A team of researchers from the Institute of Atmospheric Physics at the Chinese Academy of Sciences, working in collaboration with other institutions, has developed a new AI-based probabilistic typhoon forecasting system called FuXi-CNOPs. Built on the FuXi AI weather model, the system can predict typhoon trajectories with greater precision while also identifying and explaining the physical factors responsible for route deviations, making forecasts both more accurate and more transparent.
According to Duan Wansuo, the study's corresponding author, the technology allows artificial intelligence to move beyond functioning as a 'black box', enabling it to also explain the physical processes that drive its predictions.
Duan notes that typhoons are particularly difficult to forecast due to the chaotic nature of the atmosphere, where small changes in initial conditions can trigger significant shifts in storm trajectories. Rather than producing a single probable path, the new system generates multiple possible scenarios, allowing for a more comprehensive assessment of risks.
While leading international systems rely on 51 simulations to produce probabilistic forecasts, FuXi-CNOPs achieves superior results with just 31 simulations, reducing computational costs while improving accuracy.
One of the project's key innovations is the integration of a technique known as Conditional Nonlinear Optimal Perturbation, or CNOP, developed by the research team itself. The method identifies the primary sources of uncertainty influencing typhoon trajectories, making the AI-generated forecasts more transparent and interpretable.
The scientists found that these uncertainties are concentrated mainly in the storm's core, in its spiral rainbands, and in zones where the typhoon's circulation interacts with subtropical high-pressure systems — factors that are decisive in determining the movement of these weather formations.
The performance of FuXi-CNOPs was evaluated against 62 real typhoon cases through 91 forecasting experiments. Tests showed that for short-range forecasts of up to 24 hours, the system performs comparably to leading international models. Beyond 24 hours and up to 120 hours, however, its superiority becomes evident.
'The data also show a maximum reduction of 32.33% in trajectory forecast errors and an increase of up to 29.2% in the ability to quantify uncertainties associated with meteorological phenomena,' the research team stated.
For the researchers, FuXi-CNOPs represents a significant advance in the application of artificial intelligence to meteorology, paving the way for forecasting systems that are more explainable, transparent, and reliable — and better equipped to improve community preparedness in the face of extreme weather events.
Source: Diário Económico
Original article: https://www.diarioeconomico.co.mz/?p=525746
