RIME-RF-RIME: A novel machine learning approach with SHAP analysis for predicting macroscopic permeability of porous media

Abstract

Predicting the macroscopic permeability of porous media is critical in various scientific and engineering applications. This study proposes a novel model that combines Random Forest (RF) and rime-ice (RIME) optimization algorithm, denoted RIME-RF-RIME, to predict permeability based on six key features covering fluid phase dimensions, geometric characteristics, surrounding phase permeability, and media porosity. After the input space simplification process using RIME, the RF model achieves high predictive accuracy with a coefficient of determination (R²) of 0.980. Furthermore, SHapley Additive exPlanations (SHAP) values are employed to decipher these features' importance and interaction effects on the model's predictions. The analysis reveals that porosity, permeability of the porous phase, and the size of the fluid phase perpendicular to the flow direction exert the most significant individual influences. This study not only unveils crucial insights into the underlying mechanisms governing permeability in porous media but also contributes to developing interpretable and reliable predictive models for related applications.