A Data-Driven Approach for Efficient Prediction of Permeability of Porous Rocks by Combining Multiscale Imaging and Machine Learning

Digital rock physics has increasingly become an emerging field in which advanced numerical simulation and high-resolution imaging are combined to accurately predict rock properties. In this context, multiscale imaging is crucial for fully capturing the inherent heterogeneity of natural rocks. However, limitations in resolution and field of view (FOV) present significant challenges. Direct numerical simulations at large scales are often not computationally practical or may be too expensive. The compromise between FOV and resolution is particularly pronounced in the complex multiscale pore structures of porous rocks, including carbonates in particular. To address this issue, we propose an innovative machine learning technique that integrates multiscale imaging data at varying resolutions. For the rock sample, we used the imaging data published by Nabipour et al. (Adv Water Resour 104695, 2024) in three resolutions. Our approach employs an optimized neural network design combined with a transfer learning strategy, enabling the identification of complex cross-scale correlations that were previously unattainable with conventional methods. The results demonstrate that this multiscale neural network approach effectively estimates permeability by analyzing various aspects of pore morphology across different scales. In particular, we achieved high accuracy, as evidenced by R-squared coefficients of 0.966 for training and 0.836 for testing in low-resolution domains, and also significantly enhanced computational efficiency, reducing the overall computational time. Despite being tested for images of carbonate rocks, the developed method is adaptable to a wide range of multiscale porous materials and offers a promising solution to the persistent challenge of balancing imaging resolution with FOV.