Microsoft's AI Aurora enables accurate prediction of air pollution
Climate forecasts are a major challenge. Microsoft's Aurora predicts not only the weather but also air pollution – in less than a minute.
Lightning and thunderstorms.
(Image: eskystudio / Shutterstock.com)
A research team from the "Microsoft Research AI for Science" group, together with international research partners, recently presented the artificial intelligence Aurora. It is the first foundation model for the Earth's atmosphere, i.e. a universally applicable AI that can be adapted for various tasks. The model should be able to calculate global forecasts for the weather and air pollution faster than conventional systems. The researchers published their results on the arXiv preprint server.
Traditionally, weather forecasts are made using numerical methods, but these require powerful supercomputers. Some research groups, such as those from Huawei, Nvidia and Google DeepMind, take a different approach: they use AI-based models to calculate forecasts faster and with greater accuracy.
Microsoft's latest model, Aurora, uses 1.3 billion parameters and was pre-trained for over a million hours with data from six weather and climate datasets. The researchers then retrained the model for specific tasks. The team attributes Aurora's high performance to this approach.
This makes it possible to gain a deep understanding of the processes in the atmosphere. In addition, the model should be able to calculate precise predictions, for example for temperature, wind speed, air pollution and the concentration of greenhouse gases, even with limited training data during fine-tuning.
Predictions in under a minute
Aurora is the first AI-based model that can predict global air pollution - for five days in less than a minute. It is orders of magnitude faster than the conventional "Copernicus Atmosphere Monitoring Service" model from the European Center for Medium-Range Weather Forecasts (ECMWF) in Reading, England.
Global weather forecasts are even possible for ten days. The system requires fewer resources than traditional weather forecasts. The research team estimates that the computing speed of Aurora is 5000 times faster than that of a conventional model. In a direct comparison with Google's GraphCast, Aurora usually delivered more reliable results. However, there is still too little data available for a general comparison. Researcher Matthew Chantry from ECMWF, who was not involved in the study, told the journal Nature: "You would have to invest a lot of time and probably have access to the models themselves to really go into detail and say with certainty that model A is better than model B".
Better early warning systems thanks to AI
Accurate weather forecasts are becoming increasingly important as extreme weather events become more frequent. However, predicting events such as hurricanes is a challenge even for advanced AI forecasting models due to the rapid increase in their intensity and high wind speeds. Aurora, on the other hand, should be able to predict the extent and severity of individual weather events and could therefore serve as an early warning system, the researchers speculate in the study.
(anw)