Microsoft Research recentlyAnnounceIts open-source Aurora AI basic model, after combining deep learning with large-scale heterogeneous data processing technology, can not only significantly improve the accuracy of weather forecasts, but also promote deeper research and development of meteorological and Earth system science and technology by fine-tuning various environmental conditions such as ocean waves and air quality.
According to a paper published by Microsoft in the journal Nature, its open source model Aurora AI, which is deeply trained with more than 100 million hours of satellite, radar, and weather station computer simulation data, can now be used to estimate typhoon paths, analyze wave heights and tidal directions by fine-tuning various climate conditions, and can also further predict air pollution conditions.
In the past, the Aurora AI model predicted the path of Typhoon Doksuri, which affected Taiwan in 2023. It determined four days in advance that the typhoon would land on Luzon Island in the northern Philippines. At that time, the US Joint Typhoon Warning Center (JTWC) predicted that its path would directly deviate to the vicinity of Taiwan.
Furthermore, Aurora AI's trajectory predictions for global tropical cyclone forecasts between 2022 and 2023 were more accurate than those of seven major global meteorological centers, marking the first time that an artificial intelligence model's cross-forecast weather forecast accuracy has surpassed that of traditional numerical inference methods.
In addition to weather and cyclone analysis, Aurora AI currently also demonstrates highly accurate performance in ocean wave forecasting and air quality analysis and forecasting. It can analyze subtle changes in wave structure, enabling enhanced predictions of extreme ocean conditions caused by typhoons and other tropical cyclones. It can also analyze changes in air quality to determine the likelihood of air pollution impacts such as sandstorms.
At the same time, Microsoft emphasized that Aurora AI adopts a more flexible deep learning architecture, so it can process multiple data sources and data with different data structures simultaneously, and generate medium-term weather forecasts within seconds. Compared with the forecast systems used by traditional meteorological centers, the processing speed is 5000 times higher, and it can operate at lower time, electricity and other costs. At the same time, each data fine-tuning operation can be completed within only 4-8 weeks.
Aurora AI is currently used in Microsoft's MSN Weather weather forecast service and is also available for developer testing through Azure AI Foundry Labs. Microsoft has also released the Aurora AI model weights and some source code, allowing developers to modify, adjust, and fine-tune the model as needed, and use it to develop more weather application-related services.
