Microsoft publishes Aurora, a foundation model for the Earth system

Background

Weather and climate information is produced through operational forecasting systems, which turn observations into predictions on a regular schedule. A foundation model takes a different starting point: it is trained on broad data and then adapted to specific tasks. Aurora is presented as a foundation model for the Earth system, and its paper was published in Nature on 21 May 2025; Microsoft Research promoted the work the same month.^[1][2]

Operational forecasting is a continuous activity: models are run repeatedly, and their outputs are meant to be usable in practice. Treating operational forecasts as the comparison point matters because it sets a demanding baseline, not a one-off experimental score. A foundation model, by contrast, is built for reuse: one pretrained model can be adapted to several tasks instead of requiring a separate model to be built from scratch for each one.^[1]

What happened

The paper's claim is not built on a single narrow improvement. It rests on beating operational forecasts in four separate domains rather than one: air quality, ocean waves, tropical cyclone tracks, and high-resolution weather.^[1] That four-domain structure is the core of the reported result. Each domain asks for a different kind of output, from pollution fields to wave conditions to storm tracks to fine-scale weather, so the claim is that the same approach can be measured across several forecasting problems at once.^[1]

Key details

The cost argument is central to the paper: it reports the same or better forecasts at orders of magnitude less compute.^[1] In operational forecasting, compute is a practical constraint because predictions must be run repeatedly and at useful resolution. A claim of comparable or improved output at far lower compute therefore changes the terms on which forecasting capability is discussed, not only the accuracy score. Fine-tuning is presented as the route to new applications without retraining from scratch.^[1] In general, fine-tuning means starting from an existing model and adjusting it on a smaller, task-specific dataset, rather than building a model from zero. That is what makes the model reusable as a base for new uses. The authors also frame the work as widening access to high-quality climate and weather information.^[1]

Why it matters

Read together, the claims place three things side by side: performance across four domains, a large reduction in compute, and adaptation through fine-tuning. The paper does not present these as separate stories; the cost and fine-tuning points support the access argument.^[1] The significance claimed is therefore both technical and practical: better or equal forecasts in several areas, at lower computational cost, with a path to extend the model to new uses. For a foundation model for the Earth system, that combination is the basis on which the work is offered to readers and to the wider weather and climate community.^[1]