Improving Kinetic Prediction and Structural-Electronic Mechanistic Coherence in the Fenton Process via a Cross-Scale Machine-Learning Framework.

Accurately predicting and understanding contaminant degradation kinetics in advanced oxidation processes remains challenging due to the fragmented and even contradictory structure- and electronic-driven mechanistic interpretations, which traditional linear and isolated experimental methods fail to integrate across these multiscale drivers. This study developed a unified multiscale machine-learning framework that fused molecular fingerprints (MFs) and quantum chemical features (QCFs) of contaminants to predict the kobs of contaminant degradation in the Fenton process, aiming to link structural and electronic drivers. Compared with single-scale models, the fusion model achieved a higher predictive performance and robustness across all evaluated algorithms. More importantly, feature fusion reshaped the feature importance landscape, yielding a coherent structural-electronic mechanistic interpretation in which electronic reactivity is expressed in a structure-dependent manner rather than acting as an isolated molecular property. Interaction analyses (e.g., UMAP and t-SNE) further reveal the complementarity between MFs and QCFs by visualizing their distribution patterns. Partial dependence analyses reveal the synergistic interactions between key environmental factors (e.g., pH) and contaminant properties, defining optimal kinetic regimes for degradation kinetics. External validation experiments confirm the generalizability of the fusion model to unseen contaminants and varying reaction conditions. Overall, this proposed framework reconciles mechanistic understanding and provides a basis for rational optimization of the Fenton process.