimal NPV is 10% larger due to more options
of well locations.
Based on the optimal well locations and types, a second stage of drilling order op-
timization is carried out and NPV versus the number of generations is given in Fig. 3.19.
Similar to case A, the optimal NPV’s are lower than all the initial NPV’s in Fig. 3.18 due to
a shorter operational time when drilling order is considered. The NPV’s for the candidate
solutions in the initial generation have a relatively small range, about $0.2 × 109 , which is
consistent with case A where we select 12 wells out of 30 potential locations. If we had
chosen the “optimal” drilling order based on the best NPV of the initial population of Fig.
3.19(a), we would have obtained an “optimal” NPV equal to $10.12 × 109 whereas as drilling
order optimization increased this to $10.13 × 109 , an absolute increase of 10 million dollars
at the cost of 1000 reservoir simulation runs. However, if we simply ignored any evaluation
of NPV for a set of possible drilling order, i.e., did not even generate the initial population
of Fig. 3.19(a), we could have picked a drilling order that resulted in an NPV of $9.94 × 109
which is about 180 million lower than the best NPV from the initial population.
Table 3.14 shows the optimal solutions obtained with sequential optimization of (x, T )
and O. In this case, the number of possible selections of 12 wells among 64 potential locations
64!
is 12!52!
≈ 6 × 1010 . Hence it is difficult to comment on the quality of the optimal solutions
by comparing the labels of the selected locations. Instead, the locations of optimal wells
obtained with three seeds are shown in Fig. 3.20. From Fig. 3.20, the three runs using
different seeds lead to consistent solutions in that they all lead to (i) 8 producers and 4
injectors, (ii) similar locations of the producers, and (iii) 2 injectors in the oil zone and 2
injectors in the aquifer. For all three seeds, the first injector is drilled as the 6th or 7th well.
Hence, the optimization gives us a clear indication about the good potential locations, their
corresponding types and when to drill the injectors.

78
 Based on the optimal well locations, types and drilling order obtained with well con-
trols fixed at bounds (upper bound of the injection BHP and lower bound of the production
BHP), we carried out control optimization for the selected 12 best wells using StoSAG start-
ing from the initial guesses where at all control steps, the BHP’s at producers are 870 psi
(145 psi higher than the lower bound of production BHP) and the BHP’s at injectors are
2,640 psi (22 psi lower than the upper bound of injection BHP). Similar to case A, NPV is
not improved by control optimization. Table 3.15 presents a summary of the optimal NPV’s
obtained with both sequential optimization ((x, T ) → O) and simultaneous optimization
(x, T , O). As in case A, sequential optimization of (x, T ) → O outperformed simultaneous
optimization of (x, T , O) for all three seeds.

Best Chrom: $10.4417 # 109 Best Chrom: $10.3328 # 109 Best Chrom: $10.1629 # 109
10 10 10
NPV, 109$

NPV, 109$

NPV, 109$
5 5 5

0 0 0
0 20 40 0 20 40 0 20 40
No. of Generations No. of Generations No. of Generations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.18: NPV versus generations obtained with the optimization of well locations and
types using three seeds, Example 2-B, i.e., selecting Nw = 12 wells given Kmax = 64 potential
locations in a channelized reservoir.

Best Chrom: $10.1311 # 109 Best Chrom: $10.0284 # 109 Best Chrom: $9.8465 # 109
10.2 10.2 10.2
NPV, 109$

NPV, 109$

NPV, 109$

10 10 10

9.8 9.8 9.8

9.6 9.6 9.6

0 10 20 0 10 20 0 10 20
No. of Generations No. of Generations No. of Generations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.19: NPV versus generations obtained with drilling order optimization using three
seeds, Example 2-B.

79
 0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7

+
+ o 0.6 0.6 0.6 0.6
o o
+ oo o o + o
+ o o + + o
+ o o o 0.5 0.5 o
0.5 o
0.5
+ o oo o o o o o
o o o o o o
+ o o o 0.4 + o 0.4 o 0.4 o 0.4
o o o o + o +
o +
+ oo o o + o
+ o o o o 0.3
o
0.3 o + o
0.3 + 0.3
+ o o o o o o
o o o
+ oo o o o o o
+o o o 0.2 0.2 0.2 0.2
o
+ + + +

0.1 0.1 0.1 0.1

0 0 0 0

(a) So -initial (b) Soopt -seed 1 (c) Soopt -seed 2 (d) Soopt -seed 3

Figure 3.20: Well configurations and oil saturation fields of layer 1. (a) shows the set of
potential locations plotted on the initial oil saturation field; (b)-(d) show the optimal well
configurations obtained with sequential optimization of (x, T ) → O for three seeds, with the
corresponding oil saturation fields after 20 years of production on the top layer, Example
2-B. ⊗ represents an injector and o represent a producer.

Table 3.14: Optimal solutions obtained with sequential optimization of (x, T ) → O using
three seeds, Example 2-B.
Seeds NPV, $ × 109 Ninj Npro Wells
Seed 1 10.13 4 8 P03-6 P07-12 P16-4 P22-9 P24-2 P27-1 P31-3 P40-5 I42-11 I48-7 I56-10 I63-8
Seed 2 10.03 4 8 P03-1 P07-11 P16-5 P24-2 I25-12 P27-3 P29-4 P35-10 P40-7 I43-8 I53-6 I60-9
Seed 3 9.85 4 8 P01-6 P03-4 P13-10 P16-5 P19-1 I25-9 P27-3 P29-2 P33-12 I48-7 I60-11 I63-8

Table 3.15: Comparison of the optimal NPV’s obtained with optimization of (x, T ) → O
and (x, T , O) followed by control optimization, Example 2-B.

Optimal NPV, $ × 109

Seeds Sequential Simultaneous
(x, T ) (O) (u) (x, T , O) (u)
Seed 1 10.44 10.13 10.13 10.07 10.07
Seed 2 10.33 10.03 10.03 9.42 9.42
Seed 3 10.16 9.85 9.85 9.51 9.51
Average 10.31 10.00 10.00 9.67 9.67

3. Case C of Example 2
Based on the 64 potential locations presented in case B, we carried out a study to
select 12 wells starting from time 0 in addition to selecting 8 infill wells starting from the

80
fourth year, i.e., we select in total Nw = 20 wells. Due to the three year’s gap between
the original wells and the infill wells, it is risky to adopt a sequential procedure where the
well locations and types are determined assuming all 20 wells operate from time 0 and then
optimize drilling order based on the optimal locations and types obtained. Hence, in this
example, we only consider simultaneous optimization of well locations, types and drilling
order. In this case, the drilling order is still parameterized as a permutation of the sequence
[1, 2, · · · , j, j + 1, · · · , Nw ]. However, wells with Oj ≤ 12 are drilled at (Oj − 1)Td days, while
wells with Oj > 12 are drilled at (Oj − 13)Td + Tinf ill days, where Tinf ill is the time at which
we drill the first infill well, i.e., the 13th well. In this example, Tinf ill = 1460 days. Given
Nw = 20 and bpl = 6, x is a 120-dimensional vector, T is a 20-dimensional vector, O is a
20-dimensional vector, which sums up to 160. In this case, we used a larger population size
Np = 80 and run for 70 generations.
Table 3.16 shows a summary of the results obtained with simultaneous optimization
for case C. As one can see, based on the average NPV of three runs, the infill case ob-
tained a 29% increase compared to the initial NPV and 3.4% higher NPV compared to the
simultaneous result of Example 2-B (Table 3.15), before the subsequent control optimiza-
tion. However, simultaneous optimization only obtained a suboptimal solution, i.e., the best
NPV obtained with seed 3 ($10.07 × 109 ) is slightly less than the best NPV of Example 2-B
(Table 3.15) obtained by sequential optimization ($10.13 × 109 ). Similar to previous cases,
the subsequent control optimization only lead to slight increase in the optimal NPV; see
the result of u in Table 3.16. This is because the optimal control settings for the optimal
(x, T , O) happen to be at the bounds. Similar to Example 1, we extended the lower bound of
production pressure from 725 psi to 500 psi and the upper bound of injection pressure from
2,662 psi to 2,800 psi and re-did the well control optimization to obtain the optimal NPV’s
presented in the “û” column of Table 3.16. As one can see, after extending the bounds on
pressure, the NPV is improved by 4.5% on average. The 12 wells scheduled to drill within
the first year include 3 injectors and 9 producers on average which is similar to case B, the
8 infill wells starting at year 4 include 4 injectors and four producers on average. Fig. 3.21

81
shows the optimal well locations plotted on the remaining oil saturation field after 20 years
of production. From Fig. 3.21, generally, we still obtain two injectors in the oil zone where
the two injectors are scheduled to drill within the first year after 120 days (i.e., after the 5th
drilling sequence).

0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7

+ +
+ o 0.6 0.6 0.6 + 0.6
o o
+ oo o o o + o
+ o o
+ 0.5 o 0.5 0.5 0.5
o o o o o o o o o
+ o oo o o o + o o
o o o o o
+ o o o 0.4 + o + 0.4 o + 0.4 o + 0.4
o o o o o o o
o o o +o o
+ oo o o + +
+ o o o o 0.3 + o 0.3 0.3 0.3
+ o o o o + o o
o o o o o
o
o o
+ oo o o o + o + o
+ o o o o + + o
o 0.2 0.2 0.2 + 0.2
+ + + + + +

0.1 0.1 0.1 0.1

0 0 0 0

(a) So -initial (b) Soopt -seed 1 (c) Soopt -seed 2 (d) Soopt -seed 3

Figure 3.21: Well configurations and oil saturation fields of layer 1. (a) shows the set of
potential locations plotted on the initial oil saturation field; (b)-(d) show the optimal well
configurations obtained with simultaneous optimization of (x, T , O) for three seeds, with the
corresponding oil saturation fields after 20 years of production on the top layer, Example 2-
C. ⊗ represents an injector and o represent a producer. The first 12 wells drilled are marked
in black color and the 8 infill wells are marked in blue color.

Table 3.16: Comparison of the optimal NPV’s obtained with simultaneous optimization of
(x, T , O) followed by control optimization, i.e., (x, T , O) → (u), Example 2-C.

Optimal NPV, $ × 109 Number of wells

Seeds first 12 wells 8 infill wells
Initial (x, T , O) u û
NInj NP ro NInj NP ro
Seed 1 7.33 10.03 10.05 10.52 3 9 3 5
Seed 2 7.80 9.88 9.90 10.31 3 9 4 4
Seed 3 8.05 10.07 10.10 10.58 4 8 5 3
Average 7.73 10.00 10.02 10.47 3 9 4 4

82
3.2.3 Investigation on drilling order effect

In both examples (the channelized reservoir and the Brugge model), a sequential
optimization of (x, T ) followed by O was carried out. In the optimization of (x, T ), all wells
are assumed to start operating from time 0 and the various candidate solutions during all
generations have a large range, i.e., approximately $4 × 109 for the channelized reservoir
model and $7 × 109 for the Brugge model. Based on the optimal (x, T ) obtained, the
candidate solutions in the subsequent drilling order optimization only have a small range
which is approximately $280 × 106 (6% of the optimal NPV) for the channelized reservoir
example and $200 × 106 (2% of the optimal NPV) for the Brugge example. Nevertheless,
Leeuwenburgh et al. [44] find that the drilling order has a much larger impact on the
NPV’s than in our cases. In [44], different drilling orders can lead to a change which is
approximately 90% of the optimal NPV in one case and 17% of the optimal NPV in another
case. We believe, this is because that [44] used a very short reservoir life, i.e., 645 days
which is only 120 days after drilling 16 wells at an interval of 35 days, while in our work the
reservoir life is much longer, i.e., 20 years. Hence, the impact of drilling order depends on
the reservoir life, Tres , and the time required to drill a well, Td .
1. Investigation of the effect of Tres on drilling order
In Example 2-B, 12 paths are selected out of 64 potential paths. Based on the optimal
(x, T ) obtained, a secondary drilling order optimization considering different reservoir lives
is carried out. Besides the reservoir life of 20 years, two shorter reservoir lives are considered
to resemble the cases in [44], one is 1.5 years which is 217.5 days after all 12 wells are drilled
and the other one is 2 years which is 390 days after all 12 wells are drilled. Fig. 3.22 shows
the NPV’s versus number of generations during the drilling order optimization for the three
different reservoir lives. Table 3.17 shows the optimal drilling order obtained. The optimal
drilling orders obtained for reservoir lives of 1.5 years and 2.0 years are very similar to each
other while for the reservoir life of 20 years, the first injector is drilled 30 days earlier and
the drilling order of I56 and P03 are almost switched. Table 3.18 shows a summary of the
NPV’s, including the NPV of the worst drilling order encountered, the range of NPV’s in

83
the first generation, the optimal NPV and the optimal NPV minus the worst NPV. The
results of Table 3.18 indicate that the shorter the reservoir life, the larger the ranges of the
NPV’s in the initial generation and the larger the improvements in NPV’s using drilling
order optimization. The results of Table 3.18 partially explain why Leeuweeburg et al. [44]
found drilling order has a larger impact when the reservoir life is short. In industry, the
reservoir life is often a few decades and the time required to drill a well is usually quite short
(e.g. 1 to 3 months).

Best Chrom: $2.7635 # 109 Best Chrom: $3.6313 # 109 Best Chrom: $10.1311 # 109
10.4

2.6 3.6 10.2

NPV, 109$

NPV, 109$

NPV, 109$
2.4 3.4 10

2.2 3.2 9.8

2 3 9.6
0 10 20 30 0 10 20 30 0 10 20
No. of Generations No. of Generations No. of Generations
(a) 1.5 years (b) 2 years (c) 20 years

Figure 3.22: NPV versus generations obtained with drilling order optimization within dif-
ferent reservoir lives using the optimal well locations and types (x, T ) obtained with seed 1,
Example 2-B.

Table 3.17: Optimal drilling order within different reservoir lives using the optimal well
locations and types (x, T ) obtained with seed 1, Example 2-B.
Tres (days) Wells
547.5 P03-10 P07-12 P16-5 P22-4 P24-1 P27-2 P31-3 P40-8 I42-11 I48-9 I56-6 I63-7
730.0 P03-11 P07-12 P16-4 P22-5 P24-1 P27-2 P31-3 P40-8 I42-10 I48-9 I56-6 I63-7
7300 P03-6 P07-12 P16-4 P22-9 P24-2 P27-1 P31-3 P40-5 I42-11 I48-7 I56-10 I63-8

Table 3.18: Summary of NPV’s obtained with drilling order optimization within different
reservoir lives using the optimal well locations and types (x, T ) obtained with seed 1, Ex-
ample 2-B.

Tres NPV, $ × 109

years Initial Worst Initial Best - Initial Worst Optimal Optimal - Initial Worst
1.5 2.21 0.48 2.76 0.55
2.0 3.17 0.42 3.63 0.46
20 9.94 0.19 10.13 0.19

84
 2. Investigation of the effect of Td on drilling order
In Example 2-A, 12 paths are selected out of 30 potential paths. Based on the optimal
(x, T ) obtained, a secondary drilling order optimization considering different drilling time
Td is carried out. Besides Td = 30 days, we also consider Td = 90 days. Fig. 3.23 shows
the NPV’s versus number of generations during the drilling order optimization. Table 3.20
shows the optimal NPV’s and drilling order obtained. Candidate solutions for Td = 90 days
have lower NPV’s (approximately by 5.5%) than those for Td = 30 days. This is because
for Td = 90 days, the total operation time for 12 wells are less than that for Td = 30 days.
For both scenarios, the first injector being drilled is I01. However, for Td = 30 days, I01 is
drilled as the 7th well at 180 days while for Td = 90 days, I01 is drilled as the 5th well at
360 days. Applying the optimal drilling order obtained for Td = 30 days to Td = 90 days
gives a $94 × 106 lower NPV, indicating that both the optimal drilling order and the optimal
NPV’s are dependent on the time required to drill a well, Td . Table 3.19 shows a summary
of the NPV’s, including the NPV of the worst drilling order encountered, the range of NPV’s
in the first generation, the optimal NPV and the optimal NPV minus the worst NPV. The
results of Table 3.19 indicate that the longer the drilling time, the larger the ranges of the
NPV’s in the initial generation and the larger the improvements in NPV’s using drilling
order optimization.

Best Chrom: $9.0812 # 109

9
NPV, 109$

8.9

8.8

8.7
0 10 20
No. of Generations
(a) Td = 30 days (b) Td = 90 days

Figure 3.23: NPV versus generations obtained with drilling order optimization with different
drilling time Td using the optimal well locations and types (x, T ) obtained with seed 3,
Example 2-A.

85
Table 3.19: Summary of NPV’s obtained with drilling order optimization within different
drilling time Td using the optimal well locations and types (x, T ) obtained with seed 3,
Example 2-A.

Td NPV, $ × 109
days Initial Worst Initial Best - Initial Worst Optimal Optimal - Initial Worst
30 8.89 0.15 9.08 0.20
90 8.07 0.39 8.58 0.51

Table 3.20: Optimal drilling order with different drilling time Td using the optimal well
locations and types (x, T ) obtained with seed 3, Example 2-A.
Td (days) NPV, $ × 109 Wells
30 9.08 I01-7 P02-3 P03-2 P07-6 P08-12 I09-9 P11-5 P13-10 P16-4 P19-1 I21-11 I25-8
90 8.58 I01-5 P02-3 P03-2 P07-9 P08-12 I09-10 P11-8 P13-11 P16-4 P19-1 I21-7 I25-6

3.2.4 Investigation on population size

In a traditional genetic algorithm (GA), the population size is set by the user to
a fixed value at the beginning of the search and remains constant through the entire run.
Specification of the population size is problematic in many ways. If it is too small, then GA
may not be able to reach high quality solutions. If it is too large, then GA spends unnecessary
computational resources, and, in fact, may require more computational resources than can
be allocated. The scenarios are illustrated in Fig. 3.24.
Finding an adequate population size is a difficult task. It has been shown, both
theoretically and empirically, that the optimal population size is something that differs from
problem to problem. A somewhat widely accepted intuition towards population size is that it
should be set proportionally to the problem’s size and difficulty. However, problem difficulty
is very hard to define for real-world problems. Faced with such difficulties, many users end
up either using a so-called “standard setting” (50-100 population size) or rule of thumb that
the population size be twice the length of the chromosome [59], guessing a number, or doing
experimentation with a number of different sizes to see which one works best. In this section,
we investigate the effect of population size on GA performance based on Example 2-A.
In addition to Np = 50 used previously, three more different population sizes are
tested for the joint optimization of well locations and types in case A of Example 2. The

86
 Figure 3.24: Effect of population size on GA performance from Lobo et al. [51].

sizes tested are Np = 24, 72, and 108 which are equivalent to 13 , 1
1
and 1 12 of the chromosome
length respectively. Figs. 3.25, 3.26 and 3.27 respectively show the NPV versus generations
obtained with three seeds using Np = 24, 72, and 108. The corresponding results for Np = 50
are shown in Fig. 3.15. As one can see, seed 1 got stuck at a sub-optimal solution using
Np = 24 and the optimal results obtained using Np = 50, 72, 108 do not vary significantly
from seed to seed. Most importantly, with an increase in Np , better optimal NPV’s are
obtained. Table 3.21 shows a summary of the NPV’s obtained including the one using
Np = 50. From Table 3.21, Np = 50 is the best choice given 2000 simulation runs. The
general rule in GA literature for the optimum population size (twice the number of bits used
in a chromosome) is not necessary for this problem.

(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.25: NPV versus generations obtained for the optimization of well types and locations
using three seeds, Np =24, Example 2-A.

87
 (a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.26: NPV versus generations obtained for the optimization of well types and locations
using three seeds, Np =72, Example 2-A.

(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.27: NPV versus generations obtained for the optimization of well types and locations
using three seeds, Np =108, Example 2-A.

Table 3.21: Best NPV’s obtained by optimizing well locations and types with about 2000
reservoir simulation runs using different population sizes, Example 2-A.

Optimal NPV, ×109 $

Item
Np =24 Np =50 Np =72 Np =108
Iterations 40 80 20 40 20 40 20 40
Simulations 960 1920 1000 2000 1440 2880 2160 4320
Seed 1 8.96 9.03 9.07 9.29 9.23 9.23 9.09 9.32
Seed 2 8.91 9.39 9.20 9.32 9.27 9.32 9.23 9.46
Seed 3 9.15 9.29 8.88 9.34 9.05 9.23 9.27 9.39
Average 9.01 9.24 9.05 9.32 9.18 9.26 9.20 9.39

88
 3.3 Comparison with StoSAG Using Priority Parametrization

3.3.1 StoSAG using priority parametrization

Somewhat similar to our work, Leeuwenburgh et al. [44] selected Nw wells from a
given set of Kmax potential wells with Kmax > Nw and determined the drilling order of the
Nw wells using the priority parameterization. Recall that in the priority parametrization, (i)
each well is assigned one continuous priority variable for drilling order, and (ii) the larger
the priority variable, the earlier the well is drilled. Unlike our work, they did not consider
the well type as a distinct optimization variable. Instead, for each of the Kmax potential well
paths, they assigned a fixed well type. Then, they fix the well controls and optimize only the
set of priority variables where each priority variable corresponds to one specific well out of
the Kmax , and each priority variable is restricted to be in the interval [0, 2]. At each iteration
of the StoSAG algorithm, the Nw wells with the Nw highest priority variables are selected
to be drilled and the drilling order is based on the ordering of the priority variables where
the well with the highest value of its priority variable is drilled first. The paths of these
Nw wells and their drilling order along with the values of the fixed well controls are input
into the reservoir simulator to evaluate the NPV of production. In order to avoid having
multiple priority variables reach the same bound, [44] introduced an inequality constraint on
the sum of the priority variables and used the interior point method to solve the constrained
optimization problem. However, their numerical results show that, this methodology may
lead to lower NPV’s than the initial guess which is feasible but suboptimal; see the cyan
curve in Fig. 3.28 (or equivalently Fig. 3 of [44]).
To avoid solving a constrained optimization problem, they proposed to reset the
updated priority variables (after each iteration that improves NPV) to pre-set values with
equal intervals, such that only the ordering of values of priority variables changes at each
iteration. However, their numerical results showed that this method underperformed the one
in which the best performing perturbed vector of priorities is selected as the trial update as
opposed to using StoSAG which underscores our concern that it may be difficult to obtain

89
Figure 3.28: NPV versus iterations for the drilling order optimization process starting from
four different initial guesses (Fig. 3 in [44]).

good stochastic derivatives of the objective function when drilling order is parameterized
using continuous priority variables; see Fig. 3.29 or equivalently Fig. 6 of [44].

Figure 3.29: NPV versus iterations for the drilling order optimization of 4 injectors and 8
producers (Fig. 6 in [44].) Trial update using best performing perturbed priority sample is
shown in red line where individual samples are shown as gray dots. Trial updates obtained
with line search using StoSAG gradient estimated from normal distributed perturbations is
shown in blue line. Trial updates obtained with line search using StoSAG gradient estimated
from Bernoulli perturbations is shown in green line.

3.3.2 Numerical examples

Next, we apply a modified procedure of [44] to the two examples in last section. We
do not consider one of their examples because they do not provide enough information to
reproduce their example results and, even if they did, it would be preferable to consider a

90
more complicated problem. Leeuwenburgh et al. [44] consider two procedures for updating
the vector of priority variables at each iteration. In the first procedure, they generate a set of
perturbations of the current estimate of this priority vector and use the stochastic gradient
(StoSAG). Then they consider potential updates obtained by taking steps in the direction
of StoSAG to try to find an updated vector that increases NPV. In the second procedure,
they simply consider the perturbed vector of priorities and select as the update the one
that gives the largest NPV. In our modification, we include in the total set of trial updates
both the perturbed priorities and the updates generated by taking steps in the StoSAG
direction. From this total set, we select the one that gives the highest NPV. Since this
modification makes use of both the random search ability and the search ability along the
StoSAG gradient, it seems clear that it should outperform both individual update methods
considered in [44].
For both Example I (the channelized reservoir) and case A of Example 2 (the Brugge
model), there are 30 potential locations. Following [44], each of the 30 potential locations
is assigned one priority variable, based on which, the Nw well locations and drilling order
can be determined. Besides the 30 priority variables, we also added the type variables.
For the channelized reservoir, 30 well type variables are added. For case A of the Brugge
model, only 20 well types variables are added since 10 out of 30 potential locations are in
the aquifer which can only have “injector” type. The priority variables are preset as values
from 0.1 to 3.0 with an interval of 0.1; the type variables are preset as either 0.25 or 0.75. If
a type variable is greater than 0.5, an injector is placed; otherwise, a producer is placed. At
each iteration, the perturbation is sampled from Gaussian distribution N (0, 0.22 ) for priority
variables and N (0, 0.52 ) for type variables. The initial step size in the normalized StoSAG
direction is set as 0.5, the maximum number of step-size cuts is set as 3, and the maximum
number of perturbation resamples is set as 5.
For the channelized reservoir, Fig. 3.30 shows the NPV’s versus generations obtained
with simultaneous optimization of well locations and types using GA for three seeds. The
initial guess of the priority method is chosen as the best individual of the 50 random samples

91
of the well locations, types and drilling order generated at the first iteration of GA using seed
3 in Fig. 3.30. Fig. 3.31 shows the NPV versus the number of iterations using three seeds
following the modified methodology of [44]. The green square represents the case where the
update is the best performing perturbation, the blue star represents the scenario where the
update is along the StoSAG search direction, and the red circle represents an iteration where
no update that increases the NPV is obtained. When this occurs, new perturbations are
generated. If at one iteration, we generate five sets of perturbations without improving NPV,
the algorithm is terminated. As one can see, almost half of the iterations fail to improve
the NPV’s by searching along the StoSAG gradient direction. Moreover, the average of the
optimal NPV’s obtained in Fig. 3.31 is approximately 6% less than that generated with GA
in Fig. 3.30.

Best Chrom: $4.6341 # 109 Best Chrom: $4.53 # 109 Best Chrom: $4.6673 # 109

4 4 4
NPV, 109$

NPV, 109$

2 2 NPV, 109$ 2

0 0 0
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
No. of Generations No. of Generations No. of Generations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.30: NPV versus generations obtained with simultaneous optimization of (x, T , O)
using three seeds, Bound A, Example 1.

For case A of Example 2, Fig. 3.32 shows the NPV versus generations obtained with
simultaneous optimization of well locations, types and drilling order using GA with three
seeds. The initial guess of the priority method is chosen as the best individual of the 50
random samples generated at the first iteration of GA using seed 2 in Fig. 3.32. Fig. 3.33
shows the NPV versus the number of iterations using three seeds following the modified
methodology of [44]. As one can see, most of the NPV gain is obtained by random perturba-
tions while searching along the StoSAG direction only yields a slight increase. Moreover, the
optimal NPV’s obtained in Fig. 3.33 are approximately 8% less than the NPV’s generated
with GA in Fig. 3.32.

92
 (a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.31: NPV versus iterations obtained with simultaneous optimization of (x, T , O)
following the methodology of [44] using three seeds, Example 1. The green square represents
a successful update of the best performing perturbation, the blue star represents a successful
update along the StoSAG search direction, and the red circle represents a iteration where
no successful updates is obtained.

Best Chrom: $8.9618 # 109 Best Chrom: $9.1055 # 109 Best Chrom: $9.1078 # 109
10 10 10
NPV, 109$

NPV, 109$

NPV, 109$
5 5 5

0 0 0
0 20 40 60 0 20 40 60 0 20 40 60
No. of Generations No. of Generations No. of Generations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.32: NPV versus generations obtained with simultaneous optimization of (x, T , O)
using three seeds, Example 2-A.

(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 3.33: NPV versus iterations obtained with simultaneous optimization of (x, T , O) fol-
lowing the methodology of [44] using three seeds, Example 2-A. The green square represents
a successful update of the best performing perturbation, the blue star represents a successful
update along the StoSAG search direction, and the red circle represents a iteration where
no successful updates is obtained.

93
 3.4 Comparison with a Gradient-based Algorithm

As mentioned in Chapter 1, [83, 89, 26] used a gradient-based method to optimize

the number, types and control settings of wells given Kmax potential locations, a problem
similar to the one considered in this chapter. Hence, in this section, we provide a comparison
between GA and a gradient-based algorithm.

3.4.1 A gradient-based algorithm

In the gradient-based method in [83, 89, 26], a well is placed at each potential location
with a pre-specified type (injector or producer) and a liquid rate qj , j = 1, 2, · · · , Kmax which
applies during the whole reservoir life. The liquid rates are the optimization variables and
if qj → 0, well j is not drilled. The cost of drilling a well in location j is then approximated
with a continuous and differentiable function of the well rate, f (qj ), and the actual drilling
cost, Cw,j . The total drilling cost is K
P max
j=1 f (qj )Cw,j . It is desirable to have f (qj ) = 0 when

qj = 0 and f (qj ) = 1 when qj is large. By optimizing the control settings of Kmax wells and
by driving the rates to zero one by one to reduce the drilling cost, an approximation of the
optimal well number, types, locations and rates can be eventually obtained. If the algorithm
is terminated when the number of wells is equal to Nw , the same problem of interest is
solved, which is to select Nw wells from Kmax potential locations. Major disadvantages of
this procedure are that (i) it is only applicable to wells under rate control, (ii) only one
control step can be used and (iii) it cannot optimize the drilling order although that could
be obtained by a sequential optimization. Since only one control step is considered, this
problem can also be efficiently solved using GA by optimizing the well rates, types and
locations simultaneously. It is interesting to compare the optimal results obtained with the
gradient-based method and GA.
Wang et al. [83] proposed to define f (qj` ) using the following function

qj`
f (qj` ) = , (3.12)
qj` + β

where, β is a small constant number, 10−10 ; ` represents the `th iteration; qj is the injection

94
 β ∂f (qj` ) β
or production rate of well j at `th iteration. Hence, =
2 . When β is
∂qj` (qj` +β)2
≈
(qj` )
small, Eq. 3.12 tends to be insensitive to the change of controls which is not desirable.
Thus, Forouzanfar and Reynolds [26] proposed to use the following function for f (qj` )

!0.25
qj`
f (qj` ) = , (3.13)
qj`−1

where, qj`−1 represents the control variable (flow rate) of well j at (`−1)th iteration. However,
the above function can take values significantly different than 1.0 for qj` 6= 0. To avoid this
difficulty, Forouzanfar and Reynolds [26] simply set f (qj` ) = 1.0 if qj` 6= 0 and f (qj` ) = 0.0 if
∂f
qj` = 0. Hence there is an inconsistency in the evaluation of f (qj` ) and ∂qj`
. To ensure enough
sensitivities and keep the consistency, we propose to use a logistic form for f (qj` ) given by,

`
1 − e−ωqj
f (qj` ) = ` , (3.14)
1 + e−ωqj

∂f
where, ω is a weighting factor to control the magnitude of ∂qj`
. The logistic function is
composed of one steep part for small values of ωqj` where ωqj`  5 and one flat part where
f (qj` ) ≈ 1 for ωqj` > 5. Here, based on a computational experiment, ω is chosen to satisfy
ωqjcrit ≈ 5 (see Fig. 3.34) where qjcrit is an estimate of the minimum liquid rate necessary for
well j to be profitable. qjcrit is obtained by solving

qjcrit ∗ (ro − rw )Tres − Cw,j = 0, (3.15)

where ro is the oil price, rw is the water injection cost, and Tres is the reservoir life. Assuming
balanced injection and production, the left hand side of Eq. 3.15 represents the maximum
net revenue (non-discounted) for a well with rate qjcrit , i.e., the revenue of the oil production
(for a producer) or the displaced oil (for an injector) deducted by the water injection cost.
Besides the new form for f (qj` ), two more modifications are made based on [26] in the
implementation of the gradient-based method. The first modification is that each feasible

95
 1
f
∂f
0.8 ∂ωqj

0.6

j
f, ∂ωq
∂f
0.4

0.2

0
0 5 10 15 20
ωqj

Figure 3.34: Illustration of the logistic function and its derivative in Eq. 3.14.

potential location is assigned two wells of different types (one injector and one producer)
with same initial rates to avoid bias due to pre-specifying the type of each well as was done
in [26]. Hence, the optimization variable set is u = [q1I , q1P , q2I , q2P , · · · , qK
I
max
P
, qK max
], where
the superscript “I” represents an injector and the superscript “P ” represents a producer.
The objective is to eliminate wells one by one until only Nw wells are left while maximizing
the NPV function using the adjoint gradient ∇u J(u` ) where J(u` ) is given by

Nt
( "N NI
#)
P
`

X ∆tn X
n n n n n n

X
cnwi ·q nwi,k +cngi ·q ngi,k


J u = tn ro ·q o,j −cw ·q w,j −cg ·q g,j −
n=1 (1+b) 365 j=1 k=1
2K
(3.16)
Xmax

− f (u`j )Cw,j .
j=1

The second modification made to the algorithm used in [26] is that we removed the equality
constraint on the total injection (or production) rate so that the well control settings can be
adjusted more flexibly, i.e., only bound constraints are enforced. The lower bound, ulow , is
a 2Kmax -dimensional vector whose elements are zero and the upper bound, uup , is a 2Kmax -
dimensional vector whose elements are the maximum allowable rate for each well. The bound
constraints are enforced using the gradient projection method discussed below. Rewrite the
active constraints in the following form

Au` = b, (3.17)

where, A is the Nactive ×2Kmax -dimensional sensitivity matrix of the active bound constraints,

96
Nactive is the number of active bound constraints and b is an Nactive -dimensional vector
whose element can be either 0 if a lower bound constraint is active or uup
j if an upper bound

constraint is active. The projected search direction can be obtained as

d` = (I − AT (AAT )−1 A)∇u J(u` ). (3.18)

nP o
2∗Kmax
At each iteration, ∇u J(u` ) can be obtained by subtracting ∇u j=1 f (qj` )Cw,j
from the adjoint gradient outputted from the reservoir simulator. ∇u J(u` ) is then projected
`
onto the active bound constraints using Eq. 3.18 and an aggressive initial step size αmax along
d` is then tried to bring at least one component of u to either its lower or upper bound;
`
see Eqs. 3.19 and 3.20. If taking a step αmax d` fails to improve the objective function,
backtracking is implemented with a maximum number of step-size cuts equal to 5. If after
5 cuts, the trial update still fails in improving the objective function, we accept the initial
`
trial step αmax d` even though this decreases the NPV. We continue the iterations until there
are only Nw wells left or d` is empty. The detailed algorithm is shown in Algorithm 4.

`
αmax `
= min{αmax,j }N
j=1 .
u
(3.19)


low `
 uj `−uj , if d`j < 0,


` d
αmax,j = up
j
for j = 1, 2, · · · , Nu . (3.20)
`
 uj −u d`j
 j

`
dj
, if > 0,

3.4.2 Optimization of (x, T , u) using GA

It is computationally feasible to optimize the well control settings using GA when the
number of control steps is small. As the well control variables are continuous variables, it is
suitable to use a real value encoding. However, for the joint optimization of well locations,
types and controls using GA, well types change from generation to generation and the bounds
on controls for injectors and producers are usually different. For example, if injectors are
under rate control and producers are under BHP control where the upper bound on the

97
Algorithm 4 Pseudo-code for the gradient-based optimization

1. Place one injector and one producer at each of the Kmax potential paths. Set qjcrit , j =
1, 2, · · · , 2 ∗ Kmax and ω. Formulate the vector of the initial control variables, u0 . Set
the number of existing wells as 2∗Kmax , the maximum number of step-size cuts nsc = 5
and the number of step-size cuts ncuts = 0. Set ` = 0.

2. Calculate the gradient ∇u J(u` ) where J(u` ) is given in Eq. 3.16.

3. Find the active constraints at the estimate of u` , form A of Eq. 3.17 and then calculate
d` using Eq. 3.18.

4. If the number of existing wells is equal to Nw or d` is empty, terminate the algorithm.

`
5. Calculate αmax using Eqs. 3.19 and 3.20. Set α` = αmax
`
.

6. Take a trial step, u`+1,trial = u` + α` d` , run reservoir simulation and evaluate J(u`,trial ).

7. If J(u`,trial ) > J(u` ), set u` = u`,trial , ` = ` + 1, ncuts = 0.

8. Otherwise, check if ncuts < nsc . If yes, set α` = α` /2 and ncuts = ncuts + 1, go to step 6;
otherwise, set α` = αmax
`
and update u`+1 = u` + α` d` , set ` = ` + 1 and ncuts = 0.

9. Calculate the number of existing wells as number of potential wells with non-zero rate.
Go to step 2.

injection rate is much larger than the upper bound on the production BHP’s, it is possible
that after crossover and mutation, one producer is assigned a BHP value which is greater
than its upper bound, which thus may shut-in this producer; if both injectors and producers
are under BHP control, it is quite often that one producer may be shut-in if it switches to
an injector without changing the BHP’s in the next generation. To avoid these potential
problems, choking factors û between 0 to 1 are used as the design variables instead of well
(j)
controls u. If a well is an injector (Tj = 1), the actual control variable ui is then given by

(j) (j)
up
ui = ulow low
I + ûi (uI − uI ), (3.21)

where, ulow
I and uup
I are the lower bound and upper bound specified for the control variables

of injectors, i = 1, 2, · · · , Nw , j = 1, 2, · · · , Nc . If a well is a producer (Tj = 0), the actual

98
 (j)
control variable ui is given by

(j) (j)
up
ui = ulow low
P + ûi (uP − uP ), (3.22)

up
where, ulow
P and uP are the lower bound and upper bound specified for the control variables

of producers.
For real value encoded chromosomes, linear recombination is commonly used as the
crossover operation where the variables of the offsprings are chosen as the linear combinations
of variables of the parents (P1 and P2 ) using

C1 = aP1 + (1 − a)P2 . (3.23)

Here, a is a scaling factor randomly sampled from the uniform distribution U[−d, 1 + d]. A
value of d = 0.25 ensures (statistically) that the variable range of the offspring is the same
(j)
as the variable range spanned by the variables of the parents [61]. In the mutation, if ûi
is to be mutated, “controlled mutation” is applied, i.e., it will be replaced by a random
(j) (j)
value sampled from the uniform distribution given by U[0.9ûi , 1.1ûi ]. If the variables of
the produced offsprings are outside the specified bounds, we simply truncate them to the
nearest bounds.
To jointly optimize Nw well locations, types and control settings, we consider mixed
encodings of real values and binary values of

c = [x, T , û]. (3.24)

For the mixed encoding in Eq. 3.24, a similar procedure discussed for the crossover/mutation
operations for the chromosome of (x, T , O) is applied. If the randomly generated crossover
point falls in the part of c corresponding to the well control settings û, then the linear
recombination crossover is carried out for û; if the crossover point falls in the gene parts
corresponding to (x, T ), then chromosomes after the crossover point are exchanged using

99
the modified one-point crossover scheme discussed earlier. Similarly, the mutation operation
starts with the common gene-by-gene mutation. If any binary gene describing (x, T ) is to be
mutated, it will be flipped, i.e., 0 is changed to 1 or vice versa; if any of the gene describing
the well control settings is to be mutated, “controlled mutation” will be applied.

3.4.3 Numerical example: channelized reservoir

Based on the channelized reservoir example introduced in Section 3.2.1, we now com-
pare the performance of the gradient-based method [26] and GA. We consider the case that
both injectors and producers are under liquid rate control. The lower and upper bounds for
the injection and production rates for all wells are set as 0 STB/day and 40,000 STB/day re-
spectively. In this example, the critical rate qjcrit (see Eq. 3.15) is approximately 50 STB/day
for a vertical path and 100 STB/day for a horizontal path, the weighting factor ω is set as
0.1 and initial rate for each well at all potential locations is set as 1,200 STB/day. The
optimal rates, locations and types of Nw = 8 wells are determined. Fig. 3.35 shows the
NPV versus simulation runs (on the left) and the number of remaining wells versus simula-
tion runs (on the right) obtained with the gradient-based method. Note that the NPV does
not increase monotonically with respect to the simulation runs. When all well rates are far
greater than qjcrit , this problem reduces to a pure control optimization problem. Whenever
the back tracking line search method cannot find a step size that improves the NPV and
the number of wells is still greater than Nw , an aggressive step size is taken to bring at least
one component of u to either its lower or upper bound. Taking this aggressive step size may
decrease the NPV. From Fig. 3.35, we see that this NPV decrease happens frequently after
about 100 simulation runs. The highest NPV obtained is $7.37 × 109 with 21 wells. When
the number of wells is reduced to 8, the NPV is reduced to $6.59 × 109 .
Fig. 3.36 shows the NPV’s versus the number of generations obtained with GA. With
GA, the optimization variables are the locations of the 8 wells selected out of the 30 paths,
x, the type of the 8 wells, T , and the vectors of the scaled well controls, û. A chromosome
thus has the form (x, T , û). Binary encoding is used for (x, T ). For û, real value encoding

100
is used with a linear recombination crossover and controlled mutation. The optimal NPV
obtained with GA is $7.69 × 109 which is 14% higher the NPV obtained using the gradient-
based method. Moreover, the optimal NPV of GA is also higher than the best NPV that
the gradient-based algorithm achieved ($7.37 × 109 when the number of well is 21). Fig.
3.37 shows the optimal well placement and the remaining oil saturation field of layer 7. Fig.
3.38 shows the optimal well injection/production rates obtained with the gradient-based
method (left) and GA (right) respectively. The gradient-based algorithm finds one injector
and seven producers while the optimal result obtained with GA indicates that we should
select 2 injectors and 6 producers.

8 $7.37 # 109 60
Nw=21 $6.59 # 109
Nw=8 No. of Wells
NPV, 10 9 $

6 40

4 20 $7.37 # 109
Nw=21 $6.59 # 109
Nw=8
2 0
0 100 200 300 0 100 200 300
No. of Simulations No. of Simulations
(a) NPV versus simulation runs (b) Number of wells versus simula-
tion runs

Figure 3.35: NPV and number of wells versus simulation runs for the joint optimization of
well locations, types and rates obtained with the gradient-based method, Example 1.

Best Chrom: $7.6917 # 109

8
NPV, 109$

6
4
2
0
0 10 20 30
No. of Generations

Figure 3.36: NPV versus generations for the joint optimization of well locations, types and
rates obtained with GA, Example 1.

101
 (a) Gradient-based (b) GA

Figure 3.37: Optimal well configurations obtained with the gradient-based method and GA
method respectively, plotted on the oil saturation fields of layer 7 at the end of 4,000 days,
Example 1.

40 40
Rate, # 103 bbl

Rate, # 103 bbl

30 30

20 20

10 10

0 0
P02 I04 P09 P10 P13 P18 P24 P28 P09 P10 P12 I13 P17 P22 P23 I28

(a) Gradient-based (b) GA result

Figure 3.38: Optimal well injection/production rates obtained with the gradient-based
method and GA, Example 1.

3.5 Summary and Discussion

In this chapter, we considered the problem of maximizing the NPV by optimizing the
well locations, types and drilling order given a set of potential drilling paths and the number
of wells that are authorized to be drilled. After the categorical variables are optimized, we
optimize well controls. It is important to note the approach presented here would likely
become too computationally expensive to be practical if the potential drilling paths are
not restricted to a reasonably small number. Up to now, we have considered a maximum
of 64 potential drilling paths. Increasing the number to 128 or more would likely require a
computational cluster to be computationally feasible. Because we only consider selecting Nw
well paths from Kmax potential well paths, the optimization of the paths, types and drilling
order can be accomplished with the Genetic Algorithm (GA). If, however, the well paths
are allowed to be anywhere in the reservoir, it would be more natural to parametrize well

102
trajectories (paths) as continuous variables and use StoSAG as the optimization algorithm
[52, 90]. This is discussed in Chapter 5.
In implementations of GA, mixed encodings are applied where the well locations and
types are parameterized using binary encoding and the drilling sequence is parameterized
using ordinal numbers. Suitable mutation and crossover operations for the mixed encodings
are proposed and applied. The sets of encoded variables are optimized both simultaneously
and sequentially. Results indicate that GA gives good solutions in the following sense: (i)
the NPV produced is significantly larger than the NPV of any member of a set of initial
guesses; (ii) different runs of GA produce a variety of choices of optimal well paths, but
the variation in the estimated optimal NPV’s is relatively small, and (iii) GA outperformed
the method proposed by Leeuwenburgh et al. [44] for the simultaneous optimization of well
locations, types and drilling order, and GA also outperformed a gradient-based method by
[83, 89, 26] for the simultaneous optimization of well locations, types and rates.
.

103
 CHAPTER 4

JOINT OPTIMIZATION OF WELL NUMBER, TYPES AND LOCATIONS

In addition to the problem considered in last chapter, here, we consider a more general
problem where no prior information on potential well paths is given and the number of wells
to be drilled is not specified. Thus, the problem becomes as follows: given a maximum of
Kmax wells authorized to be drilled, determine how many wells actually should be drilled,
the path of each well and the type of each well (i.e., an injector or a producer). Again the
cost function to be maximized is the net present value (NPV) of production over the life the
reservoir. We firstly consider the joint optimization of well number, types and locations for
a deterministic reservoir model. Then we develop an efficient robust optimization procedure
when a large number of realizations are used to characterize the reservoir uncertainty.

4.1 Problem Description and Methodology

4.1.1 Problem description

Optimizing the number of wells explicitly may change the size of design variables (well
types, locations, control settings and so on) from iteration to iteration of the optimization
algorithm. To fix the size of the design variables, we specify the maximum number of wells
authorized to be drilled as Kmax , and then assign each well a “drill-or-not” status variable,
sj , j = 1, 2 · · · , Kmax . We use sj = 1 to denote “to-drill” well j and sj = 0 to denote “not-
to-drill” well j. Hence, the number of actual drilled wells can be obtained as Nw = K
P max
j=1 sj

by determining sj for Kmax wells. Similarly, assign each of the Kmax wells a type indicator,
Tj . As a well can only be an injector or a producer, Tj is also parameterized as a binary
variable, where, Tj = 1 represents an injector and Tj = 0 represents a producer. Let xj
denote the location of a well where the size of xj depends on the trajectory of well j. In this

104
chapter, we only consider placement of fully-penetrating vertical wells. Hence, xj is simply
the (i, j) indicies of the gridblock that well j is placed in. Given pre-specified operating
conditions for injectors and producers, well j can be characterized by wj = (sj , Tj , xj ). Note
that, (Tj , xj ) has an impact on the development plan only if well j is drilled, i.e., sj = 1;
otherwise, the (Tj , xj )’s only act like place holders in the design variable set.
With this parametrization, it is difficult to consider simultaneous optimization of
drilling order, O, and (s, T , x). Denote the time required to drill a well as Td days, then
following the last chapter, O could be parameterized as a Kmax -dimensional ordinal number,
i.e., a permutation of {1, 2, 3, · · · , Kmax }, and well j would be put into operation at time
(Oj − 1)Td days. However, even if well j is not drilled (i.e., well status of well j corresponds
to do-not-drill), it still holds (reserves) Td days of drilling time (i.e., the drilling rig does not
work for the next Td days) and because of the “rig down time,” the solution obtained will
be at least somewhat suboptimal. [35, 76] assign each of the Kmax wells a label, preset their
drilling order and opening time, and optimize the well status (“an injector,” “do-not-drill”
and “a producer”), locations and controls for these Kmax wells. In their work, to switch the
drilling order of well i and well j (i 6= j), the locations and controls for well i and j have to
be switched which is difficult for Mesh Adaptive Direct Search (MADS) or Particle Swarm
Optimization (PSO) (optimizers used in [35, 76]) when well i and well j are far away from
each other. Moreover, in their parameterization, the total number of control steps should be
Kmax (Kmax +1)
no less than 2
which can be computationally intensive for stochastic algorithms
(e.g., PSO and GA). Considering that in practice, well control optimization is usually not
conducted until all wells have been drilled and that it is difficult to find the optimal solution
of a high-dimensional optimization problem, we only consider optimizing the drilling order
and controls after the optimal (s, T , x) has been determined. In this way, the dimension of
the drilling order and control variables can be reduced to account for only Nw = K
P max
j=1 sj

wells. Since in last chapter, we have discussed the optimization of drilling order and well
controls, in this chapter, we only focus on the joint optimization of well number, types and
locations.

105
4.1.2 Constraints handling

Constrained optimization problems contain two classes of constraints: hard con-

straints (sometimes referred to as mandatory constraints) that must be satisfied by any
solution, and soft constraints, that may or may not be satisfied (Jiang et al. [38]). In
the optimization problem considered in this chapter, we consider four types of constraints,
including one hard constraint and three soft constraints.

1. All Kmax wells must be placed in the active gridblocks of a reservoir.

Denoting the set of active gridblocks as Ωactive , this constraint is given by

xj ∈ Ωactive , j = 1, 2, · · · , Kmax . (4.1)

This is a hard constraint that traditional constrained optimization methods (e.g.,

penalty method, augmented Lagrangian method, barrier method and so on) cannot
handle. Violation of this type of constraints may result in failed simulation runs where
there is no way to evaluate the NPV function. In this work, whenever a well is placed
in an inactive gridblock after a trial update or a perturbation (see the case of xj,trial
k+1

in Fig. 4.1), it is projected to the nearest bound of the active gridblocks (see the case
of xjk+1 in Fig. 4.1). Note that since xjk+1 represents the (i, j) gridblock indices which
are discrete, xjk+1 has to be truncated to the nearest integer values.

2. The inter-well distance between any pair of wells has to be larger than a predefined
value of the minimum allowable well spacing, which is denoted by Rmin . Let ci,j (s, T , x)
denote the well spacing constraint corresponding to well i and well j, where ci,j is
defined by
ci,j x(i) , x(j) = max {0, −(ri,j − Rmin )} ≤ 0,


(4.2)

where ri,j is the areal distance between well i and well j. In this chapter, ri,j is measured
Kmax (Kmax −1)
in terms of number of gridblocks. There are 2
well spacing constraints at
maximum.

106
Figure 4.1: Constraint handling when a well is placed in an inactive gridblock.

Even though the ci,j ’s are soft constraints, it is difficult to couple the traditional con-
strained optimization methods designed for continuous variables into the mixed-integer
problem in an efficient manner. In this work, we consider two methods to enforce the
constraints in Eq. 4.2. One is the repair method used in Genocop III [57], which re-
pairs infeasible solutions by combining them with feasible solutions. As an alternative,
we propose a method based on a heuristic mapping. The idea is simple: whenever a
pair of wells violate the well spacing constraints, they are moved away from each other
along the line through the two locations until the distance between them is greater
or equal to Rmin while both wells still fall in active gridblocks. If either well hits the
boundary of the active gridblocks, stop moving it and continue to move the other well
until the distance is greater or equal to Rmin . Note that this operation may lead to
constraint violations of other well pairs containing either of these two wells. Repeating
this operation many times (in the extreme case, it can be even an infinity number of
times) does not guarantee the well spacing constraints can be honored. Hence, we set
the maximum number of repeats as 20 Kmax (K2max −1) where Kmax (Kmax −1)
2
is the maxi-

107
 mum number of well pairs. Afterwards, we classify the Kmax wells into two sets, one
feasible set whose element does not cause any violation of the minimum well spacing
constraints and one infeasible set whose element cause at least one violation of the
minimum well spacing constraints. The active gridblocks within Rmin gridblocks away
from each well in the feasible set are marked as occupied gridblocks (the rest of the
gridblocks are marked as unoccupied gridblocks). The well with the maximum number
of violations, well w, whose location is x(w) , is then moved to the unoccupied active
(w)
gridblock nearest to x(w) , denoted as xnew , and well w is removed from the infeasible
set and added to the feasible set. The gridblocks within Rmin gridblocks away from
(w)
the current location of well w, xnew , are also marked as occupied and removed from
the unoccupied gridblocks. If the infeasible set is not empty, i.e., there exist wells that
cause violations, we continue to move the well with the maximum number of violations
to the unoccupied active gridblock nearest to its original location following the same
procedure. A detailed description of this procedure is given in Algorithm 5.

3. A producer cannot be placed in the aquifer. In this work, we explicitly convert a

producer in the aquifer to an injector.

4. The gas-oil-ratio and water cut of a producer cannot exceed prescribed upper limits.
These are two nonlinear state constraints and are enforced through the embedded well
management procedure in commercial simulators, e.g., a producer is shut-in if its water
cut exceeds a prescribed upper limit, or the oil rate is cutback by a predefined factor
if the gas-oil-ratio is higher than the predefined maximum value.

4.1.3 Joint optimization of well status, types and locations

We propose two algorithms to simultaneously optimize well status, types and loca-
tions. Note that it is possible that the aforementioned constraints are violated after each
update of the design variable (e.g., crossover, mutation, StoSAG perturbations, backtrack-
ing, GPS update and so on). Hence, the constraints on the design variables are enforced

108
Algorithm 5 Algorithm to enforce well spacing constraints
Set Nmove = 0, f easible = f alse.
Do while (Nmove < 20 Kmax (K2max −1) and f easible = f alse )
– Nmove = Nmove + 1;
– Find the infeasible well pair with minimum distance among all well pairs, move
this pair of wells apart along the line through the two locations; stop moving a well
if it hits the boundary of the active gridblocks and continue to move the other well
until the distance is greater or equal to Rmin .
– Check the well spacing constraint violations for all well pairs. If there is no violation,
set f easible = true; otherwise, set f easible = f alse.
EndDo
If f easible = f alse,
– Find all the infeasible pair of wells and put them in an infeasible set, W I , and put
the rest of the Kmax wells in a feasible set, W F .
– Do while (f easible = f alse)
∗ Put the active gridblocks within Rmin gridblocks away from each well in W F
as Ωoccupied and the rest of the active gridblocks as Ωunoccupied .
∗ Find the well with the largest number of constraint violations in W I and denote
it as well w. Move well w to a gridblock in Ωunoccupied nearest to its original
location and move well w from the infeasible set W I to the feasible set W F .
∗ If W I is empty, set f easible = true; otherwise, set f easible = f alse.
∗ If f easible = f alse and Ωunoccupied is empty, report warning message “Rmin is
specified too large” and break the loop.
– EndDo
EndIf

before each evaluation of the objective function using the strategies presented in last section.
Both algorithms we develop are compared with the Genocop III method used in [21].
The first algorithm is an iterative sequential optimization where we alternate opti-
mizing well status and types using GA with pure binary encoding and optimizing locations
of wells drilled in the chromosome with the highest objective function, x, using StoSAG or
GPS if a StoSAG step fails to improve the NPV. This algorithm is designed to improve the
convergence speed of GA and is presented in Algorithm 6. Since we start with the GA step,
a reasonable set of locations for all Kmax wells have to be provided as a prior. In the GA
step, we use the Stochastic Universal Sampling method [4] (discussed in Section 3.1.1) to

109
select the parent pool. Since the chromosome only consists of binary variables, we simply
use the one-point crossover operator and invert the bit of a gene if it is to be mutated (i.e.,
change 1 to 0 or vice versa). To apply this method, we need to pre-specify (i) the GA related
parameters including population size Np , maximum number of generations Ng , crossover
rate pc and mutation rate pm and number of consecutive GA generations NGA ; (ii) StoSAG
related parameters including number of perturbations Npert , initial step size α0 , maximum
number of step-size cuts is nsc , maximum number of resampling Nres and the number of
consecutive StoSAG iterations NSto . Although it is not clear how to set NGA and NSto , our
heuristic settings of NGA = 4 and NSto = 5 give a significant increase in the convergence
speed. We intentionally avoid using NGA = 1 because it is not likely that the best chromo-
some gets updated after only one generation of GA evolution and we want to balance the
contribution of optimizing (sj , Tj ) and xj . As the computational cost for the GPS method
can increase dramatically with respect to the dimensions of the optimization variables, we
limit the number of coordinate directions traversed in each GPS step to Nx . Since GPS is
used after the StoSAG algorithm converges to fine-tune the well locations, a fixed and small
mesh size is used, i.e., ∆0 = 1 gridblock.
A second algorithm which simultaneously optimizes well status, types and locations
using GA with mixed encoding of binary variables (for well status and types) and discrete
variables (for well locations) is proposed. Different crossover and mutation operations are
required for different encodings. When doing crossover, for each pair of parents (P1 , P2 ),
sample a random number from the uniform distribution, r ∼ U [0, 1]. If r < pc , then (i)
carry out one-pint crossover for genes corresponding to well status and types; (ii) obtain
the location information of an offspring using arithmetic crossover. When doing mutation,
for each gene, sample a random number r from U [0, 1]. If r < pm , invert bit for genes
corresponding to well status and types or move wells to one of its neighboring gridblocks for
genes corresponding to locations. A detailed description is provided in Algorithm 7.
Both the iterative sequential method and mixed-encoded GA are compared with the
Genocop III method used in Emerick et al. [21], see Algorithm 8. Similar to the mixed-

110
Algorithm 6 Iterative sequential algorithm to jointly optimize well number, types and
locations

1. Preset Np , Ng , pc , pm , NGA , NSto , ∆0 = 1 and the maximum number of simulation

runs, Nsim . Set F LSto = f alse and F LGP S = f alse. In general, F LSto = true
indicates that StoSAG has failed to improve the NPV in Nres consecutive iterations
and F LGP S = true indicates that GPS has failed to improve the NPV after searching
along Nx coordinate directions. Generate the initial locations for all Kmax wells based
on engineering judgment and generate the initial population where each chromosome
has the form c = [s1 , T1 , s2 , T2 , · · · , sKmax , TKmax ].

2. Continue for NGA generations of GA to optimize well status and types.

3. If F LSto = true (the StoSAG step failed), F LGP S = true (the GPS step failed) and the
best chromosome c remains the same during the last NGA iterations, go to step 4.

Otherwise,

• if F LSto = f alse (the StoSAG step did not fail), optimize the locations of wells
being drilled in the chromosome which gives the highest NPV among current
population, x, using StoSAG for NSto iterations. If StoSAG fails to improve the
NPV for Nres iterations, set F LSto = true (otherwise, F LSto remains f alse).
• else, optimize x using GPS for Nx coordinate directions. If GPS fails to improve
the NPV for Nx coordinate directions, set F LGP S = true (otherwise, F LGP S
remains f alse).
• If NPV is increased, evaluate each chromosome of the current population with
the updated well locations x. Go to Step 4.

EndIf

4. Termination criteria check. If the number of generations is greater than Ng , or the number
of simulations is greater than Nsim , terminate; otherwise, go to step 2.

encoded GA, the Genocop III method is designed for simultaneous optimization of well
status, types and locations. However, it is different from the mixed-encoded GA in two
aspects. Firstly, the Genocop III defines two populations, one search population where each
chromosome is only required to satisfy the hard constraints and one reference population
in which all chromosomes are required to satisfy all constraints considered. The population
size of the search population Nsp and the population size of the reference population Nrp can
be different, though we use Nsp = Nrp in the numerical examples of this chapter. At each
generation, a new search population is generated based on the existing search population

111
using GA operators (i.e., selection, crossover, mutation and replacement). However, it is
possible that some of the newly generated chromosomes are infeasible. Different from the
mixed-encoded GA which simply uses the heuristic mapping discussed earlier, the Genocop
III method repairs an infeasible chromosome S in the search population by combining it
with a feasible chromosome R which is randomly selected from the reference population.
Once a feasible chromosome Z is obtained after the repair of S, we evaluate the objective
function J(Z) and set J(S) = J(Z). If Z gives a larger objective function than R for a
maximization problem, R is then replaced by Z, i.e., Z is added to the reference population.
Hence, the best feasible solution is always given in the reference population. Moreover, Z
replaces S with the replacement ratio of pr . Whether to replace an individual S with its
repaired version Z is related to Lamarckian inheritance [78]: that is, can an organism pass
on traits that it acquires during its lifetime to its offspring? Some researchers never replace
individuals with their repaired versions (pr = 0), others always replace individuals with their
repaired versions (pr = 1), and others recommend that value of pr between 5% and 20%
gives good results (Schoenauer and Michalewicz [75], Orvosh and Davis [67]). In this work,
we use the same value for pr as Emerick et al. [21], which is 25%. Secondly, Genocop III uses
the arithmetic crossover (crossover operators designed for real values) while mixed-encoded
GA use different crossover operations for different types of variables (arithmetic crossover
for well locations and one-point crossover for binary variables).
In this work, the convergence criterion for the iterative sequential method is set as two
consecutive failures in the GA step followed by the GPS step or the GPS step followed by the
GA step; the convergence criterion for both the mixed-encoded GA method and the Genocop
III method is set as failure to improve the NPV in 10 consecutive generations. The iterative
sequential method combines the global search ability of GA and the local search ability of
StoSAG or GPS and is more likely to converge faster than either the mixed-encode GA or
the Genocop III method where only GA is utilized. The three algorithms are terminated
either at convergence or when the number of simulation runs exceeds a pre-defined value
(2,500 for deterministic optimization in this work).

112
Algorithm 7 Simultaneous optimization of well status, types and locations using mixed-
encoded GA

1. Preset Np , Ng , pc , pm and the maximum number of simulation runs Nsim .

Generate the initial population where each chromosome has the form c =
[s1 , T1 , x1 , s2 , T2 , x2 , · · · , sKmax , TKmax , xKmax ].

2. Evaluate the NPV for each of the chromosomes in the current population.

3. Termination criteria check. If the number of generations is greater than Ng , or the

number of simulations is greater than Nsim , terminate.

4. Selection. A portion of current population is selected to form the parent pool using the
Stochastic Universal Sampling (SUS) method.

5. Crossover. Select random pairs of parents from the parent pool without replacement. For
each pair of parents (P1 , P2 ), sample a random number from the uniform distribution,
r ∼ U [0, 1]. If r < pc , then

• randomly select a crossover point and exchange (sj , Tj ) on the right of the
crossover point between P1 and P2 .
• Sample a random number a from U [−0.25, 1.25], and obtain the location informa-
(C ) (P ) (P )
tion of an offspring by xj 1 = axj 1 +(1−a)xj 2 , j = 1, 2, · · · , Kmax where x(P1 )
and x(P2 ) represent the location information of P1 and P2 respectively; repeat this
(C ) (C ) (C )
procedure to obtain xj 2 . Note, xj 1 and xj 2 need to truncated to the nearest
integer value and bounded inside the active gridblocks. If well spacing constraints
are violated, apply heuristic mapping given in Section 4.1.2.
• Combine (sC C1 C1 C2 C2 C2
j , Tj ) and xj , (sj , Tj ) and xj to obtain two offspring chromo-
1

somes, c1 and c2 .

6. Mutation. For each gene, sample a random number r from U [0, 1], if r < pm , invert bit
for genes corresponding to (sj , Tj ) and move well xj to neighboring gridblock for genes
corresponding to xj . Goto step 2.

Following all three methods given in Algorithms 6, 7 and 8, we perform a second-stage

optimization which fine tunes the locations of wells that are drilled using StoSAG followed by
GPS. For simplicity of notation, the three two-stage optimization processes are still referred
to as the iterative sequential optimization method, mixed-encoded GA and the Genocop III
method respectively. In the iterative sequential method, if the first-stage optimization has
converged, then the second-stage optimization is not required. In the mixed-encoded GA and
the Genocop III method, even if the first-stage optimization has converged, a second-stage

113
Algorithm 8 Genocop III method

1. Preset the replacement probability pr , the search population size Nsp , the reference
population size Nrp . Preset Ng , pc , pm and the maximum number of simulation runs,
Nsim . Generate the initial search population which contains Nsp individuals that satisfy
the hard constraints (given in Eq. 4.1 in our case). Chromosomes in the search
population are not required to satisfy the soft constraints (given in Eq. 4.2 in our
case). Generate the initial reference population which contains Nrp individuals that
satisfy all the constraints (in our case, they are Eqs. 4.1 and 4.2). Each chromosome
has the form c = [s1 , T1 , x1 , s2 , T2 , x2 , · · · , sKmax , TKmax , xKmax ]. Evaluate the objective
function for each individual in the initial reference population.

2. Evaluate the objective function for each individual in the search population. Note, each
individual, if infeasible, has to be repaired according to Algorithm 9.

3. Termination criteria check. If the number of generations is greater than Ng , or the

number of simulations is greater than Nsim , terminate.

4. Selection. A portion of the current search population is selected to form the parent pool
using the Stochastic Universal Sampling (SUS) method.

5. Crossover. Select random pairs of parents from the parent pool without replacement.
For each pair of parents,

• sample a random number r from the uniform distribution U [0, 1];

• if r < pc , then do simple arithmetic crossover to generate a pair of offsprings, e.g.,

c1 = aP1 + (1 − a)P2 , (4.3)

where, P1 and P2 are the selected pair of parents, c1 is one of the gener-
ated offsprings, a is a random variable generated from a uniform distribution
U [−0.25, 1.25]. Truncate the genes corresponding to (sj , Tj ) in c1 to the nearest
integer in {0, 1}. Truncate the genes corresponding to xj in c1 to the nearest
integer and bound xj inside the active gridblocks using the procedure given in
Section 4.1.2. c2 is generated in the same way as for c1 .

7. Mutation. For each gene, sample a random number r from U [0, 1], if r < pm , invert bits
for genes corresponding to (sj , Tj ) and move well j to a random grid among the four
neighboring gridblocks for genes corresponding to xj . Goto step 2.

optimization is still necessary considering that it is difficult for GA to find the optimal well
location variables.

114
Algorithm 9 Pseudo-code for evaluation of the search population in Genocop III
For i = 1 : Nsp

• Denote S as the ith individual in the search population.

• If S does not satisfy constraints defined in Eq. 4.2,

– Set Z = S;
– Choose an individual, R, from the reference population;
– Do while (Z is not feasible)
∗ generate a random variable a from U [0, 1];
∗ Z = aZ + (1 − a)R;
– EndDo
– Evaluate the objective function of the repaired chromosome, J(Z);
– Set J(S) = J(Z);
– If J(Z) > J(R) for a maximization problem, set R = Z;
– Sample r from U [0, 1]. If r < pr , replace S by Z.

• EndIf

EndFor

4.2 Numerical Examples of Deterministic Optimization

In this section, the three proposed algorithms are tested on two numerical examples,
the PUNQ model and the Brugge model, where the maximum number of wells is set equal
to 20 and 30 respectively. For both examples, we determine the optimal well status, types
and locations with well controls fixed at their minimum BHP’s for producers and maximum
BHP’s (or rates) for injectors, respectively. For both examples, we generate a reference
scenario of the well status, types and locations based on engineering judgment and this
reference scenario is included as one of the candidate solutions in the initial population, i.e.,
the best NPV of the initial population is no less than the NPV obtained with the reference
scenario. For both examples, the optimal results obtained by the proposed three algorithms
are also compared with a pure well placement optimization of the reference scenario using
StoSAG followed by GPS.
In this work, when using GA, we set population size Np = 50, maximum number of
generations Ng = 50, maximum number of simulation runs in each stage of the optimization

115
process Nsim = 2, 500, the crossover probability pc = 0.9 and the mutation probability
pm = 0.01; when using StoSAG, we set the number of perturbations for each gradient
estimation Npert = 10, the initial step size α0 = 5 gridblocks, the maximum number of step-
size cuts is nsc = 4 and the maximum allowable number of consecutive failures in improving
the NPV, Nres , as 2. In the iterative sequential algorithm, the number of generations in
each GA step is set as NGA = 4 and the number of iterations in each StoSAG step is set
as NSto = 5. In both examples, the economic parameters are set as: ron = $80.0/STB,
cnw = $5.0/STB, cnwi = $5.0/STB, b=10.0%. In the PUNQ model, we set the drilling cost of
a well Cw,j = $30 × 106 while in the Brugge model, we set Cw,j = $6 × 106 . Moreover, in
both examples, the minimum well spacing is set equal to Rmin = 3 gridblocks.

4.2.1 Example 1: PUNQ-S3 model

The PUNQ-S3 reservoir is a synthetic model based on an actual North Sea reservoir.
The original simulation grid for the PUNQ-S3 problem consists of 19 × 28 × 5 grids with
∆X = ∆Y = 590.6 ft and 1,761 active gridblocks. The top structure of the field in Fig. 4.2,
shows that the field is bounded to the east and south by a fault, and links the north and
west to a fairly strong aquifer. A small gas cap is located in the center of the dome shaped
structure. The production life is set as 20 years. In this example, we consider the case
that all wells are vertical and a well can only be placed at the center of active gridblocks.
Producers operate under specified bottomhole pressure control at 1,740.45 psi. However, if
the gas-oil-ratio of a producer is greater than 1,121.92 SCF/STB, the oil rate will be cut back
to 75% of the oil flow rate at last time step and the well control will switch to oil rate control
for this time step. Water injection wells operate under rate control at 314.5 STB/day. In
this model, we preset the maximum number of wells authorized to be drilled, Kmax , equal to
20. The reference scenario contains 10 injectors and 10 producers, as shown in Fig. 4.3(a).
1. Well placement optimization of the reference scenario
We carried out a well placement optimization of the 20 wells given in the reference
scenario to illustrate the efficacy of StoSAG followed by GPS search and to benchmark the

116
Figure 4.2: The top layer of PUNQ model. Left: the depth of the top layer. Right: the
initial pressure of the top layer. The black curve delineates the water-oil contact and the
red curve delineates the gas-oil contact.

three proposed algorithms. Note in the following GPS search step, we set the initial mesh
size ∆0 equal to one gridblock since GPS is used to fine-tune the solutions obtained by the
StoSAG. Fig. 4.4 shows NPV versus the number of simulation runs where o’s represent
the StoSAG iterations and ∗’s represent the GPS iterations. As one can see, the StoSAG
algorithm increased the NPV from $1.88 × 109 to $2.07 × 109 in 44 simulations, and after
two consecutive failures in improving the NPV with StoSAG, the optimization switched to
the GPS algorithm after 75 simulation runs. The GPS algorithm obtained a similar increase
in the NPV, from $2.07 × 109 to $2.28 × 109 , in 375 simulation runs. This confirms our
previous discussion that StoSAG is efficient in optimizing numerical discrete variables, but
cannot do a good local search. Thus, a secondary optimization using GPS can give another
boost in the optimal NPV. Fig. 4.3(b) shows the optimal locations for the 20 wells after
the well placement optimization. Compared with the initial locations in Fig. 4.3(a), only
7 wells have moderate changes in the locations while the rest of the wells only have minor
changes. This also indicates that the locations given in the reference scenario are somewhat
reasonable.
Considering that there are only 20 vertical wells (40 location variables) and the PUNQ
model only contains 19 × 28 × 5 active gridblocks (i.e., the search domain is relatively small),
it is computationally feasible to use GPS alone for the well placement optimization following

117
 +

+ +
o
o +
+ o +
+ o
o +
o
o o
+ o o
o
+ o
+ o
+
+ o +
o o
+ +
o o o
o
+
+
+ +

(a) Reference (b) Optimal

Figure 4.3: Locations of 20 wells in the reference scenario (left) and their locations after
well placement optimization (right), plotted on the top layer of the PUNQ model. Injector
locations are marked in brown color and labeled with ⊗; producer locations are marked in
yellow color and labeled with o; the gas cap is marked in green color; the aquifer is marked
in dark blue and the oil zone is marked in light blue color.

Optimal NPV:2.2838# 108$

2.6
NPV, # 109$

2.4

2.2

1.8
0 100 200 300 400 500
Simulation Runs
Figure 4.4: NPV versus the number of simulation runs in the well placement optimization
of the 20 wells in the reference scenario of the PUNQ model. o’s represent the StoSAG
iterations and ∗’s represent the GPS iterations.

Algorithm 3. Hence, besides using StoSAG followed by GPS, we also consider the well
placement optimization of the reference scenario using only GPS. Three different initial
mesh sizes are used, i.e., ∆0 = 1 gridblock, ∆0 = 2 gridblocks and ∆0 = 3 gridblocks. In
the application of GPS, the mesh size is reduced by one gridblock when the NPV’s cannot
be improved after searching all coordinate directions and GPS is terminated only when the
minimum mesh size (i.e., 0 gridblock) is achieved. For each element of the variable vector at

118
`th iteration x` , e.g., the ith element, if either point of x` + ∆` ei and x` − ∆` ei gives higher
NPV than x` , then 5 follow-up trial updates along the direction that improves the NPV are
taken until the NPV stops improving where the size of each trial update equals ∆` . Fig. 4.5
shows the NPV versus number of simulation runs using GPS with three different ∆0 ’s. From
Fig. 4.5, with a small initial mesh size, GPS tends to take small steps which may cause the
algorithm to be trapped at a local solution which is far from the global optimum, see the
case of ∆0 = 1. An optimal NPV of $2.24 × 109 is obtained using GPS with ∆0 = 1, which
is 2% less than the NPV obtained using StoSAG follow by GPS with ∆0 = 1. Increasing
the initial mesh size can produce higher optimal NPV’s. For example, when ∆0 is increased
from one gridblock to two gridblocks, the optimal NPV is increased by 7% to $2.39 × 109 .
A further increase in ∆0 to 3 gridblokcs gives an optimal NPV of $2.38 × 109 which is
slightly less than that obtained with ∆0 = 2 gridblocks because a different local solution is
obtained. Increasing the initial mesh size would inevitably increase the computational cost.
For ∆0 = 1, 2, 3 gridblocks, the optimization processes cost approximately 400, 550, 850
simulation runs, respectively.

2.6
NPV, # 109$

2.4

2.2
"0 =1
2 "0 =2
"0 =3
1.8
0 500 1000
Simulation Runs
Figure 4.5: NPV versus number of simulations for the well placement optimization of the
reference scenario, using GPS with three different initial mesh sizes, PUNQ model.

To investigate the effect of well controls, a subsequent control optimization is carried

out based on the optimal well locations presented in Fig. 4.3(b). The upper and lower
bounds of the water injection rate is set as 314.5 STB/day and 0 STB/day respectively.

119
The upper and lower bound of the bottomhole pressure for a producer is set as 3,379 psi
and 1,740.45 psi respectively. The reservoir life is divided into 20 equal control steps where
each control step has the length of one year. The temporal correlation length is set to be
730 days. During all control steps, the initial guesses for the injection rates are set as 283
STB/day and the initial guesses for the production BHP’s are set as 1,885 psi. i.e., the initial
guesses for controls are somewhat different than but close to the fixed values used in the
well placement optimization. Fig. 4.6 shows the NPV versus the number of simulation runs
during the control optimization where the NPV is increased by approximately 2% compared
to setting the controls at the pre-fixed controls used to optimize the well locations, i.e.,
the upper bound of the injection rates and the lower bound of the production BHP’s. As
illustrated in Chapter 3 where we select Nw wells out of Kmax potential paths, whether the
subsequent control optimization can significantly increase the NPV’s or not depends on the
controls used to optimize the well locations and the bounds of the control settings. Fig.
4.7 shows the optimal well controls where most of the wells remain operating close to the
upper bound of the injection rates and the lower bound of the production BHP’s, but that
“P4” and “P8” tend to produce less in the later stage of the production life. It is worth to
mention that the optimal controls look a bit rough due to the fact that we set the temporal
correlation length small, i.e., equivalent to two control steps.

Optimal NPV:2.3318# 109$

2.4
NPV, # 109$

2.3

2.2

2.1
0 50 100 150 200 250
Simulation Runs
Figure 4.6: NPV versus number of simulation runs during the well control optimization
using the optimal well locations of the reference scenario, PUNQ model. The red dashed
line represents the optimal NPV’s obtained after the well placement optimization.

120
 1 1 300
2 2
3000 250
3 3
Producer index

Injector index
4 4 200
5 5
6 2500 6 150
7 7 100
8 8
9 2000 9 50
10 10
0
5 10 15 20 5 10 15 20
Control steps Control steps
(a) Production Control (b) Injection Control

Figure 4.7: Optimal controls obtained for the reference scenario with optimal well locations,
PUNQ model.

2. Joint Optimization of well status, types and locations

The definition of the initial population may impact the performance of the optimiza-
tion process. It is likely that a population of poor individuals will require several generations
until it produces a good solution. In this sense, employing the engineer’s knowledge to set
up the initial population is highly desirable. Hence, for all three algorithms to optimize
the well status, types and locations, the reference scenario developed based on engineering
judgment (given in Fig. 4.3(a)) is included in the initial population. The reference scenario
gives an NPV of $1.88 × 109 , which, as shown later, is higher than the rest of the randomly
generated chromosomes in the initial population, indicating that reference scenario defines
a good initial case.
In the iterative sequential algorithm, chromosomes within a population have identi-
cal well locations, but different well status and types. The initial population of the iterative
sequential method uses the Kmax = 20 well locations given in Fig. 4.3(a) and contains 50
candidate solutions of the well status and types where one chromosome is the same as given in
Fig. 4.3(a) and the rest of the 49 chromosomes are generated randomly by sampling the well
status sj and well type Tj from {0, 1} for each of the Kmax wells. Considering that it is pos-
sible that we sample a producer in the aquifer and the fact that we force it to be an injector,

121
we avoid placing more injectors than producers by increasing the probability of a well being
a producer (Tj = 0) in the initial population. As a reservoir generally has more producers
than injectors, we pre-set the the probability of a well being a producer (Tj = 0) is equal to
twice the probability of a well being an injector (Tj = 1). In the mixed-encoded GA method,
a chromosome contains information of the well status, types and locations of all Kmax wells,
and chromosomes within a population are different from each other. The initial population
of mixed-encoded GA method again consist of one reference chromosome given in Fig. 4.3(a)
and 49 random chromosomes. For each random chromosome, (sj , Tj ), j = 1, 2, · · · , Kmax
are sampled in the same way as in the iterative sequential algorithm while the location of
well j is randomly sampled from the 35 locations given in Fig. 4.8. For the Genocop III
method, both the initial search population and the initial reference population are generated
using the same way as for the mixed-encoded GA method, but that the chromosome given
in Fig. 4.3(a) is repeated 25 times in the initial reference population. Emmerick et al. [21]
show that starting from a reference population in which, the reference scenario (engineer
suggested case) represents 50 % of the individuals, leads to better solutions than starting
the optimization with a completely random reference population. By repairing the infeasible
chromosomes in the search population with the feasible chromosomes in the reference popu-
lation, the Genocop III method tends to generate an optimal solution similar to the reference
scenario, i.e., an “improved engineer defined case.” Since the distances between each pair of
wells in Fig. 4.8 and Fig. 4.3(a) are all greater than Rmin = 3 gridblocks, all chromosomes
of the initial population satisfy the minimum well spacing constraints.
The optimal NPV’s obtained with the three different algorithms using three random
seeds are plotted in Fig. 4.9 and summarized in Table 4.1. As one can see, the iterative
sequential algorithm converged in the first stage of optimization (i.e., a secondary well place-
ment optimization is not used) while both the mixed-encoded GA and Genocop III method
increased the NPV using the second-stage optimization to fine-tune the optimal well loca-
tions. During the first-stage optimization, the iterative sequential algorithm gives the highest
NPV, the mixed-encoded GA gives the second highest NPV while the Genocop III method

122
Figure 4.8: The set of potential locations (marked in yellow and red colors) for the initial
population, plotted on the top layer of the PUNQ model.

gives the worst NPV. After the second-stage optimization, the Genocop III method still gives
the lowest NPV on average, and the mixed-encoded GA gives a slightly higher NPV than
the iterative sequential algorithm because the first-stage optimization of the mixed-encoded
GA simultaneously optimized the well status, types and locations and hence finds a better
solution than the iterative sequential optimization procedure. However, as we found in last
chapter, simultaneous optimization does not always gives a higher estimated optimal NPV
than the a sequential method. The Genocop III method converges the slowest among all
three algorithms during the first-stage optimization since the linear combination process to
repair infeasible chromosomes tends to produce similar chromosomes to the reference sce-
nario (the engineer defined case), considering that the candidate solutions in the search
population have a good chance to be combined with the reference scenario. However, this
does not indicate that starting from a randomly generated reference population, Genocop
III could converge faster since the randomly generated chromosomes generally have lower
NPV’s than the reference scenario and linear combination with these poor individuals may
lead to an even slower converge. In any case, the Genocop III method produces the worst
NPV. Fig. 4.10 shows the NPV’s versus the number of simulations obtained with the best
seed of each of the three algorithms during the first-stage optimization. From Fig. 4.10, the
reference scenario given in Fig. 4.3(a) has the largest NPV among the initial populations
generated for all three algorithms, indicating the reference scenario is indeed a good initial

123
guess. In the iterative sequential algorithm, ∗’s represent the NPV’s of current population
re-evaluated with the updated locations after a StoSAG step, ∗’s represent the NPV’s of
current population re-evaluated with the updated locations after a GPS step. Even though
only the locations drilled in the best chromosome are updated in the StoSAG and GPS steps,
the NPV’s for the whole population are shifted up with a similar variance to NPV’s before
the well location updates. The first StoSAG step increased the NPV by $0.17 × 109 while
the second StoSAG step failed to improve the NPV for Nres = 2 iterations which invoked the
GPS step after NGA = 4 GA generations. The iterative sequential algorithm converged at
around 1600 simulation runs. The best NPV’s of each generation for the mixed-encoded GA
and the Genocop III method increase slowly until they are terminated at 2,500 simulation
runs. All three algorithms obtained higher optimal NPV’s compared to the well placement
optimization of the reference scenario, respectively by %11, %12 and %5. Table 4.2 shows
the simulation cost for the three different algorithms using three random seeds. From Table
4.2, the iterative sequential method requires the least number of simulation runs while both
the mixed-encoded GA and the Genocop III method require over 2,500 simulation runs.
Since StoSAG and GPS are better at local search compared to GA, the simulation runs for
the second-stage optimization is much less than the first-stage optimization.

Table 4.1: Comparison of the optimal NPV’s obtained with three different algorithms using
three seeds, PUNQ model.

Iterative sequential Mixed-encoded GA Genocop III

Seeds Stage #1 Stage #2 Stage #1 Stage #2 Stage #1 Stage #2
Seed 1 2.55 - 2.47 2.53 2.29 2.36
Seed 2 2.46 - 2.48 2.56 2.12 2.48
Seed 3 2.54 - 2.48 2.60 2.35 2.36
Average 2.52 - 2.48 2.56 2.25 2.40

The number of injectors and producers in the optimal solutions obtained with the
three different algorithms using three seeds are summarized in Table 4.3 and plotted in
Fig. 4.11. The iterative sequential algorithm and the mixed-encoded GA obtained similar
number of wells, i.e., 9 or 10 wells on average for three different runs. Due to the existence

124
 2.8

2.6 W1-x
NPV, # 109 $ W2-x
2.4 W3-x
W1-(s,T,x)
2.2 W2-(s,T,x)
W3-(s,T,x)
2 Reference

1.8
Seed 1 Seed 2 Seed 3 Average

Figure 4.9: Comparison of the optimal well numbers obtained with three different algorithms
using three seeds, PUNQ model. “W1-(s,T,x) and “W1-x” respectively represent the first-
stage and the second-stage optimization of the iterative sequential algorithm, “W2-(s,T,x)
and “W2-x” respectively represent the first-stage and the second-stage optimization of the
mixed-encoded GA, “W3-(s,T,x) and “W3-x” respectively represent the first-stage and the
second-stage optimization of the Genocop III method, and the black dashed line represents
the NPV of the reference scenario.

Best Chrom: $2.5492 # 109 Best Chrom: $2.4832 # 109 Best Chrom: $2.3529 # 109
NPV $ # 109

NPV $ # 109
NPV $# 109

2 2 2

1 1 1

0 0 0
0 1000 2000 0 1000 2000 0 1000 2000
No. of Simulations No. of Simulations No. of Simulations
(a) Iterative sequential - stage 1 (b) Mixed-encoded GA - stage 1 (c) Genocop III - stage 1

Figure 4.10: NPV’s versus the number of simulation runs obtained with the best seed of each
of the three algorithms, PUNQ model. In the iterative sequential algorithm, ∗’s represent
the NPV’s of current population re-evaluated with the updated locations after a StoSAG
search step, ∗’s represent the NPV’s of current population re-evaluated with the updated
locations after a GPS search step.

of the strong aquifer which provides pressure support for producers, the optimal solutions
tend to have no more than two injectors. As the Genocop III method tends to follow the
guidance of the reference scenario which has 20 wells, it obtained 15 wells on average, based
on three different runs, including 10 producers and 6 injectors. Fig. 4.12 shows the optimal
well locations obtained with the best seed of the three runs for each algorithm. From Fig.

125
Table 4.2: Comparison of simulation cost obtained with three different algorithms using
three seeds, PUNQ model.

Iterative sequential Mixed-encoded GA Genocop III

Seeds Stage #1 Stage #2 Stage #1 Stage #2 Stage #1 Stage #2
Seed 1 1619 - 2500 128 2500 164
Seed 2 2004 - 2500 187 2500 559
Seed 3 2295 - 2500 175 2500 225
Average 1973 - 2500 163 2500 316

4.12, even though the total number of wells is different, the optimal number and locations
of producers are somewhat similar to each other. Note that the inter-well distances are all
greater than the pre-defined minimum well spacing, Rmin = 3 gridblocks.

Table 4.3: Comparison of optimal well numbers obtained with three different algorithms
using three seeds, PUNQ model.

Iterative sequential Mixed-encoded GA Genocop III

Seeds Nw NP ro NInj Nw NP ro NInj Nw NP ro NInj
Seed 1 9 8 1 11 9 2 13 9 4
Seed 2 9 8 1 10 10 0 18 10 8
Seed 3 10 9 1 10 10 0 15 10 5
Average 9 8 1 10 10 1 15 10 6

20

W1-NInj
15
W2-NInj
W3-NInj
Nw

10
W1-NPro
W2-N
Pro
5
W3-NPro

0
Seed 1 Seed 2 Seed 3 Average

Figure 4.11: Comparison of optimal well numbers obtained with three different algorithms
using three seeds, PUNQ model. “W1” represents the iterative sequential algorithm, “W2”
represents the mixed-encoded GA, and “W3” represents the Genocop III method.

126
 + +

o
o o +
+
o o
o o
o o o
o o o o o o
o

o o
+ o + o + +
o
o o
o o o o

(a) Iterative sequential (b) Mixed-encoded GA (c) Genocop III

Figure 4.12: Optimal well locations obtained with the three algorithms using the best seed,
plotted on the top layer of the PUNQ model.

3. Investigation of the effect of drilling cost

In the preceding optimization of well status, types and locations, the drilling cost of a
well was set equal to $30 × 106 . To investigate the effect of the drilling cost, we consider two
more different drilling cost of a well, $100 × 106 and $8 × 106 . The mix-encoded GA which
performs the best during the optimization of (s, T , x) is used in this comparison. Figs. 4.13
and 4.14 respectively show the optimal NPV’s and the optimal number of wells obtained for
three different drilling cost.
When drilling a well is relatively cheap (e.g., Cw,j = $8 × 106 ), the total drilling
PKmax
cost (i.e., j=1 sj Cw,j ) represents a smaller penalty on the number of wells, and hence

the optimal solution tends to drill more wells. The case of Cw,j = $8 × 106 obtains the
largest number of wells (i.e., 9 producers and 5 injectors) along with the highest optimal
NPV among all three cases. Since there is a strong aquifer in the PUNQ model to provide
pressure support, with an increase in the drilling cost from $8×106 to $30×106 , injectors are
eliminated and only 10 producers are drilled. The fact that the two cases of Cw,j = $30 × 106
and Cw,j = $8 × 106 give similar optimal NPV’s may indicate that 10 producers are more
or less enough to deplete the PUNQ reservoir and the increase in well number beyond 10
producers only has a negligible affect on the overall production. With a further increase

127
in the drilling cost, i.e., from Cw,j = $30 × 106 to Cw,j = $100 × 106 , four more producers
are eliminated, i.e., only 6 producers are obtained which significantly reduced the optimal
NPV. The decrease in the optimal NPV is because of (i) the reduced production due to the
reduced number of producers and (ii) the increase in the total drilling cost.

3
NPV, # 109$

0
Cw,j=8#10 6$ Cw,j=30#10 6$ Cw,j=100 #10 6$

Figure 4.13: Optimal NPV’s obtained with different drilling cost using the mixed-encoded
GA algorithm, PUNQ model.

15
NInj
NPro
10
Nw

0
Cw,j=8#10 6$ Cw,j=30#10 6$ Cw,j=100 #10 6$

Figure 4.14: Optimal number of wells obtained with different drilling cost using the mixed-
encoded GA algorithm, PUNQ model.

4.2.2 Example 2: Brugge model

A brief introduction of the Brugge model was provided in Chapter 3. In this example,
we consider the case that all wells are vertical and a well can only be placed at the center of
active gridblocks. Both injectors and producers operate under bottomhole pressure control
where the BHP’s for producers are specified as 725 psi and BHP’s for injectors are specified

128
as 2,662 psi. However, a producer is shut in when its water cut is greater than 0.94. In this
model, we preset the maximum number of wells authorized to be drilled, Kmax , equal to
30. The original 30 wells (including 20 producers and 10 injectors) are used as the reference
scenario shown in Fig. 4.15(a). The reference scenario gives an NPV of $8.33 × 109 , which,
as shown later in Fig. 4.18, is higher than most of the randomly generated chromosomes in
the initial population, indicating that the reference scenario defines a good initial case. The
initial population for all three algorithms contains one chromosome of the reference scenario
given in Fig. 4.15(a) and 49 random scenarios generated in a similar way as for the PUNQ
model. For the mixed-encoded GA method and the Genocop III method, the well locations
of the 49 random chromosomes are randomly sampled from the 64 locations given in Fig.
3.20(a). Since the distances between each pair of wells in Fig. 3.20(a) and Fig. 4.15(a) are
all greater than Rmin = 3 gridblocks, all chromosomes of the initial population satisfy the
minimum well spacing constraints.
1. Well placement optimization of the reference scenario
We first carried out a well placement optimization of the 30 wells given in the reference
scenario (shown in Fig. 4.15(a)) to illustrate the efficacy of a GPS search after the StoSAG
search and to benchmark the three proposed algorithms. Fig. 4.16 shows the NPV versus the
number of simulation runs where o’s represent the StoSAG iterations and ∗’s represent the
GPS iterations. As one can see, the StoSAG algorithm increased the NPV from $8.33 × 109
to $11.18 × 109 in 283 simulations, and after two consecutive failures in improving the NPV,
the optimization switched to the GPS algorithm at 313 simulation runs. The GPS algorithm
increased the NPV by $109 , to $12.15×109 , in 616 simulation runs. This observation is quite
consistent with the result for the PUNQ model. Fig. 4.15(b) shows the optimal locations
for the 30 wells after the well placement optimization. Compared with the initial locations
in Fig. 4.15(a), two injectors are moved into the oil zone and two producers are moved into
the aquifer. Since a producer will be shut-in if its water cut exceeds 0.94, the two producers
placed in the aquifer are shut-in for the whole reservoir life to avoid extra water production.
The fact that two producers are placed in the aquifer suggests that 20 producers are enough.

129
 0.8 0.8

0.7 0.7

+ +
+ 0.6 0.6
+
+ o o o
o o + o
+ o 0.5 o o 0.5
o
+ o + o
o o o
+
o 0.4 0.4
+ o o oo
o o o o
+ o o +
+ +
+ o
0.3 o o 0.3
o o
o
oo o
o
+ o o
0.2 o + 0.2
+ +o

0.1 0.1

0 0

(a) Reference (b) Optimal

Figure 4.15: Locations of the 30 wells in the reference scenario (left) and their locations
after well placement optimization (right), plotted on the top layer of the Brugge model. ⊗
represents an injector; o represents a producer.

Optimal NPV:12.1498# 109$

13
NPV, # 109$

12
11
10
9
8
0 500 1000
Simulation Runs
Figure 4.16: NPV versus the number of simulation runs in the well placement optimization
of the 30 wells in the reference scenario of the Brugge model. o’s represent the StoSAG
iterations and ∗’s represent the GPS iterations.

2. Optimization of well status, types and locations

The optimal NPV’s obtained with three different algorithms using three seeds are
summarized in Table 4.4 and plotted in Fig. 4.17. For the iterative sequential method, only
the first seed requires a second-stage optimization where the NPV is increased slightly, while
for the other two seeds, the last GPS step of the first-stage optimization found local solutions

130
of the well locations and subsequent GA generations before termination did not change the
optimal well status and types. Both the mixed-encoded GA and the Genocop III method
require the second-stage optimization where the NPV’s are increased by more than $109
on average. During the first stage of optimization, the iterative sequential algorithm gives
the highest NPV, the mixed-encoded GA method gives the second highest NPV while the
Genocop III method gives the worst NPV. After the second-stage optimization, the mixed-
encoded GA gives a higher NPV than the iterative sequential algorithm while the Genocop
III method still gives the worst NPV. This again is quite consistent with the results from
the PUNQ model. The run obtained with seed 3 gives a relatively poor performance for all
three algorithms. Considering only the runs obtained with seed 1 and seed 2, the optimal
NPV’s obtained with the three different algorithms are similar and comparable to the optimal
NPV’s obtained with well placement optimization of the reference scenario, indicating that
optimizing well status and types only marginally improved the NPV. This is because the
drilling cost of a well is relatively cheap in this example, i.e., $6 × 106 , which imposes a small
penalty on number of wells. Hence, the optimal well status are somewhat similar to the
reference scenario.
Fig. 4.18 shows the NPV’s versus the number of simulation runs obtained with the
best seed of each of the three algorithms in the first-stage optimization. In the iterative
sequential algorithm, the fourth StoSAG search step failed to improve the NPV for Nres =
2 which invoked the GPS search step after NGA = 4 GA generations. Even though we
terminated the iterative sequential algorithm after 2,500 simulation runs, its NPV stabilized
at around 1,500 simulation runs. The best NPV’s of each generation of the mixed-encoded
GA and the Genocop III method increase slowly until they are terminated at 2,500 simulation
runs. Table 4.5 shows the simulation cost for the three different algorithms using three
random seeds. From Table 4.2, the iterative sequential method requires the least number
of simulation runs while both the mixed-encoded GA and the Genocop III method require
over 3,000 simulation runs on average of three runs. Since the Brugge model has many more
active gridblocks and more wells in the optimal solution compared to the PUNQ model, the

131
second-stage optimization requires more simulation runs for the Brugge model than for the
PUNQ model.

Table 4.4: Comparison of optimal NPV’s obtained with three different algorithms using
three seeds, Brugge model.

Iterative sequential Mixed-encoded GA Genocop III

Seeds Stage #1 Stage #2 Stage #1 Stage #2 Stage #1 Stage #2
Seed 1 12.27 12.32 11.00 12.23 11.28 12.32
Seed 2 12.28 – 11.39 12.60 10.63 11.53
Seed 3 11.94 – 11.35 12.46 11.22 12.12
Average 12.16 12.18 11.25 12.43 11.04 11.99

13

12 W1-x
NPV, # 109 $

W2-x
11 W3-x
W1-(s,T,x)
10 W2-(s,T,x)
W3-(s,T,x)
9 Reference

8
Seed 1 Seed 2 Seed 3 Average

Figure 4.17: Comparison of optimal well numbers obtained with three different algorithms
using three seeds, Brugge model. “W1-(s,T,x) and “W1-x” respectively represent the first-
stage and the second-stage optimization of the iterative sequential algorithm, “W2-(s,T,x)
and “W2-x” respectively represent the first-stage and the second-stage optimization of the
mixed-encoded GA, “W3-(s,T,x) and “W3-x” respectively represent the first-stage and the
second-stage optimization of the Genocop III method, and the black dashed line represent
the NPV of the reference scenario.

The number of injectors and producers in the optimal solutions obtained with three
different algorithms using three seeds are summarized in Table 4.6 and plotted in Fig. 4.19.
All three algorithms obtained a similar number of wells on average (25 or 26), but the number
of wells varies widely from seed to seed when the Genocop III is used. Fig. 4.20 shows the
optimal wells obtained with the best of the three runs for each algorithm. From Fig. 4.12,
even though the number of wells are different for each algorithm, in all cases, we obtain

132
 Best Chrom: $12.2671 # 109 Best Chrom: $11.3516 # 109 Best Chrom: $11.2615 # 109

NPV $ # 109
10 10
NPV $# 109
10

NPV $# 109
5 5 5

0 0 0
0 1000 2000 0 1000 2000 0 1000 2000
No. of Simulations No. of Simulations No. of Simulations
(a) Iterative sequential - stage 1 (b) Mixed-encoded - stage 1 (c) Genocop III - stage 1

Figure 4.18: NPV’s versus the number of simulations obtained with the best seed of each
of the three algorithms, Brugge model. In the iterative sequential algorithm, ∗’s and ∗’s re-
spectively represent the NPV’s of current population re-evaluated with the updated locations
after a StoSAG search step and a GPS search step.

Table 4.5: Comparison of the simulation cost for the three different algorithms using three
seeds, Brugge model.

Iterative sequential Mixed-encoded GA Genocop III

Seeds Stage #1 Stage #2 Stage #1 Stage #2 Stage #1 Stage #2
Seed 1 2500 433 2500 698 2500 720
Seed 2 2500 – 2500 630 2500 432
Seed 3 2500 – 2500 653 2500 500
Average 2500 144 2500 660 2500 552

Table 4.6: Comparison of optimal well numbers obtained with three different algorithms
using three seeds, Brugge model.

Iterative sequential Mixed-encoded GA Genocop III

Seeds Nw NP ro NInj Nw NP ro NInj Nw NP ro NInj
Seed 1 25 17 8 24 16 8 28 15 13
Seed 2 26 16 10 26 18 8 29 15 14
Seed 3 23 14 9 26 17 9 20 13 7
Average 25 16 9 25 17 8 26 14 12

several injectors in the oil zone. As the Genocop III method tends to follow the guidance
of the reference scenario, its optimal well placement is quite similar to that of the reference
scenario. The inter-well distances are greater than or equal to the pre-defined minimum well
spacing, Rmin = 3 gridblocks, for all results.

133
 30

W1-NInj
W2-N
20 Inj
W3-NInj
Nw

W1-NPro
10 W2-NPro
W3-NPro

0
Seed 1 Seed 2 Seed 3 Average

Figure 4.19: Comparison of the optimal well numbers obtained with three different algo-
rithms using three seeds, Brugge model. “W1” represents the iterative sequential algorithm,
“W2” represents the mixed-encoded GA, and “W3” represents the Genocop III method.

0.8 0.8 0.8

0.7 0.7 0.7

+
0.6 0.6 + 0.6
+ +
o + o
ooo + oo o
0.5 0.5 o 0.5
o + o + + o o
+ oo + o
o o + o
o 0.4 o 0.4 0.4
oo o + o +
+ + + o +
o o o
o o oo o +
+ +
+ o
0.3 o + 0.3 o o 0.3
o + o +
o o
oo o + o
o o
o 0.2 0.2 + + 0.2
+

0.1 0.1 0.1

0 0 0

(a) Iterative sequential (b) Mixed-encoded GA (c) Genocop III

Figure 4.20: Optimal well locations obtained with the three algorithms using the best seed,
plotted on the oil saturation field of the top layer after 20 years of production, Brugge model.
⊗ represents an injector; o represents a producer.

4.3 Field Development Optimization Considering Geological

Uncertainty

There is significant geological uncertainty in the reservoir description due to the

134
limited knowledge about the underground formation. The common approach to account for
geological uncertainty is to use multiple plausible geostatistical realizations of the reservoir
model. However, the computational cost when a large ensemble of realizations is used may
be prohibitive. Hence, only a small ensemble is affordable in robust optimization, which may
not be sufficient to capture the subsurface uncertainty. Thus, in this section, we propose a
relatively efficient procedure for robust optimization of well status, types and locations when
a large number of representative realizations are considered.
Denote the number of realizations to represent the geological uncertainty by Ne , and
denote the kth realization (a vector of the model parameters) by mk , then the full-set
ensemble of reservoir models is given by

Ωf = {mk |k = 1, 2, · · · , Ne }.

The most common approach for handling geological uncertainty is to maximize the expected
value of the objective function (NPV in this work) over the ensemble of Ne realizations
[86, 22, 20, 49, 52], i.e.,

Ne
Ω 1 X
JE f (s, T , x) = J(mk , s, T , x), (4.4)
Ne k=1

which is commonly referred to as robust optimization in the literature. The computational

cost for robust optimization is much higher compared to deterministic optimization since
each objective function evaluation requires Ne reservoir simulation runs. Hence, to reduce
the computational cost, besides developing efficient optimization algorithms, we also focus
on adaptively selecting a smaller subset of representative models and optimizing over the
subset ensemble.
Denote the subset ensemble of the representative models by

Ωs = {mjk |mjk ∈ Ωf , k = 1, 2, · · · , Nes },

135
where Nes is the number of realizations in the subset. We assume the realizations of the
reservoir models in the full ensemble Ωf could be represented by a subset of realizations Ωs
in the sense that optimizing the expected NPV over the subset gives a good approximation
of that would be obtained by maximizing the expected NPV over Ne realization. Note this
representative subset may change as the estimate of (s, T , x) changes. Under the preceding
assumption, the expected NPV over the full ensemble can be approximated by

Nes
1 X
JEΩs (s, T , x) = J(mjk , s, T , x). (4.5)
Nes k=1

Considering that the computational cost is the main concern of robust optimization,
and that the iterative sequential algorithm is the most efficient of the three algorithms pro-
posed for deterministic optimization of well status, types and locations (see Tables 4.2 and
4.5), the iterative sequential algorithm is adapted to robust optimization. Theoretically, at
each iteration, whether an updated estimate of (s, T , x) obtained using the subset ensemble
should be be accepted or not, depends on whether the expected NPV of the whole ensem-
Ω
ble models, JE f (s, T , x), is improved or not. However, considering that each evaluation of
Ω
JE f (s, T , x) requires Ne simulation runs, we avoid evaluating it every iteration of the opti-
mization algorithm. We emphasize that the base algorithm is an iterative sequential method
which combines GA, StoSAG and GPS. If the population size of GA is set equal to 50, the
number of GA iterations is set to 50, and the number of realizations in the subset Ωs is set
equal to 5, then it would require at least 50 ∗ 50 ∗ 5 = 12, 500 simulation runs to carry out the
optimization of (s, T ), without considering the simulation runs spent on adaptively selecting
the subset realizations and on the StoSAG and GPS search steps to update x. Hence, it
is desirable to keep the the size of the subset ensemble small. Thus, in this chapter, the
maximum number of realizations in Ωs is set equal to 3. The representative models in the
subset are selected as the models corresponding to the equally spaced percentiles of the cu-
mulative distribution function (cdf) of the NPV’s for Ne realizations. Suppose Ne = 50 and
Nes = 3, then Ωs includes the realizations corresponding to the P2, P50 and P100 (2th, 50th

136
and 100th percentiles) of the cdf of NPV in Ωf .
Below, we provide three different “procedures” in the context of Algorithm 6 to
determine when to apply the estimates of (s, T , x) obtained by maximizing JEΩs to all Ne
realizations and how to update the subset ensemble. Procedures 1 and 2 are used as the
benchmark for Procedure 3.

• Procedure 1. The cdf of NPV for the full ensemble, Ωf , is only constructed once based
on the initial settings of the well status, types and locations (i.e., the reference scenario
generated using engineering judgment) and the realizations corresponding to P2, P50
and P100 are selected to form the subset ensemble. Throughout the optimization
process, these three realizations are used to represent the geological uncertainty and
each evaluation of the objective function (given in Eq. 4.5) requires 3 simulation runs.

• Procedure 2. The cdf of NPV for the full ensemble Ωf is initially constructed based
on the initial settings of (s, T , x) and reservoir models corresponding to P2, P50 and
P100 of the initial cdf of NPV are selected to form the initial subset. For each search
step (the GA step, the StoSAG or the GPS step), we use the best estimate of the
optimal solution (the estimate of (s, T , x) that gives the largest expected NPV for the
Ne realizations, denoted as wopt ) as the initial guess. Denote the estimate of (s, T , x)
after carrying out the kth search step using the selected representative subset, Ωk−1
s ,
as v k,opt . After each search step, the NPV’s of the Ne realizations in the full ensemble
Ω
are evaluated using v k,opt and the expected NPV of the Ne realizations, JE f (v k,opt ),
Ω
is calculated and a trial cdf of NPV is built. If JE f (v k,opt ) is no greater than the
Ω
one calculated using the best estimate of the optimal solution, JE f (wopt ), we do not
accept the updated solution, i.e., we replace the updated solution by the solution at
the beginning of the step. Otherwise, we (i) accept v k,opt as the best estimate of the
optimal solution and accept the trial cdf of NPV, and (ii) select a new representative
subset of realizations (reservoir models corresponding to P2, P50 and P100 of the newly
built cdf of NPV), Ωks , to represent the geological uncertainty for the next GA, StoSAG

137
 or GPS search step, whichever comes first.

• Procedure 3. In Procedure 2, to decide if an estimate of (s, T , x) obtained with the

subset is accepted as the best estimate of the optimal solution, the NPV’s of the Ne
realizations are evaluated after each search step (the GA step, StoSAG step or GPS
step). However, it is possible that the preceding search step leads to a big change to
the design variables (s, T , x) with which the subset realizations are no longer repre-
sentative. Hence, to decide if the subset needs to be re-selected, we also evaluate the
Ne realizations within each search step provided that a specified criterion is satisfied.
The cdf of the full ensemble Ωf is initially constructed based on the reference scenario
(the initial settings of the (s, T , x)). A set of reference realizations (Ω0co ) are chosen as
the reservoir models corresponding to P2, P25, P50, P27 and P100 of the initial cdf of
NPV and one representative realization, the P50 of the initial cdf of NPV, is selected
to form the representative subset. Denote the estimate of (s, T , x) obtained with the
subset realizations at the `th iteration as v ` and denote the best estimate of the op-
timal solution obtained till `th iteration as wopt . If v ` improves the expected NPV of
the subset ensemble compared to last iteration, then the NPV’s of the realizations in
Ω`co are evaluated using v ` . If the ordering of the NPV’s of the realizations in Ω`co re-
mains the same as the ordering at the best estimate of the optimal solution, the subset
realizations are still considered to be representative and we move on to next iteration;
otherwise the NPV’s of the Ne realizations are evaluated using v ` and the expected
Ω Ω
NPV, JE f (v ` ), is calculated. If JE f (v ` ) is not greater than the one obtained with the
Ω
best estimate of the optimal solution, JE f (wopt ), we do not accept the updated solu-
tion, i.e., we replace the updated solution by wopt , and enlarge the size of the subset to
3 if there is only one realization in the subset, i.e., the reservoir models corresponding
to the P2, P50 and P100 of the best estimate of the cdf are selected. Otherwise, we
(i) accept v ` as the next estimate of the optimal solution (set wopt = v ` ) and build
the cdf of NPV based on Ne realizations using v ` , and (ii) select a new representative
subset of realizations (models corresponding to P2, P50 and P100 of the newly built

138
 cdf of NPV), Ω`s , and a new reference set, Ω`co , for next iteration.

For cases where the representative models are relatively independent of the design vari-
ables, (s, T , x), using only one realization (the adaptively selected P50 realization) can
greatly save the simulation cost; for cases where the representative models heavily rely
on the design variables, (s, T , x), the adaptively selected three realizations (P2, P50
and P100 realizations of the newly built cdf) can roughly represent the full ensemble.
To estimate the StoSAG search direction, if the subset ensemble size is increased, the
number of perturbations per realization, Npert , is reduced in order to keep the com-
putational cost low, i.e., we set Npert = 15 for Nes = 1 and Npert = 5 for Nes = 3.
Algorithm 6 coupled with Procedure 3 is summarized in Algorithm 10. In the context
of Algorithm 10, Procedure 3 is given in Algorithm 11.

In Algorithm 10, ALidx represents the indices of the three search step, i.e., ALidx = 1
represents the GA search step, ALidx = 2 represents the StoSAG search step, ALidx = 3 rep-
resents the GPS search step; F LSto = true and F LGP S = true respectively represent failures
in StoSAG search step and the GPS search step; evalΩf = true represents evaluating the
NPV’s of the Ne realizations in Ωf ; s0 , T 0 and x0 respectively represent the initial settings
of well status, types and locations (i.e., the reference scenario generated using engineering
judgment); the superscript opt denotes the best estimate of the optimal solution; xopt , copt
and P OP opt respectively represent the best estimate of well locations, status and types, and
the population which contains copt ; ` denotes the number of iterations within each search
step; Ω`s and Ω`co respectively represent the subset ensemble to account for the uncertainty
and the reference ensemble to check the ordering of NPV’s at `th iteration of a search step;
P OPs` represent the population at `th iteration; c`i and c`b respectively represent the ith
chromosome and the best chromosome in P OPs` ; x`,trial and x` respectively represent the
well locations after a trial update and the actual value of well locations at `th iteration.
Robust optimization of well status, types and locations given the well controls, using
the iterative sequential method coupled with Procedure 1, Procedure 2 and Procedure 3 are
referred to Robust-1, Robust-2 and Robust-3, respectively. In this section, the convergence

139
Algorithm 10 Pseudo-code for adaptive robust optimization using Procedure 3

Preset Np , Ng , pc , pm , NGA , NSto and ∆0 = 1. Set F LSto = f alse and F LGP S = f alse.
Generate the initial settings of well status s0 , types T 0 , and locations x0 . Set c0b =
Ω
[s01 , T10 , s02 , T20 , · · · , s0Kmax , TK0 max ], copt = c0 and xopt = x0 . Evaluate JE f (copt , xopt ) and
form cdf opt . Set the size of the initial subset, Nes = 1, select the P50 realization of
cdf opt to form Ω0s and select the “P2, P25, P50, P75 and P100” realizations of cdf opt
to form Ω0co . Set the number of perturbations used to estimate the StoSAG gradient
at each iteration, Npert = 15.
1. Generate the initial population, P OPs0 , with Np random chromosomes and replace one
random chromosome by c0 . Evaluate P OPs0 and set P OP opt = P OPs0 .
2. Set ` = 0, P OPs` = P OP opt , c`b = copt and x` = xopt . If ALidx = 1, evaluate each
chromosome in P OPs` using xopt , where the fitness of a chromosome is the expected
NPV of the realizations in Ω`s , and go to step 3; if ALidx = 2, go to step 4; if ALidx = 3,
go to step 5.
Ω`
3. Continue for NGA generations of GA to maximize JE s by adjusting the well status
and types. At each iteration, (i) set Ω`+1 s = Ω`s , Ω`+1
co = Ω`co and ` = ` + 1; (ii)
`−1
generate P OPs` using GA operators based on P OPs`−1 ; (iii) evaluate JEΩs (c`i , xopt )
`−1
for c`i in P OPs` and find the chromosome with the largest JEΩs (c`i , xopt ), c`b ; (iv) if
`−1 `−1
JEΩs (c`b , xopt ) > JEΩs (c`−1 opt
b , x ), apply Procedure 3 given in Algorithm 11.
Ω`
4. Maximize JE s by adjusting x, using StoSAG for NSto iterations. The number of pertur-
bations used to estimate the StoSAG gradient per realization is set equal to NNpert es
. At
`+1 ` `+1 `
each iteration, (i) set Ωs = Ωs , Ωco = Ωco and ` = `+1; (ii) search along the StoSAG
`−1 `−1
gradient to obtain a trial update x`,trial ; (iii) if JEΩs (copt , x`,trial ) > JEΩs (copt
b ,x
`−1
),
` `,trial ` `−1
set x = x and apply Procedure 3 given in Algorithm 11, otherwise set x = x
`−1
and go to next iteration. If StoSAG failed to improve JEΩs for Nres iterations, set
F LSto = true; otherwise, set F LSto = f alse. Go to Step 7.
Ω`
5. Maximize JE s by adjusting x along Nx coordinate directions. At each iteration, (i) set
Ω`+1
s = Ω`s , Ω`+1 `
co = Ωco and ` = `+1; (ii) search along the coordinate direction to obtain
`−1 `−1
a trial update x`,trial ; (iii) if JEΩs (copt , x`,trial ) > JEΩs (copt
b ,x
`−1
), set x` = x`,trial
and apply Procedure 3 given in Algorithm 11, otherwise set x = x`−1 and go to next
`

iteration. If GPS failed to improve JEΩs for Nx coordinate directions, set F LGP S = true;
otherwise, set F LGP S = f alse.
6. If ALidx 6= 1, set ALidx = 1; otherwise, set ALidx = 2 if F LSto = f alse, and set ALidx = 3
if F LSto = true and F LGP S = f alse. Go to step 2.

criterion for Robust-1 is set as failure to improve the expected NPV of the three fixed
realizations in a GA step followed by a GPS step or a GPS step followed by a GA step. The
convergence criterion for Robust-2 and Robust-3 is set as consecutive failures in updating

140
Algorithm 11 Procedure 3 for adaptive robust optimization of well status, types and loca-
tions

• Set evalΩf = f alse.

• If ALidx = 1,

– set vb` = (c`b , xopt );

– if ` = NGA , set evalΩf = true.

• If ALidx = 2,

– set vb` = (copt , x` );

– if ` = NSto or F Lsto = true, set evalΩf = true.

• If ALidx = 3,

– set vb` = (copt , x` );

– if ` = Nx , set evalΩf = true.

• If evalΩf = f alse,

– evaluate J(mk , vb` ), ∀ mk ∈ Ω`−1

co ;

– if the ordering of the NPV’s in Ω`−1

co change, set evalΩf = true.

• If evalΩf = true,
Ω
– evaluate JE f (vb` ) and build a trial cdf, cdf trial ;
Ω Ω
– if JE f (vb` ) > JE f (copt , xopt ),
∗ if ALidx = 1, set copt = c`b and P OP opt = P OPs` ;
∗ if ALidx 6= 1, set xopt = x` ;
∗ set cdf opt = cdf trial and re-select realizations from cdf opt to form Ω`s and Ω`co ;
Ω`
∗ evaluate JE s (vb` ).
– otherwise,
∗ set x` = xopt , c`b = copt and P OPs` = P OP opt ;
∗ if Nes = 1, set Nes = 3, and re-select the “P2, P50 and P100” realizations of
cdf opt to form Ω`s .

• Otherwise, move to next iteration.

the cdf of the Ne realizations in a GA step followed by a GPS step or a GPS step followed
by a GA step. The three methods (i.e., Robust-1, Robust-2 and Robust-3) are terminated
either at convergence or when the number of simulation runs exceeds a pre-defined value.

141
 4.4 Numerical Examples of Robust Optimization

In this section, the three robust optimization procedures (i.e., Robust-1, Robust-2 and
Robust-3) are applied to both the PUNQ model and the Brugge model. For both examples,
50 realizations of the reservoir model are generated to represent the geological uncertainty.
The maximum number of simulation runs is set as 10,000 for the PUNQ model and 8,000
for the Brugge model.

4.4.1 Example 1: PUNQ model

The PUNQ model is considered where 50 realizations of the horizontal and vertical
permeability fields are used to represent the geological uncertainty. The same reference well
placement scenario, well control settings and economic parameters are used as in determin-
istic optimization presented in Section 4.2.1. The only difference is that we consider the
minimum well spacing as one gridblock, i.e., two wells cannot be placed in the same grid-
block. The NPV’s for all 50 realizations are calculated using the reference scenario given
in Fig. 4.3(a) where the cdf of NPV is shown as the black curve in Fig. 4.21. The P2,
P50 and P100 realizations correspond to realization #48, #8 and #27 respectively, whose
log-permeability fields are shown in Fig. 4.22. From Fig. 4.22, the permeability fields of
these three realizations are quite different, indicating the large geological uncertainty in the
reservoir models. We first carry out the adaptive robust optimization using iterative sequen-
tial steps of GA, StoSAG and GPS, and then we consider the case where the StoSAG search
step is not used.
1. Robust optimization using the iterative sequential algorithm with GA, StoSAG
and GPS
Since GA and StoSAG are stochastic algorithms, three initial seeds are tried for each
of the three robust optimization procedures. Fig. 4.21 shows the cdf’s of NPV obtained
with the reference scenario and the optimal well placement using three seeds. The cdf’s
of NPV are generated based on evaluating the NPV of all Ne = 50 reservoir models given
an estimate of (s, T , x). The green, blue and red circles respectively represent the P2,

142
P50 and P100 realizations of the initial cdf of NPV (the black curve) obtained with the
reference scenario. From Fig. 4.21, the subset realizations selected from the initial cdf no
longer correspond to the 2th, 50th and 100th percentiles of the cdf’s of NPV obtained with
the optimal solutions, indicating that the representative realizations heavily depend on the
optimization variables for the PUNQ model and that it is necessary to adaptively re-select
the representative realizations.

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

cdf

cdf

cdf
0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
1 1.5 2 2.5 3 3.5 1 1.5 2 2.5 3 3.5 1 1.5 2 2.5 3 3.5
NPV, # 109 $ NPV, # 109 $ NPV, # 109 $
(a) Robust-1 (b) Robust-2 (c) Robust-3

Figure 4.21: Cdf’s of NPV obtained with the reference scenario (black curve) and the optimal
well placements for the three robust procedures using three seeds (blue curve for seed 1, red
curve for seed 2 and green curve for seed 3), robust optimization of the PUNQ model. The
green, blue and red circles respectively represent the P2, P50 and P100 of the cdf of NPV
obtained with the reference scenario.

From Fig. 4.21(a), using Robust-1 (where the subset is fixed during the optimization
process), the optimal cdf’s obtained with the three seeds are quite different in that seed 2
obtained lower optimal NPV’s for almost all 50 realizations compared to those of the other
two seeds. From Figs. 4.21(b) and (c), the optimal results obtained with Robust-2 (where
the subset is re-selected at each switch of the search steps) and Robust-3 (where the subset
is re-selected following Algorithm 11), are almost independent of the seed. For example, the
difference between the best and worst expected NPV of the full ensemble for the optimal
solutions obtained with Robust-2 and Robust-3 are respectively $50 × 106 (2% of the average
optimal NPV) and $30 × 106 (1.2% of the average optimal NPV). Besides being consistent,
the optimal expected NPV obtained with Robust-2 and Robust-3 are respectively 3.6% and
7.7% higher than the one obtained with Robust-1, on average of three seeds.

143
 8
5 5 5
6
10 10 10

15 15 15 4

20 20 20 2

25 25 25 0

5 10 15 5 10 15 5 10 15

(a) ln kh - Real.(b) ln kh - Real. (c) ln kh - Real. #27

#48 #8

8
5 5 5
6
10 10 10

15 15 15 4

20 20 20 2

25 25 25 0

5 10 15 5 10 15 5 10 15

(d) ln kv - Real.(e) ln kv - Real. #8 (f) ln kv - Real. #27

#48

Figure 4.22: Realizations of the log-horizontal permeability fields (upper row) and the log-
vertical permeability fields (bottom row), corresponding to P2, P50 and P100 of the cdf of
NPV obtained with the reference well placement scenario, PUNQ model. kh and kv represent
the horizontal and vertical permeability respectively.

Fig. 4.23 shows the expected NPV’s of the subset versus number of simulations using
three fixed subset realizations (realizations #48, #8 and #27) during three runs of Robust-1.
We apply the optimal solutions obtained with the subset ensemble to all Ne reservoir models
and presented the expected NPV’s of the full ensemble using dashed lines in color in Fig.
4.23. From Fig. 4.23, the optimal solution obtained with seed 2 gives an expected NPV of
$2.5 × 109 for the subset, but only $2.18 × 109 for the full ensemble which is 13% less, due
to the loss of representativeness of the subset realizations.
Fig. 4.24 shows the expected NPV’s for the full ensemble with respect to number
of simulation runs using Robust-2 where each point represents a switch of the search steps.
From Fig. 4.24, both GA and GPS search step increased the expected NPV of the full

144
 2.8
Seed 1

Expected NPV, # 109$

2.6 Seed 2
Seed 3

2.4

2.2

1.8
0 2000 4000 6000
Simulations
Figure 4.23: Expected NPV of the susbet ensemble versus number of simulation runs ob-
tained with Robust-1 using three seeds (blue curve for seed 1, red curve for seed 2 and
green curve for seed 3), robust optimization of the PUNQ model. The solid lines represent
the expected NPV’s of the subset ensemble; the dashed lines in color represent the optimal
expected NPV’s of the full ensemble; the black dashed line is the expected NPV of the full
ensemble obtained with the reference scenario.

ensemble significantly, however, the StoSAG search step only improved it slightly. For seed
2, the optimization got stuck from around 4,000 simulation runs to 9,000 simulation runs,
but then provided a sharp increase in the expected NPV of the full ensemble. Fig. 4.25 shows
both the expected NPV’s over the subset and the full ensemble during the StoSAG and GPS
search step using seed 2. From Fig. 4.25, the StoSAG step could efficiently improve the
expected NPV of the subset (the blue dots), but not the average NPV of the full ensemble
(the blue stars). This is because the StoSAG search direction is only a stochastic uphill
direction, and NSto = 5 iterations along the StoSAG search direction lead to a big change in
the well locations with which the selected subset realizations are no longer representative,
even though the improvement in the expected NPV’s of the subset seems to be less than that
in the GPS step. Hence, optimizing over the subset realizations using StoSAG only slightly
improves the expected NPV of the full ensemble.
Fig. 4.26 shows the expected NPV’s for the full ensemble with respect to number of
simulation runs using Robust-3 where it is less frequent for the expected NPV of the full
ensemble to get stuck. This is because that Robust-3 procedure not only constructs the cdf’s

145
Expected NPV, # 109 $ 2.6 2.6 2.6

Expected NPV, # 109 $

Expected NPV, # 109 $

2.4 2.4 2.4

2.2 2.2 2.2

2 2 2

1.8 1.8 1.8

0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Simulations Simulations Simulations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.24: Expected NPV of the full ensemble versus number of simulations obtained for
Robust-2 using three seeds, robust optimization of PUNQ model. o represents an update by
GA, ∗ represents an update by StoSAG, and  represents an update by GPS.

2.6
Expected NPV, # 109 $

2.4

2.2

1.8
0 2000 4000 6000 8000 10000
Simulations
Figure 4.25: Expected NPV versus number of simulations obtained with Robust-2 using seed
2, robust optimization of PUNQ model. o, ∗, and  respectively represent the expected NPV
of the full ensemble obtained using GA, StoSAG and GPS; . and . respectively represent the
expected NPV of the subset ensemble obtained using StoSAG and GPS.

of NPV at each switch of the search steps, but also when the ordering of NPV’s of a set of
reference realizations changes. This tends to prevent large updates to the well status, types
and locations between two neighboring cdf constructions, which may cause the representative
subset to become non-representative. It is worth to mention that Robust-3 procedure starts
with a subset containing only one realization, however, the size of the subset is increased
to 3 after the first generation of GA, since as we have observed, the representativeness of
realizations for the PUNQ model heavily depends on the optimization variables.

146
Expected NPV, # 109$ 2.6 2.6 2.6

Expected NPV, # 109$

Expected NPV, # 109$

2.4 2.4 2.4

2.2 2.2 2.2

2 2 2

1.8 1.8 1.8

0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000 0 2000 4000 6000 8000 10000
Simulations Simulations Simulations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.26: Expected NPV of the full ensemble versus number of simulations obtained for
Robust-3 using three seeds, robust optimization of PUNQ model. o represents an update by
GA, ∗ represents an update by StoSAG, and  represents an update by GPS.

Figs. 4.27, 4.28 and 4.29 respectively show the optimal wells drilled, their types and
locations obtained with Robust-1, Robust-2 and Robust-3 where each optimization procedure
is repeated three times using different seeds. Due to the existence of a strong aquifer which
provides pressure support for producers, the optimal solutions tend to avoid drilling any
injectors except that one injector is drilled in the optimal solution of Robust-3 using seed 1.
Meanwhile, the number of producers obtained using three seeds ranges from 7 to 10, which
are quite consistent with each other.
The expected NPV’s of the full ensemble and the subset ensemble obtained with
optimal solutions, the optimal number of producers along with the simulation costs to obtain
convergence are summarized in Table 4.7. Note that the maximum number of simulation
runs is set to be 10,000 and seed 2 of Robust-2 didn’t reach convergence before the algorithm
termination. From Table 4.7, based on the average performance of three stochastic runs, (i)
Robust-1 costs the least number of simulation runs, but provides the lowest expected NPV’s
of the full ensemble; (ii) Robust-2 gives higher expected NPV’s of the full ensemble compared
to Robust-1, but costs the highest number of simulation runs; (iii) Robust-3 requires 1,200
simulation runs less than Robust-2, but 500 simulation runs more than Robust-1, however,
it provides the highest expected NPV’s of the full ensemble; (iv) the expected NPV’s of the
subset ensemble are generally higher than those of the full ensemble at the optimal solutions

147
since the former is the objective that we actually maximize, and the difference between the
former and the latter is much larger for Robust-1 compared to Robust-2 and Robust-3.

(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.27: Optimal types and locations of drilled wells obtained with Robust-1 using three
seeds, robust optimization of the PUNQ model. Injector locations are marked in brown
color and labeled with ⊗; producer locations are marked in yellow color and labeled with o;
the gas cap is marked in green color; the aquifer is marked in dark blue and the oil zone is
marked in light blue color.

(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.28: Optimal types and locations of drilled wells obtained for Robust-2 using three
seeds, robust optimization of PUNQ model. Injector locations are marked in brown color
and labeled with ⊗; producer locations are marked in yellow color and labeled with o; the
gas cap is marked in green color; the aquifer is marked in dark blue and the oil zone is
marked in light blue color.

148
 (a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.29: Optimal types and locations of drilled wells obtained for Robust-3 using three
seeds, robust optimization of PUNQ model. Injector locations are marked in brown color
and labeled with ⊗; producer locations are marked in yellow color and labeled with o; the
gas cap is marked in green color; the aquifer is marked in dark blue and the oil zone is
marked in light blue color.

Table 4.7: Summary of the optimal solutions and computational cost obtained with the three
robust optimization procedures using three seeds, robust optimization of PUNQ model.

Robust-1 Robust-2 Robust-3

Sim. $ × 109 Sim. $ × 109 Sim. $ × 109
Seeds Ω NP ro Ω NP ro Ω NP ro
# JE f JEΩs # JE f JEΩs # JE f JEΩs
Seed 1 5950 2.43 2.56 8 4465 2.42 2.51 9 4031 2.55 2.60 9
Seed 2 5500 2.18 2.50 7 10000 2.46 2.54 8 7443 2.53 2.58 10
Seed 3 5500 2.43 2.63 9 7547 2.41 2.46 8 6984 2.52 2.60 10
Average 5650 2.35 2.56 8 7337 2.43 2.50 8 6153 2.53 2.59 10

2. Robust optimization using the iterative sequential algorithm with GA and GPS
Due to the observation that the StoSAG optimization for the subset realizations
after NSto = 5 iterations only slightly improves the expected NPV of the full ensemble, we
investigate Robust-1, Robust-2 and Robust-3 optimization procedures using only the GA
search step and the GPS search step. Figs. 4.30 and 4.31 show the expected NPV of the full
ensemble versus number of simulations obtained with Robust-2 and Robust-3 respectively.
The convergence speed of Figs. 4.30 and 4.31 has been greatly improved compared to Figs.
4.24 and 4.26, i.e., saving over 1,300 simulation runs per seed. The expected NPV’s of

149
the full ensemble and the subset ensemble obtained with the optimal solutions, the optimal
number of producers along with the simulation cost to reach convergence are summarized
in Table 4.8. From Table 4.8, based on the average performance of three stochastic runs,
(i) Robust-3 remains giving highest expected NPV’s of the full ensemble while Robust-1
still gives the worst expected NPV for Ne = 50 realizations; (ii) the expected NPV’s of the
subset ensemble are higher than those of the full ensemble at the optimal solutions and the
difference in between is much larger for Robust-1 compared to Robust-2 and Robust-3; (iii)
the optimal number of producers are consistent using different optimization procedures and
different seeds, i.e., 7 to 10 producers are drilled. Note almost no injectors are drilled, though
we do not show it.

2.6 2.6 2.6

Expected NPV, # 109 $

Expected NPV, # 109 $

Expected NPV, # 109 $

2.4 2.4 2.4

2.2 2.2 2.2

2 2 2

1.8 1.8 1.8

0 2000 4000 6000 8000 0 2000 4000 6000 8000 0 2000 4000 6000 8000
Simulations Simulations Simulations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.30: Expected NPV of the full ensemble versus number of simulations obtained for
Robust-2 using the iterative sequential method with GA and GPS steps, robust optimization
of the PUNQ model. o’s and ’s respectively represent updates by GA and GPS.

Table 4.8: Summary of the optimal solutions and computational cost obtained with the three
robust optimization procedures with GA and GPS, robust optimization of PUNQ model.

Robust-1 Robust-2 Robust-3

Sim. $ × 109 Sim. $ × 109 Sim. $ × 109
Seeds Ω NP ro Ω NP ro Ω NP ro
# JE f JEΩs # JE f JEΩs # JE f JEΩs
Seed 1 6100 2.39 2.59 9 6055 2.52 2.60 9 4905 2.54 2.59 10
Seed 2 3100 2.36 2.57 7 5915 2.40 2.46 9 4912 2.44 2.50 8
Seed 3 6400 2.44 2.59 8 5135 2.52 2.60 9 4749 2.55 2.60 9
Average 5200 2.40 2.58 8 5700 2.48 2.55 9 4855 2.51 2.56 9

150
Expected NPV, # 109 $ 2.6 2.6 2.6

Expected NPV, # 109 $

Expected NPV, # 109 $

2.4 2.4 2.4

2.2 2.2 2.2

2 2 2

1.8 1.8 1.8

0 2000 4000 6000 8000 0 2000 4000 6000 8000 0 2000 4000 6000 8000
Simulations Simulations Simulations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.31: Expected NPV of the full-ensemble versus number of simulations obtained for
Robust-3 using the iterative sequential method with GA and GPS steps, robust optimization
of PUNQ model. o’s and ’s respectively represent updates by GA and GPS.

4.4.2 Example 2: Brugge model

In the original Brugge model, 104 geological realizations of the horizontal perme-
ability, vertical permeability, porosity, connate water saturation and the net-gross ratio are
generated to describe the reservoir uncertainty. However, due to our limited computational
resources, we only use 50 realizations where 10 realizations are generated using each of the
5 geo-statistical methods (see [71] for more details). The same reference well placement
scenario, well control settings, economic parameters and minimum well spacing are used as
in deterministic optimization presented in Section 4.2.2. The NPV’s for all 50 realizations
are calculated using the reference scenario given in Fig. 4.15(a) where the cdf of NPV is
shown as the black curve in Fig. 4.32. The P2, P50 and P100 realizations correspond to
realization #32, #14 and #9 respectively, whose petrophysical parameters for layer 1 are
shown in Fig. 4.33. From Fig. 4.33, the petrophysical parameters of these three realizations
are quite different, indicating the large geological uncertainty in the reservoir models.
1. Robust optimization using the iterative sequential method with GA, StoSAG and
GPS
Since each simulation run for the Brugge model costs approximately 2 minutes and
it has been observed for the PUNQ model that the inclusion of StoSAG search step in the
iterative sequential method may lead to a slow convergence speed, only one run is tried for

151
each of the three robust optimization procedures using GA, StoSAG and GPS search steps.
Fig. 4.32 shows the cdf’s of NPV obtained with the reference scenario (given in Fig. 4.15(a))
and the optimal well placement for Robust-1, Robust-2 and Robust-3 respectively. The cdf’s
of NPV are generated based on evaluating the NPV of all Ne = 50 reservoir models given an
estimate of (s, T , x). The green, blue and red circles respectively represent the P2, P50 and
P100 realizations of the cdf of NPV obtained with the reference scenario (the black curve).
From Fig. 4.32, the subset models selected from the initial cdf still roughly correspond to
the 2th, 50th and 100th percentiles of the cdf’s of NPV obtained with the optimal solutions,
indicating that the representative realizations are relatively independent of the optimization
variables (well status, types and locations) for the Brugge model, and that it is unnecessary
to frequently re-select the representative realizations.

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

cdf

cdf

cdf
0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
6 8 10 12 6 8 10 12 6 8 10 12
NPV, # 109 $ NPV, # 109 $ NPV, # 109 $
(a) Robust-1 (b) Robust-2 (c) Robust-3

Figure 4.32: Cdf’s of NPV’s obtained with the reference scenario (black curve) and the
optimal well placement for three robust optimization procedures, robust optimization of
Brugge model. The green, blue and red circles, respectively, represent the P2, P50 and P100
of the cdf of NPV obtained with the reference scenario.

With Robust-1 procedure where the subset ensemble (containing 3 realizations) are
fixed during the optimization process, the optimal solution gives an expected NPV of $10.77×
109 for the subset and $10.47 × 109 for the full ensemble which is 3% less; see Fig. 4.34.
With Robust-2 procedure where the subset ensemble (containing 3 realizations) is re-selected
at each switch of the algorithms, i.e., from GA to StoSAG/GPS or vice versa, the optimal
solution gives an expected NPV of $10.10 × 109 for the full ensemble, which is far less than
the value obtained with Robust-1; see Fig. 4.35. This is because the StoSAG step converges

152
 0.3 1
0.5
8 8
0.25
0.8
0.4
6 6
0.2
0.6 0.3
4 4
0.15

0.4 0.2
2 2
0.1

0 0 0.2 0.1
0.05

-2 -2 0 0 0

con
(a) ln kh - Real. #32 (b) ln kv - Real. #32 (c) φ - Real. #32 (d) NTG - Real. #32 (e) Sw - Real. #32

0.3 1
0.5
8 8
0.25
0.8
0.4
6 6
0.2
0.6 0.3
4 4
0.15

0.4 0.2
2 2
0.1

0 0 0.2 0.1
0.05

-2 -2 0 0 0

con
(f) ln kh - Real. #14 (g) ln kv - Real. #14 (h) φ - Real. #14 (i) NTG - Real. #14 (j) Sw - Real. #14

0.3 1
0.5
8 8
0.25
0.8
0.4
6 6
0.2
0.6 0.3
4 4
0.15

0.4 0.2
2 2
0.1

0 0 0.2 0.1
0.05

-2 -2 0 0 0

con
(k) ln kh - Real. #9 (l) ln kv - Real. #9 (m) φ - Real. #9 (n) NTG - Real. #9 (o) Sw - Real. #9

Figure 4.33: Petrophysical parameters for the top layer of the realizations corresponding to
P2, P50 and P100 of the cdf of NPV obtained with the reference well placement scenario,
Brugge model. kh and kv represents the horizontal and vertical permeability respectively, φ
is the porosity, N T G is the net to gross ratio and Swcon is the connate water saturation.

153
slowly when it comes near a local solution, but the slight improvements of the expected NPV’s
of the subset ensemble and the full ensemble are sufficient to avoid triggering the convergence
criteria for the StoSAG search step until about 8,000 simulations and hence delayed the GPS
fine-tune step for Robust-2. Given the maximum number of simulation runs as 8,000 for
robust optimization, the inclusion of StoSAG search steps seems disadvantageous. In fact
the resulting slow convergence could happen to all three robust optimization procedures and
Robust-2 happens to be an unlucky run. With Robust-3 procedure where the subset ensemble
(starting with ensemble size 1) is re-selected following Algorithm 11, the optimal solution
gives the highest expected NPV of the full ensemble, $10.67×109 , and the fastest convergence
speed; see Fig. 4.36. For Robust-3, the subset ensemble only contains one realization in the
early iterations and hence can save simulation costs and search more thoroughly within the
specified number of simulation runs.
The optimal expected NPV’s for the full ensemble and subset ensemble, the optimal
number of injectors and producers along with the simulation costs to achieve convergence are
summarized in Table 4.9. Note we terminate the optimization process when the number of
simulation runs exceeds 8,000 even if it hasn’t converged yet. From Table 4.9, (i) Robust-3
requires less simulation runs compared to Robust-1 and Robust-2, and provides the highest
expected NPV’s for the full ensemble; (ii) the expected NPV’s of the subset ensemble are
generally higher than those of the full ensemble at the optimal solutions since the former is
the objective that we actually maximize, and the difference between the former and the latter
is quite similar for all three optimization procedures which is consistent with Fig. 4.32; and
(iii) the optimal number of wells are similar using different optimization procedures, i.e., 23
to 25 wells are drilled. Figs. 4.37, 4.38 and 4.39 respectively show the optimal wells drilled,
their types and locations obtained with the three optimization procedures, along with the
oil saturation of the top layer after 20 years of production obtained with realization #14.
2. Robust optimization using the iterative sequential method with GA and GPS
Due to the observation that the inclusion of StoSAG optimization may lead to a
slow convergence speed, we investigate the Robust-1, Robust-2 and Robust-3 optimization

154
Table 4.9: Summary of the optimal solutions and computational cost obtained with three
robust optimization procedures, robust optimization of the Brugge model.
Robust-1 Robust-2 Robust-3
Sim. $ × 109 Sim. $ × 109 Sim. $ × 109
Seeds Ω NInj NP ro Ω NInj NP ro Ωf NInj NP ro
# JE f JEΩs # JE f JEΩs # JE JEΩs
Seed 1 7798 10.47 10.77 8 15 8000 10.10 10.20 14 11 7159 10.67 10.88 15 8

11
Expected NPV, # 109 $
10

7
0 2000 4000 6000 8000
Simulations
Figure 4.34: Expected NPV of the subset ensemble obtained with the best candidate so-
lution of a population versus number of simulation runs obtained with Robust-1, robust
optimization of Brugge model. The solid line in blue represents the expected NPV’s of the
subset ensemble, the blue dashed line represents optimal expected NPV of the full ensemble,
and the black dashed line represents the expected NPV of the full ensemble obtained with
the reference scenario.

procedures using the iterative sequential method with GA and GPS search steps only. Three
runs are tried for each of the robust optimization procedures. Fig. 4.40 shows the cdf of
NPV for the 50 realizations obtained with the reference scenario given in Fig. 4.15(a)
and the optimal solutions for the three optimization procedures. Similar to Fig. 4.32, the
subset models selected from the initial cdf still roughly correspond to the 2th, 50th and
100th percentiles of the cdf’s of NPV obtained with the optimal solutions, indicating that
the representative realizations are relatively independent of the optimization variables (well
status, types and locations) for the Brugge model. This indicates that the optimal expected
NPV’s of the full ensemble obtained with Robust-1, Robust-2 and Robust-3, should be more
or less the same, since it is not necessary to re-select the representative subset ensemble
frequently.

155
 11

Expected NPV, # 109 $

10

7
0 2000 4000 6000 8000
Simulations
Figure 4.35: Expected NPV versus number of simulations obtained with Robust-2, robust
optimization of the Brugge model. o, ∗ respectively represent the expected NPV of the
full ensemble obtained using GA and StoSAG; . represents the expected NPV of the subset
ensemble obtained using StoSAG.

11
Expected NPV, # 109 $

10

7
0 2000 4000 6000 8000
Simulations
Figure 4.36: Expected NPV of the full ensemble versus number of simulations obtained
for Robust-3, robust optimization of the Brugge model. o represents an update by GA, ∗
represents an update by StoSAG, and  represents an update by GPS.

The expected NPV’s of the full ensemble and the subset ensemble obtained with
optimal solutions, the optimal number of injectors and producers along with the simulation
costs to obtain convergence are summarized in Table 4.10. Note that the maximum number
of simulation runs is set to be 8,000 and seed 1 of Robust-1 didn’t reach convergence before
the algorithm termination. From Table 4.10, based on the average performance of three
stochastic runs, (i) Robust-1 costs the least number of simulation runs because the runs

156
 0.8 0.8

0.7 0.7

0.6 0.6
+ o + o
o o
o 0.5 o 0.5
+ o o + o o
o o
o + o +
o o 0.4 o o 0.4
++ + ++ +
+ o + o
o 0.3 o 0.3
o o
o o
o o
o o
0.2 0.2
+ +

0.1 0.1

0 0

(a) Layer 1 (b) Layer 9

Figure 4.37: Optimal types and locations of drilled wells obtained for Robust-1, plotted on
the oil saturation field at the top layer of realization #14 after 20 years of production, robust
optimization of the Brugge model. ⊗ represents an injector and o represents a producer.

0.8 0.8

0.7 0.7

+
0.6 +
0.6
o o
oo oo
0.5 0.5
+ o + o
+ o + o
o + o +
o o o
0.4 o o o
0.4
+ +
+ +
o 0.3 o 0.3
o o o o
+ +
o o
o o
0.2 0.2

0.1 0.1

0 0

(a) Layer 1 (b) Layer 9

Figure 4.38: Optimal types and locations of drilled wells obtained for Robust-2, plotted on
the oil saturation field at the top layer of realization #14 after 20 years of production, robust
optimization of the Brugge model. ⊗ represents an injector and o represents a producer.

with seed 2 and seed 3 converged at local solutions which are far from the global solution;
(ii) as expected, Robust-2 and Robust-3 give almost the same optimal expected NPV’s of
the full ensemble which is higher than the value of Robust-1, but Robust-3 requires 700

157
 0.8 0.8

0.7 0.7

+ 0.6 + 0.6
o o
+ o + o
+ oo 0.5 + oo 0.5
+ o + o
o o
o o
+ + 0.4 + + 0.4
o o o o
o o
+ +
o 0.3 o 0.3
o o
o o
o o
o 0.2 o 0.2
+ +

0.1 0.1

0 0

(a) Layer 1 (b) Layer 9

Figure 4.39: Optimal types and locations of drilled wells obtained for Robust-3, plotted on
the oil saturation field at the top layer of realization #14 after 20 years of production, robust
optimization of the Brugge model. ⊗ represents an injector and o represents a producer.

1 1 1

0.8 0.8 0.8

0.6 0.6 0.6

cdf

cdf

cdf

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
6 8 10 12 6 8 10 12 6 8 10 12
NPV, # 109 $ NPV, # 109 $ NPV, # 109 $
(a) Robust-1 (b) Robust-2 (c) Robust-3

Figure 4.40: Cdf’s of NPV’s obtained with the reference scenario (black curve) and optimal
well placement for the three robust optimization procedures with GA and GPS search steps,
robust optimization of Brugge model. The green, blue and red circles respectively represent
the P2, P50 and P100 of the cdf of NPV obtained with the reference scenario.

simulation runs less than Robust-2; (iii) the expected NPV’s of the subset ensemble are
generally higher than those of the full ensemble at the optimal solutions since the former is
the objective that we actually maximize, and the difference between the former and the latter
is almost the same for Robust-1, Robust-2 and Robust-3; (iv) the number of wells obtained
with the optimal solutions whose expected NPV’s are larger than $10.50 × 109 , are quite
similar, i.e., 8 to 9 injectors and 13 to 19 producers. Figs. 4.41 and 4.42 respectively show

158
the expected NPV of the full ensemble versus number of simulations obtained with Robust-2
and Robust-3. Compared to Figs. 4.35 and 4.36, the convergence speed of Robust-2 and
Robust-3 without the StoSAG search step have been greatly improved, i.e., saving about
1,000 simulations per run.
Table 4.10: Summary of the optimal solutions and computational cost obtained with the
three robust optimization procedures using the iterative sequential method with GA and
GPS search steps, robust optimization of Brugge model.
Robust-1 Robust-2 Robust-3
Sim. $ × 109 Sim. $ × 109 Sim. $ × 109
Seeds Ω NInj NP ro Ω NInj NP ro Ωf NInj NP ro
# JE f JEΩs # JE f JEΩs # JE JEΩs
Seed 1 8000 10.55 10.80 9 13 5894 10.66 10.83 9 19 5054 10.68 10.87 7 14
Seed 2 5345 10.14 10.36 9 11 7795 10.49 10.76 11 14 6247 10.63 10.82 9 17
Seed 3 4796 10.31 10.52 7 13 7429 10.66 10.89 7 18 7227 10.50 10.63 10 12
Average 6047 10.33 10.56 8 12 6908 10.60 10.79 9 17 6176 10.60 10.77 9 14

11 11 11
Expected NPV, # 109 $

Expected NPV, # 109 $

Expected NPV, # 109 $

10 10 10

9 9 9

8 8 8

7 7 7
0 2000 4000 6000 8000 0 2000 4000 6000 8000 0 2000 4000 6000 8000
Simulations Simulations Simulations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.41: The expected NPV of the full ensemble versus number of simulation runs
obtained for Robust-2 using the iterative sequential method with GA and GPS steps, robust
optimization of Brugge model. o’s and ’s respectively represent updates by GA and GPS.

4.5 Summary and Discussion

In this chapter, we considered maximizing the NPV by optimizing the well status,
types, locations given a maximum number of wells authorized to drill. The optimization
problem considers both deterministic reservoir models and stochastic reservoir models whose
uncertainty is represented by a large number of realizations of the reservoir properties. For
deterministic optimization, three algorithms are designed to combine the global search abil-
ity of GA with the local search ability of StoSAG and GPS methods, including an iterative

159
Expected NPV, # 109 $ 11 11 11

Expected NPV, # 109 $

Expected NPV, # 109 $

10 10 10

9 9 9

8 8 8

7 7 7
0 2000 4000 6000 8000 0 2000 4000 6000 8000 0 2000 4000 6000 8000
Simulations Simulations Simulations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 4.42: The expected NPV of the full ensemble versus number of simulation runs
obtained for Robust-3 using the iterative sequential method with GA and GPS steps, robust
optimization of Brugge model. o’s and ’s respectively represent updates by GA and GPS.

sequential procedure and two simultaneous procedures followed by StoSAG and GPS search
steps, one using mixed-encoded and one using the Genocop III method. Numerical results
on the PUNQ model and the Brugge model show that both the iterative sequential method
and the mixed-encoded GA outperform the Genocop III method and that the iterative se-
quential method has the fastest convergence speed. For the optimization under geological
uncertainty (also referred to as robust optimization), we developed an efficient subset real-
ization selection procedure and coupled it with the iterative sequential method. Numerical
results on the PUNQ model and the Brugge model show that with the adaptive procedure,
we can achieve robust optimization within a few thousand simulation runs, given a large
ensemble to characterize the geological uncertainty. Besides, we find it desirable to exclude
StoSAG from the iterative sequential method for robust optimization.

160
 CHAPTER 5

WELL PLACEMENT OPTIMIZATION WITH COMPLICATED WELL

TRAJECTORIES

In this chapter, we consider well placement optimization of complicated well trajec-

tories, e.g. slanted wells and multi-segmented wells, given the well number, their types and
controls. Corresponding parameterizations for the well trajectories are firstly proposed and
applied to deterministic optimization using a hybrid of StoSAG and GPS. Afterwards, bi-
objective well placement optimization under geological uncertainty is investigated where the
two objectives are to maximize the expected value of the life-cycle net-present-value (NPV)
and to maximize the minimum NPV of the set of realizations representing the geological
uncertainty, using an efficient implementation of the lexicographic method.

5.1 Well Trajectory Parameterization

5.1.1 Slanted well trajectory parameterization

Here, a slanted well trajectory refers to the case where the center-line of a well is
a straight line in the 3D space. Emerick et al. [21] proposed to use the (i, j, k) gridblock
indices of the heel point and toe point of a well to characterize the trajectory of a slanted
well . We propose to use the (x, y, z) coordinates of the two end points of a well inside the
reservoir, which is a slightly more general parameterization. We consider a procedure to
obtain the path of well iw given the vector

Xiw = [xiw iw iw iw iw iw
0 , y0 , z0 , x1 , y1 , z1 ], (5.1)

where, (xiw iw iw iw iw iw
0 , y0 , z0 ) and (x1 , y1 , z1 ) are the coordinates of the two end points of well

161
iw; see Fig. 5.1(a). The entry point (xiw iw iw iw iw iw
0 , y0 , z0 ) and the exit point (x1 , y1 , z1 ) are

required to fall in the active gridblocks, Ωactive . Two types of the most commonly considered
constraints are the minimum well spacing constraints where the distance between any pair
of well trajectories, d(Xiw , Xjw ) should be greater than a pre-specified value Rmin and the
maximum/minimum well length constraints, these two constraints are given by

d(Xiw , Xjw ) > Rmin iw, jw = 1, 2, · · · , Nw , iw 6= jw, (5.2a)

low up
liw ≤ liw ≤ liw iw = 1, 2, · · · , Nw , (5.2b)

where, d(Xiw , Xjw ) is the Euclidean distance between the iwth and jwth wells (see Appendix
A for the detailed calculation), Rmin is the minimum allowable inter-well distance, liw is the
p
length of the iwth well and is calculated as liw = (xiw iw 2 iw iw 2 iw
1 − x0 ) + (y1 − y0 ) + (z1 − z0 ) ,
iw 2

low up
liw and liw correspond to the minimum and maximum allowable well lengths for the iwth
Nw (Nw −1)
well respectively, and Nw is the number of wells. There are 2
well spacing constraints
and 2Nw well length constraints in total. For the parameterization given in Eq. 5.1, both
constraints in Eq. 5.2 are nonlinear constraints on the design variables, i.e., the evaluations
of d(Xiw , Xjw ) and liw do not require simulation runs.
In order to evaluate the objective function, we simulate the reservoir dynamic perfor-
mance using the commercial simulator, CMG, where deviated perforations can be specified
by the (x, y, z) coordinates of their entry and exit points running through each perforated
gridblock. With this information, CMG is able to calculate an accurate deviated well pro-
ductivity/injectivity index using a Peaceman type well model ([68, 69, 70]). Thus, a routine
which identifies the grid cells crossed by each well and calculates the Cartesian coordinates
of the entry point and exit point of each perforation of a well in its perforated cell is imple-
mented, see Appendix C. In this dissertation, this routine is only developed for reservoirs
with (transformed) Cartesian grids.
Denote the full set of the variable vector for vertical well placement optimization as
x = [X1 , X2 , · · · , XNw ], where the dimension of x is Nx = 6Nw . The net present value

162
(NPV) of production is defined as

 hP 
PNt ∆tn NP
i
ron ·q no,j −cnw ·q nw,j −cng ·q ng,j − N − N

P I n n n n
P w
J(m, x) = n=1 tn j=1 k=1 c ·q
wi wi,k +c ·q
gi gi,k iw=1 Cw (Xiw ),
(1+b) 365

where n, Nt , ∆tn , NP , NI , ron , cnw , cng , cnwi , cngi , q no,j , q nw,j , q ng,j , q nwi,k , q ngi,k denote the same
terms as in Section 2.1, and Cw (Xiw ) denotes the drilling cost for the iwth well which
depends on the trajectory variables defining this well, Xiw . Given a vector of the reservoir
model parameter, m, the trajectory optimization problem of vertical wells considered in this
chapter is stated as follows,

max J(m, x) (5.3a)

x∈RNx

s.t. (xiw iw iw
0 , y0 , z0 ) ∈ Ωactive , iw = 1, 2, · · · , Nw , (5.3b)

(xiw iw iw
1 , y1 , z1 ) ∈ Ωactive , iw = 1, 2, · · · , Nw , (5.3c)

d(Xiw , Xjw ) > Rmin iw, jw = 1, 2, · · · , Nw , iw 6= jw, (5.3d)

low
liw ≤ liw iw = 1, 2, · · · , Nw , (5.3e)
up
liw ≤ liw iw = 1, 2, · · · , Nw . (5.3f)

5.1.2 Multi-segmented well trajectory parameterization

For a multi-segmented well, using the (x, y, z) coordinates of the end points of each
segment as design variables may destroy the smoothness of the well trajectory after the
variable updates in the optimization process. Hence, we propose a different parameterization
(see Fig. 5.1(b) and Eq. 5.4 below) to transform the drillability (smoothness) constraint to
the bound constraints on optimization variables which then can be effectively enforced by
truncation. Specifically, the vector of parameters for well iw is given by,

Xiw = [xiw iw iw iw iw iw iw iw iw iw iw iw
0 , y0 , z0 , z1 , l1 , θ1 , z2 , l2 , δθ2 , · · · , zNseg , lNseg , δθNseg ], (5.4)

where, (xiw iw iw
0 , y0 , z0 ) are the coordinates of the entry point of the iwth well into the reservoir

163
((xiw iw iw
0 , y0 , z0 ) is required to fall into an active gridblock of the reservoir), see Fig. 5.1(b);

θ1iw is the azimuth angle of the first segment (the angle between the x-axis and the projection
of the first segment onto the x-y plane); δθiiw , i = 2, 3, · · · , Nseg represents the change in
azimuth angle of the ith segment compared to the (i − 1)th segment so the azimuth angle
iw
of each segment can be calculated recursively using θiiw = θi−1 + δθiiw , i = 2, 3, · · · , Nseg ;
ziiw , i = 1, · · · , Nseg represents the z coordinate of the end point of the ith segment of
the iwth well; liiw , i = 1, · · · , Nseg represents the length of the ith segment. The (x, y)
coordinates of the end point of the ith segment of the iwth well can be calculated as xiw i =
q q
2 iw 2 2 iw 2



xiw
i−1 + (liiw ) − ziiw − zi−1 ∗ cos(θiiw ) and yiiw = yi−1
iw
+ (liiw ) − ziiw − zi−1 ∗ sin(θiiw ).
Similar to the azimuth angle concept, we define ϕiw
1 as the inclination angle of the first

segment (the angle between z-axis and the first segment of the well trajectory), and define
δϕiw
i , i = 2, 3, · · · , Nseg as the change in inclination angle of the ith segment compared to

the (i − 1)th segment. Note that ϕiw iw

1 and δϕi are not in the well trajectory parameter Xiw .

For a multi-segmented well, there may be two segments of perforations in the gridblock
containing the end point of a well segment (except the last well segment). Since in CMG as
well as other commercial simulators, one can only define a perforation with one line segment,
thus a slanted pseudo-perforation is required. We propose to adjust the completion factor
of the pseudo-perforation based on CMG [55] so that equal injectivity/productivity can be
obtained. The calculation of the coordinates of the entry and exit point and the completion
factor of this pseudo-perforation is presented in Appendix C.2.
In fact, the multi-segmented well trajectory depicted in Fig. 5.1(b) is not drillable
using current directional drilling techniques, and a good-quality wellbore should be smooth
in practice. Hence, when applying the optimal multi-segmented well trajectory, we smooth
the ith and the (i + 1)th well segments as a circular arc located in the plane of these two
segments. Fig. 5.2 illustrates the smoothing of two consecutive well segments for a multi-
segmented well where points A and C denote the two end points of the ith well segment,
points C and E denote the two end points of the (i + 1)th segment and points B and D
denote the points on segments AC and CE such that |BC| = |CD| = min( |AC|
2
, |CE|
2
) and O

164
 (a) Slanted well

(b) Multi-segmented well

Figure 5.1: Illustration of well trajectory parameterization for a slanted and a multi-
segmented well.

is the center of the arc BD.

d The calculation of the segment end points and the smoothing

of the trajectory of well iw are given in Appendix B.

To be drillable, the dogleg severity (DLS) of the well trajectory in Fig. 5.2 has to
stay below a maximum value. DLS ([58]) is the dogleg angle (defined as the overall change
in inclination and azimuth of a borehole from the ith segment to the (i + 1)th segment) per
100 feet of the measured depth. Following the notation in Fig. 5.2, DLS is given by

100βiiw 100 180

DLSiiw = = , (5.5)
BD
d OB π

165
Figure 5.2: Illustration of the smoothed trajectory of two consecutive well segments for a
multi-segmented well.

where,
BC
OB = βiiw
, (5.6)
tan 2

and s 2 iw 2
δϕiw
 
i δθi
βiiw = 2arcsin sin + sinϕiw iw iw
i sin(ϕi + δϕi ) sin . (5.7)
2 2

Substituting Eq. 5.6 into Eq. 5.5, we obtain

100 180 β iw
DLSiiw = tan( i ). (5.8)
BC π 2

iw
liiw li+1
From Eq. 5.8, the dogleg severity is low if βiiw is low or BC (i.e., min( 2
, 2 )) is large.
Since the thickness of a reservoir is usually much smaller than its width or length and the
perforated length of a directional well is usually quite large, the change in inclination δϕiw
i

is usually kept at a small value if a well trajectory is bounded within the a reservoir (or a
specific target zone). To avoid handling the nonlinear constraints that the dogleg severity
has to be below a maximum value, we put bound constraints on the changes in the azimuth
angle δθiiw and the segment length liiw . The lower and upper bounds of δθiiw and liiw are

166
denoted as δθilow,iw and δθiup,iw , and, llow,iw and lup,iw , respectively. Thus the total set of
bound constraints on optimization variables are given by

(xiw iw iw
0 , y0 , z0 ) ∈ Ωactive , iw = 1, 2, · · · , Nw ,

zilow,iw < ziiw < ziup,iw , iw = 1, 2, · · · , Nw , i = 1, 2, · · · , Nseg ,

lilow,iw < liiw < liup,iw , iw = 1, 2, · · · , Nw , i = 1, 2, · · · , Nseg , (5.9)

θ1low,iw < θ1iw < θ1up,iw , iw = 1, 2, · · · , Nw ,

δθilow,iw < δθiiw < δθiup,iw , iw = 2, · · · , Nw , i = 2, 3, · · · , Nseg .

Note that even though the bound constraints in Eq. 5.9 can be efficiently enforced by
truncation, they only roughly enforce the drillability constraint. If the nonlinear constraints
given in Eq. 5.2 are considered in the multi-segmented well optimization, then d(Xiw , Xjw )
is defined as the minimum distance between any pair of well segments of the iwth and jwth
wells (iw 6= jw).
Denote the full set of the variable vector for multi-segmented well placement as x =
[X1 , X2 , · · · , XNw ], where the dimension of x is Nx = N iw
P w
iw=1 3(Nseg + 1). Given a vector

of reservoir model parameter, m, the well trajectory optimization of multi-segmented wells

considered in this chapter is stated as follows,

max J(m, x) (5.10a)

x∈RNx

s.t. (xiw iw iw
0 , y0 , z0 ) ∈ Ωactive , iw = 1, 2, · · · , Nw , (5.10b)

zilow,iw < ziiw < ziup,iw , iw = 1, 2, · · · , Nw , i = 1, 2, · · · , Nseg , (5.10c)

lilow,iw < liiw < liup,iw , iw = 1, 2, · · · , Nw , i = 1, 2, · · · , Nseg , (5.10d)

θ1low,iw < θ1iw < θ1up,iw , iw = 1, 2, · · · , Nw , (5.10e)

δθilow,iw < δθiiw < δθiup,iw , iw = 2, · · · , Nw , i = 2, 3, · · · , Nseg (5.10f)

d(Xiw , Xjw ) > Rmin iw, jw = 1, 2, · · · , Nw , iw 6= jw. (5.10g)

167
 5.2 Constrained Optimization using Augmented Lagrangian Method

The problems described in Eqs. 5.3 and 5.10 are maximization problems under both
bound constraints and nonlinear constraints on the design variables. For the vertical well
placement optimization, Eqs. 5.3b to 5.3c are bound constraints and Eqs. 5.3d to 5.3f are
nonlinear constraints. For the multi-segmented well trajectory optimization, Eqs. 5.10b
to 5.10f are bound constraints and Eq. 5.10g is a nonlinear constraint. Denote the upper
and lower bounds of x respectively as xup and xlow , then the bound constraints can be
re-written in a general form that xlow
i ≤ xi ≤ xup low up
i , i = 1, 2, · · · , Nx where xi , xi and xi ,

respectively denote the ith component of the vectors xlow , x and xup . In this chapter, the
bound constraints are enforced by truncating the design variables beyond the bounds to their
nearest bounds while the nonlinear constraints are handled using the augmented Lagrangian
method (ALM).
As the nonlinear constraints in Eqs. 5.3d, 5.3e, 5.3f and 5.10g may be of different
scales, and C. Chen [10] observed that the convergence rate of ALM can be slowed appreciably
by poor scaling of the constraints based on numerical examples, we scale these nonlinear
constraints so that the scaled constraints are more or less on the same magnitude. The
scaled nonlinear constraints in Eqs. 5.3f and 5.10g are given below,

Rmin − d(Xiw , Xjw )

c (x) = ≤ 0, iw, jw = 1, 2, · · · , Nw , iw 6= jw. (5.11)
Rmin

The scaled constraints of Eqs. 5.3d and 5.3e are given by

low
liw − liiw
c (x) = low
≤ 0, iw = 1, 2, · · · , Nw ,
liw
up (5.12)
liiw − liw
c (x) = up ≤ 0, iw = 1, 2, · · · , Nw .
liw

The components of the design variable vector x are also of different scales, e.g., for a slanted
well, z0iw and z1iw may be of different scales with xiw iw iw iw
0 , x1 , y0 and y1 ; for a multi-segmented

well, δθiiw are of completely different scales with xiw iw iw

0 , y0 and z0 . Similar changes to each

168
component of x may produce different magnitude of changes in the NPV function J(m, x).
Hence, we introduce a new variable vector, v, which has similar magnitude in all its com-
ponents and then solve the optimization problem in terms of v. The ith component of v is
given by
xi − xlow
i
vi = up , for i = 1, 2, · · · , Nx , (5.13)
xi − xlow
i

where 0 ≤ vi ≤ 1. Given vi , the ith component of x can be obtained by

xi = xlow
i + vi (xup low
i − xi ), for i = 1, 2, · · · , Nx . (5.14)

Denote the total number of nonlinear constraints as Nnc and the nonlinear constraints
as ci (v) ≤ 0, i = 1, 2, · · · , Nnc , the problems defined in Eqs. 5.3 and 5.10 can be written in
a general form of

max J(m, v)
v∈RNx

s.t. vi ∈ [0, 1], i = 1, 2, · · · , Nx , (5.15)

ci (v) ≤ 0, i = 1, 2, · · · , Nnc .

With ALM, the solution of a constrained optimization problem is obtained by solving

a series of bound constrained optimization sub-problems with a penalized objective function
defined by

Nnc  
X 1
LA (m, v, λki L , µkL ) = J(m, v) − λki L (ci (v) 2
+ si ) + k (ci (v) + si ) , (5.16)
i=1
2µ L

where λki L is the Lagrangian multiplier within the kL th inner loop for the nonlinear constraint
ci (v), µkL is the penalty parameter within the kL th inner loop iterates where all ci ’s share
the same µkL , and si is a positive slack variable used to convert the ith inequality constraint
(ci (v) ≤ 0) into an equality constraint. The Lagrangian function of Eq. 5.16 is a concave
quadratic function of the slack variable si . Denote the term in the bracket of Eq. 5.16 as

169
ψci , then
Nnc
X
LA (m, v, λki L , µkL ) = J(m, v) − ψci . (5.17)
i=1

Following Nocedal and Wright [62], we eliminate the slack variables for the inequality con-
∂ψci
straints, si , by solving si from ∂si
= 0 and substitute si back to ψci so that


k
 2
− µ L λki L ci < −µkL λki L


2
ψci (v, λki L , µkL ) = . (5.18)
c2i
λki L ci + −µkL λki L


2µkL
ci ≥

Eq. 5.18 indicates that in ALM, if ci (v) is less than the switch-point, −µkL λki L , ψci is a
constant that has no effect on the optimization; otherwise, ψci is a quadratic function of
ci (v).
The optimization process using ALM consists of two coupled loops. In the inner loop,
an iterative optimization procedure is used to maximize LA (m, v, λki L , µkL ) given in Eq. 5.17
with the value of the Lagrangian multiplier λki L and penalty parameter µkL fixed. In the outer
loop of the optimization process, the value of λki L and µkL are updated in order to minimize
the violation of constraints. When the constraint violation is small, λki L ’s are updated and
µkL is unchanged; otherwise, λki L ’s are kept the same and µkL is updated (decreased), see
Chen and Reynolds [11]. In this chapter, following [11], the initial Lagrangian multipliers
are estimated as
 
ci (v0 )
λ0i = max 0, , i = 1, 2, · · · , Nnc , (5.19)
µ0

where v0 is the initial guess of the scaled design variables. It is very important to choose a
reasonable value for the initial penalty parameter µ0 . From Eq. 5.16, if µ0 takes a large value
which imposes negligible penalty to the NPV, then the first few inner loops, perform like
unconstrained optimization which may lead to a great constraint violation. When this occurs,
many extra simulations are needed to eliminate the constraint violation in the following inner
loops so that the updated µkL is small enough. However, if µ0 takes a very small value which
corresponds to a large penalty term, we may not be able to increase the NPV once the

170
the current estimate is infeasible or close to the boundaries of the feasible region, since the
objective function will be dominated by the penalty term. In this work, we chose µ0 by
ensuring that the average penalty term is on the order of a reasonable portion of the initial
NPV value.
Nnc
γJ(v0 ) = ĉ, (5.20)
2µ0

where, γ is the percentage of the average penalty term over the initial NPV value (we use
γ = 10%), ĉ is the average value of the constraints ci ’s (we use ĉ = 1). From Eq. 5.20, we
can obtain
Nnc
µ0 = ĉ. (5.21)
2γJ(v0 )

Please see Algorithm 12 for a detailed implementation of the ALM method.

Due to the gradient estimation error in StoSAG and the fact that the objective
function surface of the well placement optimization is rough, ALM is applied twice. In the
first application of ALM, the inner loop optimization represented by Eq. 5.17 is solved
using StoSAG. Then starting from the optimal estimates of v, λi ’s and µ obtained, ALM
is applied again where the inner loop optimization is solved using GPS. Both applications
of ALM are terminated either at convergence or when the maximum number of simulation
runs exceeds a specified number. Algorithms 2 and 3 in Chapter 2, respectively, present the
procedures of StoSAG and GPS algorithms. The convergence criteria for StoSAG is set as five
consecutive failures in improving LA (m, v, λkL , µkL ) along the StoSAG search direction using
backtracking where the maximum number of step size cuts is set equal to 5. However, the
perturbation size and the initial step size are chosen according to the optimization problem
and are specified for each numerical example presented later. The convergence criteria for
GPS is that the relative change of LA (m, v, λkL , µkL ) is less or equal to 10−4 after searching
along Nx coordinates directions or a maximum number of simulations is reached. In both
StoSAG and GPS, the bound constraints are solved using truncation, i.e., truncating the
design variables beyond the bounds to their nearest bounds.

171
Algorithm 12 Augmented Lagrangian method to solve Eq. 5.15

• Initialization: kL = 0. Choose x0 , v0 , λ0i and µ0 . Set β = 0.2, α = 0.1, τ = 0.25, η ∗ =

0.01, η kL = 0.1.

• Step 1: Find the approximate optimum of the augmented Lagrangian function defined
in Eq. 5.17 subject to the bound constraints using StoSAG or GPS, starting from v0 .
The optimal solution is denoted by vk∗L in the transformed domain or x∗kL in the original
domain.

• Step 2: Check constraint violation, σcv . If σcv < η ∗ , set s∗ = x∗kL and terminate the
optimization algorithm; otherwise go to Step 3.
v
u Nnc
uX 

2
σcv = t max 0, ci vk∗L
i=1

• Step 3: If σcv ≤ η kL , update the Lagrangian multipliers and tighten the constraint
violation tolerance by
ci (v ∗ )
λi = max{0, λi + µkkLL } and η kL +1 = max{η kL ∗ min (µkL )β , 0.5 , η ∗ },


and set
µkL +1 = µkL , v0 = vk∗L and kL = kL + 1.

• Step 4: If σcv > η kL , update the penalty parameter and tighten the constraint violation
tolerance by
λikL +1 = λki L and η kL +1 = max{η kL ∗ min (µkL )α , 0.5 , η ∗ },


and set
µkL +1 = τ µkL , v0 = vk∗L , kL = kL + 1.

Where, η kL and η ∗ are the constraint violation tolerance at the kL th outer loop and that at
convergence, respectively.

5.3 Bi-objective Well Placement Optimization

Given an ensemble of realizations which reasonably covers the uncertainty of the

model parameters, it is desirable to maximize the average NPV over all realizations and,
in order to minimize downside risk, to maximize the minimum NPV. The lexicographical
method optimizes the two objectives sequentially according to the order of importance where
the most important objective is firstly optimized and then the less significant objective is
optimized with the value of previous objective treated as an additional constraint.

172
 In this section, the primary objective is to maximize the expected value of NPV over
the set of Ne realizations of the reservoir model parameter chosen to represent the geological
uncertainty, {mk }N
k=1 , i.e., the optimization problem pertaining to the primary objective is
e

given by

Ne
1 X
max JE (v) = J(v, mk ) (5.22a)
v∈RNx Ne k=1

s.t. 0 ≤ vi ≤ 1, i = 1, 2, . . . , Nx (5.22b)

ci (v) ≤ 0, i = 1, 2, · · · , Nnc (5.22c)

where Ne is the number of realizations representing the geological uncertainty, v is the

scaled well trajectory variables, ci (v)’s represent the nonlinear constraints, and Nnc is the
total number of nonlinear constraints. The primary optimization problem defined in Eq.
5.22 is simply the robust optimization under nonlinear constraints which can be solved
using StoSAG followed by GPS coupled with ALM. When the number of reservoir models,
Ne , is greater or equal to 10, following [24], we use only one perturbation per reservoir
model to estimate a StoSAG gradient which requires Ne simulation runs. In GPS, each
coordinate search requires at least 2 ∗ Ne simulation runs. By solving Eq. 5.22, we obtain
the estimate of the optimal (scaled) vector of well trajectories, denoted by vE∗ , and the
corresponding expected NPV, denoted by JE∗ . To minimize the uncertainty in the NPV’s
resulted from the geological uncertainties, previous researchers have tried to minimize the
standard deviation of the optimal NPV’s, or to maximize the minimum NPV. To avoid the
situation that minimization of the standard deviation is achieved by decreasing the highest
NPV, as observed for fluvial models by [49], we maximize the minimum NPV in the secondary
optimization, subject to the constraint that the average NPV should not be smaller than
JE∗ or some percentage of JE∗ , i.e., ηJE∗ where 0.95 ≤ η ≤ 1.0. Note that the constrained
optimization problem starts from vE∗ , which is a feasible point.

173
 The second optimization problem is given by

max JR (v) = min{J(v, mk )}Ne

k=1 (5.23a)
v∈RNx

s.t. JE (v) ≥ ηJE (vE∗ ), (5.23b)

0 ≤ vi ≤ 1, i = 1, 2, . . . , Nx , (5.23c)

ci (v) ≤ 0, i = 1, 2, · · · , Nnc (5.23d)

and solved using ALM. To avoid poor scaling of the constraints, Eq. 5.23b is reformulated
as
ηJE (vE∗ ) − JE (v)
cJ = ≤ 0. (5.24)
JE (vE∗ )

The corresponding augmented Lagrangian function is defined similar to Eq. 5.17, but with
an extra term. Specifically,

Nnc
X
LA (v, λkciL , λkcJL , µkL ) = JR (v) − ψci (v, λki L , µkL ) − ψcJ (v, λkcJL , µkL ), (5.25)
i=1

where 
− µ L λkc L 2
k

cJ < −µkL λkcJL


kL kL 2 J
ψcJ (v, λ , µ ) = . (5.26)
c2J
λkc L cJ + cJ ≥ −µkL λkcJL


J 2µkL

Here, λkcJL is the Lagrangian multiplier associated with the nonlinear constraint cJ (v) ≤ 0
within the kL th inner loop iterates. Following a a similar approach as presented in Eq. 5.19,
the initial value of λcJ is set to 0 since

cJ (vE∗ )
 
λ0cJ = max 0, = 0. (5.27)
µ0

Maximization of the minimum NPV over all realizations, JR (v), is a max-min prob-
lem. JR (v) is continuous with respect to the optimization variables v, but not differentiable
everywhere. When the NPV of the risk realization becomes higher than or equal to another
realization, the function JR (v) becomes non-differentiable and the gradient becomes unde-

174
fined. Liu and Reynolds [50] used the adjoint gradient to solve a bi-objective well control
optimization problem. They had to transform the maximization of the expected shortfall,
JR (v), to a constrained optimization problem which maximizes a newly introduced param-
eter α subject to the constraints that NPV for each realization is greater than α. However,
with the stochastic approximate gradient coupled with steepest ascent algorithm, the non-
differentiability is not an issue.
Denote JˆR (v, λki L , µkL ) = JR (v) − N kL kL
P nc
i=1 ψci (v, λi , µ ), Eq. 5.25 can be rewritten as

LA (v, λkciL , λkcJL , µkL ) = JˆR (v, λki L , µkL ) − ψcJ (v, λkcJL , µkL ). (5.28)

Denote the realization with the lowest NPV value among the Ne realizations at `th iteration
as mr , then mr is given by

mr = min {J(v ` , mk )}N

k=1 .
e
(5.29)
k

mr is not fixed and may change whenever the (scaled) design variable v is updated.
When maximizing Eq. 5.28, a basic simplex search direction dsim can be obtained in
a similar way as presented in Eq. 2.8, i.e.,

 
dsim ≡ dLA = (∆V ∆V T )+ ∆V ∆LA v, λki L , λkcJL , µkL , (5.30)

where ∆V = [δv1 , δv1 , . . . , δvNpert ], and ∆LA = [δLA,1 , δLA,2 , . . . , δLA,Npert ]T . To obtain a
reliable stochastic gradient of LA , five to fifteen perturbations have to be used (i.e., 5 ≤
Npert ≤ 15). Since the evaluation of Eq. 5.28 requires Ne simulation runs, Eq. 5.30 requires
Ne ∗ Npert simulation runs, which is quite expensive.
Note that at beginning of each iteration of an optimization algorithm, mr (Eq. 5.29) is
computed so JˆR (v, λki L , µkL ) only involves a single model mr while ψcJ (Eq. 5.26) involves the
entire ensemble of realizations used to represent the geological uncertainty. When computing
the stochastic gradient of cJ (v, λkcJL , µkL ) in order to obtain the stochastic gradient of ψcJ ,

175
it is appropriate to use one perturbation per realization as discussed previously in Section
2.2.2. However, JˆR (v, λki L , µkL ) depends only on mr so we will need to use five to fifteen
perturbations to compute the stochastic gradient of JˆR where the simulation will need to
be run for each perturbation but based on the single model mr . In essence, the stochastic
gradient of the augmented Lagrangian function can be approximated using the sum of two
stochastic gradients. The details of this procedure are formulated below.
The gradient of Eq. 5.28 (or equivalently Eq. 5.25) is given by

∇v LA (v, λkciL , λkcJL , µkL ) = ∇v JˆR (v, λki L , µkL ) − ∇v ψcJ (v, λkcJL , µkL ), (5.31)

where 
cJ ≤ −µkcJL λkL

0

∇v ψcJ (v, λkcJL , µkL ) =   , (5.32)
 −λkc L + cJ
cJ > −µkL λkcJL


J µkL
∇ v cJ

and from Eq. 5.24,

∇v JE (v)
∇v cJ (v) = − . (5.33)
JE (vE∗ )

We propose to separate the StoSAG direction dLA into two parts where we estimate StoSAG
gradients for the two terms JˆR and ψcJ separately, denoted as dJˆR and dψcJ respectively.
dLA obtained in this way is denoted by

dsep = dJˆR − dψcJ . (5.34)

PNnc
dJˆR is the stochastic gradient of the summation of JR and − i=1 ψci , and is approximated
using Npert perturbations (we use Npert = 10 in this chapter):

dJˆR = (∆V ∆V T )+ ∆V ∆Jˆ v, mr , λkL , µkL .

(5.35)

176
dψcJ is obtained as 
cJ ≤ −µkL λkL

0

dψcJ =   (5.36)
 −λkL + cJ kL kL


µkL
dcJ cJ > −µ λ ,

where dcJ is approximated using one perturbation per realization given by

Ne
1 X
+ kL kL


dcJ = (δvk δvk )δvk δJ v, m k , λ , µ . (5.37)
Ne ∗ JE∗ k=1

The number of simulation runs required to approximate dsep is Npert if cJ ≤ −µkL λkcJL and
Ne + Npert if cJ > −µkL λkcJL . If Ne = 10 and Npert = 10, estimating dsep requires only 10 to
20 simulation runs while estimating dsim requires 100 simulation runs.

5.4 Numerical Example: Channelized Reservoir

In this section, we present numerical results for the channelized reservoir model whose
description is given in Section 3.2.1. Recall that the channelized reservoir is box-shaped
with 50 × 50 × 14 gridblocks where each gridblock is 200 ft × 200 ft × 10 ft. We firstly
consider the well trajectory optimization of slanted wells and multi-segmented wells for
deterministic optimization. Then we consider the bi-objective well placement optimization
where we maximize the expected NPV over a set of realizations characterizing the geological
uncertainty and maximize the minimum NPV.
Ten realizations are used to represent the geological uncertainty where the facies
depositions are generated using object-based modeling and the petrophysical parameters
for each facies (i.e., permeability and porosity) are generated using sequential Gaussian
Simulation; see related parameters in Table 3.1. The horizontal permeability along the x and
1
y directions are set equal while the vertical permeability is set equal to 10
th of the horizontal
permeability. The reservoir life is set to be 3,000 days throughout where all wells operate
under BHP control. The BHP’s for injectors are 6,000 psi and the BHP’s for producers are
2,000 psi. We consider placing 8 wells including 6 producers initialized as horizontal wells
of 2, 000 ft in length which are completed at different layers and 2 injectors initialized as

177
slightly slanted wells of 616 ft in length which fully penetrate all 14 layers. Fig. 5.3 shows
the initial well trajectories plotted on the horizontal permeability field (upper row) and the
porosity field (bottom row) of layer 7 for three different realizations. For easier view, the
projected initial paths are plotted on layer 7 in Figs. 5.6(a) and 5.11(a) without any value
of petrophysical parameters. In Figs. 5.3, the squares, circles and black dots represent the
projections of the entry points, exit points and perforations of a well respectively; the black
solid line represent the projections of the perforations in the current layer. PRO-01 and PRO-
06 are initially completed in layer 1, PRO-03 and PRO-04 are initially completed in layer 5,
and PRO-02 and PRO-05 are initially completed in layer 11. The economic parameters to
calculate the NPV are set as: ron = $50.0/STB, cnw = $5.0/STB, cnwi = $5.0/STB, b=10.0%.
The drilling cost for well iw is defined as

Nseg
X q q 
Cwiw = iw iw 2 iw iw 2 iw iw 2
1, 524 (zi − zi−1 ) + 3, 048 (xi − xi−1 ) + (yi − yi−1 ) . (5.38)
i=1

A minimum well spacing of 500 ft is also enforced, see Eq. 5.11. Meanwhile, we also enforce
the maximum length of a well be no greater than 4,000 ft for producers and 2,000 ft for
injectors, see Eq. 5.12.
We consider the optimization of two scenarios, where in one case, all eight wells are
considered as slanted wells and in another case, the six producers are parameterized as multi-
segmented wells and the two injectors remain as slanted wells. Each multi-segmented well is
characterized by 8 segments where the entry point of a well (x0 , y0 , z0 ) is bounded into the
reservoir, the lower and upper bounds on the azimuth angle of the entry segment θ1iw are
set equal to 0◦ and 360◦ , the lower and upper bounds on the change in the azimuth angle
between two consecutive well segments, δθiiw , are set equal to −20◦ and 20◦ , and the lower
and upper bounds on the segment length, liiw , are set equal to 200 ft and 500 ft.
All the optimization variables are normalized to [0,1]. In StoSAG, the perturbations of
the ith design variable are sampled from the normal distribution N (0, σi2 ), i = 1, 2, · · · , Nx ,
where σi ’s are chosen so that a perturbation causes a reasonable change in the objective

178
 (a) kh - Realization #1 (b) kh - Realization #8 (c) kh - Realization #10

(d) φ - Realization #1 (e) φ - Realization #8 (f) φ - Realization #10

Figure 5.3: Projections of the initial well trajectories on the horizontal permeability field
(upper row) and the porosity field (bottom row) of layer 7 for three different realizations,
channelized model.

function that is being maximized. Here, σi ’s are chosen so that the two end points of each
well move by approximately one gridblock when the perturbation of each parameter have a
magnitude of σi . For example, if the ith design variable is a coordinate variable and there
1
are N gridblocks along the coordinate direction, we approximately take σi = N
. Hence,
σi is set to 0.02 in the x and y directions and 0.07 in the z direction. Denote the initial
0
length of each multi-segmented well as liw , iw = 1, 2, · · · , Nw and denote the minimum value
of the length and width of a gridblock as ∆X. The perturbations in θ1iw , δθiiw and liiw
in the real domain should move the end point of a multi-segmented well roughly by one
gridblock. σi ’s for θ1iw ’s are chosen so that liw
0
sin(σi (θ1up,iw − θ1low,iw )) ≈ ∆X, σi ’s for δθiiw ’s
0
are chosen so that liw sin(σi (δθiup,iw − δθilow,iw )Nseg ) ≈ ∆X, and σi ’s for liiw ’s are chosen so
that σi (liup,iw − lilow,iw )Nseg ≈ ∆X. So for variables θ1iw , δθiiw and liiw , the standard deviations
σi ’s are set as 0.01, 0.02 and 0.05 respectively, which correspond to 3.6◦ , 0.8◦ and 32 ft.

179
The maximum step size is set as 0.1, and the maximum number of step size cuts and the
maximum number of allowable resamples are both set equal to 5. After the application of
ALM using StoSAG, ALM using GPS is used to fine tune the optimal well locations. In
ALM using GPS, a fixed mesh is used where ∆i ’s are set equal to σi ’s. GPS is considered to
converge if NPV cannot be improved after searching all coordinate directions of the design
variable. Note that there is still a possibility that higher NPV can be obtained if we continue
the optimization process with a smaller mesh size ∆i . However, we avoid doing that to save
computational resources. In deterministic optimization, the maximum number of simulation
runs is set as 3,000. In bi-objective well placement optimization, the maximum number of
simulation runs is set as 4,500 for each stage of the optimization runs (the robust optimization
and the secondary optimization).

5.4.1 Deterministic optimization using realization #1

In this section, deterministic optimization using realization #1 of the channelized

reservoir example (see Fig. 5.3(a) and (d)) is conducted starting from the initial well place-
ment given in Fig. 5.3. We consider the optimization of both slanted wells and multi-
segmented wells. For slanted well parameterization, (i) bound constraints on the design
variables ((x, y, z) coordinates of the two end points of wells) enforce the condition that all
wells fall in the reservoir and (ii) the nonlinear constraints include the well spacing con-
straints and well length constraints. However, for multi-segmented well parameterization,
(i) the bound constraints on design variables do not guarantee that the whole well lays in-
side the reservoir and well trajectories that fall outside the reservoir are not considered as
perforations, and (ii) since the maximum well length constraints can be enforced by bound
constraints on liiw , minimum well spacing constraint is the only type of nonlinear constraints
in this example. Different than the initial guess for slanted wells, the circle symbols of PRO-
03 and PRO-06 in Fig. 5.3 are selected as the entry points of the multi-segmented wells, in
order to avoid the truncation of the well trajectories when the segment lengths are increased
so that PRO-03 and PRO-06 go beyond the eastern boundary of the reservoir.

180
 1. Slanted trajectory optimization followed by multi-segmented trajectory optimiza-
tion
The 8 wells are initially parameterized as slanted wells and StoSAG followed by GPS
coupled with ALM is used to optimize the well trajectories where the total number of design
variables (coordinates of well end points) is 48. Then the optimal solutions of the producers
found in the first optimization are parameterized as multi-segmented wells with 8 segments
of equal lengths while the injectors are parameterized as one-segmented wells, i.e., slanted
wells, then using the reparameterized optimal solution for the slanted well case as the initial
guess, StoSAG followed by GPS is applied. The total number of design variables in this
case is 174. Three optimization runs with different seeds are tried. Fig. 5.4 shows the
NPV versus number of simulation runs obtained where the blue and red solid lines represent
the slanted well trajectory optimization using StoSAG and GPS respectively, and the blue
and red dashed lines represent the subsequent multi-segmented well trajectory optimization
using StoSAG and GPS respectively. When optimizing the slanted well trajectories, StoSAG
improved the NPV significantly, by approximately 55% on average, within 600 simulation
runs. However, the difference between the optimal NPV’s obtained with three seeds can be
up to $0.17×109 , which is equivalent to 4.4% of the initial NPV, indicating that StoSAG can
suffer from pre-mature convergence due to the roughness of the objective function surface
and the gradient estimation error. A subsequent optimization using GPS is able to improve
the NPV by approximately another 8% and the difference between the optimal NPV’s for the
three seeds is reduced to $0.15 × 109 , or equivalently 3.7% of the initial NPV. The follow-up
multi-segmented trajectory optimization only improves the NPV slightly (from $6.41 × 109
to $6.49 × 109 on average of three seeds, which is equivalent to 2% of the initial NPV),
indicating that it may be unnecessary to use over complicated parametrization for wells in
a box-shaped reservoir.
Fig. 5.5 shows the cumulative oil production, cumulative water production and in-
jection for the initial and optimal well placement obtained using seed 3 (the one with the
highest NPV among three seeds) after the optimization using slanted well parameteriza-

181
 Optimal NPV:6.4088# 109 $ Optimal NPV:6.4943# 109 $ Optimal NPV:6.5554# 109 $
7 7 7

6 6 6
NPV, 109 $

NPV, 109 $

NPV, 109 $
5 5 5

4 4 4

3 3 3
0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000
No. of Simulations No. of Simulations No. of Simulations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 5.4: NPV versus number of simulation runs obtained for slanted well trajectory op-
timization followed by multi-segmented trajectory optimization, deterministic optimization
using realization #1, channelized model. The solid lines represent the slanted well trajectory
optimization (StoSAG in blue and GPS in red) and the dashed lines represent the subsequent
multi-segmented well trajectory optimization (StoSAG in blue and GPS in red).

tion followed by multi-segmented well parameterization. In Fig. 5.5, Qcum cum

o , Qw and Qcum
wi

respectively represent the cumulative oil production, cumulative water production and injec-
tion. From Fig. 5.5, after the optimization using slanted well parameterization, the amounts
of injected and produced water are both greatly reduced while the amount of produced oil
remains similar. Consistent with Fig. 5.4 that the multi-segmented trajectory optimization
only improves the NPV slightly, there is only slight change in Qcum cum
o , Qw and Qcum
wi after

the follow-up optimization.

In Fig. 5.6, the well trajectories of the initial and optimal well placement (obtained
with seed 3) are projected onto layer 7 without any value of petrophysical parameters. These
paths are also plotted on the oil saturation field of layer 7 after 3,000 days of production in
Fig. 5.7. Since in Figs. 5.6 and 5.7, the projections of the optimal trajectories of the multi-
segmented wells are close to straight, we only present the coordinates of the starting and end
points for all eight wells in Table 5.1. From Fig. 5.7 and Table 5.1, in the optimal solution
of the slanted wells, (i) after using StoSAG, both injectors are moved in the north direction
with shorter trajectories where the lengths of INJ-01 and INJ-02 are reduced to 315 ft and
324 ft respectively from 616 ft, and all producers are moved closer to the reservoir bounds;
(ii) the subsequent application of ALM with GPS tends to increase the lengths of some

182
wells, i.e. PRO-01, PRO-02 and PRO-04, and also causes PRO-04 to be moved closer to the
southwest corner. Considering that the subsequent multi-segmented trajectory optimization
only slightly increases the optimal NPV, the optimal well trajectories look quite similar to
the ones obtained after the slanted well optimization. Fig. 5.8 shows the projections of the
initial and optimal injector perforations obtained with slanted followed by multi-segmented
trajectory optimization using seed 3, onto the x − z plane of the horizontal permeability field
of realization #1. From Fig. 5.8, after optimization, the two injectors not only have shorter
trajectories, but also perforate gridblocks with relative low permeabilities compared to the
initial well placement given in Fig. 5.8(a), leading to a significant decrease in the amount of
cumulative water injection and production (see Fig. 5.5).

1200 1200 1200

Cumlative Production, 10 6 bbl

Cumlative Production, 10 6 bbl

Cumlative Production, 10 6 bbl

Qcum Qcum Qcum Qcum
o
Qcum
w
Qcum
wi
Qcum
o
Qcum
w
Qcum
wi
o w wi
1000 1000 1000

800 800 800

600 600 600

400 400 400

200 200 200

0 0 0
0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000
Time, days Time, days Time, days
(a) Initial locations (b) Slanted Opt. (c) Slanted - multi-segmented Opt.

Figure 5.5: Cumulative oil production, water production and water injection obtained with
the initial and optimal well placements after slanted well trajectory optimization followed
by multi-segmented trajectory optimization using seed 3, deterministic optimization using
realization #1, channelized model.

2. Multi-segmented trajectory optimization

To compare with the optimization process of slanted wells followed by multi-segmented
wells, we also carried out optimization of multi-segmented wells by itself starting from the
initial well trajectories given in Fig. 5.7(a). Fig. 5.9 shows the NPV versus number of
simulation runs obtained using StoSAG (solid blue line) and GPS (solid red line) respec-
tively. From Fig. 5.9, StoSAG improves the NPV significantly, by approximately 52% on
average, within 600 simulation runs. However, the difference between the optimal NPV ob-
tained with three seeds can be up to $0.34 × 109 , or equivalently 9% of the initial NPV,

183
 (a) Initial locations (b) Slanted Opt. - StoSAG (c) Slanted Opt. - GPS

(d) Multi-segmented Opt. - (e) Multi-segmented Opt. - GPS

StoSAG

Figure 5.6: Projections of the initial and optimal well placements after slanted well trajectory
optimization followed by multi-segmented trajectory optimization on layer 7 using seed 3,
deterministic optimization using realization #1, channelized model.

Table 5.1: Summary of the trajectories of all 8 wells obtained with the initial and opti-
mal well placements after slanted well trajectory optimization followed by multi-segmented
trajectory optimization using seed 3, deterministic optimization using realization #1, chan-
nelized model.
Initial setting Slanted Opt. Slanted-multi-segmented Opt.
Wells
(xiw iw iw
0 , y0 , z0 ) (xiw iw iw
Nseg , xNseg , xNseg ) (xiw iw iw
0 , y0 , z0 ) (xiw iw iw
Nseg , xNseg , xNseg ) (xiw iw iw
0 , y0 , z0 ) (xiw iw iw
Nseg , xNseg , xNseg )
PRO-01 (300, 700, 4806) (2300, 700, 4806) (100, 218, 4800) (4092, 0, 4800) (100, 218, 4800) (4099, 202, 4800)
PRO-02 (3900, 700, 4905) (5900, 700, 4905) (5194, 2707, 4880) (6958, 0, 4888) (4994, 3107, 4880) (6976, 252, 4888)
PRO-03 (7700, 700, 4845) (9700, 700, 4845) (7549, 236, 4800) (9900, 0, 4850) (7550, 220, 4800) (9887, 1, 4850)
PRO-04 (300, 9300, 4845) (2300, 9300, 4845) (0, 9202, 4831) (1322, 9400, 4826) (0, 9202, 4831) (1694, 9259, 4805)
PRO-05 (3900, 9300, 4905) (5900, 9300, 4905) (3701, 9654, 4860) (5553, 9366, 4823) (3701, 9654, 4863) (5659, 9435, 4800)
PRO-06 (7700, 9300, 4806) (9700, 9300, 4806) (7335, 9223, 4800) (10000, 9471, 4800) (7320, 9184, 4800) (9800, 9450, 4800)
INJ-01 (900, 4900, 4940) (1500, 4900, 4800) (1047, 4564, 4940) (1349, 4685, 4804) (1052, 4556, 4940) (1349, 4685, 4801)
INJ-02 (8500, 4900, 4800) (9100, 4900, 4940) (9164, 3991, 4936) (9211, 4213, 4800) (9164, 3991, 4936) (9211, 4213, 4800)

again indicating that StoSAG by itself cannot find the optimal solution of the well place-
ment optimization problem. A subsequent optimization using GPS is able to improve the

184
 (a) Initial locations (b) Slanted Opt. - StoSAG (c) Slanted Opt. - GPS

(d) Multi-segmented Opt. - (e) Multi-segmented Opt. - GPS

StoSAG

Figure 5.7: Oil saturation after 3,000 days of production plotted on layer 7, obtained with
the initial and optimal well placements after slanted well trajectory optimization followed
by multi-segmented trajectory optimization using seed 3, deterministic optimization using
realization #1, channelized model.

NPV by another 10% and the difference between the optimal NPV of any pair of results for
two distinct seeds is narrowed to $0.09 × 109 , or equivalently 2.4% of the initial NPV. More
importantly, the average optimal NPV obtained with 3 runs for the multi-segmented well
optimization is 3.3% less than the optimization process of slanted trajectories followed by
multi-segmented trajectories within 3,000 simulation runs, which again indicates that it is
unnecessary to use over-complicated well trajectory parameterization. In fact, the number
of optimization variables for a multi-segmented well trajectory in this example is 3.5 times
more than the number needed for slanted well trajectory parameterization so the multi-
segmented case requires more simulation runs to converge. This explains the observation
that a smaller NPV is obtained for a given number of total reservoir simulation runs when

185
Figure 5.8: Projections of the initial and optimal injector perforations obtained with slanted
followed by multi-segmented trajectory optimization using seed 3, onto the x − z plane of the
horizontal permeability field, deterministic optimization using realization #1, channelized
model.

using the multi-segmented parameterization throughout the optimization process.

Optimal NPV:6.4165# 109 $ Optimal NPV:6.3219# 109 $ Optimal NPV:6.3267# 109 $

7 7 7

6 6 6
NPV, 109 $

NPV, 109 $

NPV, 109 $
5 5 5

4 4 4

3 3 3
0 1000 2000 3000 0 1000 2000 3000 0 1000 2000 3000
No. of Simulations No. of Simulations No. of Simulations
(a) Seed 1 (b) Seed 2 (c) Seed 3

Figure 5.9: NPV versus number of simulations obtained for multi-segmented trajectory
optimization, deterministic optimization using realization #1, channelized reservoir. The
blue lines represent the StoSAG algorithm and the red lines represent the GPS algorithm.

Fig. 5.10 shows the cumulative oil production, cumulative water production and in-
jection for the initial and optimal well placement obtained using seed 1 (the one with the
highest NPV among three seeds) after the optimization using multi-segmented well param-
eterization throughout the optimization process. From Fig. 5.10, similar to the previous
case, after the optimization, the amounts of injected and produced water are both greatly
reduced while the amount of produced oil remains similar. In Fig. 5.11, the well trajectories
of the optimal well placement (obtained with seed 1) after StoSAG and GPS are projected

186
onto layer 7 without any value of petrophysical parameters. These paths are also plotted
on the oil saturation field of layer 7 after 3,000 days of production in Fig. 5.12. Since in
Figs. 5.12, the projections of the optimal trajectories of the multi-segmented wells are close
to straight, we only present the coordinates of the starting and end points, the total lengths,
azimuth angles of the entry segments, and the intervals of the perforated layers for all eight
wells in Table 5.2.
From Fig. 5.12, (i) similar to the optimal solution of slanted wells, both injectors are
moved in the north direction and all producers are moved closer to the reservoir bounds,
after using StoSAG; (ii) the subsequent application of ALM with GPS significantly changes
some well trajectories, i.e., the lengths of both PRO-01 and PRO-02 are increased, PRO-
02 is shifted downward by almost 2,000 ft, and PRO-04 whose trajectory lays outside of
the reservoir after StoSAG is now shifted back into the reservoir again. Note that with the
multi-segmented well parameterization, the bound constraints on the design variables do not
guarantee that the whole well lays inside the reservoir and trajectories outside of the reservoir
boundaries are not considered as perforations, which is the case of PRO-04 in Figs. 5.11(b)
and 5.12(b). Fig. 5.13 shows the projections of the initial and optimal injector perforations
obtained with the multi-segmented trajectory optimization using seed 1, onto the x−z plane
of the horizontal permeability field of realization #1. From Fig. 5.13, after optimization, the
two injectors not only have shorter trajectories, but also perforate gridblocks with relative
low permeabilities compared to the initial well placement given in Fig. 5.13(a), which is
similar to Fig. 5.8.
For all the optimal well placement scenarios shown in Figs. 5.7 and 5.12, both the
minimum well spacing constraints and the maximum well length constraints are satisfied. In
fact, these two types of constraints are rarely violated since maximizing the NPV tends to
penalize long well trajectories and to avoid wells intersecting with each other.

5.4.2 Bi-objective optimization using 10 realizations

In this section, we consider the bi-objective well placement optimization where we

187
 1200 1200

Cumlative Production, 10 6 bbl

Cumlative Production, 10 6 bbl

Qcum Qcum Qcum Qcum
o
Qcum
w
Qcum
wi
o w wi
1000 1000

800 800

600 600

400 400

200 200

0 0
0 1000 2000 3000 0 1000 2000 3000
Time, days Time, days
(a) Initial (b) Multi-segmented Opt.

Figure 5.10: Cumulative oil production, water production and water injection obtained for
the initial and optimal well placements after multi-segmented well trajectory optimization
using seed 1, deterministic optimization using realization #1, channelized model.

(a) Initial locations (b) Multi-segmented Opt. - (c) Multi-segmented Opt. - GPS
StoSAG

Figure 5.11: Projections of the initial and optimal well placement after multi-segmented
trajectory optimization onto layer 7 using seed 1, deterministic optimization using realization
#1, channelized model.

Table 5.2: Summary of the trajectories of all 8 wells obtained for multi-segmented trajec-
tory optimization using seed 1, deterministic optimization using realization #1, channelized
model.
Initial Setting Slanted-multi-segmented Opt.
Wells
(xiw iw iw
0 , y0 , z0 ) (xiw iw iw
Nseg , xNseg , xNseg ) θ1iw liw K1-K2 (xiw iw iw
0 , y0 , z0 ) (xiw iw iw
Nseg , xNseg , xNseg ) θ1iw liw K1-K2
PRO-01 (300, 700, 4806) (2300, 700, 4806) 0 2000 1 (200, 0, 4800) (4198, 32, 4810) 4.7 4000 1-2
PRO-02 (3900, 700, 4905) (5900, 700, 4905) 0 2000 11 (5159, 1968, 4926) (7898, 1219, 4940) -10.5 2868 5-14
PRO-03 (9700, 700, 4845) (7700, 700, 4845) 180 2000 5 (9400, 200, 4860) (6858, 191, 4807) 180.0 2463 1-7
PRO-04 (300, 9300, 4845) (2300, 9300, 4845) 0 2000 5 (200, 9789, 4909) (1650, 9971, 4837) 7.5 1467 4-11
PRO-05 (3900, 9300, 4905) (5900, 9300, 4905) 0 2000 11 (4202, 9564, 4921) (5929, 9890, 4940) 3.9 1813 6-14
PRO-06 (9700, 9300, 4806) (7700, 9300, 4806) 180 2000 1 (9919, 9416, 4820) (7105, 9640, 4800) 171.2 2917 1-3
INJ-01 (1500, 4900, 4800) (900, 4900, 4940) 180 2000 1-14 (883, 4567, 4800) (437, 4568, 4940) 179.8 467 1-14
INJ-02 (8500, 4900, 4800) (9100, 4900, 4940) 0 2000 1-14 (9005, 3930, 4800) (9528, 3933, 4905) 0.3 534 1-11

188
 (a) Initial locations (b) Multi-segmented Opt. - (c) Multi-segmented Opt. - GPS
StoSAG

Figure 5.12: Oil saturation after 3,000 days of production plotted on layer 7, obtained for
the initial and optimal well placement after multi-segmented trajectory optimization using
seed 1, deterministic optimization using Realization #1, channelized model.

Figure 5.13: Projections of the initial and optimal injector perforations obtained with multi-
segmented well trajectory optimization using seed 1, onto the x − z plane of the horizontal
permeability field, deterministic optimization using Realization #1, channelized model.

maximize the expected NPV over ten realizations characterizing the geological uncertainty
and maximize the minimum NPV, see Eqs. 5.22 and 5.23, using Lexicographical method.
We parameterize all 8 wells as slanted wells so the total number of design variables is 48
starting from the initial guess given in Fig. 5.7(a) and considering minimum well spacing
constraints and maximum well length constraints, see Eq. 5.2. Fig. 5.14 shows a summary
of the bi-objective well trajectory optimization using StoSAG followed by GPS coupled with
ALM. According to the lexicographic method, the primary objective (the average NPV over
10 realizations, JE ) is optimized first where JE is improved from $3.76 × 109 to $6.14 × 109 ,

189
which represents a 63% increase in JE . The primary optimization is terminated after 4,500
simulation runs where GPS has not converged yet. During the primary optimization, the
nonlinear constraints are rarely violated. We denote the optimal solution of the primary
optimization as vE∗ . Starting from vE∗ , a secondary optimization is carried out where the
minimum NPV is maximized subject to an extra constraint that the average NPV is no less
than 0.99JE (vE∗ ), or equivalently $6.08 × 109 . In the secondary optimization, Npert = 10
perturbations are used to estimate the StoSAG search direction of JR , dJˆR , using Eq. 5.35
while one perturbation per realization is used to estimate the StoSAG search direction of ψcJ ,
dψcJ , using Eq. 5.36 when cJ > −µkL λkL . Fig. 5.14(b) shows the changes in the augmented
Lagrangian Function (LA ), the minimum NPV (JR ) and the average NPV (JE ) with respect
to the number of simulation runs. It is important to note that LA increases with respect
to number of simulations monotonically only in the inner loop of the ALM. The updates
to the Lagrangian multipliers and the penalty parameter in the outer loop will change the
initial value of LA in the next inner loop. Both JR and JE do not increase monotonically with
respect to the number of simulation runs. From Fig. 5.14(b), the minimum NPV is increased
from $5.74 × 109 to $6.04 × 109 , or by approximately 5%; although the result is unexpected,
the average NPV is also increased from $6.14 × 106 to $6.24 × 109 , or by approximately 1.5%.
Fig. 5.15 shows the indices of the model with the lowest NPV value at each iteration and the
number of violated constraints during the secondary optimization process. As is observed,
the secondary optimization does not appear to suffer from the non-differentiability issue
even though the “risk model” (model mr ) can change from iteration to iteration. At most
iterations, the results of Fig. 5.15 indicate that realization #8 is the risk model, but for a few
iterations, either realization #6 or #7 is the realization with the lowest NPV. Fig. 5.14(c)
shows the cumulative distribution function (cdf) of NPV’s of the 10 realizations obtained
with the well placement of the initial guess and the optimal solution of the bi-objective
optimization. From Fig. 5.14(c), robust optimization shifted the cdf of NPV to the right
and the secondary optimization makes the cdf of NPV more compact. The realizations with
the lowest, second lowest and the third lowest NPV respectively are realizations #8, #7 and

190
#6, i.e., the three realizations that have served as the risk realization at some iterations of
the secondary optimization, see Fig. 5.15(b).

7 6.5 1
LA JR JE Initial
Robust Opt.
Average NPV, 109 $

0.8
6 Bi-obj Opt.

NPV, 109 $
0.6

CDF
5 6
0.4

4 0.2

3 5.5 0
0 1000 2000 3000 4000 0 1000 2000 3000 4000 3 4 5 6 7
No. of Simulations No. of Simulations NPV, 109 $
(a) Robust Opt. (b) Secondary Opt. (c) CDF’s

Figure 5.14: Bi-objective well placement optimization using 10 realizations, channelized

model. (a) Average NPV over 10 realizations versus number of simulation runs in primary
optimization using StoSAG (blue) and GPS (red); (b) the augmented Lagrangian function,
LA , minimum NPV, JR , and average NPV, JE , versus number of simulation runs in secondary
optimization using StoSAG (blue) and GPS (red) where LA is represented by marker x, JR is
represented by the solid line, JE is represented by the dashed line and JE (vE∗ ) is represented
by the dotted line; (c) Cdf’s of NPV’s for the initial well placement and the optimal well
placement after primary and secondary optimization.

10

4
No. of violated consraints
Index of the risk realization
2

0
0 1000 2000 3000 4000

Figure 5.15: Number of violated constraints and indicies of the risk realization during the
secondary optimization of the bi-objective optimization of slanted well trajectories using 10
realizations, channelized model.

Fig. 5.16 shows the cumulative oil production, cumulative water production and
cumulative water injection after the secondary optimization for realization #1 whose NPV
is reduced by 1.3% and for realization #8 whose NPV is increased by 4.6%. From Fig. 5.16,
the secondary optimization slightly increases the volume of injected water and produced
water for realization #1 and slightly decreases the volume of injected water and produced

191
water for realization #8. For both realizations #1 and #8, the secondary optimization
causes negligible change in the cumulative oil production.

600 600

Cumlative Production, 10 6 bbl

Cumlative Production, 10 6 bbl

Qcum
o
Qcum
w
Qcum
wi
Qcum
o
Qcum
w
Qcum
wi

400 400

200 200

0 0
0 1000 2000 3000 0 1000 2000 3000
Time, days Time, days
(a) Robust Opt. - Realization #1 (b) Secondary Opt. - Realization #1

600 600
Cumlative Production, 10 6 bbl

Cumlative Production, 10 6 bbl

Qcum Qcum Qcum Qcum
o
Qcum
w
Qcum
wi
o w wi

400 400

200 200

0 0
0 1000 2000 3000 0 1000 2000 3000
Time, days Time, days
(c) Robust Opt. - Realization #8 (d) Secondary Opt. - Realization #8

Figure 5.16: Cumulative oil production, water production and water injection obtained after
the primary and secondary optimization of the bi-objective optimization, channelized model.

The minimum NPV of the cdf obtained after the secondary optimization in Fig.
5.14 corresponds to realization #8 which was the risk realization during the last several
iterations of the bi-objective optimization procedure. This means during the last iteration,
the algorithm was maximizing the NPV of realization #8 but subject to the constraint that
the expected NPV obtained during the first optimization step of bi-objective lexicographic
optimization does not decrease by more than one percent. If the constraint is removed and
we simply apply deterministic optimization to maximize the NPV of realization #8, we
obtain an NPV value equal to $6.45 × 109 , which is 6.8% higher than the minimum NPV
obtained by bi-objective optimization.
It is worthwhile to mention that if robust optimization found a solution which is
close to a global optimum, it is unlikely that the secondary optimization would improve

192
 Optimal NPV:6.4473# 109 $
7

NPV, 109 $
5

3
0 1000 2000 3000
No. of Simulations

Figure 5.17: NPV versus number of simulation runs for the deterministic optimization using
realization #8, channelized model. The blue line represents the StoSAG algorithm and the
red line represents the GPS algorithm.

the average NPV since these two objectives now conflict with each other. To illustrate the
more normal behavior, we tried a different run of the bi-objective optimization where we
obtained an average NPV of $6.25 × 109 after the primary optimization which is decreased
to $6.23×109 after the secondary optimization, even through the minimum NPV is improved
from $5.70 × 109 to $5.92 × 109 ; see Fig. 5.18.

7 6.5 1
LA JR JE Initial
Robust Opt.
Average NPV, 109 $

0.8
6 Bi-obj Opt.
NPV, 109 $

0.6
CDF

5 6
0.4

4 0.2

3 5.5 0
0 1000 2000 3000 4000 0 1000 2000 3000 4000 3 4 5 6 7
No. of Simulations No. of Simulations NPV, 109 $
(a) Robust Opt. (b) Secondary Opt. (c) CDF’s

Figure 5.18: Bi-objective well placement optimization using 10 realizations using a different
seed, channelized model. (a) Average NPV over 10 realizations versus number of simula-
tion runs in primary optimization using StoSAG (blue) and GPS (red); (b) the augmented
Lagrangian function, LA , minimum NPV, JR , and average NPV, JE , versus number of sim-
ulation runs in secondary optimization using StoSAG (blue) and GPS (red) where LA is
represented by marker x, JR is represented by the solid line, JE is represented by the dashed
line and JE (vE∗ ) is represented by the dotted line; (c) Cdf’s of NPV for the initial well
placement and the optimal well placement after primary and secondary optimization.

193
 5.5 Numerical Example: Oseberg Reservoir

We consider a problem which is based on a modification of the original reservoir

simulation model of Oseberg reservoir in the Norwegian sector of the North Sea. The modi-
fication simply uses a more uniform grid. The reservoir consists of three distinct geological
zones, Etive, Rannoch and Oseberg. Etive is the top zone and Oseberg is the bottom zone.
These two zones are separated by Rannoch which is a relatively tight layer. There is ver-
tical communication between the three zones. We use a 54 × 27 × 10 simulation grid with
non-uniformed grid sizes in the oil and gas zones along x and z directions. The gridblock
size in the x direction is equal to 328 ft in the top and middle part of the reservoir, and the
gridblock size increases gradually from 328 ft to 2,624 ft as we approach the bottom of the
reservoir boundary in the x direction. The gridblock size in the y direction is equal to 656
ft. Only one vertical gridblock is used in the Etive and Rannoch layers, and layers 3 through
10 correspond to the Oseberg zone. Gridblock sizes in the z direction are non-uniform with
values equal to 23.0 ft in Etive, 16.5 ft in Rannoch and 11.5 ft in Oseberg.
The initial reservoir pressure is 4,071 psi at the depth of 8,192 ft subsea, and the
original reservoir fluid bubble point pressure is 3,771 psi. The top structure of the Oseberg
model and the initial oil saturation field are shown in Fig. 5.19. Note that the reservoir is
tilted with a significant dip. The reservoir has a gas cap at the top and an aquifer at the
bottom. The initial gas-oil contact is at 8,192 ft subsea and the water-oil contact is at 8,918
ft subsea. The oil zone is separated from the aquifer by an impermeable tar mat which is
highlighted with a blue line in Fig. 5.19. In the oil zone, the initial oil saturation is 0.885
and the initial water saturation is equal to the irreducible water saturation which is equal
to 0.115. The initial gas saturation in the gas cap is equal to 0.885. The water saturation of
the gas cap is at the irreducible water saturation and there is no oil in the gas cap initially.
In this case study, all fluid related properties and statistical parameters to generate the
model realizations are obtained from Zhang [88]. When generating the petrophycical param-
eters, a 2D non-isotropic exponential covariance function is used where the major correlation
direction is along the x-axis direction and the minor correlation direction is along the y-axis

194
direction. The correlation length is set to 6,555 ft in the major correlation direction, 1,968
ft in the minor correlation direction and 11.64 ft in the vertical direction. The horizontal
log-permeability, vertical log-permeability and the porosity field are all generated following
normal distribution where the mean and deviation of the petro-physical parameters for each
geological zone are given in Table 5.3. Note that horizontal and vertical log-permeability in
Oseberg decrease linearly from top to bottom, i.e., from layer 3 to layer 10. The correlation
between horizontal log-permeability and vertical log-permeability is set equal to 0.8, and the
correlation between horizontal log-permeability and porosity is set equal to 0.3, following
Zhang [88]. The minimum porosity for any gridblock in the model is set to 0.03. In the tar
mat, the horizontal permeability is set to 0.03 mD, and vertical permeability is set to 0.05
mD. Ten realizations of the horizontal permeability, vertical permeability and the porosity
fields are generated to represent the geological uncertainty where realizations #1 and #7 are
shown in Figs. 5.20 and 5.21 respectively.

(a) Reservoir depth (b) Initial oil saturation

Figure 5.19: Top structure and initial oil saturation of the Oseberg model. The tar mat is
highlighted with a blue line.

Table 5.3: Petrophysical property parameters, Oseberg model.

Etive Rannoch Oseberg

Parameters Mean Mean
Mean Variance Mean Variance Variance
top bottom
ln(k) 7.5 1.2 2.1 1.8 7.8 6.3 0.4
ln(kz ) 6.3 1.8 0.15 2.2 6.4 4.4 0.8
φ 0.14 0.002 0.1 0.001 0.22 0.22 0.001

195
 We put two gas injectors under rate control in the gas cap and two oil producers
under pressure control in the oil zone. All four wells are initially vertical and are set to fully
penetrate all 10 reservoir layers. The initial locations of the two producers are set close to the
tar mat to delay gas breakthrough based on engineering experience. The two injectors are
initially placed in the middle of the gas cap. Intuitively, the locations of the gas injectors, as
long as they are in the gas cap, may not affect the reservoir development very much because
the gas density is very low compared to the densities of oil and water and the reservoir
permeabilities are high and fairly uniform with the exception of the tar mat. Due to the
low permeability of the tar mat, it seems likely that the optimal locations of the producers
are just above the tar mat. The initial well locations are shown in both Figs. 5.20 and 5.21.
Injectors are under rate control with fixed injection rate of 100 MMscf/day. Producers are
under BHP control with specified bottomhole pressure of 2,000 psi. The maximum water
cut and gas oil ratio for a producer are set to 0.98 and 561 MMscf/STB, respectively, and a
producer is shut in if these values are exceeded. Also an upper limit of 100,000 STB/day is
set for the oil production rate of each producer and a producer switches to constant oil rate
control if this value is exceeded. The maximum injection BHP is set to be 6,000 psi, and
an injector switches to bottomhole pressure control if this value is exceeded. A minimum
distance of 500 ft between any pair of the wells is enforced. During the optimization, the
bounds on the well trajectory variables of the injectors are set to be the boundaries of
the initial gas cap (the first 31 gridblocks along x direction) while the bounds on the well
trajectory variables of the producers are set to be the boundaries of the initial oil zone (the
32th to 47th gridblocks along x direction). We parameterize all four wells as slanted wells
using the coordinates of their end points and consider both deterministic optimization and
bi-objective optimization in which, there are 24 optimization variables in total. Similar to
the channelized reservoir, all the trajectory variables are scaled to the domain of [0,1] and
the StoSAG perturbations are sampled from Gaussian distributions N (0, σi2 ) so that the
perturbation of each parameter on the magnitude of σi approximately corresponds to one
1
gridblock. For a design variable corresponding to the y-coordinate, σi is set equal to 27
; for

196
 1
a design variable corresponding to the x-coordinate, σi is set equal to 31
for the injectors
1
and 16
for the producers; for a design variable corresponding to z coordinate, σi is set equal
to 0.1. In the GPS implementation, a fixed mesh size, ∆i , is used where ∆i is set equal
to σi . In this example, we set the oil price equal to $50.0/STB, the water disposal cost is
$5.0/STB, the gas injection cost is $1.50/Mscf, the gas disposal cost is zero, and the annual
discount rate is 10.0%.

(a) kh - Layer 1 (b) kh - Layer 3 (c) kh - Layer 10

(d) φ - Layer 1 (e) φ - Layer 3 (f) φ - Layer 10

Figure 5.20: Horizontal permeability (mD, upper row) and porosity field (bottom row) for
realization #1, Oseberg.

5.5.1 Deterministic optimization using realization #1

We firstly investigated deterministic optimization using realization #1. Fig. 5.22

shows the NPV versus number of simulations using StoSAG (blue line) and GPS (red line)
respectively. From Fig. 5.22, StoSAG improved the NPV significantly, by approximately
12.44% within 500 simulation runs and the subsequent GPS method improved the NPV
by another 7% within 200 more simulations. It is worthwhile to mention that due to the
roughness of the objective function for the well placement optimization problem and the
existence of the gradient estimation error in StoSAG, the NPV curve may consists of multiple
plateaus where low-quality search directions are obtained. Moreover, when the StoSAG

197
 (a) kh - Layer 1 (b) kh - Layer 3 (c) kh - Layer 10

(d) φ - Layer 1 (e) φ - Layer 3 (f) φ - Layer 10

Figure 5.21: Horizontal permeability (mD, upper row) and porosity field (bottom row) for
realization #8, Oseberg.

algorithm is terminated after 5 consecutive re-samples without obtaining an uphill direction

(see Algorithm 2), it is often useful to apply a follow-up GPS step.

Optimal NPV:6.1382# 109 $

6.5
NPV, 109 $

5.5

5
0 200 400 600 800
No. of Simulations

Figure 5.22: NPV versus number of simulations obtained for deterministic optimization of
slanted well trajectories using realization #1, Oseberg. The blue line represents the StoSAG
algorithm and the red line represents the GPS algorithm.

Fig. 5.23 shows the oil saturation after 3,900 days of production for the initial and op-
timal well placement using realization #1. In Figs. 5.23, the squares, circles and black dots
represent the projections of the entry points, exit points and perforations of a well respec-
tively; the black solid lines represent the projections of the perforations in the current layer.
After deterministic optimization, (i) the two producers which are initially fully-penetrating
vertical wells become horizontal in the 10th (bottom) layer along the tar mat because of the

198
high and fairly uniform permeability with a lack of high permeable conduits, and (ii) both
injectors still lay in the gas cap where INJ-01 perforates from layer 3 to layer 10 and INJ-02
perforates from layer 1 to layer 10. The minimum distance between these four wells are 511 ft
which satisfies the minimum distance constraint of 500 ft. However, the trajectories indicate
that effectively, we need only one long horizontal well whose length is almost equal to the
reservoir width in the y direction. As we can see from Fig. 5.23, the gas flooding front is
quite stable for the reservoir with a significant dip which indicates the geological uncertainty
does not make a big influence in the flooding process. Similar to the channelized reservoir
model, the minimum well spacing constraints are rarely violated during the optimization
procedure since maximizing the NPV tends to avoid wells intersecting with each other.

(a) Initial - Layer 1 (b) Initial - Layer 3 (c) Initial - Layer 10

(d) Deterministic Opt - Layer 1 (e) Deterministic Opt - Layer 3 (f) Deterministic Opt. - Layer 10

Figure 5.23: Oil saturation after 3,900 days of production obtained with initial and optimal
well locations, deterministic optimization using realization #1, Oseberg.

5.5.2 Bi-objective optimization using 10 realizations

Fig. 5.24 shows a summary of the bi-objective well trajectory optimization where the
average NPV and the minimum NPV of 10 realizations are maximized, subject to minimum
well spacing constraints using StoSAG followed by GPS. In primary optimization (robust
optimization), the average NPV over 10 realizations, JE , is improved from $5.31 × 109 to

199
$6.28 × 109 , by approximately 18%. The primary optimization is terminated after 4,500
simulations where GPS converged with the fixed mesh size. A secondary optimization is then
carried out where the minimum NPV is maximized subject to an extra constraint that the
average NPV is no less than 99% of the optimal NPV obtained in the primary optimization,
or equivalently $6.22 × 109 . Npert = 10 perturbations are used to estimate dJˆR and one
perturbation per realization is used to estimate dcJ in the secondary optimization, see the
discussion of Algorithm 12. Fig. 5.24(b) shows the changes in the augmented Lagrangian
Function (LA ), the minimum NPV (JR ) and the average NPV (JE ) with respect to the
number of simulation runs. From Fig. 5.24(b), we see that in this example the minimum
NPV is only increased from $6.09 × 109 to $6.14 × 109 (i.e., by approximately 0.8%) and
the average NPV is also increased from $6.28 × 106 to $6.32 × 109 (i.e., by approximately
0.6%). Fig. 5.25 shows the indices of the model with lowest NPV value and the number
of violated constraints during the secondary optimization process. Again, the secondary
optimization rarely suffers from the non-differentiable issue and the risk model changes from
realization #2 to #1. Fig. 5.14(c) shows the cumulative distribution function (cdf) of NPV’s
of the 10 realizations obtained with the well placement of the initial guess and the optimal
solutions of the bi-objective optimization. From Fig. 5.14(c), both robust optimization and
the secondary optimization shifted the cdf of NPV to the right. It is worthwhile to mention
that after bi-objective optimization, the realization with the lowest NPV is realization #1
whose NPV is $6.14 × 109 , which is equal to the optimal NPV obtained for the deterministic
optimization of realization #1.
Fig. 5.26 shows the oil saturation after 3,900 days of production for realization #1
(upper row) and #7 (bottom row) obtained with the initial guess and optimal solutions af-
ter robust optimization and secondary optimization. After robust optimization, (i) PRO-01
perforates both layer 9 and layer 10 and PRO-02 perforates only layer 10, and (ii) INJ-01
perforates from layer 2 to layer 4 and INJ-02 perforates from layer 2 to layer 10. After
bi-objective optimization, (i) both producers perforate only layer 10, and (ii) INJ-01 perfo-
rates from layer 3 and layer 4 and INJ-02 perforates layer 3 to layer 10. Realization #7 has

200
the highest NPV among all 10 realizations after both robust optimization and bi-objective
optimization. Considering the locations of the gas injectors do not affect the reservoir de-
velopment much, the major change of the optimal well trajectories is that, PRO-01, which
penetrates layers 9 and 10 after robust optimization, now only penetrates layer 10.

6.5 6.4 1
Initial
Robust Opt.
0.8
6.3 Bi-obj Opt.
6
NPV, 109 $

NPV, 109 $
0.6

CDF
6.2
0.4
5.5
6.1 0.2

5 6 0
0 1000 2000 3000 4000 0 1000 2000 3000 4000 5 5.5 6 6.5
No. of Simulations No. of Simulations NPV, 109 $
(a) Robust Opt. (b) Secondary Opt. (c) CDF’s

Figure 5.24: Bi-objective well placement optimization using 10 realizations, Oseberg reser-
voir. (a) Average NPV over 10 realizations versus number of simulation runs in primary
optimization using StoSAG (blue) and GPS (red); (b) the augmented Lagrangian function,
LA , minimum NPV, JR , and average NPV, JE , versus number of simulation runs in sec-
ondary optimization using StoSAG (blue) and GPS (red) where LA is represented by marker
x, JR is represented by the solid line, JE is represented by the dashed line and JE (vE∗ ) is
represented by the dotted line; (c) Cdf’s of NPV for the initial well placements and the
optimal well placement after primary and secondary optimization.

10
No. of violated constraints
Index of the risk realization
8

0
0 500 1000 1500 2000 2500
No. of Simulations

Figure 5.25: Number of violated constraints and indicies of the risk realization during sec-
ondary optimization of the bi-objective optimization of slanted well trajectories, Oseberg.

201
 (a) Initial - Realization #1 (b) Robust Opt. - Realization (c) Bi-obj Opt. - Realization #1
#1

(d) Initial - Realization #7 (e) Robust Opt. - Realization (f) Initial - Realization #7
#7

Figure 5.26: Oil saturation in layer 10 after 3,900 days of production obtained with initial well
locations, and optimal well locations after robust optimization and bi-objective optimization,
for realization #1 and #7, Oseberg.

5.6 Summary and Discussion

In this chapter, we proposed parameterizations for both slanted wells and multi-
segmented wells, and investigated the applicability of StoSAG and GPS for deterministic
and robust well placement optimization considering the minimum well spacing constraints
and the well length constraints. The augmented Lagrangian method is adopted to handle the
nonlinear constraints. The necessity of using complicated well trajectory is also investigated
based on a box-shaped channelized reservoir. When geological uncertainty exists, StoSAG
and GPS are also coupled with lexicographic method to maximize the average NPV and
maximize the minimum NPV of the ensemble of realizations representing the uncertainty.
An efficient implementation of the StoSAG is also developed to reduce the computational
cost of the secondary optimization. The effectiveness of the bi-objective optimization is
validated on one water-flooding example, the channelized reservoir, and one gas-flooding
example, the Oseberg example.

202
 CHAPTER 6

DISCUSSION AND CONCLUSIONS

In this work, we developed and applied a hybrid algorithm which combines the Genetic
Algorithm (GA), stochastic simplex approximate gradient (StoSAG), and General Pattern
Search (GPS) to address three problems of interest in field development optimization:

• (i) selecting wells from a given set of potential paths and determining their types
(injectors or producers), drilling order and locations,

• (ii) determining the well status (drill or not-to-drill), types (injectors or producers)
and locations for both deterministic and robust optimization,

• (iii) and considering bi-objective well placement optimization using a complicated well
trajectory parameterization.

In virtually all cases, well control optimization is done subsequent to the optimization of the
other design variables.
For the first problem, this study revealed that (i) GA can effectively solve the prob-
lem of selecting a fixed number of wells from a given set of potential locations, moreover,
for this problem GA gives higher optimal NPV compared to a gradient-based method by
[25] and a StoSAG method using priority parameterization proposed by [44]; (ii) sequential
optimization where the well locations and types are optimized in the first stage followed by
a second stage of drilling order optimization gives a value of NPV that is comparable to
or higher than the NPV generated with simultaneous optimization of well locations, types
and drilling order; (iii) the optimization of well locations, types and drilling order with fixed
BHP’s followed by the well control optimization gives the highest optimal NPV if we operate

203
injectors at their maximum BHP and producers at their minimum BHP when optimizing
the well locations, types and drilling order.
For the second problem considered, we developed an iterative sequential algorithm
(ItSA) where we alternate optimizing well status and types using GA with pure binary
encoding and optimizing locations of wells using StoSAG or GPS if a StoSAG step fails to
improve the NPV, and a sequential algorithm (SA) where we optimize well status, types
and locations using GA first and then optimize the locations of wells drilled using StoSAG
followed by GPS. ItSA and SA result in comparable values of the optimal NPV, but of these
two algorithms, ItSAG converges approximately 20% faster for deterministic optimization.
The performance of both algorithms are compared with that of Genocop III and the results
of deterministic optimization indicate that Genocop III costs similar amount of simulation
runs as SA, but gives the worst optimal NPV (6% and 7% lower than ItSA and SA for the
PUNQ model respectively, and 1.5% and 3.3% lower than ItSA and SA for the Brugge model
respectively). The ItSA coupled with an adaptive subset model selection procedure can
efficiently optimize the well status, types and locations given a large number of realizations
of reservoir models. In robust optimization where the computation cost becomes expensive,
results indicate that it is more computationally efficient to eliminate StoSAG from the hybrid
optimization algorithm when estimating the optimal value of well status, type and location
design variables, i.e., simply combine GA and GPS in the hybrid algorithm.
For the third problem considered, we proposed a parameterization for slanted and
mult-segmented well trajectories and found that for box-shaped reservoirs, there is not a
significant gain in NPV after optimization by parameterizing wells as multi-segmented wells
as opposed to slanted wells and that well placement optimization using GPS mehtod can
still improve the NPV significantly after using the StoSAG method. We also found that the
lexicographic method combined with a modified StoSAG/GPS algorithm can significantly
reduce the downside risk with negligible deterioration in the the expected (average) NPV
that is obtained by only maximizing the average NPV of life-cycle production.

204
 BIBLIOGRAPHY

[1] A. A. Awotunde. Inclusion of well schedule and project life in well placement opti-
mization. In SPE Nigeria Annual International Conference and Exhibition. Society of
Petroleum Engineers, 2014.

[2] A. A. Awotunde. On the joint optimization of well placement and control. In SPE Saudi
Arabia Section Technical Symposium and Exhibition. Society of Petroleum Engineers,
2014.

[3] A. A. Awotunde. Generalized field-development optimization with well-control zonation.

Computational Geosciences, 20(1):213–230, 2016.

[4] J. E. Baker. Reducing bias and inefficiency in the selection algorithm. In Proceedings
of the second international conference on genetic algorithms, volume 206, pages 14–21,
1987.

[5] W. Bangerth, H. Klie, M. Wheeler, P. Stoffa, and M. Sen. On optimization algorithms

for the reservoir oil well placement problem. Computational Geosciences, 10(3):303–319,
2006.

[6] B. Beckner and X. Song. Field development planning using simulated annealing-optimal
economic well scheduling and placement. In SPE annual technical conference and exhi-
bition. Society of Petroleum Engineers, 1995.

[7] M. C. Bellout, D. E. Ciaurri, L. J. Durlofsky, B. Foss, and J. Kleppe. Joint optimization

of oil well placement and controls. Computational Geosciences, 16(4):1061–1079, 2012.

[8] A. C. Bittencourt and R. N. Horne. Reservoir development and design optimization. In

SPE annual technical conference and exhibition. Society of Petroleum Engineers, 1997.

205
 [9] B. Chen and A. C. Reynolds. Ensemble-based optimization of the water-alternating-
gas-injection process. SPE Journal, 2015.

[10] C. Chen. Adjoint-gradient-based production optimization with the augmented Lagrangian

method. PhD. thesis, University of Tulsa, Tulsa, Oklahoma, USA, 2011.

[11] C. Chen, G. Li, and A. Reynolds. Robust constrained optimization of short and long-
term npv for closed-loop reservoir management. spe 141314. Proceeding SPE RSS,
Houston, TX, USA, 21–23 February, 2011.

[12] Y. Chen, D. S. Olive, and D. Zhang. Efficient ensemble-based closed-loop production

optimization. In Proceedings of the SPE Improved Oil Recovery Symposium, number
SPE 112873, 2008.

[13] Y. Chen and D. Oliver. Ensemble-based closed-loop applied to Brugge field. In Pro-
ceedings of the SPE Reservoir Simulation Symposium, number SPE 118926, 2009.

[14] D. E. Ciaurri, T. Mukerji, and L. J. Durlofsky. Derivative-free optimization for oil

field operations. In Computational Optimization and Applications in Engineering and
Industry, pages 19–55. Springer, 2011.

[15] A. R. Conn, K. Scheinberg, and L. N. Vicente. Introduction to derivative-free optimiza-

tion, volume 8. Siam, 2009.

[16] R. J. de Moraes, R.-M. Fonseca, M. A. Helici, A. W. Heemink, and J. D. Jansen.

An efficient robust optimization workflow using multiscale simulation and stochastic
gradients. Journal of Petroleum Science and Engineering, 172:247–258, 2019.

[17] S. Ding, R. Lu, Y. Xi, S. Wang, and Y. Wu. Well placement optimization using di-
rect mapping of productivity potential and threshold value of productivity potential
management strategy. Computers & Chemical Engineering, 121:327–337, 2019.

[18] Y. Ding. Optimization of well placement using evolutionary methods. In Europec/EAGE

Conference and Exhibition. Society of Petroleum Engineers, 2008.

206
[19] S. T. Do and A. C. Reynolds. Theoretical connections between optimization algorithms
based on an approximate gradient. Computational Geosciences, 17(6):959–973, 2013.

[20] L. Durlofsky, I. Aitokhuehi, V. Artus, B. Yeten, and K. Aziz. Optimization of advanced

well type and performance. In ECMOR IX-9th European Conference on the Mathematics
of Oil Recovery, 2004.

[21] A. A. Emerick, E. Silva, B. Messer, L. Almeida, D. Szwarcman, M. Pacheco, and

M.Vellasco. Well placement optimization using a genetic algorithm with nonlinear
constraints. In Proceedings of the SPE Reservoir Simulation Symposium, number
SPE 118808, 2009.

[22] G. M. V. Essen, P. M. J. V. den Hof, and J. D. Jansen. Hierarchical long-term and

short-term production optimization. SPE Journal, 16(1):191–199, March 2011.

[23] R. Fletcher, S. Leyffer, and P. L. Toint. A brief history of filter methods. Preprint
ANL/MCS-P1372-0906, Argonne National Laboratory, Mathematics and Computer Sci-
ence Division, page 36, 2006.

[24] R. R.-M. Fonseca, B. Chen, J. D. Jansen, and A. Reynolds. A stochastic simplex

approximate gradient StoSAG for optimization under uncertainty. International Journal
for Numerical Methods in Engineering, 109(13):1756–1776, 2017.

[25] F. Forouzanfar, W. E. Poquioma, and A. C. Reynolds. Simultaneous and sequential esti-

mation of optimal placement and controls of wells using a covariance matrix adaptation
algorithm. SPE Journal, 2015.

[26] F. Forouzanfar and A. Reynolds. Joint optimization of number of wells, well locations
and controls using a gradient-based algorithm. Chemical Engineering Research and
Design, 92(7):1315–1328, 2014.

[27] F. Forouzanfar and A. C. Reynolds. Well-placement optimization using a derivative-free

method. Journal of Petroleum Science and Engineering, 109:96–116, 2013.

207
[28] F. Forouzanfar, E. D. Rossa, R. Russo, and A. C. Reynolds. Life-cycle production opti-
mization of an oil field with an adjoint-based gradient approach. Journal of Petroleum
Science and Engineering (online first), 2013.

[29] D. E. Goldberg and R. Lingle. Alleles, loci, and the traveling salesman problem. In
Proceedings of an international conference on genetic algorithms and their applications,
volume 154, pages 154–159. Lawrence Erlbaum, Hillsdale, NJ, 1985.

[30] R. Hanea, P. Casanova, F. H. Wilschut, and R. Fonseca. Well trajectory optimiza-

tion constrained to structural uncertainties. In SPE Reservoir Simulation Conference.
Society of Petroleum Engineers, 2017.

[31] R. Hanea, R. Fonseca, C. Pettan, M. Iwajomo, K. Skjerve, L. Hustoft, A. Chitu, and

F. Wilschut. Decision maturation using ensemble based robust optimization for field
development planning. In ECMOR XV-15th European Conference on the Mathematics
of Oil Recovery, 2016.

[32] T. Humphries and R. Haynes. Joint optimization of well placement and control for
nonconventional well types. Journal of Petroleum Science and Engineering, 126:242–
253, 2015.

[33] T. D. Humphries, R. D. Haynes, and L. A. James. Simultaneous and sequential ap-

proaches to joint optimization of well placement and control. Computational Geo-
sciences, 18(3-4):433–448, 2014.

[34] O. J. Isebor, L. J. Durlofsky, and D. E. Ciaurri. A derivative-free methodology with local

and global search for the constrained joint optimization of well locations and controls.
Computational Geosciences, 18(3-4):463–482, 2014.

[35] O. J. Isebor, D. Echeverrı́a Ciaurri, and L. J. Durlofsky. Generalized field-development

optimization with derivative-free procedures. SPE Journal, 19(05):891–908, 2014.

208
[36] J. D. Jansen, D. R. Brouwer, G. Naevdal, and C. P. J. W. van Kruijsdijk. Closed-loop
reservoir management. First Break, 23:43–48, 2005.

[37] M. Jesmani, B. Jafarpour, M. Bellout, R. Hanea, and B. Foss. Application of simul-

taneous perturbation stochastic approximation to well placement optimization under
uncertainty. In ECMOR XV-15th European Conference on the Mathematics of Oil Re-
covery, 2016.

[38] Y. Jiang, H. Kautz, and B. Selman. Solving problems with hard and soft constraints us-
ing a stochastic algorithm for max-sat. In 1st International Joint Workshop on Artificial
Intelligence and Operations Research, page 20, 1995.

[39] A. Karim, M. Ghazali, A. Raub, and M. R. Bin. Optimizing development strategy

and maximizing field economic recovery through simulation opportunity index. In
SPE Reservoir Characterisation and Simulation Conference and Exhibition. Society of
Petroleum Engineers, 2011.

[40] B. Kayode, M. R. Yaacob, and F. A. Abdullah. Connected reservoir regions map created
from time-lapse pressure data shows similarity to other reservoir quality maps in a
heterogeneous carbonate reservoir. In International Petroleum Technology Conference.
International Petroleum Technology Conference, 2019.

[41] R. A. Khan and A. A. Awotunde. Determination of vertical/horizontal well type from

generalized field development optimization. Journal of Petroleum Science and Engi-
neering, 162:652–665, 2018.

[42] J. Kraaijevanger, P. Egberts, J. Valstar, and H. Buurman. Optimal waterflood de-

sign using the adjoint method. In SPE Reservoir Simulation Symposium. Society of
Petroleum Engineers, 2007.

[43] J. Lee, C. Park, J. M. Kang, and C. K. Jeong. Horizontal well design incorporated with
interwell interference, drilling location, and trajectory for the recovery optimization. In
SPE/EAGE Reservoir Characterization & Simulation Conference, 2009.

209
[44] O. Leeuwenburgh, A. Chitu, R. Nair, P. Egberts, L. Ghazaryan, T. Feng, and L. Hustoft.
Ensemble-based methods for well drilling sequence and time optimization under uncer-
tainty. In ECMOR XV-15th European Conference on the Mathematics of Oil Recovery,
2016.

[45] O. Leeuwenburgh, P. J. Egberts, and O. A. Abbink. Ensemble methods for reservoir

life-cycle optimization and well placement. In SPE/DGS Saudi Arabia Section Technical
Symposium and Exhibition. Society of Petroleum Engineers, 2010.

[46] L. Li and B. Jafarpour. A variable-control well placement optimization for improved

reservoir development. Computational Geosciences, 16(4):871–889, 2012.

[47] L. Li, B. Jafarpour, and M. R. Mohammad-Khaninezhad. A simultaneous perturbation

stochastic approximation algorithm for coupled well placement and control optimization
under geologic uncertainty. Computational Geosciences, 17(1):167–188, 2013.

[48] X. Liu. Multiobjective optimization of a waterflood with gradient-based algorithms. The

University of Tulsa, 2016.

[49] X. Liu and A. C. Reynolds. Gradient-based multi-objective optimization with applica-

tions to waterflooding optimization. Computational Geosciences, pages 1–17, 2014.

[50] X. Liu and A. C. Reynolds. Multiobjective optimization for maximizing expectation

and minimizing uncertainty or risk with application to optimal well control. In SPE
Reservoir Simulation Symposium. Society of Petroleum Engineers, 2015.

[51] F. G. Lobo and D. E. Goldberg. The parameter-less genetic algorithm in practice.

Information Sciences, 167(1-4):217–232, 2004.

[52] R. Lu, F. Forouzanfar, and A. Reynolds. Bi-objective optimization of well placement and
controls using StoSAG. In SPE Reservoir Simulation Conference. Society of Petroleum
Engineers, 2017.

210
[53] R. Lu, F. Forouzanfar, and A. C. Reynolds. An efficient adaptive algorithm for robust
control optimization using StoSAG. Journal of Petroleum Science and Engineering,
159:314–330, 2017.

[54] R. Lu and A. C. Reynolds. Joint optimization of well locations, types, drilling order and
controls given a set of potential drilling paths. In SPE Reservoir Simulation Conference.
Society of Petroleum Engineers, 2019.

[55] I. Manual. Version x-2009.10. Computer Modelling Group (CMG), Calgary, Canada,
2009.

[56] C. Maschio, L. Nakajima, and D. J. Schiozer. Production strategy optimization us-

ing genetic algorithm and quality map. In Europec/EAGE Conference and Exhibition.
Society of Petroleum Engineers, 2008.

[57] Z. Michalewicz and G. Nazhiyath. Genocop III: A co-evolutionary algorithm for numer-
ical optimization problems with nonlinear constraints. In Proceedings of 1995 IEEE In-
ternational Conference on Evolutionary Computation, volume 2, pages 647–651. IEEE,
1995.

[58] R. Mitchell and S. Miska. Fundamentals of drilling engineering. Society of Petroleum

Engineers, 2011.

[59] G. Montes, P. Bartolome, and A. L. Udias. The use of genetic algorithms in well
placement optimization. In SPE Latin American and Caribbean petroleum engineering
conference. Society of Petroleum Engineers, 2001.

[60] A. N. Morales, H. Nasrabadi, and D. Zhu. A new modified genetic algorithm for well
placement optimization under geological uncertainties. In SPE EUROPEC/EAGE an-
nual conference and exhibition. Society of Petroleum Engineers, 2011.

211
[61] H. Mühlenbein and D. Schlierkamp-Voosen. Predictive models for the breeder genetic
algorithm I. continuous parameter optimization. Evolutionary computation, 1(1):25–49,
1993.

[62] J. Nocedal and S. J. Wright. Numerical Optimization. Springer, New York, 2006.

[63] D. F. Oliveira and A. C. Reynolds. Hierarchical multiscale methods for life-cycle-

production optimization: A field case study. SPE Journal, 20(05):896–907, 2015.

[64] D. S. Oliver. Incorporation of transient pressure data into reservoir characterization. In

Situ, 18(3):243–275, 1994.

[65] J. E. Onwunalu and L. J. Durlofsky. Development and application of a new well pattern
optimization algorithm for optimizing large scale field development. In SPE Annual
Technical Conference and Exhibition. Society of Petroleum Engineers, 2009.

[66] J. E. Onwunalu and L. J. Durlofsky. Application of a particle swarm optimization

algorithm for determining optimum well location and type. Computational Geosciences,
14(1):183–198, 2010.

[67] D. Orvosh and L. Davis. Shall we repair? genetic algorithms combinatorial optimization
and feasibility constraints. In Proceedings of the 5th International Conference on Genetic
Algorithms, page 650. Morgan Kaufmann Publishers Inc., 1993.

[68] D. Peaceman. Representation of a horizontal well in numerical reservoir simulation.

SPE Advanced Technology Series, 1(01):7–16, 1993.

[69] D. W. Peaceman. Interpretation of well-block pressures in numerical reservoir simulation

(includes associated paper 6988). Society of Petroleum Engineers Journal, 18(03):183–
194, 1978.

[70] D. W. Peaceman. Interpretation of well-block pressures in numerical reservoir simulation

with nonsquare grid blocks and anisotropic permeability. Society of Petroleum Engineers
Journal, 23(03):531–543, 1983.

212
[71] L. Peters, R. Arts, G. Brouwer, C. Geel, S. Cullick, R. Lorentzen, Y. Chen, K. Dunlop,
F. Vossepoel, R. Xu, P. Sarma, A. Alhuthali, and A. Reynolds. Results of the Brugge
benchmark study for flooding optimisation and history matching. SPE Reservoir Eval-
uation & Engineering, 13(3):391–405, 2010.

[72] D. D. Salam, I. Gunardi, and A. Yasutra. Production optimization strategy using hybrid
genetic algorithm. In Abu Dhabi International Petroleum Exhibition and Conference.
Society of Petroleum Engineers, 2015.

[73] P. Sarma, K. Aziz, and L. J. Durlofsky. Implementation of adjoint solution for optimal
control of smart wells. In SPE Reservoir Simulation Symposium. Society of Petroleum
Engineers, 2005.

[74] P. Sarma and W. H. Chen. Efficient well placement optimization with gradient-based
algorithms and adjoint models. In Intelligent energy conference and exhibition. Society
of Petroleum Engineers, 2008.

[75] M. Schoenauer and Z. Michalewicz. Evolutionary computation at the edge of feasibility.

In International Conference on Parallel Problem Solving from Nature, pages 245–254.
Springer, 1996.

[76] M. G. Shirangi and L. J. Durlofsky. Closed-loop field development optimization under

uncertainty. In SPE Reservoir Simulation Symposium. Society of Petroleum Engineers,
2015.

[77] N. Sibaweihi, A. A. Awotunde, A. S. Sultan, and H. Y. Al-Yousef. Sensitivity studies and

stochastic optimization of CO2 foam flooding. Computational Geosciences, 19(1):31–47,
2015.

[78] D. Simon. Evolutionary optimization algorithms. John Wiley & Sons, 2013.

[79] V. Torczon. On the convergence of pattern search algorithms. SIAM Journal on opti-
mization, 7(1):1–25, 1997.

213
[80] R. R. Torrado, G. D. Paola, A. F. Perez, A. A. Fuenmayor, M. S. De Azevedo, and
S. Embid. Optimize a WAG field development plan, use case of carbonate ultra-deep
water reservoir. In Proceedings of EUROPEC 2015, Madrid, Spain, 01-04 June, number
SPE 174344, 2015.

[81] S. Vlemmix, G. J. Joosten, R. Brouwer, and J. D. Jansen. Adjoint-based well trajectory

optimization. In EUROPEC/EAGE Conference and Exhibition. Society of Petroleum
Engineers, 2009.

[82] O. Volkov and M. Bellout. Gradient-based constrained well placement optimization.

Journal of Petroleum Science and Engineering, 2018.

[83] C. Wang, G. Li, and A. C. Reynolds. Optimal well placement for production optimiza-
tion. SPE Journal, 14(3):506–523, 2009.

[84] X. Wang, R. D. Haynes, and Q. Feng. Well control optimization using derivative-free
algorithms and a multiscale approach. arXiv preprint arXiv:1509.04693, 2015.

[85] C. Yang, C. Card, L. X. Nghiem, and E. Fedutenko. Robust optimization of SAGD

operations under geological uncertainties. In SPE Reservoir Simulation Symposium.
Society of Petroleum Engineers, 2011.

[86] B. Yeten, L. J. Durlofsky, and K. Aziz. Optimization of nonconventional well type,

location and trajectory. In SPE annual technical conference and exhibition. Society of
Petroleum Engineers, 2002.

[87] M. Zandvliet, M. Handels, G. van Essen, R. Brouwer, and J. D. Jansen. Adjoint-based

well placement optimization under production constraints. SPE Journal, 13(04):392–
399, 2008.

[88] F. Zhang. Automatic History Matching of Production Data for Large Scale Problems.
Ph.D. thesis, The University of Tulsa, Tulsa, Oklahoma, 2002.

214
[89] K. Zhang, G. Li, A. C. Reynolds, J. Yao, and L. Zhang. Optimal well placement using
an adjoint gradient. Journal of Petroleum Science and Engineering, 73(3-4):220–226,
2010.

[90] Y. Zhang, R. Lu, F. Forouzanfar, and A. C. Reynolds. Well placement and control
optimization for WAG/SAG processes using ensemble-based method. Computers &
Chemical Engineering, 101:193–209, 2017.

215
 APPENDIX A

MINIMUM WELL SPACING CONSTRAINT

The enforcement of the minimum well spacing constraint requires the calculation of
the distance between two line segments in 3D. The procedure for performing this calculation
is presented here. Consider two line segments, namely AB and CD, in 3D space. The
coordinates of points A, B, C, D are denoted by xA , xB , xC and xD , respectively.
Suppose that P is a point on the line through A and B, AB, then the coordinates xP
of point P can be written as

xP = xA + sP (xB − xA ) . (A.1)

If 0 6 sP 6 1, point P is on the line segment AB. If sP < 0 or sP > 1, point P is on the

extended line of the line segment AB. Similarly, if Q is a point on the line through C and D,
CD, then the coordinates xQ of point Q can be written as

xQ = xC + sQ (xD − xC ) . (A.2)

Note Q is a point on the line segment CD if and only if 0 ≤ sQ ≤ 1. The distance between
the points P and Q is
|P Q| = kxP − xQ k2 , (A.3)

and the distance between the two lines is defined as the minimum value of |P Q|. Let

f (sP , sQ ) = kxP − xQ k22 . (A.4)

216
If f (sP , sQ ) is equal to the distance between the two lines AB and CD, then sP and sQ must
satisfy the following two equations,

 ∂f (sP ,sQ )


∂sP
= 0,
(A.5)
 ∂f (sP ,sQ )


∂sQ
= 0.

If the solution obtained for Eq. A.5 satisfies 0 6 sP 6 1 and 0 6 sQ 6 1, then the point
P is on the line segment AB and the point Q is on the line segment CD, and the distance
(calculated with Eq. A.3) between line AB and CD is the same as the distance between line
segments AB and CD, i.e., shortest distance between a pair of points, P and Q with P on
the line segment AB and Q on the line segment CD. Otherwise, we need to calculate the
distance between the point A and the line segment CD, the distance between the point B
and the line segment CD, the distance between the point C and the line segment AB and the
distance between point D and the line segment AB. Then the distance between line segment
AB and line segment CD is the smallest one of these four distances.
In the following, we briefly explain the computation of the distance of the point, P, to
the line segment, AB, in 3D. The following equation defines the plane that passes through
the point P and is perpendicular to the line AB.

(xB − xA )T (x − xP ) = 0. (A.6)

Let Q be the intersection point of the plane and line AB. It is obvious that PQ is per-
pendicular to AB, and the point Q is the perpendicular foot whose coordinates are given
by
xQ = xA + sQ (xB − xA ) . (A.7)

Substituting Eq. A.7 into Eq. A.6 and solving the resulting equation for sQ gives

(xA − xB )T (xA − xP )
sQ = . (A.8)
kxA − xB k22

217
If 0 6 sQ 6 1, the point Q is on the line segment AB, and the length of line segment PQ is
by definition the distance between the point P and the line segment AB. If sQ < 0, the point
Q is on the extended line of line segment BA, and the distance between the point P and the
line segment AB is the distance between the point P and point A. If sQ > 1, the point Q is
on the extended line of line segment AB, and the distance between the point P and the line
segment AB is the distance between the point P and point B.

218
 APPENDIX B

MULTI-SEGMENTED WELL TRAJECTORY CALCULATION

In Chapter 5 of this dissertation, the trajectory of a multi-segmented well (e.g., well

iw) is parameterized using Eq. 5.4, which is recited below

Xiw = [xiw iw iw iw iw iw iw iw iw iw iw iw
0 , y0 , z0 , z1 , l1 , θ1 , z2 , l2 , δθ2 , · · · , zNseg , lNseg , δθNseg ].

To obtain a multi-segmented well trajectory given an estimate of the variables in Eq. 5.4,
we need to calculate the coordinates of the end points of each well segment. The calculation
of the segment end points of the multi-segmented well iw with Nseg segments in Cartesian
coordinate system is presented in Algorithm 13.

Algorithm 13 Calculation of the segment end points of the multi-segmented well iw

• For i = 1 : Nseg

– Calculate the azimuth angle of the ith segment, θiiw . If i > 1, θiiw = θi−1
iw
+ δθiiw .
– Calculate δziiw = ziiw − zi−1
iw
.
– Calculate the horizontal departure of the ith segment using
q
iw 2 2
HDi = (liiw ) − (δziiw ) .

– Calculate the (x, y) coordinates of end point of ith segment using

xiw iw iw iw
i = xi−1 + HDi cos(θi ),
yiiw = yi−1
iw
+ HDiiw sin(θiiw ).

– Set i = i + 1.

• EndFor

219
 To apply the optimal estimate of Eq. 5.4, the multi-segmented trajectory obtained
with Algorithm 13 has to be smoothed in order to be drillable. In this dissertation, each pair
of consecutive well segments of well iw is smoothed as a circular arc located in the plane of
these two segments; see Fig. 5.2 for an illustration. The smoothing procedure of well iw is
presented in Algorithm 14.

Algorithm 14 Smoothing of the multi-segmented well iw

• For i = 1 : Nseg − 1

– Calculate the coordinates of the end points of the ith and (i + 1)th segments.
– Calculate the coordinates of the starting point, B, and the end point, D, of the
smoothing arc, BD.
d
If (liiw >li+1
iw
)
iw
li+1
∗ s= 2liiw
,
∗ xiw
B = xiw iw iw iw iw iw
i − s(xi − xi−1 ), yB = yi − s(yi − yi−1 ),
iw
zBiw = ziiw − s(ziiw − zi−1
iw
),
xiw +xiw y iw +y iw z iw +z iw
∗ xiw
D = i 2 i+1 , yD iw
= i 2 i+1 , zDiw
= i 2 i+1 ,
else
liiw
∗ s= iw ,
2li+1
xiw +xiw y iw +y iw z iw +z iw
∗ xiw
B =
i
2
i−1
, yBiw = i 2 i−1 , zBiw = i 2 i−1 ,
∗ xiw iw iw iw iw iw iw iw iw iw iw iw
D = xi + s(xi+1 − xi ), yD = yi + s(yi+1 − yi ), zD = zi + s(zi+1 − zi ).

– Calculate the inclination angle of the ith and (i + 1)th segments,

 iw iw   iw iw 
zi −zi−1 zi+1 −zi
∗ ϕiw
i = 90−arcsin liw
, ϕiw
i+1 = 90−arcsin liw
, δϕiw iw iw
i = ϕi+1 −ϕi .
i i+1

– Calculate the dogleg angle βiiw and the radius of curvature Riiw of the arc BD,
d
r 2  2
iw δϕiw iw iw iw δθiiw
∗ βi = 2arcsin sin 2i
+ sinϕi sin(ϕi + δϕi ) sin 2 ,
min(liiw , li+1
iw )
∗ Riiw = β iw
.
2 tan i
2

– Calculate the dogleg severity,

100 180
∗ DLSiiw = Riiw π
.
– Set i = i + 1.

• EndFor

220
 APPENDIX C

EQUIVALENT WELL INDEX CALCULATION

In this dissertation, well trajectories can be parameterized as vertical, slanted or

multi-segmented. Each well can perforate many gridblocks. In CMG (the simulator used
in this work to simulate the reservoir dynamics) as well as other commercial simulators,
one can only define a perforation within one gridblock as a single line segment. Given
the Cartesian coordinates of the entry point and exit point to a gridblock, the perforation
length, the completion factor and the perforated gridblock indices, CMG can calculate an
accurate productivity/injectivity index using a Peaceman type well model ([69, 70, 68]). In
this section, we firstly provide a routine which identifies the grid cells crossed by each well
and calculates the required information of each perforated cell for CMG based on an estimate
of the design variables given in Eq. 5.1 or 5.4.
Note that for a multi-segmented well, there may be two segments of perforations in
the gridblock containing the end point of a well segment (except the last well segment), which
can not be specified in CMG. Thus the calculation of a one-segmented pseudo-perforation
with equal productivity/injectivity index is required and we provide the calculation based on
CMG [55] in Appendix C.2. In this dissertation, this routine is only developed for reservoirs
with (transformed) Cartesian grids.

C.1 Calculation of the Inputs for CMG

If well iw is a slanted well, we can obtain the coordinates of the two end points of
well iw directly from an estimate of the variables in Eq. 5.1. If well iw is a multi-segmented
well, we can obtain the coordinates of the end points of the Nseg well segments of well iw

221
from an estimate of the variables in Eq. 5.4 using Appendix B. Consider a set of Cartesian
grids given in Fig. C.1, we provide a routine which identifies the grid cells crossed by well iw
and calculates the Cartesian coordinates of the entry and exit points, the perforation length
and the completion factor of each perforation of well iw in Algorithm 15. Since a slanted
well can be considered as a multi-segmented well with a single well segment, this routine is
general for both slanted wells and multi-segmented wells. However, we emphasize that this
routine is only developed for Cartesian grids illustrated in Fig. C.1.
In Algorithm 15, we denote the number of gridblocks of the reservoir along (x, y, z) di-
rections respectively as NGridI , NGridJ and NGridK . We define xyz as a (NGridI +1)-dimensional
array where xyz
ig is set equal to the x coordinate of the left surface of gridblock (ig , 1, 1), for

ig = 1, 2, · · · , NGridI and xyz

NGridI +1 is set equal to the x coordinate of the right surface of grid-

block (NGridI , 1, 1). Similarly, we define y xz as a (NGridJ + 1)-dimensional array where yixz
g
is
set equal to the y coordinate of the back surface of gridblock (1, ig , 1) for ig = 1, 2, · · · , NGridJ
xz
and yNGridJ +1
is set equal to the y coordinate of the front surface of gridblock (1, NGridJ , 1).
We also define z xy as a (NGridK + 1)-dimensional array where zixy
g
is set equal to the z coor-
xy
dinate of the top surface of gridblock (1, 1, ig ) for ig = 1, 2, · · · , NGridK and zN GridK +1
is set
equal to the z coordinate of the bottom surface of the (1, 1, NGridK ) gridblock. We denote
the coordinates of the entry and exit points of the ip th perforation of the ith segment of well
iw respectively as (xiw,0 iw,0 iw,0 iw,1 iw,1 iw,1
i,ip , yi,ip , zi,ip ) and (xi,ip , yi,ip , zi,ip ). We also denote its perforation

iw iw
length and the completion factor respectively as li,i p
and fi,i p
.
In the following, we describe the procedure to determine the (i, j, k) indicies of a
perforated gridblock given the coordinates of a point. Take the point (xiw iw iw
i , yi , zi ) as an
yz
example. If xiw iw
i < xNGridI +1 , then the gridblock index along x-direction, GIi , is set equal

to {ig : xyz iw
i g ≤ xi < xyz NGridI iw
ig +1 }ig =1 ; otherwise, set GIi = NGridI . If yiiw < yN
xz
GridJ +1
, then
the gridblock index along y-direction, GJiiw , is set equal to {ig : yixz
g
≤ yiiw < yixz }NGridJ ;
g +1 ig =1

xy
otherwise, set GJiiw = NGridJ . If ziiw < zN GridK +1
, then the gridblock index along z-direction,
GKiiw , is set equal to {ig : zixy
g
≤ ziiw < zixy }NGridK ; otherwise, set GKiiw = NGridK .
g +1 ig =1

222
Figure C.1: Numbering of the Cartesian grids. Numbers in parenthesis on the top surface
of each gridblock indicate the (i, j, k) gridblock indicies.

Algorithm 15 Calculation of the input parameters into CMG given the Cartesian coordi-
nates of the segment end points of well iw

• For i = 1 : Nseg

– Calculate the Cartesian coordinates of the two end points of segment i,

i.e., (xiw iw iw iw iw iw
i−1 , yi−1 , zi−1 ) and (xi , yi , zi ), and their gridblock indcies, i.e.,

iw iw iw
(GIi−1 , GJi−1 , GKi−1 ) and (GIiiw , GJiiw , GKiiw ).

iw
– Define a Nstrial -dimensional vector s where Nstrial = |GIi−1 iw
− GIiiw | + |GJi−1 −
iw
GJiiw | + |GKi−1 − GKiiw | + 3. Set m = 1.

iw
– For ig = min(GIi−1 , GIiiw ) : max(GIi−1
iw
, GIiiw )
xyz iw
ig −xi−1
∗ If xiw iw
i 6= xi−1 , set sm = xiw iw and m = m + 1.
i −xi−1

– EndFor

iw
– For jg = min(GJi−1 , GJiiw ) : max(GJi−1
iw
, GJiiw )
yjxz iw
−yi−1
∗ If yiiw 6= yi−1
iw
, set sm = g
yiiw −yi−1
iw and m = m + 1.

– EndFor

223
•
iw
– For kg = min(GKi−1 iw
, GKiiw ) : max(GKi−1 , GKiiw )
zkxy iw
−zi−1
∗ If ziiw 6= zi−1
iw
, set sm = g
ziiw −zi−1
iw and m = m + 1.
– EndFor
– Remove si , i = m, m + 1, · · · , Nstrial from s and remove repeated elements in s.
Define A = {si : si ≤ 0 and si ∈ s}, B = {si : 0 < si < 1 and si ∈ s} and
C = {si : si ≥ 1 and si ∈ s}. Denote the dimensionality of B as NB and set
N piw
i = NB + 2.
p p
– Define a N piw p
i -dimensional vector c where c1 = max{A}, ci+1 = Bi , i =
p
1, 2, · · · , NB and cN piw = max{C}. Sort cp in ascending order.
i

– For ip = 1: N piw
i −1
∗ Calculate the coordinates of the entry and exit points of the ip th perforation
using

(xiw,0 iw,0 iw,0 iw iw iw p

 iw iw iw iw iw iw
i,ip , yi,ip , zi,ip ) = (xi−1 , yi−1 , zi−1 ) + cip ∗ (xi , yi , zi ) − (xi−1 , yi−1 , zi−1 )

(xiw,1 iw,1 iw,1 iw iw iw p

 iw iw iw iw iw iw
i,ip , yi,ip , zi,ip ) = (xi−1 , yi−1 , zi−1 ) + cip +1 ∗ (xi , yi , zi ) − (xi−1 , yi−1 , zi−1 )

∗ Calculate the perforation length using

  q
iw
li,i p
= min(cpip +1 , 1) − max(cpip , 0) ∗ (xiw iw 2 iw iw 2 iw iw 2
i − xi−1 ) + (yi − yi−1 ) + (zi − zi−1 )

iw
∗ Set the completion factor fi,i p
= 1.
∗ Calculate the perforation gridblock indices based on the coordinates
xiw,0 iw,1
i,ip +xi,ip
iw,0
yi,i p
iw,1
+yi,ip
iw,0
zi,ip
iw,1
+zi,i p
( 2
, 2
, 2
).
– EndFor
– If i > 1 and cpi,1 < 0, re-calculate the required information for a pseudo-perforation
in the gridblock of the first perforation of the ith well segment,
∗ reset the coordinates of the entry and exit points of the pseudo-perforation
respectively as (xiw,0 , y iw,0 , z iw,0
i−1,N piw −1 i−1,N piw −1 i−1,N piw −1
) and (xiw,1 iw,1 iw,1
i,1 , yi,1 , zi,1 ),
i i i
∗ re-calculate the length of the pseudo-perforation using
q
iw
li,i p
= (xiw,1 iw,0 2 iw,1 iw,0 2 iw,1 iw,0 2
i,1 − xi−1,N piw −1 ) + (yi,1 − yi−1,N piw −1 ) + (zi,1 − zi−1,N piw −1 ) ,
i i i

iw
∗ re-calculate the completion factor fi,1 of the pseudo-perforation using Ap-
pendix C.2 which is discussed later.

• EndFor

224
 C.2 Two-segmented Perforation within a Gridblock

When wells are parameterized as multi-segmented wells, it is common that two line
−→ −−→
segments of perforations exist in one simulation gird; see perforation AC and CB in Fig.
C.2 as an example. Since in most of the commercial simulators, one can only define the
perforation with one segment, a pseudo-perforation with equal injectivity/productivity is
required. Thus, the calculation of the coordinates of the entry and exit point, the perforation
length and the completion factor of this pseudo-perforation is presented here.

Figure C.2: Two-segmented perforation within one simulation gridblock.

Take Fig. C.2 as an example where point C is the end point of the ith well seg-
ment of well iw, point A is the entry point of the last perforation of the ith well seg-
ment and point B is the exit point of the first perforation of (i + 1)th well segment. De-
−→ iw,0 iw,0 iw,0 −−→
note perforation AC = (xiw iw iw
i , yi , zi ) − (xi,N piw −1 , yi,N piw −1 , zi,N piw −1 ) and perforation CB =
i i i
−→ −→
(xiw,1 iw,1 iw,1 iw iw iw iw,1
i+1,1 , yi+1,1 , zi+1,1 )−(xi , yi , zi ). A pseudo-perforation AB (AB = (xi+1,1 , yi+1,1 , zi+1,1 )−
iw,1 iw,1

−→ −−→
(xiw,0 , y iw,0 , z iw,0
i,N piw −1 i,N piw −1 i,N piw −1
)) is required instead of the true perforations AC and CB so that
i i i

the flow rate of phase p remains roughly the same, i.e.,

W I−→ λ ∆p−→ = W I−→ λp ∆p−→ + W I−

AB p AB AC AC
−→ λ ∆p−
CB p
−→,
CB
(C.9)

where p denotes oil, water and gas; λp is the mobility of phase p within the perforated grid-
lock; ∆p−→ , ∆p−→ and ∆p−
AB AC
−→ respectively represent the pressure draw-down of perforations
CB
−→ −→ −−→
AB, AC and CB, and W I− → , W I−→ and W I−
AB AC
→ respectively represent the well indices of
−
CB

225
 −→ −→ −−→
perforations AB, AC and CB. Assume ∆p−→ = ∆p−→ = ∆p−
AB AC
−→ , then Eq. C.9 reduces to
CB

W I−→ = W I−→ + W I−
AB AC
−→.
CB
(C.10)

−→ −−→ −→
The well index of a perforation l (e.g., l can be AC, CB or AB) is given by

βK(l)h(l)fcomp
WI = (C.11)
ln(re (l)/rw ) + s

where β is a coefficient involving the angular fraction and the unit conversion factor, K(l) is
the completion planar averaged permeability, h(l) is the perforation length (i.e. 2-norm of
the vector l), fcomp is the completion factor, re (l) is the the drainage radius in the direction
of the wellbore l, rw is the well radius and s is the skin factor. Note for a true perforation
−→ −−→ −→
(e.g., AC and CB), fcomp is always one. However, for a pseudo-perforation (e.g., AB), fcomp
has to be adjusted so that Eq. C.10 is satisfied.
−→ −−→ −→
Denote the perforation AC as l1 , CB as l2 and AB as l. Substituting Eq. C.11 into
Eq. C.10 gives

βK(l)h(l)fcomp (l) βK(l1 )h(l1 ) βK(l2 )h(l2 )

= + , (C.12)
ln(re (l)/rw ) + s ln(re (l1 )/rw ) + s ln(re (l2 )/rw ) + s

and then fcomp (l) can be solved as,

 
K(l1 )h(l1 ) K(l2 )h(l2 ) ln(re (l)/rw ) + s
fcomp (l) = + . (C.13)
ln(re (l1 )/rw ) + s ln(re (l2 )/rw ) + s K(l)h(l)

In the following, we will briefly discuss the calculation of the drainage radius and
the completion planar averaged permeability when a perforation does not parallel one of the
gridblock coordinate axes. Take perforation l as an example. According to CMG [55], K(l)
and re (l) are obtained by interpolating K’s and re ’s along the gridblock coordinate axes.
Denote i, j and k as unit vectors pointing in the local (x, y, z) directions. When the
wellbore is parallel to the D axis (D can be either i, j and k), the drainage radius re (D) is

226
calculated as s
V
re (D) = α , (C.14)
h(D)

where, α is a coefficient involving the geometric factor, the angular fraction and the unit
conversion factor, V is the bulk volume of the perforated gridblock, h(D) is the grid block
l·i l·j l·k
thickness in the direction D. Define θi = arccos( klk 2
), θj = arccos( klk 2
) and θk = arccos( klk 2
).
Then the value of the drainage radius in the direction of the wellbore is interpolated as
 
2 2 2 2 2 2
 re (i) cos (θi ) sin (θj ) sin (θk ) + re (j) cos (θj ) sin (θi ) sin (θk )
re (l) =   /S,
+re (k) cos2 (θk ) sin2 (θi ) sin2 (θj )

where, S = cos2 (θi ) sin2 (θj ) sin2 (θk )+cos2 (θj ) sin2 (θi ) sin2 (θk )+cos2 (θk ) sin2 (θi ) sin2 (θj ). Sim-
ilarly, the completion planar averaged permeability K is calculated by

√
 
2 2 2 2
p 2 2
Ky Kz cos (θi ) sin (θj ) sin (θk ) + Kx Kz cos (θj ) sin (θi ) sin (θk )
K(l) =   /S,

+ Kx Ky cos2 (θk ) sin2 (θi ) sin2 (θj ),
p

where Kx , Ky , Kz respectively represent the permeability of the perforated gridblock along

the (x, y, z) directions.

227