Physics–informed machine learning frameworks for predicting fines migration–induced permeability decline: Implications for enhanced oil recovery

Fines migration during low-salinity water injection causes severe permeability impairment and injectivity loss, yet accurately predicting this damage remains a key bottleneck for enhanced oil recovery (EOR) optimization and underground CO₂ storage because hydrodynamic, electrokinetic, and petrophysical processes are tightly coupled. This study presents a data-driven modeling framework for forecasting fines-induced permeability reduction using two physically interpretable approaches: the Group Method of Data Handling and a hybrid Response Surface Methodology model. We compiled a literature-derived dataset of 582 coreflood observations from seven coreflood experiments reported across five published studies. The dataset includes salinity, injected pore volume, clay content, porosity, interstitial velocity, and normalized permeability as the principal variables used for model development. To incorporate electrokinetic effects, we further enhanced the hybrid-RSM model by using salinity-dependent clay and sand zeta potentials derived from published correlations, enabling representation of the electrical double-layer interactions that govern fines detachment. We benchmarked model performance against simplified mechanistic models using independent test data. The hybrid RSM model achieved high predictive accuracy (R² = 0.9838, AAPE = 7.89%), closely matched by the self-organizing GMDH network (R² = 0.9868, AAPE = 8.85%), with both significantly outperforming classical empirical models. Validation using residual diagnostics, bootstrap uncertainty quantification, and nonparametric hypothesis testing confirmed the model's robustness. Furthermore, trend analysis successfully reproduced emergent permeability-reduction behavior near the critical salt concentration (0.1 M NaCl), identifying interstitial velocity, injected pore volume, and clay content as the dominant controlling factors. By pairing high predictive precision with mechanistic transparency, this framework offers a robust, computationally efficient tool for screening LSWF strategies and managing injectivity in subsurface energy operations.