Microsoft’s foundation model treats Earth as a prediction problem

What it does Aurora is a machine learning model for Earth system forecasting. It predicts atmospheric variables like temperature, along with air pollution and ocean waves. Microsoft provides four specialized variants—medium- and high-resolution weather, air pollution, and ocean wave prediction—built atop a general pretrained backbone. The release includes the model implementation and downloadable checkpoints so researchers can run inference or build on the work.

The interesting bit The foundation-model approach means Aurora learns a broad representation of the planet first, then adapts to specific domains with relatively little extra data. The documentation is also refreshingly candid about limits, explicitly warning that the released checkpoints are streamlined versions of the internal models and may differ in subtle ways.

Key highlights

• Published in Nature (May 2025) with training data spanning ERA5, CMIP6, HRES, GFS, and CAMS reanalysis.
• Four task-tuned variants out of the box: medium-res weather, high-res weather (0.1°), air pollution (NO₂), and ocean wave direction.
• Includes pretrained weights; loading the small model triggers roughly a 500 MB download.
• Evaluated on extreme events like Storm Ciarán and rare heat/cold anomalies, not just root-mean-square error.
• Licensed for research; Microsoft explicitly notes it was not developed or tested for non-academic use.

Caveats

• The released models are “streamlined versions” of the originals and may deviate in subtle, unintended ways that affect performance.
• No accuracy guarantees: neural-network predictions can fail arbitrarily on out-of-distribution or even unlucky in-distribution inputs.
• Outputs are not operationally ready; the authors warn against using them directly for planning without additional verification tests.

Verdict Atmospheric-science and climate-ML researchers should grab this for reproducibility and fine-tuning. If you need a production weather pipeline or guaranteed accuracy, the README explicitly tells you to look elsewhere.