StoryResearch
Microsoft Research introduces Aurora, a 1.3B-parameter foundation model of the atmosphere
Microsoft Research described Aurora, an AI model trained on more than a million hours of weather and climate data. In the team's tests it beat both the leading physics-based forecast (ECMWF's IFS HRES) and Google's GraphCast on weather, and produced 5-day global air-pollution forecasts that beat state-of-the-art chemistry simulations on 74% of targets.
- 1.3 billion parameters; 3D Swin Transformer with Perceiver encoders and decoders.
- About 5,000 times faster than ECMWF's Integrated Forecasting System.
- Beat GraphCast on 94% of evaluated targets.