New upgrades to HRRR-Cast: the experimental AI-powered regional model

HRRR-Cast Version 1 demonstrated that an AI emulator can successfully reproduce many aspects of the HRRR forecast model while being dramatically more efficient to run. Trained on about three years of HRRR analysis data and multiple lead times (1-, 3-, and 6-hour forecasts), it proved capable of producing realistic forecasts of reflectivity, temperature, humidity, and winds across the contiguous U.S. domain. Version 1 also showed early promise for probabilistic forecasting, as ensemble-like approaches suggested the potential to represent forecast uncertainty.

Now, building on that foundation is GSL’s release of HRRR-Cast Version 2. The vertical resolution was increased from 12 to 20 levels, which is more computationally demanding but leads to much better representation of upper-level atmospheric conditions, an advance especially valuable for aviation applications. At the surface, a suite of new variables was added, including wind components at 10 m and 80 m above the surface, two-meter dewpoint, two-meter relative humidity, total cloud cover, visibility, and cloud ceiling height. Precipitation fields were extended to provide hourly accumulations, and convective diagnostics such as CAPE and CIN were incorporated to support thunderstorm forecasting.

Just as impactful as the calculation of new forecast fields, Version 2 introduces the HRRR-Cast ensemble, with 9 members running every hour, producing forecasts out to 48 hours. This ensemble system lays the foundation for probabilistic forecasting and improves the model’s skill in storm placement and intensity.

With a bigger model comes bigger training sets. Four years of HRRR data were used to train Version 2 of HRRR-Cast in total. However, even with more data, the model has been optimized for efficiency. Through model design and algorithm improvements, it continues to run several orders of magnitude faster than current, physics-based models.

HRRR-Cast is being integrated into NOAA’s broader Project EAGLE (Experimental Artificial intelligence Global and Limited-area Ensemble), aiming to enhance the forecasts, physical representativeness of its fields, technical capabilities to further improve its computing speed, verification techniques, and development of probabilistic forecasts.

As of September 2025, HRRR-Cast Version 2 runs experimentally at the National Weather Service (NWS) Environmental Modeling Center (EMC) for testing and evaluation. Verification is being done using the novel “WxVx” tool, based in the Developmental Testbed Center’s Model Evaluation Tools (MET) system: a light, portable tool that creates model performance statistics, and the same tool being used to evaluate other components of Project EAGLE. GSL’s Model Analysis Tool Suite (MATS) is used to compare HRRR-Cast’s forecasts to those from the operational HRRR in real time. The ensemble forecasts will be accessible in GSL’s DESI visualization tool, including its individual members in a “postage stamp” display format. This will also play a crucial role in evaluating the model’s performance, while also making it available to NWS meteorologists and hydrologists in the field.

HRRR-Cast Version 3 is already in development and will be trained on a much larger dataset spanning roughly nine years of HRRR data at the native 3 km resolution, giving the model a stronger foundation. Another key focus will be on ensemble calibration, with the goal of producing more reliable spread and better probabilistic guidance. Release of Version 3 is expected in late 2025 or early 2026, marking the next step in evaluating HRRR-Cast for a future transition to NWS operations.

Key collaborators in the development of HRRR-Cast include EMC, the Cooperative Institute for Research in Environmental Sciences (CIRES, at the University of Colorado Boulder), and Cooperative Institute for Research in the Atmosphere (CIRA, at Colorado State University). Further collaborations in OAR extend to the National Severe Storms Laboratory (NSSL), Physical Sciences Laboratory (PSL), and Weather Program Office (WPO). HRRR-Cast’s integration with Project EAGLE will be done in coordination with the Earth Prediction Innovation Center (EPIC). The work is made possible with strategic investments from GSL and the NOAA Office of the Chief Information Officer’s (OCIO) High Performance Computing & Communications Incubator and Scientific Engineering and Novel Architecture (SENA) programs.