Multi-resolution transfer learning for rapid prediction of pore-scale multiphase flow

Deciphering multiphase flow patterns at the pore-scale is fundamentally essential for upscaling to determine macro-scale flow parameters. Direct numerical simulations provide detailed insights related to pore-scale flow physics but are computationally expensive. To reduce computational costs, coarser meshes may be used at the expense of accuracy. This study presents a deep learning framework that leverages multi-resolution data for two-dimensional pore-scale two-phase flow at fixed capillary number: low resolution simulations generate large training datasets, while high resolution simulations offer the required supervision to capture the pore-scale flow physics. The model effectively transfers the flow physics learned from low resolution dataset to the high resolution cases, requiring only limited high-fidelity data for adaptation. By combining computational efficiency with predictive accuracy, the proposed framework facilitates rapid and accurate pore-scale flow analysis, addressing a critical need in multiphase flow research.