Microsoft (MS) reported that its weather forecasting artificial intelligence (AI) model "Aurora World," launched in May last year, has demonstrated performance that surpasses existing Earth system forecasting models. It performed accurately and efficiently in weather forecasting, air quality, typhoon paths, and marine wave predictions.
An international research team, including Microsoft and the University of Pennsylvania, published this information in the international journal "Nature" on the 22nd.
Aurora World has drawn attention since its public release. MS stated at the time that "Aurora World is a tool optimized for rapid prediction of air pollution," noting that it can predict global air pollution for five days and the weather for ten days at a speed about 5,000 times faster than traditional forecasting models.
The international research team has now broadened the application scope of Aurora World to the "Earth system" unit. Earth system forecasting is a field that integrates predictions of natural phenomena across the globe, including weather, ocean currents, sea ice, hurricanes (typhoons), and air quality. It plays a crucial role in early warnings about extreme weather events.
The problem is that such predictions require vast computations. Existing Earth system forecasting models were complexly built on decades of accumulated data, requiring supercomputers and dedicated personnel for operation. The variables included in the predictions, such as temperature, precipitation, ocean currents, air pollution, and sea ice, are numerous, leading to significant time and expenses.
Recently, with the rapid advancement of AI technology, expectations have arisen that Earth system forecasting can also be improved to be faster and cheaper. The research team trained Aurora World with about 1 million hours of geophysical data for Earth system unit forecasting.
Aurora World exhibited superior performance in air quality, wave, typhoon path, and high-resolution weather forecasting compared to existing models. On the other hand, the resources required for computation have significantly decreased. Notably, in predicting typhoon paths over five days, it was more accurate than all seven major forecasting centers globally. Its 10-day weather predictions showed that 92% were better than those from other centers.
Another strength of Aurora World is its rapid learning speed. The total training time took about 4 to 8 weeks, which is a significant difference compared to the years it takes to develop existing weather models. The research team explained that "the ability to improve performance in such a short period was due to the vast data accumulated through existing approaches."
The research team assessed that Aurora World could establish itself as a "foundational model" capable of predicting the entire Earth's system, extending beyond simple weather forecasting. It can be widely utilized for various environmental issues, such as monitoring air pollution and predicting climate change. They added that reducing the computation expenses required for predictions using AI could facilitate the easy utilization of precise climate information, greatly aiding in responding to the climate crisis.
References
Nature (2025), DOI: https://doi.org/10.1038/s41586-025-09005-y