Jun 06, 2024 14:00:00

Microsoft announces 'Aurora,' a super-fast AI that predicts 10-day global weather and air pollution in just one minute

With the development of AI, weather forecasts that once required long hours of running supercomputers can now be done in a short time with high accuracy. Meanwhile, Microsoft has announced ' Aurora ', an AI model that can instantly predict air pollution at the same time as the weather forecast.

Introducing Aurora: The first large-scale foundation model of the atmosphere - Microsoft Research

https://www.microsoft.com/en-us/research/blog/introducing-aurora-the-first-large-scale-foundation-model-of-the-atmosphere/

Superfast Microsoft AI is the first to predict air pollution for the whole world
https://www.nature.com/articles/d41586-024-01677-2

The unexpected destructive power of the bomb cyclone Ciaran that struck Europe in November 2023 left deep scars in areas centered around the UK and France, and demonstrated the limitations of existing weather forecasting models.

by

Xfigpower

To address these challenges, Microsoft has developed Aurora, an AI that can quickly analyze huge amounts of atmospheric data. Aurora is a foundational model with 1.3 billion parameters trained on more than 1 million hours of diverse weather data and climate simulations, and can perform a 10-day global weather forecast in one minute.

Aurora is also approximately 5,000 times faster than the current state-of-the-art numerical weather forecast model, the Integrated Forecast System (IFS), and is more accurate, efficient and versatile, allowing it to forecast a wide range of atmospheric variables, from temperature and wind speed to air pollution and greenhouse gas concentrations.

The figure below compares the total amount of nitrogen dioxide in the atmosphere, an indicator of air pollution, with Aurora's forecast (top) and data from the Copernicus Atmosphere Monitoring Service (CAMS) operated by the European Centre for Medium-Range Weather Forecasts (ECMWF) (bottom). Nitrogen dioxide tends to be concentrated in densely populated areas of East Asia, but Aurora was able to accurately capture both the local and global distribution of this pollutant.

The following shows Aurora's

root mean square error (RMSE) against CAMS, with the more negative values (blue) indicating that Aurora is better. For various metrics such as carbon monoxide (CO) and nitric oxide (NO), Aurora performed as well as or better than CAMS for 74% of the targets.

According to a study by Microsoft Research AI for Science AI researcher Paris Perdicaris and his colleagues, Aurora was able to predict the five-day global trends of key air pollution indicators (carbon monoxide, nitrogen oxides, nitrogen dioxide, sulfur dioxide, ozone, and particulate matter) within one minute, and this prediction was achieved at an order of magnitude lower computational cost than CAMS.

Matthew Chantry, machine learning researcher at ECMWF, said of Microsoft's research results: 'Until now, researchers in this field have relied on traditional mathematical models and machine learning. Aurora is the first AI model that can generate forecasts of global air pollution, a task that is much more complex than weather forecasting. The advantage of AI models is that they require fewer calculations to make predictions than traditional models.'

Microsoft hopes Aurora will serve as a model for the future development of AI-based environmental forecasting technologies, which it says could democratize accurate weather information in data-scarce areas, such as developing countries and remote regions.