Elucidating the role of reactive nitrogen species in micropollutant degradation using machine learning.

Reactive nitrogen species (RNS, e.g., NO2•, NO•, and NH2•) are generated in many natural and engineered water systems but their roles on micropollutant degradation are largely overlooked. Experimental determination of micropollutants' reactivity towards RNS requires tedious procedures. We herein developed machine learning (ML) models to establish the quantitative relationship between the chemical structures of micropollutants (described by Molecular Access System (MACCS) and morgan fingerprints (MF)) and their reactivities toward two representative RNS (NO2• and NH2•). We also applied Bayesian Optimization for tuning models' hyperparameters. Among the 10 ML algorithms, the XGBoost model coupled with MACCS achieved the best balance of precision and generalization (R2train = 0.9140, R2test = 0.8871). The model's accuracy in prediction was also validated against experimental results using sulfanilamide, ciprofloxacin, and diphenhydramine as representative micropollutants. The validated model was employed to understand the key moieties of chemical structures that are reactive towards RNS. It is also used to predict the bimolecular reaction rate constants of 413 micropollutants (excluding in the dataset) toward RNS. A web interface was developed to make our model open-access for the engineers and practitioners in water treatment fields. This work not only enhances our fundamental understanding of RNS dynamics but also provides a predictive framework for designing advanced water treatment strategies.