GENERATION OF MULTIPLE

HISTORY MATCH MODELS USING

MULTISTART OPTIMIZATION

A REPORT SUBMITTED TO THE DEPARTMENT OF ENERGY

RESOURCES ENGINEERING
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE
DEGREE OF MASTER OF SCIENCE

By
Manish K Choudhary
June 2012

I certify that I have read this report and that in my opinion it is fully adequate, in scope
and in quality, as partial fulfillment of the degree of Master of Science in Petroleum
Engineering.

__________________________________
Dr. Tapan Mukerji

iii

Abstract
Uncertainty in the geological model presents a key challenge in development decisions.
Production data from the field are acquired only at limited locations and are sparse.
Time-lapse seismic data is available field-wide but has limited resolution. In addition,
increasingly production logging data is being recorded in wells, which provide
information regarding vertical heterogeneity between wells. The available data set is still
sparse for accurately modeling spatial distribution of reservoir properties. Hence,
multiple geological realizations can exists which match the given production history and
generate varying forecast, all of which should be analyzed for decision-making.
In my research thesis, two optimization algorithms have been tested for generating
multiple history- matched geological models. The reservoir inversion problem has been
formulated using optimization technique, with an objective of minimizing the variance
between observations and output of numerical models using one, two and all three
datasets as described above. Optimization is carried out in reduced model space. Model
reduction is achieved by spatial principal component analysis (PCA), where optimization
search space is projected to a subspace of much smaller dimension.
Local optimizers often tend to find solutions faster than global methods, though they can
be trapped in local minima. Randomly generated multiple initial points can be optimized
in parallel to locate multiple models matching history. Hook-Jeeves direct search (HJDS)
algorithm, simultaneous perturbation stochastic approximation (SPSA) algorithm has
been used for optimization and results are compared with rejection sampler. The minima
points identified through optimization represent geological models that are consistent
with the production history of the field. The methodology has been tested on three
different synthetic case studies with both categorical variable and continuous variables
The optimization process locates geological models that are consistent with production
history but present a varying forecast which can help in decision analysis.

Acknowledgments
I wish to express my sincere gratitude to my advisor Dr. Tapan Mukerji for his guidance,
and support during my Masters research. His deep and insightful suggestions were the
building blocks of this research. His continuous patience, motivation and enthusiasm
helped me at all times in my two years at Stanford. I deeply appreciate his tireless reading
of my thesis.
I am greatly indebted to the faculty, staff, and students of Department of Petroleum
Engineering for their support and contribution to my academic achievements.
I would also like to express my thanks to David Echeveria for his computer code, which
was the starting point of this research. I would also like to thank Mehrdad Shirangi for
explaining SPSA algorithm to me.
I would like to thank the companies supporting the Stanford Center for Reservoir
Forecasting (SCRF) affiliate program at Stanford University for their financial
contribution that made this research possible.
Finally, I owe my greatest gratitude to my family for their love and devotion. Without
their support, I would have never made it this far. I am grateful to my friends for their
support and constant motivation to achieve my objective.

vii

Contents
Abstract ............................................................................................................................... v
Acknowledgments............................................................................................................. vii
Contents ............................................................................................................................. ix
List of Tables ..................................................................................................................... xi
List of Figures .................................................................................................................. xiii
1

Introduction ................................................................................................................. 1
1.1 Methodology ........................................................................................................ 2
1.2 Principal Component Analysis ............................................................................. 3
1.3 K-Medoid Clustering............................................................................................ 3
1.4 Optimization Algorithm ....................................................................................... 3
1.4.1
Hook Jeeves Algorithm................................................................................. 4
1.4.2
SPSA Algorithm ........................................................................................... 4

Testing the Idea - Case Study I & II ........................................................................... 5

2.1 Model Description ................................................................................................ 5
2.2 Dimension Reduction ........................................................................................... 6
2.3 Observables and Objective Function.................................................................... 6
2.4 K - Medoid Clustering.......................................................................................... 7
2.5 Model Set - 1: Continuous Property based Model ............................................... 8
2.5.1
Model Generation ......................................................................................... 8
2.5.2
Dimension reduction and Clustering ............................................................ 9
2.5.3
Optimization Results..................................................................................... 9
2.5.4
Optimization Algorithm Efficiency ............................................................ 11
2.5.5
Forecast Uncertainty ................................................................................... 12
2.6 Model Set - 2: Categorical Property based Model ............................................. 12
2.6.1
Model Generation ....................................................................................... 12
2.6.2
Dimension reduction and Clustering .......................................................... 13
2.6.3
Inclusion of Seismic Tomography Data ..................................................... 15
2.6.4
Optimization Results................................................................................... 16
2.6.5
Optimization Efficiency .............................................................................. 18
2.7 Conclusions ........................................................................................................ 19

Expanding the Idea Case Study III ........................................................................ 21

3.1 Model Description .............................................................................................. 21
3.1.1
Geology ....................................................................................................... 21
3.1.2
Petrophysics ................................................................................................ 22
3.1.3
Relative Permeability & Capillary Pressure ............................................... 23
3.1.4
Fluid Properties ........................................................................................... 23

ix

3.1.5
Rock Physics Model ................................................................................... 24
3.1.6
Production Strategy..................................................................................... 24
3.2 Model Decomposition & Reconstruction ........................................................... 25
3.2.1
Principal Component Analysis ................................................................... 25
3.2.2
Dimension Reduction.................................................................................. 26
3.2.3
Model Generation ....................................................................................... 27
3.3 History ................................................................................................................ 28
3.4 Optimization ....................................................................................................... 30
3.4.1
Workflow .................................................................................................... 30
3.4.2
Observables & Response Function ............................................................. 30
3.5 Results ................................................................................................................ 31
3.5.1
Flow Only ................................................................................................... 31
3.5.2
Flow & PLT Data........................................................................................ 32
3.5.3
Flow, PLT and Seismic Data Optimization ................................................ 34
3.6 Rejection sampling algorithm ............................................................................ 36
3.7 Optimization Efficiency ..................................................................................... 37
3.8 Forecast Uncertainty .......................................................................................... 38
3.9 Conclusions ........................................................................................................ 39
4

Summary ................................................................................................................... 41

Nomenclature .................................................................................................................... 43
References ......................................................................................................................... 45

List of Tables

Table 2-1: Dynamic properties of the model ...................................................................... 6

Table 2-2: Porosity Variogram: Model - 1 ......................................................................... 8
Table 3-1: Fluid Properties ............................................................................................... 24
Table 3-2: Rock Physics Model ........................................................................................ 24
Table 3-3: Weight for Optimization ................................................................................. 31

xi

List of Figures

Figure 1-1: Schematic of Hook-Jeeves pattern search algorithm ....................................... 4

Figure 2-1: Porosity-Permeability Relationship ................................................................. 5
Figure 2-2: Relative Permeability Curve ............................................................................ 5
Figure 2-3: Proxy distance computation for clustering....................................................... 7
Figure 2-4: PDF and CDF of porosity: Model Set - 1 ........................................................ 8
Figure 2-5: K-medoid clustering: Model Set - 1 ................................................................. 9
Figure 2-6: Cumulative oil production vs. Time: Model Set - 1 ........................................ 9
Figure 2-7: Cum. oil production vs. cum. water injection: Model Set - 1 ........................ 10
Figure 2-8: Layer 5 - History matched models: Model Set - 1 ......................................... 10
Figure 2-9: Optimization function efficiency: Model Set - 1 ........................................... 11
Figure 2-10: MDS view of initial & final models: Model Set 1 .................................... 12
Figure 2-11: Forecast uncertainty of history matched models: Model Set 1 ................. 12
Figure 2-12: Principal components (1st, 2nd, 5th & 30th): Model Set 2 ........................... 13
Figure 2-13: Variance of principal component: Model Set - 2 ......................................... 13
Figure 2-14: Model reconstruction: Model Set 2........................................................... 14
Figure 2-15: K-medoid clustering: Model Set 2 ............................................................ 14
Figure 2-16: Tomography Survey: Model Set - 2 ............................................................. 15
Figure 2-17: Layer 5: Flow only history matched models: Model Set 2 .................... 16
Figure 2-18: Layer 5: Flow & seismic history matched models: Model Set 2 ........... 17
Figure 2-19: E- Type of final solution models: Model Set 2 ......................................... 17
Figure 2-20: Variance of final solution models: Model Set 2 ....................................... 18
Figure 2-21: Optimization efficiency : Model Set - 2....................................................... 18
Figure 2-22: Forecast variability: Model Set 2 .............................................................. 19
Figure 3-1: Reservoir structure and Channel Training Image .......................................... 21
Figure 3-2: Porosity distribution in the reservoir .............................................................. 22
Figure 3-3: Histogram of porosity for Sand and Shale ..................................................... 22

xiii

Figure 3-4: Relative permeability curveand Capillary pressure curve ............................. 23

Figure 3-5: Zamba Production Strategy ........................................................................ 25
Figure 3-6: Variance of first 3000 principal components ................................................. 26
Figure 3-7: Cumulative variance of first 3000 principal components .............................. 26
Figure 3-8: Reconstruction of model using limited set of principal components ............. 27
Figure 3-9: Bounds for PCA coefficients ......................................................................... 27
Figure 3-10: Randomly generated models using 3000 major principal components ....... 28
Figure 3-11: Production history of the Zamba reservoir .................................................. 28
Figure 3-12: Fluid saturations over time.......................................................................... 29
Figure 3-13: Spinner survey data for well - 4 and well - 6 ............................................... 29
Figure 3-14: Optimization workflow ................................................................................ 30
Figure 3-15: Water Cut - flow only Optimization ............................................................ 32
Figure 3-16: Layer 5 of history matched models: Flow only optimization ...................... 32
Figure 3-17: Water cut - flow & PLT data optimization .................................................. 33
Figure 3-18: PLT data match - End of 3rd year for Well - 6 (Injector) ............................. 33
Figure 3-19: PLT data match - End of 3rd year for Well - 4 (Producer) ........................... 34
Figure 3-20: History matched models: Flow & PLT data optimization ........................... 34
Figure 3-21: Water cut - flow, PLT and seismic data optimization.................................. 35
Figure 3-22: P-wave velocity match - flow, PLT and seismic data optimization............. 35
Figure 3-23: History matched models: Flow, PLT and seismic data optimization .......... 36
Figure 3-24: Comparison with rejection sampler ............................................................. 36
Figure 3-25: Optimization Code Efficiency for Hook Jeeves algorithm ....................... 37
Figure 3-26 : Optimization Code Efficiency for SPSA algorithm .................................... 38
Figure 3-27: Forecast of history matched models............................................................. 38

xiv

Chapter 1
1 Introduction
Selection of the best development plan for an oil field is one of the significant decisions
undertaken by oil-field operators. The decision is based on analysis of oil & gas forecast
from multiple reservoir models generated using reservoir simulation. Numerical
simulation models require reliable estimates for various reservoir parameters that cannot
be measured directly (Slater & Durrer, 1971; Friedmann, Chawathe, & Larue, 2003). Of
many reservoir parameters, sub-surface heterogeneity i.e. spatial distribution of porosity
and permeability is one of important sources of uncertainty (Stephen & MacBeth, 2008;
Srinivasan & Deutsch, 2004). It is why operators rely on forecasts from multiple
reservoir models to quantify the uncertainty and risk associated with development plan
(Cruz, Horne, & Deutsch, 2004; Rivera, et al., 2007).
Limited data is available to classify the spatial distribution of reservoir properties
accurately, resulting in multiple geological models that match historical observations.
Production data from the field is acquired only at limited locations and is sparse. Timelapse seismic data is available field-wide but has limited resolution. Increasingly spinner
surveys (Kading, 1976)) are being recorded in wells. The data though being sparse,
provides information about connectivity between wells and can be used for history
matching (Yoelin & Howald, 1970; Vogelij, Leach, & Kapteyn, 1993; Panda &
Nottingham, 2011). These independently recorded observations from the field provide
information about spatial distribution of rock parameters at different resolution, all of
which should be integrated for history matching and uncertainty quantification (Stephen,
Soldo, MacBeth, & Christie, 2006; Litvak, Christie, Johnson, Colbert, & Sambridge,
2005). Decisions pertaining to selection of best plausible development option should be
based on performance analysis of all history matched subsurface realizations.
The process of history matching involves adjustment of model parameters to match
model response with the observed dataset. The traditional history matching approach or
cascaded approach, involved independent adjustment of model parameters and is usually
ad-hoc. As such, it can result in models that are quite different from prior models. The
modern history matching approach calls for a closed-loop workflow where facies
distribution is perturbed in a geological consistent manner to obtain a good match
between model response and observed dataset. This ensures that posterior models have
similar features as prior models (Caers, 2011). Different methodologies have been
proposed to minimize the mismatch between model prediction and observed data simulated annealing, pattern search, gradual deformation, Markov chain, iterative resampling, gradient based optimization, etc (Caers, 2011; Himmelblau, 1972).
The mismatch or error minimization problem can also be set up as an optimization
problem. Many authors have proposed formulation of simultaneous matching of
1

production and seismic data as an optimization problem (Huang, Meister, & Workman,
1998; Sarma, Durlofsky, Aziz, & Chen, 2006). The formulation can also be extended to
other data set like spinner survey. A petroleum reservoir can have close to million
unknowns (porosity, permeability, facies etc.) and any optimization in the original space
will be inefficient and would have extremely high computation costs. (Echeveria &
Mukerji, 2009) suggest using principal components to reduce the optimization search
space while maintaining geological consistency. The process of optimization would
locate a model which has minimum variance across observed data points and estimated
values, in the local space (local minima) or across the entire solution space (global
minima). This process would only locate one solution model and hence does not provide
any mean to capture production forecast uncertainty. This research thesis builds on earlier
proposed optimization scheme of (Echeverria & Mukerji, 2009) to locate multiple history
matched models.
1.1

Methodology

The inversion process is formulated as an optimization problem, with the aim to reducing
the error between observations and calculated dataset, similar to the one proposed
(Echeverria & Mukerji, 2009). Let denotes the dimensional space in which the
reservoir models are defined, is the set of admissible geological models which closely
matches observed data and g denotes the indicator property (facies) or reservoir property
(porosity and/or permeability) which is used to define the geological model. The system
of models can be expressed as . The optimization problem can be defined as
in Equation (1-1)
= ()

(1-1)

where m U n comprises all observed data in the inversion, and () U n represents

the numerically computed observables for the model . Varying weights/normalization
for different components in the observables (production, seismic, flow survey etc.) can be
included in the optimizer function (Euclidean norm in this case) to take into account
variable uncertainty in the data acquisition & observations. The optimization problem is
ill conditioned due to much larger number of inversion parameters than observations
( >> ) and model reduction using principal component analysis have been proposed
earlier. Optimization is carried out in subspace of lower dimension (principal
components: < ), statistical information for which is obtained from prior knowledge
of the reservoir. The principal components also can be used to randomly generate
geologically consistent reservoir model from randomly generated coefficients for each
principal component. Local optimizers are strongly dependent on the initial guess but
tend to be more efficient than the global optimizer. Multiple optimization run starting
from different initial guesses can help to obtain multiple minima across the solution
space, while using efficiency of local optimizers.

1.2

Principal Component Analysis

Principal component analysis (PCA) is a mathematical procedure that linearly transforms

an original set of variables into a substantially smaller set of linearly uncorrelated
variables called principal components. The number of principal components is equal to
the dimensionality of the space that can be much smaller than the dimension of original
variable set. The limited set of principal components represents all information of the
original data set (Jolliffe, 2002). The transformation generates principal components in
the decreasing order of variance i.e. the first principal component has the largest possible
variance and accounts for maximum variability among the variable set, the second
principal component has the second largest variance and is orthogonal to the first
principal component. Each succeeding component add less and less variability to the
model and is orthogonal to all other principal components (Gass, 2007).
The principal components are independent of each other only if the variables are
normally distributed. PCA is a linear technique and any extension to non-linear variables
approximates the method. PCA is sensitive to the relative scaling of the original
variables. The PCA technique is also called by other names such as discrete Karhunen
Love transform (KLT), the Hotelling transform and proper orthogonal decomposition.
1.3

K-Medoid Clustering

Reservoir models are large dimensional models and it is often difficult to visually
compare models with each other. Prior models generated using geostatistical technique
look similar visually but can produce widely varying production response. likewise,
models which look different may have similar oil & gas forecast. Since the local
optimizers are very sensitive to initial guess, it is important that the starting models be
spaced as far as possible to prevent convergence to same minima.
K-Medoid clustering algorithm can be used in a multi-dimensional space (Caers, 2011) to
select initial models for optimization. A forward model computation may be too
expensive so a proxy distance mapping can be used for clustering. The K-Medoid
algorithm attempts to cluster the dataset by minimizing the non-metric distance between
points and selects models that are deemed center of the cluster. The selected prior models
are likely to be spaced far apart and will generate different posterior models. (Caers,
2011; Maulik, Bandyopadhyay, & Mukhopadhyay, 2011).
1.4

Optimization Algorithm

Local optimizers tend to be efficient than global optimizers as they quickly converge to a
local minima. A gradient-based optimization can further increase speed, but gradients are
costly to compute. Adjoint based simulator can help in accelerating the search but require
access to source code. In most cases, adjoints are unavailable in commercial simulators or
they are too complicated to work with. Two gradient free algorithms were used for
optimization during the study and are briefly described in the section below.

1.4.1

Hook Jeeves Algorithm

Hook Jeeves Direct Search (HJDS) is a gradient free optimization method, where
approximate numerical slopes are calculated from function calls (Hooke & Jeeves, 1961;
Gottfried & Weisman, 1973). The pattern search method consists of sequence of
exploratory moves about a base point that, if successful, are followed by pattern moves.
The purpose of exploratory moves is to acquire information about the function in the
neighborhood of the base point. The base point is perturbed one dimension at a time and
function value is evaluated resulting in general into a new base point. If no function
reduction is achieved, the step length is reduced and procedure is repeated.
A pattern move attempts to speed up search using information from preceding base
points. A translation in the best search direction is made and function is reevaluated. The
alternating sequence of exploratory and pattern moves is continued until convergence or
if the step length has reduced to specified small value. Figure 1-1 shows a schematic of
Hook-Jeeves methodology in two-dimension space

Figure 1-1: Schematic of Hook-Jeeves pattern search algorithm

1.4.2

SPSA Algorithm

Simultaneous perturbation stochastic approximation (SPSA) is a modified version of the

Kiefer-Wolfowitz algorithm (Kiefer & Wolfowitz, 1952), where stochastic gradient for
the function is estimated by perturbing one parameter at a time. Kiefer-Wolfowitz
algorithm is unfeasible as evaluation of objective function for perturbation in every
parameter is impractical in case of history matching.
A stochastic approximation (SA) for minimization uses some sort of random process to
select the search direction for every iteration (Robbins & Monro, 1951). SPSA improves
on the SA algorithm by simultaneously perturbing all model parameters randomly to
generate a search direction for every iteration. The search direction can be random walk
but is chosen downhill for every iteration. The algorithm behaves similar to steepest
descent direction (Gao, Li, & Reynolds, 2007) and can be implemented with any
simulator.

Chapter 2
2 Testing the Idea - Case Study I & II
The multistart optimization methodology was tested on two different synthetic box
models to test its applicability for continuous property based models and categorical
parameter based models. The reservoir models constructed for both case studies were of
identical dimensions and had same fluid properties. First set of models used continuous
property as primary variable, (in this case porosity) with a defined two point variogram.
Porosity models were generated using SGSIM 1 with statistical data was extracted from
well logs. The second set of models used categorical variable, in this case facies to model
reservoir heterogeneity. Reservoir facies models were generated in SNESIM 2 using a
channel-training image. The facies model used for the exercise was extracted from
Stanford VI reservoir (Castro, Caers, & Mukerji, 2005).
2.1

Model Description

The reservoir box model consisted of 4000 grid cells with 20 cells in X-direction, 20 cells
in Y-direction and 10 cells in Z-direction. A bi-modal histogram was used to populate
porosity in grid cells the first case. For the other case, a uniform distribution was used to
populate porosity for each facies. Permeability was assigned to grid cells using KozneyCarman based porositypermeability relationship (Mavko, Mukerji, & Dborkin, 2009).
Figure 2-1 plots the porosity-permeability relationship used in the models

1000

Oil
0.8

0.8

Water

0.6

0.6

0.4

0.4

0.2

0.2

100
0.1

0.2

0.3

Porosity,

0.4

0.5

Figure 2-1: Porosity-Permeability Relationship

SGSIM Sequential Gaussian Simulation

SNESIM Single Normal Equation Simulation

Relative Permeability, kr

Relative Permeability, kr

Permeability, k

10000

0
0

0.2

0.4

0.6

0.8

Water Saturation, Sw

Figure 2-2: Relative Permeability Curve

The reservoir model represented an inverted five-spot pattern. All five wells were
modeled fully penetrating. Both sets of models were constrained using porosity/facies at
the well locations. PVT properties for the fluids in the reservoir were modeled as black
oil with no free gas in the system. The reservoir was considered saturated with oil and
connate water at the start of simulation. The reservoir model was simulated for water
flood scenarios. Oil-water relative permeability curves were generated using Corey
functions (Ahmed, 2010; Dake, 2001). Figure 2-2 shows the relative permeability curves
used in the model. The key dynamic parameters of the model are listed in Table 2-1.
Table 2-1: Dynamic properties of the model

Property
Oil Viscosity
Water Viscosity
Rock Compressibility
Well Control
Production Strategy
(Water Flood)

2.2

Value
1.20 cp
0.325 cp
5 x 10-6 psi-1
Constant pressure injector and
Constant pressure producer
4 Water injectors in the Corner
1 Oil Producer in the Center

Dimension Reduction

Large set of realizations are required for extracting principal component using PCA
technique. An ensemble of 1000 realizations was generated for both model set using
SGSIM (Model 1: Porosity based model) and SNESIM (Model 2: Facies based
model). PCA was performed using MATLAB software on both ensembles to generate
principal components and their associated coefficients. The variance for each principal
component and cumulative variance was analyzed to select optimum number of principal
components for model reconstruction and thereby achieve dimension reduction.
Additionally, original model realizations were compared with models reconstructed using
limited set of principal components to verify the optimum number of principal
components.
Lower and upper bounds of the coefficients were also computed to determine the range of
coefficient for each principal component. The range of coefficients was further extended
by (+/-) twice the standard deviation of coefficients. New realizations for each model
were constructed using limited set of principal components by selecting coefficients
randomly sampled from the uniform distribution of the lower and upper bound.
2.3

Observables and Objective Function

The reservoir models were simulated for water flood for oil recovery. Four water
injectors were located near the corners of the model that inject water at constant bottomhole pressure. An oil producer was located at the center of the grid that is also
constrained by constant bottom-hole pressure. The five-spot well pattern described above

is simulated using Stanford's GPRS 3 simulator for a period of 90 days of reservoir

exploitation. Flow data at wells is sampled at an interval of 10 days.
=

=
=0

[ () () ]2
()

= 0.5

[ () ()]2
()

(2-1)

= 0.5

The observable () for a model refers to the cumulative production and injection data of
the wells over time, whereas () referred to the production and injection history dataset.
The historical data set was generated by simulating the true reference model for water
flood. The response function () used for optimization was the normalized variance of
model response with respect to observed data set. Equal weighs for production and
injection data was used as a part of optimization algorithm. Equation (2-1) shows the
response function used for optimization algorithm. The optimization algorithm attempted
to minimize the response function by perturbing the coefficients for each principal
component.
2.4

K - Medoid Clustering

A large set of random realizations were generated for both types model. The large set of
realization were clustered together to select 10 initial realizations for multi-start
optimization. Clustering was performed using multidimensional scaling (MDS) KMedoid clustering (Scheidt & Caers, 2009) which helps in selecting models for
optimization, which are distinctively different from each other.

Figure 2-3: Proxy distance computation for clustering

A proxy distance was used for purpose of MDS clustering. The proxy distance that was
used for both model set was calculated using harmonic average of permeability in grid
3

GPRS General Purpose Reservoir Simulator

http://pangea.stanford.edu/researchgroups/supri-b/research/research-areas/gprs

cells between each injector-producer combination. Figure 2-3 shows a schematic of proxy
distance computation. Permeability across the layers was arithmetically averaged in
vertical direction prior to computing harmonic averages along the injector producer.
This represented a four-dimensional vector for each realization that was mapped to a twodimensional space for clustering
2.5

Model Set - 1: Continuous Property based Model

2.5.1

Model Generation

The first model set used porosity as primary variable. Porosity was modeled as a
Gaussian variable using the histogram derived from well data and spatially correlated by
a known variogram. The histogram of the porosity data modeled is shown in the Figure
2-4. The parameters of the variogram model used for generating realizations are listed in
Table 2-2.
An ensemble of 1000 realizations was generated using SGSIM in SGeMS 4 constrained to
five well locations for principal component analysis and dimension reduction.
0.2

0.9

0.16

0.8

0.14

0.7

0.12

0.6

0.10

0.5

0.08

0.4

0.06

0.3

0.04

0.2

0.02

0.1

0
0

0.05

0.1

0.15

0.2

Porosity [Fraction]

0.25

0.3

Cumulative Frequency

0.18

Frequency

0
0.35

Figure 2-4: PDF and CDF of porosity: Model Set - 1

Table 2-2: Porosity Variogram: Model - 1

Parameter
Nugget
Range
Model

SGeMS - Stanford Geostatistical Modeling Software:

Value
10%
Max 10; Med 5, Min 2
Exponential

http:// sgems.sourceforge.net/

2.5.2

Dimension reduction and Clustering

PCA was carried out on model ensemble in MATLAB software to generate principal
components and coefficients. The number of dimension was reduced to 50 after
analyzing the variance associated with each principal component with intent to capture
approximately 50% of the total variance. A set of 100 new realizations was generated to
using limited set of principal components to select 10 starting models for optimization.
Clustering was performed by using the proxy distance defined earlier (Section 2.4). The
MDS plot of the selected 10 models and all other models is shown in Figure 2-5.

Figure 2-5: K-medoid clustering: Model Set - 1

2.5.3

Optimization Results

Figure 2-6: Cumulative oil production vs. Time Initial and final models: Model Set - 1

Figure 2-7: Cum. oil production vs. cum. water injection - Initial & final models: Model Set - 1

Hook-Jeeves algorithm was used as an optimization algorithm for perturbing models for
history matching. Figure 2-6 plots the cumulative oil production vs. time and Figure 2-7
plots the cumulative oil production vs. cumulative water injection for the initial starting
models and the final solution models for the period of history match. As it can be
observed, optimization results in reduction of mismatch with production history.

Figure 2-8: Layer 5 - History matched models: Model Set - 1

10

The process of multistart optimization results in 10 different realization that match the
production history. These realizations represent possible geological scenarios all of which
match production data within the chosen tolerance.
Figure 2-8 shows a cross section view (Layer 5) of the final models after optimization.
As it can be observed, they represent widely different scenarios.
2.5.4

Optimization Algorithm Efficiency

The variance of flow response with the historical data was computed (Response
Function) at end of every iteration to monitor efficiency of optimization. Figure 2-9 plots
the response function for the realizations as it is perturbed through the iterations. At the
end of Hook-Jeeves iteration, 80% of the initial models converged to a cost function
value of less than 10-2.

Figure 2-9: Optimization function efficiency: Model Set - 1

Multidimensional scaling can be also used to analyze the difference between the initial
guess models and the final optimized solution models. Figure 2-10 shows a
multidimensional-scaled plot of initial models and final perturbed models along with the
true solution mapped to two-dimension space. The response function value across ten
time steps was used as distance between the models.

11

Figure 2-10: MDS view of initial & final models: Model Set 1

2.5.5

Forecast Uncertainty

The final models were simulated for an additional period of 180 days (Total - 270 days)
to ascertain the uncertainty in forecast captured through these models. Figure 2-11 shows
the forecast of ten final models compared with the true solution. Oil production forecast
of the ten models vary from each other and provide a measure of uncertainty

Figure 2-11: Forecast uncertainty of history matched models: Model Set 1

2.6
2.6.1

Model Set - 2: Categorical Property based Model

Model Generation

The second model set was generated with an indicator property facies. Facies value of
one at a grid location represented sand while zero represented shale. A channel training
12

image was used to generate multiple realizations constrained to well locations using
multi-point geostatistics algorithm (Strebelle, 2002) - SNESIM in SGeMS for Stanford
VI reservoir (Castro, Caers, & Mukerji, 2005). A subset of Layer-3 of Stanford VI
reservoir model was extracted for the purpose of study. An ensemble of 1000 realizations
was used for PCA and dimension reduction. This model set has been used for previous
studies by Echeverria et.al. (Echeveria & Mukerji, 2009; Echeverria & Mukerji, 2009).
2.6.2

Dimension reduction and Clustering

Facies is a binary parameter and hence model reduction using PCA approximates the
process. Facies reconstructed with limited set of principal components would no longer
be binary but instead convert to a continuous property with values greater than one and
less than zero. Rescaling and thresholding was used to convert the reconstructed facies to
an indicator property. The principal components were extracted in MATLAB along
with associated coefficients and variance.
This image cannot currently be displayed.

Figure 2-12: Principal components (1st, 2nd, 5th & 30th): Model Set 2
(Yellow- Sand; Black Shale)

Figure 2-12 shows X-Y cross-section of few the principal components. The higher
principal components provide less and less information. Figure 2-13 plots the variance
associated with principal components. The first 30 principal components were selected
for model reconstruction as it captured over 50% of the total variance.

Figure 2-13: Variance of principal component: Model Set - 2

13

Figure 2-14 compares the reconstructed models generated using 10, 30 and 100 principal
components. As it can be observed, model reconstructed with 100 principal components
is very similar to model reconstructed with 30 principal components.

Figure 2-14: Model reconstruction: Model Set 2 (Yellow- Sand; Black Shale)

A set of 100 new random models were generated by using first 30 major principal
components. These models were clustered using proxy distance defined earlier (Section
2.4) to select ten starting models for optimization. The MDS plot of the selected ten
models and all other models is shown in Figure 15. The reconstructed values were
mapped to porosity and permeability as inputs for flow simulation

Figure 2-15: K-medoid clustering: Model Set 2 (Yellow- Sand; Green Shale)

14

2.6.3

Inclusion of Seismic Tomography Data

An additional set of optimization was performed to test the impact of inclusion of seismic
data set. The travel time data associated with cross-well seismic tomography (Stewart &
Domenico, 1991; Lo & Inderwiesen, 1994) is used as the second observable. Seismic
tomography estimates sub-surface properties (seismic velocity or attenuation) by
analyzing elastic wave field propagation from sources to receivers. Travel-time data for
cross-well tomography are governed by rock physics relations (Mavko, Mukerji, &
Dborkin, 2009) that are dependent on rock porosity, saturation and facies type. Figure
2-16 shows a schematic of cross-well tomography and tomography survey set-up for the
study.

Figure 2-16: Tomography Survey: Model Set - 2

Cross-well tomography travel time was calculated for the true reference model and was
used for optimization by modifying Equation (2-1) by including the normalized variance
between model response and recorded data of reference model. Equal weight for flow
data and seismic data was used for optimization algorithm. Equation (2-2) shows the
modified response function used for flow & seismic data optimization.
=

=
=0

[ () () ]2
=

()

+
=0

= 0.25

[ () ()]2

[ () () ]2

[ ()]2
= 0.25

()

(2-2)

= 0.5
15

2.6.4

Optimization Results

Hook-Jeeves algorithm was used as an optimization algorithm for history matching. The
optimization algorithm reduces the mismatch between model response and history
2.6.4.1 Flow Only

True
Figure 2-17: Layer 5: Flow only history matched models: Model Set 2
(Yellow- Sand; Green Shale)

Figure 2-17 shows the X-Y cross section of the final solution models along with the true
reference model. Here yellow color represents the sand and green represents the shale
facies. As it can be observed, final solution models vary with each other but at the same
time all models contain the trace of bottom channel facies in Layer 5as in true
reference model.
2.6.4.2 Flow and Seismic Data
Inclusion of cross-well tomography data provides additional information across the interwell section. The data is not descriptive for the entire reservoir hence the impact of
seismic data does not drastically change the final solution models. Figure 2-17 shows the
X-Y cross section of the final solution models along with the true reference model. Here
yellow represents the sand facies while green represents the shale facies. As it can be
observed, the solution models are different from true reference model and all models
contains majority of sand facies along the lower edge of the model similar to true model.

16

True

Figure 2-18: Layer 5: Flow & seismic history matched models: Model Set 2
(Yellow- Sand; Green Shale)

2.6.4.3 Ensemble average and Variance

The ensemble of final solution was combined to compute E-type and variance. Figure
2-19 shows the E-type of first, fourth, seventh and ninth layer and Figure 2-20 displays
the variance. The inclusion of cross-well seismic data slightly reduces the variance.
True Model

Production Data Only

Production & Seismic Data

Layer - 1
Layer - 4
Layer - 7
Layer - 9
Figure 2-19: E- Type of final solution models: Model Set 2 (Yellow Sand; Green Shale)

17

True Model

Production Data Only

Production & Seismic Data

Layer - 1

Layer - 4

Layer - 7

Layer - 9

Figure 2-20: Variance of final solution models: Model Set 2: (Yellow Sand; Green Shale)

2.6.5

Optimization Efficiency

Figure 2-21 plots the response function over the function evaluations. The optimization
algorithm is able to reduce the response function value to less than 10-2. In addition, it can
be observed that the inclusion of the cross-well seismic tomography data slows the
convergence algorithm and it takes longer to converge to values less than 10-2.

Figure 2-21: Optimization efficiency : Model Set - 2

18

Forecast Uncertainty
The converged models were further simulated to quantify the forecasts and their
variability. Figure 2-22 compares the mean and spread of prediction runs for both sets of
inversion with true model forecast. As expected, inclusion of seismic data in optimization
reduces the forecast uncertainty.

Figure 2-22: Forecast variability: Model Set 2

2.7

Conclusions

Multi-start optimization in reduced model dimension works effectively with both

continuous property and categorical variable for generating multiple geological models.
The resulting models are consistent with production history of the field but have different
forecast and hence represent possible subsurface realizations. Model dimension reduction
using PCA does cause of loss of geological information associated with higher order
Eigen vectors, but offers a reasonable model approximation. The technique also offers
flexibility to include additional data such as seismic data or well survey data for reducing
solution space uncertainty.

19

Chapter 3
3 Expanding the Idea Case Study III
The observations from the box model guided us to test the idea on a larger field-scale
model with complex features. The intent for a larger model was to ascertain the
usefulness and practicality of dimension reduction using PCA where reservoir models are
complex and forward simulations are computationally demanding. Additionally, a larger
field-scale model envisaged development of workflow that integrated geostatistical
simulation technique within the framework of optimization. A large synthetic reservoir
model was hence created with over 90,000 grid cells for this study. Multistart
optimization algorithm was used got history matching with multiple data sources like
field production data, seismic data and spinner survey data
3.1
3.1.1

Model Description
Geology

The reservoir is a box model made of 75 cells in Xdirection, 50 cells in Y-direction and
25 cells in Z-direction. The reservoir was named Zamba. The reservoir structure dips
towards south at an angle of 20 degrees with respect to horizontal. The reservoir consists
of two facies representing sand and shale. A channel-training image was used to generate
the reservoir structure using SNESIM 5 algorithm of SGeMS 6. Figure 3-1 shows a
schematic of the reservoir structure and the training image used for generating the
reservoir facies. Sand is shown as red while background shale is shown in blue. The
reservoir has eight wells placed in 3 rows. The top row of well consists of 3 producers
(Well 1, 2 and 3), middle row contains 2 producers (Well 4 and 5) and bottom row
contains 3 injectors (Well 6, 7 & 8)

Figure 3-1: Reservoir structure (Left) and Channel Training Image (Right): (Red Sand; Blue
Shale)
5

SNESIM - Single Normal Equation Simulation

SGeMS - Stanford Geostatistical Modeling Software:

http:// sgems.sourceforge.net/

21

3.1.2

Petrophysics

Both sand and shale have been considered porous and permeable in the reservoir model.
The porosity in sand and shale facies were populated independently using SGSIM 7
algorithm of SGeMS. Porosity variogram in sand was considered anisotropic with
continuity along the channel direction, while shale variogram was less anisotropic with
maximum, medium and minimum direction comparable with each other. Figure 3-2
shows the porosity distribution in the reservoir. Figure 3-3 shows the porosity histogram
of the sand and shale facies. The effective porosity in sand varies from 16% to 33% while
it varies from 3% to 16% in shale.

Figure 3-2: Porosity distribution in the reservoir

Figure 3-3: Histogram of porosity for Sand (Left) and Shale (Right)

Permeability in the grid cells was computed using a porosity permeability relationship. A
modified porosity permeability relationship from Norne field (Suman & Mukerji, 2012)
was used for the model. Equation (3-1) and Equation (3-2) represents the porosity
permeability relationship used in the reservoir model for sand and shale respectively.
Sand
Shale
7

log = 5.095 + 2.0580

log = 14.610 + 0.1745

SGSIM - Sequential Gaussian simulation

22

(3-1)
(3-2)

3.1.3

Relative Permeability & Capillary Pressure

The reservoir model is simulated for water flood scenario with both oil-water contact and
gas-oil contact. A three-phase relative permeability model was used to model flow across
grid cells. Oil-water relative permeability for sand was generated using Corey functions
(Ahmed, 2010; Dake, 2001) with Corey exponent of 2.75 and 4 for oil and water
respectively. The Corey exponent for oil and water in case of shale are 3.25 and 3.75
respectively. Connate water saturation in sand and shale was assigned as 12% and 30%
respectively.
Modified capillary pressure data from Norne field was used for sand and shale facies.
The relative permeability model and capillary pressure model used in the model are
shown in Figure 3-4.

Relative Permeability

0.8
0.7

0.9
0.8
0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0.0

0.0
0.0

0.2
0.4
0.6
0.8
Water Saturation [fraction]

1.0

Shale
Sand

30
Capillary Pressure [psi]

0.9

35

1.0

Kro - Shale
Kro - Sand
Krw - Shale
Krw - Sand

Relative Permeability

1.0

25
20
15
10
5
0
0.0

0.2
0.4
0.6
0.8
Water Saturation [fraction]

1.0

Figure 3-4: Oil Water relative permeability curve (Left) and Capillary pressure curve (Right)

3.1.4

Fluid Properties

The reservoir simulation model included a three phase fluid system - Oil, Water and Gas.
An isothermal black oil formulation was used to describe fluid behavior. Oil in the
reservoir was assigned an API gravity of 30o with a saturation pressure of 3600 psi. The
gravity of the gas in the reservoir was assigned as 0.85. The salinity of water was
assumed 50000 ppm. The reservoir temperature was set at 180 o F. The model was
initialized with an oil-water contact (OWC) within the structure at 6800 feet. No free gas
existed in the model at the time of initialization.
The PVT properties for oil and gas were generated for different reservoir pressures using
Standing correlation (Standing, 1947). PVT properties for water were calculated using
McCain correlation (McCain, 1989). The key fluid properties are listed in Table 3-1

23

Table 3-1: Fluid Properties

Fluid Property
Oil API
Gas Gravity
Separator Gas Gravity
Reservoir Temperature
Saturation Pressure
Water Salinity
3.1.5

Value
30o API
0.85
0.90
180o F
3600 psi
50,000 ppm

Rock Physics Model

Rock physics model defines how seismic attributes change due change in pressure and
saturation of the rock. An uncemented-sand (or soft-sand) is used for sand facies. The
model assumes that the sand grains were deposited as a dense random pack of identical
spherical grains and average number of contacts per grain between five and nine (Mavko,
Mukerji, & Dborkin, 2009). Seismic attributes are defined for this type of setting using
Hertz-Mindlin theory (Mindlin, 1949). Sand is assumed to be made of Quartz, Feldspar
and fragments, while shale is made of clay and quartz.
Seismic attributes for shale facies are estimated using Gardners power law for ( ) P
wave velocity (Gardner, Gardner, & Gregory, 1974) and Mudrock line for ( ) S wave
velocity (Castagna, Batzle, & Eastwood, 1985). In-situ fluid properties are computed
using Batzle-Wang relations (Batzle & Wang, 1992) which are used to compute seismic
properties for rock saturated with fluid mixture using Gassmann fluid substitution
(Gassmann, 1951; Mavko, Mukerji, & Dborkin, 2009). The key parameters of rock
physics model are listed in Table 3-2.
Table 3-2: Rock Physics Model

Parameter
Sand Composition
Shale Composition
Sand Vp & Vs
Shale Vp & Vs
3.1.6

Value
80% Quartz, 15% Feldspar, 5 % Rock fragments
85% Clay, 15% Quartz
Uncemented Soft Sand Model
Gardners relation (Vp) & Castagna relation (Vs)

Production Strategy

The Zamba reservoir model was simulated for water flood scenario. The top and
middle rows of well were assigned as producers. The top row of the wells (Well 1, 2
and 3) were completed with preformation in all 25 layers, while the middle row of wells
(Well 4 and 5) were completed in top 20 layers. The middle group of wells was
constrained at a rate of 13000 bbl. per day, while top group wells were constrained with a
maximum liquid rate of 8000 bbl. per day each. A lower rate constraint was assigned on
top group of wells to minimize formation of secondary free gas due to pressure drop. All

24

producers had a minimum bottom-hole pressure (BHP) constraint 50 psi. Figure 5 shows
the location of wells in the reservoir

Figure 3-5: Zamba Production Strategy

Water was injected through the bottom row of wells (Well 6, 7 and 8) to push the oil to
the producer and to maintain producers. The injectors were completed in all 25 layers and
were set to inject at a rate of 15000 bbl. per day with a maximum injection pressure of
3500 psi. Injected water replaced the voidage created in the reservoir by the producers
thereby maintaining pressure, at the same time pushing the water towards producers. The
reservoir model was simulated for 4 years of water flood.
Oil, gas and water production rate were recorded from the well at every month.
Production rates from every layer were recorded once every year simulating a production
logging survey spinner survey in the well.
3.2
3.2.1

Model Decomposition & Reconstruction

Principal Component Analysis

The multistart optimization methodology for history matching described in Section 1.1 is
carried out in a reduced model space. A large set of reservoir models 10,000 models
were generated with the same training image using SNESIM 8 algorithm in SGeMS
software. The training image is indicative of the reservoir geology is a prior information
available with geoscientists
Principal component analysis was carried out the matrix of reservoir model to compute
the major principal components. The principal components are set of values of linearly
uncorrelated variables, such that first principal component has the largest possible
variance and each succeeding component in turn has the next highest variance. PCA of

SNESIM Single Normal Equation Simulation

25

the model set resulted in 10,000 principal components. Figure 3-6 and Figure 3-7 shows
a plot of variance and cumulative variance of first 3000 principal components

Figure 3-6: Variance of first 3000 principal components

Figure 3-7: Cumulative variance of first 3000 principal components

3.2.2

Dimension Reduction

Few of the models were reconstructed using limited set of principal components and were
visually compared with true model for consistency. Figure 3-8 shows a reconstruction of
layer 5 of the model number 1000 using 30, 70, 1000, 3000 and 5000 principal
components. As it can be observed, the reconstructed model closely resembles the true
model when it is reconstructed using 3000 principal components. The first 3000 principal
components represent 90% of the total variance that can be inferred from Error!
Reference source not found. & Figure 3-7. This is in line with observation made by in
other papers (Choudhary, Mukerji, & Echeverria, 2011; Echeverria & Mukerji, 2009)
which suggests that models reconstructed with 25% of non-zero principal components
closely match the true model.
Additionally, it can be visualized that model reconstructed with 70 principal components
reproduce the significant channels. The higher order principal components only add to
finer details to these channels.

26

Figure 3-8: Reconstruction of model using limited set of principal components.

(Yellow Sand; Black Shale)

3.2.3

Model Generation

Figure 3-9: Bounds for PCA coefficients

The computational capacity limits the number of dimensions in which optimization can
be performed, hence a smaller subset needs to be selected, which at the same time should
be able to reproduce the key features. For the purpose of optimization for this case study,
models were reconstructed using first 3000 principal components, while optimization was
only performed in the first 70 principal components. Coefficients for each principal
component are selected sampled from a uniform distribution between minimum and
maximum coefficient value obtained as part of principal component analysis. The

27

minimum and maximum bound for coefficients of the 3000 major principal vectors is
shown in Figure 3-9 along with coefficients for a randomly generated model.

Figure 3-10: Randomly generated models using 3000 major principal components
(Yellow Sand; Black Shale)

Sand and shale facies is assigned in the model generated using with limited components
using a threshold equal to prior sand proportion. Figure 3-10 represents two-dimensional
X-Y slices (Layer 1) for three random realizations generated using 3000 principal
components.
3.3

History

Figure 3-11: Production history of the Zamba reservoir

The Zamba reservoir was simulated using Eclipse for water injection for a period of 4
years with wells on rate and pressure constraint. Injected water maintains the pressure of
the reservoir and pushes the oil towards the producers. The simulated production for 4
years is used as production history for purpose of optimization. Additionally, the layer
wise flow rate data for the wells was recorded annually to simulate PLT survey. Figure
11 shows the field production history of the reservoir.

28

Figure 3-12: Fluid saturation at Initial Condition, End of first year and at end of third year

Figure 3-12 shows the oil, water and gas saturation in Layer 5 of the true reference model
at initial condition, at end of 1st year of water flood and at end of 3rd year of water flood.
As it can be observed, water injected is gradually pushing oil towards producers up-dip.
The saturation data combined with rock physics model are input for synthetic seismic
computation, which is used for history matching.
Due to heterogeneity in the reservoir, production and injection in the well do not occur
uniformly across layer. Spinner surveys (or PLT Data) capture the vertical trend of
heterogeneity by measuring the oil, water and gas flow rate from each layer of the well.
Figure 3-13 plots the spinner survey of the Well - 4 (producer) and Well - 6 (injector) at
end of 1st year, 2nd year and 3rd year. The data shall be used for optimization algorithm

Figure 3-13: Spinner survey data for well - 4 and well - 6

29

3.4
3.4.1

Optimization
Workflow

Figure 3-14: Optimization workflow

An automated workflow was developed in MATLAB for the purpose of optimization.

Facies for the reservoir is constructed using randomly generated coefficients and limited
set of principal vectors. The facies model is used an input in SGeMS for populating
porosity and permeability in the grid. Porosity for each facies is generated using SGSIM 9,
while permeability is assigned using porosity-permeability relationship. The generated
reservoir model is simulated for water flood in Eclipse and observables compared
against history. The mismatch between simulation output and history is used to perturb
PCA coefficients, which in turn generates an updated reservoir model. Figure 3-14 shows
a schematic of the workflow
3.4.2

Observables & Response Function

The observables for reservoir model include both well level data and grid level data. Oil,
water and gas phases are recorded for wells at the end of each month, while pressure and
saturation data for entire grid is recorded on a yearly basis. The grid level data (pressure
and saturation) is used to generate normal-incidence impedance grid that is also used for
optimization. Additionally, layer wise phase rates are also recorded for every well on a
yearly basis to simulate PLT spinner surveys.
Equation (3-3) represents the cost function used for optimization. Here p (t) denotes the
model response for given parameter, p (t) denotes the observed history data,
p represents the weight for given data type, WC represents water cut and PLT
represents spinner survey data for wells. A weighted sum of the normalized variance of
9

SGSIM Sequential Gaussian Simulation

30

the model response with respect to observed data set is used as a response function for
optimization. The weight for each data type can be altered based on data quality.
=

=0

p = Field WC, Well WC and PLT data

() ()
()

(3-3)

Optimization was performed for three different scenarios - flow data only, flow data &
PLT data and flow, PLT and seismic data. As the Zamba reservoir is primarily under
water flood, only water cut at field level and well level was included for purpose of flow
only optimization. In case of flow and PLT data optimization, layer wise phase rates (oil,
water and gas) for all eight wells combined with field level and well-level water cut was
used computing the response function. For the third scenario, Born-approximation
(Mukerji, Mavko, & Rio, 1997; Lo & Inderwiesen, 1994) of P-wave velocity was for
every grid cell was also included in the response function computation for optimization.
The seismic data is a field-wide data set can help in constraining facies away from wells.
The weights used for all the three scenarios are listed in Table 3-3.
Table 3-3: Weight for Optimization (Normalized by Total in Algorithm)

Flow Only
Field Water Cut 10
Well Water Cut 20

Flow & PLT Data

Field Water Cut 10
Well Water Cut 20
Well PLT Data 30

Total - 30

Total - 60

3.5

Flow , PLT and Seismic Data

Field Water Cut 10
Well Water Cut 20
Well PLT Data 30
Seismic Data - 30
Total - 90

Results

A set of 10 randomly generated initial models was optimized using the optimization
algorithms. The results for each of the scenario are discussed below.
3.5.1

Flow Only

The response function for "Flow only optimization" was computed using water cut
measurement at field and well level. The process of optimization resulted in seven
models with appreciable match. Figure 3-15 shows the water cut response of the reservoir
and individual wells. At the end of 3rd year, water breakthrough has already occurred
through the middle row of producers (Well - 4 & Well - 5). No water breakthrough has
occurred in top row of producers (Well - 1, Well - 2 & Well - 3) as such have not been
plotted in Figure 3-15. The blue lines denotes models which match water-cut history,
grey lines denotes model which did not match water-cut history and red points denotes
the history dataset.

31

Well - 4

0.6

0.5

0.5

0.4
0.3
0.2
0.1

0.4
0.3
0.2

500

1000

Time (Days)

1500

0.2
0.15
0.1
0.05

0.1
0

Field

0.25

Water Cut (fraction)

0.6

Well - 5

0.7

Water Cut (fraction)

Water Cut (fraction)

0.7

500

1000

Time (Days)

1500

500

1000

Time (Days)

1500

Figure 3-15: Water Cut - flow only Optimization

Red History, Blue Models matching History, Grey Models not matching History

Figure 3-16 shows the XY slices of the fifth layer of models, which converged during
iterations along with the True Model. As it can be observed, the converged solutions
are quite different from the true solution.

Figure 3-16: Layer 5 of history matched models: Flow only optimization

3.5.2

Flow & PLT Data

The second set of realization included the PLT dataset. Spinner survey data provides
layer wise phase-rates that provides information about vertical heterogeneity and spatial
connectivity of channels between well pairs. A good match of PLT in turn improves the
water cut match since well-level water-cut is an aggregated sum of spinner survey data. A

32

higher weight to PLT data should also be avoided due to inherent inaccuracies in
measuring phase rates down-hole. The process of optimization resulted in 11 models with
appreciable match. The water cut match of the model with respect to the history dataset is
shown in Figure 3-17.
Well - 4

0.6

0.5

0.5

0.4
0.3
0.2
0.1

0.4
0.3
0.2

500

1000

Time (Days)

1500

0.2
0.15
0.1
0.05

0.1
0

Field

0.25

Water Cut (fraction)

0.6

Well - 5

0.7

Water Cut (fraction)

Water Cut (fraction)

0.7

500

1000

Time (Days)

1500

500

1000

Time (Days)

1500

Figure 3-17: Water cut - flow & PLT data optimization

Red History, Blue Models matching History, Grey Models not matching History

The PLT data match for an injector (Well - 6) and a producer (Well - 4) is shown in
Figure 3-18 and Figure 3-19 respectively. The optimization process results in models that
capture all significant vertical flow rate variation in the well.

Figure 3-18: PLT data match - End of 3rd year for Well - 6 (Injector)

33

Figure 3-19: PLT data match - End of 3rd year for Well - 4 (Producer)

The XY slices of the fifteenth layer of solution models are shown in Figure 3-20 along
with the True Model. The converged models do contain the major channel trends as in
the true reference models.

Figure 3-20: Layer 15 of history matched models: Flow & PLT data optimization

3.5.3

Flow, PLT and Seismic Data Optimization

The third set of optimization was performed using all three data set. The P-wave velocity
cube provides field-wide information at lower resolution. The change in P-wave velocity

34

at each grid cell is a complex function of pressure and saturation. Changes in P-wave
velocity over time can only be resolved for a significant change in saturation and
pressure. The water front movement in the reservoir represents a transition from oilsaturation to water-saturation and affects P-wave velocities in the grid cells. Seven
models converged to solution models. Figure 3-21 shoes the water-cut match of middle
row producers and field level water-cut with respect to history data set
Well - 4

0.5
0.4
0.3

0.2
0.1

0.6
0.5
0.4

0.3
0.2

500

1000

Time (Days)

1500

0.2
0.15
0.1
0.05

0.1
0

Field

0.25

Water Cut (fraction)

0.6

Well - 5

0.7

Water Cut (fraction)

Water Cut (fraction)

0.7

500

1000

Time (Days)

1500

500

1000

Time (Days)

1500

Figure 3-21: Water cut - flow, PLT and seismic data optimization
Red History, Blue Models matching History, Grey Models not matching History

The P-wave velocity match of selected models is shown in Figure 3-22. The converged
models capture the major channel trends in the reservoir that are seen in the true model.

Figure 3-22: P-wave velocity match - flow, PLT and seismic data optimization

35

The XY slices of the fifth layer of models that converged during iterations are shown in
Figure 3-23 along with the True Model. As it can be observed, the converged solutions
capture major channel trends, but they still differ from each other due to different sand
channel placement in the reservoir.

Figure 3-23: Layer 15 of history matched models: Flow, PLT and seismic data optimization

3.6

Rejection sampling algorithm

Figure 3-24: Comparison with rejection sampler

36

The results of optimization are compared with rejection sampler to estimate the efficiency
of optimization algorithms. The "rejection sampler" or acceptance-rejection method is
called an "exact" sampler as it perfectly follows the Bayes rule. A rejection sampler
requires exhaustive evaluation of very large set of prior models. All models that do not
match the historical data set are rejected, while retaining subset of models which history
(Caers, 2011).
For rejection sampling, 10,000 prior models generated for PCA were simulated for water
flood using Eclipse and model response compared with the history. Rejection sampling
was carried out for all three scenarios - flow only and flow & PLT data and flow, PLT
and seismic data optimization. Over 1000 models were selected during rejection sampling
of flow only case scenario, 71 models were only selected for flow & PLT optimization
and 42 models matched the history data within tolerance for flow, PLT and seismic data
optimization. Figure 3-24 compares the X-Y cross-section of the E-type of models
obtained from rejection sampling with E-type obtained from multi-start optimization.

3.7

Optimization Efficiency

Figure 3-25: Optimization Code Efficiency for Hook Jeeves algorithm

Convergence is slow with both the algorithm for a large model used in case study. While,
Hook- Jeeves algorithm moves step-by-step towards the minima, SPSA algorithm uses
stochastic gradient computed in randomly generated directions. The convergence for
Hook-Jeeves algorithm is shown in Figure 3-25 for first 40 iterations, while Figure 3-26
shows convergence of SPSA algorithm for four convergence cycles. The lines in green
denote the flow & PLT optimization scenario while lines in red represent the flow, PLT
and Seismic data optimization. The convergence rate for a large model like "Zamba" was
slow, due to complex spatial distribution of facies. Inclusion of additional dataset further
slows the convergence.

37

Figure 3-26 : Optimization Code Efficiency for SPSA algorithm

3.8

Forecast Uncertainty

Figure 3-27: Forecast of history matched models

The history-matched models obtained from multi-start optimization were further

simulated in forecast mode for three more years to ascertain the variability in the forecast.
Figure 3-27 shows the water cut forecast obtained from history -matched models. As it
can be observed, water cut prediction varies within these models. The variation in
forecast is less for flow only optimization due to tighter tolerance used for history matching. The tolerance of flow & PLT data optimization and flow, PLT and seismic
data optimization was same. It can be observed that the uncertainty expressed by the
models from flow, PLT and seismic data is slightly less than the PLT data optimization
similar to the observation made in Section 2.6.4. Inclusion of additional data set further

38

constraints the reservoir parameters resulting in models that are more comparable to
"true" reference model. The break in the water cut trend at the end of third year is due to
change of group oil-rate constraint for the top row of producers from 30,000 STB/day to
33000 STB/day.

3.9

Conclusions

The technique of multi-start optimization in reduced dimension is a simple to use

methodology for generating multiple geological models. The resulting models are
consistent with production history of the field hence represent possible subsurface
realizations. This method thus helps in capturing non-uniqueness and variability in
subsurface modeling and is able to locate possible realizations, which should be
investigated prior to any development decision.

39

Chapter 4
4 Summary
This thesis presents a workflow for generating multiple history-matched models using
optimization in reduced dimensions. In this thesis, it has been demonstrated through
multiple case studies that field observations i.e. production data, seismic data, spinner
data etc. are too sparse for defining reservoir parameters accurately and multiple solution
models exists which closely match the historical observation.
In the thesis, it has been shown that principal component analysis (PCA) can be used to
reduce the dimension of the reservoir model to less than ten percent of original
dimension. It is also shown that random model generated using PCA can be optimized in
in a multistart approach to generate history-matched models that are different from each
other produce varying forecast for hydrocarbon production.
In the thesis, it has also been demonstrated that inclusion of spinner data representing
injector-producer connectivity and fieldwide seismic can help in reducing the
uncertainty of the posterior models. Any economic analysis of the field should be carried
out only after analyzing production forecast from all models that matched history.
It can also be concluded from Case III that PCA technique cannot accurately reproduce
fine scale features and other model reduction technique should be tested in cases where
facies description is non-linear. The choice of model reduction technique and number of
reduced dimensions is a subjective discussion. The section should be based on analysis of
unconstrained prior models.

41

Nomenclature

Porosity

Reservoir property
Posterior model space: Space defined by history matched models

Dimension of observations i.e. history data set

Observation time

()

Model observables - History data

Permeability
Weight for optimization

Reservoir model

()

Number of dimensions
Reduced number of dimensions for optimization

End of history
Prior model space: Reservoir model space made of dimensions

Model response - Simulation output

43

References
Ahmed, T. (2010). Petroleum Engineering Handbook. Gulf Professional Publishing.
Batzle, M., & Wang, Z. (1992). Seismic properties of pore fluids. Geophysics, 57(11),
1396-1408.
Caers, J. (2011). Modeling Uncertaininty in Earth Sciences. John Wiley and Sons.
Castagna, J. P., Batzle, M. L., & Eastwood, R. L. (1985, April). Relationships between
compressional-wave and shear-wave velocities in clastic silicate rocks.
Geophyiscs, 50(4), 571-281.
Castro, S., Caers, J., & Mukerji, T. (2005). The Stanford VI Reservoir. 18th Annual
Affiliate Meeting. Stanford: SCRF.
Choudhary, M., Mukerji, T., & Echeverria, D. (2011). Generation of Multiple History
Matched Models using multistart optimisation. 24th Affiliate Meeting. Stanford:
Stanford Center of Reservoir Forecasting.
Cruz, P. S., Horne, R., & Deutsch, C. V. (2004). The Quality Map: A Tool for Reservoir
Uncertainty Quantification and Decision Making. SPE Reservoir Evaluation &
Engineering, 7(1), 6-14.
Dake, L. (2001). The practice of reservoir engineering. Elsevier.
Echeveria, D., & Mukerji, T. (2009). Robust scheme for inversion of seismic and
production data for reservoir facies modeling. International Exposition and
Annual Meeting. Houston: SEG.
Echeverria, D., & Mukerji, T. (2009). A Robust Scheme for Spatio-Temporal Inverse
Modeling of Oil Reservoirs. 18th World IMACS/MODSIM Congress, (pp. 42064212). Cairns, Australia.
Friedmann, F., Chawathe, A., & Larue, D. (2003). Assessing Uncertainty in Channelized
Reservoirs Using Experimental Designs. SPE Reservoir Evaluation &
Engineering, 6(4), 264-274.
Gao, G., Li, G., & Reynolds, A. C. (2007). A Stochastic optimization algorithm for
automatic history matching. SPE Journal, 196 - 208.
Gardner, G. H., Gardner, L. W., & Gregory, A. (1974). Formation velocity and density the diagnostic basic for stratigraphic traps. Geophysics, 39, 770-780.
Gass, P. C. (2007). Principal component analysis. Northridge: Calfiornia State university.
Gassmann, F. (1951). ber die elastizitt porser medien. Viertel. Naturforsch. Ges.
Zrich, 96, 1-23.
Gottfried, B. S., & Weisman, J. (1973). Introduction to Optimization Theory. Englewood
Cliffs: Prentice Hall.
Himmelblau, D. M. (1972). Applied Non Linear Programming. New York: McGraw-Hill.
Hooke, R., & Jeeves, T. A. (1961, April). "Direct Search" Solution of numerical and
statistical problems. Journal of the ACM (JACM), 8, 212-229.
Huang, X., Meister, L., & Workman, R. (1998). Improving production history matching
using time-lapse seismic data. The Leading Edge, 1430-1433.
Jolliffe, I. T. (2002). Principal Component Analysis (Second ed.). New York: SpringerVerlag.

45

Kading, H. W. (1976). Horizontal-Spinner, A new Production Logging Technique. The

Log Analyst, 3-7.
Kiefer, J., & Wolfowitz, J. (1952). Stochastic estimation of a regression function. The
annals of mathematical statistics, 23, 462-466.
Litvak, M., Christie, M., Johnson, D., Colbert, J., & Sambridge, M. (2005). Uncertainty
Estimation in Production Predictions Constrained by Production History and
Time-Lapse Seismic in a GOM Oil Field. SPE Reservoir Simulation Symposium
(p. 93146). The Woodlands, Texas: SPE.
Lo, T.-W., & Inderwiesen, P. L. (1994). Fundamentals of Seismic Tomography. Tulsa:
Society of Exploration Geophysicists.
Maulik, U., Bandyopadhyay, S., & Mukhopadhyay, A. (2011). Multiobjective Genetic
Algorithms for Clustering: Applications in Data Mining and Bioinformatics.
Heidelberg: Springer.
Mavko, G., Mukerji, T., & Dborkin, J. (2009). The Rock Physics Handbook (Second ed.).
Cambridge University Press.
McCain, W. D. (1989). The Properties of Petroleum Fluids (Second ed.). Tulsa:
PennWell Books.
Mindlin, R. D. (1949). Compliance of elastic bodies in contact. Journal of Applied
Mechanics, 16, 259-268.
Mukerji, T., Mavko, G., & Rio, P. (1997). Scales of Reservoir Heterogeneities and
Impact of. Mathematical Geology, 29(7), 933-950.
Panda, M., & Nottingham, D. (2011). Systematic Surveillance Techniques for a Large
Miscible WAG Flood. SPE Reservoir Evaluation & Engineering, 299-309.
Rivera, N., Meza, N., Kim, J., Clark, P., Garber, R., Fajardo, A., & Pea, V. (2007).
Static and Dynamic Uncertainty Management for Probabilistic Production
Forecast in Chuchupa Field, Colombia. SPE Reservoir Evaluation & Engineering,
10(4), 433-439.
Robbins, H., & Monro, S. (1951). A stochastic approzimation method. The Annals of
Mathematical Statistics, 22, 400-407.
Sarma, P., Durlofsky, L. J., Aziz, K., & Chen, W. H. (2006). Efficient real-time reservoir
management using adjoing-based optimal control and model updating.
Computational Geosciences, 3-36.
Scheidt, C., & Caers, J. (2009). Uncertainty Quantification in Reservoir Performance
Using Distances and Kernel Methods - Application to a West Africa Deepwater
Turbidite Reservoir. SPE Journal, 14(4), 680-692.
Slater, G., & Durrer, E. J. (1971). Adjustment of Reservoir Simulation Models to match
field performance. SPE Journal, 11(3), 295-305.
Srinivasan, S., & Deutsch, C. V. (2004). Data Sufficiency for Reservoir Development
Decision-Making in the Presence of Uncertainty. Journal of Canadian Petroleum
Technology, 43(3), 52-60.
Standing, M. B. (1947). A Pressure-Volume-Temperature Correlation For Mixtures Of
California Oils And Gases. Drilling and Production Practice, 275-287.
Stephen, K. D., & MacBeth, C. (2008). Reducing Reservoir Prediction Uncertainty by
Updating a Stochastic Model Using Seismic History Matching. SPE Reservoir
Evaluation & Engineering, 11(6), 991-999.

46

Stephen, K. D., Soldo, J., MacBeth, C., & Christie, M. (2006, December). MultipleModel Seismic and Production History Matching: A Case Study. SPE Journal,
11(4), 418-430.
Stewart, R. R., & Domenico, S. N. (1991). Exploration Seismic Tomography:
Fundamentals. Tulsa: Society of Exploration Geophysicists.
Strebelle, S. (2002). Conditional simulation of complex geological structures using multipoint geostatistics. Mathematical Geology, 34, 1-21.
Suman, A., & Mukerji, T. (2012). Sensitivity Study of Rock Physics Parameters for
Modeling the Time Lapse Signature of the Norne Field. 25th Annual Affiliate
Meeting. Pacific Grove: SCRF.
Vogelij, H. N., Leach, M. J., & Kapteyn, P. K. (1993). Multilayer Flowmeter Testing
Combined With 3D Field Modeling To Enhance the Ameland Field Development.
SPE Formation Evaluation, 5-10.
Yoelin, S. D., & Howald, C. D. (1970). Production Logging as Used To Solve Water
Injection Problems in the Huntington Beach Offshore Field. Journal of Petroleum
Technology, 1083-1088.

47