Machine Learning with Iterative Error-Driven Refinement for Predicting PFAS-Coated Surface Wettability

Abstract

PFAS coatings impart strong water and oil repellency on materials, but their environmental persistence necessitates sustainable alternatives. The development of alternative chemistries is hindered by limited mechanistic understanding of how molecular structure and surface organization govern intrinsic water and oil wettability. Here, a machine learning-driven molecular dynamics (MD) simulation framework was developed to predict and rationalize the intrinsic water and oil wettability of PFAS coatings on nylon for both water and oil. By systematically sampling PFAS head group types (−CHO, −COOH, −OH, and −SO3H), chain lengths from C1–C7, surface packing densities (1.42–4.65 nm–2), and both linear and branched isomers, a comprehensive database of MD-derived contact angles was constructed. A data-efficient, error-driven refinement strategy enables accurate prediction of water and oil contact angles with mean absolute errors below 3° for linear PFAS and below 5° for branched isomers. Analysis shows that intrinsic water and oil repellency peaks at intermediate packing densities rather than at the highest packing density. Moreover, water repellency depends more on chain length than oil repellency, and head group chemistry strongly modulates these effects. Oil contact angles are typically ∼10–20° lower than water contact angles, consistent with experimental observations. This work provides molecular-level insight into the intrinsic water and oil wettability of PFAS coatings and establishes a data-driven framework that may support the future rational design and screening of environmentally benign, non-PFAS coating chemistries.