AirMPA: a model predicts 13 air pollutants on a global scale and outperforms the Copernicus service in 90 per cent of cases

By separating weather data from pollutant chemistry, this Swin-Transformer-type model forecasts air quality up to five days in advance and outperforms CAMS on all variables beyond 48 hours.

Numerical air pollution forecasting systems are expensive to compute and rely on incomplete parameterizations, while deep learning approaches often mix weather and pollutant variables in the same model. AirMPA, presented on October 19, 2026 by Yiheng Wang and his co-authors, takes another path: an adapter that uses weather fields, emission inventories and past concentrations to specifically predict pollution.

Five days ahead of schedule at 0.4° resolution

The model estimates 13 pollutants with a 0.4° grid and a 12-hour step. It is based on a modified Swin-Transformer, with a focus on transport along the latitude bands and a two-branch fusion. Trained with the European Centre's ERA5 reanalysis as a weather forcing, it consumes forecasts from external weather models in operation, which allows it to chain its predictions for up to five days.

According to the authors, AirMPA consistently outperforms Aurora and outperforms the Copernicus Atmosphere Monitoring Service (CAMS) in about 90% of pollutant-timeline combinations, and on all variables beyond 48 hours.