Hong Kong, Jan 28, 2026, 21:14 HKT
- HKUST team says new “DDMS” system can forecast thunderstorms and heavy rain up to four hours ahead
- Researchers say the model updates about every 15 minutes and improved accuracy by more than 15% in tests
- China’s Meteorological Administration and Hong Kong Observatory are working to integrate it into forecasting
Hong Kong scientists have developed an artificial intelligence weather system that can predict thunderstorms and heavy downpours up to four hours ahead, widening the window for storm warnings, the team said on Wednesday. “We hope to use AI and satellite data to improve prediction of extreme weather so we can be better prepared,” said Su Hui, a chair professor at the Hong Kong University of Science and Technology who led the work. Reuters
The announcement lands after a year of unusually wet weather in 2025 that scientists said was linked to climate change. Hong Kong issued its highest rainstorm warning five times and the second-highest 16 times last year, setting records, its observatory said.
Those alerts often come with little room to move when convective storms — fast-growing thunderstorms that can dump intense rain — flare up near the city. Existing short-range forecasts for these systems are often limited to 20 minutes to two hours, HKUST said, leaving emergency services and the public with a narrow margin.
The new framework, known as DDMS, refreshes forecasts about every 15 minutes and improved accuracy by more than 15% in testing, the researchers said. They developed it in collaboration with China’s weather authorities and said it is aimed at predicting heavy rainfall.
DDMS uses a diffusion model, a type of generative AI that learns by adding “noise” to training data and then reversing the process to produce more precise predictions. The system was trained on infrared brightness temperature data — satellite readings linked to cloud-top temperature — collected by China’s Fengyun-4A satellite from 2018 to 2021.
The team combined the satellite data with meteorological expertise to track how convective cloud systems evolve, and validated the model using spring and summer samples from 2022 and 2023. Su said satellites can pick up cloud formation earlier than radar-based systems.
Dai Kuai, first author of the study, said radar signals can be affected by terrain and precipitation and often detect changes only after convective clouds have already formed. “By leveraging satellite data that monitor cloud evolution from space, the new AI model can detect signs of convective development much earlier,” he said. Su said “the algorithm can be applied to data from different satellites, expanding its coverage.” Edu
HKUST said the model produced forecasts at a 48-kilometre resolution and performed steadily across scales down to 4 km. It called DDMS the first AI system able to forecast thunderstorm development four hours ahead over a region of about 20 million square kilometres, spanning China, Korea and much of Southeast Asia.
Both the China Meteorological Administration and the Hong Kong Observatory are working to incorporate the model into forecasts, the researchers said. HKUST said earlier, more precise warnings could also help industries such as energy and insurance by sharpening risk assessments.
The Hong Kong work comes as tech groups pitch AI as a faster alternative to numerical weather prediction, or NWP — physics-based simulations that solve fluid-dynamics equations and can be costly to run. Nvidia released three open-source “Earth-2” weather models this week, including one designed for severe-storm forecasts of up to six hours over the United States. “Once trained, AI is 1,000 times faster,” said Mike Pritchard, Nvidia’s director of climate simulation research. Reuters
But DDMS still has to prove itself in routine forecasting, where agencies blend satellite feeds, radar and multiple models under tight deadlines. The team trained it on 2018–2021 data and tested it on 2022–2023 samples; forecasters will be watching how consistently it performs when the next run of severe storms arrives.
The researchers described the work in a paper titled “Four-hour thunderstorm nowcasting using a deep diffusion model for satellite data,” which HKUST said was published in the Proceedings of the National Academy of Sciences. Pnas
