Extreme storms and rains: HRRRCast, emulator IA of the American weather model HRRR, does better on several measurements after a year in real time

Driven over nine years of 3-km-resolution analysis, HRRCast has been running in near real time for the past year. Over seven months of independent verification, its overall average exceeds the operational model on several surface variables.

AI-based weather models shine on a global scale, but still struggle on the scale of thunderstorms, where the most dangerous events take place. A team led by Daniel Abdi and Isidora Jankov presents a new version of HRRRCast, an emulator trained on the analyses of the High-Resolution Rapid Refresh (HRRR), the high-resolution regional model used in the United States.

Native resolution and rare events

Several choices explain the progress: training at the native resolution of 3 km rather than on a subsampled grid at 6 km, which avoids artifacts and significantly improves the score on convective scales; a weighting of the loss function that gives more weight to heavy precipitation and radar reflectivities, without degrading common situations; and a training archive extended from three to nine years (2016-2025), with more vertical levels and output variables.

HRRRCast has been running in near real-time for the past year. Over seven months of independent verification based on observations, the overall average improves several near-surface indicators compared to the operational HRRR. Its rainfall forecasts remain competitive up to about 25 mm per hour, and up to nearly 75 mm per hour in the best recent configuration. Limitation noted by the authors: the model generally places thunderstorms better than it reconstructs their structure.