Li08 PDF

HIERARCHIC MODELING AND HISTORY MATCHING OF
MULTI-SCALE FLOW BARRIERS IN CHANNELIZED

RESERVOIRS
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ENERGY
RESOURCES ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Hongmei Li
August 2008
c Copyright by Hongmei Li 2008
All Rights Reserved
ii
I certify that I have read this dissertation and that, in my opinion, it
is fully adequate in scope and quality as a dissertation for the degree
of Doctor of Philosophy.
(Dr. Jef Caers) Principal Adviser
(Dr. Hamdi Tchelepi)
(Dr. Stephan Graham)
Approved for the University Committee on Graduate Studies.
iii
iv
Abstract
Deep-water channelized reservoirs often consist of multi-scale architectures contain-
ing large-scale channel belts, middle-scale individual channels and small-scale channel
inll components. These architectural elements are dened between bounding sur-
faces. Along these hierarchical bounding surfaces, thin shale drapes may be present as
ow barriers that compartmentalize the reservoirs. Because the distribution of shale
drapes is dominated by these bounding surfaces, they themselves form multi-scale
reservoir heterogeneity. The key problem addressed in this thesis arises in the situ-
ation where multi-scale shale drapes are present along channel, channel belt and/or
valley bounding surfaces, but the channel locations are uncertain or unknown. In
order to reduce this uncertainty, rst a realistic representation of the channel distrib-
ution should be obtained and constrained to hard data; then the channel and drape
locations should be calibrated to the production data.
We propose a coupled geologic modeling and history matching method where the
reservoir architecture composed of channels are simulated with pre-dened stack-
ing patterns, the shale drapes are then simulated along the bounding surfaces using
multiple-point statistics techniques; channel and hole locations are gradually per-
turbed until the corresponding ow responses match the eld production data. The
perturbation during the history matching honors the individual channel geometry
and the conceptual channel stacking patterns. In other words, the perturbation is
geologically consistent.
The reservoir architecture modeling in this work involves dening channel depo-
sition fairways (valleys) based on seismic data, modeling long sinuous channels and
placing them into dened fairways such that all data are matched. This work adopts
v
a stratigraphic-based modeling approach. In this stratigraphic-based approach, indi-
vidual channels are simulated using the YACS method (Alapetite et al., 2005). This
method is fast and conditions to well data under the assumption that channel sand
can be identied in the well data. To stack multiple channels reproducing the con-
ceptual stacking pattern model, the migration ratio and overlap ratio are used as
input parameters in the simulation process. To match the production data, channel
location is perturbed using the gradual deformation method, the continuity of shale
drapes is perturbed using the probability perturbation method.
Tests on a reservoir analog of deepwater conned-channel systems demonstrated
the feasibility of proposed modeling and history matching workow. The sensitivity
study showed that the hole proportion along individual channels and channel loca-
tion are the most important parameters to ow. The history matching results also
revealed that performing one step perturbation of channel location and hole propor-
tion is enough to obtain desired history-matched geologic models. The prediction
results have shown that the proposed workow can be used to generate geologic mod-
els with better prediction power than randomly selected geologic models.
Application of the proposed workow to a realistic deepwater conned-channel
reservoir showed that, in case of high NTG (07-08) reservoirs, both the scheme of
perturbing channel and hole location and the scheme of perturbing hole location
along xed channels can achieve history matching. However, the former perturbation
scheme is more ecient since the perturbation is in full 3D and consistent to geologi-
cal conceptual model. Case study also demonstrated that including prior information
of variability of shale drape leads to more realistic geologic models.
vi
Acknowledgements
First and foremost, I wish to express my deepest gratitude to my advisor, Prof. Jef
Caers, who has guided me over the years. Jef is one of the best teachers I have ever
had. He has been a consistent source of knowledge, advice and encouragement. I have
benetted greatly from his thoughtful guidance and many insightful suggestions. For
his careful reading of many drafts of this dissertation and other SCRF papers, for his
unwavering enthusiasm and support of this work, for his many contributions to my
education at Stanford, I owe him many thanks.
I would like to thank Prof. Andre Journel for his wisdom and for imparting some of
his knowledge to me through our many meetings and classes. It has been an honor
and a privilege to have learned Geostatistics from him.
This work would not have been possible without the contributions of Frans van der
Vlugt, Mark Barton, Carlos Pirmez and Omer Alpak from Shell. Shell provided
invaluable model dataset and constructive suggestions regarding channel stacking
pattern simulation and history matching.
I would like to thank Prof. Hamdi Tchelepi and Prof. Steve Graham for reading this
dissertation and providing valuable comments and suggestions. I am also thankful
to Prof. Roland Horne for being my committee member. Prof. Mark Zoback kindly
chaired my Ph.D. oral defense and is gratefully acknowledged.
I would like to express my gratitude to the SCRF research team for the valuable advice
and help provided and for making research a lot of fun. Dr. Scarlet Castro shared
with me her PPM framework written in python. Her help is greatly appreciated. My
thanks also go to Lisa Stright for providing me with details on shale drape modeling
and perturbation, and to Alex Bourcher and Ting Li for their help and guidance with
vii
SGEMS.
I owe my greatest gratitude to the Department of Energy Resources Engineering, all
the faculty, sta and my fellow students. In particular, I would like to thank Prof.
Margot Gerritsen who is not only providing excellent engineering courses, but also
hosting fun womens dinner. I am also thankful to Ginni Savalli and Thuy Nguyen.
They are always there to help me through all kinds of paper work. The SCAPE
friends have made my stay at Stanford truly pleasant. Their help and friendship are
very much appreciated.
Before I came to US, I was a geologist. My master thesis advisor Prof. Chris White
at Louisiana State University help me make a smooth transition from a pure geologist
to a reservoir engineer and led me into the door of geostatistic reservoir modeling.
Without his guidance and recommendation, I could not have had the chance to study
at Stanford. I feel lucky to have had Prof. Chris White as my master thesis adviser
at the very beginning of my engineer adventure, and I am deeply grateful to him.
Special thanks goes to my dear friends Wenjuan Lin and Junpeng Yue for their
friendships. Wenjuan shared with me the challenging experience of the rst year. I
thank her for always being there to support. Junpeng always encourage me and give
me condence to go through many tough situations. Each one of you is an amazing
friend and I have been fortunate to have met you.
Last but not least, I would like to thank my family. My husband Deqiang Wang has
been always there for me during the ups and downs. I thank him for his love, support
and encouragement. I would like to express my gratitude to my parents for their
unconditional love and condence in me, and my brother Shicheng for taking part of
my responsibility to take care of our parents.
viii
Contents
Abstract v
Acknowledgements vii
1 Introduction 1
1.1 Reservoir architecture modeling . . . . . . . . . . . . . . . . . . . . . 3
1.2 Flow barriers modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Geologically consistent history matching . . . . . . . . . . . . . . . . 11
1.4 Proposed modeling approach . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2 Proposed modeling workow 19
2.1 Reservoir architecture modeling . . . . . . . . . . . . . . . . . . . . . 20
2.1.1 Individual channel modeling . . . . . . . . . . . . . . . . . . . 21
2.1.2 Channel simulation parameters . . . . . . . . . . . . . . . . . 25
2.1.3 Channel stacking pattern modeling . . . . . . . . . . . . . . . 25
2.1.4 Well data conditioning . . . . . . . . . . . . . . . . . . . . . . 33
2.1.5 Shale drapes modeling . . . . . . . . . . . . . . . . . . . . . . 40
2.1.6 Summary of the architecture modeling . . . . . . . . . . . . . 43
2.2 Geologically consistent history matching . . . . . . . . . . . . . . . . 45
2.2.1 Channel stacking pattern perturbation . . . . . . . . . . . . . 45
2.2.2 Behavior of a chain of stacking pattern realizations . . . . . . 56
2.2.3 Shale drape perturbation . . . . . . . . . . . . . . . . . . . . . 60
2.2.4 Perturbation procedure for history matching . . . . . . . . . . 62
ix
3 Workow testing on a synthetic data set 67
3.1 Introduction to the synthetic dataset . . . . . . . . . . . . . . . . . . 68
3.2 Sensitivity study of shale drapes parameters . . . . . . . . . . . . . . 69
3.3 Modeling parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.4 History matching results . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.5 Chapter summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4 Applications to a realistic turbidite reservoir 97
4.1 Information available for history matching . . . . . . . . . . . . . . . 98
4.2 Setting up true cases . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.3 Perturbing shale drapes in high NTG reservoir . . . . . . . . . . . . . 107
4.4 Perturbing shale drapes by regions . . . . . . . . . . . . . . . . . . . 109
4.4.1 Region sensitivity study . . . . . . . . . . . . . . . . . . . . . 113
4.4.2 History matching results . . . . . . . . . . . . . . . . . . . . . 115
4.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5 Conclusions and future work 122
5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.2 Recommendations for future work . . . . . . . . . . . . . . . . . . . . 125
Bibliography 130
x
List of Tables
3.1 Seven factors and level set up . . . . . . . . . . . . . . . . . . . . . . 80
3.2 Plackett-Burman design with 7 factors . . . . . . . . . . . . . . . . . 80
3.3 Water cut(%) response @ 900 days for PB experiment runs and analysis 81
3.4 Water breakthrough time (days) for PB experiment runs and analysis 82
3.5 Field oil recovery eciency(%) @900 days for PB experiment runs and
analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.6 Global NTG for dierent channel belts of reservoir analog . . . . . . 85
4.1 NTG for dierent channel belts . . . . . . . . . . . . . . . . . . . . . 100
4.2 Variograms used for channel porosity and permeability simulation . . 101
4.3 Description of simulation model for true production simulation . . 105
4.4 Four factors and level set up . . . . . . . . . . . . . . . . . . . . . . . 115
4.5 The two level 2
41
design . . . . . . . . . . . . . . . . . . . . . . . . . 115
xi
List of Figures
1.1 Hierarchy for a deepwater channelized reservoir. Shale drapes could
be deposited on the channel belt margins, on channel margins within
individual channel belts and on cross-laminations within individual
channels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Schematic diagram showing the inuence of shale drapes on reservoir
connectivity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Example result from MPS. Where two channels cross, MPS can not
delineate the ne scale ow barriers that separate these two channels. 7
1.4 A synthetic 2D fracture training image (left, 400x300 grid) with 13%
fracture (blue lines) and one realization using snesim (right, 200x200
grid). The fracture patterns shown in training image are not repro-
duced in the simulated realization. . . . . . . . . . . . . . . . . . . . 8
1.5 A hierarchic workow to simulate multi-scale ow barriers. . . . . . . 17
1.6 Coupled geologic modeling and history matching of multi-scale ow
barriers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1 Denition of parameters used to describe the channel geometry . . . . 21
2.2 The workow for individual channel modeling . . . . . . . . . . . . . 24
2.3 The example showing curvilinear channel belt case . . . . . . . . . . . 24
2.4 The original potential map in Figure 2.3 showing the channel geometry
before adding noise map . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.5 Channel potential maps with dierent noise histogram variances . . . 26
xii
2.6 Channel potential maps with dierent noise variogram ranges: top
row is for ranges along channel orientation direction, bottom row is for
ranges perpendicular to this direction . . . . . . . . . . . . . . . . . . 26
2.7 The potential maps (bottom) and their corresponding original poten-
tial maps (top) generated with dierent potential gradients . . . . . . 27
2.8 Two parameters used for channel stacking pattern modeling . . . . . 29
2.9 Schematic graph showing channel stacking patterns with dierent pat-
tern parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.10 A channel belt body (left) and its bounding surface (right). Channels
will be lled into the space within the channel belt limit . . . . . . . 29
2.11 The channel stacking pattern modeling process showing how channel
is lled into belt (continue to next page) . . . . . . . . . . . . . . . . 30
2.12 The channel stacking pattern modeling process showing how channel
is lled into belt. Left column is the architecture model, middle col-
umn are the realizations for migration ratio, and right column are the
realizations for overlap ratio . . . . . . . . . . . . . . . . . . . . . . . 31
2.13 Three channel belts case. Individual channels are lled into each belt
until its net-to-gross reached. All the channels are conned by channel
belt limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.14 Channel bounding surfaces extracted from Figure2.11 . . . . . . . . . 32
2.15 Schematic graph showing the well facies data and their interpreted
channel sections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.16 Synthetic uniform distributions of two pattern parameters . . . . . . 35
2.17 Synthetic example demonstrating the stochastic interpretation of well
facies data. The same well data can result in multiple channel section
stacking pattern realizations . . . . . . . . . . . . . . . . . . . . . . . 36
2.18 Schematic example showing the compatible/incompatible interpreta-
tion when multiple wells present . . . . . . . . . . . . . . . . . . . . . 37
2.19 Well channel section data (left) and its original potential map (right).
The channel section geometry-maximum width and thickness- should
be the same as the dened channel cross section geometry . . . . . . 38
xiii
2.20 Well conditioning for interpreted channel sections. Well 1,2,3 could be
connected with one channel (left); However, well 3,4,5 can be connected
within another channel (right). Hence two interpretations are possible. 39
2.21 Three conditional realizations (bottom) of channel potential condi-
tioned to 10 wells (top) . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.22 Three conditional realizations (bottom) of channel potential condi-
tioned to 20 wells (top) . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.23 One realization (right column) of a channel complex conditioned to
interpreted channel stacking pattern (left column top) at well location 42
2.24 One example of shale drape training image. Red color represents holes,
and blue color indicates shales . . . . . . . . . . . . . . . . . . . . . . 44
2.25 The workow of shale drape modeling . . . . . . . . . . . . . . . . . . 44
2.26 Dierent noise maps (middle row) added to the same original potential
map (top row) result in dierent channel geometries (0-potential line
in bottom row) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.27 Same noise maps (top row) added to dierent original potential map
(middle row) result in dierent channel geometries (0-potential line in
bottom row) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.28 Channel pattern parameters sampled from their distribution function 49
2.29 A uniform distribution is transferred into Gaussian distribution . . . 50
2.30 Two Gaussian realizations are perturbed to generate one new realization 51
2.31 Transfer the perturbed realization back to uniform distribution . . . . 51
2.32 A chain of channel realizations with dierent perturbations applied to
the pair of channel stacking pattern parameters . . . . . . . . . . . . 52
2.33 Four realizations of individual channel potential maps with dierent
perturbation parameters applied to their noise maps. Note the channel
centerlines are roughly same. . . . . . . . . . . . . . . . . . . . . . . . 54
2.34 Channel complex realizations with dierent perturbation parameters
applied to the individual noise maps. Note the channel locations are
roughly the same (left column) but the channel cross-section changes
with dierent perturbations (right column, y=20). . . . . . . . . . . . 55
xiv
2.35 Channel realization (a) and the shale drapes (b) along channel bound-
aries (blue color represents shale drapes and red color is for scour holes);
wells are located in the reservoir thickness map(c); (d) is the plot of
water cut for two producers. . . . . . . . . . . . . . . . . . . . . . . . 57
2.36 Channel realization (a) independently generated from the one shown
in Figure 2.35 and the shale drapes (b) along channel boundaries (blue
color represents shale drapes and red color is for scour holes); (c) is
the reservoir thickness map, wells are located at the same positions as
shown in Figure 2.25c; (d) is the plot of water cut for two producers. 58
2.37 Realizations with dierent perturbation parameters . . . . . . . . . . 59
2.38 Water cut curves for a chain of 10 realizations. The initial realization
is for the model in Figure2.35a . . . . . . . . . . . . . . . . . . . . . . 59
2.39 Objective function of a chain of 10 realizations . . . . . . . . . . . . . 60
2.40 A synthetic reservoir water saturation model with 20 channels and 3
producers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.41 The region dening procedure . . . . . . . . . . . . . . . . . . . . . . 64
3.1 A deep-water channelized reservoir analog. The brown and yellow col-
ors are for channel ll facies, green color is for shale drapes and blue
color is for scour holes. The model is constructed using surface-based
grid. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.2 Analog model (upper) is converted into Cartesian grid (middle) and
upscaled to a coarse scale model (bottom). The color reprsents dierent
channel objects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.3 Coarse channel belt geometry identied from reference analog model. 71
3.4 The channel belt geometry is restored into original shape and their
boundaries are traced for later channel simulation . . . . . . . . . . . 72
3.5 Two level of heterogeneities: channel belts and channels . . . . . . . . 73
xv
3.6 Shale drape training images (top 2 rows) and two reservoir models. In
training images, the red color represents scour holes and blue back-
ground the shale drapes. Round vs. ellipsoid is for hole geometry,
small vs. large is for hole size, and 0.1 vs. 0.5 is for hole proportion. . 78
3.7 Relative permeability for ow simulation model. . . . . . . . . . . . . 79
3.8 Eect charts for ow responses. . . . . . . . . . . . . . . . . . . . . . 79
3.9 Channel geometry parameter derived from analog . . . . . . . . . . . 86
3.10 Channel pattern parameter distribution: the experimental CDF (blue
dots) and the analytic CDF function obtained by regression (solid lines). 86
3.11 The 3D hole distribution models (middle row) and the cross sections
(bottom row) for belts and channels (middle row) using the correspond-
ing hole training images (top row). The dark red indicates holes. The
training image grid dimension is 100 by 100 and the hole models are
50 50 50. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.12 Observed production responses from the synthetic reservoir. . . . . . 89
3.13 Optimization performance for two step perturbations. . . . . . . . . . 91
3.14 History match results for two step perturbation. . . . . . . . . . . . . 92
3.15 Reference model vs. history match geologic model. . . . . . . . . . . 92
3.16 Well data vs. simulated facies at well locations. Dierent color repre-
sents dierent channel section. . . . . . . . . . . . . . . . . . . . . . . 93
3.17 Well water cut and eld oil recovery predictions using the history
matched models and randomly picked models. . . . . . . . . . . . . . 94
3.18 Well bottom hole pressure predictions using the history matched mod-
els and randomly picked models. . . . . . . . . . . . . . . . . . . . . . 95
3.19 A channel hole training image (right) with proportion dierent with
the reference (left). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.20 The optimization performance for the case that both hole location
and proportion are perturbed. The hole proportion converges to the
reference with the mismatch decreases. . . . . . . . . . . . . . . . . . 96
3.21 Flow response Predictions using the history matched models and ran-
domly picked models. . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
xvi
4.1 Stratigraphic interpretation of WC reservoir. Gray color is for valley
region. Other colors are for dierent channel belts. . . . . . . . . . . 99
4.2 Scour hole training images for valley, belt and channel hole simulation.
The red color objects are scour holes, and blue background is shale
drapes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.3 Scour hole realizations for valley, belt and channel using the training
images shown in Figure 4.3. . . . . . . . . . . . . . . . . . . . . . . . 102
4.4 Oil and water relative permeability curves. . . . . . . . . . . . . . . . 103
4.5 Geologic properties used for ow simulation. . . . . . . . . . . . . . . 104
4.6 Reservoir model showing reservoir structure, the geologic regions, and
well conguration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.7 Water cut and bottom hole pressure proles for true case 1 and 2. . 107
4.8 Oil saturation distribution after 5 year production. . . . . . . . . . . 108
4.9 Ten Water cut and bottom hole pressure proles for both non-history
matched models and history matched models performed using pseudo
3D perturbation scheme. Blue line is for non-history matched model,
red line is for history matched model, and black line is for true data. 110
4.10 Ten Water cut and bottom hole pressure proles for both non-history
matched models and history matched models performed using true
3D perturbation scheme. Blue line is for non-history matched model,
red line is for history matched model, and black line is for true
production data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.11 Optimization performances corresponding to 10 history matching re-
sults shown in Figure 4.9 and Figure 4.10. Blue color is for pseudo-3D
perturbation scheme, purple color is for full 3D scheme. . . . . . . . . 112
4.12 The eect chart of examined factors listed in Table 4.4 on ow re-
sponses. Red lines are decision limits. If the eect bar exceed red,
this means that factor is signicant to the corresponding ow response
with 95% condence level. . . . . . . . . . . . . . . . . . . . . . . . . 116
4.13 Streamlines showing the water ow path. . . . . . . . . . . . . . . . . 117
4.14 History matching results assuming two regions in the reservoir. . . . . 118
xvii
4.15 Optimization performance corresponding to the history matched re-
sults shown in Figure 4.14. . . . . . . . . . . . . . . . . . . . . . . . . 119
4.16 History matching results assuming the reservoir as one region. . . . . 119
4.17 Optimization performance corresponding to the history matched re-
sults shown in Figure 4.16. . . . . . . . . . . . . . . . . . . . . . . . . 120
4.18 Reference and history-matched geologic models. . . . . . . . . . . . . 120
4.19 Oil and water saturation distribution after 5 years production. . . . . 121
xviii
Chapter 1
Introduction
One of the primary goals of 3D geological modeling is to provide a geologically reason-
able, numerical representation of the geology for input into ow simulators, in order to
predict the behavior of the subsurface reservoir uid ow under various hydrocarbon
recovery scenarios. In most ow simulations, the permeability spatial distribution is
the most inuential input because it conditions the ow path (Weber, 1986; Hewett,
1986). It has long been recognized that ow in heterogeneous porous media is mainly
controlled by the degree of continuity and connectivity of permeability - ow barri-
ers, e.g. shales, or high permeability channels or faults (Giordano, 1985; Lake, 1986).
Randomly disconnected small shale bodies may not aect uid ow, whereas a small
proportion of connected shale drapes may control ow and thus sweep eciency and
recovery. Therefore, a realistic representation of the distribution of high-ow channels
or ow barriers is often essential to obtain reliable production prediction.
Shale drapes have been increasingly reported and studied as ow barriers for
channelized reservoirs, including uvial channels (MacDonald, 1993) and turbidite
channels (Weimer et al., 2000). Channelized reservoirs often show multi-scale archi-
tectures such as large-scale channel belts, middle-scale single channels and small-scale
channel inll facies (Figure 1.1). These architectural elements are dened between
bounding surfaces (Clark and Pickering, 1996b). Along these hierarchical bounding
surfaces, thin shale drapes may be present as permeability barriers that compart-
mentalize the reservoirs. Studies have shown that shale drapes have stronger impact
1
2 CHAPTER 1. INTRODUCTION
Figure 1.1: Hierarchy for a deepwater channelized reservoir. Shale drapes could be
deposited on the channel belt margins, on channel margins within individual channel
belts and on cross-laminations within individual channels.
on the ow behavior than the shale facies distributed inside channels (Weimer et al.,
2000). Figure 1.2 is a schematic diagram showing the eect of shale drapes on reser-
voir connectivity. Due to the drape presence as barriers (green curvilinear curve), oil
in the disconnected channels (brown-colored) is left unswept even when these channels
are stacked and appear connected to producers (yellow-colored). Since the geometry
and distribution of shale drapes are governed by the erosional surfaces of channel belts
and the channel-body margins which are at dierent scales, they represent multi-scale
heterogeneities. Characterizing the distribution of these multi-scale ow barriers and
representing them properly in a ow simulation model are critical to reservoir behav-
ior prediction and management. The multi-scale nature of ow barriers calls for a
hierarchical modeling approach in which the large-scale reservoir architecture is mod-
eled rst, the shale drapes are then simulated within this architecture framework.
1.1. RESERVOIR ARCHITECTURE MODELING 3
Figure 1.2: Schematic diagram showing the inuence of shale drapes on reservoir
connectivity.
1.1 Reservoir architecture modeling
A channelized reservoir consists of deposits formed by multiple channels. Depend-
ing on the variation of interaction between lateral and vertical amalgamation during
the growth of channel systems, channel deposits demonstrate dierent stacking pat-
terns(Clark and Pickering, 1996a, b). These three-dimensional mosaics of permeable
(reservoir) and impermeable (non-reservoir) rock bodies of various sizes, shapes and
arrangements are termed the sedimentary architecture. The individual rock bodies
such as channels are architectural elements( Mutti and Normark, 1987). In the case
that shale drapes are deposited as basal lags along the individual channel bounding
surfaces, the proper representation of the spatial distribution of architectural elements
is important to account for the connectivity of reservoir sand bodies.
Reservoir architecture modeling in this work refers to modeling the hierarchic
stacking patterns of multi-scale architectural elements. The architecture hierarchy of
channelized reservoir ranges from seismic-scale (canyons, belts) to sub-seismic scale
(channels and internal facies). In order to attach shale drapes to the architectural
element boundaries, the simulated architectural elements should have clearly dened
geometry. Hence, the desirable architecture modeling technique should be such that
it is capable to simulate channel stacking patterns delineated by bounding surfaces.
At the same time it should be exible for data conditioning.
For seismic-scale architecture elements such as channel belts, it is easy to include
them deterministically into reservoir models. This is because we can determinis-
tically identify and interpret these large-scale object geometry from seismic data.
Therefore we can input the interpreted channel belt geometry into reservoir models
as a container for ner scale channels. For sub-seismic channels, a stochastic mod-
eling approach is required to model uncertainty and integrate various types of data.
Channels have been an important target for stochastic modeling because they control
the uid ow in the reservoir as conduits. Various channel simulation methods have
been developed to capture the geologically realistic channel geometry. This section
will review important stochastic simulation approaches used for channel simulation
and their limitations will be addressed correspondingly.
Object-based methods are also referred to as the Boolean methods. Object-based
techniques were pioneered by Haldorsen and Lake (1984), Haldorsen and Chang
(1986) and Stoyan et al. (1987). In object-based modeling, the geological hetero-
geneities such as sand and shale are dened as a set of objects. For example, in
the case of channel-type reservoirs (uvial or submarine), the sand facies can be rep-
resented by sinuous channels and the shale facies can be represented by elliptical
objects. The dimensions of each object are dened using a limited set of parameters,
usually in terms of a distribution of thickness, width, sinuosity, etc. After these ob-
jects are dened, their mutual spatial relationship such as object erosion rule, object
attraction/repulsion rule will be established. Then, the following stochastic simu-
lation simply consists of placing these objects into reservoir model with specied
spatial relationship, and moving these objects around to match local data (wells and
seismic). Compared to pixel-based methods, object-based methods provide more re-
alistic shapes such as curvilinear channels. However, with this approach it is harder,
if not impossible, to honor exactly geobodies interpreted from seismic. Another long-
standing disadvantage with object-based reservoir modeling methods lies in the con-
ditioning to well data, particularly in the presence of many wells. To overcome this
problem, Shmaryan and Deutsch developed a rule-based approach (Shmaryan and
Deutsch, 1999) for fast conditioning to well data. In the rule-based approach, the
speed is achieved through direct conditioning of the undulation channel centerline to
the sand intervals by means of 1-D conditional sequential Gaussian simulation. The
disadvantage is that this approach is unable to create large channel complexes in the
same manner as it creates individual channels within a channel complex.
Surface-based modeling approach has been introduced recently (Deutsch and Xie,
2001; Pyrcz and Deutsch, 2005) to better simulate internal and external geome-
tries and stacking patterns of architectural elements associated with turbidites. In a
surface-based approach, a conformable grid is used to explicitly model the location
and the thickness of the channel and channel drapes. With this modeling approach,
the thin shale layer could be represented with surfaces, and the cells contained between
the two surfaces created as ow barriers could be represented either with transmissi-
bility modier or zero vertical permeability values during ow simulation (Begg and
King 1985; Jackson and Muggeridge, 2000; Willis and White, 2000; Weimer et al.,
2000). However, three limitations exist. First, adequate horizontal discretization is
required to preserve a smooth surface shape, which will result in a large number of
cells and make the ow simulation time-consuming; Secondly, as the lever of condi-
tioning increases or channel size increases, it becomes more dicult to match well
data. As a result, greater computational eort is needed to generate models; Thirdly,
because both the channel location and shale drapes location are uncertain, they need
to be perturbed during history matching of the reservoir model, which would mean
changing the conformable grid automatically. Currently, such gridding is not yet at
an adequate level of robustness to be made automatic.
Alternatively, one could look at process-based methods for generating channels
and shale drapes (Bridge and Leeder, 1979; Lopez, 2003). Process-based methods
attempt to simulate fundamental geological processes to produce a numerical repre-
sentation of the reservoir geology. These approaches include the rigor of the physics
of sedimentation and depositional processes thus have the advantage to generate re-
alistic geological heterogeneity. However, enormous diculties arise when it comes to
conditioning process-based models to well data.
More recently, pixel-based multiple-point statistics (MPS) techniques have been
developed by Strebelle (2002) to simulate geologically realistic models. This approach
has been applied to capture the curvilinear features in the deepwater settings (Stre-
belle et al., 2003). It allows simulations of 3D facies geometries and distribution
using multiple-point statistics and training images. Multiple-point statistics can cap-
ture correlations that are geologically complex. Training images are non-conditional
and purely conceptual depictions of the geological patterns deemed relevant for a
particular reservoir. It can be derived from outcrop observation, expert knowledge
or geophysics (see an example in Caers et al. 2003). The MPS algorithms extract
geological patterns from the training image and reproduce them in simulated models
at the same time being constrained by seismic and well data. MPS methods make
it possible to simulate complex geometries such as channels, meanders and preserve
the relations between facies, while honoring well information as well as any other sec-
ondary constraints such as a seismic-derived probability map of facies presence. The
problems with this method are: (1) the facies patterns are reproduced as a whole,
hence in the nal model one has no knowledge of an individual object (eg., an individ-
ual channel). As a result, we can not identify single channels from MPS realizations
due to channels cross-cutting each other. For example, Figure 1.3a shows one MPS
simulated realization with channels (red) coded as 1 and non-channel (gray back-
ground) as 0. Figure 1.3b is a part of this realization where two channels cross-out
each other. We can notice the cross-cutting in the horizontal direction, however in
the vertical direction, it is impossible to identify the channel boundary because two
channels have same numerical code (color). One possible solution is to code dierent
channels with dierent numbers thus boundaries can be identied (Figure 1.3c). To
achieve that, we need to build a large and pattern-rich training image to guarantee
enough repetition of dierent cross-cutting patterns present in the training image,
however the algorithm then requires several GBs of memory (Strebelle, 2002). (2)
ne-scale low-proportion features such as shale drapes are dicult to model with
MPS. This is because low-proportion patterns do not repeat frequently enough in the
training image. Therefore, the simulated geometry and distribution of low-proportion
shale drapes may not follow the patterns shown in the training image (Figure 1.4).
In 2005, Alapetite et al. developed Yet Another Channel Simulation(YACS) ap-
Figure 1.3: Example result from MPS. Where two channels cross, MPS can not
delineate the ne scale ow barriers that separate these two channels.
proach to generate geologically realistic uvial channels while allowing for relatively
easy conditioning to multiple well data. The core of this approach lies in the associa-
tion of a fairway with the channels to be simulated. In this approach, a potential eld
is rst dened within the fairway, then this potential map is mapped into thickness
resulting in generating a channel inside the fairway. To add the necessary sinuosity to
the channel, a correlated noise is stochastically simulated and added to the potential
eld. Conditioning to well data is obtained by transforming well observations into
channel thicknesses which will constrain the simulation of noise. This method is fast
and can condition to well data under the assumption that individual channel sand
bodies can be identied in the well data. But as with other Boolean approaches, it
cannot necessarily reproduce accurately the channel stacking patterns observed from
outcrop analog or from training image.
As stated above, these modeling methods can either simulate geologically realis-
tic individual channels or channel patterns as a whole. None of them is capable of
Figure 1.4: A synthetic 2D fracture training image (left, 400x300 grid) with 13% frac-
ture (blue lines) and one realization using snesim (right, 200x200 grid). The fracture
patterns shown in training image are not reproduced in the simulated realization.
simulating channel stacking patterns delineated by bounding surfaces.
1.2 Flow barriers modeling
Flow barriers are impermeable deposits or very low permeability deposits. They
develop in many depositional environments, such as uvial, deltaic, shoreface or shal-
low/deep marine. These impermeable barriers can form by depositional processes
such as mud draping, or by digenetic processes such as calcite cementation. In this
work, we focus on thin shale drapes formed along erosionally conned channel bases.
Field and outcrop study have shown that repeated episodes of erosional cutting -
lling with coarse clastics - draping by mud during periods of quiescence - partial
excavation of mud drapes are a feature of some erosionally conned channels (Gard-
ner and Borer, 2000; Mayall et al., 2006). In contrast, some other channels systems,
or segments of channels, lack drapes, depending on a variety of factors (e.g., Lowe,
2004; Hubbard et al., 2008a, b). In this thesis, we deal with the channel end-member
characterized by impermeable shale drapes. The basal shale drapes within erosionally
conned channels have been increasingly described from a number of outcrops (e.g.,
Gardner and Borer, 2000; Gardner et al., 2003; Eschard et al., 2003). These drapes
generally are very thin (cm to m) and below seismic resolution, but they could form
1.2. FLOW BARRIERS MODELING 9
permeability barriers. Therefore, presence of channel base drape can have a signif-
icant impact on oil recovery because they represent one of the main uncertainties
in the development of a turbidite channel reservoir. Outcrop data and production
data from real eld have provided ample evidence of the presence of holes in shale
drapes. Due to the presence of holes, shale drapes are discontinuously distributed
along the channel bounding surfaces. These discontinuous thin shales often are not
correlated between wells. Even the hole distribution trend may be predictable if
they related to the locus of strong current ow within a channel, their exact location
within channels and the individual channel location are uncertain. Hence their loca-
tion will be modeled using stochastic techniques.
Stochastic simulating the spatial distribution of shale drapes is dicult. The rst
diculty is related to data resolution. Sparse well data can not provide sucient
information on where these shales are located or on their complex geometry as well
as continuity. The vertical resolution of seismic surveys (above 10 m) is too coarse
to image these thin shale drapes. As a result, the coverage and geometry of shales
are dicult to infer clearly from well or seismic data alone. The second diculty
is related to gridding these features in a high resolution geostatistical model or in
a coarsened ow simulation model: even if we can gain sucient knowledge about
shale drapes distribution and gridding, it is dicult to capture them with traditional
modeling techniques. The thickness of shale drapes could be on the order of centime-
ters whereas a simulation model grid is usually 10s or 100s of meters in resolution.
The details of the channel geometry and ne scale ow barriers can potentially be
captured with a very high-resolution highly unstructured grid (as shown in Figure
1.1), however the ultimate goal of reservoir modeling is ow prediction hence high-
resolution models need to be coarsened to make ow simulation feasible. In such
coarsened model, even with the use of unstructured grids, the continuity of shale
barrier may be destroyed.
Beside the modeling diculties, there are multiple sources of uncertainties that
need to be taken into account. A rst source of uncertainty is the shale drape pro-
portion/location. The proportion and location govern the amount of communication
between each of the separate channels. As previous stated, the location, geometry
and continuity of shale drapes are dicult to infer from well data and seismic data
alone. Although outcrop data may provide a conceptual idea of shale drape prop-
erties, such information is often 2D. Another uncertainty source is the location of
the individual channel body and channel belt itself. Quite often channel belts can
be identied from seismic data, but individual channel bodies within these belts are
below seismic resolution ( Slatt and Weimer, 1999; Abreu, 2003). Outcrop studies
may provide knowledge about the type of channel stacking, channel width, depth and
sinuosity. However, the actual location of channels in the reservoir is case-dependent
thus requires one to build a high-resolution geological model constrained to actual
reservoir data. Because the geometry of shale drapes is governed by the geometry
of channel boundaries along which they were deposited, the location of the channel
body will have strong impact on the reservoir performance, even in high net-to-gross
system in which channels have a strong degree of lateral migration due to the absence
of mud-rich levees (Clark and Pickering, 1996a).
Previous studies on discontinuous shale drapes have used either simple 2D generic
models, in which shales are located randomly within a homogeneous background
(Richardson, 1978; Martin, 1984; Begg et al., 1985; Jackson and Muggeridge, 2000)
or detailed 3D models of specic depositional environments derived directly from out-
crop data or modern analog (White and Barton, 1997; Willis and White, 2000; Li
and White, 2003; Pranter, 2007). All these studies concentrated on planar or inclined
bedding surfaces. Stright et al. (2005) used a semi-stochastic technique to simulate
shale drapes along the curvilinear surfaces of channels. The model rst specied the
geometries of channel bounding surfaces, then stochastically placed the ellipsoid holes
(with background as shale drapes) on each surfaces using conditional multiple-point
geostatistical simulation. All of these studies focus on a single hierarchy. And the
governing surfaces of architectural elements are xed or deterministically included
into the models.
Other approaches for modeling thin shales rely on more ad hoc engineering
approaches. For example, thin shales are implicitly modeled by altering relative per-
meability, kv/kh ratios or pore volume and transmissibility multipliers. In reality,
1.3. GEOLOGICALLY CONSISTENT HISTORY MATCHING 11
engineers often perturb these properties by trial and error to match production his-
tory. The success or failure of the history matching endeavor largely depends on the
experience, fortitude and sometimes good fortune of the engineer. The main limita-
tion of such an approach is that geological information is often ignored or worse, the
single deterministic model generated becomes geologically inconsistent.
Given these diculties, there is a need for a coupled geologic modeling and his-
tory matching engineering methodology which integrates both static and dynamic
data simultaneously. We propose an approach where thin shale barriers referring to
dierent scales are modeled hierarchically in a manner that is directly linked to a
reservoir ow simulator transmissibility, rather than attempting to explicitly repre-
sent them by gridding. This approach will prove to be appealing in the context of
matching the production history since channel location, shale proportion and shale
location need to be perturbed until the historical production data is matched.
1.3 Geologically consistent history matching
The lack of geologic information to appropriately reproduce the reservoir heterogene-
ity is one of the most signicant sources of uncertainty. To reduce this uncertainty,
production data are often integrated into geologic models. Integration of production
data or history matching is an optimization process which involves the denition and
minimization of an objective function. This objective function measures the dier-
ence between the historical production data and the equivalent simulated responses.
It is minimized through iteratively adjusting the geological model. Once a model has
been history matched, it can be used to simulate future reservoir behavior with a
higher degree of condence, especially if the adjustments are consistent with geolog-
ical information.
In order to obtain a history matched geological model, most methods rely on
modifying the initial geostatistical realization, proceeding by trial and error to ob-
tain a history match. As we know, the history matching process is essentially an
inversion of reservoir parameters from the historical production behavior. Because
this inversion problem is ill-posed, the history matched geological models are not al-
ways consistent with the data, other than dynamic, or do not always preserve the
geological concept believed to exist. To overcome this problem, geologically consis-
tent history-matching methods have been introduced. During geologically consistent
history-matching process, the objective function is minimized through systematical
perturbation of the local conditional distributions with a perturbation parameter, r
D
that is calibrated using the available production data. Based on the way the reser-
voir realization is perturbed, there are two perturbation methods: one is the gradual
deformation method (GDM) (Hu, 2000), the other is the probability perturbation
method (PPM), developed by Caers (2003).
GDM
The objective of the gradual deformation is to create continuous perturbation of an
initial realization, such that the perturbed realization matches better the production
data. All the perturbed realizations are sampled from the same prior model as the
initial realization. The perturbation is achieved by linearly combining two indepen-
dent standard Gaussian random functions Y
1
and Y
2
with identical covariance:
Y (r
D
) = Y
1
cos(r
D
) +Y
2
cos(r
D
) (1.1)
Whatever the r
D
perturbation parameter, Y has the same mean, variance and co-
variance as Y
1
and Y
2
. Furthermore, Y is also a Gaussian random function because
it is a sum of two Gaussian random functions (Hu, 2000). Two independent realiza-
tions y
1
and y
2
of Y
1
and Y
2
can form a continuous chain of realizations y(r
D
), which
depends on the parameter r
D
. This feature can be used to calibrate realizations to
production data. The basic idea is to optimize r
D
based on some mismatch between
eld response and the one simulated on y(r
D
).
In geostatistical practice, sequential simulation is often used to generate condi-
tional realizations. This takes two steps: rst, one builds a probability distribution
model conditioned to the hard data and pre-simulated values; secondly, one draws a
value from this probability model using the random number with a uniform distribu-
tion in [0, 1]. In this sequential simulation framework, GDM modies the realizations
by perturbing the random numbers used to draw from the conditional probability dis-
tribution model. The idea is to gradually perturb a realization of a uniform vector.
This vector contains the random numbers that are used for drawing. The pertur-
bation is achieved through Eq.1.1. Suppose u is a vector of random numbers with
uniform distribution in [0, 1]. The vector u is rst transformed into Gaussian domain
as y. Next, y is perturbed using Eq.1.1, generating a perturbed Gaussian vector
y(r
D
). Then y(r
D
) is back-transformed into a uniform vector of perturbed random
numbers u(r
D
). This u(r
D
) is a perturbation of u, so when u(r
D
) is used in sequential
simulation, a perturbation is achieved.
GDM was initially presented as a tool to deform realizations of Gaussian-related
stochastic models. By deforming gradually Gaussian realizations, the objective func-
tion varies smoothly with a gradual change in the perturbed realizations. As the
resulting variations are continuous, this method is of interest for gradient-based op-
timizations. Later, the gradual deformation method was extended to Boolean or
object-based simulations. Hu (2000, 2003) suggested applying the gradual deforma-
tion method to modify the object locations and the number of objects in a simulation.
However, the sudden appearance and disappearance of objects induces strong discon-
tinuity in the objective function. Such behavior is not desirable for gradient-based
optimizations. To alleviate this undesirable feature, Le Ravalec et al. (2004) refor-
mulated the gradual deformation method for Poisson probabilities. The reformulated
GDM allows to now smoothly add or remove objects from a Boolean simulation, thus
reduces the discontinuity of the objective function, but does not fully eliminate them.
PPM
Unlike GDM that perturbs the random number used to draw from probability distrib-
ution, PPM perturbs the probability model itself that is used to generate conditional
simulation. This method is developed by Caers (2003) and Caers and Homann
(2006). It relies on the development of multiple-point statistics (MPS). MPS is a
new geostatistical technique that allows generating conditional stochastic simulations
reproducing the training image-based geological heterogeneity (Strebelle, 2002). The
PPM oers solutions by extending the data integration framework of MPS to produc-
tion data: accounting for production data yet being constrained by prior geological
information. The idea is to convert both static data B and dynamic data D into
probabilities P(A|B) and P(A|D) respectively, where A is the property to be sim-
ulated (such as channel occurs at specic location), then combine these individual
probabilities into a single probability model P(A|B, D); nally, the reservoir models
are generated by sequentially drawing from P(A|B, D). To combine individual prob-
abilities into one model, Jounels tau model is used, and this method is expressed as
follows:
x
a
=
_
b
c
_
1 _
c
a
_
2
(1.2)
where
x =
1 P(A|B, D)
P(A|B, D)
, b =
1 P(A|B)
P(A|B)
, c =
1 P(A|D)
P(A|D)
, and a =
1 P(A)
P(A)
P(A) is the global proportion of A occurring; The parameters
1
and
2
account for
redundancy between the data B and D, but generally, they are chosen as
1
=
2
=1,
which amounts to a form of standardized conditional independence. The combined
probability model, P(A|B, D), for unit values is equal to the following expression:
P(A|B, C, D) =
1
1 +x
=
a
a +bc
(1.3)
While P(A|B) can be directly inferred by scanning the training image for equivalent
data events A|B, P(A|D) is dened as follows considering a binary spatial variable
case:
P(A|D) = (1 r
D
)i
0
(u) +r
D
P(A) [0, 1] (1.4)
Where r
D
is a parameter between [0,1] that controls how much the model is perturbed
and i
0
(u) is the initial realization of I(u) with
1.4. PROPOSED MODELING APPROACH 15
I(u) =
_
1 if a given facies occurs at u
0 else
Similar to gradual deformation method, PPM parameterizes the perturbation us-
ing a single parameter r
D
that can be optimized. The optimization uses the Brent
method but it could be any one-dimensional optimization routine. The advantage of
Brent is that it does not require derivatives, and it is ecient and robust (Homan,
2005).
Although PPM was initially based on MPS, the method is not limited to multiple-
point statistics or even sequential simulation. Any technique that uses conditional
probabilities allows one to use the probability perturbation method.
1.4 Proposed modeling approach
To model a multi-scale reservoir architectural geometry, a hierarchical workow (Fig-
ure 1.5) is adopted which is routinely used in reservoir modeling. Within this work-
ow, large-scale architectural elements (for example channel belt) are modeled rst,
their corresponding bounding surfaces are then identied and the shale drapes are
simulated on these surfaces; next, the second level hierarchic architecture elements-
individual channels within channel belts are simulated, the corresponding bounding
surfaces are captured, and shale drapes are modeled; then, lithofacies are simulated
within each channel, nally continuous petrophysical properties such as porosity and
permeability are assigned on a by-facies basis. Within the hierarchical workow,
available or known data must be honored at all times to obtain a reliable geological
model, that is, each level of hierarchy model must be consistent with all available
data with corresponding scales.
In this work, for the purpose of studying shale drapes, the lithofacies within chan-
nel will only take into account sand which means channels are fully lled with sand
facies. Properties such as porosity and permeability are modeled using traditional
sequential Gaussian simulation. The shale drapes along the bounding surfaces will be
simulated using MPS and treated as edge property without volume. For reservoir ar-
chitectural modeling, this involves dening channel deposition fairways (belts) based
on seismic data, modeling long sinuous channels and placing them into dened fair-
ways such that all data are matched. Because we assume that channel belts can often
be identied from seismic data (with some associated uncertainty), the second level
hierarchic architecture modeling- individual channel geometry and stacking pattern
modeling are the more challenging part in this workow.
To integrate production data into the reservoir modeling workow, this research
will adopt the probabilistic perturbation methods (PPM and GDM) because these
geologicaly consistent perturbation methods are capable of accounting for geologi-
cal continuity consisting of strongly connected, curvi-linear geological objects such
as channels. PPM/GDM couples the geological modeling and history matching in
an automatic workow. The stochastic nature of PPM/GDM makes assessment of
the uncertainty of reservoir connectivity possible. The main challenge in the history
matching part is how to modify channel locations to match production history while
honoring the observed or believed stacking patterns.
Figure 1.6 shows the proposed coupled modeling and history matching workow
based on the previous discussion. The probabilistic perturbation is performed in a
hierarchical framework which is consistent with previously stated hierarchic workow,
that is, it starts modeling shale drapes associated with channel belts, then individ-
ual channels are simulated within channel belts and the associated shale drapes are
simulated for each channel. To match the production data, the shale drape locations
along both belt and channel boundaries as well as the channel locations are perturbed
stochastically in an order dened by sensitivity study.
1.5 Thesis outline
This dissertation consists of ve chapters. The proposed hierarchic modeling and
history matching workow is presented in Chapter 2. This chapter also explains the
1.5. THESIS OUTLINE 17
Figure 1.5: A hierarchic workow to simulate multi-scale ow barriers.
well data conditioning process and probabilistic perturbation process. The synthetic
examples are provided to illustrate each of them.
Chapter 3 tests the proposed workow on a conned-channel reservoir analog.
The sensitivity study is rst performed to determine the most sensitive parameters
such as shale drape geometry, proportion and location. Then, the proposed workow
is applied to check its feasibility.
Chapter 4 presents a realistic case study based on an oshore West Africa turbidite
reservoir, where the modeling and history matching workow has been applied with
the purpose of determining the location of channels and shale drapes in the reservoir.
Chapter 5 discusses the major ndings and contributions of this dissertation, as
well as suggestions for future work.
Figure 1.6: Coupled geologic modeling and history matching of multi-scale ow bar-
riers.
Chapter 2
Proposed modeling workow
The proposed modeling workow in Figure 1.6 is composed of two parts: (1) static
geologic modeling by integrating various static data such as well logging, seismic, out-
crop analog; (2) dynamic data integration using a probabilistic perturbation scheme
that ensures consistency between reservoir models developed from one stage to the
next. The nal goal is to provide multiple geologic models that can be used to predict
the reservoir performance with a higher degree of condence .
The major contribution in this proposed modeling workow consists of a new
channel stacking pattern modeling technique as well as perturbation method that
modies channel location consistently with the conceptual stacking pattern model.
For channelized reservoirs, the rst level reservoir architecture - channel belt/complex
is generally observable directly from seismic data, while the individual channels are
below seismic resolution. Therefore, the channel stacking pattern modeling or the
second level reservior architecture modeling is the most challenging one. The pro-
posed modeling method makes use of the distribution function of pattern parameters
to simulate and perturb stacking patterns.
This chapter will rst introduce the static geologic modeling process which in-
cludes individual channel modeling, channel stacking pattern modeling, shale drape
modeling and well data conditioning. Following that, the geologic consistent pertur-
bation process will be presented and a 3D synthetic example is provided to test the
feasibility of proposed perturbation scheme.
19
20 CHAPTER 2. PROPOSED MODELING WORKFLOW
2.1 Reservoir architecture modeling
The static geologic modeling in this dissertation focuses mainly on reservoir architec-
ture modeling. This is because ne-scale reservoir architecture at sub-seismic scale
is challenging to model. In this work, we assume the rst level hierarchic reservoir
architecture-channel belt/complex is above seismic resolution and can be integrated
into reservoir model deterministically. On the other hand, individual channels are be-
low seismic resolution and require a stochastic modeling approach. Hence the reservoir
architecture modeling here refers to the second level hierarchic reservoir architecture
modeling or channel stacking pattern modeling.
The second level hierarchic reservoir architecture modeling (channel stacking pat-
tern modeling) workow consists of two components: individual channel modeling
and channel stacking pattern modeling. The criteria for choosing a specic modeling
approach include the ability to match hard data, to dene object (channel) bound-
aries (for the purpose of attaching shale drapes), to reproduce a given stacking pattern
and to be CPU ecient. The present work adopts a stratigraphic-based modeling ap-
proach. In this stratigraphic-based approach, individual channels are simulated using
the YACS method (Alapetite et al., 2005) because this method allows the simulated
channels to be continuous throughout the modeling area, to have dened boundaries
and to stay within an observed fairway (valley). The method is stochastic, rather
than genetic or process-based technique; it only attempts to mimic the latter. But as
will be shown, this method is fast and conditions to well data under the assumption
that individual channel sand bodies can be identied form well data. To put multiple
channels together reproducing desired stacking pattern, stacking pattern parameters
such as overlap ratio and migration ratio are simulated stochastically.
2.1.1 Individual channel modeling
To perform the YACS approach (Alapetite, et al., 2005), the following parameters
are required to specify channel geometry (Figure2.1):
Figure 2.1: Denition of parameters used to describe the channel geometry
Channel orientation (): the direction a channel extends;
Channel wavelength (): the distance between adjacent peaks or troughs;
Channel amplitude (A): the distance between adjacent trough and peak;
Channel width (W): the maximum width of the channel; it is located at the
center of the channel;
Channel thickness (H): the maximum thickness of the channel; it is located at
the center of channel;
Channel cross-section geometry is dened by a parabola
h(x) = H
_
1
_
2(x x
0
)
W
_
2
_
where x
0
is the location of channel center point and x is the position on the
cross-section to be computed.
For simplicity, in this dissertation we assume all the channels have symmetric cross
section. We also assume all the channels have regular sinesoid geometry described
by a set of geometry parameters listed above and dont consider the unanticipated
change in sinuosity. The core of the YACS method lies in the association of a
fairway with the channels to be simulated. In this method, a single channel simulation
starts by selecting a particular depositional surface. On this surface, a channel belt
is dened and a so-called potential eld is computed within this belt. In order to
create sinuosity and realistic channel geometry, a stochastic perturbation is applied
to this original potential eld. The variations then control channel geometry such
as sinuosity, amplitude. The perturbed potential eld is then mapped to channel
thickness using a transform function. Placing the thickness underneath the selected
depositional surface results in generating a channel inside the channel belt. The main
steps for single channel modeling are:
1. Select a deposition surface and dene the channel belt boundaries on this surface
(Figure 2.2a);
2. Compute the original potential values between two boundaries by interpolation
(Figure 2.2b). One boundary is set to negative and the other one is positive.
The absolute number of these two extremes corresponds to the bottom of the
channel belt at which the channel is located;
3. A correlated noise is simulated using Sequential Gaussian simulation (sgsim)
(Figure 2.2c); The simulated noise is normal distributed with mean 0. Its
standard deviation is linearly related to the amplitude of simulated channel.
The principle direction of the variogram is the orientation of the channel belt.
The range in that direction is linearly related to channel wavelength;
4. A perturbed potential map (Figure 2.2d) is obtained by adding the simulated
noise to the original potential map; Keep the positive potential values in Figure
2.2d and ip the negative potential values into positive, the 0-isopotential line
is the channel centerline (Figure 2.2e);
5. Once the channel centerline is located, the channel region is dened (Figure
2.2f) since we know the channel width;
6. Apply a transfer function
h(d) = H
_
1
_
2d
W
_
2
_
on the channel region to obtain a channel thickness map (Figure 2.2g); At the
0-potential point on Figure 2.2f, d is 0, and d increases towards the channel
boundary along the potential gradient direction.
7. Paint the thickness underneath the depositional surface; a channel in 3D space
forms (Figure 2.2h).
In the case that the channel belt is curvilinear ( Figure 2.3a), the original potential
map is interpolated between the boundaries of channel belts (Figure 2.3b), then a
correlated noise eld is simulated (Figure 2.3b). The principal axes of the noise vari-
ogram are dened by the potential gradient direction. The principal direction of the
variogram is the orientation of the channel belt. Performing Step 4 to add the simu-
lated noise to the original potential map, the 0-potential values within the perturbed
potential map delineates the desirable channel centerline (dark blue) (Figure 2.3e).
Comparing Figure 2.3e with the channel geometry showing in the original potential
map (before adding the noise map) (Figure 2.4), the perturbed channel geometry is
more realistic. Then we can perform Step 5 - 7 to generate a 3D channel body within
this curvilinear channel belt.
Figure 2.2: The workow for individual channel modeling
Figure 2.3: The example showing curvilinear channel belt case
Figure 2.4: The original potential map in Figure 2.3 showing the channel geometry
before adding noise map
2.1.2 Channel simulation parameters
In this modeling method, the channel geometry is controlled by the simulated noise
(Step 3) which is generated using Sequential Gaussian Simulation. To obtain a de-
sirable channel geometry, the noise statistics such as the variogram and histogram of
the noise can be obtained to generate desirable channel geometry parameters such as
channel wavelength, amplitude and sinuosity. Figure 2.5 shows the perturbed chan-
nel potential maps with dierent noise histogram variances. As the noise variance
increases, channel amplitude and sinuosity increases. The channel wavelength is re-
lated to the noise variogram range along the channel orientation direction. From
Figure 2.6 top row we observe that the channel wavelength increases with noise vari-
ogram ranges along the channel orientation direction. Figure 2.6 bottom row shows
channel geometries with dierent noise variogram ranges perpendicular to the chan-
nel orientation direction. The channel amplitude decreases with an increase in the
variogram range perpendicular to the channel orientation.
Besides geostatistical parameter related to noise simulation, the potential gradi-
ent in the original potential map also aects the nal channel geometry (Figure 2.7).
Figure 2.7 top row are the original potential maps generated using dierent poten-
tial gradients. The 0-isopotential line is located at the same position on these maps.
The bottom row shows the perturbed potential maps depicting channel centerline
geometry after adding the same noise. It shows that with potential gradient increase,
channel amplitude decreases.
2.1.3 Channel stacking pattern modeling
Reservoir architectural geometry is reected in the channel stacking pattern. This
dissertation uses two parameters to dene the stacking pattern, i.e. the overlap ra-
tio and the migration ratio. The reason for choosing these two parameters is that
the channel stacking architecture is controlled by the interaction between a lateral
and vertical amalgamation process (Clark & Pickering, 1996). The migration ratio
(Figure 2.8 left) is the ratio of horizontal distance (x) between two adjacent channel
center points and the channel amplitude. It constrains the lateral distance between
Figure 2.5: Channel potential maps with dierent noise histogram variances
Figure 2.6: Channel potential maps with dierent noise variogram ranges: top row is
for ranges along channel orientation direction, bottom row is for ranges perpendicular
to this direction
Figure 2.7: The potential maps (bottom) and their corresponding original potential
maps (top) generated with dierent potential gradients
two adjacent channels. The overlap ratio (Figure 2.8 right) is the ratio of vertical
overlap thickness (h) between two channels and the channel maximum thickness (H).
It constrains the vertical distance between two adjacent channels. Figure 2.9 depicts
a cartoon showing channel stacking patterns with dierent pattern parameter combi-
nations. In this work, these two parameters are combined to determine the location
of the channel object centerline relative to an adjacent channel. We assume that
the probability distribution functions of these two parameters can be obtained either
from outcrop study or process-based models.
For demonstration purposes, a uniform distribution for migration ratio and over-
lap ratio is used in the following synthetic cases. We also assume the net-to-gross
ratio for each channel belt can be obtained. Based on the superposition, the younger
channel (deposited on upper surface) erodes the older one (deposited on lower surface)
if they are in contact with each other. Given the predened pattern parameter distri-
butions and net-to-gross ratio as well as the erosion rules, the architecture modeling
is performed as follows:
1. In a channel belt (Figure 2.10), if there are no well data to be conditioned to,
rst simulate a single channel at the channel belt top center using the simulation
method stated in section 2.1.1 (Figure 2.11a); if there are wells passing through
this belt, then rst generate channels tting all the interpreted well channel
sections;
2. Draw a value for the migration ratio and overlap ratio from their corresponding
distribution functions, and use these ratios to obtain the location relative to the
previously simulated channel; simulate a new channel centered at this location
(Figure 2.11b).
3. If the simulated channel does not fully stay within the channel belt, then it
is rejected and step 2 is repeated until the new channel is completely within
channel belt;
4. Repeat step 2-3 to generate a new channel within the channel belt until the
given net-to-gross ratio is approximately reached (Figure 2.12c - h);
5. Repeat step 1-4 for each channel belt in the reservoir (Figure 2.13);
The architecture modeling is performed from top to bottom which appears to contra-
dict the sequence of deposition. There is no doubt that the channels can be generated
from base to top following the deposition rule. However, the proposed modeling se-
quence is more favorable when a vertical proportion curve needs to be taken into
account. The vertical proportion curve obtained through well and seismic data spec-
ies the proportion of all sand as a function of vertical elevation (or depth). This
information will provide a constraint on the number of channel to be simulated for
each layer. Because channel thickness simulated in the upper layer of the grid will
contribute to the sand proportion for the current layer of the grid, it is reasonable to
generate channels from top to bottom in order to honor the vertical proportion curve.
We should notice that all the individual channels are continuous throughout the
simulation domain if they are not eroded. These channels stay within the prede-
ned belts as expected. Once these channel complexes are generated, their bounding
surfaces can be traced for shale drapes modeling (Figure 2.14).
Figure 2.8: Two parameters used for channel stacking pattern modeling
Figure 2.9: Schematic graph showing channel stacking patterns with dierent pattern
parameters
Figure 2.10: A channel belt body (left) and its bounding surface (right). Channels
will be lled into the space within the channel belt limit
Figure 2.11: The channel stacking pattern modeling process showing how channel is
lled into belt (continue to next page)
Figure 2.12: The channel stacking pattern modeling process showing how channel
is lled into belt. Left column is the architecture model, middle column are the
realizations for migration ratio, and right column are the realizations for overlap
ratio
Figure 2.13: Three channel belts case. Individual channels are lled into each belt
until its net-to-gross reached. All the channels are conned by channel belt limits
Figure 2.14: Channel bounding surfaces extracted from Figure2.11
2.1.4 Well data conditioning
Well data conditioning is performed in the individual channel modeling process. The
individual channel modeling approach adopted in this dissertation is exible to condi-
tion to well data because the only stochastic engine is pixel-based Sequential Gaussian
Simulation. The basic idea of conditioning well data is that the simulated noise at
the well locations must be such that the resulting potential, consisting of a noise com-
ponent and the original potential (which maps back into channel thickness) is close
to zero (i.e., close to channel centerline). Based on this idea, the interpreted channel
thicknesses at the well locations are rst converted into potential values, then trans-
formed to noise values. These noise values are used to condition Sequential Gaussian
Simulation (Alapetite, 2005). Unlike the Boolean simulation move-until-t process,
this approach directly places the generated object to the locations corresponding to
the well data. In other words, the channels in this approach are generated directly,
not through iterative type perturbation, to match the well data; this makes the con-
ditioning fast. However, the method of conditioning is not as general as traditional
sequential simulation. Certain assumptions need to be made. Most importantly, the
well data need to be interpreted in terms of architectural elements. Such interpreta-
tion is subject to uncertainty (not considered in this dissertation), as further detailed
in the next section.
Well data in this case are facies values known along the well path. The rst step
in the well data interpretation is to separate channel and non-channel facies from well
data. Next, the channel facies are assigned to dierent channel sections (Figure 2.15).
This assignment is essentially a process of geological interpretation that takes into
account all the geological information available about the stacking patterns (this dis-
sertation assumes channel has determined and constant dimensions). Incorporating
the channel stacking pattern information into well data interpretation is necessary
because the stacking architecture of channel sand bodies has a strong control on the
interconnectivity between channels. Similar as Viseurs method (Viseur et al., 1998),
this dissertation uses two parameters to dene the stacking pattern, i.e. the overlap
ratio and the migration ratio. As a result, the interpretation of the well data is sto-
chastic. For each interpretation, we draw pattern parameters from their distribution
functions and then convert them into a channel section center position relative to the
adjacent channel section. In other words, for the same well data set, we obtain multi-
ple channel sections with dierent stacking patterns. All of their pattern parameters
follow the same given distributions. Next, we will use a synthetic example to explain
this interpretation process.
Suppose we know the channel stacking pattern parameter probability distribution
functions for the reservoir under study (Figure 2.16) and we also know individual
channel geometry parameters: wavelength = 25m, amplitude A=12m, orientation
=0, thickness H=9m and width W=15m. All the channels have constant geometry
parameters.
For the well column shown in Figure 2.17, the sand thickness is 20 feet which
Figure 2.15: Schematic graph showing the well facies data and their interpreted
channel sections
is much larger than an individual channel thickness H (9 feet in this case). This
indicates that the well passes through multiple channels. Therefore an interpretation
is required to assign channel sand section in well to channel bodies. A pair of stack-
ing pattern parameter values is drawn using their respective probability distribution
functions shown in Figure 2.16. This pair of values is used to obtain the location
of a channel section centerline relative to its adjacent one. Once we have simulated
Figure 2.16: Synthetic uniform distributions of two pattern parameters
channel section centerline position, a channel section is added into the well sand facies
with this simulated centerline position. We repeat the random drawing of pattern
parameter ratios and the channel section generation process until the sand facies are
fully lled by channel sections at the well location. The nal result is one stochastic
interpretation (eg. Figure 2.17, Case1). If we repeat the above process several times,
we can obtain multiple interpretation results (Figure 2.17). Note that in this case we
assume no knowledge of the channel boundaries in the well data. If such information
were available, the uncertainty in the interpretation could be further reduced.
For a case with multiple wells, the individual well facies data interpretation
follows the previously stated interpretation process. In addition, the interpretation
should consider well correlation data and channel geometry parameters such as sin-
uosity and amplitude. This will ensure that the interpreted multiple well data are
compatible. In other words, if two wells are close to each other (distance less than
the channel amplitude) and have sand facies at the same interval, the interpreted
channel sections from these two wells should not conict with the channel sinuosity
information if they are on the same depositional surface. For example, in Figure 2.18,
the blue dash line represents a predened channel geometry in map view. Well A and
B are within the channel amplitude region in the x direction and very close to each
other in the y direction. If these two wells have interpreted channel sections on the
same depositional surface (such as purple channel section for well A and the green
section for well B), then these two channel sections are candidates to belong to the
Figure 2.17: Synthetic example demonstrating the stochastic interpretation of well
facies data. The same well data can result in multiple channel section stacking pattern
realizations
same channel. However, in the left interpretation the channel section centerline loca-
tions are not consistent with the channel sinuosity information thus is incompatible
data, while the right one is compatible. In practice, it may not be trivial to generate
compatible data, especially when many wells are present. In this work we assume
the interpreted channel sections are compatible when multiple wells are present.
Having the interpreted channel sections along the well path, the tops and center
Figure 2.18: Schematic example showing the compatible/incompatible interpretation
when multiple wells present
points as well as channel thicknesses at the well locations for these channel sections
are recorded. Next, the hard noise data will be computed based on these interpreted
results. The computed hard noise data will be honored during noise simulation using
sgsim in order to t a channel object to the interpreted channel section. The main
steps are explained using an example in Figure 2.19:
1. Calculate the distance (d) between the well location (point A in Figure 2.19)
and the channel centerline (point C in Figure 2.19) using
d =
W
2
_
1
h
H
, so for data in Figure 2.19, d =
20
2
_
1
8
15
= 6.8
2. Generate the original potential map with predened potential gradient dP
(=0.75). The 0-isopotential line (original channel centerline) passes through
the interpreted channel center point (point C). Read the original potential value
P
org
(=10.5) at the well location A;
3. Calculate the potential value at the well location A: P
w
= dP d = 0.756.8 =
5.1
4. Calculate the noise at the well location: N
w
= P
w
P
org
= 5.1 10.5 = 5.5
The result in Step 4 is hard noise data for noise simulation. In this example, the
simulated noise is normal distributed with mean 0 and variance 25. When a shale
interval occurs at well location, simply assign a large P
w
to avoid the shales falling
close to the 0-potential line (the channel centerline) on the perturbed potential map.
For the multiple-well case one rst identies all channel sections whose top fall
Figure 2.19: Well channel section data (left) and its original potential map (right).
The channel section geometry-maximum width and thickness- should be the same as
the dened channel cross section geometry
within the same depositional surface (or layer). These channel sections potentially
belong to the same channel object. To check if these channel sections belong to the
same channel, one calculates the centerline distances between channel sections in the
direction perpendicular to the channel orientation direction. If channel sections are
within a channel amplitude range, they belong to the same channel (Figure 2.20 left).
However, the same channel section (such as the section of well 3 in Figure 2.20) could
belong to a dierent channel (Figure 2.20 right). In this specic case we need to make
a decision to deterministically set the channel section to a channel. This means that
two interpretations are possible leading to two dierent interpreted hard data sets
(hard data uncertainty).
Once we know the conditioning data points for one channel, the actual channel
Figure 2.20: Well conditioning for interpreted channel sections. Well 1,2,3 could
be connected with one channel (left); However, well 3,4,5 can be connected within
another channel (right). Hence two interpretations are possible.
centerline is equated to the average value of channel sections center point locations.
Then the previous single channel simulation approach is applied to generate a channel
passing through these channel sections by means of conditional sequential Gaussian
simulation of the noise. Figure 2.21 shows the case where 10 channel sections inter-
preted from 10 wells belong to one channel (on the same depositional surface). In
the top row gure, 10 points indicate well locations and their color represent channel
section thicknesses at the well locations. The bottom gures are three realizations
conditioned to the top well thickness data. We observe that the larger thickness the
point has, the closer it is located to the 0-potential line which represents the channel
centerline. Figure 2.22 shows the 20-well case, the well data are perfectly honored
and multiple realizations can be obtained. Figure 2.23 shows the case where one
well vertically passes through dierent layers, and there are three interpreted channel
sections overlapping one another. Based on the interpreted thickness information,
three channels are simulated and the interpreted channel stacking patterns at the
well location is reproduced.
Figure 2.21: Three conditional realizations (bottom) of channel potential conditioned
to 10 wells (top)
2.1.5 Shale drapes modeling
As previously stated, shale drapes are associated with erosional bounding surfaces.
Although shale drapes may exhibit a varying degree of coverage, they are often very
thin (cm to m) compared to sand bodies. Therefore, ignoring their volume eects,
shale drapes will be simulated on their associated bounding surfaces in 2D space. The
simulated 2D shale drapes can be easily converted into 3D transmissibility multipli-
ers in ow simulation model for shale drapes eects study (Stright, 2005). In this
work, multiple-point statistics (MPS) program snesim will be used for shale drapes
modeling. To perform MPS, a shale drape training image representing the reservoir
conceptual shale drape distribution model is required. Figure 2.24 is one example of
Figure 2.22: Three conditional realizations (bottom) of channel potential conditioned
to 20 wells (top)
Figure 2.23: One realization (right column) of a channel complex conditioned to
interpreted channel stacking pattern (left column top) at well location
the shale drape training image. The red color indicates scour holes and the back-
ground represents shale drapes covered on the bounding surface. In fact, if shale
drapes only cover a small portion of the bounding surfaces (less than 40% areally),
the reservoir connectivity is not aected by these shale drapes (Li and White, 2003).
On the contrary, if bounding surfaces are fully draped by shales, and only a small por-
tion of scour holes are present among the shale drapes, the location and proportion of
scour holes will have signicant impact on reservoir connectivity (Li, 2003; Stright et.
al., 2006). This is because the holes may connect two channels if they are in contact
with each other. Therefore, instead of simulating shale drapes, we will simulate holes
distribution on the bounding surfaces. The modeling process is as follows (Figure
2.25):
1. Extract the bounding surfaces of the architecture elements (such as belts and
channels) from the simulated architectural model
2. Flatten these bounding surfaces into multiple 2D surfaces
3. Apply snesim on these 2D surfaces to generate holes
4. Fold back 2D surfaces with simulated holes into their original 3D space
This process is applied for each hierarchy. Finally a multi-scale shale drapes model
can be generated.
2.1.6 Summary of the architecture modeling
This section presented static geologic modeling techniques to simulate channel stack-
ing patterns and shale drapes distribution along multi-scale bounding surfaces in a
channelized reservoir. The complete static geologic modeling workow is shown in
Figure 1.5. In this workow, the large-scale reservoir architecture is modeled rst, the
shale drapes are then simulated within this architecture framework. The architecture
modeling approach adopted in this dissertation is essentially a stratigraphic-based
modeling approach. It is fast and can condition to well data, at the same time the
Figure 2.24: One example of shale drape training image. Red color represents holes,
and blue color indicates shales
Figure 2.25: The workow of shale drape modeling
channel stacking patterns can be reproduced. However, as we mentioned before, this
modeling process works under some assumptions. As a consequence, the quality of
results will be dependent on the agreement between the stated geological modeling
hypothesis and the actual observed data. The rst assumption is that the proba-
bility distribution functions of pattern parameters are available or can be derived
from some sources. We also assume that the net-to-gross ratio of each channel belt
is known. As a result, the number of channels within channel belts is determined
based on this net-to-gross ratio input. In practice, obtaining the pattern parameter
distribution functions that reect the geological pattern features is not an easy job;
and determining the net-to-gross ratio for each channel belt may be challenging.
2.2 Geologically consistent history matching
In the previous section, the reservoir architecture modeling technique and the shale
drapes modeling technique are introduced to integrate various geologic information
from dierent sources. However, building geologic models is just the rst step. The
next important step is to perturb these geologic models to match production data.
The perturbation should be consistent with the geologic concept and the static data.
Once a model has been history matched, it can be used to predict reservoir perfor-
mance or control reservoir performance with higher reliability. From the proposed
workow in Figure 1.6 we can see that there are two types of perturbation: channel
stacking pattern perturbation and shale drapes perturbation. The next section will
describe the perturbation techniques for these two properties and the history match-
ing procedure based on these geologically consistent perturbation mechanisms.
2.2.1 Channel stacking pattern perturbation
Before developing our geologically consistent perturbation technique, we need to state
clearly what is meant by geological consistency. For stacking channels, geological
consistency mainly means preserving the channel stacking pattern and channel geom-
etry. In this work, individual channels are continuous over the entire simulation
domain if they are not eroded by other channels. If a portion of channel is moved
from one location to another, other portions of the same channel should be moved in
order to preserve user-dened channel geometry parameters such as wavelength and
amplitude. Furthermore, because deep-water conned channel reservoirs often exhibit
a wide variety of stacking patterns (Clark and Pickering, 1996a, b), the perturbation
of the channel locations should not destroy the interpreted or observed stacking pat-
tern if such pattern is deemed correct. In other words, when we modify individual
channels, the modied channel should keep the same geometry parameters, and the
perturbed channel complex should preserve the stacking pattern believed to exist.
From the previous individual channel simulation process described in Section 2.1.1
we can observe that there are two components that potentially aect channel loca-
tions: (1) the noise map which has impacts on individual channel geometry; (2) the
channel centerline (0-isopotential line) location in the initial potential map which
directly controls where the channel is located. Figure 2.26 shows dierent channel
potential maps (bottom row) obtained by adding dierent noises (middle row) to the
same original potential map (top row). The 0-isopotential lines in the bottom row
potential maps of Figure 2.26 are the channel centerlines, and they can be used to
represent channel shapes. Because the same original potential map is used, the chan-
nel centerlines in the perturbed potential maps have similar location. But they have
dierent geometry due to the dierent noise realizations used.
Figure 2.27 shows the same noise map (top row) being added to the dierent
original potential maps (middle row). The nal curvilinear channels exhibit dierent
geometry and are located at dierent positions. In the proposed channel stacking
pattern simulation workow, the channel centerline locations are simulated by ran-
dom drawing of stacking pattern parameters. More precisely, a pair of migration ratio
and overlap ratio is rst simulated from their corresponding distribution models, then
converted to a channel centerline whose location is placed relative to the previously
generated channel. Therefore each channel is associated with a pair of simulated
stacking pattern parameters. These pairs of pattern parameter realizations can be
perturbed in order to change channel locations during history matching, at the same
time preserving the stacking pattern. The stacking pattern is fully dened by the
distribution of migration ratio and overlap ratio.
As observed from these two examples, modifying the channel centerline location
will result in a signicant change of the channel position, whereas deforming the noise
map will cause slight variations of channel geometry, but the channel centerline loca-
tion remains the same.
As previously stated, modifying a pair of pattern parameter realizations will
Figure 2.26: Dierent noise maps (middle row) added to the same original potential
map (top row) result in dierent channel geometries (0-potential line in bottom row)
Figure 2.27: Same noise maps (top row) added to dierent original potential map
(middle row) result in dierent channel geometries (0-potential line in bottom row)
change channel locations. This modication results in an entire channel object be-
ing modied because a pair of pattern parameter realizations determines one channel
centerline location. Obviously PPM is not suitable for moving channel centerline loca-
tions because PPM modications are essentially pixel-based not object-based. Hence
the gradual deformation method will be applied to modify channel location. To per-
turb channel geometry by modifying the noise map, this work will adopt the PPM
approach. These two perturbations will share the same optimization processes, i.e.
they will share the same perturbation parameter. The optimization process adopts
the Brent method because it does not require derivatives, hence the discontinuity in
the objective function behavior in GDM is not an issue here.
Gradual deformation of channel locations The key idea is to perturb channel
locations, but maintain the interpreted or observed channel stacking patterns. GDM
is applied to deform two channel stacking pattern parameters which are then used
to determine the deformed locations relative to adjacent channels. This perturbation
process is explained using a synthetic example with two uniformly distributed stack-
ing pattern parameters (see Figure 2.12).
1. Record the migration ratio (u
M
i
) and overlap ratio (u
O
i
) for each channel i
(here assuming these parameters are uniform distributed) and store them into
a parameter vector u=(u
M
1
,u
M
2
...u
M
n
,u
O
1
,u
O
2
...u
O
n
) (Figure 2.28)
Figure 2.28: Channel pattern parameters sampled from their distribution function
2. Transfer the recorded channel pattern parameters from uniform distribution
into Gaussian distribution (Figure 2.29) ;
Figure 2.29: A uniform distribution is transferred into Gaussian distribution
3. Generate another set of Gaussian realization for channel pattern parameters
and combine them with the ones in Step 2 (Figure 2.30) using equation:
y(r
D
) = y
1
cos(r
D

2
) +y
2
sin(r
D

2
) (2.1)
4. Transfer the deformed realization y in Step 3 back to uniform distribution (Fig-
ure 2.31)
5. Use these deformed ratios to obtain the locations relative to the adjacent sim-
ulated channel; simulate new channels centered at this location using the same
noise maps as the initial patterns.
Figure 2.32 shows a synthetic example of how the initial channel positions are de-
formed when dierent perturbation parameter values are applied. We notice that
with an increase in r
D
, the magnitude of perturbation of the channel position in-
creases, but the stacking pattern is maintained during perturbation. This is because
the deformed pattern parameters always have the same distribution functions as the
initial ones
Figure 2.30: Two Gaussian realizations are perturbed to generate one new realization
Figure 2.31: Transfer the perturbed realization back to uniform distribution
Figure 2.32: A chain of channel realizations with dierent perturbations applied to
the pair of channel stacking pattern parameters
Probability perturbation of channel geometry As we mentioned before, the
noise map aects the channel geometry. Probability perturbation can be applied to
modify the noise map. This modication will change channel geometry but keep the
channel centerline location xed. The noise is a continuous property, while tradi-
tional PPM is often applied to the categorical case, where a parameterized soft data
P(A|D) is used to create the perturbation. Hence, instead of calculating P(A|D)
and combining it with prior data using the tau model, collocated sequential Gaussian
simulation (cosgsim) is performed using a correlation coecient, with as the soft
data the previous best realization. Similar to the perturbation parameter, the cor-
relation coecient quanties the amount by which this current best realization is
perturbed. The correlation coecient and the perturbation parameter are related
through = 1 r
D
. When = 1 (r
D
= 0), the soft data is honored exactly, and
the current best realization is reproduced. Conversely when = 0 (r
D
= 1), the soft
data provides no information, and the realization is maximally perturbed into a new
equiprobable realization.
Figure 2.33 shows how the initial individual channel geometry changes with dif-
ferent perturbation parameter values (r
D
values). The channel centerline locations in
dierent perturbed maps (0-isopotential line in the left column gures) have roughly
the same location, but the channel geometry varies. As r
D
increases, the channel
geometry variation becomes larger.
Figure 2.34 shows multiple channel geometry deformations with dierent pertur-
bation parameters. The initial channel complex realization is the same as shown in
Figure 2.32. By perturbing the noise maps of each individual channel, the channel
geometries change. Although channel centerline locations remain almost constant as
demonstrated in Figure 2.33, the channel geometry variations will cause the channel
cross-section to move from one position to another. This results in the channel cross-
section changes (right column).
Figure 2.33: Four realizations of individual channel potential maps with dierent
perturbation parameters applied to their noise maps. Note the channel centerlines
are roughly same.
Figure 2.34: Channel complex realizations with dierent perturbation parameters
applied to the individual noise maps. Note the channel locations are roughly the
same (left column) but the channel cross-section changes with dierent perturbations
(right column, y=20).
2.2.2 Behavior of a chain of stacking pattern realizations
Two independent realizations (Figure 2.35 & 2.36) are provided to illustrate the pre-
vious methodology to perturb channel locations. The objective of the study is to
generate a chain of realizations between these two reference realizations with dier-
ent perturbation parameters and check the objective function variation.
Figure 2.35a and 2.36a show channel complex realizations with 7 channels contin-
uously distributed within a valley. The reservoir net-to-gross is 0.35. The dimension
of the reservoir grid is 70 100 60 with grid scale 25 25 1 m. Channels are uid
conduits with permeability 1000m and porosity 25%. The permeability and porosity
of the background are 5md and 5%, respectively. Shale drapes are simulated and
added onto the channel bounding surfaces to increase reservoir heterogeneity (Figure
2.35b & 2.36b). There are two producers and one injector (Figure2.35c). Due to
reservoir heterogeneity, these two producers have dierent water cut curves (Figure
2.35d and 2.36d).
Next the proposed perturbation method is applied to generate perturbed real-
izations with r
D
ranging from 0.1 to 0.9 (Figure2.37). In more detail, the vectors
u
1
and u
2
of pattern parameters used for reservoir models in Figure 2.35a and 2.36a
respectively are transformed into Gaussian domain as y
1
and y
2
. Then a series of
perturbations are applied to these two vectors using Eq.2.1 with r
D
varying between
0.1 and 0.9. The perturbed pattern parameters are then back-transformed into their
original distribution domain and used to determine channel locations. This results in
a chain of reservoir models with varying amount of perturbations. The shale drapes
along the channel bounding surfaces within each perturbed realization are simulated
using the same random seed as the reference models. Figure 2.38 shows the water
cut curves of two wells for these 10 cases. The objective function is dened as the
sum of the square dierences between the values of water cut at each time step of the
initial realization (Figure 2.35a) and those of each perturbed realization. Figure 2.39
shows the evolution of the objective function vs. the r
D
values. The variation of the
objective function is reasonably smooth which means that the perturbation can be
eectively applied to history match the model.
In summary, A method is presented in this section for history matching of the
Figure 2.35: Channel realization (a) and the shale drapes (b) along channel bound-
aries (blue color represents shale drapes and red color is for scour holes); wells are
located in the reservoir thickness map(c); (d) is the plot of water cut for two produc-
ers.
Figure 2.36: Channel realization (a) independently generated from the one shown in
Figure 2.35 and the shale drapes (b) along channel boundaries (blue color represents
shale drapes and red color is for scour holes); (c) is the reservoir thickness map, wells
are located at the same positions as shown in Figure 2.25c; (d) is the plot of water
cut for two producers.
Figure 2.37: Realizations with dierent perturbation parameters
Figure 2.38: Water cut curves for a chain of 10 realizations. The initial realization is
for the model in Figure2.35a
Figure 2.39: Objective function of a chain of 10 realizations
channel stacking patterns in channelized reservoirs. This method allows constrain-
ing channel stacking patterns to production data using gradual deformation method
and/or probability perturbation method. By adjusting certain aspects of geostatis-
tical channel reservoir model such as channel stacking pattern parameters and noise
map while history matching, the method makes it possible to reduce the uncertainty
in geostatistical models. Although gradual deformation method is not suitable for
discontinuous objective function, the optimization algorithm - 1D Brent method-
adopted in probability perturbation method will make it work for non-dierentiable
objective functions.
2.2.3 Shale drape perturbation
Shale drapes along channel belts and channels are simulated using MPS, where shale
drapes vs scour holes are modeled as a binary indicator variable ( Section 2.1.5). Since
PPM is a suitable perturbation method for MPS facies simulation (Caers, 2002; Ho-
man, 2005), this work will use PPM to perturb shale drapes for both channel belts
and individual channels. Stright (2006) demonstrated that shale drapes can be his-
tory matched using the Single-Region PPM method. In her work the whole reservoir
was treated as one single region and a single perturbation parameter r
D
is used for
entire reservoir. As a result, the whole reservoir is perturbed by a similar amount
(Homan, 2005). The same approach will be adopted in this dissertation when the
eld has few wells.
In the case that a eld has many wells, production data from one well (area) may
match history whereas production data from another well (area) may not. Hence per-
turbing the reservoir model globally may not be ecient. In fact, if many wells are
present, the single region PPM may not be ecient at all. To overcome this problem,
a Multi-Region PPM was proposed (Homan, 2005), which allows dierent amounts
of perturbation to be applied to various areas of the reservoir.
The workow of Multiple-Region PPM is discussed fully in Homans dissertation
(Homan, 2005). The idea of Multi-Region PPM is to perturb the properties that
inuence production data for one region dierently than the properties that aect
production for another region by optimizing perturbation parameter for each region.
The workow for Multi-Region PPM is similar to the Single-Region PPM, that is,
it also consists of an inner loop and an outer loop. The inner loop nds the op-
timum realizations between the initial realization and an equiprobable realization.
The outer loop consists of replacing the initial realization with the previous optimum
realization and changing the random seed. However, there are two dierences in op-
timization procedure: the rst dierence is that the objective function is calculated
dierently since multiple perturbation parameters are dened instead of one parame-
ter; consequently, the second dierence is that Multi-Region PPM optimizes multiple
parameters jointly while Single-Region PPM optimize one parameter.
In order to perform Multiple-Region PPM, the region geometry should be dened
rst. One option is to use a streamline method to dene regions. For a synthetic
reservoir model shown in Figure 2.40, it takes three steps (Figure 2.41):
1. The streamline simulation is run using the simulation model described previ-
ously. This will provide us the streamlines dening ow paths to production
wells .
2. Having the streamline distribution, we can identify a set of streamlines that go
to the same production well. All gridblocks hit by this set of streamlines are
assigned to that well. These gridblocks form a region that principally aect
this production well. For example, if we have three production wells, this cor-
respondes to three sets of streamlines and three regions are assigned to these
wells.
3. After we have identied regions in 3D model, the next step is to assign regions
to channel boundaries along which the shale drapes are present. This is done
by simply taking the channel boundaries in 3D channel region model.
After the regions are dened, the holes distribution can be perturbed on these bound-
ing surfaces using a Multi-Region PPM. One should notice that, becuase the hole
locations will change per realization, this region dening procedure should be done
at every iteration.
In some cases, dierent channel belts are formed in very dierent depositional
settings, such as dierent erosion energy, dierent bypass period. This could result
in dierent shale drape coverage for each channel belt and geological regions of reser-
voir model. In this case, the regions are xed for whatever channel and shale drapes
are distributed. In order to perform history maching eciently, we may need to as-
sign dierent r
D
to dierent geological regions or channel belts. This could be done
by performing sensitivity study to check which production well is most sensitive to
which geologic region, then assign r
D
to that region based on the objective function
calculcted from that well(s).
2.2.4 Perturbation procedure for history matching
The calibration of the perturbation parameter r
D
and the subsequent geologic model
updating requires the implementation of an interface between the geologic modeling
algorithm and the ow simulator. In this dissertation, the geologic modeling algo-
rithm is the main program that includes all the tasks in the probabilistic perturbation
approaches for dynamic data integration. The main program is implemented within
SGEMS framework using the Python script. The ow simulator (@ Eclipse) is also
executed within the main program.
Because of the hierarchic nature of the shale drapes in channelized reservoirs, the
Figure 2.40: A synthetic reservoir water saturation model with 20 channels and 3
producers.
perturbation will be performed hierarchically. For example, the shale drapes along
channel belt and the distribution of individual channels within channel belts are per-
turbed simultanously until a predened threshold A is reached, then the shale drapes
along the individual channels are perturbed to reach the nal threshold B which is
smaller than A. In this case, the perturbation has two stages: perturbation of larger
scale geologic features and perturbation of the small scale ones. Both stages share
the same probabilistic perturbation procedure with two-loop Markov chain (Homan,
2005). The inner loop nds the optimum realizations between the initial realization
and another equiprobable realization. The outer loop consists of replacing the initial
realization with the previous optimum realization and changing the random seed.
Within the inner loop, the perturbation is performed through the optimization of
a 1-D free parameter r
D
, that is between [0,1] and controls how much the initial
model is perturbed. Once we have a r
D
value, the shale drapes distribution will be
updated using PPM (Eqn.1.3) and the channel distribution will be updated using
Figure 2.41: The region dening procedure
GDM (Eqn.1.1). Then a new multiscale shale drape realization is generated that is a
perturbation from the original realization. There may exist a value of r
D
, such that
the corresponding realization matches the production data better than the initial re-
alization. Hence nding an optimum realization is equivalent to nding the optimum
r
D
value. The optimal r
D
value is found by minimizing the objective function,
r
Dopt
= min
_
O(r
D
) =
_
_
D
S
(r
D
) D
_
_
_
(2.2)
Where D
S
(r
D
) is the simulated data and D the observed eld data. O(r
D
), the
objective function, measures the mismatch between the simulated production data
and observed eld data. For N production variables over T time steps, the objective
function is :
O(r
D
) =
_
T
1
N
1
_
(Sim
i,t
Hist
i,t
)
Hist
i,end
_
2
(2.3)
At the end of each inner loop, the optimum realization will be used as the starting
realization for the next outer loop, and the random seed will be changed to introduce
new realizations. The iterative loop is repeated until a tolerance in the ojective func-
tion (mismatch) is met or a x number of outer iteration is reached. The implemented
code for history matching can be described in the following steps:
1. Perform proposed geologic modeling algorithim to stochastically generate an
initial realization;
2. A Markov chan iterative updating process is started with the initial realization.
The Markov chan forms the outer loop of updating. Every outer loop includes
following procedures.
Change random seed
Making an initial guess of r
D
;
Update probability model for shale drapes (PPM) and the channel pattern
parameters (GDM) for channels using the same r
D
. Generate a updated
realization;
Perform a 1-D optimization process to get the r
D
that result in best match
to the production data. This is a calibration process of r
D
with the pro-
duction data and it is called the inner loop. The step of inner iteration is
controlled by a x number.
Update the stochastic realization using the best model (with the minimum
objective function).
3. Repeat Step 2 (outer loop) until a tolerance in the objective function has been
reached or a x number of outer iteration is met.
Chapter 3
Workow testing on a synthetic
data set
The proposed modeling workow in Figure 1.6 is an iterative process of geologic mod-
eling and perturbation for history matching. The results of the workow depend on
the input modeling parameters such as the distribution functions of pattern parame-
ters, individual channel object geometry parameters or shale drape proportion since
these parameters will determine whether the geologic model built is realistic. It is
also dependent on the perturbation scheme for establishing the optimal r
D
because
dierent perturbation scheme may results in dierent rate of convergence. The sen-
sitivity of geologic parameters to the objective function also aects the convergence
rate, hence the nal result of the workow.
Synthetic reservoirs are useful for such an integrated modeling and ow simulation
study because the shale drape distributions are known. Using synthetic reservoirs, we
can check if the proposed geologic modeling technique reproduces channel stacking
patterns given the geologic reasonable input parameters. We can perform sensitivity
study to nd the most sensitive geologic parameters. Based on sensitivity study, we
can also evaluate the impact of dierent perturbation schemes on the convergence rate.
67
68 CHAPTER 3. WORKFLOW TESTING ON A SYNTHETIC DATA SET
3.1 Introduction to the synthetic dataset
Figure 3.1 is an analog of a deep-water incised canyon ll reservoir from oshore West
Africa, where well log and core data exhibit signicant vertical and lateral facies vari-
ation and channel-base shale drapes presence. This analog has 5 channel belts and 20
channels. The belt 1 to 5 is numbered from young (top in the model) to old (bottom
in the model). Repeated episodes of erosion, bypass, and deposition result in a com-
plex heterolithic ll and multiple signicant bypass surfaces that increase potential for
baes and barriers to ow within the canyon ll complex. In this synthetic reservoir,
shale drapes deposited during bypassing periods are discontinuously distributed along
individual channel boundaries and channel belt boundaries. These thin (cm to m),
multi-scale shale drapes are potential ow barriers that could compartmentalize the
reservoirs. Therefore this analog is an excellent tool to study modeling of multi-scale
shale drape distribution.
The size of this analog model is 3000m 2000m 64m. The grid dimension is
100 100 209. As shown in Figure 3.1, this analog is created using surface-based
gridding. The shale drapes are treated as thin layers and the individual channels
are thicker layers. Such exible gridding allows explicit capture of the shale drapes.
However, this ne scale gridding results in a large number of cells and will make the
ow simulation very time-consuming, and since both the channel location and shale
drapes location are uncertain, they need to be perturbed during history matching of
the reservoir model, which would mean changing the conformable grid automatically.
Currently, such gridding is not yet at an adequate level of robustness to be made
automatic. As a result, this surface-based ne-scale geologic model is rst converted
into a Cartesian grid, and then upscaled to a coarse grid with 50 50 50 cells.
The coarse grid is chosen based on the rule that the geometry and continuity of in-
dividual channels are deformed as little as possible (Stright, 2005). Figure 3.2 is the
upscaled geologic model with Cartesian grid. This model contains 20 channels. In
order to study the multi-scale shale drape distribution, the geometry of individual
channel belts is rst traced explicitly from the analog model (Figure 3.3) and then
their original shape is restored if they are eroded by the younger belts (Figure 3.4).
3.2. SENSITIVITY STUDY OF SHALE DRAPES PARAMETERS 69
The assumption here is that the channel belt geometry is known, for example, they
are identied from seismic data. Once the channel belt model is created, we have two
levels of heterogeneities caused by shale drapes: channel belts and individual channels
(Figure 3.5). Then we can simulate shale drapes along belt and channel boundaries,
and treat this model as a reference model. Next, we can test the proposed workow
on this synthetic case: simulate channel distribution within channel belt regions, sim-
ulate shale drapes along belt and channel boundaries, perturb channel location and
drape location to match the production data generated from our reference model.
Figure 3.1: A deep-water channelized reservoir analog. The brown and yellow colors
are for channel ll facies, green color is for shale drapes and blue color is for scour
holes. The model is constructed using surface-based grid.
3.2 Sensitivity study of shale drapes parameters
Shape drape size, geometry, proportion and location for both channel belts and chan-
nels have the potential to aect ow responses. In reality, all of these parameters
are uncertain with varying degree. Understanding their impact on uid ow behavior
will improve the reliability of geologic modeling and the eciency of history match-
ing. Once the sensitivity of these parameters are identied, this information can be
Figure 3.2: Analog model (upper) is converted into Cartesian grid (middle) and
upscaled to a coarse scale model (bottom). The color reprsents dierent channel
objects.
Figure 3.3: Coarse channel belt geometry identied from reference analog model.
Figure 3.4: The channel belt geometry is restored into original shape and their bound-
aries are traced for later channel simulation
Figure 3.5: Two level of heterogeneities: channel belts and channels
used to guide the data acquisition or collection to obtain reliable geologic modeling
parameters; it also can be used to help design the perturbation scheme by focusing
perturbation on the most sensitive parameters, resulting in an ecient history match-
ing.
Experimental design methodology has been used in reservoir engineering applica-
tions such as uncertainty modeling (Damsleth et al., 1992; Egeland, et al., 1992; Bu
and Damsleth, 1996), sensitivity studies (Jones et al., 1995; Willis and White, 2000;
White et al., 2001, 2003), performance prediction (Chu, 1990), upscaling (Narayanan,
1999), history matching (Eide et al., 1994) and development optimization (Dejean and
Blanc, 1999). Experimental design is a strategy in which the factors are varied si-
multaneously in a series of experimental runs (3D geologic modeling) according to a
predened design matrix to obtain the experimental response. Therefore, a design
is a set of factor-value combinations for which responses are measured (Myers and
Montgomery, 1995; Box and Draper, 1987). The input parameters that are varied
are called factors. The responses are obtained by measurement or modeling. Once
the designed simulation is completed, the response surface modeling can be used to
test the relative importance of the input factor to the output response statistically.
In this way, the experimental design methods ensure that accurate conclusions can
be drawn about the entire experiment with just a few experimental runs.
The rst step in a designed simulation study is to identify the factors that may
inuence responses. Factor ranges should include all feasible factor values. Factor
values are usually scaled and coded as factor levels (Montgomery, 2000) such as [-1,1]
for two-level factorial design. Then these coded factors are combined with specied
design scheme. 3D geologic models are generated based on all the possible factor com-
binations. Next, the responses are measured or modeled using these geologic models.
Factor-responses often have linear relationship. Therefore, this technique works well
for non-spatial factors because the response often behaves linearly. For spatial factor
such as facies type, experimental design is not suitable due to the discrete nature of
some of the input geostatistical parameters and the possible non-linear variation of
the ow response.
Because the primary goal in this work is to select or screen out the few important
main eects from the many less important ones, assuming their interaction an order
of magnitude less important, the screening design that is intended to nd the few
signicant factors from a list of many potential ones (Barrentine, 1999; Montgomery,
2001) is a suited choice. The routinely used screening design is the Plackett-Burman
(PB) design (Friedmann et al, 2001, 2003). The PB design can be used to study n 1
factors in n runs in which n is dividable by four. That is, 7 factors can be studied
in 8 runs, 11 factors in 12 runs, etc. Therefore, such design can practically handle
a large number of factors. It is a two-level design which allows ecient estimation
of main eects of all factors being studied, ignoring the factor interactions. In this
two-level factorial design, each factor is assigned to its maximum or minimum value
(1) in all possible combinations with other factors.
Factors considered Seven non-spatial geologic parameters were examined (Ta-
ble 3.1). Channel belt scour hole size (BS), channel scour hole size (CS): the area
of individual scour holes for both belts and channels are varied from small to large
(Figure 3.6). The small hole size is coded as lower level -1, and the large hole size is
coded as higher level +1. Hole shape for belts (BG) and channels (CG): round shape
and ellipsoid shape are considered for both belts and channels. The round shape is
arbitrarily coded as level -1 and the ellipsoid as level +1. Hole proportion for belt
(BP) and channels (CP): the areal coverage of shale drapes along bounding surfaces
is varied from 0.1 to 0.5. The lower limit 0.1 is coded as level -1 and the upper limit
0.5 is coded as level +1. Channel location (CH): except the analog model, another
channelized reservoir model is simulated which also has 20 channels but channels
have dierent distribution than the analog (Figure 3.6). The reference analog model
is assigned as lower level -1, and the new generated one as higher level +1.
Flow responses examined Three ow responses were checked: water break-
through time (WT) which is the days when the water cut reaches 0.1; water cut (WC)
after 900 days production and eld oil recovery (FOE) after 900 days production.
Experiment design An eight-run (factor combinations) Plackett-Burman was
selected for the experimental design study, and 6 realizations are generated for each
run. Table 3.2 shows the PB design for these 7 factors. The symbol + indicates
higher level (+1), and - refers to lower level (-1).
Flow simulation model The reservoir simulation models have the same di-
mension as the coarse analog model described in Section 3.1. In order to perform
stochastic simulation of shale drapes along bounding surfaces, shale drape training
images are rst generated according to the combination shown in Table 3.2 (Figure
3.6). The MPS program snesim was performed to simulated shale drape distribution
along channel belt boundaries and channel boundaries.
The shale drapes property is explicitly included into simulation model. If channel
boundary and belt boundary overlaps (share the same grid) and both boundaries
have holes, the transmissibility multiplier for this grid will be 1, otherwise 0. But
if channel boundary does not tie to belt boundary, when holes are present on belt
boundaries or channel boundaries, the grid multiplier will be 1.
One injector injects water into one side along the channel direction (end of north
side), one producer produces oil from the other side of the model. The channel
reservoir has constant porosity (20%) and permeability (1000 mD). The model is a
two-phase oil-water system. The viscosities of oil and water are 1.0 centipoise (cp).
The density of the oil is 40lb/ft
3
and the water density is 62.24lb/ft
3
. The oil-water
relative permeability curves are shown in Figure 3.7. To observe the holes eect, the
oil and water mobility is set to 1 and there is no gravity and capillary pressure are
present.
Analysis The rst step in analyzing this experiment is to convert the data table
into a spreadsheet format. Table 3.3 ~ Table 3.5 are the responses for dierent factor
combination runs and the corresponding analysis results. The rst left column num-
ber 1, 2...8 is for 8 runs or factor combinations. Each factor combination is recorded
in the following column as + and -. For each run or factor combination, 6 statisti-
cally generated geological models are ow simulated, the corresponding responses are
recorded afterwards as Obervations 1, 2...6. Based on these recorded responses, the
sensitivity analysis are performed as follows:
Step1: Calculate the mean and variance of the 6 sample data point for each run.
The means are recorded in the row of WC in Table 3.3, WT in Table 3.4 and FOE
in Table 3.5. The variances are recorded in the row of S
2
in the three tables.
Step2: Sum up the mean values that correspond to higher level(+) for each fac-
tor, and record them in the row of
WC
+
,
WT
+
and
FOE
+
in Table 3.3 ~
Table 3.5 correspondingly; Sum up the mean values that correspond to lower level(-)
for each factor, and put them in the row of
WC
WT
and
FOE
.
Step3: Calculate the mean of summed means for level low and high for each fac-
tor, and put them into the row of WC
+
and WC
in Table 3.3, WT
+
and WT
in
Table 3.4, FOE
+
and FOE
in Table 3.5.
Step4: Calculate main eect from analysis tables, which is the dierence between
the mean of summed means for higher level (+) and lower level (-). For example,
the main eect of belt hole size (BS) for water cut (WC) in Table 3.3 is: 0.730-0.745
=-0.015.
Step5: Create the Pareto chart of eects (Figure 3.8).
Step6: Calculte the estimated standard deviation S
e
:
S
e
=
_
_
k
i=1
S
2
i
/k
_
(3.1)
where k is the number of runs, in this case k=8.
Step7: Calculate the eective standard deviation S
eff
:
S
eff
= S
e
_
4/N (3.2)
where N is the total number of sampled data. It is the product of the number of runs
(k) and the replicates for each run. In this case N = 8 6 = 48.
Step8: Determine degrees(df) of freedom and t-statistic.
df=(# of replicates per run -1)(# of runs)
In this case, df=(5)(8)=40 and t
df=40;=0.05
=2.021
Step9: Compute decision limit based on t and S
eff
.
DL = tS
eff
(3.3)
In this work, the decision limits are calculated at the 95% level of condence
(=0.05). Any eects shown absolute value greater than the decision limit(DL) was
considered statistically signicant on the responses. As shown in Table 3.3 ~ Table
3.5, for water cut, hole shape for channel belt(BG), hole size for channels(CS), hole
proportion for channels (CP) and channel locations (CL) are signicant factors that
aect individual well water cut prole at the 95% level of condence. Among these
four factors, the hole proportion for channels (CP) and channel locations (CH) are the
most signicant factors ( Fig 3.8). For the water breakthough time, except hole size for
channels (CG), other six factors are signicant at the 95% level of condence, and the
hole proportion for channels(CP) is the most signicant factor, the channel locations
(CH) the second signicant one (Fig 3.8). For eld oil recovery rate, all these seven
factors are signicant at the 95% level of condence, and the most signicant factor
is hole proportion for channels (CP), the scond signicant one is channel location
(CH) (Fig 3.8). In summary, the hole proportion for individual channels and the
individual channel location are the most sensitive factors that aect reservoir uid
ow behavior in this case. This information will help the following history matching
process performed in a ecient manner: only perturb the most sensitive geologic
factors, that is, hole proportion for channel boundaries and channel location.
Figure 3.6: Shale drape training images (top 2 rows) and two reservoir models. In
training images, the red color represents scour holes and blue background the shale
drapes. Round vs. ellipsoid is for hole geometry, small vs. large is for hole size, and
0.1 vs. 0.5 is for hole proportion.
Figure 3.7: Relative permeability for ow simulation model.
Figure 3.8: Eect charts for ow responses.
Table 3.1: Seven factors and level set up
Factor
Factor level
-1 1
Hole size of belt(BS) Small Large
Hole size of channel(CS) Small Large
Hole shape of belt(BG) Round Ellipsoid
Hole shape of channel(CG) Round Ellipsoid
Hole proprtion of belt(BP) 0.1 0.5
Hole proportion of channel(CP) 0.1 0.5
Channel location(CH) Analog model Modied analog model
Table 3.2: Plackett-Burman design with 7 factors
Run BS BG BP CS CG CP CH
1 + - - + - + +
2 + + - - + - +
3 + + + - - + -
4 - + + + + - +
5 + - + + + - -
6 - + - + + + -
7 - - + - + + +
8 - - - - - - -
T
a
b
l
e
3
.
3
:
W
a
t
e
r
c
u
t
(
%
)
r
e
s
p
o
n
s
e
@
9
0
0
d
a
y
s
f
o
r
P
B
e
x
p
e
r
i
m
e
n
t
r
u
n
s
a
n
d
a
n
a
l
y
s
i
s
R
u
n
B
S
B
G
B
P
C
S
C
G
C
P
C
H
O
b
s
e
r
v
a
t
i
o
n
s
W
C
S
2
1
2
3
4
5
6
1
+
-
-
+
-
+
+
0
.
6
2
7
0
.
6
3
3
0
.
6
4
4
0
.
5
9
9
0
.
6
0
0
0
.
6
4
8
0
.
6
2
5
0
.
0
0
0
5
2
+
+
-
-
+
-
+
0
.
7
0
6
0
.
6
9
2
0
.
7
3
9
0
.
8
0
6
0
.
7
1
3
0
.
7
7
1
0
.
7
3
8
0
.
0
0
1
9
3
+
+
+
-
-
+
-
0
.
7
2
4
0
.
7
4
5
0
.
6
9
8
0
.
6
9
5
0
.
7
1
8
0
.
7
5
2
0
.
7
2
2
0
.
0
0
0
6
4
-
+
+
+
+
-
+
0
.
7
4
7
0
.
7
7
9
0
.
7
9
1
0
.
7
9
3
0
.
7
8
3
0
.
7
4
2
0
.
7
7
3
0
.
0
0
0
5
5
+
-
+
+
+
-
-
0
.
6
8
8
0
.
8
9
8
0
.
8
8
2
0
.
8
6
0
0
.
8
6
7
0
.
8
0
4
0
.
8
3
3
0
.
0
0
6
1
6
-
+
-
+
+
+
-
0
.
7
9
7
0
.
7
9
0
0
.
7
5
8
0
.
7
3
5
0
.
7
4
7
0
.
8
0
9
0
.
7
7
3
0
.
0
0
0
9
7
-
-
+
-
+
+
+
0
.
5
9
8
0
.
6
0
4
0
.
6
0
6
0
.
6
3
0
0
.
6
0
9
0
.
5
8
4
0
.
6
0
5
0
.
0
0
0
2
8
-
-
-
-
-
-
-
0
.
7
1
9
0
.
8
9
7
0
.
8
8
8
0
.
8
8
4
0
.
7
6
7
0
.
8
1
9
0
.
8
2
9
0
.
0
0
5
4
W
C
+
2
.
9
1
8
3
.
0
0
5
2
.
9
3
3
3
.
0
0
4
2
.
9
4
9
2
.
7
2
5
2
.
7
4
1
W
C
2
.
9
7
9
2
.
8
9
3
2
.
9
6
5
2
.
8
9
4
2
.
9
4
9
3
.
1
7
3
3
.
1
5
7
W
C
+
0
.
7
3
0
0
.
7
5
1
0
.
7
3
3
0
.
7
5
1
0
.
7
3
7
0
.
6
8
1
0
.
6
8
5
W
C
0
.
7
4
5
0
.
7
2
3
0
.
7
4
1
0
.
7
2
4
0
.
7
3
7
0
.
7
9
3
0
.
7
8
9
E
e
c
t
-
0
.
0
1
5
0
.
0
2
8
-
0
.
0
0
8
0
.
0
2
7
0
.
0
0
0
-
0
.
1
1
2
-
0
.
1
0
4
S
e
0
.
0
4
5
S
e
f
f
0
.
0
1
3
d
f
4
0
0
.
0
5
t
2
.
0
2
1
D
L
0
.
0
2
6
T
a
b
l
e
3
.
4
:
W
a
t
e
r
b
r
e
a
k
t
h
r
o
u
g
h
t
i
m
e
(
d
a
y
s
)
f
o
r
P
B
e
x
p
e
r
i
m
e
n
t
r
u
n
s
a
n
d
a
n
a
l
y
s
i
s
R
u
n
B
S
B
G
B
P
C
S
C
G
C
P
C
H
O
b
s
e
r
v
a
t
i
o
n
s
W
T
S
2
1
2
3
4
5
6
1
+
-
-
+
-
+
+
7
1
2
7
0
8
6
8
8
7
0
6
7
0
4
7
1
9
7
0
6
1
0
7
.
4
2
+
+
-
-
+
-
+
4
3
0
4
1
8
4
4
1
4
0
9
4
3
7
3
6
8
4
1
7
7
2
2
.
2
3
+
+
+
-
-
+
-
5
2
6
5
1
6
5
2
8
5
1
6
4
9
8
5
1
6
5
1
7
1
1
3
.
1
4
-
+
+
+
+
-
+
4
7
5
4
1
0
4
5
8
4
0
4
3
7
7
4
2
7
4
2
5
1
3
1
2
.
6
5
+
-
+
+
+
-
-
3
0
0
3
3
5
3
1
6
3
3
8
2
0
8
3
1
1
3
0
1
2
2
9
9
.
9
6
-
+
-
+
+
+
-
4
5
5
4
6
4
4
1
7
4
1
7
4
4
8
3
8
0
4
3
0
9
8
8
.
6
7
-
-
+
-
+
+
+
7
2
3
7
4
0
7
2
4
6
9
7
7
1
4
7
2
8
7
2
1
2
0
9
.
6
8
-
-
-
-
-
-
-
3
5
3
2
7
6
2
6
5
3
0
7
2
7
2
2
6
1
2
8
9
1
2
4
7
.
6
W
T
+
1
9
4
1
1
7
8
9
1
9
6
4
1
8
6
3
1
8
7
0
2
3
7
4
2
2
7
0
W
T
1
8
6
5
2
0
1
8
1
8
4
3
1
9
4
4
1
9
3
7
1
4
3
3
1
5
3
7
W
T
+
4
8
5
4
4
7
4
9
1
4
6
6
4
6
7
5
9
4
5
6
7
W
T
4
6
6
5
0
4
4
6
1
4
8
6
4
8
4
3
5
8
3
8
4
E
e
c
t
1
9
-
5
7
3
0
-
2
0
-
1
7
2
3
5
1
8
3
S
e
2
9
.
6
S
e
f
f
8
.
5
d
f
4
0
0
.
0
5
t
2
.
0
2
D
L
1
7
.
3
T
a
b
l
e
3
.
5
:
F
i
e
l
d
o
i
l
r
e
c
o
v
e
r
y
e
c
i
e
n
c
y
(
%
)
@
9
0
0
d
a
y
s
f
o
r
P
B
e
x
p
e
r
i
m
e
n
t
r
u
n
s
a
n
d
a
n
a
l
y
s
i
s
R
u
n
B
S
B
G
B
P
C
S
C
G
C
P
C
H
O
b
s
e
r
v
a
t
i
o
n
s
F
O
E
S
2
1
2
3
4
5
6
1
+
-
-
+
-
+
+
0
.
5
7
6
0
.
5
7
3
0
.
5
7
7
0
.
5
7
8
0
.
5
7
8
0
.
5
8
1
0
.
5
7
7
0
.
0
0
0
2
+
+
-
-
+
-
+
0
.
4
7
2
0
.
4
8
5
0
.
4
7
0
.
4
3
1
0
.
4
7
3
0
.
4
2
1
0
.
4
5
9
0
.
0
0
0
7
3
+
+
+
-
-
+
-
0
.
5
5
8
0
.
5
5
6
0
.
5
7
5
0
.
5
6
6
0
.
5
6
3
0
.
5
3
8
0
.
5
5
9
0
.
0
0
0
2
4
-
+
+
+
+
-
+
0
.
4
7
9
0
.
4
3
6
0
.
4
5
4
0
.
4
1
7
0
.
4
2
7
0
.
4
6
2
0
.
4
4
6
0
.
0
0
0
5
5
+
-
+
+
+
-
-
0
.
4
5
5
0
.
4
0
9
0
.
3
9
7
0
.
4
3
3
0
.
2
8
9
0
.
4
2
6
0
.
4
0
2
0
.
0
0
3
4
6
-
+
-
+
+
+
-
0
.
5
1
0
0
.
5
1
0
0
.
5
1
4
0
.
5
0
0
0
.
5
1
5
0
.
4
6
3
0
.
5
0
2
0
.
0
0
0
4
7
-
-
+
-
+
+
+
0
.
5
9
1
0
.
5
8
9
0
.
5
8
7
0
.
5
8
1
0
.
5
8
9
0
.
6
3
0
.
5
9
5
0
.
0
0
0
3
8
-
-
-
-
-
-
-
0
.
4
5
3
0
.
3
5
7
0
.
3
3
7
0
.
3
3
9
0
.
3
8
1
0
.
3
5
8
0
.
3
7
1
0
.
0
0
1
9
F
O
E
+
1
.
9
9
7
1
.
9
6
6
2
.
0
0
1
1
.
9
2
7
1
.
9
5
7
2
.
2
3
3
2
.
1
7
7
F
O
E
1
.
9
1
3
1
.
9
4
4
1
.
9
0
9
1
.
9
8
3
1
.
9
5
3
1
.
6
7
7
1
.
8
3
4
F
O
E
+
0
.
4
9
9
0
.
7
3
7
0
.
5
0
0
0
.
4
8
2
0
.
4
8
9
0
.
5
5
8
0
.
5
4
4
F
O
E
0
.
4
7
8
0
.
4
8
6
0
.
4
7
7
0
.
4
9
6
0
.
4
8
8
0
.
4
1
9
0
.
4
5
8
E
e
c
t
0
.
0
2
1
0
.
0
0
5
0
.
0
2
3
-
0
.
0
1
4
0
.
0
0
1
0
.
1
3
9
0
.
0
8
6
S
e
0
.
0
0
0
9
S
e
f
f
0
.
0
0
0
3
d
f
4
0
0
.
0
5
t
2
.
0
2
D
L
0
.
0
0
0
5
3.3 Modeling parameters
As described in Chapter 2, in order to perform the proposed geologic modeling
method, the input parameters such as channel geometry parameters (wavelength,
amplitude etc.) and channel stacking pattern parameters (migration ratio, overlap
ratio) need to be specied rst. Generally, it is not easy to obtain information on
these parameters from seismic data because the objects are often below seismic res-
olution. Outcrop study can provide some data but outcrop data is often limited to
2D. For this synthetic study, we are in the ideal situation that parameters required
for geologic modeling are available.
In order to obtain the information about the channel geometry, the individual
channels are extracted from the 3D synthetic model. For each channel, ve geome-
try parameters including amplitude (A), wavelength (), orientation (), maximum
width (W) and thickness (H) are measured. In this synthetic model, there are 20
channels. Hence, there are 5 sets of 20 sample measurements for these ve geometry
parameter. The averaged values of each set samples for these geometry parameters
are shown in Figure 3.9. For stacking pattern parameters, each individual chan-
nel centerline is identied, their locations in x- direction (channel orientation is in
y-direction) are recorded. Then, the horizontal distance (x) between two adjacent
channels is calculated which is the dierence between their horizontal location val-
ues. And the migration ratios are derived by dividing the channel amplitude (A) to
horizontal distance data (x). To obtain the overlap ratio data, every channel top
surface z is recorded. Then, the vertical distance h
between two adjacent channels is

calculated which is the dierence between their surface z values. The vertical overlap
ratio is calculated by dividing maximum thickness (H) to (H-h
). For 20 channels
within 5 belts, there are 6 sample values for both overlap ratio and migration ratio.
The cumulative distribution functions of these two parameters are calculated using
each 6 data points (Figure 3.10 black dots). A regression is applied to obtain analytic
cdf functions for these two parameters (Figure 3.10 solid line). Therefore, the cdf
3.4. HISTORY MATCHING RESULTS 85
function for migration ratio (MR) used in this example is as follows:
p = 0.7113ln(MR) + 1.9486, MR [0.06, 0.27]
The cdf functions for overlap ratio (OR) is:
p = 4.9693(OR) 3.9576, OR [0.79, 1]
In addition of channel geometry and pattern parameters, the net-to-gross (NTG)
ratio is required for each channel belt. For simplicity, channel is treated as sand
facies (NTG=1) and background shale is non-sand (NTG=0). After restoring the
eroded parts of the channels, the corresponding net-to-gross is calculated for channel
belt 1 ~ 5 respectively (Table 3.6).
Table 3.6: Global NTG for dierent channel belts of reservoir analog
Channel belt Global NTG
Belt 1 0.68
Belt 2 0.65
Belt 3 0.65
Belt 4 0.68
Belt 5 0.60
3.4 History matching results
The proposed modeling and history matching procedure is applied to perturb multi-
scale shale drape/hole locations to match the historical production data for this syn-
thetic reservoir. The reference channel distribution model is shown at the bottom in
Figure 3.2. The shale drape models for channel belts and channels are generated us-
ing the shale drape modeling procedure presented in Chapter 2 (Figure 3.11). With
the same ow simulation parameters as sensitivity study, the reference ow model
is set up with two injectors and three producers. The injectors were controlled by
Figure 3.9: Channel geometry parameter derived from analog
Figure 3.10: Channel pattern parameter distribution: the experimental CDF (blue
dots) and the analytic CDF function obtained by regression (solid lines).
surface injection rate, and the producers were controlled by total liquid rate. Figure
3.12 shows the eld oil saturation model (top) and three producers ow responses
(bottom) for 1000 days production. These production proles are treated as the ob-
served data. The rst 600 days production proles will be history matched, and the
last 400 days production prole will be used to test the prediction reliability of the
history matched geologic models. The objective function for the history match was
dened using the sum of dierence in observed and simulated water cut and bottom
hole pressure at the three producers. During the perturbation process, the geologic
model is generated using the modeling parameters derived in previous section. The
geologic modeling will be constrained to hard data from 5 wells (Figure 3.16).
The perturbation is performed in two steps: rst the channel location and hole
location along belt boundaries are modied simultaneously (using same r
D
). Note
that the hole location along individual channels is not perturbed, but since each outer
loop changes the random seed for hole simulation, their location changes randomly
due to a change of random seed input into the MPS algorithms that generate the
hole; Secondly, the hole location along channel boundaries is perturbed while chan-
nel location and hole distribution along belts remain xed. During the perturbation
of hole location along channel boundaries, the hole models are simulated using the
hole training images shown in the top row of Figure 3.11. This means the pertur-
bation only modies hole location, the hole proportions is taken as the same one of
the reference. Both steps have 10 outer iterations. During the whole perturbation
process, the hole proportion for channels and belts are the same as the reference. In
other words, the hole proportion is not modied during history matching. Figure
3.13 shows the optimization performance for these two steps. In the rst step (purple
dots), the mismatch drops down from 4.2 to 0.87, while in the second step (green
dots), the mismatch slightly decreases from 0.87 to 0.85. The reason why the rst
step objective function drops considerably more than the second one is because the
channel location mainly controls the hole location along the channel boundaries. In
the rst step perturbation, when the channel location is modied to optimally match
production data, as a consequence, the hole location is optimized with the channel
location. As a result, the mismatch decreases rapidly. The perturbation in the second
Figure 3.11: The 3D hole distribution models (middle row) and the cross sections
(bottom row) for belts and channels (middle row) using the corresponding hole train-
ing images (top row). The dark red indicates holes. The training image grid dimension
is 100 by 100 and the hole models are 50 50 50.
Figure 3.12: Observed production responses from the synthetic reservoir.
step is to ne tune the hole location based on the rst perturbation result to further
match the production data. This ne tuning process often requires a large portion of
the CPU time to reduce a small amount of mismatch. Therefore, it is suggested that
in most case the rst step perturbation may be enough to get a satisfactory history
match.
Figure 3.14 shows the history matched results corresponding to the optimization
performance shown in Figure 3.13. Figure 3.15 compares the history matched channel
distribution model with the reference. These two models have very similar stacking
patterns. The reference has 20 channels while the history matched one has 21 chan-
nels. Figure 3.16 checks the well data conditioning in the history matched model.
It shows the simulated facies at ve well locations are consistent with the well data.
Note that this is true because the well hard data interpretation is exact (not true in
reality).
Once the history matched geologic models are obtained, prediction is assessed for
the next 400 days. Figure 3.17 and Figure 3.18 are prediction plots of well water
cuts, bottom hole pressure and eld oil recovery for multiple history matched cases
(left column) and non-history matched cases (right column). Comparing with the
non-history matched cases, the history matched models provide narrow ranges of the
future water cuts and eld oil recovery around the reference curves. This clearly
demonstrates that the history matched models have the capability to provide a more
reliable prediction.
In the above example, we assume the hole proportions are known and only the
hole location is uncertain. In practice, it would be dicult to obtain an accurate
target hole proportion from data. Moreover, as shown in previous sensitivity study,
the proportion of holes along individual channels is one of the most sensitive factors
for ow responses. Therefore, one will often need to perturb both hole location and
proportion to match history. In the next history matching example, with the same
production data and well congurations as the example above, a hole training im-
age (Figure 3.19) with target proportion dierent from the reference is used. Hence,
for channels both hole location and proportion need to be perturbed, while for belts
only the hole location is perturbed since it is less sensitive to the ow response. The
history matching procedure now consists only of the rst perturbation step (with 10
outer iterations), that is, perturbs the hole location for belts, the channel location
and the hole proportion for channel boundaries. Figure 3.20 shows the optimization
performance. The blue curve reects the inner loop mismatch, the purple curve the
outer loop mismatch and the green curve the channel hole proportion evolution during
perturbation. The mismatch between the simulated ow responses and the historical
data decreases considerably. Meanwhile, the channel hole proportion converges from
its original 20% to the reference 10%. The initial guess and nal history matched
results, as well as the corresponding predictions are shown in Figure 3.21. Again, the
history matched model has prediction results closer to the reference compared with
the randomly picked models.
Figure 3.13: Optimization performance for two step perturbations.
Figure 3.14: History match results for two step perturbation.
Figure 3.15: Reference model vs. history match geologic model.
3.5. CHAPTER SUMMARY 93
Figure 3.16: Well data vs. simulated facies at well locations. Dierent color represents
dierent channel section.
3.5 Chapter summary
The synthetic example studied in this chapter is a very realistic representation of reser-
voir architecture for deep-water conned channel system from oshore West Africa.
Hence, it can be used to test the feasibility of proposed modeling and history match-
ing workow. We can also derive the channel stacking pattern information from this
analog for future real case. The sensitivity study will help set up ecient perturba-
tion scheme in the real history matching process.
The sensitivity study shows that the hole proportion along individual channels
and channel location are the most sensitive parameters for ow responses, though
other parameters are also signicant with 95% level of condence. This means when
we perturb the geologic model for history matching, the ecient perturbation pa-
rameters are channel location and channel hole proportion. The history matching
results also show that performing one step perturbation of channel location and hole
proportion is enough to obtain desired history-matched geologic models. The second
step perturbation of channel hole location has very slow convergence rate because
the channel location which is associated with hole location is already optimized in
the rst step perturbation. The history matching and prediction results have demon-
strated that the proposed workow can be used for modeling and history matching
of multi-scale shale drapes. The nal history matched geologic models have better
prediction power than randomly selected geologic models.
Figure 3.17: Well water cut and eld oil recovery predictions using the history
matched models and randomly picked models.
Figure 3.18: Well bottom hole pressure predictions using the history matched models
and randomly picked models.
Figure 3.19: A channel hole training image (right) with proportion dierent with the
reference (left).
Figure 3.20: The optimization performance for the case that both hole location and
proportion are perturbed. The hole proportion converges to the reference with the
mismatch decreases.
Figure 3.21: Flow response Predictions using the history matched models and ran-
domly picked models.
Chapter 4
Applications to a realistic turbidite
reservoir
The dominant reservoir types in oshore West Africa (WA) are large, erosionally
conned deepwater channel complexes developed within slope valley (Abreu, et al.,
2003). High resolution seismic data and extensive outcrop studies have improved our
understanding of these types of deepwater channels. Numerous and comprehensive
studies have focused on the specic aspects of channel morphology, deposition process
and stacking patterns. These large erosinally conned channels are typically 50-100m
thick, several kilometers wide and long. They are characterized by a repeated cutting
and lling process. This process has very important implication for reservoir connec-
tivity (Mayall and OByrne, 2002). The facies at the base of each channel is critical
to the connectivity across larger erosional belts. If these facies are composed of shale
drapes they may result in barriers within reservoir. Another feature for deep-water
channelized reservoir is that they tend to be sand-rich. They often have very high
(0.6-0.8) net-to-gross ratio (NTG).
This chapter will study a reservoir model based on a real turbidite reservoir in
oshore West Africa. The valley and channel belt geometries are interpreted from
3D seismic data. However, the depositional facies lling the belts cannot be easily
inferred from the seismic data due to the limited resolution. The net-to-gross ratios
for channel belts are obtained from the company database of analog reservoirs. The
97
98 CHAPTER 4. APPLICATIONS TO A REALISTIC TURBIDITE RESERVOIR
production data and core data from the real reservoir provide ample evidence of shale
drape presence. These drapes act as barriers to compartmentalize the reservoir. But
the proportion and spatial distribution of shale drapes in the reservoir are uncertain.
The objective of this chapter is to apply the proposed geologic modeling and history
matching workow to perform the history matching study.
First, the information available for this study is described. Next, we demonstrate
that, in case of high NTG reservoir, perturbing shale drapes in the true 3D domain
while remaining consistent with its geologic conceptual description leads to more e-
cient and favorable history matching results as compared to perturbing in pseudo-3D
domain. Finally, a process of perturbing the geologic model by regions with unknown
shale drape proportion in each region is conducted and the results are presented. To
assign dierent perturbation parameters r
D
to dierent regions, a sensitivity study is
performed to relate the producers to dierent geologic regions.
4.1 Information available for history matching
In this WA reservoir, the seismic data quality is of sucient quality to identify strati-
graphic architecture including valley and belts framework (Figure 4.1). Figure 4.1
(a) shows the surface-based model of valley and belt geometry interpreted from
seismic data. The grid dimension is 100 200 34. The reservoir dimension is
4200 7200 100m. Figure 4.1 (b) shows the upscaled Cartesian grid model with
grid dimension as 505050. This Cartesian grid will be used for geologic modeling
and history matching. The reservoir volume and individual channel belt volumes are
deemed known and retain for this study. In other words, these channel belts are taken
as containers for the following individual channel lling with pre-dened conceptual
staking patterns. Based on geologic studies of well logs, this reservoir is classied into
two regions. Region 1 is the within valley region. This region has higher NTG (0.8)
and less shale drapes along belt and channel edges. Region 2 is formed by the parts
outside of the valley. This region has slightly less NTG (0.7) and more shale drapes.
In order to ll the individual channels into channel belt containers, the channel
4.1. INFORMATION AVAILABLE FOR HISTORY MATCHING 99
Figure 4.1: Stratigraphic interpretation of WC reservoir. Gray color is for valley
region. Other colors are for dierent channel belts.
100CHAPTER 4. APPLICATIONS TO A REALISTIC TURBIDITE RESERVOIR
belts NTG need to be known. Based on the company database of analog reservoirs
as well as well logs interpretation, the NTG for each belt are obtained and listed in
Table 4.1. These NTG values are assumed certain. This study will make use of the
channel geometry information (Figure 3.9) and channel stacking pattern information
(Figure 3.10) derived from reservoir analog in Chapter 3. The only changes are chan-
nel dimensions. To adapt the channel dimension to this realistic reservoir dimension,
the channel maximum thickness is increased from 15m to 18m, the maximum width
is increased from 300m to 500m.
Well-log interpretation from the real reservoir provides the range of porosity
Table 4.1: NTG for dierent channel belts
Channel belt NTG
Belt 1 0.8
Belt 2 0.8
Belt 3 0.8
Belt 4 0.75
Belt 5 0.75
Belt 6 0.75
Belt 7 0.75
Belt 8 0.7
and permeability for channel facies. Based on these information, the histograms of
porosity and permeability are generated. The porosity for channels follows normal
distribution with mean 0.28 and standard deviation 0.01. The permeabilities are also
normally distributed with mean 2200md and standard deviation 200md. The back-
ground shales have very low permeability (mean 25md) and porosity (mean 0.03). In
this study porosity and permeability are simulated using Sequential Gaussian Sim-
ulation (sgsim) by facies (channel/non-channel). The variograms listed in Table 4.2
are used to simulate both porosity and permeability distribution, for both channel
and non-channel facies. The nal porosity and permeability models are obtained after
cookie-cutting the porosity and permeability realizations using channel distribution
realization.
4.2. SETTING UP TRUE CASES 101
Table 4.2: Variograms used for channel porosity and permeability simulation
Channel Background shale
Type Exponential Exponential
Nugget 0 0
Ranges 2000m/800m/20m 1000m/1000m/10m
Angle 0/0/0 0/0/0
Shale drapes distribution along valley, belt and channel boundaries have large
uncertainty since there is few quantitative information available from the real eld.
Based on well-log data, it was interpreted that the average areal proportions of shale
drapes along valley/belt/channel are 0.55/0.65/0.65. A shale drape parameter sensi-
tivity study (Chapter 3) has shown that the most important parameter is the shale
drape (or scour hole) proportion. The scour hole geometry and shape play a lesser
role for reservoir connectivity. Therefore, this study will use three dierent scour
hole training images for valley, belt and channel (Figure 4.2). Figure 4.3 shows the
scour hole realizations for valley, belt and channel using these training images. The
production data from real reservoir is not available. Instead, the PVT table, the
oil-water relative permeability curves (Figure 4.3) and uid and rock properties were
provided. These information will be used to build true production proles for our
study. The next section explains the process of obtaining true production data.
4.2 Setting up true cases
The production data from the actual reservoir were not made available by the oper-
ating companies. In order to perform history matching study, synthetic production
data are generated using the PVT information from the real reservoir. The geologic
model characterizing channel distribution is rst simulated using the proposed geo-
logic modeling approach (Figure 4.5a). Next, the shale drapes are generated along
Figure 4.2: Scour hole training images for valley, belt and channel hole simulation.
The red color objects are scour holes, and blue background is shale drapes.
Figure 4.3: Scour hole realizations for valley, belt and channel using the training
images shown in Figure 4.3.
Figure 4.4: Oil and water relative permeability curves.
valley, belt and channel edges. Then the porosities and permeabilities are simulated
based on facies model (Figure 4.5 b and c). The shale drapes properties are included
into the simulation model as multipliers. If channel boundary, belt boundary and
valley boundary overlap (share the same grid) and three overlapped boundaries have
holes, the transmissibility multiplier for this grid will be 1, otherwise 0. Same for the
situation that two of these three boundaries overlap. But if these three boundaries do
not overlap, then when holes are present on channel boundaries, belt boundaries or
valley boundaries, the multiplier will be 1. Figure 4.5d is an example of the transmis-
sibility distribution in x direction accounting for the shale drapes presence. Figure
4.5d shows that even in a reservoir with high poro/perm properties, reservoir connec-
tivity may be reduced due to the presence of drapes.
The simulation grid is dipping 4 degree to the west (along x direction) and
6 degree to the south (along y direction) (Figure 4.6). There is one producer and
one injector in Region 1 (purple colored region in Figure 4.6), one producer and one
injector in Region 2 (gray-colored region in Figure 4.6). The model is a two-phase oil-
water system. Table 4.3 summarizes the simulation model description for generating
production data. The same water saturation, well conguration, uid properties and
relative permeability curves will be used for each waterooding simulation. Therefore
the signicant variables are reservoir architecture and shale drape distribution.
In this work, there are two true cases created, i.e. One-region case and Two-
Figure 4.5: Geologic properties used for ow simulation.
Figure 4.6: Reservoir model showing reservoir structure, the geologic regions, and
well conguration.
Table 4.3: Description of simulation model for true production simulation
Simulation property and description Value
Simulation model Black oil
Simulation period, years 5
Grid (Cartesian) 50 50 50
Active grid block 28364
Grid block dimensions, m
3
84 144 2
Permeability K
x
= K
y
, K
z
/K
x
= 0.1
Reference depth, m 3000
Initial pressure at 3000m, psi 4500
Water-Oil contact, m 4000
Oil gravity (API) 30
Water injectors 2
Injection-control rate, Stb/day 200000
Injection-BHP upper limit, psi 6000
Oil producers 2
Production- Liquid rate control, Stb/day 175000
Production- BHP lower limit, psi 1800
regions case. These two cases share the same channel facies model, porosity and
permeability models shown in Figure 4.5 a, b and c, but have dierent shale drape
distribution along belt and channel boundaries. Since there are 4 wells drilled (2
producers and 2 injectors), the geologic models are constrained to channel facies data
from these 4 wells.
One-region true case: The two geologic regions contain similar hole proportions for
both channels and belts. More specically, scour hole proportions are 0.45/0.4/0.3 for
valley/belt/channel. In this case we will compare two perturbation schemes: (1) full
3D perturbation: both channel location and hole location are perturbed with known
target hole proportions; (2) psuedo-3D perturbation: channel locations are xed, only
scour hole location is perturbed with known target hole proportions. The question
we will address concern is: In high NTG reservoir, is perturbing shale drapes enough
for history matching?
Two-region true case: Two geologic regions have dierent scour hole proportions
for both belts and channels. Region 1 has hole proportions as 0.4 /0.35 for belts and
channels respectively; while Region 2 has a lesser proportion, 0.2/0.15 for belts and
channels respectively. This case is used to test the importance of including prior (re-
gion) information for history matching. Two perturbation schemes are performed: (1)
perturbation by region: channel location and hole location as well as proportion are
perturbed assuming two regions; (2) perturbation without region: channel location
and hole location as well as proportion are perturbed assuming one region. Both per-
turbation schemes assuming channel hole proportion is uncertain or unknown. This
means the target hole proportion along channels will also be perturbed.
The ow simulations are run for 5 years. Figure 4.7 shows true production
data- water cut and bottom hole pressure proles for both cases. The oil saturation
distributions after 5 years of production and water injection are shown in Figure 4.8.
Compared with the no-shale-drape case, the presence of shale drapes renders some oil
in the reservoir unswept. Since the two-region case is more heterogeneous than the
one-region case, there is more oil left in the reservoir in former case. Next sections
will perform the proposed geologic modeling and history matching workow to match
these synthetic production data using various perturbation schemes.
4.3. PERTURBING SHALE DRAPES IN HIGH NTG RESERVOIR 107
Figure 4.7: Water cut and bottom hole pressure proles for true case 1 and 2.
4.3 Perturbing shale drapes in high NTG reservoir
If one believes that, in high NTG reservoirs, almost every channel connects with
one another, then only shale drape location matters to uid ow, hence, perturbing
shale drape distribution is sucient for history matching. However, the stratigraphic
architecture could play an important role since channel location controls the loca-
tion of shale drapes. In order to check the impact of facies architecture (with which
shale drape are associated) on ow responses, this section conducts two perturbation
schemes to match the true production data for One-region case: pseudo-3D pertur-
bation vs. full 3D perturbation. Shale drape target proportions are assumed known.
As mentioned in the previous section, the same well conguration, uid proper-
ties and relative permeability curves are used for each waterooding simulation. This
Figure 4.8: Oil saturation distribution after 5 year production.
ensures that the signicant variables are reservoir architecture and shale drape dis-
tribution.
Two perturbation schemes start with the same initial geologic model including
channel facies distribution, porosity/permeability distribution and shale drape distri-
bution. To make these two perturbation schemes comparable, each follows the same
random path when performing stochastic optimization (PPM). Ten initial geologic
models are provided, and 10 history matching runs are performed for both perturba-
tion schemes, each has 15 outer iterations.
Figure 4.9 shows the production proles simulated from 10 initial geologic models
(blue-colored lines) and the history matched models (purple-colored lines) for pseudo-
3D perturbation scheme. Figure 4.10 is for the full 3D perturbation scheme. These
two gures show the same production proles for the initial geologic models (blue
lines) since they start from the exact same initial guess. We can observe that some
initial models match the production data pretty well without any history match-
ing process. This is because in this case we assume the shale drape proportion for
4.4. PERTURBING SHALE DRAPES BY REGIONS 109
both belts and channels are certain. In fact, from the sensitivity study performed
in Chapter 3, we retained that the shale drape proportion is the most important
factor for ow. By assuming the shale drape proportions known, the uncertainty of
geologic models is reduced in a great degree. Although the uncertainty is greatly re-
duced, comparing the history matched results (purple lines) with non-history matched
ones (blue lines), the former still provides smaller uncertainty for both perturbation
schemes. Comparing the history matching results of these two perturbation schemes,
full 3D perturbation provides more eective results given the same number of outer
iterations. Figure 4.11 displays the optimization performances for these two pertur-
bation schemes. Even though the mismatch of the results obtained by the pseudo-3D
perturbation scheme drops faster in the rst couple iterations, the nal mismatches
are slightly less optimal than the ones obtained from full 3D perturbation scheme.
This demonstrates full 3D perturbation is more ecient. In the following section, all
the history matching processes are performed using a full 3D perturbation scheme.
4.4 Perturbing shale drapes by regions
As introduced in Section 4.1, this realistic reservoir has two geologic regions (Figure
4.6). These two regions have dierent amount of scour holes (or shale drape) for
both belts and channels. Each region has one producer and one injector. With few
wells available, there is a high degree of uncertainty on shale drape proportions for
valley, belts and channels. This is often the case in actual reservoir applications. To
improve prediction power of the history matched models, the perturbation should
consider uncertainty in hole proportions. In other words, the perturbation should not
only modify the hole location, but also the hole proportions. Furthermore, dierent
regions may require dierent amount of perturbation to eciently converge to dier-
ent hole proportions present in each region.
This section is to conduct the history matching process to match the produc-
tion data of the two-regions true case generated in Section 4.2. Two perturbation
schemes are performed: multi-region perturbation vs. single-region perturbation.
Both schemes automatically modify the hole proportion during history matching
Figure 4.9: Ten Water cut and bottom hole pressure proles for both non-history
matched models and history matched models performed using pseudo 3D perturbation
scheme. Blue line is for non-history matched model, red line is for history matched
model, and black line is for true data.
Figure 4.10: Ten Water cut and bottom hole pressure proles for both non-history
matched models and history matched models performed using true 3D perturbation
scheme. Blue line is for non-history matched model, red line is for history matched
model, and black line is for true production data.
Figure 4.11: Optimization performances corresponding to 10 history matching results
shown in Figure 4.9 and Figure 4.10. Blue color is for pseudo-3D perturbation scheme,
purple color is for full 3D scheme.
process. The purpose of this study is to check the importance of including prior
information on the presence of region for history matching.
In order to perform multi-region perturbation, producers need to be assigned to
dierent regions for the purpose of objective function calculation (Homan, 2005).
In the multiple region history matching approach, the number of objective functions
should be consistent with the number of regions such that the optimization process
perturbs optimally for each regions. Meanwhile, we need to know which factor(s) is
most sensitive to which ow response(s), and which ow response can be used as an
indicator for hole proportion adjustment (increase or decrease). Therefore, an exper-
imental design is performed.
4.4.1 Region sensitivity study
Table 4.4 lists four factors examined. A one-half fraction design is performed and Ta-
ble 4.5 shows the factor combinations for 8 runs. Each run has 5 realizations. Hence
a total of 40 ow simulations are performed. Flow responses in both producer 1 (in
Region 1) and producer 2 (in Region 2) are checked, including water breakthrough
time (WT), water cut (WC) at the end of 5 years production, well bottom hole pres-
sure (BHP) at the end of 5 year production. The sensitivity analysis is conducted in
the same manner as the one described in Section 3.2. Figure 4.12 displays the eects
of four factors on these three ow responses. From this sensitivity analysis result, we
can obtain the following information:
Hole proportion along channels in Region 2 is the most sensitive factor. This
is because the number of channels in Region 1 is less than the one in Region 2
(12 vs.20), same as the number of channel belts. This means Region 2 is more
heterogeneous than Region 1. Therefore, hole proportions in Region 1 have less
impacts on water movement as compared to Region 2. As shown in Figure 4.13,
water injected by Injector 2 ows towards Producer 1 (Region 1) or Producer
2 (Region 2), while almost all the amount of water injected by Injector 1 ows
towards Producer 1 (Region 1).
BHP can be used as evidence to assign the producers to dierent regions. As
shown in Figure 4.12, the channel hole proportion in Region 1 has the strongest
impact on BHP for Producer 1 in Region 1, and channel hole proportion in
Region 2 for Producer 2 in Region 2.
BHP is used as an indicator for channel hole proportion perturbation. This is
because BHP is most sensitive to channel hole proportion. In this study, the
channel hole proportion is modied by the following equation:
PropHole
Update
= PropHole
Current
+ll r
D
/10
Where r
D
is the PPM perturbation parameter, ll is an indicator (1 or -1). The
sign of ll is determined using simulated BHP information. If the simulated BHP
is higher than reference BHP data, this means the simulated reservoir model
has more energy than reference reservoir. The reason of having higher energy
is because the simulated reservoir model has less shale drapes (or higher hole
proportion) such that the injected water front moves more homogeneous (hence
slower) than the one in reference reservoir, and pushes more oil towards the pro-
ducer, resulting in higher bottom hole pressure for that producer. Hence the
channel hole proportion for the region that this producer located is decreased.
If the simulated BHP is lower than reference BHP data, this means the injected
water front moves faster towards the producer due to higher shale drape pro-
portion (lower hole proportion), hence we need to increase hole proportion in
the region that this producer located. In this study we choose threshold -10
psi and 10 psi as match criteria. If the dierence between reference BHP and
simulated BHP is smaller than -10 psi, ll =-1; if the dierence is larger than 10
psi, ll=1; otherwise ll=0.
BHP and WC are used to calculate objective function since both are sensitive
ow responses. The objective function is calculated as:
Table 4.4: Four factors and level set up
Factor
Factor level
-1 1
Hole proportion of belt in Region1 (HPBR1) 0.2 0.4
Hole proportion of channel in Region1 (HPCR1) 0.15 0.35
Hole proprtion of belt in Region2 (HPBR2) 0.2 0.4
Hole proportion of channel in Region2 (HPCR2) 0.15 0.35
Table 4.5: The two level 2
41
design
Run HPBR1 HPCR1 HPBR2 HPCR2
1 - - - -
2 + - - +
3 - + - +
4 + + - -
5 - - + +
6 + - + -
7 - + + -
8 + + + +
O(r
D
) =
_
T
1
_
(WC
S
t
WC
H
t
)
WC
H
end
_
2
+
_
T
1
_
(BHP
S
t
BHP
H
t
)
BHP
H
end
_
2
4.4.2 History matching results
After the sensitivity study, history matching processes are performed. Figure 4.14
shows the history matched production proles using multi-region perturbation. Fig-
ure 4.15 is the optimization performance for this perturbation scheme. With the
Figure 4.12: The eect chart of examined factors listed in Table 4.4 on ow responses.
Red lines are decision limits. If the eect bar exceed red, this means that factor is
signicant to the corresponding ow response with 95% condence level.
Figure 4.13: Streamlines showing the water ow path.
optimization process proceed, the objective function converges to the tolerance (1.0)
(Figure 4.15a) and the hole proportions converge from the initial guesses (0.5 for Re-
gion 1 and 0.3 for Region 2) to the references (0.35 for Region 1 and 0.15 for Region
2) respectively (Figure 4.15b). If we assume the whole reservoir as one region, and
apply single-region perturbation, the historical data can also be matched. Figure 4.16
shows the history matched production prole using one region perturbation scheme.
Figure 4.17 shows the corresponding optimization performance for this perturbation.
The channel hole proportion in the reservoir converges to an average value (0.25)
of the channel proportion in true reservoir, while the objective function decreases
towards the tolerance. Hence both perturbation schemes can provide us satisfactory
history matching. Figure 4.18 are reference and history-matched channel distribution
models. Both perturbation schemes provide very similar history-matched geologic
models in terms of the number of channels and the channel stacking patterns. If we
compare the water and oil saturation at the end of 5 years production from both
history matched models with the one from true model (Figure 4.19), the history
matched model obtained using multi-region perturbation scheme (Figure 4.19 b) has
similar water and oil distribution as the true model (Figure 4.19a), while the uid
distribution within history matched model using single-region perturbation (Figure
4.19c) is quite dierent than the true model, though there are many oil left in the
history matched reservoir model. The results demonstrate that including the prior
information of variability of holes by regions improves the prediction power of the
history matched models.
Figure 4.14: History matching results assuming two regions in the reservoir.
4.5 Chapter Summary
In this chapter, we applied proposed geologic modeling method to built channel facies
models based on the available static information of a real oshore West Africa reser-
voir. In order to obtain true production data, two true shale distribution cases,
including One-region case and Two-region case, are generated and the corresponding
production data are simulated. Then the proposed modeling and history matching
method was applied to simulate and perturb the geologic models until their ow re-
sponses match true production data. Dierent perturbation schemes have been
conducted to address various issues. The learning from this case study is summarized
as follows:
Figure 4.15: Optimization performance corresponding to the history matched results
shown in Figure 4.14.
Figure 4.16: History matching results assuming the reservoir as one region.
Figure 4.17: Optimization performance corresponding to the history matched results
shown in Figure 4.16.
Figure 4.18: Reference and history-matched geologic models.
Figure 4.19: Oil and water saturation distribution after 5 years production.
In very high NTG (07-08) reservoirs, the presence of shale drapes has signicant
impact on reservoir connectivity;
For very high NTG (07-08) reservoirs, both the scheme of perturbing chan-
nel and hole location and the scheme of perturbing hole location along xed
channels can achieve satisfactory history matching. However, the former per-
turbation scheme is slightly more ecient since the perturbation is in true 3D
and consistent to geological conceptual model;
In the case that the shale drape proportion is dierent in dierent region of
reservoir, it is suggested to apply multi-region perturbation to obtain more re-
alistic geologic models. Even single-region perturbation can obtain satisfactory
history matching, the history matched models may not have reliable prediction
power due to ignorance of the important prior information.
Chapter 5
Conclusions and future work
5.1 Conclusions
A methodology for modeling and history matching of multi-scale ow barriers in
channelized reservoirs has been presented. With this methodology, reservoir models
containing multi-scale facies architecture and associated ow barriers are constructed
that match production data and consistent to geologic data, such as well logs and
conceptual channel stacking patterns.
Within the geologic modeling process, individual channels are simulated using an
object-based approach. Unlike traditional object-based method in which a 3D object
geometry is directly created based on given geometry parameters, the object-based
modeling method adopted in this work rst generates a 2D pixel-based thickness map
that reects the channel geometry parameters, then paints the thickness in 3D form-
ing an individual channel. Compared to traditional pixel-based approach, the adopted
modeling method can provide explicitly the channel geometry needed for attaching
shale drapes. Compared to traditional object-based approach, the introduction of a
pixel-based property into object simulation makes the hard data conditioning more
exible. Furthermore, this idea makes the perturbation of individual channel geome-
try possible. By perturbing the probability model used to generate a channel thickness
map, the channel geometry in 3D is modied. However, the underlying Boolean sto-
chastic model is preserved.
122
5.1. CONCLUSIONS 123
One of the most novel aspects of this methodology is the channel stacking pattern
modeling technique, as well as perturbation method that modies channel location
consistently with the conceptual stacking pattern model. A channel stacking pattern
in this work is generated using two pattern parameters: migration ratio (MR) and
overlap ratio (OR). By simulating a set of pairs of realizations of MR and OR from
their corresponding probability distribution functions, one channel stacking pattern
is constructed, and the number of channels is the same as the number of pair of
MR and OR. With this modeling approach, the channel location can be modied by
changing the pattern parameter realizations using the gradual deformation method.
The perturbed channels still follow the same conceptual stacking pattern model.
Using the presented channel stacking pattern modeling and perturbation tech-
nique, discontinuous shale drapes associated with curvilinear channels can be per-
turbed in true 3D space. As shown in Chapter 2, shale drapes are simulated on
multiple 2D surfaces, each surface corresponds to one channel erosional boundary.
Because shale drapes are associated with channel boundaries, perturbing shale drape
distribution by moving channel location provides more freedom to the perturbation,
at the same time the perturbation remains consistent with the geologic concept. The
realistic case study in Chapter 4 has demonstrated that even in high NTG reservoirs
where almost all the channels touch with each other, true 3D perturbation scheme is
still more ecient and robust as compared to a pseudo-3D perturbation (perturbing
shale drape distribution with xed channel location). For medium-low NTG reservoirs
where channel location is the rst-order controlling factor for reservoir connectivity,
perturbing shale drape distribution with movable channel location is crucial for e-
cient history matching.
The success of the presented modeling and history matching approach largely re-
lies on the availability and quality of input information, that is, conceptual channel
stacking pattern model, channel geometry parameters and net-to-gross ratio (NTG)
for individual channel belt. In this dissertation, we assume NTG for each belt is avail-
able from geologic study. Tests on a synthetic case in Chapter 3 has demonstrated
how channel geometry parameters and probability distribution functions of channel
stacking pattern parameters can be obtained from reservoir analog. This information
124 CHAPTER 5. CONCLUSIONS AND FUTURE WORK
is used as input for a realistic reservoir study in Chapter 4.
To match production data for reservoirs with multi-scale heterogeneity, a sensitiv-
ity study is often required to determine which scales most signicantly impact ow,
and what combination of ow responses should be used for the objective function.
In the case that the reservoir has dierent geologic regions, a sensitivity study is a
useful tool to assign producers to dierent regions for objective function calculation
purposes. The synthetic belt-channel system reservoir study in Chapter 3 has showed
the hole proportion along individual channels and channel location are the most sen-
sitive parameters for ow responses. The realistic valley-belt-channel reservoir study
in Chapter 4 used a similar sensitivity study to assign producers to dierent geologic
regions. The analysis revealed that the hole proportion along channels is the most
signicant factor that impact production data, hence reservoir connectivity.
The proposed modeling and history matching workow in this dissertation is ini-
tially developed for simulating shale drapes in deep-water conned channel reservoirs.
These thin shale drapes are discontinously distributed along hierarchically erosional
surfaces. However, not all turbidite channel reservoirs have shale draped along the
channel boundaries. Only if the available data from eld, such as cores, well-logs,
production logging test (PLT), pressure information and 4D seismic data infer the
presence of shale drapes or other type of barriers that are associated with reservoir ar-
chitecture, the presented workow can be applied to generate ow barrier distribution
models with reduced uncertainty. From this point of view, the proposed modeling
and history matching workow is general and can be applied to many other deposi-
tional settings such as uvial channel system and deltaic distributary channel system,
or dierent type of barriers such as fracture systems and various cements, that can
be associated stratigraphically with reservoir architectures. In the case that we are
uncertain about the shale draped channel models, scenario-based geological modleing
should be performed, that is, shale draped channel scenario and non-shale draped
channel scenario. For the shale draped channel scenario, the proposed workow can
be appropriately applied to simulate shale distribution models.
5.2. RECOMMENDATIONS FOR FUTURE WORK 125
5.2 Recommendations for future work
Data integration is a fundamental concept in reservoir modeling. The objective is
to explicitly account for all of the available data including geologic data, geophysical
data and production data. The nal goal is to provide multiple realistic geologic
models with prediction power. A large part of the ongoing research in geostatistic
modeling is to devise techniques that can accommodate a great variety of data. This
dissertation has been devoted to developing and testing proposed modeling technique
to integrate well data, conceptual stacking patterns from reservoir analog and pro-
duction data. Suggestions for future work to improve and broaden this modeling
workow are discussed below.
Further data conditioning
Other data can be integrated into reservoir models, including vertical sand pro-
portion curve, areal sand proportion map, well-testing interpreted shale drapes
or channel width, trend map of shale drapes distribution, and seismic-derived
probability cube. In practice, some or all of these information are often avail-
able through extensive studies on well, seismic and production data. Developing
conditioning techniques to integrate these data will make the proposed model-
ing workow more practical.
A vertical proportion curve species the proportion of all channel sand as a
function of vertical elevation z or stratigraphic time, and an areal proportion
map species the sand proportion as a function of areal location (x,y). The
basic idea of constraining these two types of proportion data is as follows: in
the channel complex simulation process, each simulated channel is rasterized
on a grid, and the current vertical proportion curve and areal proportion map
are updated. Because the channels passing through wells are simulated rst, at
some point there will be no channels needed in the current depositional layer
for well data. Then extra channels outside of well data can be added into the
belt until the vertical proportion at the current layer is met. Next, the current
areal proportion map is updated and the simulation goes to next layer. This
process is repeated until the vertical proportion curve and areal proportion map
are honored. This idea however requires some modication of the channel com-
plex modeling method described in this dissertation, for example, certain rules
should be added to ensure the simulated channels follow conceptual stacking
pattern, at the same time the given proportion curve and areal map are hon-
ored.
The interpreted channel thickness, width or shale drape geometry from well-
testing data can be constrained directly using the presented modeling technique.
The thickness and width can be constrained in the same way as well data con-
ditioning described in Chapter 2. Channel thickness is rst converted to noise
hard data, and then this noise hard data is conditioned during correlated noise
simulation. Channel width can be constrained in Step 5 of individual channel
simulation (Chapter 2 Section 2.1.1) by using the interpreted channel width to
dene channel regions instead of using a predened width. Shale drapes geom-
etry or trend map can be easily conditioned when performing pixel-based MPS
simulation.
The most challenging part of data conditioning is for seismic-derived sand prob-
ability cube. The rst challenge is how to correlate pixel-based probabilities to
objects, the second challenge is how to honor probability cube and conceptual
channel stacking pattern simultaneously. As described in Chapter 2, essentially
channels are simulated using an object-based method. More specically, a 2D
thickness map is rst simulated, and then painted into 3D forming channel body.
The location of the channel is determined by stacking pattern parameter prob-
ability distribution functions. One possible avenue of integrating probability
cube is as follows:
In an explicitly dened channel belt container, start from the top surface
of this belt region, check the probability cube around this location and nd
a point (x,z) that has maximum probability, then treat this point as the
rst channel initial centerline location and generate one channel. Channel
geometry is simulated in a way that is consistent with the probability map
underlying the surface z. This could be done by (1) sum up the prob-
ability below surface z down to surface z H (H is channel maximum
thickness) and normalize it, then we get a 2D probability map. (2) inte-
grate this 2D probability map when simulating correlated noise map for
channel thickness.
To determine the next channel location, rst the probability map of mi-
gration ratio Prob
MR
and probability map of overlap ratio Prob
OR
are
calculated around point (x,z) using their representative probability distri-
butions, then a cross section of seismic-derived probability cube Prob
seis
passing point (x,z) is extracted, next we combine these three 2D probabil-
ity maps and nd the most likely point (x
,z
) (with maximum probability

value) on the combined probability map. This point is the location that
the next channel initial centerline passes through. Once we have the ini-
tial centerline, we can simulate another channel. The next channel location
will be based on point (x
,z
).
Testing shale eects in other ow regimes
In this dissertation, shale eects are considered in 3D, that is, if shale drape
is present, the transmissibility multiplier in x, y and z directions could
be 0 depending on the specic channel conguration. In order to test shale
drape impacts on reservoir connectivity, only immiscible ow with favorable
mobility ratio (M 1) is considered in ow simulation model. There is no
gravity, no capillary pressure included. As we know, the eect of a given geologic
heterogeneity on reservoir performance during immiscible ow depends not only
on the nature of that heterogeneity, but also upon the uid properties and the
ow regime. For example, the geometry and evolution of the waterfront during
water ooding depends on the relative mobility of the oil and water and the
relative importance of viscous and gravity force. Future work on shale drape
eects can extend to (1) unfavorable case (M > 1); (2) gravity dominated case;
(3) including capillary pressure.
Gridding and upscaling
Using surface-based gridding, each sedimentary body is dened between two
stratigraphically consecutive sharp boundaries. Hence, each layer represents
one object. The advantages of using a surface-based grid to characterize multi-
scale ow barriers are: (1) thin shale drape can be explicitly modeled as one
layer in surface-based reservoir model, instead of implicitly treated as surface
property in structured coarse Cartesian grid; (2) because the vertical grid cell
size varies with object (channel or drape) thickness, surface-based gridding has
potential to preserve a smooth channel surface shape with aordable number of
cells for ow simulation; (3) surface-based gridding can be directly linked to the
concept of stratigraphic sequence. For example, hierarchic sequence boundaries
interpreted from seismic data and well data can be used to construct multi-
scale surfaces in reservoir model. This creates reservoir model grids reecting
geologic architecture. However, as described in Chapter 1 Section 1.1, there are
some limitations in terms of history matching. One major concern is that dur-
ing history matching, this conformable grid need to be automatically updated
with the reservoir model perturbed. But such gridding is not yet robust enough
to be made automatic. Hence, future research work on robust surface-based
gridding is a promising direction in building geologically realistic reservoir with
aordable computation cost for ow simulation.
Instead of using a uniform Cartesian grid with uniform upscaling ratio, another
alternative is to use upscaling adaptive gridding approach in a Cartesian grid.
Within this approach, no upscaling is performed in the areas of ne-scale reser-
voir model with high degree of permeability variation, such as channel bound-
aries attached by drapes. If permeability variation is small, such as within a
single channel region, upscaling is performed in this area. This adaptive upscal-
ing approach can provide a coarse-scale model that contains fewer grid cells than
would be required when using uniform upscaling in Cartesian grid to achieve
the same accuracy during simulation.
Extend to other systems
One immediate extension is to add multi-scale fractures into reservoirs with
multi-scale shale drapes. With two types of ow barriers, each has multi-scale
nature, the geologic modeling, gridding and upscaling is getting much more com-
plicated and challenging. However, studying the eects of these ow barriers
will provide important guidance for similar reservoir prediction and manage-
ment.
Another interesting and challenging deposition system to be studied is deep-
water distributary lobes. Some authors (e.g., Sprague et al., 2002) argue that
lobes show a similar stratigraphic hierarchy as channel deposits. If scouring
(erosion) is signicant in the upper lobe environment, ne-grained deposits
could sometimes be removed, such that good quality sand bodies within lobes
will be connected. Taken in this sense, stacking pattern of lobes will control
reservoir connectivity. Therefore, knowing individual lobe geometry will help
locating ow barriers and good-quality sand bodies. For this reason, the pre-
sented channel stacking pattern modeling technique can be extended to simulate
lobe distribution. However, unlike the conned channel system studied in this
dissertation, distributary lobes exhibit non-stationarity, that is, lobe complex
is diverging from proximal to distal, and individual lobe also shows diverging
geometry. This non-stationary nature makes the geologic modeling more chal-
lenging compared to channel system. For example, how to dene a transfer
function to convert simulated potential map into lobe thickness while condition
to well data? How to characterize erosion vs. deposition within individual lobes?
How to integrate progradation and retrogradation trend into geologic models?
All of these are very interesting and practical questions to be addressed..
Bibliography
[1] Abreu V., Sullivan M., Pirmez C. and Mohrig D.: Lateral accretion packages
(LAPs): an important reservoir element in deep water sinuous channels, Marine
and Petroelum Geology 20, p. 631-648, 2003.
[2] Alapetite, J., Leon B., Grinarten E. and Mallet J-L : Stochastic modeling of u-
vial reservoirs: the YACS approach, paper SPE 97271 presented at SPE Reservoir
Simulation Symposium, Dallas, Texas, October 9-12, 2005.
[3] Barrentine, L.B.: An introduction to design of experiments: a simplied ap-
proach, American Society for Qualit (1999), 114.
[4] Begg, S. H., and P. R. King: Modeling the eects of shales on reservoir perfor-
mance: Calculation of eective permeability, paper SPE 13529 presented at SPE
Reservoir Simulation Symposium, Dallas, Texas, February 10-13, 1985.
[5] Box, G.E.P. and Draper, N.R.: Empirical model building and response surfaces,
Wiley, New York City (1987), 699.
[6] Bridge, J. and M. Leeder: A simulation model of alluvial stratigraphy. Sedimen-
tology, 26(5), p. 617-644, 1979.
[7] Bu, T. and Demsleth, E.: Errors and uncertainties in reservoir performance
prediction, SPEFE (Spt. 1996) 194.
[8] Caers, J., Strebelle S., Payrazyan K.: Stochastic integration of seismic data and
geologic scenarios: a West Africa submarine channel saga, Leading Edge 22(3), p.
192-196, 2003.
130
BIBLIOGRAPHY 131
[9] Caers, J.: Geostatistical history matching under training-image based geological
model constraints, Society of Petroleum Engineers Journal, 8(3), p. 218-226, 2003.
[10] Caers, J. and Homan, T.: The probability perturbation method: a new look
at Bayesian inverse modeling, Mathmatic Geology, v. 38, no. 1, p. 81-100, 2006
[11] Clark, J.D. and K.T. Pickering: Submarine channels: processes and architec-
tures, Vallis Press, London, 231p., 1996a.
[12] Clark, J.D. and K.T. Pickering: Architectural elements and growth patterns of
submarine channels: applications to hydrocarbon exploration. American Associ-
ation of Petroleum Geologists V.80(2), p.194-221, 1996b.
[13] Chu, C.F.: Prediction of teamood performance in heavy oil reservoirs using
correlations developed by fractorial design method, paper SPE 2002 presented at
California regional meeting, Ventura, California, April 4-6, 1999.
[14] Damsleth,E., Hage, A,Volden, R.: Maximun information at minimum cost: A
North Sea Field development study with experiment design, Journal of Petroelum
Technology, Dec. 1992, 1350.
[15] Dejean, J.P. and Blanc, G.: Managing uncertainties on production predictions
using integrated statistical methods, paper SPE 56696 presented at the SPE An-
nual Technical Conference and Exhibition, Houston, TX, Oct. 3-6, 1999.
[16] Deutsch, C.V., Y., Xie, and A.S., Cullick: Surface geometry and trend modeling
for integration of stratigraphic data in reservoir models, paper SPE 68817 at SPE
Western region meeting, Bakeseld, Califonia, March 26-30, 2001.
[17] Edie, A.L. et al.: Automatic history matching by use of response surfaces and
experimental design, paper presented at the European Conference on the Mathe-
matics of Oil Recovery, Roros, Norway, June 7-10, 1994.
[18] Egeland, T. et al.: Designing better decisions, paper SPE 24275 presented at
the SPE European Petroleum Computer Conference, Stavanger, May 25-27, 1992.
132 BIBLIOGRAPHY
[19] Eschard, R., Elbouy, E., Deschamps, R., Euzen, T. and Ayub A.: Down-
stream evolution of turbiditic channel complexes in the Pab Range outcrops (Maas-
trichtian, Pakistan), Marine and Petroleum Geology 20, p. 691-710, 2003.
[20] Friedmann, F., Chawathe, A., and Larue, D.K.: Assessing uncertainty in chan-
nelized reservoirs using experimental design, SPEREE, p.264, 2003.
[21] Friedmann, F., Chawathe, A., and Larue, D.K.: Uncertainty assessment of
reservoir performance using experimental designs, Paper 2001-170 presented at the
Petroleum Societys 2001 Canadian International Petroleum Conference, Calgary,
12-14 June, 2001.
[22] Gardner, M.H. and Borer, J.M.: Submarine channel architecture along a slope
to basin prole, Brush Canyon Formation, west Texas, In: Bouma, A.H., Stone,
C.G. (Eds), Fine-grained turidite systems: AAPG Memoir 72 and SEPM Special
Publication 68, p195-214, 2000.
[23] Gardner, M.H., Borer, J.M., Melik, J.J., Mavilla, N., Dechesne, M. and Wagerle,
R.D.: Stratigraphic process-response model for submarine channels and related
features from studies of Permian Brushy Canyon outcrops, West Texas, Marine
and Petroleum Geology 20, p. 757-788, 2003.
[24] Giordano, R.M., Salter, S.J. and Mohanty, K.K: The eects of permeability
variations on ow in porous media, paper SPE 14365 presented at the SPE Annual
Technical Conference and Exhibition, Las Vegas, Sept. 22-25, 1985.
[25] Haldorson, H.H. and L.W. Lake: A new approach to shale management in
eld-scale model, Society of Petroleum Engineers Journal, p447-457, 1984.
[26] Haldorsen, H.H., and D.W. Chang: Notes on stochastic shales: from outcrop
to simulation model, in L. Lake and H.B. Caroll, eds., Reservoir characterization:
London, Academic Press, p. 445-485, 1986.
[27] Hewett, T.A.: Fractal distributions of reservoir heterogeneity and their inunce
on uid transport, paper SPE15386 presented at the SPE Annual Technical Con-
ference and Exhibition, New Orleans, Oct. 5-8, 1986.
BIBLIOGRAPHY 133
[28] Homan, T.: Geologically consistent history matching while perturbing facies,
Ph.D dissertation, Stanford University, Stanford, 2005.
[29] Hu, L.Y.: Gradual deformation and iterative calibration of Gaussian-related
stochastic models, Journal of Mathematical Geology, 32(1), p. 87-108, 2000.
[30] Hu, Lin Y. and Sandra Jenni: History matching of object-based stochastic
reservoir models, Society of Petroleum Engineers Journal, September, p. 312-323,
2003.
[31] Hubbard, S.M., Romans, B.W., and Graham, S.A.: Deepwater channel margin
architecture, Cerro Toro Formation, Cerro Mocho, Chile, in Nilsen, T., Shew, R.,
Steens, G, and Studlick, J., eds., Deep-Water Outcrops of the World Atlas, Amer-
ican Association of Petroleum Geologists Special Publication, p. 128-131, 2008a.
[32] Hubbard, S. M., Romans, B.W., and Graham, S.A.: Deep-water foreland basin
deposits of the Cerro Toro Formation, Magallanes basin, Chile: Architectural ele-
ments of a sinuous basin axial channel belt, Sedimentology, in press, 2008b.
[33] Jackson, M.D., and A.H. Muggeridge: The eects of discontinuous shales on
reservoir performance during horizontal waterooding, Society of Petroleum En-
gineers Journal, v. 5, no. 4, p. 446-455, 2000.
[34] Jones, A., Doyle, J., Jacobsen, T. and Kjonsvik D.: Which sub-seismic het-
erogeneities inuence waterood performance? a case study of a low net-to-gross
uvial reservoir, in: de Haan, H.J. (ed.) New Developments in Improved Oil Re-
covery. Geological Society, London, Special Publication, N.84, 5-18, 1995.
[35] Larue D.K. and Friedmann F.: The controversy concerning stratigraphic archi-
tecture of chanelized reservoirs and recovery by waterooding, Petroleum Geo-
science, vol. 11, p. 131-146, 2005.
[36] Le Ravalec-Dupin M. and Hu L.Y.: Gradual deformation of Boolean simu-
lations, in: Leuangthong O. and C.V. Deutsh eds., Quantitative Geology and
Geostatistics, Geostatistics Ban 2004, p. 939-948, 2004.
134 BIBLIOGRAPHY
[37] Li, H. and C. White: Geostatistical models for shales in distributary channel
point bars (Ferron Sandstone, Utah): From ground-penetrating radar data to three-
dimensional ow modeling, AAPG Bulletin, 87(12): 1851-1868, 2003.
[38] Li, H. , Caers, J.: Geologically consistent perturbation of channel location
for history matching, 20th SCRF annual meeting, Stanford Center for Reservoir
Forecasting, Stanford, May, 2007.
[39] Lopez, S.: Modelisation de reservoir chenalises meandriforme: Approche gene-
tique et stochastique, Ph.D Thesis, Ecole des Mines de Paris, 2003.
[40] Lowe, D.R.: Deep-water sandstones: Submarine canyon to basin plain, western
California, American Association of Petroleum Geology Special Publication, 2004.
[41] MacDonald, A.C. and Halland, E.K.: Sedimentology and ahle modeling of
a sandstone-rich uvial reservoir: Upper Stratfjord Formation, Stratfjord eld,
Northern North Sea, AAPG Bulletin, V. 77, 1016-1040.
[42] Martin, J.H. and Cooper, J.A.: An intehrated approach to the modeling of
permeability barrier distribution in a sedimentologically complex reservoir, paper
SPE 13051 presented at SPE Annual Technical Conference, Houston, Sept.16-19,
1984.
[43] Mayall, M. and OByrne, C.: In:Reservoir prediction and development chal-
lenges in turbidite slope channels, OTC conference proceedings, Contribution No.
14209, 2002.
[44] Mayall M., Jones E. and Casey M.: Turbidite channel reservoirs-key elements
in facies prediction and eective development, Marine and Petroleum Geology 23,
p821-841, 2006.
[45] Myers, R.H. and Montgomery, D.C.: Response surface methodology: process
and product optimization using designed experiments, Wiley, New York City
(1995), 700.
BIBLIOGRAPHY 135
[46] Montgomery, D.C.: Design ang analysis of experiments, John Wiley and Sons,
Inc., New York City (2001).
[47] Mutti, E. and Normark, W.R.: Comparing examples of modern and ancient
turbidite systems: problems and concepts. In: Leggett, F.K. and Zua, G.G.(eds)
Marine clastic sedimentology. Graham & Trotman, London, p.1-38, 1987.
[48] Mutti, E. and Normark, W.R.: An integrated approach to the study of tur-
bidite systems. In: Weimer, P. & Link, M.H.(eds) Seismic facies and sedimentary
processes of submarine fans and turbidites systems. Springer Verlag, New York,
p.75-106, 1991.
[49] Narayanan, K. et al.: Response surface methods for upscaling heterogeneous
geologic models, paper SPE 51923 presented at the Reservoir Simulation Sympo-
sium, Houston, TX, February 14-17, 1999.
[50] Pickering, K., Clark, J., Smith, R., Hiscott, R., Ricci Lucchi, F. and Kenyon, N.:
Architectureal element anaysis of tubidite systems, and selected topic problems for
sand-prone deepwater systems. In: Pickering, K., Hiscott, R., Kenyon, N., Ricci
Lucchi, F. and Smith, R. (eds) Atlas of deepwater environments: architectural
styles in turbidite systems. Chapman & Hall, London, p.1-11, 1995.
[51] Prather, B.F.: Controls on reservoir distribution, architecture and stratigraphic
traping in slope setting, Marine and Petroleum Geology 20, p. 529-545, 2003.
[52] Pranter M., Ellison, A.I,Cole, R.D. and Patterson P.E.: Analysis and modeling
of intermediate-scale reservoir heterogeneity based on a uvial point-bar outcrop
analog, Williams Fork Formation, Piceance Basin, Colorado, AAPG Bulletin,
91(7), p. 1025-1051, 2007.
[53] Pyrcz, M.J., O. Catuneanu and C.V. Deutsch: Stochastic surface-based mod-
eling of turbidite lobes, AAPG Bulletin, 89(2), p177-191, 2005.
[54] Richardson,J.G., Harris, D.G., Rossen, R.H. and Van Hee, G: The eects of
small, discontinuous shales on oil recovery, Journal of Petroleum Technology, Nov.
1978, 1531-1537.
136 BIBLIOGRAPHY
[55] Samantray,A., Kraaijveld, M., Hognestad, J. and Bulushi W.: Reservoir char-
acterization and modeling of stacked uvial/shallow marine reservoirs: what is
important for uid-ow performance and eective reservoir prediction?, paper
SPE 11034 presented at the SPE Annual Technical Conference and Exhibition,
Anaheim, California, Nov. 11-14, 2007.
[56] Shanmugam, G.: 50 years of the turbidite paradigm (1950s-1990s): deep-water
processes and facies models-a critical perspective, Marine and Petroleum Geology,
V.17, p.285-342.
[57] Shmaryan, L.E. and C.V. Deutsch: Object-based modeling of uvial/deepwater
reservoirs with fast data conditioning: methodology and case studies, paper SPE
56821 at the SPE Annual Technical Conference and Exhibition, Houston, Texas,
Oct. 3-6, 1999.
[58] Slatt R.M. and Weimer P.: Turbidite systems Part 2: Subseismic-scale reservoir
characteristics, The Leading Edge, May, p.562-567, 1999.
[59] Stow, D.A.V. and Johansson, M.: Deep-water massive sands: nature, origin
and hydrocarbon implications, Marine Petroleum Geology, V.17, p.145-174, 2000.
[60] Stoyan, D., W.S. Kendall and J. Mecke: Stochastic geometry and its applica-
tions, New York, John Wiley & Sons, 345p., 1987.
[61] Strebelle, S.: Conditioning simulation of complex geological structures using
multiple point statistics, Mathematic Geology, 34(1): p. 1-22, 2002.
[62] Strebelle, S. and Payrazyan K.: Modeling of a Deepwater turbidite reservoir
conditional to seismic data using principal component analysis and multiple-point
geostatistics, Society of Petroleum Engineers Journal, V.8, No.3, p. 227-235, 2003.
[63] Stright, L.: Re-adressing the missing scale using edges, 18th SCRF annual
meeting report, Stanford Center for Reservoir Forecasting, Stanford, May, 2005.
BIBLIOGRAPHY 137
[64] Stright, L. , Caers, J. and Li, H.: Coupled geological modeling and history
matching of ne-scale curvilinear ow barriers, 19th SCRF annual meeting, Stan-
ford Center for Reservoir Forecasting, Stanford, May, 2006.
[65] Tillman, R.W. and Pittman, E.D.: Barriers to ow in tidal sandstone reservoirs,
Sun Ranch, Wyoming, paper SPE 27966 presented at the University of Tulsa
Centennial Petroleum Engineering Symposium held in Tulsa, OK, Aug. 29-31, 1994.
[66] Viseur, S.: Stochastic Boolean simulation of uvial deposits: A new approach
combining accuracy with eciency, paper SPE 56688 presented at the SPE Annual
Technical Conference and Exhibition, Houston, Texas, 3-6 October 1999.
[67] Weber, K.J.: How heterogeneity aects oil recovery, Reservoir Characteri-
zation, L. Lake and H. Carroll (eds.), Academic Press, Washington, DC (1986),
487-544.
[68] Weimer, P., Slatt, R.M., Dromgoole, P., Bowman, M. and Leonard A.: Devel-
oping and managing turbidite reservoirs: case histories and experiences: results of
the 1998 EAGE/AAPG research conference, AAPG Bulletin, V. 84, No. 4(April
2000), 453-465.
[69] White, C.D. and M.D. Barton: Translating outcrop data to ow models, with
applications to the Ferron Sandstone, paper SPE 38741 presented at SPE Annual
Technical Conference, San Antonio, Oct.5-8, 1997.
[70] White, C.D., B.J. Willis, K. Narayanan and S.P. Dutton: Identifying and es-
timating signicant geologic parameters with experimental design, Society of Pe-
troleum Engineers Journal, p. 311-324, 2001.
[71] White, C.D., S.A. Royer: Experimental design as a framework for reservoir
study, paper SPE 79676 presented at the SPE Reservoir Simulation Symposium,
Houston, Texas, U.S.A., February 3-5, 2003.
[72] Willis, B.J. and White, C.D.: Use of quantitative outcrop description for reser-
voir modeling, Journal of Sedimentary Research, no. 70, v. 4, p. 788-800, 2000.
138 BIBLIOGRAPHY
[73] Willis, B.J., and C.D. White: Quantitative outcrop data for ow simulation:
Journal of Sedimentary Research, v. 70, p. 788-802, 2000.

Li08 PDF

Uploaded by

Copyright:

Available Formats

You might also like

Li08 PDF

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Li08 PDF

Uploaded by

Copyright:

Available Formats

HIERARCHIC MODELING AND HISTORY MATCHING OF

MULTI-SCALE FLOW BARRIERS IN CHANNELIZED

between two adjacent channels is

) (with maximum probability

You might also like