2026/06/19 by Evan E. Sudler, Aaron J. Hill, Cameron R. Homeyer
Computer Science · Earth and Planetary Sciences · #Earthquake Detection and Analysis #Seismology and Earthquake Studies #Tropical and Extratropical Cyclones Research
paper · doi:10.1175/waf-d-25-0175.1
openalex created_date 2025/10/10 · openalex publication_date 2026/06/19 · openalex updated_date 2026/07/15
Abstract The recent emergence of machine learning in the field of weather prediction has shown remarkable promise. Deep learning models have proven to be a promising alternative to traditional physics and dynamics-based models with significantly lower computational costs given their high efficiency. While the performance of new artificial intelligence weather prediction (AIWP) models has demonstrated continuous improvement for general weather patterns, their efficacy in predicting high-impact, extreme weather events remains underexplored. To evaluate the performance of AIWP models in the face of extreme weather events, we compare short-range forecasts of three open-source machine learning weather prediction models [Pangu-Weather, Fourier forecasting neural network (FourCastNet) version 2, and GraphCast] and the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) for Hurricane Helene (2024). The AIWP models and GFS predict similar environments that support Hurricane Helene’s intensification stage, including deep-layer vertical wind shear and high midlevel relative humidity. In addition, the AIWP models accurately capture the track of the hurricane. However, the AIWP models underestimate cyclone intensity as measured through low-level winds and surface pressure, in contrast with the physics-based GFS model. Our case study aims to complement existing research on the fidelity of AIWP forecasts of extreme events, particularly in assessing their utility for operational forecasting where successive initialization updates provide critical guidance for refining hazard predictions.