Microsoft has enhanced its foundational AI model, Aurora, expanding its capabilities beyond advanced weather forecasting to include accurate air quality predictions. Developed by Microsoft Research, Aurora leverages deep learning to deliver precise forecasts for a range of meteorological phenomena, including hurricanes and typhoons, outpacing traditional forecasting methods in both speed and accuracy.
In a blog post released earlier this week, Microsoft confirmed that Aurora’s source code and model weights are now open-source, making the model accessible to researchers and developers globally. A specialized variant of Aurora that generates hourly forecasts, including cloud coverage, has already been integrated into the MSN Weather app.
What distinguishes Aurora is its foundation model architecture, which allows it to be fine-tuned for various forecasting tasks. This flexibility enables it to deliver insights into areas beyond conventional weather prediction, such as air pollution levels. According to Microsoft, Aurora was trained on over a million hours of satellite imagery, radar data, weather station readings, and historical simulations, enabling it to learn patterns across a broad spectrum of atmospheric variables.
Unlike traditional systems, Aurora doesn’t rely on predefined meteorological rules. Instead, it uses a deep-learning approach to determine which data interactions are most useful for accurate forecasting. “We’re just giving a large deep-learning model the option to learn whatever is most useful,” said Megan Stanley, Senior Researcher at Microsoft.
The model’s performance has been noteworthy. Microsoft claims that Aurora accurately predicted Typhoon Doksuri’s landfall in the Philippines four days in advance, outperforming several expert forecasts. It also successfully anticipated a sandstorm in Iraq two years ago and delivered five-day tropical cyclone path predictions in 2022 and 2023 that exceeded those of the US National Hurricane Center.
Powered by GPUs, Aurora produces forecasts in seconds, compared to the hours traditional systems running on supercomputers require. While its training phase demanded substantial computational resources, Microsoft noted that Aurora’s operational costs are significantly lower than those of conventional forecasting infrastructures.
Aurora joins a growing trend of AI-based weather models, with others like Google DeepMind’s WeatherNext also pushing the boundaries of meteorological forecasting. By combining massive datasets and deep-learning techniques, these models aim not only to improve the accuracy of forecasts but also to support efforts in climate change mitigation and disaster preparedness.
