Spe 182660 Pa

J182660 DOI: 10.
2118/182660-PA Date: 23-March-18 Stage: Page: 367 Total Pages: 29
A Physics-Based Data-Driven Model for

History Matching, Prediction, and
Characterization of Waterflooding
Performance
Zhenyu Guo and Albert C. Reynolds, University of Tulsa; and Hui Zhao, Yangtze University
Downloaded from http://onepetro.org/SJ/article-pdf/23/02/367/2115144/spe-182660-pa.pdf/1 by Indian Institute of Technology (ISM) Dhanbad user on 28 March 2023
Summary
We develop and use a new data-driven model for assisted history matching of production data from a reservoir under waterflood and
apply the history-matched model to predict future reservoir performance. Although the model is developed from production data and
requires no prior knowledge of rock-property fields, it incorporates far more fundamental physics than that of the popular capacitance–
resistance model (CRM). The new model also represents a substantial improvement on an interwell-numerical-simulation model
(INSIM) that was presented previously in a paper coauthored by the latter two authors of the current paper. The new model, which is
referred to as INSIM-FT, eliminates the three deficiencies of the original data-driven INSIM. The new model uses more interwell con-
nections than INSIM to increase the fidelity of history matching and predictions and replaces the ad hoc computation procedure for
computing saturation that is used in INSIM by a theoretically sound front-tracking procedure. Because of the introduction of a front-
tracking method for the calculation of saturation, the new model is referred to as INSIM-FT. We compare the performance of CRM,
INSIM, and INSIM-FT in two synthetic examples. INSIM-FT is also tested in a field example.
Introduction
For a large-scale-or complex-reservoir-simulation model, a single forward-simulation run can take on the order of 1 hour to 1 or more
days to complete. Because constructing a history match requires dozens of simulation runs, the computational efficiency of history
matching is of paramount importance. To accelerate the history-matching procedure, many people have considered replacing the full-
scale reservoir-simulation model with a far more computationally efficient surrogate or proxy model. Use of a proper orthogonal
decomposition is a popular procedure (van Doren et al. 2006; Cardoso and Durlofsky 2010; Gildin et al. 2013; He and Durlofsky 2014)
for developing a surrogate simulation model. These reduced-order models provide surrogate simulators for full-scale reservoir simula-
tors and significantly enhance the computational efficiency by reducing the computational time spent on a forward-simulation run.
However, the development of the reduced-order proper-orthogonal-decomposition models requires comprehensive geological modeling
and the full-order reservoir-simulation model has to be run several times to construct the reduced-order model.
As opposed to reduced-order models, data-driven models require no knowledge of petrophysical properties or other specific geologi-
cal information. However, data-driven models offer the possibility of providing good history matches and future performance predic-
tions as well as reliable prediction of the movement of injected water to enable the monitoring of a waterflood. In addition, the
development and running of a data-driven model are expected to require a small fraction of the man-hours, engineering experience, and
computational costs that are associated with the development and running of a reduced-order model. Some relatively early work on
data-driven models (Refunjol 1996; Heffer et al. 1997; Jansen and Kelkar 1997) were correlation-based and were developed to distin-
guish the interwell geomechanics and flow trends derived from the Spearman correlation coefficient, and infer geological features using
a simple statistical analysis without consideration of reservoir properties. The application of these models has been somewhat limited,
in part because the statistical results can be easily influenced by data noise. Moreover, because of the nature of correlation-based meth-
ods, the future reservoir performance is not quantitatively predictable. Data-driven models avoid this deficiency.
Perhaps the most widely used data-driven model is referred to as the CRM (Yousef et al. 2005). The name stems from the analogy
of flow in porous media to the flow of electric charge in a resistor–capacitor circuit, where the compressibility and transmissibility,
respectively, are analogous to capacitance and resistance. Inspired by the work of Albertoni and Lake (2003), the CRM estimates the
interwell connectivity by a series of allocation factors. These allocation factors are approximately equal to the percentage of the total
water-injection rate at each injection well that flows to each producer connected to that injection well. Gentil (2005) extended the model
of Albertoni and Lake (2003) by providing a more physical explanation of the allocation factor and putting forth an empirical power-
law model to estimate oil cut. By introducing a model parameter called the time constant in CRM, Yousef et al. (2005) overcame the li-
mitation of the Albertoni and Lake (2003) model, which ignored reservoir compressibility. CRM can efficiently estimate and match the
total-production-rate data by knowing the injection-rate information and, if available, the bottomhole-pressure (BHP) data at producers.
Lake et al. (2007) applied the empirical power-law model (Gentil 2005) in CRM to estimate the oil-phase rate and perform history
matching and life-cycle-production optimization. Since then, more empirical fractional flow models for miscible and immiscible flows
(Sayarpour 2008; Weber 2009; Nguyen 2012; Cao et al. 2015) have been developed for use with CRM to estimate the rates of
different phases.
One defect of CRM is that it assumes a constant productivity index (PI) for each producer during the entire production history,
which is not strictly physically correct in a multiphase-flow system. For example, in waterflooding, as water is constantly injected into a
reservoir, the PIs of different producers must change with time because saturation profiles at producers change with time. Recent work
on CRM (Cao 2014; Cao et al. 2014, 2015) mitigates this problem by only matching a certain period of production data, during which
the PIs remain approximately constant. However, the partial match of historical data evidently degrades the accuracy of predictions
generated with CRM. More importantly, the allocation factors, which are the most important model parameters in CRM, are assumed to
Copyright V
C 2018 Society of Petroleum Engineers
This paper (SPE 182660) was accepted for presentation at the SPE Reservoir Simulation Conference, Montgomery, Texas, USA, 20–22 February 2017, and revised for publication. Original
manuscript received for review 28 November 2016. Revised manuscript received for review 29 July 2017. Paper peer approved 3 August 2017.
April 2018 SPE Journal 367
ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

J182660 DOI: 10.2118/182660-PA Date: 23-March-18 Stage: Page: 368 Total Pages: 29
be constant during the entire production life. This assumption is tenuous because allocation factors change with flow pattern and injec-
tion rates (Thiele and Batycky 2006). In addition, the fractional-flow model integrated with CRM originates from an empirical correla-
tion, which—at best—provides only an approximation of the correct physics of flow. However, because of this empirical model for
fractional flow, CRM does not compute the saturation distribution, or require explicit knowledge of the initial saturation distribution or
relative permeability curves. Instead, CRM uses a correlation to compute the oil cut at producing wells.
Lerlertpakdee et al. (2014) proposed a flow-network model, which represents reservoir flow by a coupled-network model in which
each pair of wells is connected with a one-dimensional (1D) finite-difference reservoir-simulation model. Two sets of model parame-
ters, absolute permeability and pore volume (PV), are defined at each gridblock in the 1D flow-network system. Relative permeabilities
are assumed to be known. This coupled system is solved for pressures and saturations semi-implicitly, effectively by a set of 1D reser-
voir-simulation runs. To perform production optimization, a robust training scheme was developed to simultaneously history match sev-
eral sets of simulated data obtained by running a full-scale simulator with different sets of well controls, which were expected to cover
the range of control variables that are expected to be encountered during the iterations of a life-cycle production-optimization algo-
rithm. Subsequently, the trained flow-network model is applied for production optimization. Compared with CRM, the Lerlertpakdee
et al. (2014) model uses the true relative permeability curves, which is physically correct but requires a priori knowledge of the relative
permeability curves. The discretization of 1D interwell connections makes the flow-network model act as a simplified grid-based full-
scale simulator, yet it may result in a large number of model parameters for a large-scale reservoir with a large number of wells.
Another data-driven model, the INSIM, was developed by Zhao et al. (2015) to history match the production data from a reservoir
under waterflooding. Somewhat reminiscent of a streamtube (Thiele et al. 1996) or a streamline (Thiele and Batycky 2006) model,
INSIM (Zhao et al. 2015; 2016a, b) approximates a reservoir by a set of 1D interwell connective units, but unlike streamline simulation,
INSIM defines the streamtubes a priori before any computations are performed. INSIM solves for water saturation along each connec-
tive unit, which is somewhat similar to a streamtube, but in INSIM, streamtubes are defined a priori. However, instead of discretizing
the 1D interwell connections by a series of finite-difference gridblocks, as in Lerlertpakdee et al. (2014), INSIM only defines a pair of
parameters for each connection, which significantly reduces the number of model parameters. Although INSIM has been shown to
work well for many examples, its original formulation contains three defects. First, unlike CRM and the new INSIM-FT formulation
presented here, INSIM requires a priori knowledge of the relative permeability curves and fluid viscosity so that the phase rates at pro-
ducing wells can be computed during an INSIM run. Second, INSIM generally uses a single connection between an injector and neigh-
boring producers, which may not always provide a sufficient number of flow paths to obtain sufficient accuracy in history matching and
prediction [although Zhao et al. (2016a) were aware of the potential benefits of adding additional connections, they did not carefully
pursue this modification]. Third, and most importantly, there is a fundamental theoretical issue with INSIM; namely, that the method it
uses to calculate water saturation is theoretically flawed for complex cases where the saturation distribution along a connection between
wells cannot be computed from the standard Riemann solution; in such situations, the INSIM procedure for calculating water saturation
is ad hoc. In particular, the INSIM method for calculating water saturation is theoretically incorrect whenever a producer is converted
to a water-injection well and would also be incorrect if, as in INSIM-FT, we use imaginary wells to provide more flow paths for injected
water. Therefore, adding more flow paths in INSIM would invoke the incorrect ad hoc procedure for computing saturation more often.
The new data-driven model considered here uses a theoretically correct front-tracking procedure to calculate water saturation, hence the
name INSIM-FT. In addition to introducing a theoretically correct procedure to compute water saturation, INSIM-FT remedies the other
two potential deficiencies of INSIM by introducing the parameters defining power-law relative permeability curves as additional his-
tory-matching parameters so that prior knowledge of relative permeabilities is no longer required, and by introducing imaginary wells
and their associated interwell connections (streamtubes) to enable more potential flow paths. Intuitively, one expects that INSIM-FT
will be more robust than INSIM because of these modifications. We note, however, that even if cases exist where INSIM-FT does not
significantly improve the quality of the history match or future-reservoir-performance predictions obtained from INSIM, INSIM-FT
would still be preferable to INSIM because its procedure for computing water saturation is theoretically sound, whereas the saturation-
calculation procedure used in INSIM becomes ad hoc under the circumstances noted previously. An overall comparison of the assump-
tions and abilities of INSIM-FT, INSIM, and CRM is shown in Table 1. Table 1 also applies when the interior-point optimizer used in
the original INSIM code is replaced by the ensemble smoother with multiple data assimilation (ES-MDA) to obtain a more robust
implementation referred to as “INSIM (ES-MDA).”
CRM INSIM INSIM-FT

Requires known relative permeabilities No Yes No
Provides accurate saturations No No Yes
Add imaginary wells/additional flow paths No No Yes
Adherence to correct physics of flow No Partially Largely
Approximates reservoir connectivity Yes Yes Yes
Table 1—Overall comparison among CRM, INSIM, and INSIM-FT.
To compare the performance of INSIM-FT with INSIM and CRM, we consider two examples. The first example tests the ability of
the methods to identify the presence of a sealing fault. The second example pertains to a highly heterogeneous synthetic reservoir with
the simulation model dependent on a 225 2251 grid. To test the ability of the methods to estimate reservoir connectivity and moni-
tor waterfloods, we compare results generated with the methods with those obtained with FrontSim (Schlumberger 2013b). We also test
a field example using INSIM-FT to illustrate the practical applicability of the method.
Methodology
Like INSIM (Zhao et al. 2015), INSIM-FT is an interwell-numerical-simulation model that characterizes the reservoir as a set of 1D
connective flow volumes (streamtubes), as shown in Fig. 1. In Fig. 1, each red circle represents the static bulk volume controlled by the
well at its center, and this volume does not change with time. However, because of the rock compressibility, the associated PV, which
is denoted by Vp,i, is a function of the average pressure in this volume and thus can change with time. Similarly, the dark-gray regions
in Fig. 1 represent the static constant bulk volumes between well nodes. The PV associated with the bulk volume between wells i and j
368 April 2018 SPE Journal

is denoted by Vp,i, j, and in the INSIM and INSIM-FT models, we assume that any change in this volume from its initial value is entirely
caused by rock-compressibility effects. The transmissibility of connection (i, j) is denoted by Ti, j. The sets of Vp,i, j and Ti, j that charac-
terize these interwell PVs are model parameters in the INSIM and INSIM-FT models. INSIM involves no other parameters, but INSIM-
FT also includes in its parameter set some of the parameters that define power-law relative permeability curves.
2
Ti,2,Vp,i,2 Ti,1,Vp,i,1
i 1
Ti,3,Vp,i,3
3
Fig. 1—Connective units between wells of INSIM.
In the following context, we provide the key formulations of INSIM-FT, which partially but not entirely follows the work of Zhao et al.
(2015). Oilfield units are used throughout this manuscript. In INSIM and INSIM-FT, pressure equations are solved implicitly as implicit
pressure/explicit saturation (IMPES) to obtain the pressures at the well nodes. INSIM-FT replaces the INSIM procedure for calculating
water saturation, which is not theoretically sound under all conditions, with a theoretically rigorous procedure called front tracking to calcu-
late the saturation distribution along the interwell connection. The front-tracking procedure used is presented in detail in Appendix A.
We define total mobility by
kro ðSw Þ krw ðSw Þ
kt ðSw Þ ¼ þ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð1Þ
lo lw
where lm, m ¼ o,w denote the oil and water viscosities (in cp), and krm, m ¼ o,w represent the oil and water relative permeabilities. The
transmissibility at well connection (i, j) is defined as
ki; j Ai; j kt ðSw;i; j Þ
Ti; j ¼ 1:127 103 ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð2Þ
Li; j
where the water saturation of connection (i, j), Sw,i, j, is decided by upstream weighting; i.e., if pi > pj, then Sw,i, j ¼ Sw,i, with Sw,i being
the water saturation at well node i. Otherwise Sw,i, j ¼ Sw, j, with Sw, j being the water saturation at well node j; Ai, j is the average cross-
sectional area (in ft2) of the flow-connective unit between wells i and j; Li, j is the distance from well i to well j (in ft); ki, j is the absolute
permeability in the associated volume (in md). The total compressibility at control volume i at time tn is defined by
cnt;i ¼ Sno;i co þ Snw;i cw þ cr ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð3Þ
where we assume throughout that the rock compressibility (cr) and fluid compressibilities (cm, m ¼ o,w) (in 1/psi) are known constants;
So,i and Sw,i are defined as the oil and water saturations on Vp,i, where Vp,i is the PV corresponding to the volume associated with node i
(Fig. 1). From the IMPES formulation, the discrete form of the pressure equation for INSIM-FT, which is derived from a material bal-
ance on the control volume of well i, is given by
nc;i
X 1 cn1 n1
t;i Vp;i
Ti;n1 n n n
j ðpj pi Þ þ qt;i ¼ ðpni pn1
i Þ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð4Þ
j¼1
5:615 Dtn
where nc,i represents the number of wells that are connected to well i; pressures are in psi; the subscript n denotes the timestep; qnt;i in
RB/D is the total rate of well i at tn, where a positive value denotes injection and negative value denotes production; and the transmissi-
bilities (Ti;n1
j ) in Eq. 4 are defined as
3
kt ðSn1
w;i; j Þ kt ðSn1
w;i; j Þ
Ti;n1
j ¼ 1:127 10 ki; j Ai; j ¼ Ti;0 j : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .ð5Þ
Li; j kt ðS0w;i; j Þ
Throughout, kt ðS0w;i; j Þ is specified as the total mobility at the initial water saturation S0w;i; j , and Sn1
w;i; j is equal to the saturation of the upstream
well node between the connection (i, j) at time level n–1; i.e., Sn1 n1 n1
w;i; j ¼ Sw;i if pi > pn1
j , and otherwise Sn1 n1
w;i; j ¼ Sw; j . In the examples, we
assume that the initial saturation is equal to the irreducible water saturation Siw, and that the relative permeability curves are obtained by nor-
malizing effective permeabilities by the oil relative permeability at irreducible water saturations so that kro(Siw) ¼ 1. It follows that
1
kt ðS0w;i; j Þ ¼ : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð6Þ
lo
n1
In the INSIM and INSIM-FT formulations, Vp;i in Eq. 4 is assumed to satisfy the relation
Xnc;i
n1 n1
Vp;i ¼ 0:5 Vp;i; j ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð7Þ
j¼1
n1
where Vp;i; j is approximated by the first-order Taylor-series expansion given by

n1 0 n1 0
Vp;i; j ¼ Vp;i; j ½1 þ cr ðpi; j p Þ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð8Þ
where p0 is the initial reservoir pressure, and
pn1 n1
i; j ¼ 0:5ðpi þ pn1
j Þ:
With the assumption of Eq. 7, it follows that Eq. 4 involves only two types of parameters—Ti, j and Vp,i, j —which vary with time,
with Ti, j and Vp,i, j given by Eqs. 5 and 8, respectively. INSIM assumes that relative permeabilities, viscosities, compressibilities, and
initial pressure are all known so that Ti;0 j and Vp;i;
0
j are the only unknown parameters. (INSIM-FT adds parameters defining power-law
relative permeability functions as parameters.) According to Eq. 4, there is one pressure equation per well node. By solving the linear
system of pressure equations, the pressure of each well-control volume at time tn is obtained, and the total-liquid-flow rate in RB/D
along the connective volume (i, j) can be computed by
qnt;i; j ¼ Ti;n1 n n
j ðpj pi Þ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .ð9Þ
where a positive value of qt,i, j means well j (or node j) is upstream of well i (or node i) and a negative value means well i (or node i) is
upstream of well j (or node j). For the calculation of water saturation, we assume incompressible flow, so the partial-differential equa-
tion governing the saturation distribution for linear flow through one of the connective volumes between a pair of nodes is in the form
of the Buckley-Leverett equation, given by
@Sw ðx; tÞ 5:615qt;i; j ðtÞ @fw ðx; tÞ

þ ¼0 for 0 x Li; j ; tn1 t tn ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð10Þ
@t /i; j Ai; j @x
where qt,i, j(t) is assumed equal to qnt;i; j from tn1 to tn; fw is the water fractional flow; /i, j is the average porosity of the connective vol-
ume between wells i and j; and the x-direction, which varies from connection volume to connection volume, is in the direction of the
line segment that connects well i to j. By neglecting gravity and capillarity, fw, as a function of water saturation, is given by
1
fw ðSw Þ ¼ : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð11Þ
kro ðSw Þ lw
1þ
krw ðSw Þ lo
Using
Vp;i; j
/i; j Ai; j ¼ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð12Þ
Li; j
Eq. 10 can be rewritten as

@Sw ðx; tÞ 5:615qt;i; j Li; j @fw ðx; tÞ
þ n ¼ 0; for 0 x Li; j ; tn1 t tn : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð13Þ
@t Vp;i; j @x
Eq. 13 is solved semianalytically from tn1 to tn with the front-tracking procedure of Appendix A, assuming that qt;i; j ðtÞ ¼ qnt;i; j for
tn1 t tn . The initial condition for Eq. 13 is the saturation distribution for 0 x Li, j at tn1 along the well connection (i, j), which
is given by
Sw ðx; tn1 Þ ¼ Sn1

w ðxÞ; for 0 x Li; j : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð14Þ
The boundary condition for Eq. 13 is the water saturation, Suw;i; j , at time tn1, at the upstream node (either x ¼ 0 or x ¼ Li, j) for flow
between wells i and j. This upstream saturation is assumed to be constant from tn1 to tn. If the upstream node is a water injector, then
Suw;i; j ¼ 1 Sor , where Sor is the residual oil saturation and fw ðSuw;i; j Þ ¼ 1 at all times.
The saturation expression in Eq. 13 applies to each connected well pair. The computation of the saturation profile at tn along each
connection by solving Eq. 13 is independent of the solutions along the other connections. After the saturation computation, the satura-
tion profile of each connection at time level n has the form of
Sw ðx; tn Þ ¼ Snw ðxÞ; for 0 x Li; j ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð15Þ
along each pair of connected nodes.
Relative Permeability Functions. In INSIM-FT, we use power-law relative permeability functions but estimate major parameters in
these functions as part of the ES-MDA history-matching procedure (Emerick and Reynolds 2012; 2013a, b), whereas in INSIM the rela-
tive permeabilities are assumed known a priori. Moreover, INSIM-FT allows one to use a different set of relative permeability curves
along each connection, although our experience with this application is limited. In all cases, the relative permeability curves are given
by Corey-type relations. Specifically,
krw ðSwn Þ ¼ a Snwnw ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð16Þ
and
kro ðSwn Þ ¼ Snwno ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð17Þ

where
Sw Siw
Swn ¼ : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð18Þ
1 Siw Sor
In our applications, we assume that Siw and Sor are known accurately and a, nw, and no are included as parameters to be estimated by
history matching. The fluid viscosities, compressibilities, and initial conditions are assumed to be known and fixed.
Front Tracking. The problem defined by Eq. 13 is, in general, a Cauchy problem and reduces to the standard Buckley-Leverett prob-
lem only if the water saturation at the boundary condition is Suw;i; j ¼ 1 Sor at all times, and the initial condition is uniformly the irre-
ducible water saturation along an interwell connection. Specifically, in the application of INSIM-FT, the standard Buckley-Leverett
solution applies whenever we need to solve the saturation profile along a single connection between an injector and a downstream well
(a producer or an imaginary well). However, the standard Buckley-Leverett solution is the solution of a Riemann problem, and when
well A is connected to multiple upstream wells and at least a downstream well, well B, the saturation between well A and well B, which
still must satisfy Eq. 13, can no longer be represented as a Riemann problem because the saturation at well A is continually changing
with time after water breakthrough at well A. INSIM uses an ad hoc procedure to compute water saturation when the standard Buckley-
Leverett solution does not apply. Here, we provide a general procedure for computing saturations, and we implement a front-tracking
method conceptually similar to the one proposed by Holden et al. (1988) to solve the Cauchy problem. This method divides the Cauchy
problem into a set of Riemann problems, which have analytical solutions. To avoid computing curved shock paths and to simplify the
problem, the rarefaction waves of the analytical solutions of Riemann problems are approximated by a set of small shocks. By connect-
ing all the solutions of sub-Riemann problems, we obtain the global well-posed solution for Eq. 13. This method has the advantage of
being unconditionally stable and introduces little numerical dispersion. A complete derivation of the procedure is given in Appendix A.
Water Saturation and Oil-Production Rate at Well Nodes. The front-tracking algorithm gives the saturation profile along any pair
of well connections. However, when flow into (or out of) the well node occurs along multiple paths that are directly connected to the
well node, we need a way to compute the node saturation. Specifically, if one well, Wi (a producer or an imaginary well introduced
later), is directly connected to multiple wells, the water fractional flow of Wi is computed as the ratio of the sum of the water-inlow rates
to the sum of the total-inflow rate; i.e.,
Nc;i n
X
n
qt;i; j f nw;i; j
fw;i ¼ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð19Þ
j¼1
qnt;i; j
n
where Nc,i is the number of wells that are connected to Wi and are also upstream of Wi, and fw;i; j is the water fractional flow measured at the
downstream node i along connection (i, j) at time level n. The saturation at Wi is obtained by inverting the fractional-flow function; i.e.,
Snw;i ¼ fw1 ð f nw;i Þ: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð20Þ
Accordingly, the oil-production rate at producer i is estimated as

Nc;i
X
qno;i ¼ qnt;i; j ð1 fw;i;
n
j Þ: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð21Þ
j¼1
Demonstration Case. Here, we provide an example to demonstrate that the original INSIM algorithm does not always provide an
accurate saturation distribution when the Buckley-Leverett solution does not apply along all connections at all times. The example is a
simple four-node, three-connection reservoir with four wells. The geometry of the reservoir is shown is Fig. 2. Note that there are four
well nodes and three connected well pairs. Although the example in Fig. 2 seems quite simple, this example was chosen because it is
representative of part of the more complete connection map that we use for cases with far more wells, as will be clear when we discuss
Example 1. The reservoir has homogeneous absolute permeability and porosity fields. The three connections have the same average
cross-sectional areas of 1,000 ft2. Other properties are defined in Table 2. The two injectors, W1 and W2, both inject water at the same
constant rate of 100 RB/D. W3 is a dummy well that does not inject or produce, and W4 is producing at a constant rate equal to 200 RB/D.
We run the example with Eclipse (Schlumberger 2013a), INSIM-FT, and INSIM and observe the water cut at W4. For this example, we
assume that the correct relative permeability curves are known a priori.
W1 W3 W2
1,000 ft
500 ft
W4
Fig. 2—The geometry for a three-connection reservoir.
This is an example where the water fronts from two injectors meet at a node; i.e., the water fronts from W1 and W2 meet at W3. In
this situation, the analytical solution of the Buckley-Leverett equation does not give the correct saturation distribution between W3 and
W4 after water breakthrough at W3; thus, we expect that INSIM cannot give the correct saturation profile between W3 and W4. As shown

in Fig. 3, the water cut obtained with INSIM-FT is in good agreement with that obtained with Eclipse, but INSIM cannot predict the
correct breakthrough time and water cut for W4. Therefore, INSIM-FT provides a more accurate saturation prediction than does INSIM.
Property Values
Absolute permeability (md) 1,000
Porosity 0.2
–1 –10
Oil compressibility (psi ) 3.4×10
–1 –10
Water compressibility (psi ) 3.4×10
–1 –10
Rock compressibility (psi ) 4.3×10
Water viscosity (cp) 1
Oil viscosity (cp) 20
Initial reservoir pressure (psi) 3,675
Table 2—Properties of the three-connection reservoir.
0.8
WWCT: W4
0.6
0.4
0.2
0
0 20 40 60 80 100
Time (days)
Fig. 3—Water cut in W4. The black line denotes the results obtained from INSIM, the blue line denotes the results obtained from
INSIM-FT, and the red dashed line denotes the results obtained from Eclipse. WWCT 5 well water cut.
INSIM-FT vs. INSIM

In addition to a more reliable procedure for calculating saturation at nodes than is available in the original INSIM model and adding
terms defining power-law relative permeability curves as parameters, INSIM-FT includes two other features that make it more robust
and flexible than INSIM. These enhancements are discussed in the following three subsections.
Imaginary Wells. In INSIM, all nodes represented the locations of actual wells and the only connections were between existing well
pairs. In INSIM-FT, it is possible to add imaginary wells and associated connections to increase the possible flow paths along which
injected fluid can travel. INSIM-FT approximates a reservoir by a set of 1D connective units along which conservation equations are
solved, which is similar to streamtube or streamline simulators. Thus, increasing the number of connections by adding imaginary well
nodes effectively increases the number of streamtubes or streamlines along which fluid can be transported.
Another benefit of adding imaginary wells is that this avoids the direct connection between two producers or two injectors, as done
in Zhao et al. (2015). In INSIM, the flow pattern is always 1D and the flow direction between a connected well pair is from the well
node with high pressure to the well node with low pressure, no matter the well type. Thus, if there is a direct connection of two injection
wells, one of the injectors must serve as a sink, which is not consistent with the physics. Similarly, if two producers are connected, one
of the producers most serve as a source, whereas both producers should be sinks. Adding at least one imaginary well between each pro-
ducer pair and each injector pair avoids the aforementioned issue. It is important to note the total-flow rate at any imaginary well is zero
(i.e., no fluid is produced or injected by means of an imaginary well).
It may be worthwhile to note that the aforementioned problem of 1D flow between a pair of producers or a pair of injectors could
seemingly be eliminated by deleting all direct connections between producer pairs and injector pairs, but such a modification potentially
eliminates the flow path or streamtube that enables the production of oil (or displacement of oil) from the interwell volume directly
between the producer pair (injector pair). Thus, we do not generally recommend the elimination of such connections.
The basic guidelines for adding imaginary wells are as follows. First, at least one imaginary well must be added along a direct con-
nection (streamtube) between two producers (or two injectors) because otherwise there is no way to accurately model the physics of
flow between the two wells. For example, without adding an imaginary well between two producers, one producer must be upstream of
the other and flow in the interwell region would flow from the upstream well to the downstream well, whereas in reality, part of the fluid
between the two producers would flow to each producer. Second, there should be a sufficient number of imaginary wells added so that
at least there are two distinct flow paths between each injector/producer pair. These guidelines have proved satisfactory for the exam-
ples presented in this paper, including the field example as well as other synthetic examples we have tried. However, there is no guaran-
tee that examples will not arise where the user may need to increase flow paths between injectors and producers to obtain an acceptable
history match.
Total Rate Constraint on Well Control Volume. Another feature introduced in INSIM-FT is a rate constraint that states that the sum
of the liquid rates flowing into the volume controlled by each producer is equal to the specified total rate of production from that

producer. In the case where the control volume of a well is very large, it may be possible for fluid to flow through the well control vol-
ume without being produced by the well, and in this case the rate constraint does not represent the true physics. Although our INSIM-
FT code allows the user to choose whether to impose the rate constraint, all examples presented here impose this constraint because, by
means of computational experiments not shown, we found that when the rate constraint is added to the optimization problem, we gener-
ally obtain better history matches than are obtained without the constraints.
If well i is a producer, the rate constraint becomes
Nc;i
X
qnt;i ¼ qnt;i; j ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð22Þ
j¼1
where qnt;i; j is computed from Eq. 9 and Eq. 22 is applied at each producing well. qnt;i is the total-liquid-production rate specified at well
i from tn1 to tn. Nc,i of the summation in Eq. 22 is over only connections that have a positive value of qt,i, j, which is the case if and
only if the producing well i is downstream of well j; i.e., flow is from node j into the control volume of well i. By enforcing Eq. 22 as a
XNc;i
constraint, the outward-flow rate from control volume i is approximately equal to qn qnt;i , which is zero.
j¼1 t;i; j
Interwell Connectivity. For waterflood management, it is important to quantify which wells are “fed” by which injectors, and one of
the common ways is through the interwell connectivity. Interwell connectivity is defined in the CRM as an allocation factor fi, j, which
represents the percentage of the total-water-injection rate at each injection well i that flows toward producer j. Note that the subscripts i
and j in CRM have different meanings than in INSIM-FT. These allocation factors are used to reflect large-scale geological characteris-
tics. Unlike CRM, INSIM (Zhao et al. 2016a) considers the allocation factor as a time-dependent variable so that the allocation factor
for each injector/producer pair is varying at each INSIM timestep. Different from INSIM-FT, INSIM does not have imaginary wells in
the connection map, and there is only one streamtube between an injector/producer pair. Within the time interval [tn1, tn], the alloca-
tion factor fi, j(t) for the connection between injector j and producer i is assumed constant and equal to fi;nj , with fi;nj given by
qnt;i; j
fi;nj ¼ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð23Þ
qnt; j
where qnt;i; j is the total-flow rate from j toward i and qnt; j is the total injection rate of j at the nth timestep. For comparison purposes, to
obtain a single-value presentation of the allocation factor for each injector/producer pair, the allocation factors in INSIM over all the
INSIM steps are averaged by
Xnt
ð f n qn Þ
n¼1 i; j t; j
f i; j ¼ : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð24Þ
qnt; j
In INSIM-FT, we provide a procedure to compute representative terms similar to the average allocation factors in INSIM that can
reflect reservoir connectivity between injector/producer pairs. Specifically, we provide the means to compute the interwell-total-flow
rate from an injector that flows to a connected producer over the time interval [tn1, tn]. The average total-flow rate between an injector/
producer pair can then be computed by averaging the interwell-total-flow rates calculated at the INSIM-FT timesteps over the produc-
tion history. When computing the connectivity, unlike INSIM, which only considers the flow caused by the direct connection between
an injector and producer, INSIM-FT also considers the flow from an injector that passes through intermediate nodes represented by
imaginary wells to a producer.
To illustrate how to compute this “pass-through flow rate,” we assume that the water from an injector will at most pass through one
imaginary well before arriving at a producer. Consider the situation shown in Fig. 4, where injected water from I2 must either directly
pass through the volumes connecting the injector/producer pair, I2 and P1, or pass through either node Im1 or Im2, or more correctly,
through the volumes associated with these imaginary wells before reaching P1. We let qt,P1,I2 represent the total-flow rate along the
direct path connecting node I2 to P1, let qt,Im1,Ij for j ¼ 1, 2 be the total-flow rate between injector Ij and imaginary well Im1, and let
qt,P1,Im1,I2 be the total flow rate from I2 to P1 that passes through the volume associated with imaginary node Im1. Finally, we let qt,P1,Im1
be the total flow rate through the volume directly connecting Im1 and P1. Then, the flow rate from I2 that passes through Im1 and finally
arrives at P1 is approximated by
qIm ;I
qt;P1 ;Im1 ;I2 ¼ qP1 ;Im1 X2 1 2 : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð25Þ
q
i¼1 Im;Ii
Flow direction
P2
Im1
I1 P1
I2
Im2
Fig. 4—Schematic used to compute the flow rate from I2 passing through the intermediate imaginary wells to P1.
The pass-through flow rate from I2 to P1 through the volume associated with imaginary well Im2 can be calculated in a similar way.
The total-interwell-flow rate between I2 and P1, which is denoted by q^t;P1 ;I2 , is given by

X2
q^t;P1 ;I2 ¼ qt;P1 ;I2 þ j¼1
qt;P1 ;Imj ;I2 ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð26Þ
where q^ is used to indicate that this rate includes the pass-through rates plus the rate along the direct connection between I2 and Pj.
After the total-interwell-flow rate is obtained at every INSIM-FT time level, the average total-interwell-flow rate over the entire pro-
duction history is used to represent the interwell connectivity. In the examples presented here, to validate the preceding procedure for
determining how the water-injection rate of a particular injection well is allocated among producers, we compare the connectivity
obtained from INSIM-FT with that calculated from the true reservoir model with the FrontSim reservoir simulator of Schlumberger
(Schlumberger 2013b). FrontSim can calculate and generate liquid-flow rates between an injector/producer pair at every FrontSim sim-
ulation time level. By averaging these flow rates over the production history, we obtain the interwell connectivity in the same format as
in INSIM-FT. Because FrontSim runs based on the same true geological model as Eclipse 100, which is used to generate the production
data, the connectivity calculated with FrontSim reflects the connectivity of the true model. We also compare connectivity results
obtained with FrontSim with those obtained with CRM and INSIM, where the allocation factors are converted to the same format as
used by FrontSim. Letting q t;i; j be the connectivity used by INSIM-FT or FrontSim, we convert f i; j in Eq. 24 to q t;i; j by
q t;i; j ¼ q t; j f i; j ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð27Þ
where q t; j is the average total injection rate over the historical period. fi, j of CRM can be converted in a similar way.
History Matching
Parameters. In the original INSIM, we assume relative permeability curves are known, whereas in INSIM-FT, we assume relative per-
meabilities are of the form given by Eqs. 16 and 17, where the exponents no and nw and the coefficient a are unknown parameters to be
determined by history matching. INSIM-FT includes the option of estimating one global set of relative permeabilities or estimating a
different set along each well-pair connection. The common parameters of INSIM and INSIM-FT are transmissibilities and PVs of con-
nections at time zero. For i ¼ 1,2,…Nw1, j ¼ i þ 1,i þ 2,…Nw, Ti;0 j represents the transmissibility of connection (i, j) at time zero
0
and Vp;i; j represents the PV of connection (i, j) at time zero. Note that Nw includes all the actual wells and imaginary wells. In practice,
not every two wells are connected to each other and a connection map describing how wells are connected to each other will be given a
priori for a reservoir before history matching. However, as discussed previously, we do not allow a direct connection between two injec-
tors or between two producers to prevent a producer from becoming a source or an injector from becoming a sink. If wells i and j are
not connected in the connection map, then Ti;0 j and Vp;i;
0
j will not be included as parameters for history matching. To satisfy the volume
balance of the reservoir, the PV of each connection at time zero must sum to the initial total PV of a reservoir:
w 1 X
NX Nw
0 0
Vp;i; j ¼ Vp;tot ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð28Þ
i¼1 j¼iþ1
0
where Vp;tot is the total PV of a reservoir at time zero and the two summations are over all the connected well pairs (i, j) in a predefined
0
connection map. As an extension to INSIM, INSIM-FT has the option of whether to include Vp;tot as a parameter to be estimated by his-
tory matching.
INSIM and INSIM-FT are both rate-control-based data-driven models, which means injection rates and total-production rates should
be specified at each timestep. In INSIM, the observed oil-production rates are the data that are history matched. Letting the vector dobs
be the vector of all observed oil-production rates, the objective function that we minimize in INSIM and INSIM-FT to obtain a history
match is given by
1
OðmÞ ¼ ½gðmÞ dobs T C1
D ½gðmÞ dobs ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð29Þ
2
where CD is the covariance matrix of measurement error in diagonal and dobs is the vector of the observed oil-production rates. In
INSIM, m is the parameter vector that only includes Ti;0 j and Vp;i; 0
j for all the connected well pairs, whereas in INSIM-FT, a, no, nw (Eqs.
0
16 through 18), and possibly Vp;tot are also included in the set of history-matched parameters. In our applications of INSIM-FT, the first
0 0
two examples consider Vp;tot is known and the field example considers Vp;tot is a parameter. The constraints of the optimization problem
for INSIM-FT are given by
8 0
> Ti; j 0; for all connected well pairs ði; jÞ; ð30aÞ
>
>
>
> 0 0
>
> 0 Vp;i; j Vp;tot ; for all connected well pairs ði; jÞ; ð30bÞ
>
>
>
> XN
>
>
c;i
>
> qnt;i; j ðmÞ ¼ qnt;i ; for i ¼ 1; 2; …Npro ; ð30cÞ
>
>
>
> j¼1
>
>
< NX w 1 X Nw
0 0
> Vp;i; j ¼ Vp;tot ; summations only over connected well pairs ði; jÞ; ð30dÞ
>
>
>
> i¼1 j¼iþ1
>
>
>
> 0 aði; jÞ 1; for all connected well pairs ði; jÞ or for a global set of relative permeabilities ; ð30eÞ
>
>
>
> 1 no;ði; jÞ 6; for all connected well pairs ði; jÞ or for a global set of relative permeabilities ; ð30fÞ
>
>
>
>
>
> 1 nw;ði; jÞ 6; for all connected well pairs ði; jÞ or for a global set of relative permeabilities ; ð30gÞ
>
>
: 0 0 0 0
Vp;tot;low Vp;tot Vp;tot;up ; if Vp;tot is included as parameter ; ð30hÞ
0 0
where Npro is the number of producers in the reservoir and Vp;tot;low and Vp;tot;up , respectively, are the lower bound and upper bound for
0
Vp;tot , which must be specified a priori.
In INSIM, the constraints for the history-matching optimization problem (Eq. 29) only involve Eqs. 30a, 30b, and 30d. Zhao et al.
R
(2016a) used an interior-point optimization algorithm in the MATLABV (MathWorks 2011) optimization toolbox to solve this

constrained optimization problem. However, the optimization tool from MATLAB estimates the components of the gradient by finite-
difference approximations, which can introduce errors as well as computational complexity if the number of parameters exceeds a few
hundred. Here, we use ES-MDA (Emerick and Reynolds 2012; 2013a, b) for history matching with INSIM-FT. ES-MDA requires the
generation of an ensemble of Ne initial models (realizations of m), and these models actually provide a rank-deficient approximation of
a prior covariance matrix, which provides regularization during the history-matching process. ES-MDA effectively represents one itera-
tion of the Gauss-Newton method with a matrix of average sensitivity coefficients for the case where the objective to be minimized is
given by the O(m) of Eq. 29, plus a regularization term (Reynolds et al. 2006).
Eqs. 30a, 30b, 30e, 30f, 30g, and 30h only involve the simple bound constraints, whereas Eq. 30d is a linear-equality constraint and
Eq. 30c is a nonlinear-equality constraint. First, we consider how to deal with the bound constraints and the linear-equality constraint.
During the history-matching procedure using ES-MDA, after each data-assimilation step, the ensemble of reservoir models (mi,
i ¼ 1,2,…Ne) is updated using an update equation (Emerick and Reynolds 2012, 2013a,b), which does not ensure that each m satisfies
the bound constraints or linear-equality constraint. Here, we propose to enforce the two types of constraints after each data-assimilation
step. For each updated reservoir model in the ensemble, the simple bound constraints are enforced by truncation; i.e., when an updated
history-matching variable is greater than its upper bound, we set the variable equal to its upper bound, and whenever a variable is less
than its lower bound, we set that the variable equal to its lower bound. The constraint of Eq. 30d should be enforced after Eq. 30c is
0 0;a
enforced. For one reservoir-model m in the ensemble, letting Vp;i; j after enforcing Eq. 30c at each ES-MDA step be Vp;i; j , we enforce
0;a 0;aþ
Eq. 30d by replacing Vp;i; j with Vp;i; j , which is defined as
0;a
0;aþ 0
Vp;i; j
Vp;i; j ¼ Vp;tot XNw 1 XNw ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð31Þ
i¼1
V 0;a
j¼iþ1 p;i; j
which ensures that

w 1 X
NX Nw
0;aþ 0
Vp;i; j ¼ Vp;tot ;
i¼1 j¼iþ1
for this specific realization of m. The procedure of Eq. 31 is repeated for all the other reservoir models at each ES-MDA step to ensure
that Eq. 30d is satisfied. Next, to deal with the nonlinear constraint of Eq. 30c, dobs in Eq. 29 is expanded to include the specified total-
production rate as the observed data in addition to the observed oil-production rate where the predicted total-production rate at well i at
tn is given by Eq. 22, for i ¼ 1; 2; …Npro . CD in Eq. 29 is also expanded to include the variance of the measurement error of the specified
XNc;i
total rates, which must be specified a priori. qn ðmÞ is history matched to qnt;i . Notice that by performing the match of total-
j¼1 t;i; j
XNc;i
production rate, qn ðmÞ is not guaranteed to be exactly equal to qnt;i for each of the history-matched reservoir models, because
j¼1 t;i; j
the history-matching procedure cannot give the exact match. However, this match ensures that most inflow rates are produced from the
producer. For all examples presented later, we obtain acceptable matches and predictions even though Eq. 30c is not exactly satisfied.
Application of INSIM-FT
To illustrate the accuracy of INSIM-FT, we consider three examples. The first example pertains to a homogeneous reservoir with a sealing
fault, which was initially used by Cao et al. (2014) to show that CRM can identify the existence of the fault. This example is included to
demonstrate that INSIM-FT can identify large-scale reservoir connectivity features such as a sealing fault, which is similar to CRM.
Because one can always raise the issue that our implementation of CRM is imperfect, we use an in-house CRM implementation (Sayarpour
2008) that is similar to the use of Cao et al. (2014, 2015). In the second example, we consider a channelized reservoir and again show that
INSIM-FT outperforms CRM. The third example is a true field example where bottomwater drive is the primary recovery mechanism.
To obtain a quantitative measurement of the quality for history matching and future predictions for the methods considered, we
define the normalized data mismatch for matching and predictions. The normalized data mismatch for a given model m is defined by
1
ONd ðmÞ ¼ ½gðmÞ dobs T C1
D ½gðmÞ dobs ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð32Þ
Nd
where for the data mismatch in the historical period, Nd is the number of the observed oil-production rate measured during the historical
period, and dobs and C1
D are the corresponding vector of observed data and the covariance matrix for measurement error, respectively.
For the data mismatch in the prediction period Nd, dobs and C1 D in Eq. 32 are defined similarly for the prediction period. For history
matching with INSIM and INSIM-FT, the average value of normalized data mismatches is computed over all the posterior realizations
of the models.
Example 1: Homogeneous Reservoir With a Sealing Fault. We consider a reservoir with a sealing fault, as shown in Fig. 5. Except
for the fault, the reservoir has a homogeneous permeability field of 200 md and the porosity field is also uniform with / ¼ 0.2. The reser-
voir has five injectors and four producers in a five-spot pattern. The completely sealing fault, shown in white as inactive gridblocks,
divides this reservoir into two noncommunicating regions. The performance of our in-house CRM and INSIM-FT software with/without
adding imaginary wells are compared. To verify our implementation of CRM, the same reservoir properties used by Cao et al. (2014) are
entered. The reservoir properties specified in Table 3 are the same properties used by Cao et al. (2014). The Eclipse reservoir-simulation
model representing the true model is dependent on a 33 331 grid with grid dimensions being Dx ¼ Dy ¼ 77.5 ft and Dz ¼ 193.75 ft. In
the Eclipse reservoir-simulation model, the five injectors are operated at a sequence of specified rates that change with time, and four pro-
ducers are produced at a constant BHP of 250 psi. The true data for history matching with INSIM-FT or CRM are obtained by running
Eclipse 100 for 3,000 days with the true reservoir-simulation model where the varying injection rates and constant producing BHP are
specified as well-operating conditions. CRM requires the injection rate to be specified as well-operating conditions and only total-produc-
tion rate is history matched. For history matching with CRM, the observed total-liquid-production rates and oil-production rates are
obtained by adding uncorrelated Gaussian noise to the true data where the standard deviation of each measurement error is set equal to
2% of its true value. For history matching with INSIM-FT, the true total-liquid-production rates are specified as well-operating condi-
tions at each INSIM-FT timestep and the observed data, the oil-production rates, are the same ones as used in history matching with

CRM. Note that in INSIM-FT, because the sum of the total-inflow rate into the control volume of one producer is required to be matched
to the specified total-production rate of that producer, we must also specify the standard deviation of the measurement error of each speci-
fied total-production rate, which is 2% of the specified rate. Here, the data from the first 2,250 days of the Eclipse run are used for history
matching and the remaining 750 days for future predictions.
J1 J3 J2
I1 P3 I2
J1 P 1 JI 55 P 2J 2
I3 P4 I4
J3 J4 J4
Fig. 5—Fault geology.
Property Values
–1 –5
Oil compressibility (psi ) 3×10
–1 –6
Water compressibility (psi ) 1×10
–1 –6
Rock compressibility (psi ) 1×10
Table 3—Properties of the fault-segmented reservoir.
The CRM-generated history match and future predictions of the total production are shown in Fig. 6. In Fig. 6 and similar figures
presented later, the vertical line separates the historical and future-prediction time periods. The result of Fig. 6 is consistent with that
from Cao et al. (2014), which suggests that our implementation of CRM is correct. It should be noted that in CRM, the oil cut is fit by
linear regression after history matching the total-production rate and then the oil rate is calculated by multiplying the total-production
rate by the oil cut, whereas in INSIM-FT, the oil-production rate is directly computed and history matched. Despite the empirical nature
of the CRM calculation of the oil rate, Fig. 7 shows that the oil rate computed with CRM is in good agreement with the historical oil-
rate data and gives a very close prediction.
1,500
4,000
WLPR:P4 (RB/D)
WLPR:P3 (RB/D)
3,000 1,000
2,000
500
1,000
0 0
0 1,000 2,000 3,000 0 1,000 2,000 3,000
Time (days) Time (days)
(a) CRM:P3 (b) CRM:P4
Fig. 6—The total-production rates obtained from the history-matched CRM model. Red circles are observed total rates; red lines
are true total-production rates; and gray lines are estimated total rates with the history-matched CRM model. (a) CRM for P3; (b)
CRM for P4. WLPR 5 well liquid-production rate.
INSIM-FT Without Adding Imaginary Wells. The simplest possible well connections for INSIM-FT are shown in Fig. 8 for two
0 0
producing wells. The parameters required to be estimated include the values of Vp;i; j and Ti; j over all connections and a global set of pa-
0
rameters defining power-law relative permeabilities with Vp;tot known a priori. The ensemble size for ES-MDA, Ne, is 250. It is assumed

that the prior probability density function (PDF) for each parameter is given by a normal distribution. For each connected well pair
0
(i, j), the mean value of Vp;i; j for this normal distribution is estimated as
0 Li; j 0
V p;i; j ¼ XNw 1 XNw Vp;tot : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð33Þ
k¼1
L
l¼kþ1 k;l
1,500
4,000
WOPR:P4 (RB/D)
WOPR:P3 (RB/D)
3,000 1,000
2,000
500
1,000
0 0
0 1,000 2,000 3,000 0 1,000 2,000 3,000
(a) CRM:P3 (b) CRM:P4
Fig. 7—The oil-production rates calculated from the history-matched CRM model. Red circles are observed oil rates; red lines are
true values of the oil rates; and gray lines are estimated oil-production rates by the history-matched CRM model. (a) CRM for P3;
(b) CRM for P4. WOPR 5 well oil-production rate.
I1 P3 I2
P1 I5 P2
I3 P4 I4
Fig. 8—Connection map generated without adding imaginary wells.
0;k
Next, each prior ensemble member, Vp;i; j , for k ¼ 1; 2; …Ne is randomly generated by sampling the Gaussian distribution
0 0 2
ðV p;i; j ; ð0:2V p;i; j Þ Þ . According to Eq. 2, for each connected well pair (i, j),
0 0
ki; j Vp;i; j kt ðSw;i; j Þ
Ti;0 j ¼ 1:127 103 : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð34Þ
/0i; j L2i; j
Replacing /0i; j in Eq. 34 with /0, where /0 is the true porosity of the uniform porosity field, and substituting Eq. 6 into Eq. 34 yields
0
ki; j Vp;i; j
Ti;0 j ¼ 1:127 103 : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð35Þ
lo /0 L2i; j
0 0
Letting T i; j be the mean value of Ti;0 j ; T i; j is estimated by replacing ki, j in Eq. 35 with k (the initial guess of the mean value of ki, j),
and is given by
0
0 kV p;i; j
T i; j ¼ 1:127 103 ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð36Þ
lo /0 L2i; j
where k ¼ 500 md, a value that is 150% higher than the true absolute permeability value of the uniform-permeability field, which
is equal to 200 md; /0 ¼ 0.2; and lo ¼ 1 cp. For k ¼ 1,2,…Ne, the prior ensemble member, Ti;0;kj , is generated by sampling
0 0
ðT i; j ; ð0:2T i; j Þ2 Þ. The relative permeability model in this example is given by Eqs. 16 through 18, and the true values of the relative
permeability parameters are given by atrue ¼ 0:6; no;true ¼ 1:5; nw;true ¼ 1:5, Siw;true ¼ 0:2, and Sor;true ¼ 0:2.

To history match the relative permeabilities, Siw and Sor are assumed to be known and the other three parameters are estimated with
the history-matching procedure. The mean values of a, no, and nw are not equal to their true values, and, more specifically, are given as
a ¼ 0:653; n o ¼ 1:70; and n w ¼ 1:46. The prior PDFs of the three relative permeability parameters are specified by ða; 0:052 Þ;
ðn o ; 0:12 Þ, and ðn w ; 0:12 Þ, where these PDFs must also be sampled to obtain the 250 initial realizations of relative permeability pa-
rameters. The oil-production rates from multiple prior realizations of the INSIM-FT parameters are shown in Fig. 9. The history match
obtained with ES-MDA using INSIM-FT as the forward model is shown in Fig. 10. As can be seen, the match of oil-production rate is
not particularly good and in fact is much worse than the “match” of oil rates obtained with CRM. The prior and posterior ensemble of
the oil/water relative permeability curves are shown in Figs. 11 and 12a, respectively. The posterior oil/water relative permeability
curves obtained by history matching are quite different from the true relative permeabilities. (However, as shown in Fig. 12b, much bet-
ter history matches can be obtained if we introduce imaginary wells to provide additional connections and flow directions in the
INSIM-FT model.)
6,000 3,000
5,000 2,500
WOPR:P3 (RB/D)
WOPR:P4 (RB/D)
4,000 2,000
3,000 1,500
2,000 1,000
1,000 500
0 0
0 1,000 2,000 3,000 0 1,000 2,000 3,000
(a) INSIM-FT:P3 (b) INSIM-FT:P4
Fig. 9—The estimated oil-production rates obtained from the prior INSIM-FT models. Red circles are observed oil-production rates;
red lines are true values of the oil-production rates; and gray lines are estimated oil rates obtained by running the prior INSIM-FT
models. Example 1 is shown, a faulted reservoir with no imaginary wells. (a) INSIM-FT for P3; (b) INSIM-FT for P4.
4,000 1,500
WOPR:P3 (RB/D)
WOPR:P4 (RB/D)
3,000
1,000
2,000
500
1,000
0 0
0 1,000 2,000 3,000 0 1,000 2,000 3,000
Fig. 10—The oil-production rates obtained with the history-matched INSIM-FT models. Red circles are observed oil rates; red lines
are true oil-production rates; and gray lines are estimated oil rates with the history-matched INSIM-FT models. Example 1 is
shown, a faulted reservoir with no imaginary wells. (a) INSIM-FT for P3; (b) INSIM-FT for P4.
0.8
0.6
0.4
0.2
0.2 0.3 0.4 0.5 0.6 0.7 0.8

Sw
Fig. 11—Prior oil/water relative permeability curves. The red solid lines represent the true relative permeability curves, and blue
lines represent the prior models of relative permeabilities. Example 1 is shown, a faulted reservoir with no imaginary wells.

1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Sw Sw
(a) No imaginary wells (b) Imaginary wells added
Fig. 12—Posterior oil/water relative permeability curves. The red solid lines represent the true relative permeability curves, and
blue lines represent the history-matched models of relative permeabilities. Example 1 is shown, a faulted reservoir with no imagi-
nary wells. (a) No imaginary wells; (b) imaginary wells added.
INSIM-FT With Imaginary Wells. Four imaginary wells are now added to this reservoir, as shown in Fig. 13, with the new well-
connection map built by use of interwell distances (Zhao et al. 2015) shown in Fig. 14. Note that the part of the full-connection map of
Fig. 14 that connects only the four wells I3, Im3, I5, and P4 is equivalent to the structure of the four-well-connection map of Fig. 2 pro-
vided that we equate wells I3, Im3, I5, and P4, respectively, in Fig. 14 to W1, W3, W2, and W4, respectively in Fig. 2.
JI1 PJ33 I2J2

1
Im1 Im2
P
J11 JI55 P2J2
Im3 Im4
JI3 PJ44 I4J4

3
Fig. 13—INSIM-FT well placement with adding imaginary wells. Example 1 is shown, a faulted reservoir.
I1 P3 I2
Im5 Im2
P2 I5 P2
Im3 Im4
I3 P4 I4
Fig. 14—INSIM-FT connection map with imaginary wells and connections added. Example 1 is shown, a faulted reservoir.
The prior ensemble of vectors of model parameters is generated exactly the same way as for the case without imaginary wells. The
oil-production rates from multiple realizations of the vectors of model parameters that are based on the prior models are shown in
Fig. 15. Fig. 16 shows the results from three methods where the INSIM (ES-MDA) results are generated by replacing the interior-point
optimizer used in the original INSIM code by ES-MDA. We made the change of history-matching algorithms because the finite-differ-
ence approximations used to compute the gradient for the interior-point optimizer sometimes yield a relatively poor gradient approxi-
mation, and because of this, the interior-point optimizer can fail to give a good estimate of a minimizing set of model parameters. The

INSIM method obtained by replacing the interior-point optimizer with ES-MDA is referred to as INSIM (ES-MDA), whereas the results
labeled INSIM in tables or figures refer to results generated using the original INSIM code, which implements a MATLAB interior-
point optimizer with all necessary derivatives computed by finite-difference approximations. It is important to note that imaginary wells
are added only when INSIM-FT is used. The results of Fig. 16 illustrate that the history matches and predictions obtained from all three
methods shown are fairly similar. Table 4, however, provides a precise quantitative comparison of the history-matching results from
CRM, INSIM-FT, INSIM, and INSIM (ES-MDA) dependent on the normalized data match. For this example and all examples that we
have tried, INSIM (ES-MDA) gives a data match that is better than the one obtained with INSIM. For this example, the INSIM (ES-
MDA) gives a substantially superior history match than is obtained with INSIM but a worse history match than is obtained with
INSIM-FT. The results of Table 4 indicate that INSIM-FT produces the best history match and INSIM (ES-MDA) produces a better his-
tory match than CRM. The original INSIM algorithm produces a poor history match. It is important to note that the application of CRM
gives a virtually exact prediction with the prediction error equal to 4.8, but this result is an abnormality because one should generally
expect that the prediction error will be greater than or equal to the history-match error, a result that occurs with INSIM, INSIM (ES-
MDA), and INSIM-FT, as shown in Table 4. For further illustration, we reconsider the same example but use only the first 440 days of
data when performing the history match and use all the other data to monitor prediction accuracy; the quantitative results are shown in
Table 5. As the results indicate, the prediction error for CRM is no longer close to zero, although it is still less than the prediction error
obtained with INSIM-FT. However, CRM still yields a worse data match than is obtained with INSIM-FT or INSIM (ES-MDA).
4,000 2,000
WOPR:P3 (RB/D)
WOPR:P4 (RB/D)
3,000 1,500
2,000 1,000
1,000 500
0 0
0 1,000 2,000 3,000 0 1,000 2,000 3,000
Fig. 15—The oil-production rates obtained from the prior INSIM-FT models. Red circles are observed oil rates; red lines are true oil
rates; and gray lines are estimated oil rates with the prior INSIM-FT models. Example 1 is shown, a faulted reservoir with imaginary
wells. (a) INSIM-FT for P3; (b) INSIM-FT for P4.
4,000 4,000
4,000
WOPR:P3 (RB/D)
WOPR:P3 (RB/D)
WOPR:P3 (RB/D)
3,000 3,000
3,000
2,000 2,000
2,000
1,000 1,000 1,000
0 0 0
0 1,000 2,000 3,000 0 1,000 2,000 3,000 0 1,000 2,000 3,000
Time (days) Time (days) Time (days)
(a) CRM:P3 (b) INSIM (ES-MDA):P3 (c) INSIM-FT:P3
1,500 1,500 1,500

WOPR:P4 (RB/D)
WOPR:P4 (RB/D)
WOPR:P4 (RB/D)
1,000 1,000 1,000
500 500 500
0 0 0
0 1,000 2,000 3,000 0 1,000 2,000 3,000 0 1,000 2,000 3,000
(d) CRM:P4 (e) INSIM (ES-MDA):P4 (f) INSIM-FT:P4
Fig. 16—Comparison of oil-production rates calculated with the CRM, INSIM (ES-MDA), and INSIM-FT history-matched models
where only INSIM-FT uses imaginary wells. Example 1 is shown, a faulted reservoir. Red circles are observed data; red lines are
true data of oil-production rate; and the gray curve is posterior oil-production rate. (a) CRM for P3; (b) INSIM (ES-MDA) for P3; (c)
INSIM-FT for P3; (d) CRM for P4; (e) INSIM (ES-MDA) for P4; (f) INSIM-FT for P4.

CRM INSIM INSIM-FT INSIM (ES-MDA)

Historical period 74.3 752.2 54.5 62.2
Prediction period 4.8 1,202 89.4 85.2
Table 4—Comparison of data mismatch for the historical period (2,250 days) and the prediction
period for the CRM, INSIM, INSIM-FT, and INSIM (ES-MDA) history-matched models; Example 1,
faulted reservoir.

Prediction period 237.3 3,208.9 295 501
Table 5—Comparison of data mismatch for the historical period (440 days) and the prediction
period for the CRM, INSIM, INSIM-FT, and INSIM (ES-MDA) history-matched models; Example 1,
faulted reservoir.
Finally, by adding imaginary wells, we obtain a far more reasonable estimate of relative permeability curves, comparing the results
of Fig. 12b, where imaginary wells are added, with the INSIM-FT estimates of Fig. 12a, which uses the connection map of Fig. 8 with
no imaginary wells added.
A referee posed a question about the overfitting of data, so this topic may be worthy of a brief discussion. Because adding imaginary
wells increases the number of parameters in INSIM-FT, the concern is that by continuing to add imaginary wells and hence connections
and parameters, we will at some point encounter overfitting of data, which corresponds to fitting noise, and when this happens, future
predictions may become unreliable. To begin, we note that if m in Eq. 32 is set equal to the true model and dobs is the random observa-
tion vector, then the expected value of ONd ðmÞ is equal to unity; thus, the history-matching errors shown in Tables 4 and 5 indicate that
we have not overfit the observed data. A theoretical basis for the expected values of ONd can be found in Tarantola (2005), and the dis-
crepancy principle is discussed in Vogel (2002) and Iglesias and Dawson (2013). Nevertheless, we cannot claim with certainty that if
the number of interwell connections, and hence the number of parameters in INSIM-FT, were increased, the data would not be overfit;
that is, the normalized objective function evaluated at the history-matched model would not be reduced to less than unity. In fact, the
possibility of overfitting using ES-MDA with INSIM-FT may be more likely than with normal ES-MDA applications because there is
no spatial correlation between the prior model parameters in the INSIM-FT and INSIM (ES-MDA) implementations. However, we
should perhaps also note that adding imaginary wells is akin to adding streamlines, and to the best of our knowledge, overfitting issues
have not arisen when using streamline simulators in history matching even though there are parameters associated with each streamline.
Interwell Connectivity. In all INSIM algorithms, interwell connectivity is represented by the average interwell-total-flow rate
between injector/producer pairs (see the discussion centered on Eqs. 25 and 26). Here, the accuracy of connectivities estimated with the
three INSIM algorithms is investigated by comparison with connectivity generated by running the FrontSim simulator using the true
reservoir model as input. In FrontSim, the interwell-total-flow rate between a well pair can be generated at each timestep of the entire
history-matched period. The average total-flow rate is calculated as the average of all these output-flow rates over the history-matched
period. With CRM, however, the interwell connectivity is represented by allocation factors that are defined in Yousef et al. (2005);
INSIM (Zhao et al. 2015) used a connectivity formulation that is equivalent to the allocation factor in CRM. Thus, to obtain a valid
comparison, the allocation factor between each injector/producer well pair obtained with CRM, INSIM, or INSIM (ES-MDA) is con-
verted to the average interwell-total-flow rate by simply multiplying the allocation factor by the average total-injection rate of the asso-
ciated injector over the production history.
To facilitate better comparison, the interwell connectivities obtained from FronSim, CRM, INSIM (ES-MDA), and INSIM-FT are
shown in Figs. 17a, 17b, 17c, and 17d, respectively. In these and similar figures, the length of each narrow red triangle represents the
magnitude of the interwell connectivity and the direction points to the producer associated with the injector/producer well pair. The
number alongside each red triangle indicates the value of the average total-interwell-flow rate (over the history-matching period)
between the injector/producer pair (in RB/D). For brevity, the values of interwell-flow rate are shown in Fig. 17 only for the connections
with a large interwell-flow rate. As shown in Fig. 17b, the existence of the fault is reflected by the fact that the interwell-flow rate trans-
lated from the CRM allocation factor is small (on the order of 10 RB/D or fewer) between every injector/producer pair that is separated
by the fault (i.e., with wells lying on different sides of the fault). However, unlike the CRM result of Fig. 17b, the FrontSim result of
Fig. 17a indicates that there is no flow between (I2, P1) and (I3, P3). In fact, the FrontSim streamline field, which is not shown here, indi-
cates that there is no streamline from I3 to P3, or from I2 to P1. Thus, the CRM result showing a large flow rate from I3 to P3 and I2 to
P1 seems to be incorrect. The INSIM results, which are not shown, are unreasonable in the sense that there is a considerable amount of
flow between injector/producer pairs that crosses the sealing fault; for example, the average flow rate between injector I5 and P4 from
the INSIM results is 265.6 STB/D, as opposed to the correct value of zero, which follows from the fact that I5 and P4 are on opposite
sides of the fault. Because the INSIM connectivity results are inaccurate, we have included only the INSIM (ES-MDA) results in Fig.
17. Note the INSIM (ES-MDA) results are consistent with the presence of the fault and overall at least as good as those obtained with
INSIM-FT. The connectivity estimates obtained with CRM are clearly inferior to those obtained from INSIM-FT and INSIM (ES-
MDA) compared with the FrontSim results.
Example 2: Channelized Reservoir. In this example, history matching is performed with INSIM-FT, INSIM, and CRM for a channel-
ized reservoir. The log-permeability field is generated using a 25 251 geological grid, which is shown in Fig. 18. The red/yellow
regions represent high-permeability channels. However, the reservoir-simulation model is dependent on the refined grid obtained by
subdividing each geological gridblock into a 3 31 subgrid with each of the nine subgrid blocks having the same value of permeabil-
ity as the host coarse-grid cell. Thus the reservoir-simulation model, which represents the true reservoir, is dependent on a
225 2251 grid where the size of each of these gridblocks is 90 9010 ft, where 10 ft is the thickness of the gridblock. The reser-
voir has a homogeneous porosity field with / ¼ 0.2. Other reservoir properties are shown in Table 6. The well locations shown in Fig.

18 indicate that we have introduced 12 imaginary wells to increase the pathways along which water can flow from injectors to pro-
ducers. The resulting connection map for INSIM-FT with imaginary wells is shown in Fig. 19. We use 800 days of production data for
history matching and then predict reservoir performance for an additional 200 days. The true total-production rates and oil-production
rates are obtained by running the 225 2251 Eclipse model with the well-operating schedule defined in Table 7. For history matching
with CRM, the total-production rates are the observed data to be history matched and the oil-production rates are predicted with the his-
tory-matched CRM model. The corresponding observed total- and oil-production rates for matching with CRM are obtained by adding
uncorrelated Gaussian noise with standard deviation equal to 2% of the true data. However, for history matching with INSIM-FT, the
true total-production rates are specified as well controls and the observed oil-production rates (the same data used for matching with
CRM) are history matched. To match the sum of the inflow rate into the control volume of one producer to the total-production rate of
that producer with ES-MDA, the standard deviation of the measurement error of each specified total-production rate is set to 2% of the
specified rate.
I1 P3 I2 I1 P3 I2
778.6 1648.1 712.9 1057.0
556.8
667.5 883.8 624.3 13.8
771.9
P1 I5 P2 P1 I5 P2
671.4 714.7
1502.6 767.3 916.5 695.4

570.2
I3 P4 I4 I3 P4 I4
803.8 719.6
(a) FrontSim (b) CRM
I1 P3 I2 I1 P3 I2
783.8 1629.0 810.0 1528.1
662.2 804.6 636.1 oIm1 771.6 oIm2
P1 I5 P2 P1 I5 P2
723.2 664.4
1484.8 753.2 1397.0 oIm3 oIm4 777.6
I3 P4 I4 I3 P4 I4
817.9 793.5
(c) INSIM (ES-MDA) (d) INSIM-FT
Fig. 17—Interwell connectivity obtained from FrontSim, CRM, INSIM (ES-MDA), and INSIM-FT for the fault case. The length of
the narrow red triangle denotes the magnitude of total-interwell-flow rate between an injector/producer pair and the direction of
the triangle indicates which producer belongs to this injector/producer well pair. (a) FrontSim; (b) CRM; (c) INSIM (ES-MDA);
(d) INSIM-FT.
P1 IM1 P2 P3
IM2
10
IM5
IM3 IM4 9
I1
I2
8
P6
P4 P5
IM6
7
IM7
6
IM8 IM9 IM10
I3 I4
5
IM11 IM12 P9 4
P7 P8
Fig. 18—Log-permeability field, channelized reservoir.

Property Name Values

Porosity 0.2
–1 –5
Oil compressibility (psi ) 3.4×10
–1 –5
Water compressibility (psi ) 3.4×10
–1 –5
Rock compressibility (psi ) 4.3×10
Table 6—Properties of the channelized reservoir.
P1 Im1 P2 Im2 P3
Im3 I1 Im4 I2 Im5
P4 Im6 P5 Im7 P6
Im8 I3 Im9 I4 Im10
P7 Im11 P8 Im12 P9
Fig. 19—Connection map, channelized reservoir.
Well Name Well Control Values

I1–I4 Rate control (RB/D) 1,000
P1 BHP control (psi) 500
P2–P4 BHP control (psi) 2,000
P5 BHP control (psi) 1,000
Table 7—Well-operating schedule for the Eclipse simulation model.
For ES-MDA, the initial ensemble of model parameters includes initial guesses for transmissibilities and PVs at time zero defined
0
by connections and parameters defining power-law relative permeabilities. Vp;tot is known a priori in this example. The ensemble size
k;0
Ne is equal to 300. For all the connected well pairs (i, j), the kth prior realization of Vp;i; j is generated by sampling the normal distribu-
0 0
tion ðV p;i; j ; ð0:2V p;i; j Þ2 Þ, with V p;i; j given by Eq. 33. Similarly, for k ¼ 1,2,…Ne, the prior Ti;k;0j is sampled from the normal distribu-
0 0 0
tion ðT i; j ; ð0:2T i; j Þ2 Þ, with T i; j defined in Eq. 36, where k ¼ 1,000 md is a random guess of the average absolute permeability of the
entire reservoir, and lo ¼ 20 cp and /0 ¼ 0.2 are true values. Here, we use the same true relative permeability curves as in the first
example, but the three relative permeability parameters are distributed as a ða; 0:052 Þ; nw ðn w ; 0:12 Þ, and no ðn o ; 0:12 Þ,
with the mean values given by a ¼ 0:75; n w ¼ 1:3; and n o ¼ 1:7. Again, the values of Siw and Sor are assumed to be known. We assume
a single global set of relative permeabilities that are shared by all volume connections.
In this example, an initial ensemble of 300 realizations of the INSIM-FT model parameters is generated by sampling the aforemen-
tioned Gaussian distributions. The prior and history-matched realizations of relative permeabilities are shown in Figs. 20a and 20b,
respectively. The oil-production rates calculated with INSIM-FT using the multiple prior realizations of model parameters are shown in
Fig. 21. Fig. 22 compares oil-production rates calculated with the CRM, INSIM, and INSIM-FT history-matched models for four wells

(P1 through P4). In this case, the original INSIM algorithm gives a good history match of oil-rate data at wells P1 and P2, but the rela-
tively poor data match and predictions obtained for wells P3 and P4 are more representative of the overall INSIM results, as indicated
by the computational results shown in Table 8. Note that Table 8 includes both INSIM and INSIM (ES-MDA) results, and we see that
for this example, the INSIM (ES-MDA) data matches are significantly worse than those obtained with INSIM-FT. In fact, unlike the
Example 1 results, in this example, the INSIM (ES-MDA) history match is even inferior to the one obtained with CRM. Table 9
presents results with the same example but with the history-matching period shortened to 400 days. It is important to note that the
results of Tables 8 and 9 indicate that the performance of the original INSIM algorithm was negatively affected not only by the use of
an interior-point optimizer, but also by the less-reliable incorporation of the physics of flow and transport than is incorporated in
INSIM-FT. As in Example 1 (Tables 4 and 5), INSIM-FT gives a better data match than the other methods. In Example 2, the prediction
errors obtained using INSIM-FT for history matching are significantly less than those generated with the CRM history-matched model.
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Sw Sw
(a) Prior (b) Posterior
Fig. 20—Oil/water relative permeability curves, channelized reservoir. The red solid lines represent the true relative permeability
curves, and the blue lines represent the estimated relative permeabilities. (a) Prior; (b) posterior.
500 800
400
WOPR:P1 (RB/D)
WOPR:P2 (RB/D)
600
300
400
200
200
100
0 0
0 500 1,000 0 500 1,000
1,500 1,500
WOPR:P3 (RB/D)
WOPR:P4 (RB/D)
1,000 1,000
500 500
0 0
0 500 1,000 0 500 1,000
Fig. 21—Prior oil-production rate; Example 2 is shown, a channelized reservoir. Red circles are observed data; red lines are
true data of oil-production rate; gray lines are prior responses; and vertical black dashed lines are separators for matching
and prediction.
Figs. 23a, 23b, 23c, and 23d show the interwell connectivities obtained from FrontSim, CRM, INSIM, and INSIM-FT, respectively.
Compared with the connectivity generated with CRM (Fig. 23b) and INSIM (Fig. 23c), the connectivity calculated from the INSIM-FT
history-matched results in Fig. 23d is in better agreement with the FrontSim connectivity in Fig. 23a. In Fig. 23d, except for injector I4,
the INSIM-FT results agree well with the connectivity map generated with FrontSim dependent on the true Eclipse reservoir-simulation
model. For this example, the connectivity results from INSIM match those from FrontSim far better than those obtained from INSIM
(ES-MDA), and because of this the INSIM (ES-MDA) connectivity results are not shown.

500 500 500
400 400 400

WOPR:P1 (RB/D)
WOPR:P1 (RB/D)
WOPR:P1 (RB/D)
300 300 300
200 200 200
100 100 100
0 0 0
0 500 1,000 0 500 1,000 0 500 1,000
(a) CRM:P1 (b) INSIM:P1 (c) INSIM-FT:P1
800 800 1,000
800
WOPR:P2 (RB/D)
600 WOPR:P2 (RB/D) 600
WOPR:P2 (RB/D)
600
400 400
400
200 200
200
0 0 0
0 500 1,000 0 500 1,000 0 500 1,000
(d) CRM:P2 (e) INSIM:P2 (f) INSIM-FT:P2
1,000 1,500 1,500
800
WOPR:P3 (RB/D)
WOPR:P3 (RB/D)
1,000 WOPR:P3 (RB/D) 1,000

600
400
500 500
200
0 0 0
0 500 1,000 0 500 1,000 0 500 1,000
(g) CRM:P3 (h) INSIM:P3 (i) INSIM-FT:P3
1,200 1,000 1,000
1,000
800 800
WOPR:P4 (RB/D)
WOPR:P4 (RB/D)
WOPR:P4 (RB/D)
800
600 600
600
400 400
400
200 200
200
0 0 0
0 500 1,000 0 500 1,000 0 500 1,000
(j) CRM:P4 (k) INSIM:P4 (l) INSIM-FT:P4
Fig. 22—Comparison of oil-production rates calculated with the CRM, INSIM, and INSIM-FT history-matched models for four wells;
Example 2 is shown, a channelized reservoir. Red circles are observed data; red lines are true data of oil-production rate; and the
gray curve is the posterior oil-production rate. (a) CRM for P1; (b) INSIM for P1; (c) INSIM-FT for P1; (d) CRM for P2; (e) INSIM for P2;
(f) INSIM-FT for P2; (g) CRM for P3; (h) INSIM for P3; (i) INSIM-FT for P3; (j) CRM for P4; (k) INSIM for P4; (l) INSIM-FT for P4.


Historical period 23.28 313 10.11 62.1
Prediction period 115.87 914 37.14 185
Table 8—Comparison of data mismatch for the historical period (800 days) and the prediction period
among the CRM, INSIM, INSIM-FT, and INSIM (ES-MDA) models; Example 2, channelized reservoir.

Prediction period 137.4 2,008.4 71.7 98.4
Table 9—Comparison of data mismatch for the historical period (400 days) and the prediction period
among the CRM, INSIM, INSIM-FT, and INSIM (ES-MDA) models; Example 2, channelized reservoir.
P1 P2 P3 P1 P2 P3
922.35 352.1 733.25

283.05
I1 I2 I1 I2
605.15 360.9
P4 P5 P6 P4 P5 P6
759.0 386.85 701.6 309.2
298.3 355.95
I3 I4 I3 I4
194.0
200.25 231.65
P7 P8 P9 P7 P8 P9
(a) FrontSim (b) CRM
P1 P2 P3 P1 oIm1 P2 oIm2 P3
951.2 424.7 820.75
I1 I2 oIIm3 I1 oIm
m4 I2 oIm
I 5
435.6 281.3 481.8
827.0 477.6 786.6 588.7
I3 I4 oIm8 I3 oIm9 I4 oIm
m10
288.8 148.9 163.1
(c) INSIM (d) INSIM-FT
Fig. 23—Interwell connectivity obtained from FrontSim, CRM, INSIM, and INSIM-FT for the channelized reservoir. The length of a nar-
row red triangle denotes the magnitude of total-interwell-flow rate between a corresponding injector/producer pair, and the direction
of the triangle indicates which producer belongs to this injector/producer well pair. (a) FrontSim; (b) CRM; (c) INSIM; (d) INSIM-FT.
Example 3: Field Example. This example pertains to a true reservoir with strong bottomwater drive. The reservoir has 13 producing
wells and no injection wells. For the purposes of history matching, we assume the aquifer is at constant pressure and represent the aqui-
fer by a single node with the pressure at this node fixed equal to the constant pressure ( p0) of the aquifer, which is a procedure sug-
gested by Zhao et al. (2016a) for the original INSIM model. The value of p0 is set equal to the value of the initial reservoir pressure at
datum and does not change with time. The reservoir-connection map is shown in Fig. 24. Note that there are 26 imaginary wells and
these imaginary wells, as well as the actual wells, are all connected to the single aquifer node. There are no water-injection wells in the
reservoir, but it seems that CRM could be applied to this problem by treating the aquifer node as a water-injection well. However,
CRM requires that the total-water-injection rates are known, and we have no knowledge of the flow rate from the aquifer to the reser-
voir. Note that CRM cannot be applied in this example because we do not know the “injection rate” for the aquifer node.
For our implementation of INSIM-FT, the spatial coordinates of the virtual well are computed as the average of the corresponding
coordinates of the real wells:
1 X
nw
1 X
nw
xv ¼ xi ; yv ¼ yi ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð37Þ
nw i nw i
where xv and yv represent the x- and y-coordinate, respectively, of the virtual well node and nw is the number of real wells. If well i is
connected to this virtual node, the connection length is given by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Li;v ¼ ðxv xi Þ2 þ ðyv yi Þ2 ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð38Þ

where xi and yi represent the x- and y-coordinate of well i. The virtual aquifer node functions the same way as an injection-well node.
Parameters of the connections between the virtual well node and the real or imaginary wells are added to the set of parameters to be
estimated by history matching. By tuning the PVs and transmissibilities of these connections, the water-breakthrough time at a produc-
ing well can be adjusted and is not determined by the distance between the well in the basic connection map and the virtual well node
that represents the aquifer. When history matching the field data, we assume that all connections share a global set of relative perme-
abilities, and using available data, the true values of parameters of the power-law relative permeability curves are estimated as
atrue ¼ 0:14; nw;true ¼ 3:0; no;true ¼ 4:0, Siw;true ¼ 0:15, and Sor;true ¼ 0:33. In history matching, the values of Siw and Sor are fixed at the
estimated true values and the other three relative permeability parameters are included in the set of model parameters used for history
matching. In generating initial realizations of these parameters, it is assumed that they follow the normal distributions given by
a ða; 0:052 Þ, nw ðn w ; 0:22 Þ, and no ðn o ; 0:22 Þ, where the mean values are different from the estimated true values to test
the robustness of INSIM-FT and are given by a ¼ 0:16; n w ¼ 2:5, and n o ¼ 3:3.
W1
oIm1
W2
oIm5 oIm3
oIm4 oIm2 W3
oIm26 oIm14
W6 W4
oIm16
W5
oIm6 oIm9 oIm23
oIm21 oIm12
W7
oIm15
oIm8 oIm13
oIm7
W9
oIm17
W8
oIm20
oIm19
oIm18 oIm22
oIm10 W13
W11 oIm25
W10 oIm24
W12
Fig. 24—Connection map of the field example.
The initial guess of the average reservoir absolute permeability is set equal to 100 md, which is an estimate given by field engineers.
The prior ensemble of transmissibilities and PVs of different connections are generated in the same way as for the two synthetic exam-
0
ples. It should be noted that the total PV of the reservoir at time zero, Vp;tot , is also tuned as a history-matching parameter for this exam-
0 0
ple. The prior value of Vp;tot , which is denoted by V p;tot , is an estimate of the initial total PV supplied by the operator. The prior
0 0
0
ensemble of Vp;tot is generated from the normal distribution ðV p;tot ; ð0:2V p;tot Þ2 Þ. The lower bound of Vp;tot0 0
; which is Vp;tot;low , is set to
0 0 0 0
one-half of the prior value of Vp;tot and the upper bound of Vp;tot ; which is Vp;tot;up , is set to 150% of the prior value of Vp;tot . The ensem-
ble size for ES-MDA is 300. The first 800 days of production data are history matched, and the following 200 days are used for predic-
tions to ascertain whether the history-matched INSIM-FT model is capable of providing reasonable predictions of future reservoir
performance. All the producing wells are operated under the specified total-liquid-production rates that vary with time. The observed
oil-production rates are the observed data to be history matched.
Representative oil-production rates obtained from the prior ensemble of realizations of INSIM-FT models are shown in Fig. 25,
whereas Figs. 26a through 26d shows both the history-matching results and subsequent predictions calculated with INSIM-FT using
the history-matched ensemble of models. The prior and posterior relative permeabilities are shown in Figs. 27a and 27b. The results
suggest that the posterior relative permeabilities bound the true ones and are consistent with the production data. Although not shown
0 0
here, the history-matched values of Vp;tot also bound V p;tot with a narrower uncertainty range compared with that of the prior ensemble
0
of Vp;tot values. Overall, the history-matched results are remarkably good. For comparison purposes, the INSIM (ES-MDA) history-
matched oil rates at four wells are also shown in Figs. 26e through 26h. Although a careful examination of the results indicates that bet-
ter matches and predictions are obtained with INSIM-FT than with INSIM (ES-MDA), this point can be seen more clearly by examin-
ing the history-matching and prediction errors given in Table 10, which shows that the history-matching and predictions errors are less
than one-half the corresponding errors obtained with INSIM (ES-MDA).
Conclusions
The original data-driven model INSIM has two primary deficiencies. It assumes relative permeabilities are known a priori and uses a
method for calculation of saturations that is correct only when the Riemann solution of the standard Buckley equation is valid, which is
not always the case. By replacing the saturation calculation in the original data-driven model with a front-tracking procedure derived
from the semianalytical solution of Riemann and Cauchy problems and incorporating the parameters of power-law relative permeabil-
ities as additional model parameters, we remove the two main deficiencies of INSIM and obtain a modified data-driven model known as
INSIM-FT. Also, unlike INSIM, INSIM-FT adds additional flow paths (streamtubes) by means of imaginary wells (additional nodes) to
increase fidelity. Adding imaginary wells, however, required the development of the new method to determine the interwell connectivity,
which was presented in this work. In addition, we showed that when the interior-point optimizer used for history matching in the original
INSIM algorithm is replaced with ES-MDA to obtain an modified procedure referred to as INSIM (ES-MDA), the history-matching
errors decreased but were still significantly larger, in most cases, than the errors obtained from INSIM-FT. However, the connectivity
values estimated with INSIM (ES-MDA) were far better than those obtained from INSIM in Example 1 but far worse in Example 2.

1,200 1,200
1,000 1,000
WOPR:W11 (RB/D)
WOPR:W8 (RB/D)
800 800
600 600
400 400
200 200
0 0
0 1,000 2,000 0 1,000 2,000
1,200
1,200
1,000
WOPR:W13 (RB/D)
WOPR:W12 (RB/D)
1,000
800
800
600
600
400 400
200 200
0 0
0 1,000 2,000 0 1,000 2,000
Fig. 25—Prior oil-production rates of the field example. Red circles are observed oil-production rates; gray lines are prior
responses of the oil-production rates obtained from the prior INSIM-FT models; and vertical black dashed lines are separators for
matching and predictions.
1,200 1,200 1,200

1,200
WOPR:W11 (RB/D)
WOPR:W12 (RB/D)
WOPR:W13 (RB/D)
WOPR:W8 (RB/D)
1,000 1,000 1,000

1,000
800 800 800 800
600 600 600 600
400 400 400 400
200 200 200 200

0 0 0 0
0 1,000 2,000 0 1,000 2,000 0 1,000 2,000 0 1,000 2,000
Time (days) Time (days) Time (days) Time (days)
(a) INSIM-FT:W8 (b) INSIM-FT:W11 (c) INSIM-FT:W12 (d) INSIM-FT:W13
1,200 1,200 1,200

1,200
WOPR:W13 (RB/D)
WOPR:W12 (RB/D)
WOPR:W11 (RB/D)
WOPR:W8 (RB/D)
1,000 1,000 1,000

1,000
800 800 800
800
600 600 600 600
400 400 400 400
200 200 200 200

0 0 0 0
0 1,000 2,000 0 1,000 2,000 0 1,000 2,000 0 1,000 2,000
Time (days) Time (days) Time (days) Time (days)
(e) INSIM:W8 (f) INSIM:W11 (g) INSIM:W12 (h) INSIM:W13
Fig. 26—History-matched oil-production rates of the field example. Red circles are observed oil rates; gray lines are the oil rates
estimated with the history-matched INSIM-FT or INSIM models; and vertical black dashed lines are separators for matching and
predictions. (a) INSIM-FT for W8; (b) INSIM-FT for W11; (c) INSIM-FT for W12; (d) INSIM-FT for W13; (e) INSIM for W8; (f) INSIM for
W11; (g) INSIM for W12; (h) INSIM for W13.
We consider two synthetic examples where we use a reservoir-simulation model to represent the true reservoir to generate produc-
tion data for history matching as well as predictions of future reservoir performance. The computational results indicate that INSIM-FT
gives better history matches, lower prediction errors, and better estimates of reservoir connectivity than are obtainable with the original
INSIM, where history matching is performed with a MATLAB interior-point optimizer or with INSIM (ES-MDA), which uses ES-
MDA for history matching. This is the expected result dependent on the fact that INSIM-FT incorporates more of the correct physics of
flow than does INSIM. The results demonstrate that INSIM-FT can detect the presence of a fault, whereas INSIM cannot.

1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2
0.2
0
0
0.2 0.3 0.4 0.5 0.6
0.2 0.3 0.4 0.5 0.6
Sw Sw
(a) Prior (b) Posterior
Fig. 27—Estimated relative permeabilities, field example. The red lines denote the true relative permeability curves, and the blue
lines are the estimates of the relative permeability curves. (a) Prior; (b) posterior.
INSIM INSIM-FT
Historical period 115.5 41.2
Prediction period 213.8 96.6
Table 10—The comparison of data mismatch for the historical period

and the prediction period between INSIM (ES-MDA) and INSIM-FT;
field example.
We also implement a modern CRM and compare its performance with INSIM-FT. Overall, INSIM-FT outperforms CRM in terms
of the quality of the history match and determining reservoir connectivity. For the first example, however, the prediction error of CRM
is almost zero, and is much lower than the history-matching error. This CRM result appears to be an anomaly because we expect the
history-matching error to be smaller than the prediction error, which is always the case for the INSIM, INSIM (ES-MDA), and INSIM-
FT results obtained, and is also the case for the CRM results obtained for the second example.
The synthetic examples presented encompass only waterflooding. The field example pertains to an example where bottomwater
drive is a predominant reservoir mechanism and there are no water-injection wells. However, we speculate that INSIM-FT can be
applied for history matching, prediction, and characterization of a naturally fractured reservoir under waterflooding. Currently, we are
attempting to extend the INSIM-FT model to three dimensions with gravity effects. The primary objectives of this extension are to
allow for a gas cap and to enable the inclusion of slanted and horizontal wells.
Nomenclature
a ¼ endpoint relative permeability of water phase
A ¼ cross-sectional area, ft2
c ¼ compressibility, psi1
CD ¼ covariance matrix for measurement error
d ¼ vector of data response
dobs ¼ vector of observed data
fi,j ¼ allocation factor for well connection (i, j)
fo ¼ oil fractional flow
fw ¼ water fractional flow
g ¼ nonlinear function to evaluate the data response
L ¼ connection length, ft
k ¼ absolute permeability, md
kr ¼ relative permeability
m ¼ vector of model parameters
nc,i ¼ number of wells that are connected to well i
no ¼ exponent of power-law relative permeability for oil phase
nw ¼ exponent of power-law relative permeability for water phase
Nc,i ¼ number of wells that are connected to well i and are also upstream of well i
Nd ¼ number of observed data
Npro ¼ number of producers
Nw ¼ number of wells (including both actual and imaginary wells)
¼ normal distribution
O ¼ objective function for history matching
p ¼ pressure, psi
q ¼ flow rate, RB/D
Siw ¼ irreducible water saturation
So ¼ oil saturation
Sor ¼ residual oil saturation

Sw ¼ water saturation
Sw0 ¼ water saturation at inflection point of S-shaped fractional flow curve
Swn ¼ normalized water saturation
Sw* ¼ post-shock water saturation
t ¼ time
T ¼ transmissibility
Vp ¼ pore volume, ft3
g ¼ constant defined in Eq. A-1
kt ¼ total mobility for oil/water system, cp1
l ¼ fluid viscosity, cp
r ¼ shock speed
/ ¼ porosity
Subscripts
l ¼ left state
low ¼ lower bound
o ¼ oil phase
r ¼ rock or right state
R ¼ right
t ¼ total
true ¼ true value
tot ¼ total in reservoir scale
up ¼ upper bound
w ¼ water phase
Superscripts
0 ¼ initial condition
n ¼ time level
u ¼ upstream
Acknowledgments
The support of the member companies of the University of Tulsa Petroleum Reservoir Exploitation Projects is gratefully acknowledged.
Author Hui Zhao would like to express his gratitude for the support of the National Natural Science Foundation of China (Grant No.
51674039) and the China Important National Science & Technology Specific Projects (Grant No. 2016ZX05014). The authors wish to
thank Ying Li of China University of Geosciences (Beijing) for the INSIM (ES-MDA) history match for the field case.
References
Albertoni, A. and Lake, L. W. 2003. Inferring Interwell Connectivity Only From Well-Rate Fluctuations in Waterfloods. SPE Res Eval & Eng 6 (1):
6–16. SPE-83381-PA. https://doi.org/10.2118/83381-PA.
Cao, F. 2014. Development of a Two-Phase Flow Coupled Capacitance Resistance Model. PhD dissertation, University of Texas at Austin,
Austin, Texas.
Cao, F., Luo, H., and Lake, L. W. 2014. Development of a Fully Coupled Two-phase Flow Based Capacitance Resistance Model (CRM). Presented at
the SPE Improved Oil Recovery Symposium, Tulsa, 12–16 April. SPE-169485-MS. https://doi.org/10.2118/169485-MS.
Cao, F., Luo, H., and Lake, L. W. 2015. Oil-Rate Forecast by Inferring Fractional-Flow Models From Field Data With Koval Method Combined With
the Capacitance/Resistance Model. SPE Res Eval & Eng 18 (4): 534–553. SPE-173315-PA. https://doi.org/10.2118/173315-PA.
Cardoso, M. A. and Durlofsky, L. J. 2010. Use of Reduced-Order Modeling Procedures for Production Optimization. SPE J. 15 (2): 426–435. SPE-
119057-PA. https://doi.org/10.2118/119057-PA.
Emerick, A. A. and Reynolds, A. C. 2012. History Matching Time-Lapse Seismic Data Using the Ensemble Kalman Filter with Multiple Data Assimila-
tions. Computat. Geosci. 16 (3): 639–659. https://doi.org/10.1007/s10596-012-9275-5.
Emerick, A. A. and Reynolds, A. C. 2013a. Ensemble Smoother with Multiple Data Assimilations. Comput. Geosci. 55 (June): 3–15. https://doi.org/
10.1016/j.cageo.2012.03.011.
Emerick, A. A. and Reynolds, A. C. 2013b. History-Matching Production and Seismic Data in a Real Field Case Using the Ensemble Smoother With
Multiple Data Assimilation. Presented at the SPE Reservoir Simulation Symposium, The Woodlands, Texas, 18–20 February. SPE-163675-MS.
https://doi.org/10.2118/163675-MS.
Gentil, P. H. 2005. The Use of Multilinear Regression Models in Patterned Waterfloods: Physical Meaning of the Regression Coefficients. Master’s the-
sis, University of Texas at Austin, Austin, Texas.
Gildin, E., Ghasemi, M., Romanovskay, A. et al. 2013. Nonlinear Complexity Reduction for Fast Simulation of Flow in Heterogeneous Porous Media. Pre-
sented at the SPE Reservoir Simulation Symposium, The Woodlands, Texas, 18–20 February. SPE-163618-MS. https://doi.org/10.2118/163618-MS.
He, J. and Durlofsky, L. J. 2014. Reduced-Order Modeling for Compositional Simulation by Use of Trajectory Piecewise Linearization. SPE J. 19 (5):
858–872. SPE-163634-PA. https://doi.org/10.2118/163634-PA.
Heffer, K. J., Fox, R. J., McGill, C. A. et al. 1997. Novel Techniques Show Links between Reservoir Flow Directionality, Earth Stress, Fault Structure
and Geomechanical Changes in Mature Waterfloods. SPE J. 2 (2): 91–98. SPE-30711-PA. https://doi.org/10.2118/30711-PA.
Holden, H., Holden, L., and Høegh-Krohn, R. 1988. A Numerical Method for First Order Nonlinear Scalar Conservation Laws in One-Dimension. Com-
put. Math. Appl. 15 (6–8): 595–602. https://doi.org/10.1016/0898-1221(88)90282-9.
Iglesias, M. A. and Dawson, C. 2013. The Regularizing Levenberg–Marquardt Scheme for History Matching of Petroleum Reservoirs. Computat. Geo-
sci. 17 (6): 1033–1053. https://doi.org/10.1007/s10596-013-9373-z.
Jansen, F. E. and Kelkar, M. G. 1997. Non-Stationary Estimation of Reservoir Properties Using Production Data. Presented at the SPE Annual Technical
Conference and Exhibition, San Antonio, Texas, 5–8 October. SPE-38729-MS. https://doi.org/10.2118/38729-MS.

Juanes, R. and Patzek, T. W. 2003. Multiscale Numerical Modeling of Three-Phase Flow. Presented at the SPE Annual Technical Conference and Exhi-
bition, Denver, 5–8 October. SPE-84369-MS. https://doi.org/10.2118/84369-MS.
Lake, L. W., Liang, X., Edgar, T. F. et al. 2007. Optimization of Oil Production Based on a Capacitance Model of Production and Injection Rates. Pre-
sented at the SPE Hydrocarbon Economics and Evaluation Symposium, Dallas, 1–3 April. SPE-107713-MS. https://doi.org/10.2118/107713-MS.
Lerlertpakdee, P., Jafarpour, B., and Gildin, E. 2014. Efficient Production Optimization With Flow-Network Models. SPE J. 19 (6): 1083–1095. SPE-
170241-PA. https://doi.org/10.2118/170241-PA.
Lie, K. A. and Juanes, R. 2005. A Front-Tracking Method for the Simulation of Three-Phase Flow in Porous Media. Computat. Geosci. 9 (1): 29–59.
https://doi.org/10.1007/s10596-005-5663-4.
MathWorks. 2011. MATLAB Version 2011 User Guide. Natick, Massachusetts: MathWorks.
Nguyen, A. P. 2012. Capacitance Resistance Modeling for Primary Recovery, Waterflood and Water-CO2 Flood. PhD dissertation, University of Texas
at Austin, Austin, Texas.
Oleinik, O. A. 1957. Discontinuous Solutions of Non-Linear Differential Equations. Uspekhi Mat. Nauk 12 (3): 3–73.
Refunjol, B. T. 1996. Reservoir Characterization of North Buck Draw Field Based on Tracer Response and Production/Injection Analysis. Master’s the-
sis, University of Texas at Austin, Austin, Texas.
Reynolds, A. C., Zafari, M., and Li, G. 2006. Iterative Forms of the Ensemble Kalman Filter. Oral presentation given at ECMOR X–10th European Con-
ference on the Mathematics of Oil Recovery, Amsterdam, 4–7 September.
Sayarpour, M. 2008. Development and Application of Capacitance-Resistive Models to Water/CO2 Flood. PhD dissertation, University of Texas at Aus-
tin, Austin, Texas.
Schlumberger. 2013a. Eclipse Version 2013.1 Reference Manual.
Schlumberger. 2013b. FrontSim Version 2013.1 User Guide.
Tarantola, A. 2005. Inverse Problem Theory and Methods for Model Parameter Estimation. Philadelphia, Pennsylvania: Society of Industrial and
Applied Mathematics.
Thiele, M. R. and Batycky, R. P. 2006. Using Streamline-Derived Injection Efficiencies for Improved Waterflood Management. SPE Res Eval & Eng 9
(2): 187–196. SPE-84080-PA. https://doi.org/10.2118/84080-PA.
Thiele, M. R., Batycky, R. P., Blunt, M. J. et al. 1996. Simulating Flow in Heterogeneous Systems Using Streamtubes and Streamlines. SPE Res Eval &
Eng 11 (1): 5–12. SPE-27834-PA. https://doi.org/10.2118/27834-PA.
van Doren, J. F. M., Markovinović, R., and Jansen, J. D. 2006. Reduced-Order Optimal Control of Water Flooding Using Proper Orthogonal Decomposi-
tion, Computat. Geosci. 10 (1): 137–158. https://doi.org/10.1007/s10596-005-9014-2.
Vogel, C. R. 2002. Computational Methods for Inverse Problems. Philadelphia, Pennsylvania: Society of Industrial and Applied Mathematics.
Weber, D. B. 2009. The Use of Capacitance-Resistance Models to Optimize Injection Allocation and Well Location in Water Floods. PhD dissertation,
University of Texas at Austin, Austin, Texas.
Yousef, A. A., Gentil, P. H., Jensen, J. L. et al. 2005. A Capacitance Model To Infer Interwell Connectivity From Production and Injection Rate
Fluctuations. Presented at the SPE Annual Technical Conference and Exhibition, Dallas, 9–12 October. SPE-95322-MS. https://doi.org/10.2118/
95322-MS.
Zhao, H., Kang, Z., Zhang, X. et al. 2015. INSIM: A Data-Driven Model for History Matching and Prediction for Waterflooding Monitoring and Man-
agement with a Field Application. Presented at the SPE Reservoir Simulation Symposium, Houston, 23–25 February. SPE-173213-MS. https://
doi.org/10.2118/173213-MS.
Zhao, H., Kang, Z., Zhang, X. et al. 2016a. A Physics-Based Data-Driven Numerical Model for Reservoir History Matching and Prediction With a Field
Application. SPE J. 21 (6): 2175–2194. SPE-173213-PA. https://doi.org/10.2118/173213-PA.
Zhao, H., Li, Y., Cui, S. et al. 2016b. History Matching and Production Optimization of Water Flooding Based on a Data-Driven Interwell Numerical
Simulation Model. J. Nat. Gas Sci. Eng. 31 (April): 48–66. https://doi.org/10.1016/j.jngse.2016.02.043.
Appendix A—Derivation of the Front-Tracking Method

The Riemann problem (Juanes and Patzek 2003) consists of finding a (weak) solution to the initial value problem:

@Sw @fw Swl if x < x0
þg ¼ 0; Sw ðx; 0Þ ¼ ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-1Þ
@t @x Swr if x > x0
where g is a constant and the partial-differential equation represents a conservation law. Assuming that Swl = Swr, there is a single dis-
continuity in the initial saturation profile at x0. The solution to the problem is a wave connecting the left (Swl) and right (Swr) states. For
the problems considered here, we neglect gravity and capillarity, and assume the fractional-flow curve is S-shaped with a convex part
connected to a concave part. The characteristic wave may involve three possible wave types: a single rarefaction wave, a single shock
wave, and a composite rarefaction/shock wave (Juanes and Patzek 2003).
Single Shock. A shock is a discontinuous wave existing in the form of a weak solution. A single shock might develop when two char-
acteristics intersect and the characteristic speed of the left state is greater than that of the right state, which is the necessary condition
for a single shock to develop. Given two states, Swl and Swr, the trial shock speed r is determined by the Rankine-Hugoniot condition:
fw ðSwl Þ fw ðSwr Þ
rtrial ¼ g : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-2Þ
Swl Swr
To guarantee that the shock wave with speed of rtrial is admissible, the Oleinik (1957) entropy condition must be satisfied for all Sw
between Swl and Swr:
fw ðSw Þ fw ðSwl Þ fw ðSwr Þ fw ðSwl Þ fw ðSwr Þ fw ðSw Þ
: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-3Þ
Sw Swl Swr Swl Swr Sw
For our problems with only one inflection point in fw(Sw), there is an alternative entropy condition that is weaker than the condition
of Eq. A-3 but also is sufficient to ensure that the solution of Eq. A-1 is a single shock; that is, if the given Swl and Swr satisfy
fw0 ðSwl Þ rtrial fw0 ðSwr Þ: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-4Þ

Then there is a single shock solution of Eq. A-1, which is given by

8
> x x0
< Swl ; < rtrial
Sw ðx; tÞ ¼ t ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-5Þ
> x x0
: Swr ; > rtrial
t
where x0 is the location of the single discontinuity in the initial condition.
Single Rarefaction. A single rarefaction wave is a smooth function Sw(x,t) that connects Swl and Swr. It develops when fw0 increases
monotonically from the left state to the right state, and for an S-shaped fw curve with a single inflection point Sw0, a necessary and suffi-
cient condition for the saturation to be a rarefaction wave is given by
Swl > Swr Sw0 or Swl < Swr Sw0 ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-6Þ
where Sw0 is the inflection point of fw(Sw). If we know the analytical form of fw00 , it is convenient to replace Eq. A-6 with the simpler con-
dition given by
fw00 ðSwl Þ fw00 ðSwr Þ 0 and fw0 ðSwl Þ < fw0 ðSwr Þ: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-7Þ
The closed-form solution of a single rarefaction is given by

8 x x0
>
> Swl ; < g fw0 ðSwl Þ
>
> t
< x x0 x x0
Sw ðx; tÞ ¼ ð f 0w Þ1 ; g fw0 ðSwl Þ < < g fw0 ðSwr Þ : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-8Þ
>
> gt t
>
> x x0
: Swr ; > g fw0 ðSwr Þ
t
Composite Rarefaction/Shock. When the left and right states are in different convexity regions of fw(Sw), the solution can involve one
rarefaction wave and one shock. In this situation, a single shock and a single rarefaction connect at some intermediate point Sw* called
the post-shock value (Juanes and Patzek 2003). This is the value of Sw at which the left characteristic speed (rarefaction wave) equals
the speed of the right discontinuity (shock); i.e.,
fw ðSw Þ fw ðSwr Þ
r ¼ g fw0 ðSw Þ ¼ g ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-9Þ
Sw Swr
where r* is the trial shock speed. After obtaining the post-shock value of Sw*, we check the Oleinik (1957) entropy condition in Eq. A-10
to determine whether Sw* to Swr is an admissible shock. The shock is admissible if for all Sw between Swr and Sw*,
fw ðSw Þ fw ðSw Þ fw ðSwr Þ fw ðSw Þ fw ðSwr Þ fw ðSw Þ
: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-10Þ
Sw Sw Swr Sw Swr Sw
If the condition of Eq. A-10 is violated, then the Riemann solution must not be the composite rarefaction/shock wave. Otherwise, to
confirm that the solution is the composite wave, we have to verify that there is an admissible rarefaction wave behind the admissible
shock; i.e., f 0 ðSw Þ is monotonically increasing from Swl to Sw*. The closed form of the composite rarefaction/shock is given by
8 x x0
>
> Swl ; < g fw0 ðSwl Þ
>
> t
< x x0 x x0
Sw ðx; tÞ ¼ ð f 0w Þ1 ; g fw0 ðSwl Þ < < g fw0 ðSw Þ : . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-11Þ
>
> gt t
>
> x x0
: Swr ; > g fw0 ðSw Þ
t
A simpler way to look at the problem is to recognize that there are only three types of possible solutions: a single shock wave, a sin-
gle rarefaction wave, and a composite solution consisting of a shock wave and a rarefaction wave. If we rule out the first and second sol-
utions, the only possibility is the third (composite) solution. In general, combining the three solution types, the complete algorithm to
solve the Riemann problem defined in Eq. A-1 is described in the following, which represents the solution obtained by the method of
characteristics (MOC).
Algorithm 1: Complete Algorithm To Solve Riemann Problem. The algorithm consists of the following steps.
1. Define left and right states: Swl, Swr. fw ðSwl Þ fw ðSwr Þ

2. Compute the trial shock speed from the Rankine-Hugoniot jump condition: rtrial ¼ .
0 0
3. IF fw ðSwl Þ rtrial fw ðSwr Þ 0, THEN Swl Swr
• S: Single Shock with speed r ¼ rtrial.
ELSE IF fw0 ðSwl Þ < fw0 ðSwr Þ and fw00 ðSwl Þfw00 ðSwr Þ > 0 THEN
• R: Single rarefaction,
ELSE
• RS: Composite rarefaction/shock. fw ðSw Þ fw ðSwr Þ
• Compute the post-shock value Sw*, by solving fw0 ðSw Þ ¼ .
END Sw Swr
4. END

If we use the power-law relative permeability curve, the fractional curve has an analytical form and the first- and second-order deriv-
atives can be easily obtained analytically. Therefore, the criteria to determine the solution types can be computed easily. Eq. A-9 is
solved by the Newton-Raphson method.
Front-Tracking Algorithm. The Riemann problem (Eq. A-1) is a simple form of a Cauchy problem with the initial condition given
by a step function with a single discontinuity. In the application of the INSIM model (Zhao et al. 2015), we often encounter the Cauchy
problem with nonuniform initial conditions given by
@Sw @fw
þg ¼ 0; Sw ðx; 0Þ ¼ Sw0 ðxÞ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-12Þ
@t @x
where the initial condition is not a piecewise constant function of x. If the function Sw0(x) is replaced by a combination of an infinite
number of constant states, we could divide the Cauchy problem into an infinite number of local Riemann problems, which can be solved
with the help of Algorithm 1. By connecting solutions of local Riemann problems, we can theoretically obtain the exact solution for the
problem of Eq. A-12. In practice, the initial condition should be discretized into a finite number of piecewise constant states where ev-
ery two neighboring constant states define a sub-Riemann problem. By connecting the solutions of all sub-Riemann problems, Eq. A-12
is properly solved. However, if the initial saturation of Eq. A-12 is the initial condition of a classic Buckley-Leverett problem, the dis-
cretization of the initial condition is no longer needed.
The front-tracking method presented here is similar to the one that was originally proposed by Holden et al. (1988) to generate the
approximate solution to the Cauchy problem of Eq. A-12. In the most general situations, we need to discretize the continuous initial
condition by a piecewise constant function, as shown in Fig. A-1: The Cauchy problem can be decoupled into a set of local Riemann
problems. Each Riemann problem shares a common constant state with its neighbor. We can use the Riemann solver given in Algorithm
1 to solve these local Riemann problems to obtain the exact MOC solutions. As shown in Fig. A-2, because the Riemann solutions that
involve a rarefaction wave are spreading into a fan shape in the x–t diagram, the Riemann solutions (including the single-shock case) of
one local Riemann problem are also referred to as a Riemann fan (Lie and Juanes 2005). These Riemann fans as the solutions of a series
of local Riemann problems are well-defined until the characteristics of two neighboring Riemann fans intersect each other in the x–t
diagram. The intersection is also referred to as a collision.
Sw0
x 01
x 02 x 03
x
Fig. A-1—Approximation of the initial condition. Here the initial condition of Sw is a continuous function of x. We approximate it by
a piecewise constant function with four constant states, which produce three local Riemann problems between each two neighbor-
ing constant states.
Δt
(x 1,t 1)
0
0 x 01 x 02 x x 03 L
Fig. A-2—Solution waves of three local Riemann problems shown in Fig. A-1. Each of the dash lines represents a characteristic in
the corresponding rarefaction. Each of the solid lines represents a shock. The solutions of all three Riemann problems are the
composite rarefaction/shock waves. These three Riemann fans are valid until the shock in the second Riemann fan intersects with
the first characteristic in the third Riemann fan. A new Riemann problem generates at the intersection point (x1,t1).

The collisions of Riemann fans often yield a curved shock path in the x–t diagram, where the shock speed is changing as propagation
continues. The computation of a curved shock path is computationally expensive because it is associated with numerical integration.
The common way to avoid the complication of directly computing the curved shock path is to eliminate the rarefaction waves (Holden
et al. 1988). In our application, the continuous rarefaction wave in each Riemann fan is approximated with a sequence of small shocks.
Each small shock is regulated to have a shock length smaller than a tolerance dSw to guarantee numerical accuracy. The shock speed is
computed with the Rankine-Hugoniot jump condition in Eq. A-2, even though the Oleinik entropy condition might be violated. Assum-
ing one rarefaction wave in a Riemann fan ranges between SwL and SwR, where SwL is in upstream of SwR, then given dSw, we calculate
jðSwL SwR Þj
the largest integer N, which satisfies N .
dSw
Then, the interval between SwL and SwR is equally divided into N subintervals, which yields a vector of water saturations
½Sw0 ; Sw1 ; …SwN ; where Sw0 ¼ SwL and SwN ¼ SwR. In this vector, every pair of two connected saturation Sw, j and Sw, j þ 1 represents an
approximate shock combining the saturation between Sw, j and Sw, j þ 1. By performing this approximation, the solutions to all local
Riemann problems are composed of a set of shocks, which are referred to as shock clusters. To adapt to the front-tracking algorithm,
the Riemann solver in Algorithm 1 is revised accordingly to obtain Algorithm 2.
Algorithm 2: Calculation of Shock Clusters. Algorithm 2 includes the following steps.
1. Given left and right states: Swl, Swr, dSw, fw ðSwl Þ fw ðSwr Þ
2. Compute the trial shock speed from the Rankine-Hugoniot jump condition: rtrial ¼ .
0 0
3. IF fw ðSwl Þ > rtrial > fw ðSwr Þ THEN Swl Swr
• S: Single Shock with speed r ¼ rtrial.
ELSE IF fw0 ðSwl Þ < fw0 ðSwr Þ and fw00 ðSwl Þfw00 ðSwr Þ > 0 THEN
• Use multiple small shocks to approximate the rarefaction wave that connects Swl and Swr.
ELSE
• Use multiple shocks to approximate the rarefaction wave that connects Swl and Sw*, which follows a real shock that combines
the saturation between Sw* and Swr. fw ðSw Þ fw ðSwr Þ
• Compute the post-shock value Sw*, by solving fw0 ðSw Þ ¼ .
END IF Sw Swr
4. END
Choosing a large value of dSw tends to cause a large material-balance error but requires less computational time to obtain solutions.
By use of computational experiments, to balance the error and computational efficiency, we choose dSw ¼ 0.01. After solving all the
local Riemann problems defined for the initial condition and approximating the rarefaction by small shocks, as shown in Fig. A-3, we
obtain a set of shock clusters, which propagate until the first collision occurs. It is easy to see that collisions always happen at the
boundary between two neighboring shock clusters. By computing all the collision times between any two neighboring shock clusters,
we can find the earliest collision. For example, in Fig. A-3, there are initially three shock clusters. After computing the collision time
between each two neighboring shock clusters, it is seen that the earliest collision occurs between the second shock and third shock clus-
ter at (x1,t1). At the particular collision, the new Riemann problem is given by

@Sw @Sw Swl ; x < x1
þ g fw0 ðSw Þ ¼ 0; t > t1 ; Sw ðx; t1 Þ ¼ 1 ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-13Þ
@t @x Swr ; x > x
where the left (Swl) and the right states (Swr) are given by the values immediately to the left and to the right of the colliding shocks.
Δt
(x 2,t 2)
(x 1,t 1)
0
0 x 01 x 02 x x 03 L
Fig. A-3—A case of three shock clusters.
After solving the newly generated Riemann problem with Algorithm 2, the new shock cluster is inserted at (x1,t1), as represented by
the dark green lines in Fig. A-4. Then the next collision is found and the associated new shock cluster is found and inserted at the colli-
sion point. The entire procedure is repeated until the earliest collision happens outside the time/space domain, fðx; tÞj0 x L; 0 <
t < Dtg; where in our application of INSIM-FT, L is the connection length of the well pair where we need to solve for the saturation
and Dt is the timestep used in pressure equations. Once the loop terminates, the saturation profile at the end of the timestep Dt can be
calculated easily. The saturation profile obtained always remains piecewise constant, which is an acceptable form of the initial condition
required for front tracking over the next timestep. The front-tracking algorithm is summarized in Algorithm 3.

Δt
(x 2,t 2)
(x 1,t 1)
0
0 x 10 x 02 x x 03 L
Fig. A-4—Insertion of the new shock cluster.
Algorithm 3: Front-Tracking Algorithm. Algorithm 3 includes the following steps.

1. Define the piecewise constant initial function:
8
>
> S0w1 0 < x < x01
>
>
<
S0w ¼ S0wi x0i1 < x < x0i :
>
>
>
>
:
S0wn x0n1 < x < L
2. Obtain solutions for all the local Riemann problems with Algorithm 2.
3. Compute the earliest collision.
4. Loop while the collision happens in fðx; tÞjð0 < x < L; 0 < t < DtÞg,
• Find the new Riemann solution and insert into the collision point.
• Compute the next collision.
END
5. Compute saturation profile at Dt and make it the new initial condition for the front tracking at the next timestep.
Zhenyu Guo is a PhD candidate in petroleum engineering in the Department of Petroleum Engineering at the University of Tulsa.
His research interests include data-driven models, machine learning, history matching, uncertainty quantification, and produc-
tion optimization. Guo holds a bachelor’s degree and a master’s degree in petroleum engineering from China University of Geo-
sciences, Beijing. He is a member of SPE.
Albert C. Reynolds is a professor of petroleum engineering and mathematics, holder of the McMan Chair in Petroleum Engineer-
ing, and director of the University of Tulsa Petroleum Reservoir Exploitation Projects at the University of Tulsa, where he has been a
faculty member since 1970. Reynolds’ research interests include optimization, scientific computation, assessment of uncertainty,
reservoir simulation, history matching, and well testing. He has authored or coauthored more than 100 refereed papers in these
areas. Reynolds holds a bachelor’s degree from the University of New Hampshire, a master’s degree from Case Institute of Tech-
nology, and a PhD degree from Case Western Reserve University, all in mathematics. He has received the SPE Distinguished
Achievement Award for Petroleum Engineering Faculty, the SPE Reservoir Description and Dynamics Award, the SPE Formation
Evaluation Award, and the SPE John Franklin Carll Award. Reynolds became an SPE Distinguished Member in 1999 and an SPE
Honorary Member in 2014. He is an associate editor for SPE Journal and for Computational Geosciences.
Hui Zhao is an associate professor in the Petroleum Engineering Department at Yangtze University, China, and was a visiting
scholar at the University of Tulsa. His research interests are history matching, optimization, interwell-connectivity modeling, and
nongradient-based optimization algorithms. Hui holds bachelor’s, master’s, and PhD degrees in petroleum engineering, all from
China University of Petroleum, East China. He is a member of the Reviewing Expertise Group of the National Natural Science
Foundation of China and a member of SPE.

Spe 182660 Pa

Uploaded by

Copyright:

Available Formats

Spe 182660 Pa

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Spe 182660 Pa

Uploaded by

Copyright:

Available Formats

J182660 DOI: 10.

2118/182660-PA Date: 23-March-18 Stage: Page: 367 Total Pages: 29

A Physics-Based Data-Driven Model for

April 2018 SPE Journal 367

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

CRM INSIM INSIM-FT

Table 1—Overall comparison among CRM, INSIM, and INSIM-FT.

368 April 2018 SPE Journal

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

cnt;i ¼ Sno;i co þ Snw;i cw þ cr ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð3Þ

April 2018 SPE Journal 369

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

where p0 is the initial reservoir pressure, and

@Sw ðx; tÞ 5:615qt;i; j ðtÞ @fw ðx; tÞ

Eq. 10 can be rewritten as

Sw ðx; tn1 Þ ¼ Sn1

Sw ðx; tn Þ ¼ Snw ðxÞ; for 0  x  Li; j ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð15Þ

along each pair of connected nodes.

krw ðSwn Þ ¼ a  Snwnw ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð16Þ

kro ðSwn Þ ¼ Snwno ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð17Þ

370 April 2018 SPE Journal

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

Snw;i ¼ fw1 ð f nw;i Þ: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð20Þ

Accordingly, the oil-production rate at producer i is estimated as

Fig. 2—The geometry for a three-connection reservoir.

April 2018 SPE Journal 371

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

INSIM-FT vs. INSIM

372 April 2018 SPE Journal

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

April 2018 SPE Journal 373

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

374 April 2018 SPE Journal

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

which ensures that

April 2018 SPE Journal 375

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

Fig. 5—Fault geology.

Table 3—Properties of the fault-segmented reservoir.

376 April 2018 SPE Journal

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

(a) CRM:P3 (b) CRM:P4

Fig. 8—Connection map generated without adding imaginary wells.

April 2018 SPE Journal 377

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

0.2 0.3 0.4 0.5 0.6 0.7 0.8

378 April 2018 SPE Journal

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

(a) No imaginary wells (b) Imaginary wells added

JI1 PJ33 I2J2

JI3 PJ44 I4J4

April 2018 SPE Journal 379

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

1,000 1,000 1,000

1,500 1,500 1,500

1,000 1,000 1,000

500 500 500

380 April 2018 SPE Journal

ID: jaganm Time: 10:37 I Path: S:/J###/Vol00000/170095/Comp/APPFile/SA-J###170095

Sw ðx; tn Þ ¼ Snw ðxÞ; for 0 x Li; j ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð15Þ

krw ðSwn Þ ¼ a Snwnw ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ð16Þ

fw0 ðSwl Þ rtrial fw0 ðSwr Þ: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-4Þ

Swl > Swr Sw0 or Swl < Swr Sw0 ; . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ðA-6Þ