ML Workflows for Screening Degradation-Relevant Properties of Forever Chemicals.

The environmental persistence of per- and polyfluoroalkyl substances (PFAS) necessitates new remediation technologies, yet the vast chemical space makes traditional exploration methods for understanding degradation-relevant properties intractable. Rational design of PFAS degradation strategies requires accurate prediction of three critical molecular properties: bond dissociation energies (BDEs) to govern kinetics, polarizability to control catalytic interactions, and thermodynamic stability to govern reaction feasibility. Guided by theoretically-rooted principles, we identify that global properties (polarizability, stability) require spatially-informed features (3D electron density patterns), while bond-specific properties are governed by topological features (atomic connectivity). We developed two distinct physics-informed ML workflows implementing this principle: For global properties, two-point spatial correlations were compressed via Principal Component Analysis (PCA) and input to a Gaussian Process Regression (GPR) model. For local properties, a graph-based feature scheme was coupled with a Random Forest (RF) algorithm. Both workflows demonstrated strong predictive performance (GPR R2 ≈ 0.92 for polarizability; R2 ≈ 0.97 for enthalpy; RF R2 ≈ 0.87 for BDE) across multiple datasets, establishing robust Structure-Property linkages for PFAS. These physics-informed models provide a foundational capability for rapid, high-throughput screening of the vast PFAS library, enabling prioritization of candidate molecules and bonding motifs for subsequent experimental and process-level remediation studies, rather than constituting complete remediation workflows by themselves.