Microsoft Research announced in a May 27 postOpen SourceUnder Aurora AI Weather forecasting modelThe model combines deep learning and large-scale heterogeneous data processing techniques, and can not only accurately predict the weather, but can also be applied to a variety of natural environment monitoring fields such as ocean currents and air quality after fine-tuning.
According to the Microsoft team's paper, published in the journal Nature, the Aurora model was trained using more than 1 million hours of weather and environmental data from multiple sources, including satellites, radar, weather stations and computer simulations. Citing test data from its own research team, Microsoft claims that in the 2022-2023 global tropical cyclone prediction task, the Aurora model will be able to predict the weather and the environment in the future.Aurora's Trajectory Prediction Outperforms Industry Competitors and Traditional Observational Inference Algorithms Across the Board.
In addition to weather prediction, Aurora also demonstrates a high degree of accuracy in complex environmental scenarios such as wave forecasting and air quality prediction. In wave height prediction, Aurora claims to be able to capture detailed changes in wave structure, and has a better grasp of extreme sea conditions affected by typhoons; in air quality prediction, even though it has not specialized in learning atmospheric chemistry during the training phase, Aurora is still able to deduce air pollution events, such as dust storms, from a small amount of air quality data. In air quality prediction, Aurora is able to predict dust storms and other air pollution events from a small amount of air quality data, even though it has not specialized in atmospheric chemistry during the training stage.
at present,Aurora models integrated into Microsoft's MSN weather serviceMicrosoft has made the model weights and part of the source code available to developers for fine-tuning according to their needs. 1AI is accompanied by a project GitHub address as follows () Microsoft has publicized the model weights and part of the source code, developers can fine-tune according to their own needs, 1AI with the project GitHub address is as follows (https://microsoft.github.io/aurora/intro.html)