Extending the WoFSCast – a storm-scale weather emulator – to ensemble prediction

We have recently developed a deep learning model that skillfully emulates deterministic convection-allowing model forecasts. Adapted from the Google GraphCast and trained on archived forecasts from the NSSL Warn-on-Forecast System (WoFS), the “WoFSCast” generates 3-km predictions that well match corresponding WoFS ensemble member forecasts out to at least 2 h. Owing to the very short time step (10 min) used for training WoFSCast, forecast blurring (i.e., loss of energy at small scales) is modest even without the use of generative modeling, and WoFSCast predicts real-world storms as skillfully as does the WoFS. We are beginning to explore various avenues for building upon WoFSCast’s success, including: incorporating additional state and diagnostic variables; reducing the training time step to 5 min; and increasing the number of vertical levels (currently 17). Perhaps most importantly, we will assess different strategies for using WoFSCast to generate ensemble forecasts, including aggregating WoFS and WoFSCast forecasts, and training a diffusion model in-line with WoFSCast that generates samples of the forecast distribution. We will endeavor to develop an ensemble strategy that mitigates the systematic underdispersion suffered by the WoFS and all other convection-allowing ensembles. We will report our most important findings at the meeting.