Colorado State researchers work to link AI models for earlier hail warnings

Colorado State University researchers are working to connect two AI forecasting models to give communities more notice of damaging hail. The project joins short-term radar forecasts with longer-range storm probabilities. Researchers plan spring field experiments to check the predictions against hail on the ground—a crucial step before the improved tools can reach forecasters.

Two clocks for the same storm

The two models address different parts of the forecasting problem. Hail Storm Scope uses radar data to forecast conditions up to two hours ahead. The CSU Machine Learning Probabilities model estimates the likelihood of severe storms and hail as far as eight days into the future, CBS News Colorado reported on October 28.

Those are complementary capabilities, not competing forecasts. The longer-range model looks for combinations of conditions that support hail. The short-term model addresses what happens as storms approach and evolve. The National Science Foundation-funded project combines a short-term “nowcasting” model developed with NVIDIA with CSU’s existing machine-learning system for longer-range prediction.

The distinction matters: a forecast of an environment favorable to severe storms is not yet a prediction of where damaging hail will fall. Thunderstorms can produce several hazards, and not every storm produces large hailstones. The team wants the combined system to narrow that gap and give forecasters more locally specific threat information.

The forecast windows researchers are connecting

01Up to two hours ahead

Hail Storm Scope

Uses radar observations to forecast near-term conditions.

02Up to eight days ahead

CSU Machine Learning Probabilities

Estimates the likelihood of severe storms and hail over a longer horizon.

Learning the conditions, resolving the mountains

The project initially focuses on Colorado and Wyoming, both hail hotspots because of their elevation. Researchers are refining the models with high-resolution weather data for Colorado and the surrounding region. They say the approach could also be adapted to other regions and weather conditions.

Associate atmospheric science professor Kristen Rasmussen said the team has developed models with roughly 1.3-kilometer resolution—just under a mile. She said these better represent storm interactions around the state’s mountains. The forecasting system draws on nine years of destructive U.S. storms, prioritizing factors such as temperature and wind speed before storms form and intensify.

Researchers describe NVIDIA’s underlying AI architecture as flexible enough to combine different radar observations with numerical weather predictions and satellite data. They also say the system can learn from additional real-time observations. Rasmussen called early results encouraging, while cautioning that these technologies and AI approaches have not been extensively tested, either individually or together.

The test is hail reaching the ground

Spring field experiments will use radar and additional instruments to check whether predicted hail actually reaches the ground where the models expect it. The team hopes to deploy improved forecasting tools later in 2027. That is a deployment goal, not a release date.

More notice could give farmers time to assess threats to crops, businesses time to protect vehicles and solar panels, and event organizers time to move people to shelter. Michigan Farm News reported that current models often provide no more than 30 minutes of hail warning; CSU’s ambition is hours or potentially days.

Rasmussen frames AI as another source of information for meteorologists, rather than a replacement. Their regional experience helps them judge which models to trust and which are performing well. She also stressed that training-data quality is critical to performance—a concern the team is addressing as it compares forecasts with actual hail observations.