In Case You Missed It

In Case You Missed It

By: Qian Zhang, UL Research Institutes

Aerosols in the atmosphere play an important role in influencing air quality, weather and climate; therefore, accurate forecasting of atmospheric aerosols is essential for enhancing air pollution management, responding to emerging incidences, and tracking climate change. However, due to the complex aerosol compositions, physical dynamics and atmospheric chemical reactions, aerosol forecasting remains challenging with high uncertainties and computational costs. This recently published paper by Gui et al. developed a machine learning driven global aerosol-meteorology forecasting system (AI-GAMFS) to predict atmospheric aerosol optical components and surface concentrations.

AI-GAMFS predicts global aerosol optical depth (AOD) and surface mass concentrations of sulfate, dust, black carbon (BC), organic carbon (OC), sea salt (SS), and surface and upper-level meteorological variables for 5-day periods at 3-hourly intervals with approximately 50 km spatial resolution. By using machine learning, the model processes input meteorological and aerosol data via cube embedding to extract spatiotemporal features and then analyzes for complex features and relationships using a multiheaded self-attention mechanism, Lastly, it reconstructs high-dimensional features back into global forecast maps. AI-GAMFS was trained using 42 years of historical aerosol reanalysis data and was validated for more recent years against both reanalysis data and global observation networks. Depending on the predicted parameters, operational AI-GAMFS reduces the average root-mean-square-error (RMSE) by up to 37% compared to European forecasts for short-range, up to 25% for long-range, and up to 88% compared to US forecasts.

This study used three regional case studies, including East Asian mega dust storms, Saharan dust transport, and Central African and South American wildfires to verify the performance of AI-GAMFS. The model showed reliable forecasting of aerosol components, long-range dust and smoke transport, and emissions from wildfires that outperformed traditional models by enhancing correlation coefficient and reducing RMSE. This study presents a machine learning based operational framework that reliably predicts atmospheric aerosol parameters with feasible computational cost, which can enhance the understanding of aerosols impact on climate, air quality, and public health, as well as facilitate decision making and management during extreme weather events.

Read more in the literature:

Gui, K., Zhang, X., Che, H. et al. Advancing operational global aerosol forecasting with machine learning. Nature 651, 658–665 (2026). https://doi.org/10.1038/s41586-026-10234-y.