Air quality forecasts remain largely uncertain today: weather models struggle to incorporate the effects of millions of suspended particles, aerosols, which modify cloud formation and pollutant dispersion. An international team has published a study showing how machine learning, coupled with Copernicus satellite data, can reduce forecast uncertainty by several hours, opening the way for more reliable alerts for urban residents and farmers.
What the study reveals: Role of aerosols in predictive models
The analysis shows that aerosols - whether forest smoke, pollen, desert dust or volcanic sulfates - directly influence solar radiation and cloud microphysics. By integrating these particles into predictive models, researchers have observed a 15% improvement in 24-hour temperature forecast accuracy and a 20% reduction in the error of PM2.5 concentrations at 48 hours. These gains are mainly due to better representation of light absorption and diffusion processes, two crucial processes for calculating atmospheric energy balance.
The results come from a comparison between the ECMWF model, which traditionally uses very simplified aerosol fields, and a new hybrid model combining the GraphCast system and the Pangu-Weather neural network. The latter exploits Copernicus atmospheric data enriched by lidar and satellite spectrometry observations to create a high-resolution aerosol field. The study highlights that adding detailed information about aerosols reduces forecast uncertainty, especially during extreme events such as forest fires or sandstorms.
