Evaluation of track and intensity forecasts shows that the model had similar track forecast skill to the operational GFS, as well as the global fvGFS model run at GFDL. The intensity forecasts showed skill relative to the global fvGFS, although they were slightly worse in general than the operational HWRF model. A persistent low bias at early lead times in both versions of the model was due to the fact that the model spun up from GFS initial conditions, which led to difficulty in initialization of the extreme winds seen in cases like Irma and Maria. This indicates a need for vortex initialization in future versions of the model.
The model structure is evaluated through comparison with observational platforms including NEXRAD radar for landfalling storms including Harvey, Irma, and Maria as well as NOAA P-3 radar data for Harvey, Irma, Jose, and Maria. The results indicate that the model is capable of producing realistic inner-core structure in many cases, including the secondary eyewall that was observed in Harvey. In addition, the model’s precipitation forecast for Harvey correctly showed the extreme values near Houston associated with the initial rain bands and subsequent stalling of the TC. Looking more closely at structure in several of the cases, the model tends to have too large an RMW compared with observations, however. This may be due to a combination of dynamics, boundary-layer physics, and horizontal resolution, and motivates upgrades to the model that will likely improve both structure and intensity prediction.