Résumé
Per-and polyfluoroalkyl substances (PFAS) are highly persistent pollutants whose molecular-scale transport and interactions with water remain challenging to model accurately. We develop a machine learning potential (MLP) trained on ab initio data for a representative PFAS molecule in aqueous solution and evaluate its ability to reproduce structural, vibrational, and dynamical properties. The MLP captures essential features of PFAS-water interactions and yields diffusion coefficients in good agreement with experimental data, while also demonstrating transferability across PFAS of different chain lengths. Importantly, we critically compare the MLP against established non-reactive and reactive (ReaxFF) force fields. While the MLP offers great accuracy in reproducing ab initio energetics and structural correlations, ReaxFF proves more computationally efficient and comparably reliable for diffusion dynamics, highlighting the trade-offs between accuracy and efficiency in molecular simulations of PFAS. As a preliminary extension, we illustrate the applicability of MLP to PFAS adsorption on graphene. Overall, this study paves the way for more advanced molecular dynamics simulations to improve our understanding of PFAS behavior and aid in the development of detection and remediation strategies.