Refining irreducible water saturation predictions in reservoir rocks using machine learning models

This study introduces a data-driven framework for predicting irreducible water saturation (Swi) in reservoir rocks, addressing the limitations of traditional empirical models and capturing the complexity of heterogeneous formations. Using a comprehensive experimental dataset including porosity, grain density, permeability, and T2lm, eight machine learning algorithms were trained and evaluated. Among them, the Convolutional Neural Network (CNN) demonstrated the highest predictive accuracy, achieving superior R2 and error metrics. SHAP-based sensitivity analysis identified permeability as the dominant feature, reinforcing the model's physical relevance. By integrating core-scale measurements with advanced computational techniques, the proposed methodology offers a scalable and validated solution for Swi estimation, supporting more accurate reservoir characterization and informed hydrocarbon reserve management.