Table 1. Development of the four SHiELD configurations and their yearly revisions described in this paper. Timesteps are given in seconds; for nested simulations the format is global/nested timesteps. All configurations and versions use the same Noah LSM and RRTM, and all use SAS or SA-SAS shallow convection except 2017 and 2018 C-SHiELD.
All configurations are initialized using the real-time GFS analyses made available by NCEP following Chen et al. (2018). This “cold starting” from the hydrostatic, spectral GFS could potentially leave the convective-scale configurations (T-SHiELD, C-SHiELD) at a comparative disadvantage to models with native, specialized convective-scale data assimilation. This issue is minimized here due to the ability of FV3-based models to “spin up” their convective scales within a few hours of initialization and experience little degradation thereafter (Hazelton et al. 2018a,b; Marchok et al. 2018; Zhang et al. 2018; Harris et al. 2019).
Computational efficiency is crucial for useful simulation modeling, for both real-time and experimental applications. We present the timings for the most recent iterations of SHiELD in Table 2. The 13-km SHiELD needs only 3096 processor cores to complete one day in under 8.5 minutes, the threshold traditionally used for operational global prediction. The 25-km S-SHiELD completes 1.5 years per day with just over 1700 cores; we are hoping to improve the computational cost as part of further S-SHiELD development. C-SHiELD is necessarily more expensive owing to its nested grid but still completes a five-day simulation in under two hours on less than 3500 cores. T-SHiELD has a nested grid with twice as many columns as C-SHiELD but is only about 30% more expensive.
SHiELD is compiled with mixed-precision arithmetic: the dynamics (and the inlined components of the microphysics) use single-precision arithmetic while the physics uses double-precision. This differs from the practice used for most operational models (GFSv15 excluded) and for GFDL climate models, which use double-precision arithmetic throughout. Tests with the 2016 version of SHiELD had found no detectable difference in skill between predictions using mixed-precision and double-precision arithmetic, while leading to a cost reduction of about 40%.
3.1 SHiELD Medium-Range Weather Prediction
The flagship SHiELD configuration is designed for medium-range prediction with lead times of 24 hours to ten days. The design of SHiELD is similar to the operational GFS: a global c768 grid—a cubed-sphere with each face having 768 x 768 grid cells—with an average grid-cell width of about 13-km. The 2016 and 2017 versions of SHiELD used 63 vertical levels (Figure 2), the same as the hydrostatic GFSv14 but with the uppermost semi-infinite layer removed to permit nonhydrostatic simulation. SHiELD 2017 was then developed by NCEP and partners to become GFSv15 and its GFS Data Assimilation System (GDAS): specific implementation details can be seen at https://www.emc.ncep.noaa.gov/emc/pages/numerical_forecast_systems/gfs/implementations.php. Starting in 2018, SHiELD increased the number of vertical levels to 91, increasing the number of vertical levels below 700 mb from 19 to 23 and decreasing the depth of the lowest model layer from 45 to 33 m.