Abolfazl Dehghan Monfarad

Relative permeability, as one of the key parameters in multiphase flow analysis within hydrocarbon reservoirs, plays a crucial role in simulating and optimizing injection and production processes, and it holds particular importance in the context of carbon dioxide storage. Carbon Capture and Storage (CCS), as a modern approach to reducing greenhouse gas emissions, requires a precise understanding of CO₂ flow behavior in porous media. In this study, the relative permeability of the CO₂–water two-phase system was investigated and predicted using machine learning–based modeling approaches. Reservoir data with physical and chemical properties were extracted, screened, and divided into training and testing sets. Three algorithms, namely KNN, Random Forest, and CatBoost, were applied for modeling and optimized using K-Fold cross-validation. The results indicated that CatBoost provided the best performance in predicting relative permeability, with lower error values in both training and testing datasets compared to the other two models. Feature importance analysis further revealed that CO₂ saturation, porosity, and pressure and temperature exert the greatest influence on the model outcomes. The findings of this research highlight that employing advanced machine learning algorithms, particularly CatBoost, can substantially improve the accuracy of relative permeability predictions and contribute to more reliable simulations of underground carbon dioxide storage processes.