A collaboration between Colorado State University (CSU) and NvidiaA has produced an AI-based system capable of forecasting severe hailstorms two to three hours in advance, compared to the 30 to 60 minutes offered by previous nowcasting methods.
The project combines generative AI with high-resolution radar observations to deliver real-time forecasts for rapidly evolving storms. The technology is aimed at regions such as Colorado and Wyoming, where severe hail can cause significant damage to homes, agriculture, solar farms and critical infrastructure with little advance warning.
The CSU effort is led by V Chandrasekar, university distinguished professor of electrical and computer engineering, working alongside researchers from CSU’s Cooperative Institute for Research in the Atmosphere and Nvidia. The focus is on nowcasting – short-term forecasting designed to support emergency response and help communities prepare before dangerous storms intensify.
Over several months, the team trained and deployed Hail-StormScope, a specialized version of Nvidia’s Earth-2 StormScope architecture, to generate hyper-local severe weather predictions across the Colorado-Wyoming region. The system is designed to be deployable across multiple radar networks in the USA, Europe and Asia.
“What stands out to me most about Nvidia’s StormScope architecture is its ability to flexibly integrate diverse weather observations, including weather radar and satellite data, along with numerical weather prediction model fields,” said Chandrasekar Radhakrishnan, ECE Research Scientist III and lead developer on the project. “This makes the system adaptable to different regions, resolutions and weather scenarios.”
Radar data integration was led by Sounak Biswas, a radar scientist with CSU and the National Oceanic and Atmospheric Administration‘s (NOAA) Physical Sciences Laboratory, which developed the high-resolution, multi-radar composites used to train and evaluate the model.
“What excites me most about this project is the opportunity to combine CSU’s radar and AI expertise with Nvidia’s framework,” said Biswas. “A key part of this work is building high-resolution radar composites that allow AI models to learn storm evolution from real observations. I hope this work ultimately supports more timely and actionable guidance for severe weather decision-making.”
The partnership is framed around both public safety and economic resilience, with stated goals of protecting lives and reducing insurance losses from severe weather events.
The work is supported by the National Science Foundation ASCEND Engine in Colorado and Wyoming program.
Related news, Cloud measurement campaign targets improved climate model accuracy