In June 2024, Microsoft Research introduced Aurora, which it described as the first large-scale foundation model of the atmosphere. Rather than training one bespoke network per task, Aurora is pretrained on more than a million hours of diverse weather and climate simulations and then fine-tuned to specific jobs - the foundation-model recipe that reshaped language modeling, applied to the Earth system.
Aurora is a 1.3-billion-parameter model that produces predictions at 0.1-degree resolution (about 11 kilometers at the equator). Because it first learns broadly from data, it can be adapted with comparatively small task-specific datasets to forecast not just standard weather variables but also air pollution, atmospheric chemistry such as nitrogen dioxide, and greenhouse gas concentrations. Microsoft reported roughly a 5,000-fold speed-up over traditional numerical weather prediction and said Aurora outperformed GraphCast on the large majority of targets, with particular strength on extreme events. A peer-reviewed version was published in Nature in 2025.
Aurora generalized the AI-weather idea from single-purpose forecasters toward a reusable, multi-domain Earth-system model - a direction that lets one pretrained backbone serve weather, air quality, and ocean forecasting alike.