Seasonal differences in the response of Arctic cyclones to climate change in CESM1

2.1 CESM1 and the large ensemble

The CESM1 large ensemble (CESM-LE; Kay et al. 2015) is an ensemble of 40 ensemble members designed to investigate the role of internal variability as a source of uncertainty in projection of climate change. Internal variability is known to be a significant source of uncertainty in projections of Arctic climate change (Day et al. 2012; Hodson et al. 2012; Swart et al. 2015; Jahn et al. 2016). All simulations were conducted with the community earth system model, version 1, with the community atmosphere model, version 5 (CESM1-CAM5; Hurrell et al. 2013). It consists of coupled atmosphere, ocean, land, and sea ice components all at approximately 1° horizontal resolution. Treatment of the atmospheric dynamics in this model is similar to the previous version of the model, CAM4, described in Neale et al. (2013).

CAM5 is developed around a finite volume dynamical core, CAM-FV, based on the formulation of Lin (2004). One important aspect for the current study, due to its focus on high latitudes, is the nature of numerical filtering used, to ensure numerical stability, and how this deals with convergent meridians at high latitudes. In order to allow a longer time-step a polar filter is used to stabilise short-wavelength gravity waves, at high latitudes (Neale et al. 2010). This filter closely follows that of Suarez and Takacs (1995), using a fast fourier transform. CAM-FV also includes a fourth order divergence damping term in the momentum calculation, where the damping coefficient is propositional to the cosine of latitude, and therefore reduces to zero at the poles. This arrangement is described in detail by Lauritzen et al. (2012).

In order to generate the large ensemble, initial values from a random year of the 1850 pre-industrial control simulation were used to initialise the first ensemble member. This member was run from 1850 till 2005 following the CMIP5 historical forcing protocol and then from 2005 till 2080 under the Representative Concentration Pathway 8.5 (RCP8.5; see Taylor et al. 2012).

The full state vector of the first ensemble member on Jan 1, 1920 was used to initialize every other ensemble member. The initial values used to initialise each simulation differ only by a small random perturbation (order of 10⁻¹⁴ K) to the initial air temperature field. All other variables were initialised to identical values and the simulations were run from 1920 to 2080 using the forcing already described. Due to the chaotic nature of the climate system, these perturbations grow until initial condition memory is lost creating spread among the ensemble members. Although the exact method of perturbing the initial conditions can affect the spread early in the simulation (Hawkins et al. 2016), a few decades into the simulation the climate state space will be well sampled by the ensemble, since the limit of initial-value predictability for surface ocean and atmospheric variables is about a decade (e.g. Collins et al. 2006; Branstator and Teng 2010). A full description of the CESM-LE can be found in Kay et al. (2015).

The data from the simulations used in the study are the 6-hourly zonal and meridional winds and MSLP, required for the objective cyclone tracking, though these were not output for the entire run due to limitations of long-term storage. Instead, they were output in discrete time windows: 1990–2005 from the historical simulations and 2026–2035 and 2071–2080 during the RCP8.5 projections. As a result, we focus our analysis on these periods which are shorter than the 30-year windows, typically used to define the cyclone climatology in CMIP5 studies (e.g. Zappa et al. 2013a). This means that the contribution of internal variability to the total uncertainty in the projected climate simulated by an individual ensemble member, will be larger than under the CMIP5 protocol, where more years are used to characterise the climate in each period. Nevertheless, because we have so many ensemble members, and thus individual samples with the same model we can accurately characterise the forced response.

2.2 Cyclone tracking

In this study we employ the Hodges (1994, 1996, 1999) objective feature-tracking algorithm to identify and track cyclones. This algorithm has already been used extensively to study extratropical cyclones in both climate models and atmospheric reanalyses (e.g. Hoskins and Hodges 2002; Bengtsson et al. 2006; Zappa et al. 2013a).

The main aspects of this procedure are as follows: firstly, 850 hPa vorticity is computed from the 6-hourly zonal and meridional wind speeds. It is then spectrally filtered to a T42 resolution to reduce the small scale noise and focus on synoptic scale cyclones. The large scale background is also removed for total wavenumbers smaller than six. The spectral coefficients are also tapered to suppress Gibbs oscillations using the method of Sardeshmukh and Hoskins (1984). To identify cyclones the filtered data is first transformed to a uniform grid on a polar stereographic projection of similar resolution to the filtered data in order to prevent the kinds of bias in the identification often seen if using a cylindrical projection, caused by the converging meridians (Sinclair 1997). Cyclones are then identified as maxima in the filtered and projected vorticity that exceed a vorticity of 10⁻⁵ s⁻¹. The locations are then refined using a B-spline interpolation and steepest ascent method to compute the off-grid locations which helps to produce smoother tracks (Hodges 1994). Once the vorticity maxima are identified they are transformed back to the spherical coordinate system for the tracking. Tracks are first initialised using a nearest neighbor method and then refined by minimizing a cost function for the track smoothness, measured as changes in direction and speed, subject to adaptive constraints on track smoothness and displacement distance (Hodges 1999). The tracking is performed in the spherical coordinate system to prevent the types of biases often associated with tracking on a projection. The tracking methodology is exactly the same as that used in Hoskins and Hodges (2002), Bengtsson et al. (2006) and Zappa et al. (2013a).

To avoid the inclusion of stationary and short-lived features, tracks are filtered so that only the cyclone tracks that have a lifetime greater than 2 days and a length of 1000 km or more are retained for the final analysis. The sensitivity of this choice was tested by repeating the analysis but filtering tracks with lifetimes greater than 1 day and track length greater than 500 km and it does not significantly affect the results. These are the values which have typically been used for mid-latitude cyclones (e.g. Hodges et al. 2011) and published case studies of Arctic synoptic scale cyclones indicate that these lifetime and propagation thresholds are suitable choices as Arctic cyclones are typically longer lived than extra-tropical cyclones (Tanaka et al. 2012; Aizawa and Tanaka 2016).

The dynamical intensity of cyclones is measured by referencing the tracks to the full-resolution maximum 850 hPa wind speed by searching for the maximum value within a 6° spherical cap centered on the tracked vorticity maxima. The closest MSLP minima within a 5° spherical cap region are also added to the tracks using a B-spline fit to the MSLP fields and a steepest descent minimization starting from the vorticity center. Spatial maps of track density and of the mean wind speeds associated with cyclones are then computed using the spherical kernel method described by Hodges (1996).

3.1 Model biases

Before investigating the response of Arctic cyclones in CESM1 to climate change it is informative to investigate and describe the ability of CESM1 to recreate observed track density and dynamic intensity. To do this storm track statistics in the historical simulations (1990–2005) are compared with the ERA-Interim reanalysis (Dee et al. 2011) for the same period. However, it should be noted that reanalyses can differ with respect to the exact intensity and frequency of Arctic cyclones (Tilinina et al. 2014).

The model performs well in capturing the basic characteristics of the observed seasonal cycle of Arctic cyclone number and maximum wind speed within the Arctic region (north of 67.5°) as shown in Fig. 2 and Table 1 respectively, with cyclones being dynamically more intense and more numerous in winter than in summer in both the CESM1 and ERA-Interim. Biases in cyclone frequency are of magnitude less than 6% in all seasons. However, the area averaged picture masks significant regional biases.

Open image in new window

In winter the North Atlantic storm track is too zonal in CESM1, extending too far south into central Europe, with too few cyclones making their way into the GIN and Barents Sea region (see Fig. 1). This is a bias common to many climate models (Zappa et al. 2013b). Too few cyclogenesis events occur over northern Eurasia in CESM1, leading to a negative bias in track density in this region and over the eastern Arctic Ocean (see Figures S1, S2).

Biases in the summer track density in CESM1 are more consistent in sign than those in winter, with significant negative biases in track density across most of the Arctic and North Atlantic. These biases are partly caused by a lack of cyclogenesis over northern Eurasia and also a lack of cyclones entering the Arctic from the North Atlantic (Figure S1 & S2). Biases in track density are largest over the coastal region north of Scandinavia and along the Northwest Russian coastline. This results in a fairly weak Arctic Ocean cyclone maximum (Figure S3). This general low track density bias across the northern hemisphere is a common property of CMIP5 models, and may be related low levels of baroclinicity and weak polar front jet over Eurasia in models (Lee 2014).

The large ensemble mean shows biases in maximum wind speed, consistent across most ensemble members, in all seasons. However, regionally they are within 6% of the ERA-Interim values with DJF and MAM cyclones being too strong in the model but JJA and SON cyclones being too weak (Table 1).

3.2 Mean cyclone response

In this section the change in the cyclone statistics between the historical and RCP8.5 periods will be expressed as a percentage. This is because biases in the model mean that the absolute values are somewhat arbitrary.

Between the two periods there is a general reduction in track density (Fig. 2) and in the frequency of cyclones (Table 2) in the Arctic region (north of 67.5°N) in all seasons, however the only statistically significant change occurs in DJF, where the frequency decreases by more than 5%. There is some intra-ensemble variability in the magnitude of the change purely because each ensemble member has a slightly different evolution due to internal variability, however in DJF 90% of ensemble members agree on the sign of this change (see Fig. 2).

The winter response in track density is largest and most statistically significant in the Atlantic sector of the Arctic, with more than 80% of ensemble members agreeing on the sign of the response in this region (see Fig. 3). There is a concurrent increase in track density on the equatorward flank of the North Atlantic storm track, suggesting that these changes in the Arctic are the result of the tilt of the North Atlantic storm track becoming more zonal, which has been seen in future projections with other CMIP5 models (Zappa et al. 2013a). This goes hand-in-hand with a reduction in cyclogenesis to the east of Greenland (Fig. 3e).

Open image in new window

A significant reduction in winter cyclone intensity within the Arctic region is simulated by all ensemble members (Table 1) and this reduction extends across the Arctic basin, but is largest in a band extending from the GIN sea to the north pole, where the reduction in mean maximum wind speed is over 1.5 ms⁻¹ (Fig. 3c).

In summer, only two high latitude regions experience significant changes in track density, the region from northern Scandinavia eastward to the East Siberian Sea experiences a significant reduction in track density and the area south of Greenland experiences a significant increase. Elsewhere in the Arctic, all anomalies in track density are insignificant (see Fig. 3b).

There are also significant changes to cyclone intensity in summer within the Arctic. In contrast to the other seasons, which all experience a decrease in cyclone intensity within the AOCM region (70–87.5°N, 120–240°E), the ensemble mean increases by 2.3% (see Table 2; Figure S4) with similar increases over Alaska and the Canadian Archipelago as well. On the other hand, much of Northern Eurasia experiences a significant reduction in intensity (Fig. 3d).

3.3 Relationship to changes in the background climate

Amplified surface warming in the Arctic, compared to lower latitudes, can be seen in the simulations (Fig. 4). This effect is a well-known feature of global warming (e.g. Manabe and Stouffer 1980) and a common feature in climate model projections (Holland and Bitz 2003). In summer, increased radiative heat flux at the surface is associated with increased sea ice melt and heating of the upper ocean that is enhanced by the ice-albedo feedback. As such, the surface air temperature shows less increase during these months as compared to other seasons. Outside of the summer months this heat is released back to the atmosphere leading to enhanced warming in winter compared to summer and amplified warming in the lower troposphere (Serreze et al. 2009; Screen and Simmonds 2010). It can be seen that the warming in these CESM-LE simulations is also associated with dramatic sea ice loss (Figure S5).

Open image in new window

A feature of the warming that has been less discussed is the fact that in summer the region of strongest warming is not over the Arctic Ocean, but over land. This is caused by two factors, firstly the reduction in snowcover leading to a snow-albedo feedback over land (e.g. Hall and Qu 2006) and secondly due to the fact that the land has a lower heat capacity than the open ocean. This differential warming leads to an increase in the strength of the equatorward meridional 2 m temperature gradient, dSAT/dy, across the Arctic coastline, which is a region of strong temperature gradients in summer months, and is consequently referred to as the Arctic frontal zone (AFZ, see Fig. 4d). The results of Crawford and Serreze (2016) suggest that stronger temperature gradients in the AFZ enhance the dynamical intensity of cyclones passing through this region, so one might expect the simulated increase in this gradient in the RCP8.5 simulations to lead to an increase in cyclone intensities in the AOCM region (as seen in Fig. 3) compared to the historical period. Surprisingly, analysis of the smoothed vorticity tendency based on the 6-hourly timesteps did not reveal a significant increase in the growth of cyclones in this region in the RCP8.5 climate, but the maximum wind speed does show an increase (not shown). This suggests that changes in the large-scale background circulation are playing a more important role than the lower tropospheric baroclinic instability.

During summer, changes in the large-scale circulation are also observed. Figure 4f shows a significant increase in 250 hPa zonal wind, also present in the lower troposphere (not shown). This is particularly prominent in the AOCM and over North Alaska, where the largest increases in Arctic cyclone intensity are simulated. This region also experiences an increase in wind shear in the lower troposphere which is also likely to influence the dynamical intensity. Both Orsolini and Sorteberg (2009) and Nishii et al. (2014), suggest that the increase in zonal wind in climate models is a thermally driven response to the change in temperature gradient.

Large changes in surface temperature gradients are also seen in winter months (see Fig. 4c). There are large reductions in dSAT/dy in the North Atlantic and North Pacific associated with a northward shift of the sea ice edge as well as an amplified warming of the sea ice surface. These regions are collocated with regions of reduced 850 hPa zonal wind and reduced wind shear. These reductions in baroclinic conditions are in regions of cyclone development and dynamical enhancement (e.g. Klein and Heinemann 2002) and lead to a significant reduction in cyclogenesis in the GIN and Barents Sea as well as reduction in the strength of cyclones in the Arctic, downstream of these regions (see Fig. 3c). Although the major source regions for Arctic cyclones in winter are the North Atlantic and north Pacific (Zhang et al. 2004; Sorteberg and Walsh 2008), which experience reduced cyclone activity; local changes in the Arctic also play a role; analysis of relative vorticity tendencies suggest that the growth rates are lower and decay rates are higher within the Arctic basin. Changes in the mean large-scale circulation in the CESM-LE are consistent with those presented by Gervais et al. (2016).

3.4 Causes of ensemble variance

The multi-model study into Arctic cyclones by Nishii et al. (2014) explored the relationship between storminess in the AOCM region and changes in background climate, such as dSAT/dy and U850 across the CMIP3 and CMIP5 ensembles. If inter-model differences in the response of the background circulation are significantly correlated with inter-model differences in the response of storminess it provides some confidence that the co-existence of anomalies in these properties in the multi-model mean response to external forcing are actually causally related. Here we explore these relationships across the CESM-LE, but with a key difference. In the analysis of Nishii et al. (2014) the intra-ensemble variance is due to both inter-model differences in the forced response to climate change and internal variability, whereas in the CESM-LE differences between ensemble members are only due to internal variability.

Intra-ensemble variations in the response of cyclone statistics in the regions we describe above are indeed significantly correlated with the large-scale climate response. During JJA both the change in the average number of cyclones passing through the AOCM region, in a given ensemble member, and their maximum intensity within the region is significantly correlated with the mean 850 hPa zonal wind (see Fig. 5a, b). Further, variations in JJA U850 are also significantly correlated with changes in the surface temperature gradient within the AOCM (see Fig. 5c), but variability in AOCM cyclone frequency and intensity is not (not shown). In contrast to Nishii et al.’s analysis of CMIP5 inter-model variability, we do not find a significant correlation between the Eurasian Arctic frontal zone (EAFZ; 60–180°E, 65–75°N; see Fig. 4d) and either cyclone intensity (Fig. 4d) or U850 (not shown) in the AOCM region. This is consistent with the findings of the previous section and indicates that although, as shown by Nishii, inter-model spread in the strength of temperature gradients in the AFZ is an important factor in the strength of response of a given model, it may not be as important for the internal variability. Temperature gradients along the North American coastline in the American Arctic frontal zone (AAFZ; 180–260°E, 65–75°N) are significantly correlated with the U850 in the AOCM region, but are still not correlated with either cyclone intensity of frequency there (see Fig. 5e, f) in this model.

Open image in new window

It is easier to appreciate these relationships in the form of spatial maps of correlation coefficients. From Fig. 6 it is clear that intra-ensemble variability in both cyclone frequency and intensity within the AOCM is associated with the large scale atmospheric variability, which is Arctic oscillation-like in its structure. Ensemble variability in AOCM storminess between the present-day and end of century is negatively correlated with MSLP in the Arctic basin and positively correlated with the associated zonal winds around this region. There does not seem to be a significant correlation with dSAT/dy strength in the Arctic coastal regions.

Open image in new window

A similar AO-like pattern seems to explain much of the variability in the response of storminess in winter as well (Fig. 7). However, the link with surface temperature is much clearer in winter, with stormier conditions tending to go hand-in-hand with warmer temperatures in the north Pacific.

Open image in new window

3.5 The response of intense cyclones

In order to describe the response of cyclones associated with strong wind speeds, we consider the frequency distribution of dynamical intensity, measured by the maximum 850 hPa wind speed in different high latitude regions. These show the maximum wind speed obtained by storm tracks within each of these regions.

In DJF, there is a clear reduction in the frequency of strong events (greater than the 90th percentile, as calculated from all historical runs, for a given season) for all regions considered (Fig. 8). In the Arctic region, the ensemble mean decrease in frequency of occurrence is more than 36%, with all ensemble members agreeing on the sign of the change (Table 3). These winter changes are largest in the AOCM region but can also be seen in the GIN and Barents Sea region as well. This signal can also be seen in the filtered vorticity distribution, which shows a similar reduction in the occurrence of the strongest vorticity events, indicating that these changes are related to the dynamical strength of the cyclones themselves, rather than due to changes in background circulation at 850 hPa (Fig. S6).

Open image in new window

The picture is more complex in other seasons. In JJA, within the AOCM region there is an increase in the frequency of strong events (Fig. 8f). Their occurrence significantly increases, by more than 19%. Whereas in SON their frequency significantly reduces, with no significant change in MAM (Table 3).

Although the reduction in the frequency of strong events in the GINB region is strongest in winter (25%), significant reductions are also seen in MAM and SON, but with no significant change in summer.