Aurora

Aurora, from Microsoft Research (2024, with a Nature paper), is a large-scale foundation model for the Earth system: pretrained on over one million hours of weather and climate data, then fine-tuned for specialized tasks. A single model family covers 10-day global weather forecasting (beating ECMWF IFS on 92% of targets at 0.1° resolution), 5-day air pollution prediction, ocean wave forecasting, and tropical cyclone track prediction — each at a fraction of the compute of dedicated systems. Microsoft open-sourced the code and weights, and a specialized version already powers hourly forecasts in the MSN Weather app.

Limitations: fine-tuning for a new variable or region still requires domain expertise and quality data; a July 2026 preprint (Aurora 1.5) added ensemble capabilities, so the original deterministic release is now superseded for uncertainty quantification. Like all data-driven weather models, it inherits biases from its training reanalysis data.