PM_2.5 Assimilation within JEDI for NOAA's Regional Air Quality Model (AQMv7): Application to the September 2020 Western U.S. Wildfires

Abstract. This paper describes efforts to establish aerosol data assimilation capabilities for NOAA’s National Air Quality Forecasting Capability (NAQFC), a regional online air quality modeling (AQM) system under NOAA’s Unified Forecast System (UFS), by assimilating measurements of fine particulate matter (PM_2.5, particles with diameters less than 2.5 µm). PM_2.5 assimilation is developed within the Joint Effort for Data assimilation Integration (JEDI) framework and tested using its 3D-Var data assimilation (DA) component. The PM_2.5 observation operator is constructed by combining newly developed PM_2.5 transformation recipes in the JEDI Variable Derivation Repository (VADER) with a general spatial interpolation operator in the Unified Forward Operator (UFO). Cycled DA and forecast experiments were conducted from 1 to 21 September 2020, during a period of Western U.S. wildfires, to assess the impact of assimilating PM_2.5 observations from the AirNow and PurpleAir networks. The control and analysis variables include individual aerosol species, with background error standard deviations generated by scaling their respective background values. Prognostic variables such as aerosol particle number and total particulate surface area are updated accordingly following each analysis update. All DA experiments use a 3-hourly cycling interval, with PM_2.5 observations assimilated every 3 hours. The control experiment uses the same configuration but without any data assimilation. Results show that assimilating either AirNow or PurpleAir PM_2.5 data reduces 1–24 h forecast errors in terms of mean absolute error (MAE) and root mean square error (RMSE) compared to the control run over CONUS. Forecast skill, measured using the Critical Success Index (CSI) for PM_2.5 thresholds of 5, 12, and 35 µg/m³, also improves. AirNow observations have a greater impact overall, while PurpleAir shows its strongest impact over Nevada, northern Utah, Colorado, and southwestern New Mexico – regions with persistent underpredictions in the control run at forecast hour 1. Overall, the assimilation of PurpleAir observations in addition to AirNow data leads to a slight reduction in 3–24 h MAE.