Structural and Electronic Features-Integrated Machine Learning Framework for High-Throughput Prediction of Organic Pollutant Reactivity.

Understanding and predicting the reactivity of organic pollutants toward reactive species is crucial for designing efficient and targeted degradation strategies for advanced oxidation processes. However, the structural complexity and chemical diversity of pollutants pose challenges for developing interpretable and high-throughput predictive frameworks. Here, a machine-learning-based approach that integrates quantum chemical descriptors and molecular fingerprints descriptors to predict pollutant reactivity toward sulfate radicals is presented. By combining RDKit and conceptual density functional theory (CDFT) descriptors, key structure-activity features, including EHOMO(N), electron-donating capacity, ring structures, branching, and molecular surface areas, are identified, and their quantitative reactivity thresholds are established. Additionally, the quantitative read-across structure-activity relationship model incorporating intermolecular similarity expands the applicability domain (AD) to 74.3% across 12 pollutant classes, a 2.1-fold increase in the AD over quantitative structure-activity relationship (QSAR) approaches. Experimental validations across structurally diverse compounds demonstrate strong predictive performance (R2 = 0.811). This work provides a transparent and high-throughput predictive framework for reactivity prediction of organic pollutants, facilitating the bottom-up design of an advanced oxidation process tailored to specific pollutant profiles.