ACOM Seminar: Development of NOAA’s Dynamic Forecasting of Air Composition (DAFCOM) using an optimiz

Speaker: Beiming Tang, NOAA ARL & George Mason University

NOAA’s current operational air quality forecasting system, UFS-AQM, has two major limitations: relatively coarse spatial resolution (13km) and modest accuracy in model predictions when evaluated against observations. To address these limitations, we developed DAFCOM, an optimized machine learning framework designed to generate hourly 1 km surface air pollution forecasts for the next 24 hours with substantially improved predictive accuracy. DAFCOM results are compared to both the original UFS-AQM forecasts and the current operational bias-correction results.

In addition to forecast mode, DAFCOM also includes a reanalysis mode capable of reconstructing historical high-resolution air pollution estimates. By integrating CMAQ simulations, TROPOMI NO₂ column observations, and high-resolution NEMO anthropogenic emissions, we developed a four-year (2017–2020) daily 1 km surface NO₂ dataset over the continental U.S. domain. Compared with previous studies, one of the key advances of this dataset is its ability to resolve urban-scale features, including city hotspots and highway pollution gradients. Sensitivity experiments further demonstrate that high-resolution chemical transport model outputs are essential for maintaining strong statistical performance, while high-resolution emission inventories are critical for enabling machine learning models to capture detailed urban-scale structures. In addition, TROPOMI NO2 served as better predictor compared with OMI NO2, not only because of its finer native spatial resolution, but also due to its higher temporal frequency.

Both the forecasting and reanalysis components of DAFCOM serve as important approaches of how machine learning can reshape the future of atmospheric model development.