International Journal of Coal Geology 266 (2023) 104177

Contents lists available at ScienceDirect

International Journal of Coal Geology

journal homepage: www.elsevier.com/locate/coal

Uncertainty quantification for CO2 storage during intermittent CO2-EOR in

oil reservoirs
Mohamed Gamal Rezk a, Ahmed Farid Ibrahim a, b, *, Abdulrauf R. Adebayo a
a
Center for Integrative Petroleum Research, King Fahd University of Petroleum & Minerals, 31261 Dhahran, Saudi Arabia
b
Department of Petroleum Engineering and Geosciences, King Fahd University of Petroleum & Minerals, Dhahran 31261, Saudi Arabia

A R T I C L E I N F O A B S T R A C T

Keywords: CO2 enhanced oil recovery (CO2-EOR) process is considered the most feasible option for a secure storage of CO2
CO2 sequestration in underground formations. However, predicting amounts of CO2 stored and the oil recovered is associated with
Enhanced oil recovery high risks due to data uncertainty. Sources of uncertainty lie in the challenges of accurately characterizing
Uncertainty quantification
reservoir properties. Hence, the main objective of this study is to quantify the uncertainty of CO2 storage and oil
Machine Learning
Response surface methodology
recovery during intermittent CO2-assisted gravity drainage (CO2-GAGD) EOR process with a focus on key un­
Intermittent CO2 assisted gravity drainage certain parameters including reservoir properties (permeability heterogeneity, porosity, and compressibility);
fluid properties (CO2 diffusion coefficient and Henry’s constant); and rock-fluid properties (gas-oil relative
permeability data). Additionally, we aim to generate a rapid tool that reduces the computational expenses of the
compositional reservoir simulation process by employing artificial neural network (ANN) as a machine learning
(ML) tool. ANN was applied to generate predictive models that accurately predicted oil recovery and gas storage
during the intermittent CO2-GAGD process. A total of 173 numerical samples were simulated and the objective
functions were obtained. From the defined ranges of input variables and simulation results, cumulative distri­
bution functions (CDFs) were obtained and uncertainty bounds, i.e., P10, P50, and P90, were estimated. Addi­
tionally, the generated ANN model showed excellent predictions of the objective functions with high correlation
coefficients, higher than 0.98 in most of the cases, and with an absolute average percentage error (AAPE) range
from 2.6 to 9%. Finally, the results of sensitivity analysis, conducted using the developed ML models with
generating 10,000 realizations from the input and output parameters, showed the most influential input pa­
rameters on the performance of the CO2 storage and oil recovery factor. Reservoir porosity was found to have a
significant impact on oil recovery and CO2 storage. A response surface analysis showed that the main inde­
pendent input parameters contributing to the total uncertainty of oil recovery were horizontal permeability, oil
relative permeability, and formation porosity. This study highlights the significant impact of uncertainties of
input parameters on simulating the performance of CO2 storage and oil recovery in underground reservoirs.

1. Introduction tertiary recovery method in oil reservoirs. Furthermore, CO2-EOR pro­

vided an additional advantage in securing CO2 storage in underground
The main purpose of carbon capture, utilization, and storage (CCUS) formations. The high capability of CO2 to improve oil recovery during
projects is to slow down the increasing CO2 concentrations in the at­ the gas injection process in oil reservoirs comes from two-ways mass
mosphere and mitigate the subsequent climate change problem. One of transfer processes that occur between the injected CO2 and reservoirs
the viable options for CO2 storage is to inject the captured CO2 into fluids, i.e., water and crude oil, under reservoir conditions. As a conse­
underground formations such as saline aquifers, depleted oil and gas quence of the fluid-fluid and rock-fluid interactions during CO2 injec­
reservoirs, and unmineable coal seams (I. e. agency, 2018; Kumar et al., tion, the CO2 can cause oil swelling, reduction in the viscosity, density,
2020). CO2 can also be an associated gas with produced hydrocarbons and interfacial tension of crude oil, rock wettability alteration, and
during primary or secondary recovery processes from oil reservoirs. vaporization of light and intermediate oil components (Jia et al., 2019).
CO2-enhanced oil recovery (CO2-EOR) was found to be an efficient Hence, oil mobility is improved, and more oil can be recovered. Besides

* Corresponding author at: Center for Integrative Petroleum Research, King Fahd University of Petroleum & Minerals, 31261 Dhahran, Saudi Arabia
E-mail address: [email protected] (A.F. Ibrahim).

https://doi.org/10.1016/j.coal.2022.104177
Received 25 September 2022; Received in revised form 21 December 2022; Accepted 26 December 2022
Available online 28 December 2022
0166-5162/© 2022 Elsevier B.V. All rights reserved.
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

the improvement in oil recovery, CO2 can be stored during the CO2-EOR et al. (Seyyedsar et al., 2017) tested various injection strategies of low-
process through four primary trapping mechanisms. These mechanisms density CO2 in oil reservoirs using a sandstone core. They have noticed
include structural or stratigraphic trapping, mineral trapping, residual reductions in the core pressure during the soaking periods of the inter­
trapping, and solubility trapping (Ampomah et al., 2016). Stratigraphic mittent CO2 injection. These pressure drops indicated the diffusion of
trapping is controlled by the sealing characteristics of cap rock that the CO2 gas into fluids saturating the core. The oil recovery factor was
avoid the upward penetration of the underlying low-density CO2. CO2 also improved compared to the continuous CO2 flooding. Furthermore,
can also react with reservoir fluids and minerals resulting in permanent the intermittent CO2 injection was found to improve the oil swelling and
storage of the injected CO2 in the form of solid minerals. CO2 is also oil viscosity reduction during the soaking periods of the injected CO2 in a
trapped by the effect of capillary forces existing in the porous media core saturated with a heavy live oil (Seyyedsar and Sohrabi, 2017).
which is known as residual or capillary trapping (Iglauer, 2017). Addi­ Recently, Rezk and Foroozesh (Rezk and Foroozesh, 2022) numerically
tionally, the high ability of CO2 to be dissolved into reservoir fluids, investigated the effect of CO2 diffusion on both oil recovery and CO2
under the reservoir conditions, results in a secured CO2 storage mech­ storage during intermittent CO2-assisted gravity drainage in a light oil
anism which is known as the solubility trapping. CO2 solubility in reservoir. Their results highlighted the significant impact of the contact
reservoir oil is much high compared to its solubility in reservoir brine. time between the injected CO2 and reservoir fluids, controlled by the
However, gas solubility in a liquid phase is controlled by operating soaking period or CO2 injection rate, on the gas penetration depth, oil
pressure, operating temperature, and oil composition or brine salinity, recovery, and gas storage when the gas diffusion was considered in the
all of which change from one reservoir to another (Zhou et al., 2018; simulation model.
Samara et al., 2022). As it was previously discussed, CO2-EOR provides an option for CO2
Various injection strategies can be applied in CO2-EOR projects storage in underground formations. However, predicting amounts of
based on reservoir rock and fluid properties, reservoir conditions, CO2 stored and the recovered oil is associated with risks due to data
reservoir drainage area, location of the injectors and producers, and the uncertainty. This uncertainty lies in the challenges of accurately char­
amount of gas available for injection. CO2 is either injected on its own or acterizing the reservoir and fully understanding rock-fluid interactions
with water. During continuous CO2 injection strategy, oil is recovered (Ma et al., 2011). The reservoir is influenced by complex processes such
through different mechanisms that depend on the degree of CO2-oil as rock-fluid reactions, i.e., geochemical reactions, and alteration of rock
miscibility, i.e., miscible, near miscible, or immiscible. Although and fluid properties due to uplifts and plate tectonics movements.
continuous CO2 injection results in high amounts of CO2 storage at the Hence, an accurate estimation of reservoir properties is not a straight­
end of the flooding process, it faces a poor volumetric sweep efficiency forward process and requires a detailed uncertainty analysis. Either
due to the adverse gas-oil mobility ratio. Hence, alternative injection of deterministic or probabilistic methods can be applied for reservoir
CO2 and water in cycles, i.e., water alternating gas (WAG), can over­ quantification and uncertainty analysis. The deterministic method is
come this issue as the injected water fills some pores and reduces the providing a single outcome, however, there is a range of possible out­
CO2 mobility (Moghadasi et al., 2018). Cyclic CO2 injection, i.e., CO2 comes that arise due to the uncertainties in the data obtained from
Huff and Puff, is applied in reservoirs with limited drainage areas. Some various measurements and used to characterize the reservoir. Hence, the
other CO2 injection strategies also got attention such as CO2 simulta­ probabilistic method is more recommended to be applied as it generates
neous water and gas injection (CO2-SWAG), CO2 gas assisted gravity a range of possible outcomes based on the defined ranges of input values
drainage (CO2-GAGD), CO2 vapor extraction (CO2-VAPEX) which is (Asante et al., 2021). Additionally, probabilistic analysis is applied to
especially applied in heavy oil reservoirs, and intermittent CO2 quantify input data uncertainties, such as geological uncertainties, and
injection. their subsequent impact on the obtained results. Several previous studies
Fluid separation in any formation is based on the gravity segregation aimed to quantify the uncertainty of complex multiphase fluid transport
of the existing fluids based on their densities. Gas has the least density in porous media including CO2 geo-sequestration and EOR (Asante et al.,
compared to oil and water, and hence migrates to the top parts of the 2021; Pan et al., 2016; Vo Thanh et al., 2020; Dai et al., 2014). Different
formation forming a gas cap above the oil phase, the water phase exists methods were applied in the probabilistic analysis and uncertainty
at the bottom of the reservoir. Based on the concept of gravity segre­ quantification process including the Monte Carlo method, Polynomial
gation, gases are injected into the crest of the reservoir to maintain the Chaos Expansion (PCE) method, and Response Surface Method (RSM).
reservoir pressure and improve the oil recovery. An injection strategy Although the Monte Carlo method was found to be successful in both
named gas-assisted gravity drainage (GAGD) has been proposed as an linear and non-linear multi-phase flow in porous media, it requires a
EOR method using different types of gases including CO2 (Rao et al., large number of simulation runs that make the process computationally
2004). The GAGD process involves injecting the gas through vertical expensive. Polynomial Chaos Expansion (PCE) method was also applied
wells perforated at the top of the formation, while the oil production to quantify the risk associated with the geological storage of CO2 in
takes place through a horizontal well located at the bottom of the underground formations (Ashraf et al., 2013). However, the computa­
reservoir (Bautista et al., 2014). CO2-assisted gravity drainage improves tional cost of the PCE method limits its application in highly non-linear
the volumetric sweep efficiency and oil production through the gravity flow models (Pan et al., 2016). On the other side, the RSM is the most
segregation effects along the horizontal well. Furthermore, CO2 in the commonly applied method to quantify the uncertainty of complex sys­
GAGD process also improves displacement efficiency due to the mutual tems, especially in CO2 injection applications. The RSM is easily applied
interactions between the injected CO2 and the crude oil (Rezk et al., and clearly illustrates the relationship between independent variables, i.
2019). e., input variables, and one or more objective functions. Recently,
The intermittent CO2 injection strategy cycles between an injection/ several studies applied the RSM to CO2 geo-sequestration and cap rock
production period and a soaking period. The soaking period is intro­ failure (Wriedt et al., 2014), CO2-EOR and storage in oil reservoirs (Pan
duced to enhance the contact time between the injected CO2 and the et al., 2016; Dai et al., 2014), and CO2 sequestration in saline aquifers
reservoir fluids. Therefore, intermittent CO2 injection was applied in oil (Liu and Zhang, 2011).
reservoirs to improve CO2 diffusion and dissolution into the reservoir It is important to highlight here that the simulation work of CO2
oil. Consequently, oil properties are improved, and the recovery factor is injection application in porous media, either for EOR or sequestration or
increased. Seyyedsar et al. (Seyyedsar et al., 2015) experimentally both, requires large sets of input data and takes a long processing time,
investigated the performance of intermittent CO2 injection in a heavy oil especially for sensitivity analysis and optimization processes. These is­
reservoir. Their experimental results showed that the soaking alter­ sues can be avoided by applying Artificial Neural Network (ANN) to
nating CO2 injection method improved oil recovery, while less amount create a model that accurately describes the reservoir behavior and is
of CO2 was used in the flooding process. In a different study, Seyyedsar later used for prediction purposes. Additionally, the generated proxy

2
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

models can be employed in complicated reservoir engineering cases followed in this study. Section 3 presents the results and discussion that
when reservoir simulations cannot be used. ANN was used in several includes the results of base case model, machine learning training and
studies in petroleum engineering such as drilling engineering (Alsabaa testing, and uncertainty quantification of oil recovery and gas storage
et al., 2022; Elkatatny et al., 2019), assisted history matching (Shah­ with highlighting the main influential input parameters. At last, the
karami et al., 2014), oil production flow rates prediction (Ibrahim et al., main conclusions of this work are summarized in Section 4.
2022), water saturation estimation (Adebayo et al., 2015), pressure
losses in production wells (Adebayo et al., 2013), and permeability 2. Methods
predictions (Adeniran et al., 2019). Kim et al. (Kim et al., 2017) applied
ANN to predict the amount of CO2 storage in deep saline aquifers. Their 2.1. Reservoir simulation model
results showed that the developed ANN model can accurately predict the
storage efficiency of CO2 sequestration in saline aquifers. Ahmadi et al. A 3D reservoir model was created using CMG compositional simu­
(Ahmadi et al., 2016) used ANN to predict the viscosity and thermal lator (CMG-GEM). The compositional model has dimensions of 550 ft. ×
conductivity of carbon dioxide at a wide range of pressures and tem­ 825 ft. × 70 ft., corresponding to length, width, and height, respectively.
peratures. ANN was also applied in CO2-EOR projects. You et al. (You A Cartesian grid system was used, and the optimized numbers of grids
et al., 2020) coupled ANN and multi-objective optimizers to maximize are as follows: 55 cells in the I direction, 55 cells in the J direction, and
the volume of oil recovered, the volume of CO2 storage, and project 10 layers in the K direction. The base case model has a porosity equal to
economic outcomes of a CO2-EOR project. The application of ANN for 20%. The horizontal permeability is 30 mD, while the vertical perme­
predicting oil recovery and CO2 storage in residual oil formations was ability is 10 mD. The PVT model used in the simulation was tuned to
recently studied by Thanh et al. (Vo Thanh et al., 2020). The uncertainty match experimentally measured CO2-reservoir oil interaction (Rezk and
parameters that were used in their study include reservoir properties, e. Foroozesh, 2019a). The reservoir properties were selected based on the
g., porosity, permeability and thickness, and operating conditions, e.g., properties of core samples collected from one of the fields in Malaysia
CO2 injection rate. While the investigated objective functions are cu­ (Rezk et al., 2019). The relative permeability data are generated using
mulative oil production, cumulative gas storage, and cumulative CO2 Corey’s correlation using the following data: connate water saturation
retention. The ANN models showed excellent prediction performance equal to 20%, the irreducible oil saturation for the gas-liquid system is
and were validated across five real residual oil formations. 5%, the connate gas saturation is 5%, and Corey’s exponents were
As we illustrated above, the main goal of the uncertainty analysis of a considered 2. It should be noted that CO2 solubility in the reservoir
specific reservoir is to quantify and reduce the total uncertainty that will water, and CO2 diffusion in the reservoir oil were included in the model.
directly lead to a reliable output. Hence, the most contributing param­ CO2 solubility in the aqueous phase was described by Henry’s constant
eters to the total uncertainty need to be defined. Most of the previous calculated using Henry’s law (Li and Nghiem, 1986), while CO2 diffu­
studies investigating uncertainty quantification of CO2-EOR have sion was modeled by inputting a previously measured value of the CO2
considered either continuous CO2 injection or WAG injection strategy. diffusion coefficient in the same light crude oil sample (Rezk and For­
Additionally, previous studies mainly focus on the quantification of oozesh, 2019b). Table 1 summarizes the reservoir model properties and
geologic parameters uncertainties with less focus on the operating pa­ operating conditions. To simulate the intermittent CO2-GAGD process,
rameters and rock-fluid properties. Hence, in this study, we aim to three vertical injectors were perforated across the top three layers of the
quantify the uncertainty of intermittent CO2-assisted gravity drainage reservoir, and one horizontal producer was perforated in the 9th layer.
EOR process by focusing on key uncertain parameters that include The reservoir simulation model is shown in Fig. 1. The reservoir
reservoir properties, i.e., permeability heterogeneity, porosity, and boundary conditions are selected to be no-flow boundary conditions to
compressibility, fluid properties, i.e., CO2 diffusion coefficient and mimic the real reservoir conditions. The operating constraints of the
Henry’s constant, and rock-fluid properties, i.e., gas-oil relative injectors were surface injection rate, set at 130 Mft3/day, and bottom
permeability data. RSM was adopted in this paper to investigate the hole pressure was 7000 psi. While the production well was operating
impact of uncertainties of independent input parameters on key output with a single constraint, i.e., bottom hole pressure equal to 2500 psi. The
responses including oil recovery factor, total CO2 trapping, CO2 solu­ intermittent CO2 injection strategy was applied in the compositional
bility trapping, and mole fraction of CO2 in the oil phase. Furthermore, model to improve the oil recovery and gas storage by increasing the
ANN was applied to generate predictive models that accurately describe contact time between the injected CO2 and the reservoir fluids. The time
and predict the performance of CO2-EOR and storage during the inter­ scale of the intermittent injection process that cycles between injection
mittent CO2-GAGD process in oil reservoirs. Hence, we aim to generate a and production period and a soaking period, can be shown in Fig. 2. It
rapid tool that reduces the computational expenses of the compositional can be observed from Fig. 2 that the injection & production-soaking
reservoir simulation process. Therefore, a compositional reservoir cycle time was increased towards the end of the CO2 injection process.
simulation model of intermittent CO2-GAGD was created using PVT and This was followed to improve the contact time between the injected CO2
reservoir data of a Malaysian oil field. The equation of state (EOS) was and the residual oil in the reservoir, therefore, more oil can be recovered
tuned using lab-measured CO2-oil PVT properties. Thereafter, CMOST-
Artificial Intelligence (CMOST-AI) software from CMG was used to Table 1
quantify the uncertainties associated with the intermittent CO2 injection Reservoir model properties and operating conditions.
process. Latin Hypercube Design (LHD) was the method used in
Property Value
designing the experiments. This is followed by running the designed
experiments using the compositional reservoir simulation (CMG-GEM). Model dimensions (I × J) 550 × 825 ft
Reservoir model thickness 70 ft
Then, artificial neural network (ANN) was used to generate proxy
Grid cell number (I × J × K) 55 × 55 × 10
models that accurately describe the model performance. Finally, un­ Reservoir porosity, φ 20%
certainty quantifications for oil recovery and CO2 storage were carried Horizontal permeability, KH 30 mD
out using the generated ANN models. Vertical permeability, KV 10 mD
The next section presents the methodology, which includes a Initial reservoir pressure 2500 psi
Reservoir temperature 204 ◦ F
description of reservoir model properties and operating conditions, Total compressibility 1 × 10− 6 psi− 1
illustration of response surface methodology and workflow of the un­ Connate water saturation 20%
certainty quantification process for oil recovery and CO2 storage, iden­ CO2 injection rate 130 Mft3/day
tification of uncertain independent input parameters used in designing CO2 diffusion coefficient (D), (D coefficient in bulk oil) 80 х 10− 9 m2/s
Boundary conditions (BC) No flow BC
the experiments, and description of ANN machine learning technique

3
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

Fig. 1. Reservoir simulation model used in the intermittent CO2-GAGD process with an illustration of wells orientation and completion.

Fig. 2. Time scale of intermittent CO2-GAGD process.

by the effect of the CO2 diffusion into the reservoir fluids. models. Fig. 3 shows a workflow diagram that summarizes the steps
followed in this study; starting from creating the reservoir simulation
model for intermittent CO2-GAGD process, passing through data prep­
2.2. Response surface method (RSM) aration, designing the experiments, running reservoir simulations, and
ending with uncertainty quantification of oil recovery and CO2 storage
The response surface method (RSM) is one of the most applied using the models generated by the ANN, after they were tested and
methods in uncertainty quantification of complex multi-phase flow validated.
processes in porous media. In this study, CMOST-Artificial Intelligence
(CMOST-AI) software from CMG was used in the uncertainty quantifi­
cation for the intermittent CO2 injection process. In the RSM, multiple 2.3. Uncertain independent parameters
independent parameters can be adjusted together then the obtained
results are analyzed by fitting a response surface, e.g., artificial neural There are various sources of uncertainties associated with the pro­
network (ANN), to the results. Hence, for CO2-EOR and storage appli­ cess of quantifying CO2-EOR and storage in oil reservoirs. It should be
cations, the impact of multiple parameters on gas storage and oil re­ noted that the uncertainty in the obtained reservoir simulation results
covery can be easily highlighted using the RSM. Nevertheless, the RSM arises from the uncertainty in the data used to describe the reservoir
needs an effective numerical experiment design to generate the opti­ properties, reservoir rock-fluid interactions, and operating conditions.
mum experimental points, i.e., simulation runs, before running the nu­ In this study, twelve independent parameters were selected as uncertain
merical simulation. Several methods were used in designing the parameters. These parameters include reservoir permeability in the
numerical experiments that include central composite design, factorial horizontal direction (KH), permeability anisotropy (ratio of vertical and
design, mixture design, Doehlert design, Box-Behnken Design (BBD), horizontal permeability, KV/KH), reservoir porosity (φ), formation
and Latin Hypercube Design (LHD). More details about the methods of compressibility (Cf), gas relative permeability at residual liquid satura­
design can be found in Ferreria et al. study (Ferreira et al., 2007). Among tion (Krg), oil relative permeability at connate gas saturation (Krog), re­
the design methods, the LHD method is more flexible and widely used in sidual oil saturation to gas flooding (Sorg), connate gas saturation (Sgc),
designing experiments (Viana, 2016). The LHD was the method used in Corey’s exponent for gas (Ng), Corey’s exponent for oil (Nog), CO2
CMOST-AI for designing the simulation experiments. The workflow of diffusion coefficient in the oil phase, and Henry’s constant for gas sol­
RSM of this study after creating the reservoir simulation model is sum­ ubility in brine. It is important to mention here that these parameters
marized in the following steps: (1) uncertain independent parameters were selected due to their high impact on fluid flow behavior in porous
were determined; (2) the objective functions were defined (3) the media, which will be directly reflected on the oil recovery, CO2 storage
CMOST-AI was then used to design the experiment based on the LHD in the reservoir (overall trapping of CO2), solubility trapping of the
method; (4) the designed simulation runs were conducted using the injected CO2 in reservoir brine, and also oil compositional changes.
created intermittent CO2-GAGD simulation model by CMG-GEM Hence, the objective functions defined for this study are the oil recovery
compositional simulator; (5) ANN was used to generate proxy models factor, total gas trapping, solubility trapping of CO2, and CO2 mole
that accurately delineate the system responses after training, testing, fraction in the oil phase. Table 2 summarizes the uncertainty parameters
and validating the generated models; (6) uncertainty quantification for and their corresponding values. It is essential to highlight that the values
oil recovery and CO2 storage was carried out using the generated ANN of the CO2 diffusion coefficient in the oil phase and Henry’s constant,

4
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

Fig. 3. Workflow diagram that links the reservoir simulation (physics-based model) and ANN (pure mathematical model).

for different scenarios of input data. Table 3 summarizes the statistical

Table 2 analysis for the input and output data where the minimum and
Uncertainty-independent parameters used in designing the reservoir simulation
maximum oil recovery factors (RF) are 52% and 86%, respectively.
experiments.
In this study, ANN machine learning technique was applied to the
Uncertainty parameters Minimum Maximum Units collected data. The ANN technique was developed using MATLAB pro­
Compressibility 7.5 × 10− 7
1 × 10− 5
psi− 1 graming language. Four different models were developed, where one
Porosity 0.15 0.25 ___ model predicts one parameter of the output parameters including, oil
Horizontal permeability, KH 3 300 mD
recovery factor, dissolved CO2 volume in both oil and water phases, and
Permeability anisotropy, KV/KH 0.1 0.7 ___
Gas relative permeability at residual liquid 0.4 1 ___ total CO2 volume trapped in the reservoir. The data set (173 points) was
saturation, Krg used to develop the models after optimizing the training to testing
Oil relative permeability at connate gas 0.4 1 ___ datasets splitting ratio. The quality of the model was measured using
saturation, Krog absolute average percentage error (AAPE) which represents the error
Corey’s exponent for gas, Ng 1.5 3.5 ___
Corey’s exponent for oil, Nog 1.5 3.5 ___
between the actual output value and the estimated values from the ML
Connate gas saturation, Sgc 0.03 0.08 ___ model. Moreover, Spearman’s correlation coefficient was used to
Residual oil saturation to gas flooding, Sorg 0.04 0.15 ___ describe the goodness of fit between the actual output value and the
9 9
CO2 diffusion coefficient in the oil phase, DCO2- 40 х 10− 120 х 10− m2/s estimated values from the ML model. The following equations were used
oil
to estimate the AAPE and R values.
Henry’s constant, H 43,873 73,121 psi
∑
N
yi actual − yi predicted
yi
× 100%
shown in Table 2, were chosen to represent 50% and 25% uncertainty in
actual
AAPE = i=1
(1)
N
the measure/calculated base case values, respectively. The statistical
distributions of the uncertainty parameters, shown in Table 2, were [∑
(yi μx )(yi μy )
]
actual − predicted −
considered to have uniform distributions. After defining the uncertainty N− 1
parameters and the objective interests, the LHD method was used in R= (2)
σx σy
CMOST-AI for designing the simulation experiments. Hence, 173 simu­
lation jobs were created and then submitted to CMG-GEM compositional where, yi actual and yi predicted are the actual and the predicted output
simulator to conduct the simulation runs. In the next section, the ANN value, respectively, and N represents the number of points in the Data­
will use the simulation results to generate models for the defined set. μx, μy, and σx, σy are the mean and the standard deviation for the
objective functions. actual and the predicted output value, respectively.
Fig. 4 presents the process that was applied to develop the different
2.4. ANN (objective functions, training, testing, validation) ML models. After data collection and transformation, the data set was
divided randomly into training and testing sets to build the ML model.
This section shows the application of ANN method to predict the CO2 The model’s performance was optimized by adjusting the hyper­
injection performance including, oil recovery factor, total trapped CO2 parameters for the ML model. After reaching the satisfying model per­
volume, dissolved CO2 part in the oil and formation water. The LHD formance results, the model was tested on another data set. Finally, the
method provided different numerical simulation cases of 173 data points

5
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

Table 3
Statistical analysis for the input and output data.
Cf, E-6 DCO2-oil H Krg Kro Ng Nog KH Kv/KH ∅ Sgc SORG RF VCO2 in Oil VCO2 in Water Cum. Gas Prod.

mean 5.2 2.3E-04 5.9E+04 0.69 0.70 2.5 2.5 151 0.40 0.20 0.05 0.10 76 5.9E+05 5.6E+07 3.8E+09
std 2.9 7.4E-05 9.1E+03 0.19 0.19 0.6 0.6 95 0.20 0.03 0.02 0.04 6 2.1E+05 1.1E+07 1.6E+08
min 0.75 1.2E-04 4.4E+04 0.40 0.40 1.5 1.5 3 0.10 0.15 0.03 0.04 52 1.7E+05 3.3E+07 3.5E+09
25% 2.6 1.7E-04 5.0E+04 0.52 0.52 2.0 1.9 62 0.22 0.17 0.04 0.06 73 4.6E+05 4.8E+07 3.7E+09
50% 5.4 2.3E-04 5.8E+04 0.70 0.70 2.5 2.5 152 0.40 0.20 0.06 0.10 77 5.9E+05 5.5E+07 3.8E+09
75% 8.0 3.0E-04 6.7E+04 0.88 0.87 3.1 3.1 241 0.58 0.22 0.07 0.13 80 7.3E+05 6.2E+07 4.0E+09
max 10 3.5E-04 7.3E+04 1.00 1.00 3.5 3.5 300 0.70 0.25 0.08 0.15 86 1.2E+06 8.6E+07 4.1E+09

objective functions, we presented the base case results here in this sec­
Data collecon and tion. The base case properties were previously described in Section 2.1.
processing
Fig. 5 shows the oil recovery factor (RF), calculated as RF =
Volume of oil recovered
Volume of original oil in place × 100, due to the intermittent CO2 injection
Model training by process. The total CO2 trapping, calculated as the cumulative volume of
opmizing the Random split the data;
Tesng; Training CO2 injected − cumulative volume of CO2 produced, during the gas
hyperparmeters
injection was also shown in Fig. 5. Fig. 5 shows that CO2 trapping profile
is following the same trend as the oil recovery profile. This is expected
since by increasing the oil recovery, more pore space is available for CO2
Model tesng and adjust storage. Nevertheless, various CO2 trapping mechanisms can coexist
the spling rao
during the CO2-EOR process. Here in this study, we focus on the struc­
tural and solubility trapping mechanisms. Fig. 6 shows CO2 structural
trapping in the supercritical state and the amount of CO2 dissolved in the
reservoir brine. Since the connate water saturation is low in the reservoir
Uncertanity analysis model, the CO2 dissolution into the reservoir brine did not have a sig­
nificant contribution to the total CO2 trapping compared to the struc­
Fig. 4. Procedures of developing the different ML models. tural trapping, i.e., supercritical CO2. It can be also observed from Fig. 6
that most of the CO2 dissolution in brine occurred in the first 30 years of
optimized models were used to investigate the uncertainty of the input the simulation process, while the amount of the supercritical CO2 re­
data and its effect on the output data. mains increasing as more CO2 was injected into the reservoir. As it was
mentioned earlier, CO2 diffusion into the reservoir oil was considered in
3. Results the compositional model. Hence, oil phase compositional changes are
expected to occur even far away from the injectors and producers, due to
3.1. Base case results CO2-oil mass transfer process. This can be confirmed by monitoring the
gas saturation profile and CO2 mole fraction in the oil phase as shown in
As it was mentioned earlier, the intermittent CO2-GAGD strategy was Figs. 7 and 8, respectively. Due to the gravity segregation, the low-
adopted in this study. We aim to quantify the uncertainty of oil recovery density CO2 propagates to top parts of the reservoir and pushes the
and gas storage due to the CO2 injection in oil reservoirs. The impact of underlying oil towards the horizontal producer, as depicted in Fig. 6.
the uncertainty in the independent parameters, defined in Table 2, was However, the supercritical CO2 diffused into the reservoir oil, especially
tested in terms of four objective functions, which are: oil recovery factor, during the soaking periods, and consequently, the oil composition has
total CO2 trapping, CO2 solubility trapping, and CO2 mole fraction in the changed, even at the bottom layers, as can be shown in Fig. 8. Therefore,
oil phase. In order to have a better understanding of the defined it was essential to include the gas diffusion in the simulation process as it

70 12
Oil Recovery Factor
Total CO2 Trapping
60 10
Total CO2 Trapping, ×108 3

50
Oil Recovery Factor, %

8
40
6
30
4
20

10 2

0 0
0 5 10 15 20 25 30 35 40 45 50 55 60
Time, years

Fig. 5. Oil recovery factor and total CO2 trapping during intermittent CO2-GAGD process.

6
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

14
Supercrical CO2

Supercrical & Dissolved CO2, ×108 moles

12 Dissolved CO2 in Aqueous Phase

10

0
0 10 20 30 40 50 60
Time, years

Fig. 6. CO2 trapping mechanisms during intermittent CO2-GAGD process; CO2 at supercritical state and CO2 dissolved in aqueous brine.

highly affects the oil recovery and CO2 storage during CO2-EOR.

3.2. Machine learning model results

ANN technique was applied on 173 data points to train and test the
model. The optimum training testing ratio was found to be 70/30. The
optimized hyper-parameters were selected based on the best model
quality indicators. The optimum ANN model was developed with a
single hidden layer and 10 neurons. Figs. 9-12 presents the performance
of the different ML models in both training and testing data sets. The ML
models were able to predict the output parameters including the trapped
CO2 volume, oil recovery factor, and the dissolved CO2 volume in the oil
and water phases.
The correlation coefficient (R) value for the training set was found to
be 0.98 with an AAPE error between 1.2 and 6% compared to AAPE
between 2.6 and 9% in the testing data set.

3.3. Uncertainty quantification for oil recovery and CO2 storage

Fig. 7. A reservoir model 3D view of gas saturation distribution, at a cutting
plane of J = 34, after 28 years of intermittent CO2-GAGD injection. The result of the simulation runs conducted in this study were used to
generate probabilistic distributions of the output variables, i.e., objec­
tive functions. Fig. 13 depicts the cumulative distribution functions
(CDF) of the oil recovery factor and total CO2 trapping, at the end of the
intermittent CO2-GAGD process. While the CDF of CO2 moles in the oil
phase and dissolved in the aqueous phase are shown in Fig. 14. As we
mentioned earlier, there are many sources of uncertainties in the data
characterizing the reservoir, therefore, it is meaningful to quantify the
uncertainties using the CDF and probability percentages, i.e., P10, P50,
and P90. It is essential to highlight here that two measures of probability
were applied in uncertainty analysis: exceedance probability and cu­
mulative probabilities. For instance, if the exceedance probability of an
output parameter is Px, that means that the parameter was exceeded in x
% of the cases. On the other side, the cumulative probabilities are the
opposite of exceedance probabilities, i.e., Px in the exceedance proba­
bility is equivalent to P(1-x) in the cumulative probability. Here, in this
study, we presented the results of the objective functions in terms of
cumulative probabilities. Table 4 shows the cumulative probabilities of
the four objective functions covered in this paper. The corresponding
values of the oil recovery factor at the 10th, 50th, and 90th percentiles
Fig. 8. A reservoir model 3D view of oil mole fraction of CO2 component, at a
are 66.68%, 77.06%, and 82.58%, respectively. This wide range sug­
cutting plane of J = 34, after 28 years of intermittent CO2-GAGD injection. gests that the high uncertainty of the oil recovery propagates from the
uncertainties of the input parameters. To analyze the uncertainties of the
CO2 storage capacity and CO2 trapping mechanisms, we calculated the

7
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

1.E+09 1.E+09
R = 0.99 R = 0.98
1.E+09 AAPE = 1.2% 1.E+09 AAPE = 2.6%

Predicted Value
Predicted Value

1.E+09 1.E+09

8.E+08 8.E+08

6.E+08 6.E+08

4.E+08 4.E+08

2.E+08 2.E+08
a b
0.E+00 0.E+00
0.E+00 5.E+08 1.E+09 2.E+09 0.E+00 5.E+08 1.E+09 2.E+09

Actual Value Actual Value

Fig. 9. The actual versus estimated total trapped CO2 volume cross plot for the ANN model (a) the training and (b) the testing data sets.

1.E+02 1.E+02
R = 0.99 R = 0.96
9.E+01 AAPE = 1.5% 9.E+01 AAPE = 2.6%
Predicted Value

Predicted Value

8.E+01 8.E+01

7.E+01 7.E+01

6.E+01 6.E+01

5.E+01 5.E+01
a b
4.E+01 4.E+01
4.E+01 6.E+01 8.E+01 1.E+02 4.E+01 6.E+01 8.E+01 1.E+02

Actual Value Actual Value

Fig. 10. The actual versus estimated oil recovery factor cross plot for the ANN model (a) the training and (b) the testing data sets.

1.E+06 1.E+06
R = 0.98 R = 0.93
1.E+06 1.E+06
AAPE =6% AAPE =9%
Predicted Value

Predicted Value

1.E+06 1.E+06

8.E+05 8.E+05

6.E+05 6.E+05

4.E+05 4.E+05

2.E+05 2.E+05
a b
0.E+00 0.E+00
0.E+00 5.E+05 1.E+06 2.E+06 0.E+00 5.E+05 1.E+06 2.E+06

Actual Value Actual Value

Fig. 11. The actual versus estimated CO2 dissolved volume in oil phase cross plot for the ANN model (a) the training and (b) the testing data sets.

cumulative probabilities of total CO2 trapping, CO2 in oil phase, and ft3, 12.30 × 108 ft3, respectively. Hence, the uncertainty of the input
dissolved CO2 in the aqueous phase. The corresponding values of the parameters also resulted in significant uncertainties in the volume of
total CO2 trapping at P10, P50, and P90 are 7.95 × 108 ft3, 10.13 × 108 CO2 storage in the porous medium. More specifically, the uncertainties

8
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

1.E+08 1.E+08
9.E+07 R = 0.99 9.E+07 R = 0.97
8.E+07 AAPE = 1.5% 8.E+07 AAPE = 3%

Predicted Value
Predicted Value

7.E+07 7.E+07
6.E+07 6.E+07
5.E+07 5.E+07
4.E+07 4.E+07
3.E+07 3.E+07
2.E+07 2.E+07
1.E+07 1.E+07
a b
0.E+00 0.E+00
0.E+00 5.E+07 1.E+08 0.E+00 5.E+07 1.E+08

Actual Value Actual Value

Fig. 12. The actual versus estimated CO2 dissolved volume in water phase cross plot for the ANN model (a) the training and (b) the testing data sets.

Fig. 13. Cumulative distribution function of oil recovery factor and total CO2 trapping, at the end of the intermittent CO2-GAGD process.

of the total CO2 trapping can be explained by the uncertainties of the oil recovery factor is positively depending on the horizontal perme­
various CO2 trapping mechanisms, e.g., solubility trapping of CO2. To ability, oil relative permeability, and CO2 diffusion coefficient in oil
further explain this, the values of the 10th, 50th, and 90th percentiles phase compared to a negative relation with the gas relative permeability
cumulative probabilities of the moles of CO2 in the oil phase and dis­ and almost no dependency on the ratio of vertical to horizontal
solved in the reservoir brine were calculated and presented in Table 4. It permeability, formation compressibility, and residual gas saturation.
was also noted from the obtained results that a specific case might give a While the CO2 dissolution trapping positively related to the formation
P90 oil recovery and a P10 total CO2 trapping. Therefore, no case gives porosity and negatively related to Henry constant, H. The effect of
P90 results for more than one output variable. Based on the significant changing values of other independent parameters on the CO2 stored and
uncertainties of the oil recovery and CO2 storage, it is suggested that a oil recovered can be easily recognized from Fig. 15.
detailed reservoir characterization should be conducted to reduce the As we mentioned earlier, the total uncertainty in a system response is
uncertainties of the rock and rock-fluid properties, which will lead to a controlled by the degree of uncertainty of independent input parameters
more reliable simulation of the intermittent CO2-GAGD process. for such a system. Hence, we found that it is necessary to quantify the
Moreover, to investigate the importance of the different input pa­ impact of the most influential independent parameters on the objective
rameters on the performance of the trapped CO2 and the recovery factor, functions investigated in this study. Fig. 16 shows four pie charts that
a sensitivity analysis was conducted using the developed ML models quantify the impact of the main independent input parameters on total
with generating 10,000 realizations from the input and output param­ uncertainties in oil recovery factor, dissolved CO2 in oil, dissolved CO2
eters. Fig. 15 presents bar charts for the correlation coefficient between in water, and total trapped CO2. One can observe from Fig. 16 that the
the different input parameters and the different output parameters formation porosity, shown in orange color, is an essential parameter for
estimated using the ML models. R values varied between − 1 and 1, with accurate estimation of both oil recovery and CO2 storage. The reservoir
1 for a strong positive relationship while − 1 for a strong negative porosity is the dominant parameter controlling the total CO2 trapping in
relationship between the input and output parameters. For instance, the the reservoir, which is indicated by a percentage 87% contribution to

9
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

Fig. 14. The cumulative distribution function of CO2 moles in the oil phase and dissolved in the aqueous brine, at the end of the intermittent CO2-GAGD process.

permeability, and relative oil permeability. It shows that the porosity is

Table 4
the dominant factor in the trapped CO2 volume, where the maximum
Cumulative probabilities of the objective functions; P10, P50, and P90.
stored volume was found to be 1.25E9 ft3 at porosity of 0.25 and hori­
Objective Oil RF Total CO2 CO2 in oil Dissolved CO2 in the zontal permeability of 250md, and relative oil permeability of 0.85.
function, (%) trapping phase (×105 aqueous phase
Similarly, a three-dimensional plot was constructed for the oil RF as a
(×108 ft3) lb.mol.) (×107 mol.)
Percentile function of the porosity, horizontal permeability, and relative oil
10th 66.68 7.95 3.15 4.19 permeability (Fig. 17b). The horizontal permeability and the oil relative
percentile permeability parameters are the controlling parameters in the oil re­
(P10) covery while the RF slightly changes vertically with changing the
50th 77.06 10.13 5.85 5.49
percentile
porosity. The maximum oil RF of 83% can be found at horizontal
(P50) permeability higher than 250md and Kro higher than 0.93 which gave
90th 82.58 12.30 8.51 7.24 effective oil permeability around 240 md. Fig. 17c presents the three-
percentile dimensional plot for the amount of dissolved CO2 in the water as a
(P90)
function of porosity, Kv/Kh, and Henry constant. Henry constant has the
highest impact in the dissolved CO2 in water followed by the formation
the total uncertainty, as shown in Fig. 16d. Fig. 16a shows that the oil porosity. The highest CO2 trapped as a dissolved phase in the water was
recovery factor is controlled by five main independent inputs; with the found at Henry constant of 4.5E4 and porosity of 25%. However, the
horizontal permeability contributed the most (39%) and the CO2 diffu­ dissolved CO2 in oil phase was dominant by the formation porosity with
sion coefficient contributed the least (5%) to the total uncertainty. Be­ less effect from the formation permeability (Fig. 17d).
sides the significant effect of the reservoir porosity, the rock-fluid
interactions properties, e.g., relative permeability, have noticeable 4. Conclusions
contributions of the total uncertainty of the compositional changes of oil
due to CO2 dissolution (see Fig. 16b). It is also interesting to find that This study captured the real physics of complex multiphase flow, i.e.,
Henry’s constant and reservoir porosity together contributed about 96% CO2 injection in oil reservoirs, through creating a compositional simu­
to the total uncertainty of the dissolution trapping of CO2 in the connate lation model that includes CO2-reservoir fluids interactions that occur
water saturation. It is worth mentioning here that parameters having the during CO2-EOR, e.g., CO2 dissolution and diffusion. The response sur­
least influence on the objective functions can be removed from the face (RSM) approach was applied to provide an efficient tool to quantify
developed machine learning models. This will facilitate using the uncertainties of independent input parameters associated with CO2
developed models for EOR and gas storage applications and also avoid storage and oil recovery during intermittent CO2-GAGD process. ANN
including unnecessary data that can be challenging to collect. models were generated to represent the relation between the indepen­
Response surface design is not only used to analyze a problem in dent inputs and various system responses. Additionally, the response
which a response, i.e., objective function, is controlled by several in­ surfaces of the uncertain inputs and their subsequent impact on
dependent parameters but it is used to optimize that response as well. dependent variables were analyzed. Based on the results of this study,
Here, response surface analysis was used to present the relation between the following conclusions could be drawn:
the main inputs and the output parameters (Fig. 17).
Fig. 17a shows three-dimensional plots of the response surfaces for • CO2 diffusion into the reservoir oil had a significant impact on oil
the trapped CO2 volume as a function of porosity, horizontal compositional changes even far away from the injectors and

10
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

Oil RF Dissolved CO2 in Oil

Kro Kro
Kh Kh
D D
Ng Ng
H H

Cf Cf
Sgorg Sgorg
Nog Nog
SORG SORG
Kv/Kh Kv/Kh
Porosity Porosity

Krg Krg

-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1

Correlation Coeffecient With Oil_RF Correlation Coeffecient With Dissolved CO2 in
Oil

Dissolved CO2 in Water Trapped CO2

Kro Kro
Kh Kh
D D
Ng Ng
H H

Cf
Cf
Sgorg
Sgorg
Nog
Nog
SORG
SORG
Kv/Kh
Kv/Kh
Porosity
Porosity
Krg
Krg
-1 -0.5 0 0.5 1
Correlation Coeffecient With Dissolved CO2 in -1 -0.5 0 0.5 1
Water Correlation Coeffecient With Trapped CO2

Fig. 15. The correlation coefficient between the different input parameters and the output parameters based on the sensitivity analysis.

producers, which directly affected the calculated oil recovery and in place (OIP) can recovered (P10), 77.06% of OIP can probably
solubility trapping of the injected CO2. recovered (P50), and it is possible to store as much as 82.58% (P90).
• 173 simulation jobs were designed to generate a database that was While the P10, P50, and P90 for the total volume CO2 trapped are
used to train and test the generated ANN models. The optimum 7.95 × 108 ft3, 10.13 × 108 ft3, and 12.30 × 108 ft3, respectively.
training testing ratio was found to be 70/30. The optimized hyper- • The large intervals between the 10th percentile (P10) and the 90th
parameters were selected based on the best model quality in­ percentile (P90) of the calculated objective functions showed the
dicators. The optimum ANN models, developed with a single hidden significant impact of uncertainties of input parameters on the fore­
layer and 10 neurons, accurately predicted the objective functions casted CO2 storage and oil recovery.
including the trapped CO2 volume, oil recovery factor, the dissolved • The response surface analysis showed that the main independent
CO2 volume in water and oil phases. The correlation coefficient (R) input parameters contributing to the total uncertainty of oil recovery
for the training set with was found to be 0.98 with an AAPE error can be ordered in a sequence of decreasing importance as follows:
between 1.2 and 6% comparing to AAPE between 2.6 and 9% in the horizontal permeability > oil relative permeability > porosity > gas
testing data set. relative permeability > CO2 diffusion coefficient in oil.
• The probabilistic distributions of the output variables, i.e., objective • Formation porosity was found to have a significant impact on the
functions, generated in this study showed that at least 66.68% of oil calculated oil recovery and CO2 storage. For instance, porosity

11
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

Fig. 16. Pie charts showing the impact of various independent input parameters on total uncertainties in: (a) oil RF, (b) dissolved CO2 in oil, (c) dissolved CO2 in
water, and (d) trapped CO2.

contributed about 52% and 43% to the total uncertainty of the CO2 KH Horizontal permeability
dissolution in oil and water, respectively. Notably, CO2 dissolution Krg gas relative permeability at residual liquid saturation
trapping in reservoir brine is positively correlated with the reservoir Krog oil relative permeability at connate gas saturation
porosity and inversely correlated with Henry’s constant. KV Vertical permeability
• The maximum oil RF of 83% was found at horizontal permeability KV/KH permeability anisotropy (ratio of vertical and horizontal
higher than 250 md and Kro higher than 0.93 that gave effective oil permeability
permeability around 240 md. While the maximum CO2 dissolution in ML machine learning
water was found at porosity near to 25%. Nog,Ng Corey’s exponents
• ML and response surface analysis defined the effect of the most PCE Polynomial Chaos Expansion
influencing independent parameters on CO2 storage and EOR pro­ RF oil recovery factor
cess, which can help in reducing uncertainty in CO2 injection RSM response surface method
projects. Sorg residual oil saturation to gas flooding
σx,σy the standard deviation for the actual and the predicted output
Nomenclatures value
φ Reservoir porosity
AAPE absolute average percentage error μx,μy the mean for the actual and the predicted output value
ANN artificial neural network
CCUS carbon capture, utilization, and storage Author contributions
CDFs cumulative distribution functions
Cf formation compressibility The manuscript was written through the contributions of all authors.
CO2-EOR CO2 enhanced oil recovery All authors have approved the final version of the manuscript.
CO2-GAGD CO2 gas assisted gravity drainage
CO2-SWAG CO2 simultaneous water and gas injection
Declaration of Competing Interest
D CO2 diffusion coefficient
H Henry’s constant for gas solubility in brine
The authors declare that they have no known competing financial

12
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

Fig. 17. Three-dimensional response surfaces generated with ML models to show the relationships of the major input parameters output parameters: (a) trapped CO2,
(b) oil recovery factor, (c) dissolved CO2 in water, and (d) dissolved CO2 in oil.

interests or personal relationships that could have appeared to influence Ferreira, S.L.C., et al., 2007. Statistical designs and response surface techniques for the
optimization of chromatographic systems. J. Chromatogr. A 1158 (1), 2–14, 2007/
the work reported in this paper.
07/27/.
I. e. agency, 2018. Global Energy and CO2 Status report - 2017. Available: https://www.
Data availability iea.org/publications/freepublications/publication/GECO2017.pdf.
Ibrahim, A.F., Al Dhaif, R., Elkatatny, S., 2022. New generalized correlations for oil rate
predictions through wellhead chokes for high GOR reservoirs. Arab. J. Geosci. 15
Data will be made available on request. (12), 1137, 2022/06/08.
Iglauer, S., 2017. CO2–water–rock wettability: variability, influencing factors, and
implications for CO2 geostorage. Acc. Chem. Res. 50 (5), 1134–1142, 2017/05/16.
References
Jia, B., Tsau, J.-S., Barati, R., 2019. A review of the current progress of CO2 injection
EOR and carbon storage in shale oil reservoirs. Fuel 236, 404–427.
Adebayo, A.R., Abdulraheem, A., Al-Shammari, A.T., 2013. Promises of artificial Kim, Y., Jang, H., Kim, J., Lee, J., 2017. Prediction of storage efficiency on CO2
intelligence techniques in reducing errors in complex flow and pressure losses sequestration in deep saline aquifers using artificial neural network. Appl. Energy
calculations in multiphase fluid flow in oil wells. In: SPE Nigeria Annual 185, 916–928, 2017/01/01/.
International Conference and Exhibition. OnePetro. Kumar, S., Foroozesh, J., Edlmann, K., Rezk, M.G., Lim, C.Y., 2020. A comprehensive
Adebayo, A.R., Abdulraheem, A., Olatunji, S.O., 2015. Artificial intelligence based review of value-added CO2 sequestration in subsurface saline aquifers. J. Nat. Gas
estimation of water saturation in complex reservoir systems. J. Porous. Media 18 (9), Sci. Eng. 81, 103437.
893–906, 2015-08-31. Li, Y.K., Nghiem, L.X., 1986. Phase equilibria of oil, gas and water/brine mixtures from a
Adeniran, A.A., Adebayo, A.R., Salami, H.O., Yahaya, M.O., Abdulraheem, A., 2019. cubic equation of state and Henry’s law. Can. J. Chem. Eng. 64 (3), 486–496.
A competitive ensemble model for permeability prediction in heterogeneous oil and Liu, B., Zhang, Y., 2011. CO2 modeling in a deep saline aquifer: a predictive uncertainty
gas reservoirs. Appl. Comput. Geosci. 1, 100004, 2019/10/01/. analysis using design of experiment. Environ. Sci. Technol. 45 (8), 3504–3510,
Ahmadi, M.A., Kashiwao, T., Rozyn, J., Bahadori, A., 2016. Accurate prediction of 2011/04/15.
properties of carbon dioxide for carbon capture and sequestration operations. Pet. Ma, Y.Z., Ma, Y.Z., La Pointe, P.R., 2011. Uncertainty analysis in reservoir
Sci. Technol. 34 (1), 97–103, 2016/01/02. characterization and management: how much should we know about what we don’t
Alsabaa, A., Gamal, H., Elkatatny, S., Abdelraouf, Y., 2022. Machine learning model for know?. In: Uncertainty Analysis and Reservoir Modeling: Developing and Managing
monitoring rheological properties of synthetic oil-based mud. ACS Omega 7 (18), Assets in an Uncertain World, vol. 96. American Association of Petroleum Geologists,
15603–15614, 2022/05/10. p. 0.
Ampomah, W., et al., 2016. Evaluation of CO2 storage mechanisms in CO2 enhanced oil Moghadasi, R., Rostami, A., Hemmati-Sarapardeh, A., 2018. Chapter three - enhanced oil
recovery sites: application to morrow sandstone reservoir. Energy Fuel 30 (10), recovery using CO2. In: Bahadori, A. (Ed.), Fundamentals of Enhanced Oil and Gas
8545–8555, 2016/10/20. Recovery from Conventional and Unconventional Reservoirs. Gulf Professional
Asante, J., Ampomah, W., Rose-Coss, D., Cather, M., Balch, R., 2021. Probabilistic Publishing, pp. 61–99.
Assessment and Uncertainty Analysis of CO2 Storage Capacity of the Morrow B Pan, F., et al., 2016. Uncertainty analysis of carbon sequestration in an active CO2-EOR
Sandstone—Farnsworth Field Unit, 14, no. 22, p. 7765. field. Int. J. Greenhouse Gas Cont. 51, 18–28, 2016/08/01/.
Ashraf, M., Oladyshkin, S., Nowak, W., 2013. Geological storage of CO2: application, Rao, D.N., Ayirala, S.C., Kulkarni, M.M., Sharma, A.P., 2004. Development of gas assisted
feasibility and efficiency of global sensitivity analysis and risk assessment using the gravity drainage (GAGD) process for improved light oil recovery. In: Presented at the
arbitrary polynomial chaos. Int. J. Greenhouse Gas Cont. 19, 704–719, 2013/11/01/ SPE/DOE Symposium on Improved Oil Recovery, Tulsa, Oklahoma. https://doi.org/
. 10.2118/89357-MS, 2004/1/1/. Available:
Bautista, E.V., Barillas, J.L.M., Dutra Jr., T.V., da Mata, W., 2014. Capillary, viscous and Rezk, M.G., Foroozesh, J., 2019a. Phase behavior and fluid interactions of a CO2-Light
gravity forces in gas-assisted gravity drainage. J. Pet. Sci. Eng. 122, 754–760. oil system at high pressures and temperatures. Heliyon 5 (7), e02057.
Dai, Z., et al., 2014. Uncertainty quantification for CO2 sequestration and enhanced oil Rezk, M.G., Foroozesh, J., 2019b. Effect of CO2 mass transfer on rate of oil properties
recovery. Energy Procedia 63, 7685–7693, 2014/01/01/. changes: Application to CO2-EOR projects. J. Pet. Sci. Eng. 180, 298–309.
Elkatatny, S., Tariq, Z., Mahmoud, M., Abdulraheem, A., Mohamed, I., 2019. An
integrated approach for estimating static Young’s modulus using artificial
intelligence tools. Neural Comput. & Applic. 31 (8), 4123–4135, 2019/08/01.

13
M.G. Rezk et al. International Journal of Coal Geology 266 (2023) 104177

Rezk, M.G., Foroozesh, J., 2022. Uncertainty effect of CO2 molecular diffusion on oil Shahkarami, A., Mohaghegh, S.D., Gholami, V., Haghighat, S.A., 2014. Artificial
recovery and gas storage in underground formations. Fuel 324, 124770, 2022/09/ intelligence (AI) assisted history matching. In: SPE Western North American and
15/. Rocky Mountain Joint Meeting vol. All Days, SPE-169507-MS.
Rezk, M.G., Foroozesh, J., Zivar, D., Mumtaz, M., 2019. CO2 storage potential during Viana, F.A.C., 2016. A Tutorial on Latin Hypercube Design of experiments. Qual. Reliab.
CO2 enhanced oil recovery in sandstone reservoirs. J. Nat. Gas Sci. Eng. 66, Eng. Int. 32 (5), 1975–1985.
233–243, 2019/06/01/. Vo Thanh, H., Sugai, Y., Sasaki, K., 2020. Application of artificial neural network for
Samara, H., Al-Eryani, M., Jaeger, P., 2022. The role of supercritical carbon dioxide in predicting the performance of CO2 enhanced oil recovery and storage in residual oil
modifying the phase and interfacial properties of multiphase systems relevant to zones. Sci. Rep. 10 (1), 18204, 2020/10/23.
combined EOR-CCS. Fuel 323, 124271, 2022/09/01. Wriedt, J., Deo, M., Han, W.S., Lepinski, J., 2014. A methodology for quantifying risk
Seyyedsar, S.M., Sohrabi, M., 2017. Intermittent CO2 and viscosity-reducing gas (VRG) and likelihood of failure for carbon dioxide injection into deep saline reservoirs. Int.
injection for enhanced heavy oil recovery. Fuel Process. Technol. 164, 1–12, 2017/ J. Greenhouse Gas Cont. 20, 196–211, 2014/01/01/.
09/01/. You, J., Ampomah, W., Sun, Q., 2020. Development and application of a machine
Seyyedsar, S.M., Farzaneh, S.A., Sohrabi, M., 2015. Enhanced heavy oil recovery by learning based multi-objective optimization workflow for CO2-EOR projects. Fuel
intermittent CO2 injection. In: SPE Annual Technical Conference and Exhibition vol. 264, 116758, 2020/03/15/.
Day 3 Wed, September 30, 2015, D031S046R003. Zhou, X., Yuan, Q., Peng, X., Zeng, F., Zhang, L., 2018. A critical review of the CO2 huff
Seyyedsar, S.M., Farzaneh, S.A., Sohrabi, M., 2017. Investigation of Low-Density CO2 ‘n’puff process for enhanced heavy oil recovery. Fuel 215, 813–824.
Injection for Enhanced Oil Recovery. Ind. Eng. Chem. Res. 56 (18), 5443–5454,
2017/05/10.

14