Evaluating and elucidating the reactivity of OH radicals with atmospheric organic pollutants: Reaction kinetics and mechanisms by machine learning

The rate constants of the reactions of OH radicals with atmospheric organic pollutants (AOPs) are crucial physicochemical parameters to guide in the elucidation of the kinetics and mechanisms of the reactive landscape. The experimental and theoretical difficulties in revealing the reactivity of these degradation processes motivated us to develop a protocol based on machine learning combining molecular fingerprints to estimate their rate constants. The present workflow is based on Organization for Economic Cooperation and Development (OECD) guidelines and state-of-the-art techniques involving (i) data collection including 903 AOPs cataloged in the literature, (ii) pre-processing and structuring of data, (iii) development of models based on three machine learning algorithms, (iv) the standard reference of validation, and (v) mechanistic interpretation. The results show that the built model has a high predictive capacity - R2 cv > 0.959 and RMSEcv < 0.090 for the training set and R2ext and Q2 ext > 0.889 and RMSEext < 0.084 for the test set. Additionally, through the SHapley Additive exPlanations (SHAP) method, it was possible to establish insight into the contribution of chemical classes to reaction kinetics and mechanism and to discuss it consistently with current experimental and theoretical observations. The availability of the evaluated reaction rate constants permitted to elucidate the role of AOPs in the photochemical ozone balance. Finally, to disseminate use of our results, we have presented them in a user friendly web application that permits compilation of kinetic parameters, and that permits future implementations to account for the temperature dependence in the environmental relevant range, and the consideration of a wider class of chemicals and processes in the mechanistic networks of atmospheric reactivity.

Automated summary

The rate constants of OH reactions with atmospheric organic pollutants are important parameters for understanding the kinetics and mechanisms of these reactions. This study developed a protocol based on machine learning and molecular fingerprints to estimate the rate constants, and the results showed that the model had a high predictive capacity. Additionally, the contribution of chemical classes to reaction kinetics and mechanism was analyzed using the SHAP method.