Machine learning-based optimization strategy for comprehensive toxicity regulation in catalytic ozonation systems

Heterogeneous catalytic ozonation mitigates aquatic toxicity, yet comprehensive regulation strategies encompassing catalyst properties, reaction conditions, and pollutant molecular descriptors remain largely unexplored. In this study, three feature categories were quantitatively correlated with comprehensive toxicity using a hyperparameter-optimized random forest model. Carbon number and HOMO defined intrinsic pollutant response, whereas pH, reaction time, and catalyst metal identity represented adjustable variables for engineering regulation. Within the represented chemical space, low-carbon/high-HOMO pollutants exhibited lower predicted toxicity in model-identified regions associated with lower-atomic-number metal catalysts, near-neutral pH, and reaction times below 40 min or above 100 min. Together, these molecular, catalyst, and operational features characterize a coupled multivariable response pattern in predicted comprehensive toxicity. Metal identity, statistically represented by atomic number, was associated with distinct toxicity responses. Density functional theory calculations of two selected representative catalysts, Fe₂O₃ and SnO₂, linked their contrasting O₃ activation to differences in local electronic structure and interfacial charge transfer, with Fe₂O₃ showing stronger O₃ adsorption and charge transfer. Furthermore, carbon number and highest occupied molecular orbital (HOMO) energy jointly influence electron distribution in pollutants, governing selectivity of bond cleavage. Low-carbon/high-HOMO pollutants localize electrons at electron-rich functional groups, favoring electrophilic ¹O₂/^•OH attack and more directed transformation toward low-toxicity products. Conversely, high-carbon/low-HOMO pollutants contain dispersed electron-rich and electron-deficient moieties, requiring ^•OH/O₂^•--dominated oxidation to limit pathway branching and drive transformation to lower predicted toxicity. These findings provide a model-guided basis for identifying catalyst-condition combinations associated with lower predicted toxicity within the represented chemical space.