Predicting PFAS Diffusion Coefficients with Active Learning and Molecular Dynamics.

Area of Science:

• Environmental Chemistry
• Computational Chemistry
• Toxicology

Background:

• Per- and polyfluoroalkyl substances (PFAS) are persistent, toxic synthetic compounds with widespread environmental contamination.
• Predicting PFAS environmental fate is crucial for risk assessment but hindered by a lack of diffusion coefficient data.
• Experimental and traditional computational methods are insufficient for the vast PFAS chemical space.

Purpose of the Study:

• To develop a data-efficient computational framework for predicting PFAS diffusion coefficients.
• To enable accurate environmental transport modeling across a wide range of PFAS.
• To support informed environmental risk assessment and remediation planning.

Main Methods:

• Integrated machine learning (ML) and molecular dynamics (MD) with active learning.
• Utilized chemical graph representations and physicochemical descriptors for ML models.
• Employed uncertainty-based sampling to guide targeted MD simulations and model retraining.

Main Results:

• Achieved significant performance improvement in diffusion coefficient prediction (R² from 0.095 to 0.907, 88% reduction in mean relative error).
• Demonstrated the efficacy of uncertainty-based active learning over random sampling for efficient chemical space exploration.
• Successfully enabled property prediction for thousands of PFAS molecules.

Conclusions:

• The developed framework offers a computationally efficient and accurate method for predicting PFAS diffusion coefficients.
• This approach facilitates large-scale environmental fate assessment and the design of effective remediation strategies.
• Supports proactive management of PFAS contamination risks.