Bridging Microscopic Dynamics and Macroscopic Fate: A Molecular Dynamics-Machine Learning Approach for Predicting PFAS Solid–Liquid Distribution in Soils and Sediments

Per- and polyfluoroalkyl substances (PFAS) are persistent global contaminants, posing challenges to predicting their environmental fate. The solid–liquid distribution coefficient (logKd) is a key parameter for PFAS mobility, but current machine learning (ML) models often overlook its susceptibility to real-world water chemistry. To address this, we introduce the Phys-ML Sorp Framework, a novel multiscale approach integrating molecular dynamics (MD) simulations with ML to enhance logKd prediction. We quantified physically informed microscopic features from MD simulations, including radius of gyration (Rg), solvent accessible surface area (SASA), and a novel effective activity coefficient (logγ) that uniquely captures solute conformational responses by incorporating MD-derived Rg into an extended Debye–Hückel equation, offering a physically meaningful measure of nonideal solution effects. Leveraging 499 PFAS partitioning observations in pure water and calcium chloride (CaCl2) systems, our model achieved superior predictive performance (RPD = 2.90, RMSE = 0.32). The incorporation of MD-derived microscopic features resulted in a 14.62% improvement in RPD and a 13.52% reduction in RMSE over models relying solely on macroscopic parameters. SHAP analysis revealed molecular weight (MW, 0.32), SASA (0.28), logKow (0.23), Rg (0.08), and logγ (0.07) as dominant factors. This framework not only advances environmental pollutant modeling but also establishes a robust, mechanistically informed approach for enhanced environmental risk assessment.