Pet. Sci.

(2015) 12:114–128
DOI 10.1007/s12182-014-0005-6

REVIEW

A review of closed-loop reservoir management

Jian Hou • Kang Zhou • Xian-Song Zhang •

Xiao-Dong Kang • Hai Xie

Received: 13 July 2014 / Published online: 22 January 2015

Ó The Author(s) 2015. This article is published with open access at Springerlink.com

Abstract The closed-loop reservoir management tech- Keywords Closed-loop reservoir management 
nique enables a dynamic and real-time optimal production Automatic history matching  Reservoir production
schedule under the existing reservoir conditions to be optimization  Gradient-based algorithm  Gradient-free
achieved by adjusting the injection and production strate- algorithm  Artificial intelligence algorithm
gies. This is one of the most effective ways to exploit
limited oil reserves more economically and efficiently.
There are two steps in closed-loop reservoir management: 1 Introduction
automatic history matching and reservoir production opti-
mization. Both of the steps are large-scale complicated With the rapid development of the world economy, the
optimization problems. This paper gives a general review depletion of oil resources increases year by year. It is now
of the two basic techniques in closed-loop reservoir man- difficult to find additional large oil fields. Therefore, the
agement; summarizes the applications of gradient-based demand for exploiting the limited oil reserves efficiently
algorithms, gradient-free algorithms, and artificial intelli- and economically becomes increasingly significant and has
gence algorithms; analyzes the characteristics and appli- attracted more global attention in recent years. To achieve
cation conditions of these optimization methods; and this goal, an important technique proposed is closed-loop
finally discusses the emphases and directions of future reservoir management. It consists of two steps: automatic
research on both automatic history matching and reservoir history matching and reservoir production optimization.
production optimization. Automatic history matching is a sequential model
updating method, where the estimate of uncertain reservoir
properties is updated continuously according to the pro-
J. Hou duction measurements available at the time. Reservoir
State Key Laboratory of Heavy Oil Processing, China University production optimization is a complete or partial automation
of Petroleum, Qingdao 266580, Shandong, China
process for maximizing the development effect within the
J. Hou (&)  K. Zhou (&) lifecycle of a reservoir by optimizing operational parame-
School of Petroleum Engineering, China University of ters. The main idea is to exploit the oil reserves as near to
Petroleum, Qingdao 266580, Shandong, China the desired optimum as possible. Both automatic history
e-mail: houjian@upc.edu.cn
matching and reservoir production optimization are opti-
K. Zhou mization problems as mentioned by researchers such as
e-mail: zhoukang_upc@163.com
Brouwer and Jansen (2004), Sarma et al. (2005) and Wang
X.-S. Zhang  X.-D. Kang et al. (2009). These problems can be solved by optimiza-
CNOOC Research Institute, Beijing 100027, China tion theories.
Automatic history matching has been studied since the
H. Xie
ZTE Corporation, Shenzhen 518055, Guangdong, China
1960s (Jahns 1966; Wasserman and Emanuel 1976; Yang
and Watson 1988), but it is still a very difficult problem at
Edited by Yan-Hua Sun present. The existing history matching methods can be

123
Pet. Sci. (2015) 12:114–128 115

broken into two categories. One category is based on gra- 2 Problem descriptions
dient, including finite difference approximation of deriva-
tive, adjoint gradient-based methods, and gradient 2.1 Automatic history matching
simulator methods. The other category is based on gradi-
ent-free optimization, such as simultaneous perturbation The estimate of unknown geological properties using pro-
stochastic approximation, genetic algorithm, particle duction measurements is recognized as history matching. It
swarm optimization, and pattern search methods (PSMs). is an ill-posed inverse problem with many unknown res-
Oliver and Chen (2011) have summarized the recent pro- ervoir parameters that could be adjusted to obtain a match
gress on automatic history matching. against a relatively smaller number of measurements.
The origin of solving reservoir production problems Traditionally, the unknown parameters are adjusted man-
using optimization theories can be traced back to Lee and ually by trial and error. This method is time-consuming and
Aronofsky (1958). They used a linear programming often yields a reservoir numerical model which may be
method to maximize the net present value of production for unrealistic or not consistent with geological properties. To
a homogeneous reservoir. Later, some other papers address these problems, automatic history matching has
appeared in journals such as Operation Research, Man- been studied for several decades. As shown in Fig. 1,
agement Science, and Journal of Petroleum Technology. automatic history matching is an iterative procedure where
However, most papers published before the 1980s did not the unknown reservoir parameters are adjusted automati-
pay enough attention to optimization algorithms, and suc- cally with an optimizer to match the observed production
cessful applications were very rare (Aronofsky and Wil- or pressure data. In fact, automatic history matching is an
liams 1962; Wattenbarger 1970; McFarland et al. 1984). optimization problem and the most commonly used
With the advances in optimization algorithms and com- objective can be written as follows:

K h
X i
OðmÞ ¼ ðgk  dk ÞT ðCD Þ1 ^ k ÞT ðCM Þ1
k ðgk  dk Þ þ ðmk  m k ðmk  m
^kÞ ; ð1Þ
k¼1

puting power, research has increased greatly since the where the subscript k is the discrete time step; K is the total
1980s (Sequeira et al. 2002; Chacón et al. 2004; Barragán number of time steps; m is the vector of reservoir param-
et al. 2005; Gunnerud and Foss 2010; Knudsen and Foss eters to be estimated; d is the vector of observed historical
2013; Tavallali et al. 2013). data; g is the vector of corresponding data to be matched;
In order to find out under which operational parameters
at current reservoir conditions the oil production might be Set inial model
most efficient and profitable, automatic history matching
and reservoir production optimization should be combined
together. This combination forms a concept of closed-loop Perform simulaon
reservoir management, where the geological model will be
updated once the production measurements are available
Calculate misfit Producon data
and the operational parameters will be optimized based on
the newly updated reservoir model. Representative papers
on this concept include Brouwer and Jansen (2004), Jansen Yes
et al. (2005), Nævdal et al. (2006), and Bieker et al. (2007). Sasfactory ? Reservoir model
This paper gives a general review of research on auto- No
matic history matching and reservoir production optimi-
zation; analyzes the characteristics and application Opmizaon calculaons
Perform opmizaon Stop
conditions of gradient-based algorithms, gradient-free
algorithms, and artificial intelligence algorithms; and
finally discusses the emphases and directions of future Update model
research on both automatic history matching and reservoir
production optimization. Fig. 1 Flow chart for automatic history matching

123
116 Pet. Sci. (2015) 12:114–128

m^ is the prior model information; CD is the covariance history matching results are good enough, but the
matrix of measurement errors; CM is the covariance matrix assumptions on uncertainty are definitely different.
of prior probability density function. These matrices More and more attention has been paid to the generation
determine the weight of individual terms in the objective and history matching of geologically realistic non-Gaussian
function. reservoir models. Researchers have presented several
Although the least-square error has been used success- methods including Gaussian mixture models, truncated
fully to match the observed production data, it does not work pluri-Gaussian method, level set method, discrete cosine
well when the seismic data are history matched at the same transform, and other principal component analysis methods.
time. To address this problem, Tillier et al. (2013) presented Dovera and Rossa (2011) proposed expressions of condi-
an appropriate objective function for history matching of tional means, covariances, and weights for Gaussian mixture
seismic attributes based on image segmentation and a models, so that the ensemble Kalman filter algorithm
modified Hausdorff metric. The objective function of history (EnKF) became usable in this case. Liu and Oliver (2005a,
matching commonly has a complex shape and multiple local b) combined EnKF with a truncated pluri-Gaussian method
minima (Oliver and Chen 2011). This is mainly because for history matching of reservoir facies. Agbalaka and Oli-
unknown parameters are always much more numerous than ver (2008) extended this method to a three-dimensional (3D)
available production measurements. reservoir case. Chang et al. (2010) proposed a methodology
In order to carry out history matching in a lower space and to combine a level set method with EnKF for history
reduce the necessity of an explicit regularization term in the matching of facies distribution in a 2D reservoir model. Hu
objective function, various parameterization methods have et al. (2013) introduced a new method to update complex
been presented, including the zonation method, pilot point facies models generated by multipoint simulation while
method, subspace method, spectral decomposition, discrete preserving their geological and statistical consistency. Ja-
cosine transform, truncated singular value decomposition, farpour and McLaughlin (2008) tested a discrete cosine
and the multiscale method. Jacquard (1965) and Jahns (1966) transform method on two 2D, two-phase reservoir models.
applied the zonation method to reduce variables in their Other principal component analysis methods can also be
automatic history matching study. The gradzone method in used to deal with non-Gaussian reservoir models. However,
Bissell et al. (1994) and the adaptive multiscale method in their computational cost may be high.
Grimstad et al. (2003, 2004) were variations of the original During the development of automatic history matching,
zonation method. Marsily et al. (1984) presented the pilot various optimization algorithms have been introduced and
point method, in which variables were estimated only at the modified (Bissell et al. 1994; Lee and Seinfeld 1987; Go-
pilot points and others were calculated by the Kriging method. mez et al. 2001). Streamline-based techniques were also
Abacioglu et al. (2001) adopted a subspace method based on used to improve computational efficiency of history
segmentation of the objective function. Oliver (1996a, b) matching by researchers such as Agarwal and Blunt (2003),
applied the parameterization method based on spectral Cheng et al. (2004, 2005), and Gupta and King (2007).
decomposition of the prior covariance matrix to history match With streamline-based techniques, a reservoir simulation
a two-dimensional (2D) permeability field. Jafarpour and model was automatically decoupled into a series of one-
McLaughlin (2008), Jafarpour et al. (2010) used the discrete dimensional models along streamlines, which could mini-
cosine transform as a parameterization method in their history mize numerical dispersion and the effects of grid genera-
matching study. Tavakoli and Reynolds (2010, 2011) pro- tion while maintaining a sharp displacement front. Caers
vided a theoretical basis for parameterization based on trun- (2003) and Negrete et al. (2008) combined streamline
cated singular value decomposition of the dimensionless simulators with a deformation method and EnKF, respec-
sensitivity matrix. He et al. (2013) applied the orthogonal tively, to carry out automatic history matching. These
decomposition method to transform the high-dimensional streamline-based history matching techniques inherit the
states into a low-dimensional subspace. They also described shortcomings of streamline methods such as inability to
geological models in reduced terms by the Karhunen–Loève model very complex physics at the same time. Therefore,
expansion of the log-transmissibility field. they are not suitable to all history matching cases.
The parameterization methods should be performed In order to characterize the uncertainty of unknown
carefully, or results may sometimes mislead us. For geological properties, researchers introduced a Bayesian
example, Oliver et al. (2008) showed a small number of framework, with which one can formally construct a pos-
variables may underestimate the uncertainty in automatic terior density function. The books of Tarantola (2005) and
history matching problem. They adjusted a spatially vary- Oliver et al. (2008) provide a detailed description about
ing porosity with a uniform permeability or adjusted a Bayesian framework. Generally, Bayesian estimation
spatially varying permeability with a uniform porosity to depends on a prior Gaussian model. To address this limi-
history match a same set of well-test data. Both of the two tation, Sarma et al. (2008b) transformed this problem into

123
Pet. Sci. (2015) 12:114–128 117

feature space using kernel principal component analysis.

Set initial control
Other important research works involved Bayesian frame-
work include Chu et al. (1995), He et al. (1997), Zhang
et al. (2002, 2005), Liu and Oliver (2003), Oliver et al. Perform simulation
(2008), and Emerick and Reynolds (2012).
Although numerous papers on automatic history
matching have been published, most of them have con- Calculate NPV
centrated on a limited type of estimated parameters such as
permeability and porosity. They are impractical at present
for industrial field application in which we have to calibrate Yes
Optimal ? Optimal control
a myriad of other discrete and continuous parameters,
including fluid contacts, rock compressibility, and relative No
permeability. However, automatic history matching is a
meaningful research direction and much more effort will be Optimization
Perform optimization
calculations Stop
necessary in the future.

2.2 Reservoir production optimization Adjust control

Production optimization aims at achieving the best devel- Fig. 2 Flow chart for reservoir production optimization
opment performance for a given reservoir by optimizing
well controls. Figure 2 shows the schematic for this opti-
where J is the net present value and is a function of the
mization process. In order to evaluate the performance of
control vector u; L is the total number of simulation time
different development programs, various objectives have
steps; NP is the total number of producers; NI is the total
been proposed during the long research into production
number of injectors; ro is the oil price; rw is the produced
optimization. For example, Rosenwald and Green (1974)
water treatment cost; rwi is the water injection cost; qno;j and
minimized the difference between the production-demand
curve and the flow curve actually attained. Babayev (1975) qnw;j are the average oil and water production rates of the jth
provided minimum total cost per unit output. Lasdon et al. producer during the nth simulation time step, respectively;
(1986) maximized the deliverability of a gas reservoir at a qnwi;i is the average water injection rate of the ith water
specified time, minimized the total gas withdrawal shortfall injection well during the nth simulation time step; ic is the
between the demand schedule and the amount of gas that annual discount rate; Dtn is the length of the nth simulation
can actually be delivered in each month, and they also time step; tn is the cumulative time up to the nth simulation
optimized the weighted combinations of the above two time step in years.
objectives. When optimizing production strategies, one According to the general form of optimal control
often encounters multiple local maxima. This phenomenon problems, a mathematical model for reservoir production
may be a good thing sometimes because it means that there optimization can be written as
are extra degrees of freedom in the optimization problem, " #
X
L NP 
X  X
NI
Dtn
which can be used to accomplish other optimization max JðuÞ ¼ ro qno;j  rw qnw;j  rwi qnwi;i n
n¼1 j¼1 i¼1 ð1 þ ic Þt
objectives. For instance, Van Essen et al. (2011) incorpo-
rated short-term goals into the life-cycle optimization ð3Þ
problem and proposed a hierarchical production optimiza- Au  b ð4Þ
tion structure with multiple objectives. Chen et al. (2012)
also optimized both long-term and short-term net present ulow  u  uup ; ð5Þ
value. As more and more oilfields enter the high water cut where Eq. (4) represents the linear or nonlinear constraints;
period, the production costs increase gradually. Therefore, Eq. (5) gives the boundary constraints.
the net present value is commonly selected as the objective Asheim (1988) maximized the net present value for
function for production optimization. In terms of water waterflooding with multiple vertical injectors and a vertical
flooding projects, it is defined as producer by optimizing rate allocation based on the product
" #
XL XNP   X NI of permeability and thickness. Brouwer and Jansen (2004)
Dtn
JðuÞ ¼ ro qno;j  rw qnw;j  rwi qnwi;i n ; studied static and dynamic waterflooding optimization. For
n¼1 j¼1 i¼1 ð1 þ ic Þt
the static one, they kept inflow control valves constant
ð2Þ during the displacement process until water breakthrough.

123
118 Pet. Sci. (2015) 12:114–128

For the dynamic one, they applied gradients calculated production. Codas et al. (2012) integrate simplified well
with an adjoint method to dynamically optimize the pro- deliverability models, vertical lift performance relations,
duction performance and considered a simple constraint and flowing pressure behavior of the surface gathering sys-
where the total injection was equal to the total production. tem to develop a framework of integrated production opti-
In order to increase the displacement efficiency, Sudary- mization for complex oil fields.
anto and Yortsos (2000) optimized the front shape of The above review of production optimization showed
injected fluid by controlling injection rates. Results showed that the exploitation method involved has extended from
that the waterflooding optimization was a ‘‘bang–bang’’ waterflooding to EOR methods, the control variables
control problem, where each control variable took either its optimized have extended from simple to complex opera-
minimum or maximum allowed values. Zandvliet et al. tional parameters, and that the optimization algorithms
(2007) further investigated why and under what conditions applied have extended from gradient-based to gradient-free
waterflooding problems had optimal solutions under bang– methods. In addition, production optimization has
bang control. They concluded that waterflooding optimi- improved from an open adjustment process to a closed-
zation with simple boundary constraints sometimes had loop management workflow (Mochizuki et al. 2006), which
bang–bang optimal solutions, while problems with other will be discussed in detail in the next section.
general inequality or equality constraints would have a
smooth optimal solution. Gao and Reynolds (2006) pro-
posed a log-transformation method to deal with boundary 2.3 Closed-loop reservoir management
constraints. Alhuthali et al. (2007) also achieved optimal
waterflooding management using rate control. As more and more production and pressure measurements
Many pilot tests and commercial projects using enhanced become available, the reservoir model can be updated to
oil recovery (EOR) methods have been performed during the achieve a better estimate of the unknown geological
past few decades to improve the development effect of properties. Then the operational parameters should be
waterflooding. The challenges of huge investment, high cost, optimized again based on the newly updated reservoir
and high risk promote research into production optimization model. This cycle of model update and production opti-
for EOR methods. As early as in 1972, Gottfried (1972) mization is repeated during the whole process of the res-
proposed a nonlinear programming model for a cyclic steam ervoir development. This forms the concept of closed-loop
injection process. He maximized the net present value by reservoir management. The schematic of this closed-loop
optimizing steam injection volume and cycle length. Ra- process is shown in Fig. 3.
mirez et al. (1984) and Fathi and Ramirez (1984) tried to Nævdal et al. (2006) applied the EnKF method to update
maximize oil production at the minimum injection costs the reservoir simulation model and then optimized opera-
based on the calculus of variations and Pontryagin’s weak tional parameters with an adjoint formulation in order to
minimum principle. They optimized development strategies maximize the economic profits. Sarma et al. (2005) pre-
for waterflooding, carbon dioxide (CO2) flooding, and sur- sented a closed-loop management approach for efficient
factant flooding. Amit (1986) formulated a two-phase real-time production optimization. Their approach con-
dynamic optimization model which incorporated the rela- sisted of three key elements: adjoint models for gradient
tionships between extraction rates, investment decisions, calculations, polynomial chaos expansions for uncertainty
and cumulative oil recovery. Wackowski et al. (1992) propagation, Karhunen–Loeve expansions and Bayesian
applied rigorous decision analysis methodology to find the inversion theory for history matching. Results showed that
optimal development strategy for a 20 years CO2 flooding their approach increased the net present value by 25 %. For
project. The control variables included CO2 recycle capac- a similar problem, Saputelli et al. (2005) proposed a model
ity, CO2 purchase contract, processing rate, water-to-gas predictive control method. It was a class of computer
ratio, and slug size. Wu (1996) found that chemical flooding control algorithms that explicitly used a plant model for
performance was sensitive to operational parameters such as online prediction of future behavior, and computation of
chemical slug size, concentration and adsorption, price of oil appropriate control action subjected to various constraints
and chemicals, annual discount rate, and reservoir perme- through online optimization of a cost objective. The
ability. Results showed that the optimal design was a large method addressed the overwhelming complexity of overall
slug injection of low concentration surfactant and polymer, optimization problems by suggesting an oilfield operations
followed by a small slug of subsequent polymer injection. hierarchy which entailed different time scales. Due to the
One of the best examples about gas lift optimization has been rapid development of ensemble-based optimization algo-
achieved by McKie et al. (2001), in which gas lift injection rithms, Chen et al. (2009a, b) provided a new closed-loop
rates, compressor settings, and field fuel consumption for reservoir management method which integrated an
more than 500 wells were optimized to maximize liquid ensemble-based optimization method with the EnKF

123
Pet. Sci. (2015) 12:114–128 119

Fig. 3 Schematic of closed-

loop reservoir management Optimal control Reservoir system Production data

Yes

Production optimization
Optimal? Calculate NPV

No Perform simulation

Adjust control Reservoir model

Automatic history matching

Yes

Satisfactory? Calculate misfit

No Perform simulation

Perform optimization Update model

algorithm. Jansen et al. (2008) and Jansen (2011) discussed provided by ensemble members. Second, the objective
an emerging technique to increase oil recovery. Their function was the expectation of each ensemble member.
technique is an operational use of model-based optimiza- Therefore, the ensemble-based optimization method was
tion which requires a combination of long-term and short- fairly robust with respect to the uncertainty of the estimated
term objectives through multi-level optimization strategies. geological models.
Moridis et al. (2013) established a self-teaching expert With the rapid increase in research interests and great
system to increase oil production by improving flooding improvements of optimization techniques, the original
efficiency and reducing geological uncertainty. Brugge test case shows its weakness in the low frequency
When reviewing closed-loop reservoir management, it is of the feedback loop. In order to keep the Brugge test case
necessary to present the Brugge test case which was pre- as a challenging problem for testing and comparing dif-
pared for SPE (Society of Petroleum Engineers) Applied ferent techniques in closed-loop reservoir management,
Technology Workshop held in Brugge in June 2008. In the Peters et al. (2013) provided a lot of additional data
test case, well-log data, reservoir structure, 10 years’ pro- including well constraints and production history from
duction data, inverted time-lapse seismic data, and other individual well completions for another 20 years, as well as
information necessary are given by TNO (Netherlands the updated data of oil saturation and reservoir pressure.
Organisation for Applied Science Research) to estimate
unknown parameters such as permeability, porosity, and
net-to-gross thickness. After history-matched reservoir 3 Optimization methods
models were created, water flooding strategies for 20 pro-
ducers and 10 injectors were optimized. Peters et al. (2010) According to the techniques for determining search direc-
have summarized in detail the results of the Brugge test tion and step size, the optimization methods can be clas-
case obtained by nine research groups. Briefly, Table 1 sified into three categories: gradient-based algorithms,
compares the reservoir simulators, optimization methods, gradient-free algorithms, and artificial intelligence algo-
and the net present values optimized in Year 10. In the rithms. In fact, artificial intelligence algorithms are also
table, the net present value was obtained using the optimal independent of gradient information. This paper considers
strategy of each participant in the Brugge test model. As them as an independent category because almost all of
can be seen, three participants who applied ensemble-based them are inspired by intelligent behaviors in nature such as
methods in history matching step achieved a similar inference, designing, thinking, and learning.
highest net present value, although their optimization
algorithms used in the production optimization step were 3.1 Gradient-based algorithms
different. This is mainly because ensemble-based optimi-
zation methods had two distinct advantages. First, the The calculation of derivatives or the Hessian matrix is the
search direction was approximated through the correlations key to solving optimization problems using gradient-based

123
120 Pet. Sci. (2015) 12:114–128

Table 1 Research results of the Brugge test case

Participant Simulator History matching method Production optimization method Net present
value, 109 $

Halliburton Nexus Landmark’s DMSTM Scatter/tabu search method 3.53a

International Research ECLIPSE 100 Ensemble Kalman filter Ensemble Kalman filter 4.41
Institute of Stavanger
University of Oklahoma/ ECLIPSE 100 Randomized maximum Ensemble-based gradient method 4.42
Chevron likelihood method
Roxar/Energy Scitech Tempest MORE EnABLETM Sequential experimental design method 4.03
Shell International MoReS Ensemble Kalman filter Adjoint-based gradient method 4.12
Exploration and
Production BV
Schlumberger ECLIPSE 100 Selection of realizations based Artificial neural networks 4.10
on their fit to the data
Stanford University/ GPRS; Sequential quadratic Adjoint-based gradient method 4.26
Chevron ECLIPSE 100 programming; Hooke-Jeeves
direct search method
Texas A&M University FrontSim Streamline-based generalized Sequential quadratic programming 4.22
travel time inversion
University of Tulsa ECLIPSE Ensemble Kalman filter Adjoint-based gradient method 4.47
100/300
a
The result was achieved using one control interval per well, while others were obtained using three control intervals per well

Table 2 Comparison of calculation methods for gradients and Hessian matrix

Methods Calculation principles Characteristics

Numerical Small perturbations of the model parameters and calculation of Easy to implement; expensive computational cost;
perturbation the production responses unsuitable for large-scale optimization problems
Sensitivity Differentiation of the flow and transport equations Difficult to obtain analytical expressions for nonlinear
equation optimization problems
Adjoint method Optimal control theories and calculus of variations Easy to implement; dependent on reservoir simulators; hard
to transplant elsewhere

algorithms. In terms of production optimization and history sensitivity matrix or its transpose. Now, the adjoint method
matching, researchers have applied various calculation has become one of the most efficient methods existing
methods including numerical perturbation, sensitivity today to compute gradients for gradient-based algorithms.
equation, and adjoint method, as summarized in Table 2. Generally, gradient-based algorithms can be classified
The roots of the adjoint method can be found in Jac- into two categories. One is the first-order methods, which
quard (1965). Later, Carter et al. (1974) formulated their only require the derivative information. For example, the
work in a better way using Frechet derivatives. He et al. steepest ascent algorithm and the conjugate gradient
(1997) further extended Carter’s work to three dimensions method have been widely used by many researchers such
approximately. For single-phase flow problems, Chen et al. as Brouwer and Jansen (2004), Sarma et al. (2008a), and
(1974) and Chavent et al. (1975) proposed a method which Wang et al. (2009). The other category is the second-order
was regarded as what we call the adjoint method now. Li methods, which not only require the derivative information
et al. (2003) presented the first formulation of the adjoint but also require the Hessian matrix. Representative meth-
method for three-phase flow problems and pointed out that ods include Gauss–Newton, Levenberg–Marquardt,
the coefficient matrix of the adjoint equations is simply the sequential quadratic programming (Barnes et al. 2007), and
transpose of the Newton–Raphson Jacobian matrix used in the limited memory Broyden Fletcher Goldfarb Shanno
a fully implicit reservoir simulator. Therefore, the deriva- method (LBFGS).
tion of the individual adjoint equation can be avoided by When using the Gauss–Newton method for automatic
extracting and saving the Jacobian matrices. Rodrigues history matching problems, Wu et al. (1999) introduced an
(2006) derived the adjoint equations in a much neater way, artificially high variance of measurement errors at early
and provided a method for multiplication of a vector by the iterations to damp the changes in model parameters and

123
Pet. Sci. (2015) 12:114–128 121

thus avoid undershooting or overshooting. Tan and Kalo- unknown parameters stochastically and simultaneously to
gerakis (1991) pointed out that the standard Gauss–Newton generate a search direction at each iteration. The expecta-
and Levenberg–Marquardt methods require the computa- tion of stochastic SPSA gradients is true gradient and it is
tion of all sensitivity coefficients in order to formulate the always a downhill direction. Therefore, the SPSA method is
Hessian matrix, which seems impossible in reality due to a stochastic version of the steepest descent algorithm. Based
the large number of unknown parameters relative to limited on the simultaneous perturbation idea, Spall further pro-
available measurements. In order to eliminate this problem, vided a second-order SPSA method, which estimates the
the quasi-Newton method was introduced by researchers. Hessian matrix at each iteration. Later Bangerth et al.
This method only requires the gradient of the objective (2006) described an integer SPSA method and used this
function which can be computed from a single adjoint modified SPSA method to solve well placement optimiza-
solution as done in Zhang et al. (2002). In order to further tion problems. To the best of our knowledge, this is the first
improve the computational efficiency and robustness of the time that SPSA has been used in optimal control problems.
LBFGS method, Gao and Reynolds (2006) proposed a new In terms of closed-loop reservoir management, Wang et al.
line search strategy, rescaled the model parameters, and (2009) discussed the application of SPSA in the step of
applied damping factors to the production data. They also production optimization. Gao et al. (2007) applied a mod-
noticed that the new line search strategy had to satisfy the ified SPSA method for automatic history matching. In their
strong Wolfe conditions at each iteration, or the conver- study, the approximate Hessian matrix was calculated using
gence rate would decrease significantly. the inverse of the covariance matrix of the prior model.
The Karhunen–Loeve expansion can create a differen- Results showed that the modified SPSA performed almost
tiable parameterization of the numerical model in terms of as well as the steepest descent method. Of course, gradient-
a small set of independent random variables and deter- based algorithms like LBFGS would be preferred when the
ministic eigenfunctions. With this expansion, the gradient- gradient can be calculated. Otherwise, the SPSA method
based algorithms can be applied while honoring the two- may be a good choice. Based on SPSA, Li and reynolds
point statistics of the geological models (Gavalas et al. (2011) proposed a stochastic Gaussian search discretion
1976). In order to further extend the existing gradient- algorithm for history matching problems, and this modified
based history matching techniques to deal with complex method was successfully used in the well-known PUNQ-S3
geological models characterized by multiple-point geosta- (Production forecasting with uncertainty quantification) test
tistics, Sarma et al. (2007) applied a kernel principal case. Zhou et al. (2013) integrate the finite difference
component analysis method to model permeability fields. method and the SPSA method to optimize polymer flooding
This method can preserve arbitrary high-order statistics of in a heterogeneous reservoir. But currently the SPSA
random fields, and it is able to reproduce complex geology method has not been widely used for reservoir optimization
while retaining reasonable computational requirements. problems.
Gradient-based algorithms are widely used in research The EnKF algorithm was first proposed by Evensen in
on production optimization and history matching because 1994 as a Monte Carlo approximation of the Kalman filter
of their high computational efficiency and fast convergence in the ocean dynamics literature. This method obtains gra-
behavior. However, these algorithms require detailed dient information through correlations of ensemble mem-
knowledge of the numerical simulators. They can hardly be bers. Nævdal et al. (2002) applied EnKF to estimate near-
used without adjoint code and they are difficult to trans- well permeabilities. Gu and Oliver (2005) examined EnKF
form from one simulator to another. for combined parameter and state estimation in a stan-
dardized reservoir test case. Gao et al. (2006) found that
3.2 Gradient-free algorithms EnKF and the randomized maximum likelihood method
(RML) gave similar computational results. Reynolds et al.
In order to make full use of the advantages of commercial (2006) further presented their mathematical connections.
reservoir simulators when conducting production optimi- They also showed that EnKF may be viewed as updating
zation and automatic history matching, researchers have each ensemble member with a single Gauss–Newton iter-
introduced many gradient-free algorithms including ation. Liu and Oliver (2004, 2005a, b) investigated a highly
simultaneous perturbation stochastic approximation algo- nonlinear problem of facies estimation using the EnKF
rithm (SPSA), EnKF, PSM, new unconstrained optimiza- method. When using the EnKF method, Lorentzen et al.
tion algorithm (NEWUOA), and quadratic interpolation (2005) discussed the choice of initial ensemble members,
model-based algorithm guided by approximated gradient while Wen and Chen (2006) focused on the effect of
(QIM-AG). ensemble size. Emerick and Reynolds (2012, 2013) incor-
Spall (1998) proposed the SPSA method based on the porated automatic history matching in an integrated geo-
Kiefer–Wolfowitz algorithm. This new method perturbs all modeling workflow using the ensemble smoother method.

123
122 Pet. Sci. (2015) 12:114–128

Although it is not long since the introduction of EnKF into control variables) objective evaluations are needed to build
petroleum engineering, this method is developing rapidly the initial quadratic model before the first optimization
and shows huge potential for solving history matching procedure can be achieved. So it can hardly be used when
problems because it can deal with the uncertainties. the number of control variables is very large. In order to
However, the toy problem in Zafari and Reynolds improve the computational efficiency of NEWUOA, Zhao
(2007) showed that the standard EnKF cannot handle et al. (2011) developed a QIM-AG method. This method
multimodal nonlinear problems. To address this limitation, requires a minimum of only one interpolation point to build
Gu and Oliver (2007) applied an iterative EnKF to a quadratic model at each iteration. It is similar to the
assimilate multiphase flow measurements. Li and Reynolds quasi-Newton method and converges very fast.
(2009) provided two iterative EnKF procedures, with Taking the net present value as objective function, Zhao
which they obtained better history matching results than et al. (2011) compared the performance of various gradi-
that with the standard EnKF method in two examples ent-free algorithms. As shown in Fig. 4, the quadratic
including the toy problem. Lorentzen and Nævdal (2011) interpolation model-based algorithm guided by ensemble-
and Wang and Li (2011) also introduced an iterative based gradient (QIM-EnOpt) obtained the highest net
extension of EnKF in order to improve estimate results in present value. The quadratic interpolation model-based
cases where the relationship between the model and the algorithm guided by SPSA gradient (QIM-SPSA) obtained
observations is nonlinear. Agbalaka et al. (2013) proposed a similar net present value but converges much faster than
a two-stage ensemble-based technique to improve the QIM-EnOpt and EnOpt. The SPSA and NEWUOA meth-
performance of EnKF for history matching with multiple ods achieved similar values, but the computational effi-
modes. Chen et al. (2009a, b) addressed non-Gaussian ciency of NEWUOA was much lower at the beginning of
effects through a change in parameterization. They dem- the optimization process. In this test case, the particle
onstrated the effectiveness of the combination of their swarm optimization method and SID-PSM algorithm per-
methods with the traditional EnKF by history matching of formed much worse than the other algorithms.
multiphase flow in a heterogeneous reservoir. Supposing
that both of the multimodal priors and posteriors can be 3.3 Artificial intelligence algorithms
approximated by Gaussian mixture models, Dovera and
Rossa (2011) presented a modified EnKF method espe- Artificial intelligence algorithms have been used to solve
cially for multimodal systems. Heidari et al. (2013) com- production optimization and history matching problems for
bined EnKF with pilot points and gradual deformation to a long time. The representative methods include the sim-
preserve second-order statistical properties, so the depar- ulated annealing algorithm, genetic algorithm, artificial
ture of constrained petrophysical properties from prior neural networks, particle swarm optimization method, and
information could be greatly reduced. the tabu search method.
Custsódio and Vicente (2007) provided a modified PSM The simulated annealing algorithm was first introduced
guided by simplex derivatives (SID-PSM). This new method into reservoir engineering. Farmer (1992) applied this
has two modifications over the original PSM algorithm. algorithm to generate rock models with two-point geosta-
First, the predefined search directions are ranked using the tistics properties. Qian (1993) introduced Markov random
simplex gradient during the poll step and the search direc- field theory into the basic simulated annealing algorithm
tions closest to the simplex gradient are tried first. Second,
when there are enough objective evaluations, a quadratic 6.5
interpolation model is built in the search step and then it is
6.0
minimized by a trust-region method. The SID-PSM algo-
rithm has been introduced into reservoir engineering for 5.5

optimizing the settlement and adsorption parameters of as-

NPV, 107$

5.0 QIM -SPSA

phaltene in porous media. This method can converge to a QIM -EnOpt
global optimum using only the objective evaluations. 4.5 NEWUOA
However, its global convergence is very sensitive to the 4.0
SID -PSM

choice of initial values. Therefore, close attention should be PSO

3.5 SPSA
paid when using SID-PSM for practical problems.
EnOpt
For unconstrained optimization problems without 3.0
0 200 400 600 800 1000
derivatives or a Hessian matrix available, Powell (2008)
Simulation runs
proposed a NEWUOA method. It is a quadratic model-
based gradient-free trust-region algorithm based on qua- Fig. 4 Comparison of various gradient-free algorithms by Zhao et al.
dratic interpolation. At least Nu ? 2 (Nu is the number of (2011)

123
Pet. Sci. (2015) 12:114–128 123

and thus transformed it into a probabilistic uphill algo- 4 Future developments

rithm. When solving automatic history matching problems,
Ouenes et al. (1993) applied a simulated annealing algo- In order to solve the growing demand–supply gap, many
rithm directly while Carter and Romero (2002) combined it studies have been carried out in terms of closed-loop res-
with other techniques such as geostatistics, a pilot point ervoir management. However, most of the studies are
method, and a genetic algorithm. The convergence of the focused on academic research and they are far from large-
simulated annealing algorithm is sensitive to the choice of scale field application. As for field application, oilfield
initial temperature and reduction factor. If the reduction engineers are mainly concerned about three issues: pro-
factor is too large, many extreme points will be missed. cessing time, decision veracity, and operating convenience.
However, if the reduction factor is too small, the simulated In order to reduce the processing time, efficient optimiza-
annealing algorithm will converge very slowly. tion methods and parallel distributed systems should be
The genetic algorithm was proposed by Holland (1975) paid more attention. It is important for increasing decision
based on the idea of population evolution. This method has veracity to carry out more uncertainty analyses. Since most
been used as a global optimizer for automatic history of the oilfield engineers are not familiar with the theoretical
matching problems by Tokuda et al. (2004). Compared bases of closed-loop reservoir management, integrated
with a simulated annealing algorithm, a genetic algorithm software is necessary to improve their operating
starts with a population of realizations rather than a single convenience.
realization. However, there is no mathematical guarantee
for a genetic algorithm to reach a global optimum. In spite 4.1 Hybrid solvers
of these drawbacks, genetic algorithms still have much
appeal because they can be easily parallelized and con- Various optimization methods have been introduced into
nected to any existing reservoir simulators. reservoir development engineering during the past decades.
Artificial neutral networks are a rough approximation Each algorithm has its own advantages as well as draw-
and simplified simulation of a biological neuron network backs. For example, the gradient-based algorithms have
system. Generally, the neutral networks are used to estab- high computational efficiency and fast convergence rate,
lish proxy models. For this purpose, a small number of but the calculation of gradients requires detailed knowl-
simulations are run at first to build a nonlinear relationship edge of the numerical simulators. They can hardly be used
between the objective and unknown variables. Then the without adjoint code and they are difficult to transform
relationship can be used for reservoir production optimi- from one simulator to another. The gradient-free algo-
zation in order to reduce the computational cost (Elkamel rithms can converge to global optimum with a high prob-
1998; Saputelli et al. 2002). Artificial neutral networks ability and they can make full use of the technical
were also used directly as an optimizer by Ramgulam et al. advantages of commercial reservoir simulators. However,
(2007) in their studies of history matching. Although arti- the computational efficiency and convergence rate of gra-
ficial neutral networks have been widely used, the proce- dient-free algorithms are not satisfactory. Therefore, we
dures for network construction are still not established. should combine different algorithms together to develop
The particle swarm optimization method was proposed new hybrid solvers, which can incorporate the advantages
by Kennedy and Eberhart based on social behavior observed of different optimization methods.
in nature. It has been successfully used in well placement
optimization (Onwunalu and Durlofsky 2009), automatic
history matching, and water flooding optimization. This 4.2 Parallel algorithms and distributed systems
method can find the global optimum with a high probability
and can be connected to any numerical simulators. As well known, numerical simulations are necessary in
In fact, artificial intelligence algorithms are stochastic closed-loop reservoir management. However, for large-
gradient-free methods. These algorithms need much more scale reservoir optimization problems, even a single
numerical simulations than the deterministic methods to numerical simulation requires several days and the total
converge. In addition, the input parameters for imple- computational cost is unbearably expensive. To address
menting gradient-free algorithms are commonly deter- this problem, it is necessary to develop parallel algorithms
mined by the trial and error method. Therefore, the and distributed systems. The parallel algorithms should
gradient-free algorithms should be modified and integrated necessarily conserve the mathematical convergence
with other optimizers in practical applications. behavior of the original algorithms. Traditionally, the dis-
Based on the discussions in above three sections, the tributed systems can be classified into two categories. The
main characteristics for different kinds of optimization first category is homogeneous distributed system which is a
algorithms are summarized in Table 3. cluster of many similar computers or work stations. The

123
124 Pet. Sci. (2015) 12:114–128

Table 3 Characteristics of different optimization algorithms

Methods Representive algorithms Characteristics

Gradient-based Steepest ascent algorithm; conjugate gradient method; LBFGS High computational efficiency; dependent on adjoint
algorithms method; Levenberg–Marquardt algorithm; Gauss–Newton gradient which generally needs detailed knowledge of
method simulator numerics to implement; difficult to transform
from one simulator to another
Gradient-free SPSA; EnOpt; EnKF; SID-PSM; NEWUOA; QIM-AG Independent of detailed knowledge of simulator
algorithms numerics; capable of connecting to any reservoir
simulators; implement easily
Artificial Simulated annealing algorithm; genetic algorithm; artificial neural Low computational efficiency; independent of gradient
intelligence networks; particle swarm optimization algorithm or Hessian matrix information; expensive computation
algorithms cost; no guarantee of convergence

other category is heterogeneous distributed system which is 4.4 Integrated software

constructed by a set of work stations or computers with
different architectures. Currently, with the rapid develop- In order to exploit the limited oil reserves more efficiently
ment of multicore computers and graphic processing unit and economically, the field application of closed-loop
(GPU) technology, many parallel programming methods reservoir management becomes increasingly significant.
have been proposed on the basis of Open Multiprocessing However, most of the oilfield engineers do not know much
(OpenMp) and compute unified device architecture about its theoretical basis. Therefore, it is necessary to
(CUDA). For example, Wu et al. (2014) discussed a mul- develop integrated software which can be used conve-
tilevel preconditioner in a new-generation reservoir simu- niently as a black box. As far as we know, some com-
lator and its implementation on multicore computers using mercial simulators have done much work in this field. For
OpenMP. In order to conduct large-scale closed-loop res- example, the Computer Modelling Group Ltd has devel-
ervoir management, parallel algorithms and distributed oped a CMOST studio, which integrated history matching,
systems should be emphasized in the future studies. optimization, sensitivity analysis, and uncertainty assess-
ment tool. Some optimization methods such as the Latin
4.3 Uncertainty analyses hypercube design, brute force search, random search, and
PSO have been available for history matching and opti-
It is well known that automatic history matching is an mization tasks in CMOST studio. The Schlumberger
underdetermined inverse problem because the number of developed MEPO software to deal with production opti-
unknown parameters is much larger than the number of mization, uncertainty assessment, and semi-automatic his-
available production measurements. In other words, many tory matching. Although CMOST studio and MEPO
different parameter estimates can provide satisfactory data software did much work, automatic closed-loop reservoir
fits. However, some of the estimates may grossly lead to management has not yet been achieved, especially for
erroneous predictions of future production behavior. large-scale field applications. Therefore, much more effort
Therefore, developing methods that can quantify the should be paid on the development of integrated software,
uncertainty of parameter estimates are necessary to the which can help the engineers to manage oilfields easily and
success of automatic history matching. For example, Bar- scientifically.
ker et al. (2001) applied the Markov chain Monte Carlo
method to quantify the uncertainty. When predicting per-
meability from well logs, Olatunji et al. (2011) adopted a 5 Conclusions
type-2 fuzzy logic system which is good at handling
uncertainties in measurements and data used to calibrate Closed-loop reservoir management consisting of automatic
the parameters. Arnold et al. (2013) carried out a hierar- history matching and reservoir production optimization is
chical benchmark case study for history matching, uncer- an effective technique to improve the technical and eco-
tainty quantification, and reservoir characterization. Since nomic effect of reservoir development. Both of the steps
automatic history matching has become increasingly are complex optimization problems which can be solved by
important in the field of reservoir description, more efforts gradient-based algorithms, gradient-free algorithms, and
should be paid to uncertainty analyses to reduce the artificial intelligence algorithms. The computational effi-
decision risks in the future. ciency of gradient-based algorithms is the highest but they

123
Pet. Sci. (2015) 12:114–128 125

are difficult to transform from one simulator to another. Barragán HV, Vázquez RR, Rosales ML, et al. A strategy for
The gradient-free and artificial intelligence algorithms are simulation and optimization of gas and oil production. Comput
Chem Eng. 2005;30(2):215–27.
independent of numerical simulators but they need much Bieker HP, Slupphaug O, Johansen TA. Real-time production
more simulations to find the optimum. More research optimization of oil and gas production systems: a technology
should be conducted on hybrid optimization methods, survey. SPE Prod Oper. 2007;22(4):382–91 (paper SPE 99446).
parallel algorithms and distributed systems, uncertainty Bissell RC, Sharma Y, Killough JE. History matching using the
method of gradients: two case studies. Paper SPE 28590
analyses and integrated software in the future. presented at SPE annual technical conference and exhibition,
New Orleans, Louisiana; 25–28 September 1994.
Acknowledgments The authors greatly appreciate the financial Brouwer DR, Jansen JD. Dynamic optimization of water flooding
supports from the Important National Science & Technology Specific with smart wells using optimal control theory. SPE J.
Projects of China (Grant No. 2011ZX05024-004), the Natural Science 2004;9(4):391–402 (paper SPE 78278).
Foundation for Distinguished Young Scholars of Shandong Province, Caers J. Efficient gradual deformation using a streamline-based proxy
China (Grant No. JQ201115), the Program for New Century Excellent method. J Petrol Sci Eng. 2003;39(1–2):57–83.
Talents in University (Grant No. NCET-11-0734), the Fundamental Carter J, Romero C. Using genetic algorithm to invert numerical
Research Funds for the Central Universities (Grant No. simulations. Presented at the 8th European conference on the
13CX05007A, 13CX05016A), and the Program for Changjiang mathematics of oil recovery, Freiberg, Germany; 3–6 September
Scholars and Innovative Research Team in University (IRT1294). 2002.
Carter RD, Kemp LF, Pierce AC, et al. Performance matching with
Open Access This article is distributed under the terms of the constraints. SPE J. 1974;14(2):187–96 (paper SPE 4260).
Creative Commons Attribution License which permits any use, dis- Chacón E, Besembel I, Hennet JC. Coordination and optimization in
tribution, and reproduction in any medium, provided the original oil and gas production complexes. Comput Ind.
author(s) and the source are credited. 2004;53(1):17–37.
Chang H, Zhang D, Lu Z. History matching of facies distributions
with the EnKF and level set parameterization. J Comput Phys.
2010;229(20):8011–30.
References Chavent GM, Dupuy M, Lemonnier P. History matching by use of
optimal theory. SPE J. 1975;15(1):74–86 (paper SPE 4627).
Abacioglu Y, Oliver DS, Reynolds AC. Efficient reservoir history Chen C, Li G, Reynolds AC. Robust constrained optimization of
matching using subspace vectors. Comput Geosci. 2001;5(2):151–72. short- and long-term net present value for closed-loop reservoir
Agarwal A, Blunt MJ. Streamline-based method with full-physics management. SPE J. 2012;17(3):849–64 (paper SPE 141314).
forward simulation for history-matching performance data of a Chen WH, Gavalas GR, Seinfeld JH, et al. A new algorithm for
North Sea Field. SPE J. 2003;8(2):171–80 (paper SPE 84952). automatic history matching. SPE J. 1974;14(6):593–608 (paper
Agbalaka CC, Oliver DS. Application of the EnKF and localization to SPE 4545).
automatic history matching of facies distribution and production Chen Y, Oliver DS, Zhang D. Data assimilation for nonlinear
data. Math Geosci. 2008;40(4):353–74. problems by ensemble Kalman filter with reparameterization.
Agbalaka CC, Stern D, Oliver DS. Two-stage ensemble-based history J Petrol Sci Eng. 2009a;66(1–2):1–14.
matching with multiple modes in the objective function. Comput Chen Y, Oliver DS, Zhang D. Efficient ensemble-based closed-loop
Geosci. 2013;55:28–43. production optimization. SPE J. 2009b;14(4):634–45 (paper SPE
Alhuthali A, Oyerinde D, Datta-Gupta A. Optimal waterflood 112873).
management using rate control. SPE Reserv Eval Eng. Cheng H, Kharghoria A, He Z, et al. Fast history matching of finite-
2007;10(5):539–51 (paper SPE 102478). difference models using streamline derived sensitivities. SPE
Amit R. Petroleum reservoir exploitation: switching from primary to Reserv Eval Eng. 2005;8(5):426–36 (paper SPE 89447).
secondary recovery. Oper Res. 1986;34(4):534–49. Cheng H, Wen X, Milliken WJ et al. Field experiences with assisted
Arnold D, Demyanov V, Tatum D, et al. Hierarchical benchmark case and automatic history matching using streamline models. Paper
study for history matching, uncertainty quantification and SPE 89857 presented at SPE annual technical conference and
reservoir characterisation. Comput Geosci. 2013;50:4–15. exhibition, Houston, Texas; 26–29 September 2004.
Aronofsky JS, Williams AC. The use of linear programming and Chu L, Reynolds AC, Oliver DS. Computation of sensitivity
mathematical models in underground oil production. Manage coefficients for conditioning the permeability field to well-test
Sci. 1962;8(4):394–407. data. In Situ. 1995;19(2):179–223.
Asheim H. Maximization of water sweep efficiency by controlling Codas A, Campos S, Camponogara E, et al. Integrated production
production and injection rates. Paper SPE 18365 presented at optimization of oil fields with pressure and routing constraints:
European petroleum conference, London; 16–19 October 1988. the Urucu field. Comput Chem Eng. 2012;46:178–9.
Babayev DA. Mathematical models for optimal timing of drilling on Custsódio AL, Vicente LN. Using sampling and simplex derivatives
multilayer oil and gas fields. Manage Sci. 1975;21(12):1361–9. in pattern search methods. SIAM J Optim. 2007;18(2):537–55.
Bangerth W, Klie H, Wheeler M, et al. On optimization algorithm for Dovera L, Rossa ED. Multimodal ensemble Kalman filtering using
the reservoir oil well placement problem. Comput Geosci. Gaussian mixture models. Comput Geosci. 2011;15(2):307–23.
2006;10(3):303–19. Elkamel A. An artificial neural network for predicting and optimizing
Barker JW, Cuypers M, Holden L. Quantifying uncertainty in immiscible flood performance in heterogeneous reservoirs.
production forecasts: another look at the PUNQ-S3 problem. Comput Chem Eng. 1998;22(11):1699–709.
SPE J. 2001;6(4):433–41 (paper SPE 74707). Emerick AA, Reynolds AC. Combining the ensemble Kalman filter
Barnes RJ, Kokossis A, Shang ZG. An integrated mathematical with markov-chain monte carlo for improved history matching
programming approach for the design and optimisation of and uncertainty characterization. SPE J. 2012;17(2):418–40
offshore fields. Comput Chem Eng. 2007;31(5–6):612–29. (paper SPE 141336).

123
126 Pet. Sci. (2015) 12:114–128

Emerick AA, Reynolds AC. Ensemble smoother with multiple data Jansen JD. Adjoint-based optimization of multi-phase flow through
assimilation. Comput Geosci. 2013;55:3–15. porous media—a review. Comput Fluids. 2011;46(1):40–51.
Farmer CL. Numerical rocks, the mathematics of oil recovery. Jansen JD, Bosgra OH, Van den Hof PMJ. Model-based control of
Oxford: Clarendon Press; 1992. multiphase flow in subsurface oil reservoirs. J Process Control.
Fathi Z, Ramirez WF. Optimal injection policies for enhanced oil 2008;18(9):846–55.
recovery: part 2-surfactant flooding. SPE J. 1984;24(3):333–41 Jansen JD, Brouwer DR, Naevdal G, et al. Closed-loop reservoir
(paper SPE 12814). management. First Break. 2005;23(1):43–8.
Gao G, Reynolds AC. An improved implementation of the LBFGS Knudsen BR, Foss B. Shut-in based production optimization of shale-
algorithm for automatic history matching. SPE J. gas systems. Comput Chem Eng. 2013;58:54–67.
2006;11(1):5–17 (paper SPE 90058). Lasdon L, Coffman PE, Macdonald R, et al. Optimal hydrocarbon
Gao G, Zafari M, Reynolds AC. Quantifying the uncertainty for the reservoir production policies. Oper Res. 1986;34(1):40–54.
PUNQ-S3 problem in a Bayesian setting with RML and EnKF. Lee AS, Aronofsky JS. A linear programming model for scheduling
SPE J. 2006;11(4):506–15 (paper SPE 93324). crude oil production. J Petrol Technol. 1958;10(7):51–4.
Gao G, Li G, Reynolds AC. A stochastic algorithm for automatic Lee TY, Seinfeld JH. Estimation of two-phase petroleum reservoir
history matching. SPE J. 2007;12(2):196–208 (paper SPE 90065). properties by regularization. J Comput Phys. 1987;69(2):397–419.
Gavalas GR, Shah PC, Seinfeld JH. Reservoir history matching by Li G, Reynolds AC. Iterative ensemble Kalman filters for data
Bayesian estimation. SPE J. 1976;16(6):337–50 (paper SPE 5740). assimilation. SPE J. 2009;14(3):496–505 (paper SPE 109808).
Gomez S, Gosselin O, Barker JB. Gradient-based history matching Li G, Reynolds AC. Uncertainty quantification of reservoir perfor-
with a global optimization method. SPE J. 2001;6(2):200–8 mance prediction using a stochastic optimization algorithm.
(paper SPE 71307). Comput Geosci. 2011;15(3):451–62.
Gottfried BS. Optimization of a cyclic steam injection process using Li R, Reynolds AC, Oliver DS. History matching of three-phase flow
penalty functions. SPE J. 1972;12(1):13–20 (paper SPE 3329). production data. SPE J. 2003;8(4):328–40 (paper SPE 87336).
Grimstad AA, Mannseth T, Aanonsen S, et al. Identification of Liu N, Oliver DS. Evaluation of Monte Carlo methods for assessing
unknown permeability trends from history matching of produc- uncertainty. SPE J. 2003;8(2):188–95 (paper SPE 84936).
tion data. SPE J. 2004;9(4):419–28 (paper SPE 77485). Liu N, Oliver DS. Automatic history matching of geologic facies.
Grimstad AA, Mannseth T, Nævdal G, et al. Adaptive multiscale SPE J. 2004;9(4):429–36 (paper SPE 84594).
permeability estimation. Comput Geosci. 2003;7(1):1–25. Liu N, Oliver DS. Critical evaluation of the ensemble kalman filter on
Gu Y, Oliver DS. History matching of the PUNQ-S3 reservoir model history matching of geologic facies. SPE Reserv Eval Eng.
using the ensemble Kalman filter. SPE J. 2005;10(2):217–24. 2005a;8(6):470–7 (paper SPE 92867).
(paper SPE 89942). Liu N, Oliver DS. Ensemble Kalman filter for automatic history matching
Gu Y, Oliver DS. An iterative ensemble Kalman filter for multiphase of geologic facies. J Petrol Sci Eng. 2005b;47(3–4):147–61.
fluid flow data assimilation. SPE J. 2007;12(4):438–46 (paper Lorentzen RJ, Nævdal G. An iterative ensemble Kalman filter. IEEE
SPE 108438). Trans Autom Control. 2011;56(8):1990–5.
Gunnerud V, Foss B. Oil production optimization-A piecewise linear Lorentzen RJ, Nævdal G, Vallès B et al. Analysis of the ensemble Kalman
model, solved with two decomposition strategies. Comput Chem filter for estimation of permeability and porosity in reservoir models.
Eng. 2010;34(11):1803–12. Paper SPE 96375 presented at SPE annual technical conference and
Gupta AD, King MJ. Streamline simulation: theory and practice. exhibition, Texas, Dallas; 9–12 October 2005.
Textbook series, SPE textbook series Vol. 11. Richardson, Marsily G, Lavedan C, Boucher M. Interpretation of interference tests
Texas; 2007. in a well field using geostatistical techniques to fit the
He J, Sarma P, Durlofsky LJ. Reduced-order flow modeling and permeability distribution in a reservoir model. In: Geostatistics
geological parameterization for ensemble-based data assimila- for natural resources characterization. Dordrecht: Marechal
tion. Comput Geosci. 2013;55:54–69. Reidel Publishing Company; 1984. p. 831–49.
He N, Reynolds AC, Oliver DS. Three-dimensional reservoir McFarland JW, Lasdon L, Loose V. Development planning and
description from multiwell pressure data and prior information. management of petroleum reservoirs using tank models and
SPE J. 1997;2(3):312–27. nonlinear programming. Oper Res. 1984;32(2):270–89.
Heidari L, Gervais V, Ravalec ML, et al. History matching of McKie CJN, Rojas EA, Quintero NM. Economic benefits from
petroleum reservoir models by the Ensemble Kalman Filter and automated optimization of high pressure gas usage in an oil
parameterization methods. Comput Geosci. 2013;55:84–95. production system. Paper SPE 67187 presented at SPE produc-
Holland JH. Adaptation in natural and artificial systems: an tion and operations symposium, Oklahoma City, Oklahoma;
introductory analysis with applications to biology, control, and 24–27 March 2001.
artificial intelligence. Michigan: University of Michigan Press; Mochizuki S, Saputelli LA, Kabir CS, et al. Real-time optimization:
1975. classification and assessment. SPE Prod Oper. 2006;21(4):455–66
Hu LY, Zhao Y, Liu Y, et al. Updating multipoint simulations using (paper SPE 90213).
the ensemble Kalman filter. Comput Geosci. 2013;51:7–15. Moridis GJ, Reagan MT, Kuzma HA, et al. SeTES: a self-teaching
Jacquard P. Permeability distribution from field pressure data. SPE J. expert system for the analysis, design, and prediction of gas
1965;5(4):281–94 (paper SPE 1307). production from unconventional gas resources. Comput Geosci.
Jafarpour B, McLaughlin DB. History matching with an ensemble 2013;58:100–15.
Kalman filter and discrete cosine parameterization. Comput Nævdal G, Brouwer DR, Jansen JD. Waterflooding using closed-loop
Geosci. 2008;12(2):227–44. Control. Comput Geosci. 2006;10(1):37–60.
Jafarpour B, Goyal VK, McLaughlin DB, et al. Compressed history Nævdal G, Mannseth T, Vefring E. Near well reservoir monitoring
matching: exploiting transform-domain sparsity for regulariza- through ensemble Kalman filter. Paper SPE 75235 presented at
tion of nonlinear dynamic data integration problems. Math SPE/DOE improved oil recovery symposium, Tulsa, Oklahoma;
Geosci. 2010;42(1):1–27. 13–17 April 2002.
Jahns HO. A rapid method for obtaining a two-dimensional reservoir Negrete E, Datta-Gupta A, Choe J. Streamline-assisted ensemble
description from well pressure response data. SPE J. kalman filter for rapid and continuous reservoir model updating.
1966;6(4):315–27 (paper SPE 1473). SPE Reserv Eval Eng. 2008;11(6):1046–60 (paper SPE 104255).

123
Pet. Sci. (2015) 12:114–128 127

Olatunji SO, Selamat A, Abdulraheem A. Modeling the permeability Sarma P, Durlofsky L, Aziz K. Efficient closed-loop production
of carbonate reservoir using type-2 fuzzy logic systems. Comput optimization under uncertainty. Paper SPE 94241 presented at
Ind. 2011;62(2):147–63. SPE Europec/EAGE annual conference, Madrid, Spain; 13–16
Oliver DS. A comparison of the value of interference and well-test data June 2005.
for mapping permeability and porosity. In Situ. 1996a;20(1):41–59. Sequeira SE, Graells M, Luis P. Real-time evolution for on-line
Oliver DS. Multiple realizations of the permeability field from well optimization of continuous processes. Ind Eng Chem Res.
test data. SPE J. 1996b;1(2):145–54 (paper SPE 27970). 2002;41(7):1815–25.
Oliver DS, Chen Y. Recent progress on reservoir history matching: a Spall JC. Implementation of the simultaneous perturbation algorithm
review. Comput Geosci. 2011;15(1):185–221. for stochastic optimization. IEEE Trans Aerosp Electron Syst.
Oliver DS, Reynolds AC, Liu N. Inverse theory for petroleum 1998;34(3):817–23.
reservoir characterization and history matching. Cambridge: Sudaryanto B, Yortsos YC. Optimization of fluid front dynamics in
Cambridge University Press; 2008. porous media using rate control. I. equal mobility fluids. Phys.
Onwunalu J, Durlofsky L. Application of a particle swarm optimi- Fluids. 2000;12(7):1656–70.
zation algorithm for determining optimum well location and Tan TB, Kalogerakis N. A fully implicit three-dimensional three-
type. Comput Geosci. 2009;14(1):183–98. phase simulator with automatic history-matching capability.
Ouenes A, Brefort B, Meunier G et al. A new algorithm for automatic Paper SPE 21205 presented at SPE symposium on reservoir
history matching: application of simulated annealing method simulation, Anaheim, California; 17–20 February 1991.
(SAM) to reservoir inverse modeling. SPE general paper. (Paper Tarantola A. Inverse problem theory and methods for model
SPE 26297); 1993. parameter estimation. Philadelphia: SIAM; 2005.
Peters E, Chen Y, Leeuwenburgh O, et al. Extended Brugge Tavakoli R, Reynolds AC. History matching with parameterization
benchmark case for history matching and water flooding based on the SVD of a dimensionless sensitivity matrix. SPE J.
optimization. Comput Geosci. 2013;50:16–24. 2010;15(2):495–508 (paper SPE 118952).
Peters L, Arts RJ, Brouwer GK, et al. Results of the Brugge Tavakoli R, Reynolds AC. Monte Carlo simulation of permeability
benchmark study for flooding optimization and history match- fields and reservoir performance predictions with SVD param-
ing. SPE Reserv Eval Eng. 2010;13(3):391–405 (Paper SPE eterization in RML compared with EnKF. Comput Geosci.
119094). 2011;15(1):99–116.
Powell MJD. Developments of NEWUOA for minimization without Tavallali MS, Karimi IA, Teo KM, et al. Optimal producer well
derivatives. IMA J Numer Anal. 2008;28(4):649–64. placement and production planning in an oil reservoir. Comput
Qian X. Simulated annealing and conditional simulation. Master Chem Eng. 2013;55:109–25.
Thesis in Mathematics. New Mexico Tech, Socorro; 1993. Tillier E, Veiga SD, Derfoul R. Appropriate formulation of the
Ramgulam A, Ertekin T, Flemings PB. Utilization of artificial neural objective function for the history matching of seismic attributes.
networks in the optimization of history matching. Paper SPE Comput Geosci. 2013;51:64–73.
107468 presented at Latin American & caribbean petroleum Tokuda N, Takahashi S, Watanabe M. Application of genetic
engineering conference, Buenos Aires, Argentina; 15–18 April algorithm to history matching for core flooding. Paper SPE
2007. 88621 presented at SPE Asia pacific oil and gas conference and
Ramirez WF, Fathi Z, Cagnol JL. Optimal injection policies for exhibition, Perth, Australia; 18–20 October 2004.
enhanced oil recovery: part 1 theory and computational strate- Van Essen GM, Van den Hof PMJ, Jansen JD. Hierarchical long-term
gies. SPE J. 1984;24(3):328–32 (Paper SPE 11285). and short-term production optimization. SPE J. 2011;16(1):191–9
Reynolds AC, Zafari M, Li G. Iterative forms of the ensemble (paper SPE 124332).
Kalman filter. Presented at the proceedings of 10th European Wackowski RK, Stevens CE, Masoner LO et al. Applying rigorous
conference on the mathematics of oil recovery, Amsterdam; 4–7 decision analysis methodology to optimization of a tertiary recovery
September 2006. project: Rangely Weber Sand Unit, Colorado. Paper SPE 24234
Rodrigues JRP. Calculating derivatives for automatic history match- presented at oil and gas economics, finance and management
ing. Comput Geosci. 2006;10(1):119–36. conference, London, United Kingdom 28–29 April 1992.
Rosenwald GW, Green DW. A method for determining the optimum Wang C, Li G, Reynolds AC. Production optimization in closed-loop
location of wells in a reservoir using mixed-integer program- reservoir management. SPE J. 2009;14(3):506–23 (paper SPE
ming. SPE J. 1974;14(1):44–54 (paper SPE 3981). 109805).
Saputelli LA, Nikolaou M, Economides MJ. Real-time reservoir Wang Y, Li M. Reservoir history matching and inversion using an
management: a multi-scale adaptive optimization and control iterative ensemble Kalman filter with covariance localization.
approach. Comput Geosci. 2005;10(1):61–96. Petrol Sci. 2011;8(3):316–27.
Saputelli L, Malki H, Canelon J et al. A critical overview of artificial Wasserman ML, Emanuel AS. History matching three-dimensional
neural network applications in the context of continuous oil field models using optical control theory. J Can Petrol Technol.
optimization. Paper SPE 77703 presented at SPE annual 1976;15(4):70–7.
technical conference and exhibition, San Antonio, Texas; 29 Wattenbarger RA. Maximizing seasonal withdrawals from gas storage
September–2 October 2002. reservoirs. J Petrol Technol. 1970;22(8):994–8.
Sarma P, Chen WH, Durlofsky LJ, et al. Production optimization with Wen X, Chen W. Real time reservoir model updating using the ensemble
adjoint models under nonlinear control-state path inequality Kalman filter with confirming opinion. SPE J. 2006;11(4):431–42.
constraints. SPE Reserv Eval Eng. 2008a;11(2):326–39 (paper Wu SH, Zhang CS, Li QY, et al. A multilevel preconditioner and its
SPE 99959). shared memory implementation for new generation reservoir
Sarma P, Durlofsky LJ, Aziz K. Kernel principal component analysis simulator. Petrol Sci. 2014;11(4):540–9.
for efficient differentiable parameterization of multipoint geo- Wu W. Optimum design of field-scale chemical flooding using
statistics. Math Geosci. 2008b;40:3–32. reservoir simulation. PhD Dissertation. University of Texas,
Sarma P, Durlofsky LJ, Aziz K et al. A new approach to automatic Austin. 1996.
history matching using kernel PCA. Paper SPE 106176 presented Wu Z, Reynolds AC, Oliver DS. Conditioning geostatistical models
at SPE reservoir simulation symposium, Houston, Texas, USA; to two-phase production data. SPE J. 1999;4(2):142–55 (paper
26–28 February 2007. SPE 56855).

123
128 Pet. Sci. (2015) 12:114–128

Yang PH, Watson AT. Automatic history matching with variable- Zhang F, Skjervheim JA, Reynolds AC, et al. Automatic history
metric methods. SPE Reserv Eng. 1988;3(3):995–1001 (paper matching in a Bayesian framework: example applications. SPE
SPE 16977). Reserv Eval Eng. 2005;8(3):214–23 (paper SPE 84461).
Zafari M, Reynolds AC. Assessing the uncertainty in reservoir Zhao H, Chen C, Li G et al. Maximization of a dynamic quadratic
description and performance predictions with the ensemble interpolation model for production optimization. Paper SPE
Kalman filter. SPE J. 2007;12(3):382–91 (paper SPE 95750). 141317 presented at SPE reservoir simulation symposium, the
Zandvliet MJ, Bosgra OH, Van den Hof PMJ, et al. Bang-bang woodlands, Texas, USA; 21–23 February 2011.
control and singular arcs in reservoir flooding. J Petrol Sci Eng. Zhou K, Hou J, Zhang X, et al. Optimal control of polymer flooding
2007;58(1–2):186–200. based on simultaneous perturbation stochastic approximation
Zhang F, Reynolds AC, Oliver DS. Evaluation of the reduction in method guided by finite difference gradient. Comput Chem Eng.
uncertainty obtained by conditioning a 3D stochastic channel to 2013;55:40–9.
multiwell pressure data. Math Geol. 2002;34(6):715–42.

123