Plain Language Summary
At many weather forecasting centers where computer weather models are run, different models are run for different applications. However, each separate model multiplies the effort needed to maintain and upgrade each model and makes it difficult to move improvements between models.
We present a new “unified” weather modeling system, SHiELD, able to be configured for a variety of applications. This system uses a powerful computer code, FV3, to compute the fluid motion of the atmosphere at any scale, and also able to zoom in on areas of interest to better “see” severe storms or intense hurricanes. We show how we started from a quickly-assembled model for testing FV3 and then gradually improved the representation of different atmospheric processes and expanded into new uses for the system, including short-range severe thunderstorm prediction, hurricane forecasting, and forecasts out to as long as six weeks. We address some of the challenges that we faced and discuss prospects for future model improvements. Since many of the parts of SHiELD are used by models being developed by the National Weather Service for use by weather forecasters, the advances described here can be rapidly introduced into those models, eventually improving official forecasts.
1 Unified Modeling at GFDL
As computing power increases global atmosphere models are now capable of regular simulation at resolutions that had been the sole domain of regional atmospheric models. The Integrated Forecast System (IFS; ECMWF 2019a,b) of the European Center for Medium-Range Weather Forecasting (ECMWF) runs on a 9-km grid, and the Global Forecast System (GFS; Sela2010) of the US National Centers for Environmental Prediction (NCEP) runs on a 13-km grid. Some IPCC-class climate models now use grids with spacings as fine as 25 km (Chen and Lin 2013; Vecchi et al. 2019; Haarsma et al. 2017). Global atmosphere models lack the lateral boundary errors that contaminate the solutions of regional models after a few days of simulation. They thus allow us to extend mesoscale and storm-scale predictions into the medium range and beyond (Harris and Lin 2013, 2014; Zhou et al. 2019; Harris et al. 2019). Global modeling also brings many new challenges—one cannot “throw your garbage in the neighbor’s yard” in global modeling, so to speak. Biases and radiative imbalances must be minimized, as must errors anywhere in the atmosphere that could potentially grow and contaminate the entire domain.
A unified modeling system supports a variety of applications at a wide range of spatial and temporal scales within a single framework. These systems promise to simplify operational and research modeling suites and better exchange improvements and bug fixes between applications. The Unified Model of the United Kingdom Met Office (UKMO; Brown et al. 2012) is the most notable unified system. Variable-resolution models (Harris and Lin 2014, McGregor 2015) are particularly well-suited for unified modeling as they can efficiently reach very high resolutions over part of the earth, replacing the highest-resolution regional models (Hazelton et al. 2018a,b, Zhou et al. 2019) and potentially extending their lead times.
Here at GFDL a hierarchy of models has been developed for a variety of time and space scales, from centennial-scale earth-system simulations (Dunne et al. 2020) to very high-resolution weather prediction. The GFDL suite is unified around a single dynamical core, the GFDL Finite-Volume Cubed-Sphere Dynamical Core (FV3, or FV3; Putman and Lin 2007), and a single framework, the Flexible Modeling System (FMS; Balaji 2012), and other shared components. We describe one part of this suite, the System for High Resolution Prediction on Earth-to-Local Domains, or SHiELD. This model, previously called fvGFS, was developed as a prototype of the Next-Generation Global Prediction System (NGGPS) of the National Weather Service, and of the broader Unified Forecast System (UFS). SHiELD continues GFDL’s high-resolution global modeling program previously established using the High-Resolution Atmosphere Model (HiRAM; Zhao et al. 2009; Chen and Lin 2013). SHiELD couples the nonhydrostatic FV3 dynamical core (Lin et al. 2017) to a physics suite originally from the GFS (Han et al. 2017, and references therein) and the Noah Land Surface Model (Ek et al. 2002). SHiELD can be used for a variety of timescales but has been designed with a particular focus on short-to-medium range weather (18 hours to 10 days) and into the subseasonal to seasonal (S2S; several weeks to several months) range. Seasonal to decadal predictions and centennial-scale climate projections coupled to a dynamical ocean are performed at GFDL using the Seamless System for Prediction and Earth System Research (SPEAR, Delworth et al. 2020), the Coupled Model version 4 (CM4; Held et al. 2020), and the Earth System Model version 4 (ESM4, Dunne et al. 2020).
Since FV3 is designed to adapt to a variety of purposes and to any scale of atmospheric motion it is an ideal platform for a unified modeling system. All of the SHiELD configurations described here, as well as regional and doubly-periodic applications lying beyond the scope of this paper, use the same code base, the same executable, the same preprocessor, the same runscripts, and same post-processing tools, demonstrating a true unification for modeling on weather-to-S2S timescales. This approach also suggests how further unification with GFDL’s climate models, which use a different atmospheric physics (Zhao et al. 2018), the MOM6 Dynamical Ocean (Adcroft et al. 2019), and the GFDL LM4 land model, may proceed. Advances in SHiELD can be seamlessly moved into other UFS models, including the 2019 upgraded GFSv15, and other FV3-based models. Most notably, advances in SHiELD can migrate into UFS models slated for operational implementation at NCEP, including the FV3-based GFSv15. NASA GEOS (Putman and Suarez 2017), NASA/Harvard GEOS-Chem High-Performance (GHCP), CESM-FV3, and the Chinese Academy of Sciences’ F-GOALS all also use FV3 as their dynamical core and can benefit from the advances described below. This diversity of FV3-based models shows the advantages of using common components to leverage advances in the dynamical core but while still allowing centers to tailor their models to their own needs, the freedom to innovate new model designs, and to encourage the development of models as holistic integrated systems, rather than clumsily joining independent components.
SHiELD is designed for exploratory research into model design and development, with a focus on dynamics and physics-dynamics integration, and for research on prediction and atmospheric processes on timescales from a few hours to a few months. SHiELD is currently focused on deterministic prediction although effective S2S prediction will require the development of a simple ensemble (cf. Chen and Lin 2013). In this manuscript we use forecast skill as a principal means of establishing the scientific credibility of SHiELD as a research tool. Further research will more closely evaluate specific structures and processes within SHiELD, with some initial results described below (especially section 3.2) and in prior research (cf. Hazelton 2018a).
The design, evolution, configurations, and simulation characteristics of SHiELD are the subject of this paper. Section 2 describes the components of SHiELD and how they work together as a complete modeling system. Section 3 describes the four configurations of SHiELD for a variety of applications, including medium-range weather, continental convection, tropical meteorology and hurricanes, and S2S prediction. Section 4 summarizes the history of SHiELD development and discusses prospects for future work.
2 SHiELD Components
2.1 Nonhydrostatic FV3 Dynamical Core
All SHiELD simulations use the nonhydrostatic solver within the FV3 Dynamical Core. This core has been described in detail in other papers (Lin 2004, Putman and Lin 2007, Harris and Lin 2013, and references therein) and will only be summarized here. FV3 solves the fully-compressible Euler equations on the gnomonic cubed-sphere grid and a Lagrangian vertical coordinate. Fast vertically-propagating sound and gravity waves are solved by the semi‐implicit method; otherwise the algorithm is fully explicit. FV3 advances sound and gravity wave processes and advects thermodynamic variables on the shortest “acoustic” timestep, while sub-cycled tracer advection and vertical remapping (cf. Lin 2004) are performed on an intermediate “remapping” timestep, in turn performed multiple times per physics timestep.
FV3’s discretization along Lagrangian surfaces uses the piecewise-parabolic method, which previously used a monotonicity constraint to ensure positivity and to dissipate energy cascading to grid scale. In nonhydrostatic FV3 dynamical quantities (vorticity, potential temperature, and air mass) are advected by a non-monotonic scheme to reduce dissipation of resolved-scale modes. Previous work with nonhydrostatic FV3 had continued to use a monotonic advection scheme to avoid unphysical negative values. In this manuscript we present results using a new positive-definite but non-monotonic scheme to advect tracers, which greatly improves the representation of marginally-resolved and discontinuous features without creating computational noise at sharp gradients. This scheme is described in detail in Appendix A and applications to the representation of tropical cyclones in section 3d.
2.2 GFS/SHiELD Physics and Noah LSM
SHiELD inherits the GFS suite of physical parameterizations developed by the Environmental Modeling Center (EMC) of NCEP (2020) . The initial 2016 version of SHiELD, implemented for dynamical core testing during Phase II of NGGPS, used physics largely identical to the then-operational GFSv13: The Simplified Arakawa-Schubert (SAS) shallow and deep convection schemes described in Han and Pan (2011); the hybrid Eddy-diffusivity Mass-flux (EDMF) scheme (Han et al. 2016); the Rapid Radiative Transfer Model (RRTM; Clough et al. 2005); the microphysics of Zhao and Carr (1997) and cloud-fraction scheme of Xu and Randall (1996); the Navy’s simplified ozone scheme (McCormack et al. 2006); the GFS orographic gravity wave drag and mountain blocking schemes (Alpert 2002); and the convective gravity wave drag scheme of Chun and Baik (1998).
We have since made many changes to the physics to be able to support new applications, especially for convective scale prediction and marine phenomena, or to take advantage of new capabilities within the FV3 dynamical core. We first introduced the six-category GFDL microphysics and cloud fraction scheme (Zhou et al. 2019) with the fast microphysical processes split out of the physics driver and taking place on the shorter remapping timestep. Later, the GFDL microphysics was fully in-lined within FV3 (appendix B). Several new PBL schemes have also been used in SHiELD, including a modified hybrid EDMF PBL as per Zhang et al. (2015), and the Yonsei University scheme (YSU; Hong et al. 2006, Hong 2010, Wilson and Fovell 2018). We have also adopted the Scale-Aware SAS (Han et al. 2017) convection scheme in more recent versions of SHiELD.
The land surface model (LSM) is the Noah land-surface model (Ek et al. 2003), integrated within the physics and paired to the GFS surface-layer scheme. In 2017 Noah was upgraded to use the high-resolution land surface data (Wei et al. 2017), which greatly improves the appearance of land-surface fields in convective-scale simulations.
2.3 Mixed-layer Ocean
Initially sea-surface temperatures (SSTs) were prescribed as the climatological SST plus an SST anomaly from initial conditions which gradually decays to zero, without influence from the atmosphere. However, air-sea interactions are critical for several phenomena of interest to us, especially tropical cyclones and the Madden-Julian Oscillation (MJO) and may impact large-scale skill as well. To incorporate atmosphere-ocean interaction, we have implemented a modification of the mixed-layer ocean (MLO) of Pollard et al. 1973. This simple ocean computes the mixed layer depth and heat within that mixed layer as prognostic variables, with tendencies computed from the net surface heat flux. The SST is nudged towards the NCEP Real-Time Global Sea Surface Temperature (RTGSST; Thiébaux et al. 2003) climatology plus a fixed initial anomaly which decays with a fixed timescale. The ocean mixed layer depth is also nudged toward observed climatology (de Boyer Montégut et al., 2004). While considerably simpler than the three-dimensional dynamical oceans in CM4 (Held et al. 2020) and in the GFDL Hurricane Model (Bender et al. 2019), the MLO still represents the thermodynamic and dynamic ocean interactions of greatest significance on the timescales for which SHiELD is used (Hazelton et al. 2018b), without incurring the complexity of a three-dimensional dynamical ocean.
2.4 Interoperability with other UFS models
SHiELD was designed to work with other models that use FV3, FMS, the GFS Physics Driver, and/or the Interoperable Physics Driver (IPD). The IPD is the interface between FV3 and the GFS Physics Driver, although it can support other physics suites. Innovations within SHiELD can then be seamlessly exchanged with other models using these same components. The UFS Atmosphere led by NCEP (https://github.com/NOAA-EMC/fv3atm/) analogous to SHiELD. For example, the transition of FV3 and the GFDL Microphysics into the operational GFSv15 was accelerated by the IPD. Conversely, schemes which have been introduced into the GFS Physics Driver by the broader community can then be integrated into SHiELD, including the numerous schemes implemented by Zhang et al. (2018).
3 SHiELD Configurations
SHiELD leverages the flexibility of FV3 to be able to make accurate and efficient simulations at a variety of spatial and temporal scales. Much of the development of SHiELD (and previously, of HiRAM) has been driven by a desire to improve the simulation quality at the convection-permitting resolutions covered by the range of SHiELD configurations.
We present four different configurations of SHiELD. All configurations are global domains using either a uniform grid or a locally refined grid using nesting or stretching (Harris and Lin 2013; Harris et al. 2016; Zhou et al. 2019). SHiELD can also run on FV3’s doubly-periodic domain (Held and Zhou, 2006, Arnold and Putman, 2018) or on a regional domain using any regular quadrilateral grid (Dong et al., 2020), at spatial resolutions down to a few tens of meters (Jeevanjee 2017). These applications lie beyond the scope of this paper.
The four configurations can be fit within two “tiers”; Tier-1 configurations are the most well-tested, having originally been developed as prototypes to replace legacy NCEP models by FV3-based UFS systems, and having been run in near-real time for several years. These configurations demonstrate the capabilities of SHiELD, allow direct comparison to existing operational models, and provide robust tests of the forecast skill and reliability of SHiELD. Current real-time configurations are run twice daily and displayed at https://shield.gfdl.noaa.gov/.
The Tier-1 configurations are our flagship 13-km SHiELD, a prototype for the now-operational GFSv15 and for future upgrades of the GFS; (Tropical) T-SHiELD with a static, 3-km nest spanning the tropical North Atlantic, a prototype of the Hurricane Analysis and Forecast System (HAFS); and (Continental) C-SHiELD with a 3-km nest over the contiguous United States (CONUS), a prototype of the Regional Forecast System (RFS). Each of the Tier-1 configurations are usually refreshed every year with a new version, indicated by the year of the upgrade.
Our Tier-2 configurations address new challenges for numerical prediction and are still under development. Our 25-km (Subseasonal) S-SHiELD addresses the challenging domain of S2S prediction. Another configuration not discussed in this paper is the SHiELD global cloud-resolving model (GCRM) and addresses the frontier computational and data challenges of such simulations. This configuration was submitted to the DYnamics of the Atmospheric general circulation Modeled On Non-hydrostatic Domains (DYAMOND) intercomparison (Stevens et al. 2019, Satoh et al. 2019). Both configurations inspire the development of new functionality and capabilities within SHiELD and readily expose instabilities, climate drift, conservation issues, and other shortcomings. The advances driven by work on these frontier challenges help improve the Tier-1 configurations, demonstrating the value of a seamless prediction system. The domains for each of the four configurations plus the GCRM configuration are depicted schematically in Figure 1.