Mechanistic Study of CO₂-Based Oil Flooding in Microfluidics and Machine Learning Parametric Analysis

Abstract

CO₂-enhanced oil recovery (CO₂-EOR) has gained prominence as an effective oil displacement method with low carbon emissions, yet its microscopic mechanisms remain incompletely understood. This study introduces a novel high-pressure microfluidic visualization system capable of operating at 0.1–10 MPa without confining pressure and featuring stratified porous media with a 63 μm minimum throat size to provide unprecedented insights into CO₂ and CO₂-foam EOR processes at the microscale. Through quantitative image analysis and advanced machine learning modeling, we reveal that increasing the CO₂ injection pressure nonlinearly reduces residual oil saturation, achieving near-complete miscibility at 6 MPa with only 2% residual oil—a finding that challenges conventional thresholds for miscibility in heterogeneous systems. Our work uniquely demonstrates that CO₂-foam flooding not only mobilizes capillary-trapped oil films but also dynamically alters interfacial tension and the pore-scale fluid distribution, a phenomenon previously underexplored. Support Vector Regression (R² = 0.71) further uncovers a nonlinear relationship between the surfactant concentration and residual oil saturation, offering a data-driven framework for parameter optimization. These results advance our fundamental understanding by bridging microscale dynamics with field-applicable insights, while the integration of machine learning with microfluidics represents a methodological leap for EOR research.

1. Introduction

Conventional oil and gas reservoirs are gradually depleting, while unconventional reservoirs with characteristics such as ultra-low permeability and strong heterogeneity have garnered increasing attention from researchers [1,2,3,4]. However, water flooding exhibits low sweep efficiency and yields suboptimal development results [5,6,7]. Gas flooding has been recognized as the most effective method to enhance the recovery rate of highly heterogeneous reservoirs [8,9]. Compared to traditional water flooding, CO₂ flooding offers advantages such as a wide applicability range, favorable economic benefits, and high oil displacement efficiency. These benefits stem from CO₂’s lower viscosity compared to water, resulting in reduced injection pressures. Additionally, CO₂ exhibits better solubility in crude oil, enabling a significant reduction in crude oil viscosity and promoting recovery rates through expansion and dissolution processes [10,11,12]. Moreover, as a CCUS (Carbon Capture, Utilization, and Storage) technology, CO₂ flooding indirectly facilitates greenhouse gas storage [13,14,15,16]. CO₂-EOR technology demonstrates dual potential in enhancing oil recovery while enabling carbon sequestration. However, its environmental benefits require comprehensive assessment. On one hand, it can store industrial CO₂ emissions, contributing to carbon neutrality; on the other hand, combustion emissions from the extracted crude may partially offset these environmental advantages. Future efforts should focus on optimizing CO₂ sourcing (e.g., prioritizing industrial-captured CO₂ over natural gas-purified CO₂) and strengthening sequestration monitoring to achieve truly low-carbon oil production. However, in highly heterogeneous formations, achieving uniform advancement of the oil–gas interface using CO₂ is challenging. Early gas breakthroughs along high-permeability layers can lead to channeling, significantly reducing the sweep volume and diminishing extraction efficiency [17,18,19,20].

Conventional methods for controlling gas channeling in low-permeability, heterogeneous reservoirs include gas–water alternating injection [21], gel injection [22,23,24,25], miscible displacement [26,27,28,29], and foam flooding [30,31]. However, these methods have certain limitations. For example, in ultra-low permeability reservoirs with fractures, water will preferentially flow through the fractures, rendering the WAG (Water-Alternating-Gas) method ineffective. Gels can permanently block high-permeability layers, and their mobility becomes challenging to remove once they enter low-permeability layers, causing reservoir contamination.

CO₂ miscible displacement can partially mitigate the breakthrough and channeling of CO₂ along major pathways and high-permeability layers, effectively activating low-permeability reservoirs, adjusting interlayer conflicts, and improving sweep efficiency [26,27,28,32]. Han et al. proposed through core experiments that miscible displacement aids in dissolving pores, enlarging them, and consequently enhancing displacement efficiency [33]. In a CO₂ flooding test conducted by Willhite in an injection well in Kansas, the injection pressure was maintained above the minimum miscibility pressure for two years, resulting in a significant increase in the recovery rate, with approximately 4600 tons of crude oil displaced by CO₂ [34]. CO₂-foam flooding, on the other hand, involves using foam to plug high-permeability formations [35], aiming to improve sweep efficiency and enhance the extraction efficiency of low-permeability reservoirs [36,37,38,39]. However, existing research on CO₂ miscible displacement and CO₂-foam flooding is limited to the reservoir or mesoscale, and the mechanisms in heterogeneous porous media are not well understood. Additionally, foam flooding involves multiple factors, making it challenging for traditional analysis methods to clearly determine the effects of various parameter variables.

To address the issues of unclear micro mechanisms in CO₂ flooding and the limitations of traditional pressure-resistant chamber experiments, this study proposes and assembles a non-confined, microvisualization system suitable for studying CO₂ gas flooding and CO₂-foam flooding. Utilizing a novel high-pressure microfluidic chip, this system simulates the pore environment in heterogeneous reservoirs, enabling stable, accurate, and reproducible visualization of CO₂ flooding at the micro-scale. This enables the exploration of the micro-scale mechanisms of CO₂ gas flooding and CO₂-foam flooding. Additionally, machine learning models are employed to analyze the weightage and impact of various parameters on oil recovery. This research provides a foundation for enhancing recovery rates through CO₂ flooding in heterogeneous reservoirs.

2. Experimental Materials and Methods

2.1. Materials

Normal decane (Macklin (Shanghai, China), 99%, viscosity 0.92 mPa·s); Sudan Red (Sigma-Aldrich, St. Louis, MO, USA) used for staining n-decane for observation; deionized water; CO₂ (Shandong Hongda Biological Co. (Linyi, China), 99.999%); sodium α-olefin sulfonate (AOS) used for the generation of carbon dioxide foam.

2.2. Equipment

The heterogeneous microfluidic chip (Dolomite (Rochester, UK), 3200284) is the main device for the experiment, which can withstand 10 MPa at 21 °C without confining pressure. Figure 1 displays a 3 × 5 grid array of square channels whithin the microfluidic chip. The specific chip parameters are shown in

Table 1. Summary of the chip parameters.

Parameters	Characteristic Values
Channel cross-sectional area	100 µm × 110 μm (near-circular cross-section)
Pore throat diameter	85 µm, 63 µm
Porous volume	38 µL
Porosity	0.63
Wettability	Oil wet
Pressure resistance	10 MPa

. The computer controls the flow rate and pressure of the VP-Series high-pressure injection pump (Vindum-Engineering, San Ramon, CA, USA) to obtain different displacement conditions. The setup included a micro-sampler (GaoGe (Shanghai, China), 1 mL), a micro-injection pump (LongerPump (Baoding, China), LSP02-2A), a six-way valve for switching the injection fluid, a high-sensitivity back pressure valve (with an accuracy of 0.2 MPa) to control the displacement pressure difference, an electronic microscope, and a CCD camera to record and store experimental images in real time for further analysis.

2.3. Water Flooding and CO₂ Flooding Control Experiments

We performed microfluidic water flooding and gas flooding experiments as control groups. The experimental setup is illustrated in Figure 2. The experimental procedure consisted of the following steps:

(1) Chip preparation—The glass microfluidic chip was cleaned and saturated with ethanol at a flow rate of 5 mL/min before being assembled into the system.

(2) Oil saturation—n-Decane (99% purity, viscosity 0.92 mPa·s at 25 °C) stained with Oil Red O was injected at 1 mL/min until complete saturation was achieved (no visible bubbles in the field of view).

The selection of n-decane was based on its low miscibility pressure (<10 MPa), making it suitable for the chip’s pressure limitations; viscosity comparable to light crude oil (0.92 mPa·s); and excellent phase contrast with both aqueous and gaseous phases for visualization.

(3) Imaging setup—Continuous image acquisition was initiated with a CCD camera at 1 s intervals until the displacement process reached the steady state.

(4) Flooding process—For water flooding, deionized water was injected at constant rates of 0.1, 0.3, 0.5, 0.7, and 1.0 PV/min (0 MPa back pressure). For CO₂ flooding, high-purity CO₂ (99.999%) was injected at 1 PV/min with back pressures ranging from 2 to 9 MPa. The microfluidic chip had a pore volume (PV) of 40 μL. Residual oil saturation was quantified using our custom Python 3.10 image analysis program (see Section 2.5 for details).

Figure 2. Schematic representation of the experimental setup. (1) High-pressure pump; (2) computer; (3) six-way valve; (4) CO₂ gas cylinder; (5) middleware container (CO₂); (6) middleware container (n-decane); (7) middleware container (alcohol); (8) micro-injection pump; (9) micro-sampler; (10) CCD camera; (11) microscope; (12) porous media chip; (13) back pressure valve; (14) waste bottle.

Figure 2. Schematic representation of the experimental setup. (1) High-pressure pump; (2) computer; (3) six-way valve; (4) CO₂ gas cylinder; (5) middleware container (CO₂); (6) middleware container (n-decane); (7) middleware container (alcohol); (8) micro-injection pump; (9) micro-sampler; (10) CCD camera; (11) microscope; (12) porous media chip; (13) back pressure valve; (14) waste bottle.

2.4. CO₂-Foam Flooding Oil Recovery Experiment

Using the heterogeneous microfluidic displacement model, we studied the influences of various injection factors such as the surfactant (alpha-olefin sulfonate) concentration, gas–liquid ratio, and injection time on the flow behavior of CO₂-foam flooding, as well as the formation and distribution mechanism of residual oil. The experimental procedure is as follows:

(1)

Clean the glass chip and assemble it into the microfluidic system, saturating it with ethanol at a rate of 5 mL/min.

(2)

Saturate the stained n-decane at a rate of 1 mL/min until no bubbles are observed in the field of view.

(3)

Turn on the camera for continuous shooting until the displacement process reaches a steady state, with a shooting interval of 1 s.

(4)

The micro-injection pump drives the micro-sampler and the high-pressure injection pump to simultaneously inject surfactant and CO₂ into the microfluidic chip to generate foam, and foam displacement continues until it is stable. Record the final distribution of n-decane.

2.5. Calculation of Residual Oil Saturation

A custom Python 3.10 program has been developed to accurately identify the various phase fluids in the chip images. It can automatically calculate and quantitatively characterize the residual oil saturation of CO₂ displacement, providing an efficient processing method for transforming microfluidic image information into quantifiable numerical information. The processing flow is shown in Figure 3.

2.6. Machine Learning Analysis of Parameters

There are many variables and influencing factors in the process of CO₂-foam flooding. We select different algorithms to explore the impacts of these variables on the CO₂-foam flooding process. Through data partitioning, data processing, and model evaluation, we establish a machine learning model to predict the final oil displacement effect of CO₂-foam flooding, analyze the specific impacts of various factors on the final oil displacement effect, and discuss their impact mechanisms.

Different machine learning analysis methods require data selection and preprocessing. In this stage, we conducted the following steps:

(1) Data selection—We have selected 209 data points from the microfluidic CO₂-foam flooding experiment. The selected data exhibit good representativeness and stability across different dimensions, including four variables that are significant and important in the oil displacement process.

(2) Normalization—We used Min-Max normalization, as shown in Formula (1).

$$x_{normalization}={\displaystyle \frac{x_i-x_{min}}{x_{max}-x_{min}}}$$

(1)

In the formula, x_i is the ith value of feature X; x_max is the maximum value in the feature X; x_min is the minimum value in the feature X; and x_{normalization} is the normalized value of the feature x_i.

The initial data are normalized to values between 0 and 1, preventing features with large magnitudes from disproportionately influencing other features.

3. Results and Discussion

3.1. Analysis of the Influence of Water Flooding Parameters

The water flooding and gas flooding experiments, as control experiments for foam flooding, have fewer influencing parameters, and so it is very necessary for us to carry out a traditional chart analysis.

Figure 4 illustrates the curves of residual oil saturation over time at five different flow rates. At an oil displacement time of 100 s, the water displacement experiments at all different flow rates reached a steady state, where the area of residual oil fluctuated within a minimal range. As the flow rate of water displacement increased, the efficiency of water displacement was observed to rise as well. Specifically, at a flow rate of 1.0 PV/min, the chip reached a steady state in just 9 s, maintaining the residual oil saturation around 0.37; conversely, it took 91 s to reach a steady state at a flow rate of 0.1 PV/min, with its residual oil saturation reaching a high of 0.60.

After the water flooding control experiment, there is a relatively obvious residual oil distribution phenomenon in the chip area. According to the distribution position and residual form of the remaining oil in the chip, the residual form of n-decane in the chip channel can be divided into the following four types, as shown in Figure 5:

(1)

Residual oil present in a sheet form (green rectangular part);

(2)

Residual oil present in a strip form in the channel (orange rectangular part);

(3)

Residual oil present in a block form at the pore throat (blue rectangular part);

(4)

Residual oil adsorbed in the form of oil droplets at the wall of the chip pores (yellow rectangular part).

Figure 5. Water flooding residual oil morphology analysis. Red represents oil, while gray indicates water. The water flooding operation was carried out in a unidirectional flow pattern from the injection well (left) to the production well (right).

Figure 5. Water flooding residual oil morphology analysis. Red represents oil, while gray indicates water. The water flooding operation was carried out in a unidirectional flow pattern from the injection well (left) to the production well (right).

Due to chip’s heterogeneous structure, the remaining oil primarily exists in sheet-like formations, with block-like accumulations at pore throats as a secondary form. In contrast, strip-shaped residual oil in non-diameter-changing channels and wall-adsorbed oil are less prevalent and occupy smaller areas.

After water flooding, distinct residual oil distribution patterns were clearly observed within the microfluidic chip area. Based on the spatial distribution and morphological characteristics, the residual n-decane in the microchannels can be classified into four types, as shown in Figure 6:

(1)

Sheet-like residual oil (green circled areas) that forms continuous oil sheets and is distributed across interconnected pore networks.

(2)

Strip-shaped residual oil (orange circled areas) that appears as elongated oil threads and is aligned along the main flow channels.

(3)

Block-shaped residual oil (blue circled areas) that presents as discrete oil ganglia and is trapped at pore throat constrictions.

(4)

Wall-adhered droplet residual oil (yellow circled areas) that exists as attached oil droplets and is adsorbed onto pore wall surfaces.

Figure 6. Residual oil distribution at different flow rates. Red represents oil, while gray indicates water. (A) Saturated oil; (B) 20 μL/min water flooding; (C) 40 μL/min water flooding.

Figure 6. Residual oil distribution at different flow rates. Red represents oil, while gray indicates water. (A) Saturated oil; (B) 20 μL/min water flooding; (C) 40 μL/min water flooding.

Upon reaching a steady state in the water displacement experiments at 20 μL/min (Figure 6B) and 40 μL/min (Figure 6C), a significant reduction in residual oil type 1 was observed, while the other three types showed no significant changes. The outer edges of the sheet-like residual oil primarily consist of throats. These throats have a smaller diameter compared to channels, resulting in an increase in capillary forces. Due to the oil-wet nature of the chip, capillary forces act as resistance. Displacement energy is derived from the pressure difference generated by the flow rate; the higher the flow rate, the greater the oil displacement efficiency.

3.2. Analysis of the Influence of CO₂ Flooding Parameters

This study also explored the effects of the gas injection pressure on gas displacement efficiency. Figure 7 depicts the curves of residual oil saturation over time at different pressures. Below 6 MPa, as the pressure increased, the final oil saturation was reduced from 0.44 to 0.08. This was attributed to the two-phase flow in the chip being primarily influenced by capillary force, viscous drag, and displacement pressure. Here, viscous drag and capillary force, due to the chip wall’s affinity for oil, served as resistance. Meanwhile, the displacement pressure acted as the driving force. As the displacement pressure increased, the ratio of force to resistance in the chip channel also rose, and CO₂, acting as the non-wetting phase, accumulated enough force to invade the channel, affecting areas it could not access at lower ratios of force to resistance.

The non-monotonic behavior of residual oil saturation (ROS) at lower pressures (2–4 MPa) can be explained by the competing effects of capillary forces and displacement dynamics in heterogeneous porous media. At these pressures, the system operates under immiscible conditions, where capillary resistance (favored by the oil-wet chip) dominates the initial oil trapping. As the pressure increases, the displacement force grows, enabling CO₂ to invade smaller pore throats and mobilize trapped oil, leading to an initial ROS reduction. However, the onset of “Haines jumps”—sudden fluid rearrangements caused by localized pressure buildup—triggers intermittent oil mobilization and re-trapping. These jumps fragment the oil phase, but incomplete drainage allows upstream oil to replenish displaced volumes, temporarily increasing ROS. This dynamic is reflected in the rising standard deviations of ROS oscillations (

Table 2. Standard deviation of residual oil saturation per second at different pressures.

Pressure/MPa	Standard Deviation
2	0.016565
3	0.018305
4	0.032462
5	0.059644
6	0.106082

). In contrast, at pressures ≥ 6 MPa, near-miscibility eliminates interfacial tension, stabilizing flow and suppressing capillary-driven instabilities. The resulting viscous-dominated displacement ensures a monotonic ROS decline to minimal levels (~2%).

Interestingly, when the displacement pressure was below 6 MPa, there were oscillations in the remaining oil saturation in the observation area of the chip. Additionally, the intensity of these oscillations increased with the rise in pressure.

As shown in Table 2, the standard deviation of the residual oil saturation curve at 2 MPa and 3 MPa is only 0.017 and 0.018, respectively. As the pressure increases, this value rises to 0.106 at 6 MPa. The oscillation in the residual oil saturation curve may be due to “Haines jumps”. That is, during the displacement process, as the displacement pressure difference increases, the oil phase “enclosed” at the pore throat deforms, the curvature increases, and the capillary resistance increases. When the displacement pressure difference continues to increase, when the oil phase at the pore throat reaches the deformation limit and a sudden fluid breakthrough occurs. During high-pressure displacement, the local pressure accumulation at the pore throat is faster than during low-pressure displacement, the speed at which the oil phase at the pore throat reaches the deformation limit is faster, and fluid breakthroughs occur more frequently. The crescent-shaped liquid surface formed by the oil phase at the throat suddenly flows, pushing into the adjacent pores, the capillary resistance drops rapidly, a portion of the oil phase breaks through and flows out of the observation area rapidly, and the upstream oil phase on the open boundary follows up to replenish. The faster the pressure breaks through and replenishes, the higher the frequency of oscillation.

When the pressure reaches values above 6 MPa, the increase in injection pressure causes the CO₂ and n-decane to reach miscibility, reducing or even completely eliminating the interfacial tension between them. This greatly enhances the fluid mobility, effectively improving the sweep efficiency and bringing the recovery rate close to 100%.

As illustrated in Figure 8, in the CO₂ immiscible flooding process, the residual oil mainly consists of type 3 and type 4 residual oil. For oil-wet glass pores, the adhesion force of the pore walls to the residual oil is strong, which easily allows residual oil to adhere to the walls as an oil film. As the displacement pressure difference increases, the residual oil phase becomes more fragmented, and the residual oil in the pore throats becomes smaller. When the pressure continues to increase to reach the miscible phase, the oil phase in the core mixes with the CO₂, with a recovery rate close to 100%. Under these conditions, the residual oil becomes difficult to observe.

In the CO₂ immiscible flooding process, the residual oil mainly consists of type 3 and type 4 residual oil. For oil-wet glass pores, the adhesion force of the pore walls to the residual oil is strong, which easily allows residual oil to adhere to the walls as an oil film. As the displacement pressure difference increases, the residual oil phase becomes more fragmented, and the amount of residual oil in the pore throats becomes less. When the pressure continues to increase to reach the miscible phase, the oil phase in the core mixes with the CO₂, with a recovery rate close to 100%. Under these conditions, the residual oil becomes difficult to observe.

3.3. Analysis of the Influence of CO₂-Foam Flooding Parameters

Compared to the water flooding and CO₂ immiscible gas flooding displacement images, the foam flooding displacement front is smooth and covers a larger area. This is due to the high viscosity of CO₂-foam, which to some extent inhibits the finger phenomenon and delays the breakthrough time. In addition, during the CO₂ immiscible gas flooding and water flooding processes at lower pressures, residual oil exists in patches in the chip to varying degrees due to finger formation and other reasons. Foam flooding can expand into the affected area, acting as a profile control at a microscopic level. At the same time, the foam cuts off some of the oil phase so that there is rarely any patchy residual oil, and it further promotes the flow of oil phase in the low-permeability area downstream.

The residual oil on the wall surface is less noticeable during CO₂-foam flooding. The surfactant reduces oil–water interfacial tension and simultaneously changes the interfacial tension between the oil and the core wall surface. Consequently, the original wall residual oil transforms into mobilizable oil droplets or emulsified small oil beads, thereby enhancing the microscopic displacement efficiency.

In the uppermost region of the observation area, there is a portion of residual oil in dead-end spaces. In water flooding experiments and non-miscible CO₂ gas flooding experiments, the displacement of dead-end residual oil is challenging due to issues such as large viscosity differences, the formation of preferential flow paths around the area, and early breakthroughs of the displacing phase. During the foam flooding process, the foam has a sealing effect on the core’s pores, which increases the local displacement pressure. Consequently, the dead-end area experiences a greater shearing force, resulting in a reduction of residual oil in the dead-end spaces.

Due to the large number of variable parameters involved in foam flooding, it is difficult to analyze with traditional charts. In order to clarify the rule of each parameter variable, we used multivariate linear regression, support vector regression, and decision tree regression, three machine learning analysis methods, to analyze and model the experimental results and to analyze the oil displacement mechanism of CO₂-foam flooding.

In this paper, four evaluation metrics are used for the regression algorithm: mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²).

(1)

Multiple Linear Regression

By importing the data presented in

Table 3. Table of evaluation parameters of multiple linear regression model.

Evaluation Metrics	MAE	MSE	RMSE	R2
Training set score	0.16	0.04	0.2	0.33
Test set score	0.16	0.04	0.19	0.29

into the linear regression model, we obtained the following linear regression model:

$$Y=0.47-0.63X_1+0.06X_2-0.27X_3$$

(2)

The linear regression model exhibits considerable error, with an MAE of 0.16 and an R² of just 0.29. This suggests substantial inaccuracies when using a linear regression model for data analysis and a relatively weak linear relationship between the independent and dependent variables.

(2)

Support Vector Regression

After using linear regression to establish a model, it is evident that there is a poor linear relationship between variables. However, the overuse of high-order terms in non-linear regression may lead to overfitting. A better choice would be Support Vector Regression (SVR), opting for a polynomial kernel function, which can model more complex planes, thus effectively solving the problems inherent in polynomial regression with high-order terms. As can be seen in

Table 4. Support Vector Regression model evaluation parameters.

Evaluation Metrics	MAE	MSE	RMSE	R2
Training set score	0.11	0.02	0.14	0.68
Test set score	0.10	0.02	0.13	0.71

, the SVR model has smaller errors, with an MAE of only 0.10, and R² value further improved to 0.71. This demonstrates better performance in predicting the effects of foam flooding.

(3)

Decision Tree Regression

Importing the data into the decision tree, the decision tree regression model obtained is shown in Figure 9. The decision tree selects feature split points by minimizing the mean square error (MSE), progressively partitioning the data through multiple node levels (such as PV, concentration, and other features), with the leaf nodes ultimately outputting the predicted values.

The decision tree model also provides weights for each feature value. Among them, the weight of PV has further increased to 0.81, the weight of GWR has dropped to 0.02, and the weight of concentration has dropped to 0.17. At the same time, the coefficient of determination of the decision tree model has also improved, rising from 0.59 to 0.69.

The evaluation metrics for different regression models are presented in

Table 5. Regression model evaluation parameters.

Evaluation Metrics	MAE	MSE	RMSE	R2
Linear regression training set	0.16	0.04	0.2	0.33
Linear regression test set	0.16	0.04	0.19	0.29
SVR training set	0.11	0.02	0.14	0.68
SVR test set	0.10	0.02	0.13	0.71
Decision tree training set	0.10	0.02	0.12	0.74
Decision tree test set	0.10	0.02	0.13	0.69

.

Among the three regression models, both Support Vector Regression (SVR) and decision tree regression have low error parameters, indicating similar model quality. However, SVR exhibits a better fit between the test set and the training set. Therefore, SVR will be used for the next step of the predictive analysis.

According to the SVR model, a grid search was conducted on three feature parameters of CO₂-foam flooding. The predicted residual oil saturation patterns under different parameter combinations are analyzed in Figure 10A–F.

As shown in Figure 10A, when GWR is fixed at 0.8 and PV is varied, residual oil saturation decreases with increasing surfactant concentration (0.15–0.21), corresponding to PV values of 0.40–0.25. Figure 10B further demonstrates the inverse relationship between concentration and PV under this GWR condition.

Figure 10C reveals that at a fixed low concentration (0.06), optimal displacement occurs at GWR = 0.1 and PV ≈ 0.50, with diminishing returns when PV exceeds 0.5. Figure 10D confirms similar displacement efficiency at GWR = 0.5 and 0.7, validating the model’s robustness.

Figure 10E,F analyze low PV (0.15) scenarios: a higher GWR reduces displacement efficiency but yields lower initial residual oil saturation (Figure 10E), and a smaller GWR enhances early-stage surfactant action but risks foam breakthrough and preferential flow (Figure 10F).

4. Conclusions

Based on the above analysis and discussion, the following main conclusions can be drawn:

(1)

In water flooding experiments, increasing the water flooding flow can effectively reduce the formation of “capillary force residual oil” and significantly improve the oil displacement efficiency. In the gas flooding experiment, as the CO₂ displacement pressure increases, the residual oil saturation in the core significantly decreases. After the displacement pressure is greater than 6MPa, the CO₂ and n-decane in the core reach miscibility, greatly improving the flow capacity and effectively improving the sweep efficiency.

(2)

After water flooding, the residual oil morphology mainly consists of residual oil in sheets and residual oil at the pore throat, accompanied by a certain amount of wall residual oil. For non-miscible CO₂ flooding, the residual oil morphology mainly consists of more dispersed residual oil at the pore throat, and the higher the pressure, the more dispersed the residual oil phase. For miscible CO₂ flooding, the oil phase in the core is miscible with CO₂, the residual oil saturation is only 2%, and the residual oil is difficult to observe. For CO₂-foam flooding, there is a small amount of columnar residual oil and residual oil in the pore throat, and the residual oil in the blind end gradually decreases due to the plugging effect of the foam during the displacement process.

(3)

According to the prediction of the SVR model, when the gas–water ratio (GWR) = 0.8, the optimal solution range for the residual oil saturation is that when the surfactant’s normalized concentration (concentration of alpha-olefin Sulfonate, AOS) = 0.15~0.21, the corresponding pore volume (PV) = 0.40~0.25, both of which are negatively related. The efficiency of reducing residual oil saturation improves as the AOS concentration increases and decreases as the injected PV increases. When the AOS concentration = 0.06, the optimal conditions are GWR = 0.1 and PV ≈ 0.50, and the oil displacement effects at GWR = 0.5 and GWR = 0.7 are similar. When PV = 0.15, the larger the GWR, the lower the displacement efficiency, but the lower the initial residual oil saturation.

(4)

From an economic standpoint, CO₂-EOR demonstrates distinct cost advantages over conventional thermal recovery methods (e.g., steam flooding). The operational costs of CO₂-EOR primarily stem from gas compression and recycling systems, yet it eliminates substantial thermal energy consumption. Moreover, CO₂ can be sourced as industrial byproducts (e.g., captured from coal-fired power plants), further reducing raw material costs. Additionally, CO₂-EOR offers carbon sequestration benefits that may generate supplementary revenue through carbon trading mechanisms. In contrast, thermal methods require continuous fuel input for heating, resulting in significantly higher energy expenditures, which are particularly uneconomical for deep or extra-heavy oil reservoirs.

However, this study still has certain limitations, as outlined below.

(1)

The microfluidic experiments in this study revealed the mechanisms of CO₂ flooding at the microscale, but discrepancies exist with actual reservoir conditions (e.g., scale, pressure, fluid complexity, etc.). Future work could validate and optimize the results through multi-scale experiments (e.g., combined with core flooding), expanding parameter ranges (pressure/temperature/fluids), using real reservoir materials, and integrating machine learning with field data to enhance the practical applicability of the findings.

(2)

The high-pressure microfluidic visualization system has several key limitations. It cannot fully replicate the complex 3D heterogeneity of real reservoirs; its 0.1–10 MPa pressure range restricts studies under extreme conditions; the absence of confining pressure fails to reflect formation stress effects; the use of model fluids (e.g., n-decane) overlooks complex crude oil compositions; fixed oil-wet surfaces cannot characterize reservoir wettability variations; and small chip sizes limit observations of long-term dynamic processes. These constraints necessitate further validation through core flooding experiments and field data.