Time-resolved prediction of hydroxyl radical exposure and micropollutant abatement during catalytic ozonation in natural water using machine learning methods

Abstract

Hydroxyl radical ('OH) generated during catalytic ozonation plays a central role in the degradation of ozone (O3)-resistant micropollutants (MPs). Accurate quantification of 'OH exposure (f ['OH]dt), which serves as a key indicator of oxidative capacity under defined conditions, is crucial for effective control of MPs in catalytic ozonation. However, direct in-field measurement of f ['OH]dt with a 'OH probe compound is not practical. This study explores the application of machine learning (ML) models to predict f ['OH]dt in catalytic ozonation using readily available input variables. Using graphitic carbon nitride as a representative catalyst, catalytic ozonation experiments were conducted to quantify f ['OH]dt (the output variable) under varying operational (O3 dose, catalyst dose, contact time) and water quality (TOC, pH) parameters (input variables). Two ML models were constructed: Model-1 predicted only the final f ['OH]dt after full O3 consumption, whereas Model-2 included contact time as an added variable to estimate f ['OH]dt in a time-dependent manner. Both models demonstrated strong predictive performance. More importantly, Model-2 effectively captured the temporal profile of f ['OH]dt and yielded greater accuracy for final f ['OH]dt estimates by more comprehensively reflecting the effects of each variable. The model-predicted f ['OH]dt values were subsequently used for simulating MP degradation, which corresponded closely with experimental observations for multiple O3-resistant MPs. These results demonstrate that f ['OH]dt predicted from readily obtained parameters can provide a robust indicator for MP removal, and that ML-powered methods present a valuable framework for optimizing catalytic ozonation systems.