The methodology for our study mostly follows that in K13, K07, and Bender et al. (2010), and is described in detail in those studies. Here the methodology is presented only in abbreviated form here, where we focus mainly on aspects of methodology that differ from K13 and K07. The reader is referred to these previous studies for further details.
2.1. Present-day hurricane simulations
The simulation of Atlantic hurricanes proceeds in in several stages. Before we perform climate change simulations, we first want to assess whether our complete modeling system is able to adequately simulate Atlantic hurricane activity for a given set of environmental conditions. To explore this, we first test our system on present-day large-scale climate as defined by the time-evolving NCEP/NCAR Reanalysis I (Kalnay et al. 1996) of the atmosphere and observed sea surface temperature (SST) evolution for 1980–2016. For each year (1980–2016), we simulate the three-month period August-October with the regional 18-km grid Zetac regional model, using the reanalysis to provide lateral boundary conditions, the atmospheric initial conditions, and the time-evolving target fields for interior spectral nudging of the very large scale (zonal and meridional wavenumber 0–2 of the regional domain) atmospheric environment, with a nudging timescale of 12 hours. The simulations were limited to the peak three months of the Atlantic TC season (Aug. – Oct.) to save on computation requirements, which may introduce some uncertainty to our results, particularly if aspects of seasonality, such as the length of the Atlantic TC season, were to change with climate warming. For example, Dwyer et al. (2015) find that in model projections, models that project fewer TCs with climate warming also simulate shorter seasons, and vice versa for models projecting more TCs. This potential limitation of our Aug.-Oct. season approach should be kept in mind in interpreting our results. The model was run over specified time-evolving SSTs. The Zetac regional model is a nonhydrostatic model run here without convective parameterization (see K07 for details).
Tropical storm cases are identified in these three-month simulations using the automated procedure described in K07, including a requirement for warm-core structure, surface wind speeds for the storm feature of at least 17.5 m s-1, and total duration of at least 48 hours (not necessarily consecutive hours) at an intensity of at least 17.5 m s-1. Landfalling TCs for the Contiguous U.S. (CONUS) were defined by the intersection of the surface center of the TC (defined by the minimum in surface pressure) with the coastline (https://www.nhc.noaa.gov/aboutgloss.shtml), where the land region was part of the 48 contiguous states (excluding Alaska and Hawaii). Multiple CONUS landfalls by a single storm were counted as separate landfalls in our statistics.
Each individual tropical storm case from the Zetac regional model is then re-run as an individual 15-day case study using the GFDL Hurricane Model. The GFDL Hurricane Model has been used operationally by the National Weather Service (NWS) since 1995 (Bender et al. 2007; 2010) until it was retired from operations in 2017. The version used in this study was the model used in Bender et al. (2010), which was the version that was operational from 2006 through 2010. The model is a triply nested moveable mesh model designed for hurricane track and intensity prediction, and has grid-spacing of about 8.5 km in the innermost 5ox5o nest. The hurricane model has been coupled to the Princeton Ocean Model since 2001 (Bender et al. 2007) and for the present-day simulations uses a realistic present-day (U.S. Navy GDEM, or Generalized Digital Environmental Model) climatology of ocean subsurface temperature and salinity so that the hurricane model is able to accurately simulate the ocean response to the strong hurricane forcing and generate a realistic “cold wake” as it passes over the ocean (Bender et al. 2007). Following the procedure used operationally (2006–2010) for storm initialization, the ocean model state for each storm case is initialized by running the ocean model for two days using the GDEM climatology values for ocean temperature and salinity with SSTs prescribed. During this first step, a sharpening technique is employed in order to assimilate a reasonable climatologically based Loop Current and Gulf Stream ocean structure (Yablonsky et al. 2015). Next, starting three days prior to the storm case start date, atmospheric forcing is imposed, based on the model’s storm wind field. In this second step, the SSTs as well as ocean temperatures and salinity are allowed to evolve. This procedure is used to initialize a cold wake in the SST and ocean temperature field due to the passage of the storm. A limitation of the version of the operational GFDL Hurricane Modeling System used in this study was that the dynamical ocean domain did not cover the full North Atlantic basin but rather the coupled model used two separate ocean model domains—one covering the central to eastern Atlantic and the other the central to western Atlantic (see Fig. 5 of Bender et al. 2007). Storms traversing the central Atlantic could run off one of these grids and lose their ocean coupling. However since the two ocean grids overlapped when this occurred, we restarted the storm on the second ocean grid before the coupling was lost and continued the simulation with full ocean coupling until landfall. Since this happened only for a small subset of runs, and the transition occurred very far from U.S. landfall, and was treated the same for both present-day and warm climate experiments, we believe that this limitation of our model framework was very unlikely to affect our overall conclusions.
In addition, during the 15-day integrations, the hurricane model atmosphere tended to drift toward the model’s climatological state, which differs from that of the host (Zetac) regional model and NCEP/NCAR Reanalysis. However, the average drift that occurs is found to be similar for present-day climate and future warm climate scenarios, so that any systematic effects of the drift on the storm characteristics should be similar in the present-day and warm climate cases.
Figure 1 compares the tracks and intensities of simulated U.S. landfalling TCs from the Zetac regional model and the GFDL Hurricane Model for the years 1980–2013. The tracks from the higher resolution model resemble those of from the Zetac regional model, although the intensities extend to higher categories (up to category 5) in the hurricane model compared to almost no track segments above category 1 in the Zetac regional model.
One test of the simulation skill of our two-step downscaling model framework is a comparison of the year-to-year variability of August-October observed mean storm counts with that from the modeling framework for different categories of tropical storms and hurricanes (Fig. 2). For example, if the model framework, which was only provided large-scale information about the year-to-year variability of the Atlantic basin SSTs and atmospheric circulation, along with very large-scale interior spectral nudging, is still able to generate useful information about tropical storm, hurricane, and intense hurricane numbers and their year-to-year variation, this increases our confidence that the framework can translate some forms of atmospheric and SST variability into useful information about hurricane activity.
The time series in Fig. 2 show that the model is indeed useful for simulating Atlantic basin hurricane activity given specified large-scale atmospheric, oceanic, and SST conditions. Specifically, with only SSTs, ocean temperature climatology, and time-varying NCEP/NCAR Reanalysis boundary forcing and very large-scale interior domain nudging, the system has the following correlation with observed 37 year time series of Atlantic basin Aug-October storm counts: 1) all TCs (tropical storms and hurricanes): r = 0.77 (explained variance: 59%); 2) Category 1–5 hurricanes: r = 0.68 (explained variance: 46%); 3) major (Category 3–5) hurricanes: r = 0.56 (explained variance: 31%); 4) very intense (Category 4–5) hurricanes: r = 0.31 (explained variance: 10%). Assuming independent years, correlations above 0.329 are significant at the 0.05 level, so for all cases except Category 4–5 hurricanes the results show significant correlation. For the Category 4–5 hurricanes, the results are nearly statistically significant.
We note that there are rising trends evident in many of the time series in Fig. 2. The modeled trends are similar to the observed with the notable exception of basin-wide hurricane frequency (Fig. 2c) where the model has a more rapid rising trend than observed. The cause of these rising trends in observations remains a topic of research, as the time period (1980–2013) is relatively short for detection of greenhouse gas warming influence, and since both internal variability and changes in aerosol forcing are possible contributors to such TC-related trends (e.g., Goldenberg et al. 2001, Mann and Emanuel 2006; Zhang et al. 2013, Dunstone et al. 2013, Vecchi et al. 2017, Yan et al. 2017, Murakami et al. 2020; Bhatia et al. 2018). Our simulations indicate only that changes in the large-scale environment (including SSTs) help to explain the rising trends in observed TC and hurricane frequency but do not elucidate the causes of the environmental changes. However, we expect that the observed Atlantic hurricane frequency trends, and tropical Atlantic SST changes, since 1980 have multiple causes and several studies suggest the TC frequency increases are likely not primarily a response to increasing greenhouse gases alone. Thus, the over-prediction of the observed trend in hurricane frequency (1980–2013) evident in our model framework does not invalidate its potential use for greenhouse gas-driven future warming scenarios.
The model framework’s performance is much less skillful for U.S. landfalling storm counts. None of the simulated time series of storm count are significantly correlated with observed variations: 1) all TCs: r = 0.21 (explained variance: 4%); 2) Category 1–5 hurricanes: r = 0.15 (explained variance: 2%); 3) major (Category 3–5) hurricanes: r=-0.07 (no explained variance).
The above results provide an important caveat to our study. While the two-step model framework is relatively skillful at reproducing the year-to-year variation of basin-wide tropical storm and hurricane counts, this skill does not carry through to U.S. landfalling counts. Thus, while the basin-wide results provide model-based evidence that the year-to-year variability in the basin-wide numbers is not random “weather noise” but rather is controlled to a large extent by large-scale environmental conditions, the U.S. landfalling count variations seem much more difficult to capture using our modeling framework. While we are not aware of many other modeling systems that can successfully simulate U.S. landfalling TC frequency, one exception is the HiFLOR model (Murakami et al. 2016), which does show some skill in predicting seasonal U.S. landfalling TC frequency over the period 1980–2015, suggesting that there are large-scale controls on this metric that are not being well-captured in our two-step model framework. We have chosen to use our modeling framework to explore future U.S. landfalling behavior under global warming to take advantage of its high-resolution for simulating hurricane structure, while keeping in mind the model’s limitations for reproducing observed interannual variations of TC frequency from large-scale interannual variations of environmental conditions.
2.2. Climate Change Downscaling Experiments
Following on the 37 Atlantic hurricane seasons of present-day two-step downscaling simulations (1980–2016), we perform similar experiments, but applying a set of climate change conditions, for 27 of the 37 present-day seasons (1980–2006).
We first create a series of climate change “delta” fields, which we can add to the NCEP/NCAR Reanalysis, to create a series of warm-climate perturbation experiments that use realistic conditions (i.e., the reanalysis) as the baseline case. Specifically, we use we use changes in SST, sea level pressure (SLP), air temperature, relative humidity, and wind velocity to modify the NCEP/NCAR reanalysis fields that are used as boundary forcing and as the nudging target for the interior spectral nudging procedure with the Zetac regional model. As described in K13, for CMIP3 models, these perturbations included an 18-model ensemble average perturbation (models listed in K13), which was the August-October average of 2081–2100 minus 2001–2020 for the Special Report on Emission Scenarios A1B (SRES A1B) scenario. Then for 10 of the 18 CMIP3 models (identified later in this report), we created individual model perturbation fields by computing the linear trend of each model’s data for 2001–2100 then computing the 2081–2100 minus 2001–2020 difference of the data projected onto the linear trend. Our experimental design uses the interannual variability from the NCEP/NCAR Reanalysis, which assumes that the interannual variability of these fields does not change with climate change. Projected changes in interannual variability of these fields are generally assessed as having less confidence than changes in the large-scale time-mean environmental fields, although this adds some additional uncertainty to our TC projections. We constructed two 18-model ensemble mean CMIP5 model warm climate scenarios using the 2016–2035 (early 21st century) or 2081–2100 (late 21st century) period of the CMIP5 RCP4.5 scenario versus the baseline period of 1986–2005 of the CMIP5 historical runs (see K13 for a list of the models). The global temperature difference between present-day and the warm-climate condition was 1.69oC for CMIP3 vs. 1.70oC for the CMIP5 late-21st century case. In sensitivity experiments, we found that hurricane model intensity changes in the warm climate scenarios were relatively insensitive to the small increase in the ocean subsurface vertical temperature gradients associated with the SST warming scenarios (see also Tuleya et al. 2016). Therefore, following Knutson et al. (2013) for the ocean subsurface temperature profiles in the warm climate runs, we used the 18-model average three-dimensional ocean structure change from the CMIP3 models to represent the change in ocean temperature stratification in the warmer climate for all of the hurricane model climate change experiments.