Meet Aurora, Microsoft’s new AI model of the atmosphere

Microsoft researchers have developed a new AI foundation model, Aurora, that uses atmospheric data to help mitigate the effects of climate-related crises, according to a statement. While a number of tech companies have started developing AI weather-forecasting tools, including GraphCast from Google DeepMind and FourCastNet by Nvidia, Aurora is set apart by its quick ability to predict air pollution in less than a minute.

How does it work? Aurora, which is trained on over 1 mn hours of diverse weather climate simulations, can make predictions even in areas that lack comprehensive data or are experiencing extreme weather events, according to Microsoft. Its computational speed is estimated to be 5k times faster than the Integrated Forecasting System (IFS). Aurora is capable of predicting various atmospheric variables, such as temperature, wind speed, air pollution, and greenhouse gas concentrations. Its architecture, featuring a flexible 3D Swin Transformer with Perceiver-based encoders and decoders, allows it to handle heterogeneous inputs and generate high-fidelity predictions at different resolutions.

Not Microsoft’s first rodeo: The tech company partnered with a team of scientists from US government-backed Pacific Northwest National Laboratory back in January on a study aiming to leverage machine learning to identify and develop alternative materials to lithium metals used in EV battery production.