Aurora is a Microsoft Research foundation model of the atmosphere, with 1.3 billion parameters, that predicts global weather patterns and atmospheric processes such as air pollution. It is trained in two phases: pre-training on diverse weather and climate data, then fine-tuning for tasks like 10-day global forecasting or 5-day air pollution prediction. Aurora 1.5, developed by Microsoft Weather, adds 22 forecast variables including cloud cover, precipitation and radiation, hourly resolution and probabilistic ensemble forecasts. It is available on GitHub and Microsoft Foundry.
Atmospheric scientists and meteorologists form the main audience, particularly those studying data-sparse regions or extreme weather. Machine learning researchers can build on the foundation-model approach, where a broadly pre-trained model adapts to specific tasks using smaller sets of high-quality data. Teams building weather products or planning around air quality can use the forecasts through the published code and the Foundry listing. Organizations that need uncertainty estimates for risk decisions benefit from the ensemble forecasting in version 1.5.
Forecasting teams use it to produce high-resolution global forecasts much faster than traditional numerical weather models while matching or exceeding their accuracy. Version 1.5 ensembles outperform the ECMWF ensemble on 88.9% of evaluated targets over days 1 to 10 and cut tropical cyclone track error by 16% compared with the original model. Researchers fine-tune it for air pollution prediction, study extreme events such as Storm Ciaran, and test forecasts in regions where observation data is limited. Energy, agriculture and emergency planning groups can feed the precipitation, cloud cover and radiation variables into their own planning models.
