Ethan Campbell

Ethan Campbell, University of North Carolina at Charlotte

Finding Needles in a Haystack: Using a GOES Nowcasting Model to Predict Convective Initiation

Recorded Talk

While convection-allowing, physics-based models provide skillful guidance on severe weather threats over multi-day periods, nowcasting (short-term forecasting on the order of one to several hours) remains a problem not effectively addressed by traditional models, whose spin-up time and data assimilation latency limit their ability to capture the precise timing and location of convective initiation. Recent advances in machine learning (ML) offer a pathway to close this gap by learning patterns from observational data and generating forecasts in minutes at inference time, compared to multi-hour runtimes typical of convection-allowing numerical weather prediction models. To address this gap, the Machine Integration and Learning for Earth Systems (MILES) group at NSF NCAR has developed an ML model, built on their CREDIT (Community Research Earth Digital Intelligence Twin) architecture, trained on GOES-16 satellite imagery, along with topography and top-of-atmosphere solar radiation, to forecast all 16 spectral bands one to three hours in the future. We evaluate this model using two case studies – an isolated supercell thunderstorm in the Dakotas in August 2024, and supercells embedded within a frontal system in the Plains in April 2024 – to assess model performance across different convective scenarios and identify next steps for retraining. Preliminary results show skill in predicting large-scale atmospheric dynamics, such as jet streaks and mid-latitude cyclones, but further work is needed to improve the model’s representation of convective initiation.

Mentors: Kevin Yang, DJ Gagne, Charlie Becker

Slides and poster