Aurora - Microsoft's basic atmospheric model
Aurora is a 1.3 billion parameter atmospheric model developed by Microsoft Research. It extracts valuable information from massive amounts of atmospheric data to predict global weather patterns, air pollution, and atmospheric processes such as ocean waves. The model uses pre-trained...
What is Aurora?
Aurora is a 1.3 billion-parameter atmospheric model developed by Microsoft Research. It extracts valuable information from massive amounts of atmospheric data to predict global weather patterns, air pollution, and atmospheric processes such as ocean waves. The model uses a pre-trained and fine-tuned architecture to handle data at different resolutions and pressure levels. Aurora has demonstrated outstanding performance in multiple forecasting tasks, including high-resolution weather forecasting, air pollution forecasting, and tropical cyclone trajectory forecasting, with a computation speed approximately 5000 times faster than traditional numerical weather models. The model improves forecast accuracy, reduces computational costs, and provides a powerful tool for addressing climate change and extreme weather events.
Aurora's main functions
- High-resolution weather forecastAurora can generate high-resolution (e.g., 0.1°) global weather forecasts, covering meteorological variables for the next 10 days, including temperature, wind speed, and air pressure.
- Air pollution forecastAurora can predict global air pollution levels, including the concentrations of pollutants such as nitrogen dioxide (NO₂), sulfur dioxide (SO₂), ozone (O₃), and particulate matter (PM1, PM2.5, PM10).
- Ocean wave predictionAurora can predict the height, period, and direction of ocean waves, which can help in shipping, coastal defense, and marine energy development.
- Tropical Cyclone Track PredictionAurora can accurately predict the movement trajectory of tropical cyclones, providing important support for disaster prevention and mitigation.
- Extreme weather event predictionAurora excels at predicting extreme weather events such as storms, floods, and droughts, providing early warnings and reducing disaster losses.
Aurora's technical principles
- Basic model architectureAurora employs a two-stage training approach: pre-training and fine-tuning. In the pre-training stage, the model learns general weather and climate representations on a large amount of diverse atmospheric data; in the fine-tuning stage, the model is optimized for specific tasks (such as weather forecasting or air pollution forecasting).
- 3D Swin TransformerAt its core is the flexible 3D Swin Transformer, which supports handling atmospheric variables with different spatial resolutions and pressure levels.
- Perceiver encoder and decoderA Perceiver-based encoder and decoder that processes heterogeneous inputs and generates predictions at different resolutions and fidelities.
- Data diversityDuring the pre-training phase, multiple data sources are used, including analytical data, reanalysis data, climate simulations, and operational forecast data. This diverse data input allows the model to learn from a wider range of atmospheric dynamic patterns.
- computational efficiencyBased on an efficient model architecture and training method, it significantly improves computational efficiency, generates high-precision predictions in a short time, and is about 5,000 times faster than traditional numerical weather models.
Aurora's project address
- Project official website:https://microsoft.github.io/aurora
- GitHub repository:https://github.com/microsoft/aurora
- arXiv technical paper:https://arxiv.org/pdf/2405.13063
Aurora's application scenarios
- Weather forecastIt provides 10-day global weather forecasts with a resolution of 0.1° to meteorological departments and related industries, helping to optimize agriculture, flight scheduling and disaster prevention.
- Air pollution forecastIt provides forecasts of global air pollution levels for environmental protection departments and public health agencies, helping to issue air quality alerts in advance and reduce public health risks.
- Ocean wave predictionIt provides shipping companies and coastal agencies with predictions of wave height, period, and direction, optimizes shipping routes, improves navigation safety, and supports coastal defense and marine energy development.
- Tropical Cyclone Track PredictionIt provides meteorological and emergency management departments with accurate forecasts of tropical cyclone movement trajectories, enabling them to issue early warnings and reduce disaster losses in coastal communities.
- Extreme weather event predictionEarly warnings of extreme weather events such as storms, floods, and droughts help governments and the public take preventative measures to mitigate the impact of disasters.