Accelerate Flash Removal of PFAS from Soil by Human-Guided Bayesian Optimization and Interpretable Machine Learning.

Area of Science:

• Environmental Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Per- and polyfluoroalkyl substances (PFAS) are persistent environmental pollutants.
• Flash Joule heating (FJH) shows promise for PFAS decomposition but requires optimization.
• Complex reaction dynamics hinder efficient PFAS removal using FJH.

Purpose of the Study:

• To develop a data-driven workflow for optimizing PFAS removal from soil using FJH.
• To enhance the efficiency and interpretability of PFAS degradation processes.
• To understand the degradation mechanism of PFAS through functional group contributions.

Main Methods:

• Human-Guided Bayesian Optimization (HGBO) algorithm for efficient optimization.
• Interpretable Multibranch Neural Network (MBNN) for mechanistic insights.
• SHapley additive expansion (SHAP) and partial dependence analysis (PDA) for feature interpretation.
• Density functional theory (DFT) calculations for validation.

Main Results:

• HGBO improved PFAS removal efficiency by 60% in two iterations.
• The workflow outperformed standard Bayesian optimization and human-centered approaches.
• SHAP and PDA quantified feature significance and interactions in PFAS degradation.
• MBNN identified contributions of functional groups to FJH degradation, validated by DFT.

Conclusions:

• The integrated HGBO and MBNN workflow accelerates experimental optimization for PFAS removal.
• This data-driven approach provides crucial interpretability for complex chemical synthesis.
• Enables informed experimental decisions in challenging environmental remediation scenarios.