Severe Weather Forecasts from Artificial Intelligence Weather Prediction Models

Abstract Using recent observations of severe weather events, random forest (RF) machine learning models are trained to predict severe weather at 1–8-day lead times mimicking operational convective outlooks produced at the NOAA Storm Prediction Center. The RFs use environments simulated by three global artificial intelligence–based weather prediction (AIWP) models: PanguWeather, FourcastNet v2-small, and Graphcast. Individual hazard forecasts and any hazard forecasts are generated at all lead times and are compared to operational machine learning–based hazard guidance products from the Global Ensemble Forecast System Machine Learning Probabilities model. Skill metrics are computed for forecasts generated across 2024, and ensembles of RF forecasts are created for an example case to understand the value of AIWP-based machine learning guidance products. Skillful individual hazard forecasts were obtained out to day 7, but the operational machine learning guidance often performed better at longer lead times for the any hazard forecasts and at shorter lead times for individual hazards. Combining the AIWP inputs as a single metric (i.e., median) did overcome some of the deficiencies of single-AIWP RFs, although PanguWeather-RF forecasts were competitive with the median-based forecasts. Further, creating ensembles of AIWP predictions improved skill and generated distributions of severe weather potential, demonstrating a new avenue to exploit the utility and value of postprocessed AIWP systems. Significance Statement Machine learning has become a useful postprocessing tool for operational forecasters whereby it can generate operational-like products that are used as first-guess fields. With the innovation of artificial intelligence–based weather prediction models, we explore how artificial intelligence (AI)-simulated weather environments that support severe weather can be leveraged to generate guidance products for operations. Random forests are used to make 1–8-day predictions of severe weather, including individual hazards, and demonstrate skill and utility commensurate with existing operational machine learning guidance products. Additionally, we explore ensemble techniques to improve skill and provide more probabilistic forecasting information.