Predicting soil-water partition coefficients of PFAS using machine learning: Model development, interpretation, and validation.
Abstract
PER: and polyfluoroalkyl substances (PFAS) are environmentally persistent contaminants, yet experimental determination of their soil-water partition coefficients (Kd) remains costly and time-consuming. In this study, five machine-learning regression models were developed using 2057 literature-derived batch adsorption data points by integrating the average net charge descriptor (Zavg), the composite descriptor alert_prior_score, equilibrium aqueous concentration (Cw), PFAS structural descriptors, and soil physicochemical properties. Among the tested models, extreme gradient boosting (XGBoost) showed the best performance with 10 descriptors, achieving R2 values of 0.83 and 0.86 and ratio of performance to deviation (RPD) values of 2.42 and 2.66 for the Cw < 10 μg/L and Cw ≥ 10 μg/L datasets, respectively. Shapley additive explanations (SHAP) analysis indicated that hydrophobic interactions dominated adsorption at low concentrations (Cw < 10 μg/L), whereas headgroup-related hydrophilicity became more influential at higher concentrations. Independent sorption experiment validation using contaminated site soils showed that prediction deviations for all samples were within one order of magnitude. These results demonstrate that the proposed model provides an efficient and interpretable tool for predicting PFAS soil-water partitioning and supports environmental risk assessment and contaminated-site management.