Deep Learning Weather Prediction and Thunderstorms
Developing AI models that match or surpass the forecast skill of numerical weather prediction (NWP) systems while running significantly faster is a growing area of research. Most AI-NWP models, however, have been trained on global ECMWF Reanalysis version 5 (ERA5) data, which does not capture storm-scale evolution. By adapting Google’s GraphCast framework to a limited-area, storm-scale domain, we’ve trained a storm-scale emulator of the National Severe Storms Laboratory (NSSL) Warn-on-Forecast System (WoFS), a convection-allowing ensemble with 5-minute forecast outputs. Known as WoFSCast, this model was the first to show that AI-NWP methods can extend to rapidly evolving, small-scale phenomena like thunderstorms. In this talk, I’ll discuss details about the model, its performance compared to the WoFS, and results from real-time experiments conducted during the 2025 Hazardous Weather Testbed Spring Forecasting Experiment. I’ll also highlight ongoing research to extend WoFSCast to an ensemble-based framework.