NOAA Research develops an AI-powered sibling to its flagship weather model

The National Oceanic and Atmospheric Administration (NOAA) Research has announced a innovative development in weather forecasting with the release of HRRR-Cast, an experimental AI-powered sibling to its renowned High-Resolution Rapid Refresh (HRRR) short-term weather model. This innovative data-driven system, trained on three years of HRRR data, promises to significantly enhance forecasting capabilities, particularly for hazardous weather events, while dramatically improving computational efficiency.

For a decade, the HRRR has been the National Weather Service’s (NWS) flagship operational model, providing hourly updated forecasts across the continental United States with a remarkable 3-kilometer resolution. Its ability to assimilate radar data four times an hour has been crucial for predicting precipitation and severe weather like thunderstorms and tornadoes. However, as a traditional physics-driven model, HRRR relies on complex mathematical equations and demands immense supercomputing power.

Enter HRRR-Cast, a testament to NOAA’s commitment to leveraging artificial intelligence for environmental prediction. Developed by NOAA’s Global Systems Laboratory (GSL), the same team behind the HRRR, HRRR-Cast is NOAA’s first regional experimental AI forecast system. It’s a key component of Project EAGLE, NOAA’s long-term initiative to rapidly test, develop, and identify promising AI models for global and regional forecasting.

Unlike its physics-based counterpart, HRRR-Cast “learns” by analyzing vast amounts of historical HRRR data, identifying intricate patterns to make predictions. Early evaluations are exceptionally promising, with HRRR-Cast demonstrating performance at least as good as the operational HRRR for reflectivity forecasts up to seven hours, and comparable performance for humidity, temperature, and wind. Crucially, HRRR-Cast excels in producing realistic depictions of storm structure.

Perhaps the most astonishing feature of HRRR-Cast is its computational efficiency. While the operational HRRR requires a supercomputer, HRRR-Cast is light enough to run on a single laptop, being 100 to 1000 times more efficient. This drastic reduction in computational demands opens new avenues for more frequent updates, more ensemble members for probabilistic forecasting, and wider accessibility for weather prediction.

The introduction of HRRR-Cast signifies a pivotal shift towards a hybrid approach in weather forecasting, combining the strengths of traditional physics-based models with the speed and pattern-recognition capabilities of AI. As NOAA continues to refine and expand HRRR-Cast, including the development of an ensemble version expected next month, the future of accurate and timely weather prediction looks brighter than ever.