NOAA Global System Laboratory (GSL) has developed an experimental FV3-based limited area Rapid Refresh Forecast System (RRFS) Smoke and Dust model (RRFS-SD) that will be in operation at NCEP/EMC. This presentation describes the development of surface Particulate Matter (PM2.5 and PM10) assimilation scheme for providing accurate smoke and dust initial conditions to RRFS-SD in the framework of the Gridpoint Statistical Interpolation (GSI) three-dimensional variational (3D-Var) data assimilation system. The impact of the PM2.5 observations from the PurpleAir and AirNOW observing network on fire prediction is evaluated using the developed GSI system with the heavy fire events taking place in the US during September 2020. In general, the assimilation of PM2.5 observations reduces the bias in 24h PM2.5 simulation during the heavy fire events. The impact of the PurpleAir PM2.5 is comparable to the AirNOW observations during the peak of the fire events. Challenges in assimilating surface PM2.5 will also be discussed.