NOAA GFDL - Geophysical Fluid Dynamics LaboratoryGFDL - Geophysical Fluid Dynamics Laboratory

Closing the gap — Hurricane Prediction Advances in the US FV3-based Models

June 10th, 2025

Key Findings

  • The gap in hurricane track forecast skill between the American FV3-based models and the world-leading European Centre model has substantially decreased since the major upgrade of the U.S. models’ dynamical core.
  • The 10-year progression of the various American and the European models’ North Atlantic hurricane forecast skills are presented to demonstrate this achievement.
  • NOAA’s newly operational HAFS models and the experimental T-SHiELD showed outstanding performance in both track and intensity forecasts.

Jan-Huey Chen, Timothy Marchok, Morris Bender, Kun Gao, Sundararaman Gopalakrishnan, Lucas Harris, Andrew Hazelton, Bin Liu, Avichal Mehra, Matthew Morin, Fanglin Yang, Xuejin Zhang, Zhan Zhang, Linjiong Zhou. Bulletin of the American Meteorological Society. DOI: 10.1175/BAMS-D-24-0036.1

Hurricanes are one of the most destructive natural events on Earth. Improving the skill of hurricane forecasts, especially for those in the North Atlantic basin, has always been an important objective for government weather forecast agencies, emergency managers, and the atmospheric science research community in the United States. The model developed by the European Centre for Medium-range Weather Forecasts (ECMWF) has been recognized as providing the most skillful guidance for track forecasts for years. However, the performance of American models for hurricane forecasting has been catching up.

This research paper presents the 10-year progression of the North Atlantic hurricane forecasts in the various American Finite-Volume Cubed-Sphere Dynamical Core (FV3)-based models, including the operational Global Forecast System (GFS), the operational Hurricane Analysis and Forecast System (HAFS), and GFDL’s experimental SHiELD.

From 2019 to 2023 the gap in hurricane forecast track skill between American models and the ECMWF’s Integrated Forecasting System (IFS) was substantially decreased, coinciding with the major upgrade of the U.S. models’ dynamical core.  In addition, the FV3-based global Global Forecast System (GFS) and the research-oriented System for High-resolution prediction on Earth-to-Local Domains (SHiELD) showed much improved hurricane intensity forecast skill, compared to IFS and the previous generation of American global models – marking a significant achievement of U.S. global model development.

The next-generation FV3-based regional hurricane models showed outstanding performance  in both hurricane  track and intensity forecasts. The newly operational HAFS-A and HAFS-B successfully reproduced the intensity forecast skill achieved by NOAA’s previous generation operational Hurricane Weather Research Forecast Model. Furthermore, both HAFS models and the experimental global-nested T-SHiELD showed comparable or even better hurricane track forecasts than the global models from 2020 to 2023, marking the first time any hurricane model has had good skill for both intensity and track. This represents a considerable success for NOAA in upgrading the agency’s regional hurricane models.

The improved hurricane track and intensity forecasts described by the authors can be taken as an  indicator of the progress made by the entire Unified Forecast System community, reflecting major upgrades of the dynamical core, physics and data assimilation in the operational GFS and HAFS. To further improve the American models’ hurricane prediction, many aspects are worth exploring and approaching, including data-driven machine-learning  weather models. Properly adopting the new technologies to best complement the physical models could maximize the benefits from both, and possibly produce the next great advance in hurricane prediction.

This work was funded by the NOAA Research Global-Nest Initiative.

Fig. 1 North Atlantic TC track errors. (a) Upper panel shows the mean TC track errors (km) at 12-hr forecast lead time intervals for IFS (black) and GFS (red) during the years of 2014 to 2018. The 95% confidence levels for each model are indicated by the same transparent color shading. Numbers of homogeneous comparison cases for individual lead times are listed in the brackets at the bottom of abscissa. The vertical gray dotted lines indicate 72 and 120 hr forecast lead times. The error differences of GFS compared to IFS are shown in the lower panel. (b) As in (a), but for the years of 2019 to 2023, with the SHiELD forecast in green.

 

Fig. 2 (a) Mean TC track forecast skill relative to the CLIPER5 track forecast error, averaged over the 120-hour forecast for the 7 models during the 2019-2023 Atlantic hurricane seasons. The mean track forecast errors are computed at 6-hr intervals, ranging from 6 to 120 hr, and then the values shown are computed by averaging skill values over those lead times. Colors used for the models are the same as in Figure 1. Total numbers of homogeneous cases are listed in the brackets at the bottom of the abscissa for each hurricane season. (b) As in Fig.2, but for the mean TC intensity forecast skill relative to the SHIFOR5 10-m wind speed forecast error.