Machine learning-driven identification and predictive mapping of clogging regimes in porous media

Clogging in porous media critically limits the performance of subsurface and filtration systems, yet conventional models often rely on oversimplified, single-parameter thresholds to predict its behavior. This study develops a unified, machine learning–based framework to identify, characterize, and predict clogging behavior using dimensionless parameters representing pore structure, hydrodynamics, and particle–surface interactions. A total of 2,500 pore-scale realizations, generated with a pre-trained model informed by CFD-DEM simulations, were analyzed using four key metrics: permeability reduction, clogged fraction of throats, clogging zone length, and critical throat size of clogging. Three distinct clogging regimes emerged statistically (namely, surface, deep distributed, and sparse) each with its distinguished features. The framework further introduces high-resolution Phase Diagram and Clogging Diagnostic Maps that link input conditions to spatial clogging patterns and severity. These tools provide a scalable, interpretable foundation for optimizing system performance in managed aquifer recharge, enhanced oil recovery, groundwater remediation, and filtration system design.