AI Potluck
Back to Gap Map Model components / Scientific AI models

Aurora

Microsoft
open weights / Overall score: 2.7

Aurora is Microsoft's Earth-system foundation model for atmospheric and environmental forecasting. The published family includes specialized checkpoints for medium- and high-resolution weather, air pollution, ocean waves, and tropical-cyclone tracks, built from a common pretrained architecture and multiple public reanalysis and forecast datasets.

Verified 2026-09-10 via the Hugging Face model card and Microsoft repository.

Openness

3 high confidence
3.0
license
MIT(OSI)
weights
open(MIT ungated checkpoints)
data
documented-not-released(public ERA5, CMIP6, HRES, and GFS sources enumerated
code
partial(model and inference implementation without pretraining pipeline)

Aurora releases the multi-task checkpoint family and describes the public training sources, but not the pipeline that produced the pretrained base.

Adoption

2 medium confidence
2.0

The Hub reports zero downloads, so 1,014 GitHub stars are used as a last-resort reach signal.

Capability

3 high confidence
3.0

Two attributes. Adaptation is executable rather than merely described - docs/finetuning.md ships a Docker image and a fine-tuning loop built on AuroraPretrained - which is the bar ruling (b) sets. Generative is declined under ruling (d): Aurora predicts a deterministic future state, which is prediction rather than generation, and it is declined for the same reason aifs, pangu-weather and climax are. A list of specialized task names is not a further attribute.

Verified 2026-09-10