Objective To compare the spatio-temporal variations and their simulation efficiency of PM_2.5 concentration datasets simulated by different models.

Methods Nine sets of national PM_2.5 concentration simulation data that were published or shared by Chinese and international researchers from 2013 to 2020 were collected. The spatial and temporal distribution patterns of the nine datasets were compared by statistical analysis and ArcGIS mapping. PyCharm was used to conduct regression evaluation on four datasets simulated by daily-value models.

Results The simulation result by different models showed different levels and rangs of simulation values in local areas, but had similar spatial distributions in general, which tended to be higher in the central and eastern parts and lower in the western regions. Except GBD dataset, PM_2.5 concentrations of the other eight datasets all showed an overall decreasing trend, and showed the same seasonal trend, which was the highest in winter, followed by spring and autumn, and the lowest in summer. Among daily-value models, the random forest model demonstrated the best simulation performance (R²=0.76), with relatively low root mean square error (RMSE, 21.96). Among monthly-value models, the space-time extremely randomized tree model showed the best simulation performance (R²=0.98), with relatively low RMSE (3.26).

Conclusion The simulation datasets show similar spatio-temporal distributions of PM_2.5 concentrations. Nonlinear machine learning models have superior simulation performance to atmospheric chemistry models and linear regression models. In the future, the advantages of nonlinear and ensemble machine learning models can be combined to simulate PM_2.5 concentration data, which may further improve spatio-temporal resolution and simulation efficiency of the model.