MiRA: A Multi-Receptive Integration Framework for Predicting Pollutant Transformation Pathways in AOPs─A Case Study of UV and Far-UV Systems

Abstract

Accurate prediction of pollutant transformation pathways in advanced oxidation processes (AOPs) is critical for precise treatment design. However, conventional models inadequately capture global dependencies and key local reaction information, limiting reliable inference of complex multistep pathways. To overcome these limitations, this study introduces the multireceptive integration for reaction-pathway analysis (MiRA) framework. For the first time, it integrates “multi-receptive field chemical semantic capture + learnable fusion decision-making” into AOPdegradation pathway prediction. This enables the collaborative expression of local reaction sites, neighborhood structures, and global conditional dependencies within a single model, yielding information-rich feature representations with superior generalization capabilities. Using ultraviolet (UV) and far-ultraviolet (far-UV) data sets, MiRA framework achieves accuracies of 88.03% and 85.44%, respectively, with BLEU scores of 0.9459 and 0.9325, and Tanimoto similarity scores of 0.9405 and 0.9296, demonstrating robust predictive precision. Attention visualizations reveal that the framework automatically identifies key reaction features of substrates, demonstrating a degree of expert-level chemical reasoning capability. Additionally, an integrated toxicity estimation module provides relative screening of the toxicity potentials of transformation products. Overall, this study provides an intelligent analytical method for predicting pollutant transformation pathways and screening the toxicity potentials of transformation products in AOPs. It offers new methodological support for the precise design of AOPs.