Aurora

Aurora is a 1.3-billion-parameter foundation model of the atmosphere trained on more than 1 million hours of weather and climate simulations. It uses a flexible 3D Swin Transformer with Perceiver-based encoders and decoders and forecasts wind, temperature, air quality, and greenhouse-gas concentrations at 0.1° resolution (~11 km), running roughly 5,000× faster than the operational Integrated Forecasting System (IFS). Aurora matches or outperforms GraphCast and operational IFS across 94% of meteorological targets, and pretraining on diverse climate and reanalysis data enables few-shot adaptation to specific downstream tasks.

Aurora reframes weather and climate modeling, traditionally the domain of century-old numerical methods, as a foundation-model problem. It performs particularly well on extreme weather and upper-atmosphere forecasting and can predict multiple atmospheric variables simultaneously at high resolution. Published in Nature, Aurora is the headline result of Microsoft AI for Science’s earth-systems program and shows how large-scale pretraining can democratize accurate forecasting — including for data-sparse regions that have historically been least served by traditional NWP.