Abstract: The WRFDA-Chem system with the atmospheric chemistry three-dimensional variational (3DVAR) algorithm was developed and applied in the Rapid Refresh Multi-scale Analysis and Prediction System-Chem (RMAPS-Chem), and experiments were conducted with and without the assimilation of the hourly surface PM2.5 and PM10 mass concentration in November 2016 to analyze the impacts of data assimilation on forecasting. The 6-h cycle assimilation results demonstrate that the assimilation of the surface PM2.5 and PM10 observations significantly improved the model performance of PM2.5 and PM10 initial fields with an increase in the correlation by 0.27–0.37 and a reduction in the root mean square error (RMSE) of about 40%. The improvement of the PM2.5 and PM10 forecasts was acquired for over 24 h with the initial analyzed field; the RMSE of the 24-h forecast PM2.5 (PM10) was reduced by 25% (10%), and the correlation of PM2.5 (PM10) increased by 14% (25%), respectively. The increase in the data assimilation (DA) cycling frequency (from 6-h to hourly DA cycle) could further improve the PM2.5 and PM10 forecast. In future operational applications, additional experiments on the data quality control/filtering in the system should be considered to obtain an optimized assimilation performance. Since the biases reflected the combining results of model uncertainties from various aspects, better understanding and diagnosis of model uncertainties should be aimed to promote the synergistic development of the model and the data assimilation system in the future.