Carbonate reservoirs have become strategic targets for reserve expansion in China and the Middle East, driven by the dual goals of reducing carbon emissions and ensuring energy security. However, their significant heterogeneity, complex pore structures, and wettability changes present considerable challenges to the efficiency of CO₂-based enhanced oil recovery (EOR). At the pore level, the interaction of capillary forces, viscous forces, and the evolution of multiphase interfaces causes unstable displacement fronts and severely limits sweep efficiency in low-permeability areas.

To tackle these issues, this study aims to reveal the pore-scale multiphase seepage mechanisms of brine–CO₂ displacement in carbonate-coated heterogeneous porous media. This provides a microscopic foundation for optimizing CO₂ flooding parameters and enhancing sweep performance in actual carbonate reservoirs. A heterogeneous pore network was constructed using a microfluidic chip, and calcium carbonate was coated in situ to simulate authentic carbonate reservoir surfaces and wettability. A series of visualization experiments were conducted at a controlled temperature (40 ℃). CO₂ foam flooding and brine flooding at different injection rates were compared. A CCD imaging system was used to capture pore-scale evolution of oil, water, and gas phases, and gas saturation and residual oil distributions were quantified through image processing. To improve the accuracy of residual oil characterization, the ResNet152 deep neural network was trained on 2885 labeled microfluidic sub-images from CO₂ flooding, CO₂–water alternating flooding, and brine flooding. Using weighted cross-entropy loss, AdamW optimization, and learning rate scheduling, the model achieved high classification accuracy for dispersed, mixed, and heterogeneous residual oil.

Results showed that flooding performance was strongly affected by injection rate and pore-structure heterogeneity. At moderate flow rates (0.5–3 μL·min^–1), CO₂ foam greatly improved sweep efficiency, nearly eliminating residual oil saturation. Foam viscosity and the Jamin effect effectively suppressed viscous fingering and prevented preferential flow through high-permeability channels, forcing the displacing phase into low-permeability areas. Conversely, at very low injection rates (0.1 μL·min^–1), foam instability caused large dispersed gas bubbles, limiting gas saturation to 25%, and hindered oil droplet mobilization, resulting in a high residual oil saturation of 42%. Gas saturation displayed a parabolic relationship with flow rate, with the maximum (93%) at 1 μL·min^–1, where bubble size was smallest, and foam stability was optimal. Deep-learning-based oil classification also showed that brine flooding and CO₂–water alternating flooding primarily produced dispersed residual oil, whereas surfactant-assisted CO₂ flooding created a mixture of dispersed (49%), mixed (36%), and heterogeneous (14%) oil, reflecting foam instability and uneven sweep in highly heterogeneous zones. The model achieved a validation accuracy of 93%, confirming its effectiveness in pore-scale residual oil identification.

This study clarifies the mechanisms underlying brine–CO₂ displacement in carbonate-coated heterogeneous media. Calcium carbonate coating increases hydrophobicity, delays breakthrough in high-permeability pathways, and significantly enhances sweep in low-permeability zones, reducing residual oil by up to 28%. CO₂ foam flooding is highly sensitive to injection rate, with moderate flow rates producing stable foam, high gas saturation, and efficient oil mobilization, whereas very low or high rates reduce displacement stability. By combining microfluidic visualization and deep-learning image analysis, this research offers microscopic insights for optimizing CO₂ flooding conditions and provides technical guidance for deploying CO₂-based EOR in Middle Eastern carbonate reservoirs. The findings also support international cooperation under the Belt and Road Initiative and contribute to global efforts in the low-carbon, efficient development of carbonate oilfields.

Traditional experimental methods cannot facilitate the direct observation of the migration of microplastics within porous media. To address this issue, this study developed a microscopic visualization experimental system to investigate the migration and retention of microplastics and integrated artificial intelligence for the efficient identification and calculation of microplastics. The aim was to quantify the impact of the porous media structure on the migration behavior of microplastics and provide an intuitive, accurate, simple, and scalable experimental system suitable for innovative teaching and research in environmental and related disciplines.

This study developed a microscopic visualization experimental system to deeply investigate the migration and retention behavior of microplastics in porous media by constructing pore-scale single-channel models. In the experiment, five pore-scale single-channel models with different size parameters were developed to simulate the retention of microplastics under different porosity conditions. The experimental setup included a microsyringe, a micro-infusion pump, an optical microscope, and a high-resolution camera. The microsyringe and micro-infusion pump were used to control the injection of fluids, while the optical microscope and high-resolution camera were employed to capture the migration process of microplastics in the channels. Before the experiment, ethanol was used to expel air from the channels, followed by saturation with deionized water. Then, a microplastic suspension treated with an ultrasonic processor was injected into the channels at a rate of 10 µL/h, with images captured at a rate of 2 frames per second. The obtained images were corrected and preprocessed using the ImageJ software to eliminate halos and speckles. Aiming at the small size and large quantity of microplastic particles, this study employed machine learning algorithms for image recognition and counting and developed custom script codes, which were combined with the ImageJ's macro function, to perform the automated batch analysis of data, significantly improving the efficiency and accuracy of microplastic identification, especially with accuracy rates of over 98% for dispersed individual particles and over 95% for particles aggregated in porous media. This method, which combined microscopic visualization technology with artificial intelligence image recognition, provided a novel and efficient experimental method for the study of microplastic environmental behavior.

The experimental results of the microscopic visualization experiment on microplastic transport and retention provided the following conclusions. (1) The results revealed the impact of the porous media structure on the migration behavior of microplastics, with an increase in the media particle size and channel width leading to a significant increase in microplastic retention by 29.8%–56.0% and 14.5%–37.6%, respectively. These findings confirmed the key role of the porous media structure in the retention behavior of microplastics. (2) The experiment visually demonstrated the deposition patterns of microplastics in porous media, which were consistent with existing research findings, further validating the effectiveness of the experimental method. (3) By integrating artificial intelligence image recognition technology, this study developed an efficient method for the identification and counting of microplastics, significantly improving the accuracy and efficiency of data processing. This method not only provided new experimental means for environmental research and microplastic teaching but also showcased the potential application of artificial intelligence technology in the field of environmental science.

This study effectively quantified the impact of the porous media structure on the microplastic retention behavior through a microscopic visualization experimental system and confirmed the consistency of experimental results with existing research. By integrating artificial intelligence image recognition technology, the accuracy of microplastic identification and efficiency of data processing were significantly improved, providing an innovative experimental method for environmental science teaching and research.