Bridging Microscopic Dynamics and Macroscopic Fate: A Molecular Dynamics-Machine Learning Approach for Predicting...

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Environmental Science

Background:

• Per- and polyfluoroalkyl substances (PFAS) are persistent global contaminants.
• Predicting PFAS environmental fate and mobility is challenging.
• Existing machine learning models for PFAS mobility often ignore water chemistry effects.

Purpose of the Study:

• To develop a novel multiscale approach, the Phys-ML Sorp Framework, integrating molecular dynamics (MD) simulations with machine learning (ML).
• To enhance the prediction accuracy of the solid-liquid distribution coefficient (logKd) for PFAS.
• To incorporate physically informed microscopic features into ML models for improved environmental risk assessment.

Main Methods:

• Quantified microscopic features from MD simulations: radius of gyration (Rg), solvent accessible surface area (SASA).
• Developed a novel effective activity coefficient (logγ) using MD-derived Rg and an extended Debye-Hückel equation.
• Integrated these features into an ML model trained on 499 PFAS partitioning observations in pure water and CaCl2 systems.

Main Results:

• The Phys-ML Sorp Framework achieved superior predictive performance (RPD = 2.90, RMSE = 0.32).
• Incorporating MD-derived microscopic features improved RPD by 14.62% and reduced RMSE by 13.52% compared to models using only macroscopic parameters.
• SHAP analysis identified molecular weight, SASA, logKow, Rg, and logγ as key factors influencing PFAS partitioning.

Conclusions:

• The Phys-ML Sorp Framework provides a robust, mechanistically informed approach for predicting PFAS environmental mobility.
• This multiscale method significantly enhances the accuracy of logKd predictions, addressing limitations of current ML models.
• The framework offers improved capabilities for environmental pollutant modeling and risk assessment.