Aurora Large-scale Atmospheric Basic Model Demo
Tutorial Introduction
With global climate change and frequent natural disasters, accurate and reliable Earth system predictions are essential to mitigate the impact of disasters and support the progress of human society. Although traditional numerical models are powerful, they are extremely computationally expensive, limiting their widespread application. In recent years, artificial intelligence has shown great potential in the field of environmental prediction, especially in improving prediction performance and efficiency. However, the potential of AI in several key areas of the Earth system remains unexplored.
To meet this challenge, a research team from Microsoft and its collaborators launched the first large-scale atmospheric foundation model, Aurora, which can accurately predict a variety of earth system variables such as air quality, ocean waves, tropical cyclone paths, and high-resolution weather by pre-training on more than one million hours of diverse geophysical data and fine-tuning on multiple specific tasks. While significantly reducing computing costs, it surpasses the performance of existing operational forecasting systems and promotes widespread access to high-quality climate and weather information. It has been verified that Aurora's computing speed is about 5,000 times faster than the most advanced numerical forecast system IFS.
The following are the specific research results achieved by Aurora in different fields:
- In terms of air quality forecasting, Aurora outperformed resource-intensive numerical atmospheric chemistry simulations at 0.4° resolution in a 5-day global air pollution forecast, outperforming the 74% target;
- In the area of ocean wave prediction, it outperformed expensive numerical models in the 10-day global ocean wave forecast at 0.25° resolution on the 86% target;
- For the 5-day tropical cyclone track forecast, Aurora comprehensively outperformed the seven operational forecast centers, achieving an outperform rate of 100% on all targets;
- In 10-day global weather forecasts, Aurora outperformed the state-of-the-art numerical models on the 92% target at 0.1° resolution, while also improving the prediction performance for extreme events.
In terms of model structure, Aurora adopts the 3D Swin Transformer architecture, combined with the 3D Perceiver encoder and decoder. The model consists of three parts: encoder, processor, and decoder. The encoder converts heterogeneous inputs into a universal 3D potential representation, the processor achieves forward evolution in time through the 3D Swin Transformer, and the decoder converts the potential representation back to a physical prediction.
Related research papers are titled "A foundation model for the Earth system" has been published in Nature.
This tutorial uses resources for a single card A6000.
The "Workspace" contains the following notebook automation script demonstration examples:
- The demonstration example is "Prediction of ERA5"
English version:demo_Predictions for ERA5.ipynb
Chinese version:demo_Predictions for ERA5-cn.ipynb
- The demonstration example is "HRES T0 prediction"
English version:demo_Predictions for HRES T0.ipynb
Chinese version:demo_Predictions for HRES T0-cn.ipynb
- The demonstration example is "HRES prediction at 0.1 degree resolution"
English version:demo_Predictions for HRES at 0.1 Degrees.ipynb
Chinese version:demo_Predictions for HRES at 0.1 Degrees-cn.ipynb
- The demonstration example is "Air Pollution Forecast"
English version:demo_Predictions for Air Pollution.ipynb
Chinese version:demo_Predictions for Air Pollution-cn.ipynb
- The demonstration example is "Typhoon Nanmadol track forecast"
English version:demo_Track Predictions for TyphoonNanmadol.ipynb
Chinese version:demo_Track Predictions for TyphoonNanmadol-cn.ipynb
Model Introduction
1. aurora-0.4-air-pollution
Aurora-0.4-air-pollution represents a paradigm breakthrough in AI in the field of earth sciences, which achieves efficient modeling of complex atmospheric chemical processes through data-driven methods. The model has been proven to be reliable in actual business (such as integration into Microsoft MSN Weather Service) and provides new technical tools for environmental governance and public health.
Demo Example - Air Pollution Forecast
English version:demo_Predictions for Air Pollution.ipynb
Chinese version:demo_Predictions for Air Pollution-cn.ipynb
2. aurora-0.25-pretrained
aurora-0.25-pretrained is based on the innovative 3D Swin Transformer architecture, combined with the Perceiver encoder-decoder structure, which can flexibly process multi-scale and multi-variable atmospheric data.
Demonstration Example - Forecast for ERA5
English version:demo_Predictions for ERA5.ipynb
Chinese version:demo_Predictions for ERA5-cn.ipynb
3. aurora-0.25-finetuned
aurora-0.25-finetuned is a fine-tuned version of the Aurora model for specific tasks (such as 0.25° resolution prediction), which combines high efficiency, multi-task adaptability and high accuracy. Its technical core lies in the combination of flexible basic model architecture and large-scale data training, providing a new paradigm for prediction tasks in the field of earth science.
Demonstration example for -HRES T0 prediction
English version:demo_Predictions for HRES T0.ipynb
Chinese version:demo_Predictions for HRES T0-cn.ipynb
Demonstration example: Track forecast of Typhoon Nanmadol
English version:demo_Track Predictions for TyphoonNanmadol.ipynb
Chinese version:demo_Track Predictions for TyphoonNanmadol-cn.ipynb
4. aurora-0.1-finetuned
aurora-0.1-finetuned is a fine-tuned checkpoint file for high-performance atmospheric forecasting tasks. It can quickly generate 5-day global air pollution forecasts and 10-day weather forecasts at a resolution of 0.1° (about 11 kilometers), and its computational efficiency is about 5000 times higher than that of traditional numerical models.
Demonstration example of HRES prediction at -0.1 degree resolution
English version:demo_Predictions for HRES at 0.1 Degrees.ipynb
Chinese version:demo_Predictions for HRES at 0.1 Degrees-cn.ipynb
Exchange and discussion
🖌️ If you see a high-quality project, please leave a message in the background to recommend it! In addition, we have also established a tutorial exchange group. Welcome friends to scan the QR code and remark [SD Tutorial] to join the group to discuss various technical issues and share application effects↓
Citation Information
The citation information for this project is 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},
}