Aurora: A Foundation Model for the Earth System

Implementation of the Aurora model for Earth system forecasting.

Link to the paper.

Please see the documentation for detailed instructions and more examples. You can also directly go to a full-fledged example that runs the model on ERA5.

Cite us as follows:

@article{bodnar2025aurora,
    title = {A Foundation Model for the Earth System},
    author = {Cristian Bodnar and Wessel P. Bruinsma and Ana Lucic and Megan Stanley and Anna Allen and Johannes Brandstetter and Patrick Garvan and Maik Riechert and Jonathan A. Weyn and Haiyu Dong and Jayesh K. Gupta and Kit Thambiratnam and Alexander T. Archibald and Chun-Chieh Wu and Elizabeth Heider and Max Welling and Richard E. Turner and Paris Perdikaris},
    journal = {Nature},
    year = {2025},
    month = {May},
    day = {21},
    issn = {1476-4687},
    doi = {10.1038/s41586-025-09005-y},
    url = {https://doi.org/10.1038/s41586-025-09005-y},
}

Contents:

• What is Aurora?
• Getting Started
• Contributing
• License
• Security
• Responsible AI Transparency Documentation
• Trademarks
• FAQ

Please email AIWeatherClimate@microsoft.com if you are interested in using Aurora for commercial applications. For research-related questions or technical support with the code here, please open an issue or reach out to the authors of the paper.

What is Aurora?

Aurora is a machine learning model that can predict atmospheric variables, such as temperature. It is a foundation model, which means that it was first generally trained on a lot of data, and then can be adapted to specialised atmospheric forecasting tasks with relatively little data. We provide four such specialised versions: one for medium-resolution weather prediction, one for high-resolution weather prediction, one for air pollution prediction, and one for ocean wave prediction.

Getting Started

Install with pip:

pip install microsoft-aurora

Or with conda / mamba:

mamba install microsoft-aurora -c conda-forge

Run the pretrained small model on random data:

from datetime import datetime

import torch

from aurora import AuroraSmallPretrained, Batch, Metadata

model = AuroraSmallPretrained()
model.load_checkpoint()

batch = Batch(
    surf_vars={k: torch.randn(1, 2, 17, 32) for k in ("2t", "10u", "10v", "msl")},
    static_vars={k: torch.randn(17, 32) for k in ("lsm", "z", "slt")},
    atmos_vars={k: torch.randn(1, 2, 4, 17, 32) for k in ("z", "u", "v", "t", "q")},
    metadata=Metadata(
        lat=torch.linspace(90, -90, 17),
        lon=torch.linspace(0, 360, 32 + 1)[:-1],
        time=(datetime(2020, 6, 1, 12, 0),),
        atmos_levels=(100, 250, 500, 850),
    ),
)

prediction = model.forward(batch)

print(prediction.surf_vars["2t"])

Note that this will incur a 500 MB download.

Please read the documentation for more detailed instructions and for which models are available.

Contributing

See CONTRIBUTING.md.

License

See LICENSE.txt.

Security

See SECURITY.md.

Responsible AI Transparency Documentation

An AI system includes not only the technology, but also the people who will use it, the people who will be affected by it, and the environment in which it is deployed. Creating a system that is fit for its intended purpose requires an understanding of how the technology works, its capabilities and limitations, and how to achieve the best performance. Microsoft has a broad effort to put our AI principles into practice. To find out more, see Responsible AI principles from Microsoft.

Use of this code

Our goal in publishing this code is (1) to facilitate reproducibility of our paper and (2) to support and accelerate further research into foundation model for atmospheric forecasting. This code has not been developed nor tested for non-academic purposes and hence should only be used as such completely at your own risk.

Intended use

Aurora is intended for medium- and high-resolution weather prediction, air pollution forecasting, ocean wave prediction, and global environmental modelling tasks. The model can be adapted to specialised atmospheric forecasting tasks with relatively limited task-specific training data due to its foundation-model architecture.

Out-of-scope use cases

Aurora is not designed or evaluated for direct operational decision-making without expert review, applications requiring guaranteed forecast accuracy, or non-environmental prediction tasks. Use in safety-critical planning or automated decision pipelines should be accompanied by appropriate domain validation.

Limitations

Although Aurora was trained to accurately predict future weather, air pollution, and ocean waves, Aurora is based on neural networks, which means that there are no strict guarantees that predictions will always be accurate. Altering the inputs, providing a sample that was not in the training set, or even providing a sample that was in the training set but is simply unlucky may result in arbitrarily poor predictions. In addition, even though Aurora was trained on a wide variety of data sets, it is possible that Aurora inherits biases present in any one of those data sets. A forecasting system like Aurora is only one piece of the puzzle in a weather prediction pipeline, and its outputs are not meant to be directly used by people or businesses to plan their operations. A series of additional verification tests are needed before it can become operationally useful.

In addition to the above, the models published here are streamlined versions of the models originally developed internally. Whereas we tried to be as thorough as possible, it is possible that the models published here deviate from the original model in subtle, unintended ways. This may affect predictive performance.

Responsible AI considerations

Aurora is a research forecasting model and should not be treated as an operational weather service. While the model can match or exceed traditional numerical baselines on established benchmarks, its reliability can degrade in out-of-distribution conditions (e.g., rare extremes, regime shifts, or regions/variables with limited historical fidelity in underlying reanalyses and simulations). Outputs may also be misinterpreted if users overlook uncertainty, ensemble spread, or known limitations of the underlying training data.

The primary Responsible AI risks are (1) unintended use by non-experts and (2) downstream use in consequential decision-making without domain validation (e.g., emergency response, critical infrastructure operations, safety-of-life planning). To mitigate these risks, developers should: clearly communicate that the release is for research evaluation and reproducibility; require domain-expert review before any real-world decisions are informed by outputs; implement basic input validation to ensure initial conditions come from credible sources (e.g., established meteorological agencies and data providers); and benchmark performance against accepted physical modelling systems for the specific geography, horizon, and variable(s) relevant to the intended application.

Aurora is best integrated as a decision-support component for expert analysis, not as a fully autonomous trigger for actions. For higher-risk scenarios, apply additional safeguards such as human-in-the-loop review, conservative thresholds for alerts, calibration/verification tests, and ongoing monitoring for drift when changing input sources or pre/post-processing.

Data

The models included in the code have been trained on a variety of publicly available data. A description of all data, including download links, can be found in Supplementary C of the paper. The checkpoints include data from ERA5, CMIP6 (CMCC-CM2-VHR4 and ECMWF-IFS-HR), HRES forecasts, GFS T0, GFS forecasts, HRES T0, HRES analysis, HRES-WAM analysis, CAMS reanalysis, and CAMS analysis.

Evaluations

All versions of Aurora were extensively evaluated by evaluating predictions on data not seen during training. These evaluations not only compare measures of accuracy, such as the root mean square error and anomaly correlation coefficient, but also look at the behaviour in extreme situations, like extreme heat and cold, and rare events, like Storm Ciarán in 2023. These evaluations are the main topic of the paper.

Note: The documentation included in this file is for informational purposes only and is not intended to supersede the applicable license terms.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.

FAQ

How do I setup the repo for local development?

First, install the repository in editable mode and setup pre-commit:

make install

To run the tests and print coverage, run

make test

You can then explore the coverage in the browser by opening htmlcov/index.html.

To locally build the documentation, run

make docs

To locally view the documentation, open docs/_build/index.html in your browser.