HONG KONG — A workforce of Hong Kong scientists has developed a man-made intelligence weather-forecasting system to predict thunderstorms and heavy downpours up to 4 hours forward, in contrast with the vary of 20 minutes to two hours now.
The system will assist governments and emergency companies reply extra successfully to more and more frequent extremes of climate linked to local weather change, the workforce from Hong Kong University of Science and Technology mentioned on Wednesday.
“We hope to use AI and satellite data to improve prediction of extreme weather so we can be better prepared,” mentioned Su Hui, chair professor of the college’s civil and environmental engineering division, who led the venture.
The system aimed to predict heavy rainfall, Su instructed a press convention to describe the work printed within the Proceedings of the National Academy of Sciences in December.
Its model applies generative AI methods, injecting noise into coaching information in order that the system learns to reverse the method within the effort to produce extra exact forecasts.
Developed in collaboration with China’s climate authorities, it refreshes forecasts each quarter-hour and has boosted accuracy by greater than 15 per cent, the workforce mentioned.
Such work is essential as a result of the variety of typhoons and episodes of moist climate Hong Kong and far of southern China confronted in 2025 far exceeded the seasonal norm, scientists mentioned.
The metropolis issued its highest rainstorm warning 5 occasions final yr and the second highest 16 occasions, setting new information, its observatory mentioned.
Both China’s Meteorological Administration and Hong Kong’s Observatory are working to incorporate the model into forecasts.
The workforce’s new AI framework, known as the Deep Diffusion Model primarily based on Satellite Data (DDMS), was skilled utilizing infrared brightness temperature information collected between 2018 and 2021 by China’s Fengyun‑4 satellite tv for pc.
Satellites can detect cloud formation sooner than different forecasting programs resembling radar, Su added.
The information was mixed with meteorological experience to seize the evolution of convective cloud programs and later validated with spring and summer season samples from 2022 and 2023.
[[nid:728662]]