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

Abstract

Flash Joule heating (FJH) presents an attractive method to decompose per- and polyfluoroalkyl substances (PFAS) but suffers from an optimization challenge due to its complex reaction dynamics. In this study, we introduce a data-driven workflow that includes a Human-Guided Bayesian Optimization (HGBO) algorithm and an interpretable multibranch neural network (MBNN) to understand and optimize PFAS removal from soil. The HGBO algorithm incorporates expert intuition into the optimization cycle via a probabilistic acquisition strategy to enhance efficiency. In two iterations, HGBO improves the PFAS removal efficiency by 60%, outperforming vanilla BO and human-centered optimization. The results are well interpreted by SHapley additive expansion (SHAP) values and partial dependence analysis (PDA) to quantify feature significance and interactions. An interpretable MBNN is then developed to quantify the contributions of functional groups in various PFAS to the FJH degradation mechanism, which is further validated by density functional theory calculations. Seamless integration of HGBO and interpretable MBNN in one data-driven workflow not only accelerates experimental optimization but also provides interpretability, enabling more informed experimental decisions in complex chemical synthesis with limited data.