Predictive modeling of permeability in porous media based on structural parameters and machine learning

The study of permeability in porous media is of significant theoretical and engineering importance in geosciences. Accurately determining the permeability poses a critical challenge due to the high error margins in laboratory measurements, the limited generalizability of analytical models like the Kozeny–Carman (K–C) equation, and the difficulty of balancing accuracy with efficiency in numerical simulations. This research employs machine learning techniques to predict permeability based on the structural parameters of porous media. A large dataset of randomly generated porous media was created using the truncated Gaussian random field method, with key structural parameters, such as porosity, specific surface area (SSA), tortuosity, and fractal dimension computed. The study demonstrates that the K–C constant is not fixed but varies with porosity and medium properties, leading to a revised model. The Support Vector Machine model trained on these parameters achieved highly accurate permeability predictions. The SHapley Additive exPlanations analysis ranked the factors influencing permeability, identifying porosity as the most significant, followed by SSA, fractal dimension, and tortuosity. Additionally, the mechanisms and characterization methods by which pore complexity affects permeability were investigated from the perspectives of fluid–solid interface, flow path, and pore distribution. These findings offer a more accurate method for predicting permeability and contribute to a deeper understanding of fluid flow in porous media.