We conduct our simulations using the atmospheric component of MPAS-A, version 5.1 (Skamarock et al., 2012). MPAS-A is a non-hydrostatic global atmosphere-only model that uses unstructured Voronoi meshes (Du et al., 1999) to create variable-resolution grids. This grid structure permits localized areas of high resolution to transition gradually to lower resolutions, thus alleviating the boundary issues associated with sharp transitions between domains in traditional nesting approaches (Park et al., 2014). The focus of the simulations presented here is on Northern Hemisphere phenomena; therefore, we use a variable-resolution mesh with 15 km grid spacing over the Northern Hemisphere, expanding out to 60 km in the Southern Hemisphere (Fig. 1).

The MPAS-A atmospheric physics suite includes a subset of schemes adapted from versions of the Weather Research and Forecasting (WRF) model (Skamarock et al., 2008). Our simulations employ the following physics parameterizations: WRF Single-Moment 6-Class Microphysics (WSM6; as in WRF 3.8.1), Yonsei University (YSU; as in WRF 3.8.1) representation of the planetary boundary layer, Tiedtke (as in WRF 3.3.1) subgrid-scale convective parameterization, Community Atmosphere Model (CAM; as in WRF 3.3.1) shortwave and longwave radiation, and the Noah land surface model (as in WRF 3.3.1) for surface processes. We selected the Tiedtke convective parameterization scheme because it includes convective momentum transport (CMT), which has been shown to be important for reducing model biases in surface winds and TC intensity (Zhang and McFarlane, 1995; Han and Pan, 2006; Hogan and Pauley, 2007; Richter and Rasch, 2008). CMT also improves the representation of features such as the Intertropical Convergence Zone (Zhang and Wang, 2006; Kim et al., 2008). We completed a series of preliminary tests using a quasi-uniform 60 km mesh to further refine our physics choices (not shown).

We selected 10 simulation years with varying phases of ENSO based on the Multivariate ENSO Index (MEI) and the Oceanic Niño Index (ONI) over the TC season (Table 1). These years were also chosen to sample a range of TC activity in the North Atlantic, eastern North Pacific, and western North Pacific basins. We chose to sample phases of ENSO rather than other modes of climate variability due to its strong connection to global TC activity (e.g., Gray, 1984; Chan, 1985; Lander, 1994; Chu and Wang, 1997; Kossin et al., 2010). Each simulation is integrated for 14.5 months, from 1 March of the first year through 14 May of the following year, with the first month of each simulation discarded as spin-up; output is recorded every 6 h.

We used the ECMWF ERA-Interim Reanalysis (ERA-I; European Centre for Medium-Range Weather Forecasts, 2009; Dee et al., 2011) with a spectral T255 resolution (∼0.7∘ horizontal grid spacing) for present-day initial conditions. SST and sea ice fields are updated daily throughout the simulations. The configuration of these surface fields is discussed further in Sect. 2.3. For the future climate simulations, we modify the ERA-I initial and lower boundary conditions by adding monthly-averaged temperature changes derived from a 20-member ensemble of CMIP5 GCMs (Table 2). A similar change was applied to the deep-soil temperature. These temperature changes are calculated by subtracting the 1980–1999 average temperature from the 2080–2099 average temperature following the Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report (AR5) Representative Concentration Pathway (RCP) 8.5 emissions scenario, interpolated to the ERA-I grid and added to the existing temperature data at all atmospheric pressure and soil levels. Geopotential height and specific humidity are adjusted by the model based on the imposed temperature changes; relative humidity is held constant at the initial time in the spin-up run (see the following paragraph). We set carbon dioxide (CO2) concentrations in the future climate simulations to 936 ppm, the level projected by the RCP8.5 emissions scenario for 2100 (Meinshausen et al., 2011). Present-day CO2 concentrations are based on analyzed values set according to the year.

Rather than running the simulations in chronological order, the simulation years are sorted from the strongest La Niña year (i.e., the year with the smallest MEI and most negative ONI; Table 1) to the strongest El Niño year (i.e., the year with the largest MEI and most positive ONI; Table 1). This design aims to minimize model spin-up in response to changes in SST. With the present-day and future initial conditions set, we conduct full simulations for a neutral ENSO year (e.g., 2013) for both present-day and future environments. These single-year simulations are used as spin-up and are therefore excluded from our analysis. While we took this precaution to allow the model atmosphere to come into equilibrium with the imposed warming and adjusted CO2 for the future climate experiment, we repeated this process for the present-day simulation to maintain consistency. We then used the output from 1 March, towards the end of the initial spin-up simulation, to initialize the first simulation year. This method continues for both the present-day and future experiments by using the output from the latter part of one simulation (e.g., 1 March) to initialize the next (Fig. 2). Applying this unique “daisy-chain” technique avoids the need for excessive spin-up times for each year; instead, we discard only the output from the first month, which allows any discontinuities arising from the change in SST to equilibrate.

As with the atmospheric initial conditions, the SST fields used in the simulations are taken primarily from ERA-I. The SSTs in ERA-I have, however, been derived from several different datasets over the years (Dee et al., 2011). For reanalysis times after February 2009, ERA-I surface fields originate from the Operational Sea Surface Temperature and Sea-Ice Analysis (OSTIA; Donlon et al., 2012). To maintain consistency between all simulations, the OSTIA SST, interpolated from its native 0.05∘ horizontal grid spacing to the ERA-I grid, is used for simulation years prior to 2009. Therefore, we effectively use OSTIA SST for all simulations. For present-day soil temperature and moisture, we use the ERA-I fields.

The SSTs for the future climate simulations are altered in the same manner as the initial condition atmospheric and soil temperatures (e.g., Fig. 3a–b). The same technique of adding a GCM delta field onto existing data cannot, however, be used for sea ice. Instead, similar to Mizielinski et al. (2014), monthly-averaged CMIP5 ensemble mean sea ice fields are temporally interpolated to create daily sea ice fields for both present-day (1980–1999) and future (2080–2099 under the RCP8.5 emissions scenario) time periods. An example of these sea ice fields is shown in Fig. 3c–d. We then replaced the analyzed sea ice in the ERA-I with these climatological fields for use in all model simulations. While the climatological present-day sea ice does not entirely match the analyzed field in the ERA-I (e.g., the sea ice edge is much more diffuse), handling the sea ice in this manner ensures that it is plausibly represented in the future climate simulations. The presence of an overly diffuse ice edge could result in unrealistically weak lower tropospheric baroclinicity during warm seasons in these locations.

Our technique for simulating a future climate is similar to the PGW approach in the sense that (1) the analyzed initial and lower boundary conditions are altered by adding projected temperature changes from GCMs to represent future conditions, and (2) analyzed high-resolution SST fields are used to preserve realistic representation of ocean eddies and SST gradients. High-resolution SST is of demonstrated importance for midlatitude cyclone development and other regional climate changes (e.g., Brayshaw et al., 2011; Booth et al., 2012; Kirtman et al., 2012; He and Soden, 2016; Siqueira and Kirtman, 2016). By using a global model, however, one of the main limitations of PGW, the constraint of the lateral boundary conditions, is alleviated. Therefore, our future climate simulations are best described as time-slice experiments with prescribed high-resolution SSTs. The UPSCALE (UK on PRACE: weather-resolving Simulations of Climate for globAL Environmental risk) experiments described by Mizielinski et al. (2014) use a similar time-slice technique for simulating a future climate. By simulating a small ensemble of 26 years, UPSCALE samples a broad range of interannual variability and ENSO states; however, 25 km grid spacing is insufficient for resolving full-strength tropical cyclones (Davis, 2018). Therefore, our simulations complement UPSCALE by offering sufficiently high resolution to better capture the atmospheric mesoscale, specifically tropical cyclones.

We conducted the MPAS-A simulations on the US National Center for Atmospheric Research (NCAR) supercomputer, Cheyenne (Computational and Information Systems Laboratory, 2017). Cheyenne is a 5.34 petaflop SGI ICE XA cluster with 145 152 Intel Xeon processor cores and 313 terabytes (TB) of memory. Each 14.5-month simulation was run on 1152 cores and consumed roughly 92 000 CPU hours, including resources needed for post-processing, leading to a total of ∼1.9 million core hours used for these experiments.

We post-process model output to vertically interpolate fields to selected isobaric levels and horizontally interpolate from the native unstructured mesh to a 0.15∘×0.15∘ latitude–longitude grid. Due to storage constraints, we saved a limited number of variables for the Northern Hemisphere only; however, monthly restart files are archived, enabling replication of a particular period of time or event as needed. The post-processed output occupies approximately 50 TB of storage space for the output for all 20 simulations and is currently stored on Cheyenne's High Performance Storage Space (HPSS) and at North Carolina State University.

For our future climate simulations, we computed temperature delta values using the mean IPCC AR5 RCP8.5 emissions pathway. While other plausible scenarios exist, we selected a high emission pathway to maximize the signal of climate change in our simulations. Using the GCM ensemble mean temperature changes to alter our initial and lower boundary conditions diminishes the considerable amount of variability in the temperature changes projected by individual GCMs. Computing an ensemble mean from a set of simulations using temperature change fields from each GCM is, however, unlikely to produce significantly different results (Hill, 2010; Lackmann, 2015; Marciano, 2014). An alternate strategy to using the GCM ensemble mean SST change would be to apply changes computed as a function of GCM ENSO phase. However, given uncertainties in the GCM SST change fields, we felt that the use of the ensemble mean was best for our initial set of experiments. Likewise, our analysis does not address possible changes in the distribution of ENSO phases.

The adjustment of geopotential height based on the imposed temperature changes for the future climate simulation introduces some degree of imbalance between the model mass and wind fields. In previous studies, we utilized the digital filter initialization (DFI) capability of the WRF model to reduce these imbalances. Since this feature is not available in MPAS-A, we conducted a full 14.5-month spin-up simulation to allow time for the dynamics of the model atmosphere to restore balance. We also maintain constant relative humidity between the present-day and future simulations in the initial conditions. While this assumption may be appropriate over ocean basins, it does not necessarily hold true over land areas (e.g., Sherwood and Fu, 2014). With no constraints on the lateral boundaries, however, this constraint is only applied once, at the beginning of our spin-up simulation, and therefore is not enforced during the subsequent simulations. The relative humidity within our model domain evolves freely through the duration of the simulations; by the end the initial year-long spin-up period, we expect the distribution of water vapor to be fully equilibrated with the simulated future climate.

While our treatment of sea ice in the model allows for a plausible representation of future conditions, we use identical sea ice fields in each member of our present-day simulation set and similarly for the future set. We therefore exclude the effects of interannual variability in sea ice. Several studies have highlighted the connection between sea ice variability and atmospheric circulations in the Northern Hemisphere (e.g., Deser et al., 2000; Overland and Wang, 2010); our intention here, however, is to minimize this influence and instead focus on changes due to altered temperatures and atmospheric composition. Another limitation inherent in our methods is the assumption that future patterns of SST variability will remain similar to what they are today. Nevertheless, we believe the benefits of using high-resolution SST analyses to preserve realistic SST gradients and alleviate regional biases associated with atmosphere–ocean coupling (e.g., He and Soden, 2016) outweigh this limitation.

Many previous studies have shown that neglect of SST cooling due to cyclone passage results in TCs that are too strong (e.g., Schade and Emanuel, 2009). Use of analyzed SST fields in our simulations does not allow for TC-generated cold wakes, which could contribute towards a positive bias in TC intensities and could lead to unrealistic temporal clustering of TCs. The use of convective parameterization, however, particularly the Tiedtke scheme which adjusts momentum, tends to weaken TCs through momentum adjustment in a warm-core cyclonic structure, an effect opposite to that resulting from the neglect of SST cooling. Additionally, the presence of pre-existing cold wakes in the OSTIA SST field could erroneously weaken TCs that occur in their path; however, these pre-existing cold wakes in the OSTIA do not appear to be particularly strong in magnitude (not shown) and therefore are unlikely have a substantial impact on simulated TC strength. Ideally, a grid length of 4 km or less would be used to fully capture TC structure and intensity (e.g., Gentry and Lackmann, 2010), but computational expense does not allow this for the Northern Hemisphere region of interest for the simulation durations necessary to obtain statistically meaningful results regarding the impacts of climate change. A benefit of our configuration is that the resolution is sufficiently high to capture nearly the full range of TC intensity; preliminary testing highlighted the capability of our 15 km grid to replicate realistic TC structures, including spiral rain bands and a defined eye (not shown). We acknowledge that the neglect of sea-surface cooling, the use of parameterized convection, and potential effects of pre-existing cold wakes in the SST analysis data are limitations to our approach. These limitations are, however, consistent between present-day and future simulations, allowing any differences found in TC intensity to remain meaningful (Patricola and Wehner, 2018).

We recognize that the methods employed in this study account only for projected changes due to increased anthropogenic greenhouse gases and therefore do not represent other external climate forcings. Changes in other aspects of the climate system, such as changes in aerosols, deep soil moisture, and vegetation, are not represented. Despite the limitations discussed, our method alleviates limitations associated with regional PGW and coarse GCMs, and is much more computationally efficient than running high-resolution global models for long integration periods (e.g., centuries); the result is a set of controlled simulations suitable for examining the effects of climate change on high-impact weather events.

There are two primary midlatitude storm track regions in the Northern Hemisphere (the North Pacific and the North Atlantic) where baroclinic waves form over regions of enhanced temperature contrast linked to warm western boundary currents off the east coasts of Asia and North America and propagate eastward through downstream development (Chang et al., 2002). The extratropical cyclones in these regions play an essential role in the Earth's climate system and contribute to everyday weather, including high-impact events. Therefore, it is important that they are well represented in model simulations.

As suggested by Chang and Fu (2003), variance in daily-mean fields can be used as proxies for storm track activity. Here, we use the 24 h variance of daily-mean sea-level pressure (SLP), calculated using Eq. (2) from Chang et al. (2013): 1SLPvariance=SLPt+24h-SLP(t)2‾, where the overbar indicates the quantity is averaged over time, in this case over the winter season (December–February; DJF) when storm activity in the Northern Hemisphere is maximized (Chang et al., 2002; Brayshaw et al., 2009). Figure 4 shows the wintertime SLP variance for the MPAS-A simulations compared to the ERA-I; the ERA-I climatology in Fig. 4b is computed using only the 10 years corresponding to our simulations. The North Pacific and North Atlantic storm track regions are clearly evident in the model simulations; the overall spatial correlation coefficient is greater than 0.98, indicating that general patterns of SLP variance are well reproduced in the MPAS-A simulations. As evident from the positive biases at ∼35∘ N, 165∘ W and ∼40∘ N, 30∘ W (Fig. 4c), both storm track regions are shifted equatorward, and the North Pacific storm track is more zonally oriented in the MPAS-A simulations. Comparison with 100 random samples of 10-year means from the ERA-I record indicates that these biases, primarily the shift in the North Pacific, likely represent true differences between the MPAS-A simulations and the real atmosphere. Negative biases in simulated storm activity occur east of Greenland, over Scandinavia, and throughout central North America; these differences, however, fall within the range of observed variability (Fig. 4c).

Corresponding to the Northern Hemisphere extratropical storm tracks are the midlatitude jet features, represented by the wintertime average zonal wind speed in Fig. 5. As in Fig. 4, the ERA-I climatology in Fig. 5b includes only the 10 simulation years. As indicated by a pattern correlation coefficient of ∼0.99, the orientation and spatial extent of the North Pacific and North Atlantic jets at the 250 hPa level are well replicated by MPAS-A (Fig. 5a–b). While the North Atlantic jet maximum is slightly stronger in the MPAS-A simulations, the general strength of both features compares well between the simulations and reanalysis (Fig. 5a–b). Furthermore, examination of a cross-section of zonally averaged zonal wind shows the jet maximizes at roughly the same altitude and latitude (∼200 hPa and ∼30∘ N) in both the MPAS-A simulations and the ERA-I climatology, albeit the MPAS-A maximum is moderately weaker (Fig. 5c–d). Additionally, semi-permanent maritime SLP features, such as the Aleutian Low over the Bering Sea, the North Pacific subtropical high, and the Icelandic Low, are well captured in the MPAS-A simulations (Fig. 5a–b). The Bermuda High in the North Atlantic, however, while evident in the MPAS-A simulations, is comparatively weaker than analyzed.

Average precipitation over the 10 simulation years compared to the 19-year (1998–2016) climatology from the Tropical Rainfall Measurement Mission (TRMM; Huffman et al., 2007; Tropical Rainfall Measuring Mission, 2011) 3B42 product is shown in Fig. 6. Because 4 of our 10 simulation years occur before the TRMM record began, we opted to use the full TRMM climatology for comparison. A pattern correlation coefficient of ∼0.95 indicates that MPAS-A simulates the general spatial pattern of tropical precipitation well; the Intertropical Convergence Zone (ITCZ) in the equatorial Pacific and maxima along the west coast of India, over the Himalayas, and throughout northern South America are all well-represented by the model. The primary difference between the two precipitation fields is the overproduction of precipitation by MPAS-A in many areas (Fig. 6c), an issue common among other high-resolution modeling studies (e.g., Bacmeister et al., 2014; Small et al., 2014). The overestimation in the subtropical Pacific basin (Fig. 6c) is primarily due to overproduction of summer and fall precipitation. The summer season is also responsible for the overproduction of precipitation through the Bay of Bengal and Gulf of Thailand, suggesting an overactive summer monsoon in the MPAS-A simulations. Another notable difference between MPAS-A and TRMM annual average precipitation is the westward shift of the heaviest precipitation along the ITCZ in the Atlantic basin. This shift in precipitation is likely related to a westward shift of the summertime African easterly jet (AEJ; not shown).

We compare simulated TC characteristics to the International Best Track Archive for Climate Stewardship (IBTrACS; Knapp et al., 2010). Only the IBTrACS values for the 10 simulated years (Table 1) are considered for comparison. Consistent with similar studies (e.g., Murakami et al., 2015; Roberts et al., 2015; Yamada et al., 2017), model storms are generally weaker than observed in terms of maximum 10 m wind speed; several simulated storms do, however, attain a minimum SLP of less than 900 hPa (Fig. 7). Therefore, as in Roberts et al. (2015), storm intensity for the simulated TCs is measured by the minimum lifetime SLP of the storm in addition to maximum 10 m wind speed as defined by the Saffir–Simpson scale. Using the minimum SLP, categories are defined as >994 hPa for tropical storms (TSp) and 980–994, 965–979, 945–964, 920–944, and <920 hPa for category (Catp) 1–5 equivalent tropical cyclones, respectively (Roberts et al., 2015). The subscript p is used to discriminate the SLP-based categories from those defined by the Saffir–Simpson wind speed thresholds. For IBTrACS, the maximum 10 m wind speed and minimum SLP across all reporting centers are used as the observed storm intensity for categorization.

Figure 8 shows the average TC frequency over the 10 simulation years for the Northern Hemisphere as a whole, in addition to each basin. Our MPAS-A simulations generate excess TC activity in the Northern Hemisphere, primarily due to the overactivity in the western North Pacific basin. Simulated TC frequencies for the North Atlantic, eastern North Pacific, and Northern Indian basins are within the observed range. Across all basins, when categorizing TCs by minimum SLP (Fig. 8a), MPAS-A generally underestimates the number of weak systems (those with strengths less than Catp1), and overestimates the number of Catp1 and Catp2 storms. The frequency Catp3 TCs and stronger, on the other hand, are simulated reasonably well. With regard to TC categorization by maximum 10 m wind speed (Fig. 8b), MPAS-A simulates the frequency of TS strength TCs quite well in all basins. Strong TCs (Cat4 and Cat5), however, are universally underestimated by the model in favor of Cat1–Cat3 TCs. In the future simulations, there is an increase in TC activity in both the North Atlantic and western North Pacific basins (not shown); further investigation into these future changes in TCs will be the subject of a future paper.

Spatially, the model-simulated TC track density compares reasonably well with observations; the pattern correlation coefficient is about 0.7 (Fig. 9). The most prominent difference is the lack of TC activity in the eastern portion of the North Atlantic basin, which is common among several similar modeling studies (e.g., Bell et al., 2013; Strachan et al., 2013; Small et al., 2014; Roberts et al., 2015). TC genesis in this region typically occurs during August and September (Kossin et al., 2010; Daloz et al., 2015); comparison between the simulated atmosphere and ERA-I monthly-averaged 850–200 hPa vertical wind shear for the 10 simulation years during these months shows a strong positive bias in the model over the North Atlantic development region that is likely a primary factor in this lack of TC generation (not shown). Additionally, the westward shift in the simulated summertime AEJ (not shown) is consistent with a westward shift in the wave accumulation zone, which is likely impacting the location of TC genesis in this area (Done et al., 2011). Figure 9 does not show a strong track density bias in the Gulf of Mexico. Roberts et al. (2015) note that a steady supply of vorticity in the Caribbean contributed to their overestimation of track density in this area; thus, it is possible that tracking TCs as SLP minima, rather than maxima in 850 hPa relative vorticity, helps alleviate this bias. Unlike previous studies (e.g., Small et al., 2014; Murakami et al., 2015; Roberts et al., 2015; Yamada et al., 2017), we do not find a positive track density bias in the central North Pacific; instead, we see a slight underrepresentation of TC activity in that area around ∼150∘ W.

Aside from the underestimation in August and September (likely attributed to the lack of TC genesis in the eastern portion of the basin), MPAS-A simulates the present-day seasonal TC cycle for the North Atlantic reasonably well; TC activity increases during the spring and summer seasons and reaches a maximum in the fall (Fig. 10a). For the eastern North Pacific, MPAS-A produces too many storms in the springtime (April and May) and too few storms during the summer months (Fig. 10b). As defined by Camargo et al. (2008), cluster-2-type eastern Pacific TCs form off the coast of Mexico, travel towards the northwest along the coastline, and have a bimodal seasonal distribution with peaks in late spring/early summer and early fall, similar to the modeled cycle in Fig. 10b. Compared to the ERA-I 10-year climatology, enhanced westerlies at the 500 and 850 hPa levels between 0–20∘ N in the eastern portion of the eastern Pacific basin for April and May (not shown) suggest that our simulations may be in a regime more conducive to these cluster 2 storms. For the western North Pacific (Fig. 10c), MPAS-A correctly simulates the fall peak in TC activity; there is, however, a secondary peak in April that does not match observations. Although there is a general overestimation of storm activity in the northern Indian basin, the model does replicate the shape of the seasonal cycle with both the early summer and mid-fall peaks represented, albeit the fall peak occurs 1 month earlier than observed (Fig. 10d).