Hierarchical machine learning-based prediction for ultrasonic degradation of organic pollutants using sonocatalysts

Ultrasound-based advanced oxidation processes (AOPs) are effective for degrading organic pollutants, with hydrogen peroxide (H₂O₂) acting as a key intermediate in radical generation and overall degradation efficiency. However, conventional machine learning models often overlook the influence of reactive oxygen species (ROS), limiting their ability to accurately capture reaction dynamics. This study developed a two-stage hierarchical machine learning model that explicitly incorporates the role of H₂O₂ in predicting degradation efficiency in ultrasound-based AOPs using sonocatalysts. The first stage predicts H₂O₂ concentration generated during ultrasonic reactions. The second stage then estimates degradation efficiency using the predicted H₂O₂ levels and other experimental conditions. This stepwise structure reflect the mechanistic sequence of ROS-mediated degradation. The optimized CatBoost model exhibited robust predictive performance by achieving coefficients of determination of 0.9941 and 0.9986 and root mean squared errors of 2.0724 and 0.9054 for the first and second stages, respectively. Additionally, the analysis revealed that frequency influences degradation efficiency indirectly by modulating H₂O₂ production. Partial dependence plot analysis demonstrated a nonlinear relationship, where degradation efficiency initially increased with H₂O₂ concentration but plateaued at higher levels due to catalyst surface interactions and radical scavenging effects. This study establishes a machine learning framework that integrates ROS into predictive modeling, enhancing the mechanistic understanding and optimization of ultrasound-based AOPs with sonocatalysts.