Hong Kong scientists develop AI model that predicts severe thunderstorms four hours in advance

AI model forecasts severe thunderstorms 4 hours ahead with higher accuracy

In a critical advance for climate resilience, researchers from The Hong Kong University of Science and Technology (HKUST) have developed an AI model that can predict dangerous convective storms -- including Black Rainstorms, thunderstorms and extreme heavy rainfall like those that have hit Hong Kong -- up to four hours before they strike. This world-first technology, developed in collaboration with national meteorological institutions and powered by satellite data and advanced deep diffusion technology, improves forecast accuracy by over 15% at the 48-kilometer spatial scale compared with existing systems. This breakthrough strengthens the overall accuracy of the national weather forecasting system and promises to transform early warning systems for vulnerable communities across Asia. This research aligns with the core objectives of the State Key Laboratory of Climate Resilience for Coastal Cities (SKL-CRCC), which was established with the approval of the Ministry of Science and Technology of China last year. The laboratory operates under the directorship of Prof. Charles Ng Wang-Wai, Vice President for Institutional Advancement and CLP Holdings Professorship in Sustainability at HKUST. The research team includes Prof. Su Hui, the Climate Change and Extreme Weather Direction Lead of the SKL-CRCC, Chair Professor in the Department of Civil and Environmental Engineering, and Global STEM Professor at HKUST; Dr. Dai Kuai, Postdoctoral Fellow in the same department; as well as scholars from the Harbin Institute of Technology (Shenzhen), the China Meteorological Administration's (CMA) Institute of Tropical and Marine Meteorology (ITMM), and the National Satellite Meteorological Center (NSMC). The study was published in the Proceedings of the National Academy of Sciences under the title "Four-hour thunderstorm nowcasting using a deep diffusion model for satellite data." Why conventional forecasts fall short Extreme weather events have become increasingly frequent in recent years. Hong Kong issued four Black Rainstorm Warnings within just eight days last summer, while regions such as Bali in Indonesia, southern Thailand, and other regions also experienced severe flooding that resulted in significant casualties and economic losses. Conventional weather forecasts rely on numerical weather prediction (NWP) models, which simulate future atmospheric conditions by solving complex fluid-dynamical equations. However, NWP requires intensive computation and is highly sensitive to atmospheric chaos and limitations in observational data. For rapidly evolving, small scale convective systems -- including thunderstorms and rainstorms -- accurate forecasts are often limited to 20 minutes to two hours in advance. This short window leaves governments, emergency services, and the public with critically limited time to prepare, evacuate, or mitigate damage. How the new AI model works To address these challenges, the HKUST led team developed a new AI computational framework known as the Deep Diffusion Model of Satellite Data (DDMS). The model applies state of the art generative AI techniques: noise is added to satellite data during training, enabling the model to learn the reverse process of generating high-quality data. The team trained the model using infrared brightness temperature data collected by China's FengYun 4A satellite from 2018 to 2021, incorporating professional meteorological domain expertise to accurately capture the evolution of convective cloud structures. Model performance was validated using samples from the spring and summer seasons (May to August) of 2022 and 2023. Performance gains and key capabilities The team developed the world's first AI system capable of forecasting thunderstorm development four hours ahead, with accuracy improved by more than 15% at a resolution of 48 kilometers, compared to existing systems. Other technological breakthroughs include: * High resolution, high frequency forecasts updated at approximately 15 minute intervals, covering a region of about 20 million km, including China, Korea, Southeast Asia, and surrounding areas. * Stable performance across multiple spatial scales (4 km to 48 km) and throughout different seasons, with particularly strong accuracy in the 2-4 hour forecast window. Within this critical lead time, it excels precisely where conventional models fail most, delivering reliable forecasts with accuracy improvements ranging from 3% to 16% and averaging 8.26%. Benefits over radar-based forecasting Dr. Dai Kuai, the first author of the paper stated, "Conventional weather forecasting models rely mainly on ground based radar, but radar signals are easily affected by terrain and precipitation composition and often detect changes only after convective clouds have already formed. This results in delays in forecast lead time. "By leveraging satellite data that monitor cloud evolution from space, the new AI model can detect signs of convective development much earlier, enabling more timely warnings. DDMS represents a major advancement in atmospheric monitoring and severe-weather early warnings, enabling faster and more accurate forecasts and strengthening regional disaster preparedness and response." Broader impact and commercialization potential Prof. Su remarked, "This research is a collaboration between universities and national level institutions, including the CMA and the NSMC, and provides a valuable new reference model for operational forecasting. The algorithm can be applied to data from different satellites, expanding its coverage and enabling more countries and regions to respond effectively to rising climate risks. "The system also has strong commercialization potential: it can support industries such as energy and insurance by providing earlier and more precise risk assessments, helping organizations evaluate the potential impacts of extreme weather in advance and enhance overall resilience. We are moving from simply observing weather to intelligently anticipating it, which is a fundamental shift for safety and sustainability in a warming world."

Hong Kong scientists launch AI model to better predict extreme weather

HONG KONG, Jan 28 (Reuters) - A team of Hong Kong scientists has developed an artificial intelligence weather-forecasting system to predict thunderstorms and heavy downpours up to four hours ahead, compared with the range of 20 minutes to two hours now. The system will help governments and emergency services respond more effectively to increasingly frequent extremes of weather linked to climate change, the team from Hong Kong University of Science and Technology 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, chair professor of the university's civil and environmental engineering department, who led the project. The system aimed to predict heavy rainfall, Su told a press conference to describe the work published in the Proceedings of the National Academy of Sciences in December. Its model applies generative AI techniques, injecting noise into training data so that the system learns to reverse the process in the effort to produce more precise forecasts. Developed in collaboration with China's weather authorities, it refreshes forecasts every 15 minutes and has boosted accuracy by more than 15%, the team said. Such work is crucial because the number of typhoons and episodes of wet weather Hong Kong and much of southern China faced in 2025 far exceeded the seasonal norm, scientists said. The city issued its highest rainstorm warning five times last year and the second highest 16 times, setting new records, its observatory said. Both China's Meteorological Administration and Hong Kong's Observatory are working to incorporate the model into forecasts. The team's new AI framework, called the Deep Diffusion Model based on Satellite Data (DDMS), was trained using infrared brightness temperature data collected between 2018 and 2021 by China's Fengyun-4 satellite. Satellites can detect cloud formation earlier than other forecasting systems such as radar, Su added. The data was combined with meteorological expertise to capture the evolution of convective cloud systems and later validated with spring and summer samples from 2022 and 2023. (Reporting by Joyce Zhou; Writing by Farah Master; Editing by Clarence Fernandez)