WoFSCast: A Machine Learning Model for Predicting Thunderstorms at Watch-to-Warning Scales

Abstract

Developing AI models that match or exceed the forecast skill of numerical weather prediction (NWP) systems but run much more quickly is a burgeoning area of research. Most AI-NWP models, however, have been trained on global ECMWF Reanalysis version 5 data, which does not resolve storm-scale evolution. We have therefore adapted Google's GraphCast framework for limited-area, storm-scale domains, then trained on archived forecasts from the Warn-on-Forecast System (WoFS), a convection-allowing ensemble with 5-min forecast output. We evaluate the WoFSCast predictions using object-based verification, grid-based verification, spatial storm structure assessments, and spectra analysis. The WoFSCast closely emulates the WoFS environment fields, matches 70%–80% of WoFS storms out to 2-hr forecast times, and suffers only modest blurring. When verified against observed storms, WoFSCast produces contingency table statistics and fractions skill scores similar to WoFS. WoFSCast demonstrates that AI-NWP can be extended to rapidly evolving, small-scale phenomena like thunderstorms.