SPE-180376-MS

Optimization of CO2-EOR Process in Partially Depleted Oil Reservoirs

W. Ampomah, R. S. Balch, R. B. Grigg, and M. Cather, Petroleum Recovery Research Center; R. A. Will and
S. Y. Lee, Schlumberger Carbon Services

Copyright 2016, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Western Regional Meeting held in Anchorage, Alaska, USA, 23–26 May 2016.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
This paper presents an optimization methodology for CO2 enhanced oil recovery in partially depleted
reservoirs. A field-scale compositional reservoir flow model was developed for assessing the performance
history of a CO2 flood and optimizing oil production and CO2 storage in the Farnsworth field unit (FWU),
Ochiltree County, Texas.
A geological framework model constructed from geophysical, geological and engineering data ac-
quired from FWU was used for the reservoir modeling. A laboratory fluid analysis was tuned to an
equation of state and subsequently used to predict the thermodynamic minimum miscible pressure
(MMP). An initial history calibration of primary, secondary and tertiary recovery are conducted as the
basis for the study.
After a good match was realized, an optimization model with proxy was constructed with an objective
function that maximized both oil recovery and CO2 storage. Experimental design was used to link
uncertain parameters to the objective function. A reduced order proxy model was necessary to reduce
computational cost. Control variables considered in this study included: CO2 purchase, recycled CO2,
water alternating gas cycle and ratio, infill wells and bottomhole pressure of injectors and producers.
Polynomial response surface methodology was used to create the proxy model based on training
simulations. This involved an iterative process until a validated surrogate model was achieved. A
sensitivity analysis was first conducted to ascertain which of these control variables to include in the
reduced order model. A genetic algorithm using a mixed-integer capability optimization approach was
employed to determine the optimum developmental strategy to maximize both oil recovery and CO2
storage.
The proxy model reduced the computational cost significantly. The validation of the reduced order
model ensured accuracy in the dynamic modeling results. The prediction outcome showed the robustness
and reliability of the genetic algorithm in optimizing oil recovery and CO2 storage.
The reservoir modeling approach used in this study showed an improved way of optimizing oil
production and CO2 storage within partially depleted oil reservoirs such as FWU. This study serves as a
benchmark for potential CO2–EOR projects in the Anadarko basin and/or geologically similar basins
throughout the world.
2 SPE-180376-MS

Introduction
Carbon dioxide enhanced oil recovery (CO2-EOR) is a mature technology shown to improve oil
production in many cases. Less often considered is that much CO2 used for EOR is never recovered but
is actually stored in the reservoir permanently; this long term sequestration may help to offset greenhouse
emissions from burning fossil fuels. In recent years, miscible CO2 floods have accounted for more oil
production in United States as compared to other EOR methods (Enrick and Oslan, 2012; Dai et al., 2014).
A survey production conducted in 2014 showed US miscible CO2 floods yielded 292,735 b/d of oil from
109 projects, accounting for 38% of US oil output from EOR processes (OGJ, 2014). However, miscible
flood output had a 5% lower production rate as compared to 2012, when the OGJ survey reported 112
projects. This is mainly due to economic reasons. As oil prices continue to fall, there is a high possibility
of EOR projects facing cancellation or delay. This strengthens the need to use state of the art technology
to maximize assets at a minimal operational cost. The majority of atmospheric CO2 emissions are
generated from anthropogenic sources. Using anthropogenic CO2 for EOR where available provides an
excellent opportunity to replace naturally sourced CO2 for EOR processes (Ampomah et al., 2016a; Enrick
and Oslan, 2012). This is particularly valuable in areas with oil reservoirs that would respond favorably
to CO2 EOR but that are distant from any natural CO2 sources. There are the twin benefits of a high
potential of recovering residual oil trapped within mature hydrocarbon reservoirs and storing a high
percentage of injected CO2 to reduce atmospheric emissions (Bachu, 2016).
CO2 presents certain difficulties when used in combined EOR/sequestration processes. For any CO2
EOR project, there is a very large initial investment to build the processing facilities for CO2 injection and
handling the recycled CO2 for re-injection and a need for continuous purchase of CO2 in order to maintain
the injection process. The primary goal; however, is to produce oil, maximizing profit, and not to consider
the storage of CO2 as an asset. When EOR and sequestration are both considered as goals of a project,
the optimal situation may change; there may eventually be some financial benefit to increasing storage,
perhaps at the loss of some production. Still the project must be profitable. Thus it becomes essential for
the producers to co-optimize oil production and CO2 storage.
This work presents a framework to co-optimize oil recovery and CO2 storage using genetic optimi-
zation algorithm with proxy model. Using a proxy model for the optimization process resulted in a
shortcut with the response surface instead of running full simulations to save computational time.
Twenty-one (21) years of forecasting is generated to ascertain the optimum operational development
strategy. A genetic algorithm with mixed integer capability is used for the optimization process. The
outcome from this study is compared to that from the best scenario based model conducted by Ampomah
et.al, (2016a).

Background
There have been a number of studies addressing co-optimization of oil recovery and CO2 sequestration
(Veld et al., 2014; Ashgari and Al-Dliwe, 2004; Babadagli, 2006; Leach et al., 2011; Jessen et al., 2005;
Forooghi et al., 2009; Ettehadtoavakkol et al., 2014). Veld et al., (2014) discussed most of the limitations
of the previous works. One of the most important issues is that some of the studies did not have a
comprehensive reservoir fluid model coupled with the optimization models, which could possibly affect
the obtained optimal results. Leach et al., (2011) presented a detailed economic analysis to improve
coupled optimization processes. Forooghi et al., (2009) used numerical modeling on a sector model to
co-optimize CO2-EOR and sequestration. The study focused on sensitivity analysis of operational
parameters and how parameter interactions affected the co-optimization process. Most CO2-EOR numer-
ical simulations have focused on much on operational parameters rather than the purchase and recycle of
CO2.
SPE-180376-MS 3

Whenever a successful history match is achieved, the accepted model can be used to predict future
reservoir performance. Several methods such as scenario based modeling (Ampomah et al., 2016a) and
the optimization approach are used (Zangl et al., 2006; Goodwin, 2015). In more complex optimization
workflows an experimental design can be utilized to ascertain the relationship between several control
variables affecting a process and/or its output (White et al., 2003; Yeten, 2005; Zangl, 2006). Experi-
mental designs have been used in CO2 storage applications (Ashraf et. al., 2013). Yeten (2005) presented
overview of experimental design methods.
Experimental design methodology is often used to reduce computational time by building approxi-
mated models usually referred to as proxy or surrogate models (Gogu and Passieux, 2013; Petvipusit et
al., 2014; Goodwin, 2015). Proxy models are highly appropriate for production optimization procedures
due to their complexity (Yeten, 2005; Zubarev, 2009; Bevillion, 2015). There are four proxy models used
most often in engineering disciplines: polynomial regression, kriging, thin-plate splines, and artificial
neural network (Yeten, 2007; Zubarev, 2009). Polynomial regression methodology is most common in the
petroleum engineering discipline and is therefore used in this work for CO2 – EOR process modeling.
Algorithms based on evolution theory have gained ground and are the most widely used in the
petroleum industry for history matching and optimization (Kramer et al., 2011). For optimization, genetic
algorithms have been used in various disciplines to solve a variety of problems (Romero et al., 2000; Gates
and Chakrabarty, 2006; Erbas and Christie, 2007; Safarzadeh and Motahhari, 2014; Al Dossary and
Nasrabadi, 2015; Biagi et al., 2015; Safi et al., 2016). Safi et al., (2016) recently published a study
discussing use of genetic algorithms for optimizing CO2 utilization based on different injection strategies.
In this study the genetic algorithm is also used to co-optimize oil recovery and CO2 storage within
Farnsworth Unit (FWU).
The Southwest Regional Partnership on Carbon Sequestration (SWP) is one of the seven United States
Department of Energy sponsored regional partnerships for the study of geological sequestration of CO2
and is currently partnering with Chaparral Energy LLC (CELLC) to study the CO2 storage potential in a
commercial- scale EOR project at FWU, Texas (Ampomah et al., 2016a). FWU was discovered in 1955
with initial oil in place estimated at 120 MMStb and was unitized in 1963. Waterflood began gradually
in 1964 and intensified by 1968. CELLC acquired FWU in November 2009 and began CO2 flooding in
December 2010. As of December 2014 a total of 16.82 BScf of CO2 has been injected into the Morrow
B sandstone formation (Ampomah et al., 2016a). Out of the total injected volume, CELLC purchased ˜
13.52 BScf of anthropogenic CO2 from ethanol and fertilizer plants (Figure 1). The remaining 3.48 BScf
was recycled from produced CO2. 93% of the purchased CO2 has been stored within the target formation
(Ampomah et al., 2016a). The operator expects to increase CO2 injection patterns from 12 to 25 five-spot
patterns in the future although expansion plans are highly dependent on oil prices. There is strong
incentive for both SWP and CELLC to construct comprehensive numerical models to help optimize both
oil recovery and CO2 storage in the most cost effective way.
4 SPE-180376-MS

Figure 1—Location of the Farnsworth Unit (red diamond) in the northeast part of the Texas panhandle within the Anadarko Basin
(shaded region). Stars indicate locations of large point sources from which anthropogenic CO2 is scrubbed from emissions and
compressed for injection. The northern star is the Arkalon Ethanol Plant in Liberal, Kansas, and the southern star is the Agrium
Fertilizer Plant in Borger, Texas. Modified from Grigg and McPherson, 2012.

Methodology

This work presents a numerical optimization based methodology to co-optimize oil recovery and CO2
storage. The Eclipse E300 compositional simulator was used in all simulation runs presented in this work.
A history matched and/or scenario based model adapted from Ampomah et al., (2016a) is used as the
baseline model for this study. A forecasting development strategy is constructed similar to one of several
scenario-based models constructed by Ampomah et al. (2016a) to find optimum operational parameters
to co-optimize oil recovery and CO2 storage using a genetic optimization algorithm. The outcome from
this study is compared to a baseline case (Ampomah et al. 2016a). Figure 2 illustrates the workflow used
in this.
SPE-180376-MS 5

Figure 2—Flow chart illustrating workflow for optimization with proxy models

A multi-objective function co-optimizing oil recovery and CO2 storage was formulated based on results
from the benchmark case. The baseline simulation case, which is based on manual optimization, showed
a possibility of improving CO2 storage and/or oil recovery by increasing recycled produced gas. This
provides maximum utilization of available CO2, which translates to increasing CO2 injection without
additional purchases. The defined multi-objective function maximizes both oil recovery and CO2 injection
(dependent variables) as illustrated in Equation 1. In this work, an equal unit weight was assigned to both
vectors. The incremental oil recovery and CO2 storage are derived from the best-optimized case.
After defining the multi-objective function, an initial twenty-three operational uncertain parameters
were included in sensitivity analysis to study their effect on the overall objective function using the Latin
Hypercube sampling technique (Table 1). The outcome was ranked and twelve control variables were
selected to be included in proxy modeling and/or the optimization process. The distribution ranges for
selected control variables were adjusted based on sensitivity analysis outcomes. A quadratic response
surface (proxy or surrogate) model was constructed based on the training data from numerical simulations.
The response surface is a second-order polynomial model determined by a least-square fit between the
polynomial response and training data. A typical quadratic response surface is shown in Equation 2. The
quadratic model requires a minimum of (N⫹1)(N⫹2)/2 training data where N represents number of
independent variables. A total of 100 training samples were selected using the Latin Hypercube sampling
technique. 10 additional samples were separately prepared and used for validating the constructed proxy
model. This process is repeated until a valid proxy is realized. Once a valid proxy model is accepted, it
is subsequently used in place of full simulation for optimization process.
6 SPE-180376-MS

Table 1—Initial uncertain parameters with their possible ranges used for sensitivity analysis
Uncertain Variable Description Units Minimum Maximum

BHP_INJ1 Bottomhole Injection Pressure psia 4500 5000

BHP_PROD1 Bottomhole Production Pressure (2015-2018) psia 1000 3000
BHP_PROD2 Bottomhole Production Pressure (2019-2036) psia 2000 4000
GCA1 Group 1 Gas Cycle (2015-2018) months 1 12
GCA2 Group 2 Gas Cycle (2015-2018) months 1 12
GCA3 Group 3 Gas Cycle (2015-2018) months 1 12
GCA4 Group 4 Gas Cycle (2015-2018) months 1 12
GCB1 Group 1 Gas Cycle (2019-2036) months 2 12
GCB2 Group 2 Gas Cycle (2019-2036) months 2 12
GCB3 Group 3 Gas Cycle (2019-2036) months 2 12
GCB4 Group 4 Gas Cycle (2019-2036) months 2 12
GOR1 Gas-Oil Ratio (2015-2019) Mscf/stb 15 30
GOR2 Gas-Oil Ratio (2019-2036) Mscf/stb 18 30
PROD1 Production Group Rate (2015-2019) stb 1800 3000
PROD2 Production Group Rate (2019-2036) stb 1500 2500
WCA1 Group 1 Water Cycle (2015-2019) months 0 12
WCA2 Group 2 Water Cycle (2015-2019) months 0 12
WCA3 Group 3 Water Cycle (2015-2019) months 0 12
WCA4 Group 4 Water Cycle (2015-2019) months 0 12
WCB1 Group 1 Water Cycle (2019-2036) months 0 12
WCB2 Group 2 Water Cycle (2019-2036) months 0 12
WCB3 Group 3 Water Cycle (2019-2036) months 0 12
WCB4 Group 4 Water Cycle (2019-2036) months 0 12

A genetic optimization algorithm is used to optimize the objective function (Equation 1). The genetic
algorithm follows three main evolutionary concepts including selection, crossover and mutation (Al
Dossary, 2015). The first step involves generating an initial population that consists of individuals
corresponding to solution in the sampling space. These individuals are evaluated based on their fitness,
which depends on the constructed objective function for selection. The fitness selection operator used in
this study is the ⬙Roulette Wheel⬙ operator (Goldberg, 1991; Liposwki and Lipowska, 2012). The
individuals with less favorable fitness are discarded and those remaining proceed to a new generation
through the process of mutation. This process is repeated until a prescribed condition is met. Table 2
shows summary of genetic algorithm operators used in the optimization.

Table 2—Parameters used in the genetic optimization process

Population 100
Maximum generations 20
Mutation probability 5%

(1)

where,
w ⫽ weight assigned to vector
FOPT ⫽ Cumulative oil production, Stb
FGIT ⫽ Cumulative CO2 injection, Ib-mol
(2)
SPE-180376-MS 7

where,
y ⫽ output response
␤ ⫽ unknown coefficients
X ⫽ vector of N inputs; the first, second and third term represents linear, interaction and
quadratic terms.

Model Description
Farnsworth Unit field (FWU), located in Ochiltree County Texas, is a mature hydrocarbon reservoir
undergoing tertiary production. The field was discovered in 1955 with initial oil in place estimated at 120
MMStb. Detailed geological summaries of FWU have been presented in several publications (Munson,
1988; Czoski, 2014; Gallagher, 2014; Hutton, 2015; Ampomah et al., 2015a, 2015b, 2016a; Ross-Coss et
al., 2016). The Morrow B reservoir, which is the main target reservoir, is overlain by the Morrow shale
and Thirteen Finger limestone that serve as excellent seals for injected CO2 (Heath et al., 2015). The CO2
–EOR study for FWU has concentrated on the west half of the field, which showed a better response to
the waterflood and is where CELLC is focusing their efforts. Ampomah et al., (2015a) presented the initial
geological model for FWU, which has been used by various researchers to conduct research in hydro-
geology, reactive transport and risk assessment of the project. Reservoir rock properties such as porosity
and permeability were constructed using geostatistical simulation methods. There was not a strong
correlation between porosity and permeability based on data available at the time of model construction
(Gallagher, 2014; Ampomah et al., 2015a). The FWU model used for this work has an upscaled grid of
82⫻78⫻5 with grid block dimensions of 200 ft ⫻200 ft (Ampomah et al., 2015a).
Initial water saturation at FWU was 31% as measured from special core analysis in the 1950’s. The
initial reservoir pressure and temperature was 2217.17 psia and 168 °F recorded at datum depth of 7900
ft. The reservoir was slightly undersaturated at time of discovery with a bubble point pressure of 2073.7
psia. The initial oil in place estimated for the west half model used in this work is 60⫾4 MMStb. Relative
permeability curves used in this work was measured from special core analysis (Figure 3a). Capillary
pressure used for this study was based on Corey and Brooks correlations modified for FWU by White et
al., (2015) as illustrated in (Figure 3b).

Figure 3—(a) oil and water relative permeability and capillary pressure (b) gas relative permeability
8 SPE-180376-MS

Gunda et al. (2015) provided a comprehensive reservoir fluid analysis on a FWU oil sample which has
been used in various SWP reservoir simulation models (Ampomah et al. 2015b; Ampomah et al. 2016a).
Reservoir fluid was tuned to an equation of state (EOS) using the 3-parameter Peng Robinson EOS (Peng
and Robinson, 1976) with a Peneloux volume correction (Peneloux et al. 1982). The Lohrenz-Bray-Clark
correlation is used for the calculations of viscosities (Lohrenz et al. 1964). After the PVT tuning process,
a slimtube simulation experiment conducted on FWU fluids resulted in a minimum miscible pressure
(MMP) of about 4009 psia compared to experimental value of 4200 psia, representing a less than 5% error
(Gunda et al., 2015; Ampomah et al., 2015b). Sensitivity analyses were conducted to study the effect of
using recycled CO2 with or without impurities on MMP predictions (Gunda et al., 2015). These analyses
enabled modeling of CO2-EOR performance incorporating the real world use of both purchased and
recycled CO2.
Ampomah et al., (2016a) presented the preliminary history match model for FWU. The primary and
secondary history processes lasted for 55 years from field discovery in 1955 until December 2010. The
tertiary CO2 flood was performed between December 2010 and December 2014. Various waterflood
patterns swept to residual oil saturation of about 27%, a successful flood. This is not unexpected given the
relatively good mobility ratio of about 1.6 and high injection throughput of at least 1.7 displaceable pore
volumes. From the simulation results, nearly 6 MMBO represents 10% of OOIP produced prior to the
initiation of waterflood. A total of 25.3 MMBO of cumulative oil has been produced at the end of
secondary and tertiary recovery from the west half of FWU, which represents nearly 40% of OOIP. The
most important aspect of this work from the SWP’s perspective is the amount of CO2 that has been stored
within the Morrow B reservoir. A total of 16.82 BScf of CO2 has been injected including both purchased
and recycled gas. About 93% of the purchased CO2 has been stored as of December 2014. These results
suggest that CO2 sequestration, one of the primary SWP objectives, may be increased through optimi-
zation of the WAG process.

Baseline Model
The baseline simulation model was adapted from Ampomah et. al., 2016a. The operational strategy that
yielded the best CO2 recovery and/or oil recovery is used in this work. At FWU, there are plans to convert
seven extant water injector wells and an additional six wells to water alternating gas (WAG) wells. There
is a possible constant CO2 purchase of about 10,000 Mscf/d sourced from two locations and a compressor
facility to recycle produced gas. As of now, produced gas from the field has more than 90% CO2. The best
operational strategy reported by Ampomah et al., (2016a) stored about 75% of CO2 and produced about
34% incremental oil recovery beyond waterflood.
Newly drilled WAG injectors were queued to be introduced into the model based on CELLC’s
anticipated project schedule, which is highly dependent on oil price. A total of 25 WAG wells and 35
producers were introduced into the model by the end of 2018. The WAG wells were divided into four
groups with similar operational constraints. The strategy involves continuous injection of 10,000 Mscf/d
of purchase CO2 in addition to recycled gas until end of 2021. At year 2022, purchased CO2 was cut to
8,000 Mscf/d representing a 20% reduction. CO2 purchased was subsequently reduced by 1000 Mscf/d
each year until 2030. Afterwards, use of purchased CO2 ceased and only recycled gas was injected until
January of 2036. The prediction case was run for a total of 21 years. The other operational constraints
includes assigning 2000psia as BHP_PROD, 4800 psia as BHP_INJ and WAG cycle of 3:1.
The simulation results indicates a possibility of losing more than 5% of produced CO2 to flaring
without recycling even though the compressor may not have been utilized to its full capacity (Figure 4).
Therefore there is a possibility of potential increase in CO2 storage and/or oil recovery forecast by using
advanced optimization tools. The genetic optimization approach could predict optimum operational
parameters such as injection/production bottomhole pressures and WAG cycles to maximize the objective
function.
SPE-180376-MS 9

Figure 4 —UCO2 plot represents CO2 purchase (blue) profile in Ib-mol/d from 2015 to 2036. UCO2I (black) is the amount of gas available
for injection in Ib-mol/d. The red curve represents the actual gas injection rate for baseline case. It is clear that a lot of available CO2
are actually not fully which could be due to operating at non-optimized strategy.

Results
This section presents detailed results on proxy modeling construction and genetic optimization modeling.
A sensitivity analysis was performed to reduce the number of control variables to be used in
constructing proxy or surrogate model for optimization. Ampomah et al., (2016b) presented a detailed
description of these uncertain parameters and their effects. Figure 5 show a tornado plot summary of each
individual uncertain parameter’s effect on the multi-objective function. The top twelve sensitive param-
eters were included in the optimization procedure. The selected variables with ranges are presented in
Table 3. WAG cycle after year 2018 and injection / production well bottomhole pressures were the most
sensitive parameters and as such used as control parameters in optimization study.

Figure 5—Tornado plot ranking of uncertain parameters as a function of defined objective function. The x-axis represents uncertain
parameters and y-axis represents multi-objective function
10 SPE-180376-MS

Table 3—Selected control variables after sensitivity analysis used for optimization process
Initial
Control Variable Description Units Minimum Maximum Guess

WCB2 Group 2 Water Cycle (2019-2036) months 0 3 1

WCB3 Group 3 Water Cycle (2019-2036) months 0 3 1
GCB2 Group 2 Gas Cycle (2019-2036) months 2 10 5
PROD2 Production Group Rate (2019-2036) stb 1500 2500 1800
WCB1 Group 1 Water Cycle (2019-2036) months 0 12 1
BHP_INJ1 Bottomhole Injection Pressure psia 4500 5000 5000
WCB4 Group 4 Water Cycle (2019-2036) months 0 12 1
GCB1 Group 1 Gas Cycle (2019-2036) months 2 12 7
GCB3 Group 3 Gas Cycle (2019-2036) months 2 12 9
BHP_PROD2 Bottomhole Production Pressure (2019-2036) psia 1500 2500 2000
GCB4 Group 4 Gas Cycle (2019-2036) months 2 12 3
PROD1 Production Group Rate (2015-2019) stb 1600 3000 1800

The optimization experiment was designed using fast proxy models that approximated the full
simulations in order to reduce computational time. 100 simulations were run during the training process
to construct the surrogate model. Ten additional simulations were conducted to validate the polynomial
response surface model. The Latin hypercube algorithm was used for training, validation and evaluation
sampling. Figure 6 and 7 are cross plots of simulated vs predicted cumulative oil production and
multi-objective function training data and validation data to ascertain the viability of the constructed
surrogate model. The clustering of training and validation data points around the equiline shows the
accuracy of proxy. These cross plot shows results at the last time step of the model (January 1, 2036).
Figure 8 shows response surface pressure at various percentiles and pressure from one of ten validation
cases. The proxy model is accurately verified at a particular time-step. This figure shows less fit at the
beginning and gets better at later periods. The response was approximated once a year which could have
caused the earlier misfit. However, it shows the trend of the possible pressure profile, which is worth
mentioning since field operation at the optimum scenario could at some point be below currently known
MMP. Figure 9 shows the mean, 10th, 50th and 90th percentiles of multi-objective function. Statistical
measurements of uncertainty defined by 90th percentile-to-10th percentile ratio shows a value of 1.46 for
the computed multi-objective function last time step. A lower standard deviation (6.72E⫹8 ⫾ 9.64E⫹6)
was found as well. Calculation of (P50-Mean)/P50 was also used to assess uncertainty (Ampomah et al.,
2014; Chen, 2012). A computed value of 0.02, which is far less than 1, signifies a low uncertainty. These
statistical analyses confirmed a low uncertainty associated with the constructed proxy model.
SPE-180376-MS 11

Figure 6 —Cross plot of simulated vs. predicted results in addition to the validation points for cumulative oil production. The clustering
of the points around the equiline signifies proxy accuracy.

Figure 7—Cross plot of simulated vs. predicted results in addition to the validation points for formulated multi-objective function. The
clustering of the points around the equiline accounts for accuracy in proxy model construction.
12 SPE-180376-MS

Figure 8 —Response surface model for pressure compared with validated simulation case.

Figure 9 —Plot of multi-objective function from proxy model at different percentiles

Once the fast proxy model was accepted, it was used in place of full finite difference simulation to yield
faster optimization results. A genetic algorithm that has mixed-integer capabilities is used to co-optimize
the objective function. The process required a minimum of 1101 iterations. The optimization converged
and a best iteration was realized at 1051 runs. A full simulation run was conducted using the optimum
parameters predicted by the proxy model for comparison. This process was repeated until a minute error
between proxy results and the full simulation validation case was realized. A total of six genetic
optimization processes were required to yield an optimum solution. There was an error of 1.9% between
the proxy optimum solution and the confirmation full simulation run on the computed multi-objective
function.
SPE-180376-MS 13

Table 4 summarizes the comparison between the baseline case and the genetic optimized case. The
optimum case yielded 48.80 MMstb of oil (81% of OOIP), which is 8.7% greater recovery as compared
to the manually engineered case. More significant to the SWP research effort is the prediction that 95%
of the CO2 purchased is anticipated to be stored within FWU. If FWU is operated using the optimum
operational parameters that result from the genetic optimization process there could be a 20% increase in
CO2 storage over that predicted from manually-optimized operation.

Table 4 —Summary of comparison between baseline case and genetic optimization case
Baseline Genetic
Results Unit case Optimized case

CO2 Purchased MMIb-mol 154 154

CO2 Production Cumulative MMIb-mol 402 627
Recycle MMIb-mol 364 619
CO2 Injection Cumulative MMIb-mol 518 772
Total Storage* MMIb-mol 116 148
% Storage % 75 95
Cumulative Oil Production MMstb 43.62 48.80
% Oil Recovery % 72.70 81.40

Summary and Conclusions

With the uncertainties in operational and/or geologic parameters, optimization processes typically involve
significant numbers of runs and computational time. Furthermore, the simulation based optimization
approach has a major limitation in finding an optimum solution due to the complexity of the defined
problem. This work presented a new approach combined with proxy model and a genetic algorithm to
optimize oil recovery and CO2 storage simultaneously.
In CO2-EOR, there is certainly a trade-off between maximizing oil recovery and CO2 storage. A unique
multi-objective function has been developed to co-optimize oil recovery and CO2 storage using state-of-
the-art modeling tools. The CO2 defined term, for example CO2 injection, could be changed depending
on the problem considered. The use of proxy modeling could be valuable in reducing computation time
in CO2-EOR co-optimization cases but it has been limited in its application, perhaps due to the complexity
of the defined multiple objectives. The difficult challenge in using proxy models is that they do not
necessarily have physical justifications and unless they are thoroughly validated they have no value. A
given proxy model is also site-specific, so once validated, it can be used to perform real-time simulation
studies. In this work, several approaches including statistical tools were utilized to validate and evaluate
a constructed proxy model prior to the optimization process to ensure accuracy in future predictions.
A genetic algorithm with mixed integer capabilities was used to optimize the multi-objective function
to co-optimize oil recovery and CO2 storage. This optimizer has been utilized in finding optimum
solutions in several engineering and science disciplines. Since the genetic algorithm has no idea about the
problem one must always specify constraints about control variables to limit the space of possible
solutions. There are no direct instructions available on selecting ranges for genetic operators such as
population size and other variables. More research is needed in this area in order to take full advantage
of the robustness of this optimizer to solve difficult challenges faced by the petroleum industry. The
parameters used for this work were chosen based on limited sensitivity analysis and/or computational cost.
In this work, a population size of 100 and maximum of 20 generations required a minimum of 1101
iterations. The optimization process converged at iteration 1051. The computations lasted for few hours
due to the use of a faster proxy model. The genetic algorithm optimum solution yielded a possibility of
storing 95% of purchased CO2 and most importantly recovering 40% of remaining oil post waterflood. A
14 SPE-180376-MS

confirmation with full simulation run using optimum control variables predicted by the proxy model
resulted in a residual error of 1.9%.
In any CO2-EOR project, the MMP has to be monitored over the production life of the field. As the
percentage of lighter components such as methane is reduced within the reservoir, there is a high
likelihood of a reduction in MMP. Figure 7 showed that FWU reservoir pressure dropped to less than 4000
psia after year 2030 when CO2 purchase is ceased. Laboratory work is still ongoing to determine the
changes in MMP after nearly all lighter components have been produced from the Morrow B reservoir.
Adjustment of the WAG ratio could be necessary at the later stages of operation if the reservoir pressure
goes below anticipated MMP and has a detrimental effect on oil recovery.
This work successfully demonstrated a framework to co-optimize oil recovery and CO2 storage using
a multi-objective function and a reduced order model. The developed multi-objective function was
successfully optimized using a genetic algorithm which predicted more than 80% oil recovery and 95%
CO2 storage.
This approach can be used to optimize CO2 storage in other large-scale field projects while still
producing a significant amount of the remaining oil. The approach can also be used for optimization of
different facets of EOR projects and as well as with other engineering and/or science disciplines where
difficult problems require more complex types of analyses.

Unit Conversions
1 Mscf ⫽ 2.635 lb-mol
1 lb-mol ⫽ 1.977E-05 Metric tonnes

Acknowledgements
Funding for this project is provided by the U.S. Department of Energy’s (DOE) National Energy
Technology Laboratory (NETL through the Southwest Regional Partnership on Carbon Sequestration
(SWP) under Award No. DE-FC26-05NT42591. Additional support has been provided by site operator
Chaparral Energy, L.L.C. and Schlumberger Carbon Services.DisclaimerThis report was prepared as an
account of work sponsored by an agency of the United States Government. Neither the United States
Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied,
or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any
information, apparatus, product, or process disclosed, or represents that its use would not infringe
privately owned rights. Reference herein to any specific commercial product, process, or service by trade
name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement,
recommendation, or favoring by the United States Government or any agency thereof. The views and
opinions of authors expressed herein do not necessarily state or reflect those of the United States
Government or any agency thereof.

References
1. Ampomah, W., Balch, R. S., Grigg, R. B. Will, R., and Dai. Z., White M. D. (2016a): ⬙Farnsworth Field CO2-EOR
Project: Performance Case History⬙ paper SPE-179528-MS. SPE Improved Oil Recovery Conference held in Tulsa,
OK, USA. April 11-13 2016
2. Ampomah, W., Balch, R. S., Cather M., Will R. & Lee S. Y.. (2016b) Performance of CO2-EOR and Storage
Processes Under Uncertainty. SPE Europec at 78th EAGE Conference May 30- June 2 2016, Vienna, Austria.
SPE-180084-MS
3. Ampomah, W., Balch, R. S., and Grigg, R. B. (2015a) Analysis of Upscaling Algorithms in Heterogeneous Reservoirs
with Different Recovery Processes. SPE Production and Operations Symposium March 1-5 2015, Oklahoma USA.
http://dx.doi.org/10.2118/173588-MS
SPE-180376-MS 15

4. Ampomah, W., Balch, R. S., and Grigg, R. B., Dai Z., and Pan F. (2015b) Compositional Simulation of CO2 Storage
Capacity in Depleted Oil Reservoirs. Carbon Management Technology Conference November 16-19 2015, Sugarland,
Houston-Texas USA. http://dx.doi.org/10.7122/439476-MS
5. Ampomah, W., Chen, H. Y., & Assad, J. (2014). Comparison of Probabilistic Techniques in the Study of Fluid Effects
on Seismic Properties. International Petroleum Technology Conference. January 19-21 2014, Qatar. http://dx.doi.org/
10.2523/17452-MS
6. Ashgari K. and Al-Dliwe, A. 2004. Optimization of carbon dioxide sequestration and improved oil recovery in oil
reservoirs. In Proceedings of the 7th International Conference on Greenhouse Gas Control Technologies: Amsterdam,
1: 381–89. Elsevier.
7. Babadagli, T. 2006. Optimization of CO2 injection for sequestration/enhanced oil recovery and current status in
Canada. In Lombardi, S., Altunina, L. K., and Beaubien, S. E., editors. Advances in the Geological Storage of Carbon
Dioxide, 261–70. Dordrecht: Springer.
8. Bachu, S. (2016). Identification of oil reservoirs suitable for CO2-EOR and CO2 storage (CCUS) using reserves
databases, with application to Alberta, Canada, Int. J. of Greenh. Gas Control, 44, 152–165.
9. Dai, Z., Middleton R., Viswanathan H., Fessenden-Rahn J., Bauman J., Pawar R., Lee S. and McPherson B., (2014).
An integrated framework for optimizing CO2 sequestration and enhanced oil recovery, Environ. Sci. Technol. Lett.,
1, 49 –54.
10. Dai, Z., Middleton R., Viswanathan H., Fessenden-Rahn J., Bauman J., Pawar R., Lee S. and McPherson B., (2014).
An integrated framework for optimizing CO2 sequestration and enhanced oil recovery, Environ. Sci. Technol. Lett.,
1, 49 –54.
11. Bevillon, D., Petroleum, D., Dpe, E., Mohagerani, S., & Reservoir, B. (2015). A Miscible EOR Project in a Mature,
Offshore, Carbonate Middle East Reservoir - Uncertainty Analysis With Proxy Models Based on Fateh Field and
Mishrif Reservoir. SPE Reservoir Characterisation and Simulation Conference and Exhibition held in Abu Dhabi,
UAE, 14-16 September 2015.
12. Biagi, J., Agarwal, R., & Zhang, Z. (2016). Simulation and optimization of enhanced gas recovery utilizing CO 2.
Energy, 94, 78 –86.
13. Chen, H.-Y. (2012). Engineering Reservoir Mechanics, New Mexico Tech, Socorro New Mexico
14. Czoski, P., (2014). Geologic characterization of the Morrow B reservoir in Farnsworth Unit, TX using 3D VSP
seismic, seismic attributes, and well logs: M.S. Thesis, New Mexico Tech
15. Dai, Z., R. Middleton, Viswanathan H., Fessenden-Rahn J., Bauman J., Pawar R., Lee S. and McPherson B., (2014).
An integrated framework for optimizing CO2 sequestration and enhanced oil recovery, Environ. Sci. Technol. Lett.,
1, 49 –54.
16. Dossary, M. A. Al, A, S. A. T., Nasrabadi, H., & Texas, A. (2015). Well Placement Optimization Using Imperialist
Competition Algorithm, SPE Reservoir Characterisation and Simulation Conference and Exhibition held in Abu
Dhabi, UAE, 14-16 September 2015.
17. Enick, R. M. and Olsen, D. K. (2012). Mobility and Conformance Control for Carbon Dioxide Enhanced Oil Recovery
(CO2-EOR) via Thickeners, Foams, and Gels – A Detailed Literature Review of 40 Years of Research, National
Energy Technology Laboratory, DOE/NETL-2012/1540, Activity 4003.200.01, 267p.
18. Erbas, D., & Christie, M. (2007). Effect of Sampling Strategies on Prediction Uncertainty Estimation. SPE Reservoir
Simulation Symposium. Houston, Texas.
19. Ettehadtavakkol, A., Lake, L. W., and Bryant, S. L. 2014. CO2-EOR and storage design optimization. International
Journal of Greenhouse Gas Control, 25: 79 –92. http://dx.doi.org/10.1016/j.i-jggc.2014.04.006.
20. Forooghi, A., Hamouda, A., and Eilertsen, T. 2009. Co-optimization of CO2 EOR and sequestration in a North Sea
chalk reservoir. Paper SPE 125550 presented at the 2009 SPE/EAGE Reservoir Characterization and Simulation
Conference, Abu Dhabi, UAE, 19-21 October. http://dx.doi.org/10.2118/125550-MS
21. Gallagher, S. R. (2014). Depositional and Diagenetic Controls on Reservoir Heterogeneity: Upper Morrow Sandstone,
Farnsworth Unit, Ochiltree County, Texas. M.S. Thesis, New Mexico Tech, Socorro NM, USA.
22. Gogu, C., & Passieux, J. C. 2013. Efficient surrogate construction by combining response surface methodology and
reduced order modeling. Structural and Multidisciplinary Optimization, 47 (6), 821–837.
23. Goodwin, N., & Products, E. (2015). Bridging the Gap Between Deterministic and Probabilistic Uncertainty
Quantification Using Advanced Proxy Based Methods, SPE Reservoir Simulation Symposium held in Houston, Texas,
USA, 23-25 February 2015.
24. Goldberg, K. D. (1991) A comparative analysis of selection schemes used in genetic algorithms, in: G.J.E Rawlins
(Ed.), Foundations of Genetic Algorithms, Morgan Kaufmann Publishers, San Mateo, California, 1991, pp. 69 –93.
16 SPE-180376-MS

25. Grigg, R. B., McPherson, B. J., (2012). Phase III- Deployment Phase Farnsworth Unit CCUS, Ochiltree, Texas,
National Energy Technology Laboratory Carbon Storage RD Meeting: Developing the Technologies and Building the
Infrastructure for CO2 Storage. August 21-23, Pittsburgh, USA.
26. Gunda D., Ampomah, W., Grigg, R. B. and Balch, R. S. (2015) Reservoir Fluid Characterization for Miscible
Enhanced Oil Recovery. Carbon Management Technology Conference November 16-19 2015, Sugarland, Houston-
Texas USA. http://dx.doi.org/10.7122/440176-MS
27. Heath, J. E., Dewers, T. A., Mozley, P. S. (2015). Characteristics of the Farnsworth Unit, Ochiltree County, Texas :
Southwest Partnership CO2 Storage - EOR Project.
28. Hutton, A., (2015). Geophysical Modeling and Structural Interpretation Of A 3D Reflection Seismic Survey In
Farnsworth Unit, TX. M.S. Thesis, New Mexico Tech, Socorro NM, USA
29. Jessen K., Kovscek, A. R., and Orr, Jr. F. M.,. 2005. Increasing CO2 storage in oil recovery. Energy Conversion and
Management, 46: 293–311.
30. Kramer O., Echeverr´ıa Ciaurri D., and Koziel S. 2011. Derivative-free optimization. S. Koziel and X. S. Yang, eds.,
Computational Optimization, Methods and Algorithms, 61–83. Springer.
31. Leach, A., Mason, C. F., and van’t Veld, K. 2011. Co-optimization of enhanced oil recovery and Carbon sequestration.
Resource and Energy Economics, 33: 893–912. http://dx.doi.org/10.1016/j.rese-neeco.2010.11.002.
32. Li, S. & Zhang, Y. 2014. Model complexity in carbon sequestration: a design of experiment and response surface
uncertainty analysis. International Journal of Greenhouse Gas Control, 22, 123–138.
33. Lipowski, A., & Lipowska, D. (2012). Roulette-wheel selection via stochastic acceptance. Physica A: Statistical
Mechanics and Its Applications, 391 (6), 2193–2196. 10.1016/j.physa.2011.12.004
34. Lohrenz, J., Bray, B. G., Clark, C. R. (1964). ⬙Calculating Viscosities of Reservoir Fluids from their compositions⬙,
SPE Paper 915, Journal of Petroleum Technology, p. 1171–1176.
35. Munson, T. 1988. Depositional, Diogenetic, and Production History of the Upper Morrowan Buckhaults Sandstone,
Farnsworth Field, Ochiltree County Texas. West Texas State University.
36. Oladyshkin, S., Class, H., Helmig, R., & Nowak, W. 2011. A concept for data-driven uncertainty quantification and
its application to carbon dioxide storage in geological formations. Advances in Water Resources, 34 (11), 1508 –1518.
37. Oil and Gas Journal, April 2012 p. 56
38. Oil and Gas Journal, April 2014
39. Peneloux, A., Rauzy, E., and Freze, R.: ⬙A Consistent Correction for Redlich-Kwong-Soave Volumes,⬙ Fluid Phase
Equilibria (1982).
40. Peng, D. Y. and Robinson, D. B. (1976). A New Two-Constant Equation of State. Ind.Eng.Chem. Fundamentals, 15,
59 –64
41. Petvipusit, K. R., Elsheikh, A. H., Laforce, T. C., King, P. R., & Blunt, M. J. 2014. Robust optimisation of CO2
sequestration strategies under geological uncertainty using adaptive sparse grid surrogates. Computational Geosci-
ences, 18 (5), 763–778.
42. Romero, C. E., Carter, J. N., Gringarten, A. C., and Zimmerman, R. W. 2000. A Modified Genetic Algorithm for
Reservoir Characterisation. Paper SPE 64765 presented at the International Oil and Gas Conference and Exhibition
in China, Beijing, 7-10 November. 10.2118/64765-MS.
43. Ross-Coss D., W. Ampomah, R. S. Balch, M. Cather, P. Mozley, L. Rasmussen (2016). An Improved Approach for
Sandstone Reservoir Characterization. SPE Western Regional Meeting held in Anchorage, Alaska, USA, 23-26 May
2016. SPE-178375-MS
44. Safi, R. (2015). Numerical Simulation and Optimization of Carbon Dioxide Utilization for Enhanced Oil Recovery
from Depleted Reservoirs, Chemical Engineering Science 144, 30 –38. 10.1016/j.ces.2016.01.021
45. Safarzadeh, M. A., & Motahhari, S. M. (2014). Co-optimization of carbon dioxide storage and enhanced oil recovery
in oil reservoirs using a multi-objective genetic algorithm (NSGA-II). Petroleum Science, 11 (3), 460 –468. 10.1007/
s12182-014-0362-1
46. Shahkarami, A., Mohaghegh, S. D., & Hajizadeh, Y. (2015). SPE-173405-MS Assisted History Matching Using
Pattern Recognition Technology. SPE Digital Energy Conference and Exhibition, Woodland, TX, USA, 3-5 March.
47. Veld, K. Van, Wang X., & Alvarado, V. (2014). Economic Co-optimization of Oil Recovery and CO 2 Sequestration,
SPE ATCE, 27-29 October, 2014, The Netherlands
48. White, C., & Royer, S. (2003). Experimental Design as a framework for reservoir studies. Proceedings of SPE
Reservoir Simulation Symposium. Houston, Texas, U.S.A., 3-5 February 2003. 10.2118/79676-MS
49. White, M. D., McPherson, B. J., Grigg, R. B., Ampomah, W., & Appold, M. S. (2014). Numerical Simulation of
Carbon Dioxide Injection in the Western Section of the Farnsworth Unit. Energy Procedia, 63, 7891–7912.
http://dx.doi.org/10.1016/j.egypro.2014.11.825
SPE-180376-MS 17

50. Whitley, D., Starkweather T., and Bogart C., 1990, Genetic algorithms and neural networks: Optimizing connections
and connectivity: Parallel Computing, 14, no. 3, 347–361, http://dx.doi.org/10.1016/0167-8191(90)90086-O.
51. Yeten, B., Castellini, A., Guyaguler, B., & Chen, W. H. (2005). A comparison study on Experimental Design and
Response Surface Methodologies. SPE Reservoir Simulation Symposium. Houston, Texas U.S.A., 31 January 2005-2
February 2005. 10.2118/93347-MS.
52. Zangl, G., Graf, T., & Al-Kinani, a. (2006). Proxy modeling in production optimization. SPE Europec/EAGE Annual
Conference and Exhibition held in Vienna, Austria, 12-15 June 2006. 10.2118/100131-MS
53. Zubarev, D. I. (2009). SPE 124815 Pros and Cons of Applying Proxy-Models as a Substitute for Full Reservoir
Simulations. SPE Annual Technical Conference and Exhibition held in New Orleans, Louisiana, USA, 4-7 October
2009. 124815, 23. 10.2118/124815-MS