From flash floods and severe storms to high winds and extreme temperatures, timely and accurate weather forecasts help communities prepare and respond. Now, CIRES researchers working in NOAA’s Global Systems Laboratory (GSL) have developed a powerful new tool that could transform how those forecasts are made.

HRRRCast is an experimental artificial intelligence (AI) model designed to predict hazardous weather conditions. It delivers forecasts comparable to NOAA’s flagship models while running up to 100 times more efficiently. The system produces sharper forecasts and offers a scalable, cost-effective alternative to traditional physics-based models for regional ensemble forecasting.

NOAA’s High‑Resolution Rapid Refresh (HRRR) model has been the flagship short‑term weather forecasting model for over a decade. The model uses supercomputers to solve complex atmospheric physics equations that estimate how the atmosphere will behave over the next few hours or days. While highly accurate, running HRRR requires enormous computing power and time, limiting how many forecasts can be produced and how far into the future they can extend.

HRRRCast takes a different approach. Instead of calculating every physical process step by step, it uses AI to learn patterns from past weather data and applies that knowledge to current conditions using the power of a laptop. 

“Researchers realized they had more than a decade of incredibly rich, high‑resolution data from NOAA’s HRRR and Global Forecast System models,” said Daniel Abdi, a CIRES research scientist at NOAA GSL and lead author of the paper. “That made it possible to train an AI system to emulate the behavior of a trusted weather model.”

During testing, HRRRCast has performed as well as HRRR in predicting temperature, wind, humidity, and storm structure. It showed particular skill in capturing storm location and structure. HRRRCast can also generate ensembles—multiple plausible future scenarios—which help forecasters better understand what might happen and how confident they can be in their forecasts

HRRRCast is a key component of Project EAGLE (Experimental AI Global and Limited‑area Ensemble), NOAA’s long‑term initiative to rapidly test, evaluate, and transition AI weather models into operations.

Project EAGLE provides a real‑time testing environment where AI models can be evaluated using NOAA’s trusted forecast metrics and compared directly with operational systems. It already produces daily global forecasts using AI models such as GraphCast, and researchers are expanding it to include high‑resolution regional systems like HRRR‑Cast.

“NOAA scientists see AI and traditional physics‑based models as complementary,” said Isidora Jankov, Chief of GSL’s Scientific Computing Branch. “By improving physical models and the data they produce, we also improve the AI systems trained on that data. The future of forecasting is hybrid.”

NOAA will evaluate HRRRCast this year as a step toward transitioning the model to NOAA National Weather Service operations. 

This research was published in the American Meteorological Society journal Artificial Intelligence for Earth Systems.