Machine learning enabled kinetic prediction and inverse design of sonoelectrochemical oxidation: Molecular descriptor integration, applicability domain assessment, and experimental verification

Sonoelectrochemical oxidation (US-EC) is a promising method for removing emerging pollutants, but its degradation kinetics are influenced by the interactions between operating conditions, reactor characteristics, and pollutant properties. This study developed an machine learning framework for kinetic prediction, reliability assessment, and inverse design of US-EC oxidation. A database containing 224 experimental and literature observations was constructed. Extreme gradient boosting (XGB), artificial neural networks, and support vector regression models were compared using source hierarchical repeated cross-validation and a fixed internal retainer set. XGB showed the most stable performance in repeated validation. At the log 10 (k) scale, its mean R 2 was 0.909 ± 0.029, and the root mean square error was 0.224 ± 0.037. Characteristic ablation experiments showed that adding physicochemical and quantum chemical descriptors reduced the RMSE by 3.69% compared to using only process variables. Ranking importance analysis and the TreeSHAP method identified applied voltage, electrode material, electrode gap, and log K ow as important predictors of degradation kinetics. Domain of applicability analysis revealed that prediction errors increased for observations with weak support from the training data. Scenario analysis reproduced the general response to changes in electrode gap. However, the model showed limited sensitivity to ultrasonic frequency trends derived from the literature. This finding indicates that numerical coverage of variables alone cannot guarantee reliable predictions when relevant process conditions are not adequately reflected in the training data. Subsequently, the XGB was combined with a domain constrained particle swarm optimization algorithm to determine the near-optimal operating region for caffeine degradation. Five representative conditions were tested experimentally. The measured mean k value was 0.0554 min −1 , the predicted value was 0.0546 min −1 , and the overall mean absolute percentage error was 7.28%. These results demonstrate that combining molecular descriptors, applicability assessment, inverse optimization, and experimental validation can support data-driven prediction and operating condition optimization for US-EC processes.