摘要:
The atmospheric aqueous-phase chemistry has received increasing attention in the last decades for its non-negligible environmental significance. Yet, the insufficient experimental data on oxidative reaction rate constants (kaq) obstructs the further analysis and modeling of this system. Predictive models based on machine learning (ML) algorithms have shown potential as an effective estimation tool, however, they are restricted to the lack of training data as well. To overcome this data limitation, we developed multi-task (MT) models that could exploit the common knowledge from reactions in gas- and aqueous-phases simultaneously. Toward kaq of organic compounds with hydroxyl radical (OH), nitrate radical (NO3), and ozone (O3), the MT models showed a notably better predictive ability compared to benchmark models, while obtaining wide applicability on compounds from different chemical classes. By interpreting the models using Shapley additive explanations (SHAP), we evidenced that the MT models utilized the common knowledge in both phases and correctly identified the reaction mechanisms. This study aims to provide new insight into the estimation of necessary kinetic parameters in atmospheric aqueous-phase chemistry, as well as a reference to ML research for other predictive tasks of atmospheric interest.
访问链接