Graph Neural Network Integrating Self-Supervised Pretraining for Precise and Interpretable Prediction of Micropollutant Treatability by HO•-Based Advanced Oxidation Processes

Abstract

Machine learning (ML) has become a crucial tool to accelerate research in advanced oxidation processes via predicting reaction parameters to evaluate the treatability of micropollutants (MPs). However, insufficient data sets and an incomplete prediction mechanism remain obstacles toward the precise prediction of MP treatability by a hydroxyl radical (HO^•), especially when k values approach the diffusion-controlled limit. Herein, we propose a novel graph neural network (GNN) model integrating self-supervised pretraining on a large unlabeled data set (∼10 million) to predict the k_HO values on MPs. Our model outperforms the common-seen and literature-established ML models on both whole data sets and diffusion-controlled limit data sets. Benefiting from the pretraining process, we demonstrate that k-value-related chemistry wisdom contained in the pretrained data set is fully exploited, and the learned knowledge can be transferred among data sets. In comparison with molecular fingerprints, we identify that molecular graphs (MGs) cover more structural information beyond substituents, facilitating a k-value prediction near the diffusion-controlled limit. In particular, we observe that mechanistic pathways of HO^•-initiated reactions could be automatically classified and mapped out on the penultimate layer of our model. The phenomenon shows that the GNN model can be trained to excavate mechanistic knowledge by analyzing the kinetic parameters. These findings not only well interpret the robust model performance but also extrapolate the k-value prediction model to mechanistic elucidation, leading to better decision making in water treatment.