Physics-informed machine learning for cross-study prediction of oil–water membrane fouling

Performance decay driven by coupled transport, accumulation, and removal processes remains a central challenge in many chemical engineering systems. Membrane fouling in oil–water separation is examined here as a representative case. While numerical and empirical models can capture specific cases, their predictive performance across independent studies remains limited. In this work, we combine a reduced-order finite-element model of transport and fouling resistance with a data-driven framework that maps experimental metadata to flux-decline behavior across 112 datasets involving various membranes, oils, and operating conditions. The reduced representation reproduces flux–time trajectories with high fidelity (average R 2 ≈ 0.95). Mapping experimental metadata to curve descriptors enables stage-resolved prediction of normalized flux, yielding mean absolute errors of 0.08–0.14 under grouped cross-validation. External validation using publication-level holdouts shows that differences between studies, rather than model capacity, limit predictive performance, highlighting the need for data standardization. The analysis indicates links between crossflow and pressure and early-stage fouling, while porosity and permeance relate to steady-state flux. Based on these results, we propose a minimal checklist of parameters needed for reproducible fouling models. The results quantify the impact of metadata completeness on model transferability and provide a reproducible framework for cross-study analysis of time-dependent membrane performance.