Predicting the Sorption Capacity of Perfluoroalkyl and Polyfluoroalkyl Substances in Soils: Meta-Analysis and Machine Learning Modeling.

Abstract

Predicting the soil sorption capacity for perfluoroalkyl and polyfluoroalkyl substances (PFAS) is pivotal for environmental risk assessment. However, traditional experimental methods are inefficient, necessitating computational model development. We compiled a comprehensive data set including 44 PFAS and 405 soils from 35 literature reports, conducted a meta-analysis, and constructed robust machine learning models. Machine learning models using LightGBM with RDKit or PaDEL descriptors achieved R2 of 0.89, 0.88, and 0.72, RMSE of 0.28, 0.28, and 0.36, and MAE of 0.18, 0.19, and 0.28 for cross-validation, internal test set, and external test set, respectively. SHapley Additive exPlanation (SHAP) analysis identified PFAS properties as the primary influence on sorption, followed by environmental conditions and soil properties. We found that low SOC (<0.56%) minimally affects PFAS sorption. A pH of 6 is the boundary point where anionic PFAS are mainly attracted or repelled by electrostatic interaction, and higher pH may enhance the PFAS soil sorption through cation bridges. Although van der Waals forces and polar interactions enhance the sorption of PFAS with carbon chains ≥8, the introduction of polar structures containing oxygen, nitrogen, and sulfur into PFAS will lower hydrophobicity and sorption affinity. This study provides accurate predictive models, which are helpful for environmental decision-making.