Chinese researchers have developed a new typhoon forecasting system that integrates artificial intelligence (AI) with atmospheric dynamics to achieve more accurate track predictions, according to a study recently published in the journal Advances in Atmospheric Sciences.

Due to the chaotic butterfly effect inherent in the atmosphere, tiny perturbations in the initial state can be continuously amplified during weather evolution, ultimately leading to deviations in typhoon tracks. This has long been a challenge in medium- and long-term typhoon forecasting.

In recent years, AI-based meteorological forecasting has gained traction, significantly improving the efficiency of weather prediction through enhanced computing capabilities. However, most AI models rely on empirical pattern recognition from data and lack the support of real atmospheric physical laws, resulting in instability in medium- and long-term forecasting and making it difficult to fully meet practical needs.

A research team led by Duan Wansuo from the Institute of Atmospheric Physics under the Chinese Academy of Sciences, in collaboration with Li Hao's team from Fudan University, introduced nonlinear dynamics algorithms into the domestically developed FuXi meteorological large model, creating the FuXi-CNOPs typhoon ensemble forecasting system.

According to the researchers, the new system can precisely identify the critical and sensitive areas that influence typhoon movement, screen out the initial meteorological perturbations most likely to amplify errors and alter a typhoon's trajectory, and generate forecast data that conform to real atmospheric physical laws.

Validation using 62 typical typhoons and 91 comparative experiments showed the new system delivers steady performance improvements. In 24-hour short-term forecasting, its technical performance is largely on par with leading global operational systems. Its advantages become more pronounced in medium- and long-term forecasting from 24 to 120 hours.

The system also delivers greater efficiency. While traditional leading systems require 51 sets of computational data to complete a forecast, FuXi-CNOPs achieves more accurate and stable results with just 31 sets, significantly reducing computing resource consumption and improving efficiency.