An Informatics Approach to Recommender Systems for PFAS Molecule Replacement

Per- and polyfluoroalkyl substances (PFAS) are persistent environmental contaminants used across an enormous range of industrial and consumer applications, and identifying safe, nonfluorinated replacements remains a slow, largely trial-and-error process. In this work, we construct a recommender engine for suggesting PFAS alternative chemistries based on past replacements. We present a data-driven recommender framework that learns molecular substitution relationships from experimentally documented PFAS-to-non-PFAS pairings, and we demonstrate its viability through two case studies. In the first, a recommender approach grounded in known substitution chemistry and validated by first-principles physics is shown to accelerate the identification of over 500 potential nonregrettable PFAS alternatives across diverse application domains. In the second, we consider substitutions with silicone molecules, performing ab initio calculations on 95 unique PFAS molecules that have been replaced by silicones and 12 unique silicone molecules that have served as replacements, and we then use our recommender engine to identify chemistries that are prime candidates for replacement. Together, these results demonstrate that a data-driven recommender approach can meaningfully navigate the vast space of possible PFAS replacements and surface chemically plausible, synthesizable candidates for experimental evaluation.