Figure 9 . Hurricane Irma (2017) forecast initialized at 00 UTC 03 September 2017. Left column shows time-series plots of maximum 10-m winds (a), minimum central pressure (b), and RMW (c) compared against extended Best Track observations (Demuth et al. 2006). Right column shows time-radius plots of azimuthally averaged (d) 10-m winds (e) 5-km vertical velocity and (f) precipitation rate from forecasts of Hurricane Irma initialized 03 September 2017, from a prototype of T-SHiELD 2018 with the monotonic (CTRL) and positive-definite tracer advection schemes (PD). The RMW is denoted as a dashed black line. Note that a localized extremum (left panels) may not be visible in the azimuthal averages (right panel), especially during rapid intensification.
A more systematic comparison of wind radii between the 2017 and 2018 T-SHiELD versions (Figure 10, d) shows that the effect of the PD scheme is not limited to a single storm. Noting that the difference between the two T-SHiELD versions is more than just the PD scheme, we do see a systematic and substantial decrease in the radius of the 64-kt (33 m s-1, hurricane force) winds in the 2018 version. The 2018 version spins up the vortex such that within 36 hours of initialization, the 64-kt radii reduce to and then remain a consistent 20-25 nautical miles (nm; 37–46 km) for the rest of the forecast period. This represents a reduction of more than half at 120-h lead time compared to the 2017 version, which steadily widens the 64-kt radii during the simulation. There is also a reduction in radii forecast errors compared to Best Track estimates in T-SHiELD 2018, with the qualification that there is considerable (potentially 40% for 64-kt: Landsea and Franklin 2013) uncertainty in estimates of wind radii. This uncertainty can impact the initialization of tropical cyclones using real-time storm message files (Bender et al. 2017), and thereby of estimates of size-related impacts like precipitation and extreme winds.