Machine learning-guided identification of E_HOMO-associated pollutant reactivity in piezocatalytic persulfate systems

Piezocatalytic persulfate oxidation shows promise for refractory pollutant degradation, yet predictive kinetics remain elusive because catalyst-centric studies have largely overlooked pollutant molecular properties. To address this gap, a closed-loop framework integrating literature data mining, density functional theory calculations, interpretable machine learning, mechanistic analysis, and experimental validation was developed. A literature derived dataset containing 408 records was constructed using catalyst characteristics, mechanical stimulation parameters, reaction conditions, and pollutant molecular descriptors. Among eight regression algorithms, XGBoost showed the best performance and achieved a mean test R² of 0.848 ± 0.013, an RMSE of 0.312 ± 0.019, and an MAE of 0.226 ± 0.015 for lg k_obs across 20 repeated train/test splits. SHAP analysis identified E_HOMO as an important pollutant descriptor associated with degradation kinetics. Mechanistic analysis in an independently prepared ZnO@CuO/PMS system suggested that tetracycline oxidation involved holes (h⁺), singlet oxygen (¹O₂), and ZnO@CuO-PMS* reactive complex, with interfacial PMS activation facilitated by the Cu²⁺/Cu⁺ redox cycle and piezo-generated charges. Experimental validation using twelve pollutants, including four compounds absent from the model development dataset, showed that the model generally reproduced relative kinetic differences. Experimental k_obs was positively correlated with E_HOMO (R²=0.76, p = 0.0038). This framework provides a basis for predicting pollutant degradation kinetics and further prioritizing relative removal susceptibility under defined piezocatalytic persulfate systems.