Explained: Hong Kong scientists' AI weather model to forecast severe storms up to four hours ahead

A team of Hong Kong researchers have presented a revolutionary artificial intelligence system, which is likely to improve early warning capabilities for extreme weather occurrences, including thunderstorms and heavy rain.

Developed by the Hong Kong University of Science and Technology (HKUST) team, the new model gives emergency services and the general public far more time to prepare by extending the forecasting lead time from the conventional 20 minutes to a two-hour window, and now up to four hours before extreme weather occurs.

The discovery comes with an increase in climate-related severe weather, with recent record rainfall patterns in areas like Hong Kong. According to the HKUST team, the system will assist governments and emergency services in better responding to more common weather extremes associated with climate change.

Need for better short-term forecasting

Meteorologists have long struggled with accurate nowcasting, or short-range forecasting, of fast-emerging convective systems (such as severe thunderstorms). Conventional numerical weather prediction (NWP) models are based on the solution of physical equations of atmospheric motion, which is highly computational and often yields low accuracy at extremely short lead periods.

Conventional models often work well for more general weather patterns, but they have trouble identifying extremely localised, rapidly changing phenomena like the start of a thunderstorm or sudden rainfall. As a result, severe storm warnings typically have lead periods of only 20 minutes to two hours, giving authorities and locals limited time to prepare, evacuate, or reduce damage.

Introducing Deep Diffusion Model of Satellite (DDMS) Data

To address this limitation, the HKUST team developed the Deep Diffusion Model of Satellite Data (DDMS), a novel AI computational framework. The model uses cutting-edge generative AI techniques; to enable the mode, noise is injected into satellite data during training.

The DDMS system makes use of generative AI methods, particularly diffusion modelling, which teaches a neural model to rebuild and improve intricate patterns by introducing noise into input data.

DDMS is trained using infrared brightness temperature data from China's Fengyun-4 satellite, covering the years 2018–2021. Seasonal samples from 2022–2023 are used for validation. Compared to radar-based methods that only detect weather after convection strengthens, the model can identify minor early signs of convection and cloud formation far before they evolve into major storms via satellite data.

Performance and Collaboration

When compared to current short-range prediction systems, the DDMS framework has shown an accuracy gain of more than 15%, and updates forecasts every 15 minutes. The enhanced accuracy of the model is particularly pertinent to areas vulnerable to sudden, intense rainfall and thunderstorms, which led to record rainstorm warnings in Hong Kong in 2025.

It provides accurate forecasts with accuracy gains ranging from 3% to 16% and averaging 8.26% within the critical 2-4 hour forecast window lead time, where traditional models fall short.

The DDMS AI system was developed in collaboration with mainland China's meteorological organisations, such as the China Meteorological Administration (CMA)and national satellite centres. The Hong Kong Observatory and China's national forecasting agencies want to incorporate DDMS findings into operational predictions.

According to Dr Dai Kuai, the first author of the paper, DDMS is a significant development inatmospheric monitoring and early warning systems for severe weather, allowing for faster and more precise predictions and strengthening local readiness and response.

Commercial implications of AI model

The introduction of DDMS is a significant step towards improving regional weather resilience, particularly when climate change increases the frequency and severity of extreme weather occurrences.

According to Professor Su Hui, Climate Change and Extreme Weather Direction Lead, the approach also has a high potential for commercialisation. By offering earlier and more accurate risk assessments, it can help businesses like energy and insurance estimate the possible effects of extreme weather ahead of time and improve overall resilience.

The work was published in the Proceedings of the National Academy of Sciences in December.