Development of an integrated machine-learning and data assimilation framework for NO emission inversion.

As major air pollutants, nitrogen oxides (NO, mainly comprising NO and NO) not only have adverse effects on human health but also contribute to the formation of secondary pollutants, such as ozone and particulate nitrate. To acquire reasonable NO simulation results for further analysis, a reasonable emission inventory is needed for three-dimensional chemical transport models (3D-CTMs). In this study, a comprehensive emission adjustment framework for NO emission, which integrates the simulation results of the 3D-CTM, surface NO measurements, the three-dimensional variational data assimilation method, and an ensemble back propagation neural network, was proposed and applied to correct NO emissions over China for the summers of 2015 and 2020. Compared with the simulation using prior NO emissions, the root-mean-square error, normalized mean error, and normalized mean bias decreased by approximately 40 %, 40 %, and 60 % in NO simulation using posterior NO emissions corrected by the framework proposed in this work. Compared with the emissions for 2015, the NO emission generally decreased by an average of 5 % in the simulation domain for 2020, especially in Henan and Anhui provinces, where the percentage reductions reached 24 % and 19 %, respectively. The proposed framework is sufficiently flexible to correct emissions in other periods and regions. The framework can provide reliable and up-to-date emission information and can thus contribute to both scientific research and policy development relating to NO pollution.