Abstract
Relative permeability is a critical petrophysical parameter that controls multiphase flow behavior in porous media and significantly impacts reservoir simulation accuracy, recovery forecasting, and EOR planning. Conventional laboratory-based methods for determining relative permeability, while accurate, are time-intensive, costly, and spatially limited. This study focuses on applying machine learning techniques to estimate oil and water relative permeability using core data from a reservoir in southwest Iran. Sixteen core samples were analyzed, and seven input features—absolute permeability, porosity, irreducible water saturation, oil permeability at Swi, viscosities of oil and water, and pressure differential—were used to train four machine learning models: Extra Trees, K-Nearest Neighbors, Categorical Boosting, and Extreme Gradient Boosting. Models were optimized using Bayesian hyperparameter tuning and evaluated using R², RMSE, and MAE. For water relative permeability, the Extra Trees model delivered the best performance, achieving an R² of 0.9974 on the overall dataset, with the lowest RMSE (0.0045) and MAE (0.0007), indicating high accuracy and generalizability. KNN also performed well, especially in the 0.1–0.2 permeability range. In contrast, for oil relative permeability, the KNN model achieved the highest accuracy with an R² of 0.9973 and the lowest error metrics (RMSE = 0.0113, MAE = 0.0024), outperforming the other methods in both training and testing sets. Extra Trees performed poorly for oil permeability, especially in capturing higher permeability ranges. SHAP sensitivity analysis revealed that water saturation is the most influential factor for both oil and water models. For water, oil permeability at Swi also had a major impact; for oil, oil viscosity played the second most critical role. Overall, this study demonstrates that machine learning offers a robust, efficient alternative to laboratory experiments for estimating relative permeability, with m