Modeling PFAS Sorption in Soils Using Machine Learning

In this study, we
introduce PFASorptionML, a novel machine learning
(ML) tool developed to predict solid–liquid distribution coefficients
(Kd) for per- and polyfluoroalkyl substances
(PFAS) in soils. Leveraging a data set of 1,274 Kd entries for PFAS in soils and sediments, including compounds
such as trifluoroacetate, cationic, and zwitterionic PFAS, and neutral
fluorotelomer alcohols, the model incorporates PFAS-specific properties
such as molecular weight, hydrophobicity, and pKa, alongside soil characteristics like pH, texture, organic
carbon content, and cation exchange capacity. Sensitivity analysis
reveals that molecular weight, hydrophobicity, and organic carbon
content are the most significant factors influencing sorption behavior,
while charge density and mineral soil fraction have comparatively
minor effects. The model demonstrates high predictive performance,
with RPD values exceeding 3.16 across validation data sets, outperforming
existing tools in accuracy and scope. Notably, PFAS chain length and
functional group variability significantly influence Kd, with longer chain lengths and higher hydrophobicity
positively correlating with Kd. By integrating
location-specific soil repository data, the model enables the generation
of spatial Kd maps for selected PFAS species.
These capabilities are implemented in the online platform PFASorptionML,
providing researchers and practitioners with a valuable resource for
conducting environmental risk assessments of PFAS contamination in
soils.