AI Numerical Weather Weather Prediction (NWP) emulators like Google DeepMind’s GraphCast could revolutionize weather prediction by providing more skillful forecasts at much less computational expense. While these emulators work well for global scales since decades of reanalyses at these scales are available for training, finer scales are a challenge because accurate, consistent, and long-term analysis datasets that resolve individual thunderstorms do not exist. National Severe Storms Laboratory (NSSL) is taking on this challenge by designing a system known as WoFSCast, which is based on Google DeepMind’s GraphCast framework, but modified to emulate the Warn-on-Forecast System (WoFS), a real-time, high-resolution prediction system. The WoFS, led by the NSSL and Cooperative Institute for Severe and High-Impact Weather Research and Operations (CIWRO), is a cloud-based, rapidly-updating, 3-km ensemble designed to support watch-to-warning (0–6 hr) operations for tornadoes, flash floods, and other high-impact weather. In recognition of the WoFS, NSSL and the Office of Oceanic and Atmospheric Research Global Systems Laboratory (OAR GSL) received a 2023 Department of Commerce Gold Medal Award for “scientific and engineering excellence in developing a revolutionary prediction tool that provides short-term probabilistic thunderstorm guidance.” To create WoFSCast, the open-source GraphCast code was modified to work with WoFS’ relocatable, 900-km regional domain and then trained and evaluated using over 100 days of WoFS forecasts. In an initial evaluation focusing on 2-h deterministic forecasts, WoFSCast was found to closely emulate the WoFS while running much faster: the WoFSCast can generate an 18-member, 6-h forecast in ~30 seconds on a single GPU, whereas the WoFS requires ~10 minutes on ~1100 of CPUs. The blazing speed of WoFSCast promises to improve the already impressively low latency of 0–6-h thunderstorm forecasts from WoFS, potentially enabling longer-lead warnings and forecasts for thunderstorm hazards. Beyond the reduced latency, WoFSCast enables real-time generation of very large forecast ensembles – comprising potentially hundreds of members – which may provide more accurate, better calibrated forecasts than the 18-member WoFS. NSSL is exploring several strategies for extending WoFSCast into an ensemble prediction system, including using diffusion models, training with probabilistic loss, and adapting Google’s GenCast to WoFS emulation. Real-time WoFSCast ensemble predictions will likely be evaluated in the NOAA Hazardous Weather Testbed and/or Hydrometerology Testbed starting in 2026. In addition, new WoFSCast models will be developed to perform storm-scale data assimilation and to generate 1-km forecasts. NSSL’s long-term vision is to develop an end-to-end data-driven ensemble system that provides super low-latency, well calibrated, exceptionally high-resolution forecasts of high-impact thunderstorms.
