Graph Neural\nNetwork Integrating Self-Supervised Pretraining\nfor Precise and Interpretable Prediction of Micropollutant Treatability\nby HO<sup>•</sup>‑Based Advanced Oxidation Processes

Machine learning (ML) has become a crucial tool to accelerate\nresearch\nin advanced oxidation processes via predicting reaction parameters\nto evaluate the treatability of micropollutants (MPs). However, insufficient\ndata sets and an incomplete prediction mechanism remain obstacles\ntoward the precise prediction of MP treatability by a hydroxyl radical\n(HO<sup>•</sup>), especially when <i>k</i> values\napproach the diffusion-controlled limit. Herein, we propose a novel\ngraph neural network (GNN) model integrating self-supervised pretraining\non a large unlabeled data set (∼10 million) to predict the <i>k</i><sub>HO</sub> values on MPs. Our model outperforms the\ncommon-seen and literature-established ML models on both whole data\nsets and diffusion-controlled limit data sets. Benefiting from the\npretraining process, we demonstrate that <i>k</i>-value-related\nchemistry wisdom contained in the pretrained data set is fully exploited,\nand the learned knowledge can be transferred among data sets. In comparison\nwith molecular fingerprints, we identify that molecular graphs (MGs)\ncover more structural information beyond substituents, facilitating\na <i>k</i>-value prediction near the diffusion-controlled\nlimit. In particular, we observe that mechanistic pathways of HO<sup>•</sup>-initiated reactions could be automatically classified\nand mapped out on the penultimate layer of our model. The phenomenon\nshows that the GNN model can be trained to excavate mechanistic knowledge\nby analyzing the kinetic parameters. These findings not only well\ninterpret the robust model performance but also extrapolate the <i>k</i>-value prediction model to mechanistic elucidation, leading\nto better decision making in water treatment.