Hong Kong, Jan 28, 2026, 21:14 HKT
- The HKUST team claims their new “DDMS” system can predict thunderstorms and heavy rainfall up to four hours in advance
- According to researchers, the model refreshes roughly every 15 minutes and boosted accuracy by over 15% in testing
- China’s Meteorological Administration and the Hong Kong Observatory are collaborating to incorporate it into their forecasting systems
Scientists in Hong Kong have unveiled an AI-driven weather system capable of forecasting thunderstorms and heavy rains up to four hours in advance, extending the lead time for storm alerts, the research team announced Wednesday. “We aim to leverage AI along with satellite data to enhance extreme weather forecasts and boost preparedness,” said Su Hui, chair professor at the Hong Kong University of Science and Technology and lead researcher on the project. 1
This announcement follows a record-breaking year of heavy rainfall in 2025 that scientists connected to climate change. Hong Kong’s observatory reported issuing its top rainstorm warning five times and the second-highest warning 16 times last year, both unprecedented numbers.
Those alerts usually leave little time to react when convective storms—rapidly developing thunderstorms that can unleash heavy rain—strike near the city. HKUST noted that current short-range forecasts for these events typically cover just 20 minutes to two hours, giving emergency responders and the public a tight window to prepare.
The new framework, called DDMS, updates forecasts roughly every 15 minutes and boosted accuracy by over 15% in tests, researchers reported. Created alongside China’s weather authorities, it’s designed specifically to predict heavy rainfall.
DDMS relies on a diffusion model, a generative AI method that works by adding “noise” to training data and then undoing that noise to generate sharper predictions. Its training data came from infrared brightness temperature readings—satellite measurements tied to cloud-top temperatures—gathered by China’s Fengyun-4A satellite between 2018 and 2021.
The team merged satellite data with meteorological insights to monitor the development of convective cloud systems, validating their model against spring and summer data from 2022 and 2023. Su noted that satellites detect cloud formation sooner than radar-based methods.
Dai Kuai, the study’s lead author, noted that radar signals are often hampered by terrain and precipitation, typically picking up changes only after convective clouds form. “Using satellite data that track cloud evolution from space, the new AI model can spot early signs of convective activity much sooner,” he explained. Su added that “the algorithm works with data from multiple satellites, broadening its reach.” 2
HKUST reported that the model generated forecasts at a 48-kilometre resolution and remained consistent down to 4 km scales. They described DDMS as the first AI system capable of predicting thunderstorm development four hours in advance across roughly 20 million square kilometres, covering China, Korea, and large parts of Southeast Asia.
The China Meteorological Administration and the Hong Kong Observatory are both moving to integrate the model into their forecasts, the researchers noted. According to HKUST, improved accuracy in warnings could benefit sectors like energy and insurance by enhancing risk evaluation.
Tech companies are pitching AI as a quicker alternative to numerical weather prediction (NWP)—the physics-based simulations that solve fluid-dynamics equations but demand heavy computing power. Nvidia unveiled three open-source “Earth-2” weather models this week, including one tailored to severe-storm forecasts up to six hours ahead for the US. “Once trained, AI is 1,000 times faster,” said Mike Pritchard, Nvidia’s director of climate simulation research. 3
DDMS has yet to show its chops in everyday forecasting, where agencies juggle satellite data, radar, and several models under pressure. The team trained it on data from 2018 to 2021, then tested it on samples from 2022 and 2023. Forecasters will be tracking how reliably it performs once the next wave of severe storms hits.
The researchers detailed their findings in a paper titled “Four-hour thunderstorm nowcasting using a deep diffusion model for satellite data,” published in the Proceedings of the National Academy of Sciences, according to HKUST. 4
