AIFS-Compo: A Data-Driven Atmospheric Composition Forecasting System

Machine learning has shown great success in numerical weather prediction. Here, we extend these advances to atmospheric composition forecasting by introducing AIFS-Compo, an AI-based system for predicting aerosols and reactive trace gases. Building on ECMWF’s AI weather forecasting framework, AIFS, we develop a large-scale graph-transformer model trained in two stages: first on CAMS EAC4 reanalysis data, and subsequently on a combination of CAMS analysis and forecast data. The resulting system produces 3-hourly forecasts and jointly uses prognostic variables from both numerical weather prediction and atmospheric composition.

When verifying against observations, AIFS-Compo achieves lower errors than the operational IFS-Compo system for 5-day forecasts of aerosol optical depth (AOD) and PM2.5, while showing comparable skill for reactive gases including ozone, carbon monoxide, nitrogen dioxide, and sulfur dioxide. Overall, AIFS-Compo delivers performance competitive with the operational system at a fraction of the computational cost. This efficiency for example enables extension to longer leadtimes, such as 10-day forecasts, supporting applications including early ozone hole prediction.