Early-winter North Atlantic low-level jet latitude biases in climate models: implications for simulated regional atmosphere-ocean linkages

The main atmospheric reanalysis dataset used here for evaluating CMIP models is the NOAA-CIRES-DOE Twentieth Century Reanalysis (20CR) Version 3 (20CRv3) (Slivinski et al 2019), which is an update on the previous version 2c. Version 3 comprises an increased number of ensemble members (80 from 56) and corrects issues with version 2c such as inaccurate estimates of uncertainty and a bias in sea-level pressure globally in the mid-19th century. Due to remaining uncertainty over corrections of biased ship observations in this period (Slivinski et al 2019), only data from 1861 were included here. All 80 ensemble members were used here to provide a measure of reanalysis uncertainty. SST in 20CRv3 is derived from two different datasets, SODAsi version 3 (Giese et al 2016) before 1981 and HadISST2.2 (Titchner and Rayner 2014) for 1981 onwards. Monthly mean fields of the following variables were used: zonal wind on 850 hPa, atmospheric pressure at mean sea level, surface (skin) temperature and total turbulent HF (the sum of sensible and latent components). Note that over open ocean, surface temperature is defined as SST. For HF, positive values denote fluxes from the surface upward into the atmosphere. For calculating ocean surface areal mean indices, a land mask was used to mask out land areas. To check robustness of jet diagnostics across different reanalysis datasets, the European Centre for Medium Range Weather Forecasts (ECMWF) ERA-20C reanalysis (Poli et al 2016) was also used. ERA-20C spans the period 1900–2010, therefore its use was restricted to robustness assessments. To assess robustness of 20CRv3-derived meridional temperature gradients, the fully comprehensive ECMWF ERA5 dataset (Hersbach et al 2020) and its preliminary extension to 1950 (Bell et al 2021) were also used since ERA5 directly assimilates observations of temperature. For the meridional temperature gradient diagnostics temperature at 850 hPa (TA850) and 2 m (TAS) were assessed.

The climate (2012) model datasets used are the World Climate Research Programme's CMIP5 (Taylor et al 2012) and CMIP6 (Eyring et al 2016) datasets. The combined CMIP5 and CMIP6 datasets will be referred to hereinafter simply as 'CMIP'. The analysis is based on output form the full complexity 'historical' simulations, which are run using known major climate forcings over the period from the mid 19th century to the present. For a given model, historical simulations are in general run multiple times and these 'realizations' are important for assessing the role of internal climate variability in the historical period. All available models and realizations with the required data are included, as listed in tables 1 (CMIP5) and 2 (CMIP6). The atmospheric fields used are monthly mean zonal wind and temperature on 850 hPa (CMIP variable names 'ua' and 'ta'), atmospheric pressure at mean sea level ('psl') and surface-air (2 m) temperature ('tas'). To mirror the approach used for assessing 20CRv3, monthly mean skin temperature ('ts') was used for SST diagnostics and the sum of sensible and latent components ('hfss' and 'hfls') used for HF. Monthly data spanning years 1861–2005 were extracted, since this is the period of overlapping availability across CMIP historical simulations and 20CRv3. Daily data for the shorter 1950–2005 period were also extracted for more detailed daily jet latitude diagnostics, due to the lack of pre-1950 availability in many CMIP ensemble members, particularly in CMIP5. Land area fraction ('sftlf') was used to mask out land areas in the ocean surface diagnostics.

The diagnostics used to assess jet speed and latitude are based on the approach used by Bracegirdle et al (2018). This draws from the definition of Woollings et al (2015), but uses monthly mean data rather than daily data in order to maximise the number of models and realizations available for inclusion. In this method, for each monthly mean field the maximum in the latitudinal profile of zonally-averaged zonal wind at 850 hPa is identified. The zonal averaging is defined over the longitude range 60° W—0° and the value of the maximum defines the jet speed index (JSI) and the latitude of this maximum defines the jet latitude index. Seasonal means are assembled from jet diagnostics calculated from monthly mean fields. A comparison with results based on daily fields in Bracegirdle et al (2018) showed that annual to multi-decadal variability in the monthly-derived indices is a close analogue to the daily-derived version. As part of diagnosing CMIP jet latitude biases in more detail, jet latitude diagnostics using daily data following the method of Woollings et al (2010) were calculated for the period 1950–2005.

Area-weighted spatial means of sea surface parameters, surface temperature and HF, are defined over the SPG region (45° N–65° N, 60° W–20° W) denoted SST_SPG and HF_SPG. This box broadly spans the spatial extent of the SPG as identified by Biri and Klein (2019). The southern boundary of the box was chosen to account for annual-decadal variability in SPG location but also balanced against being too far south into other key regions of the NA, in particular the Gulf Stream. Grid boxes with more than 10% land were masked out. Note that since the SST_SPG diagnostic is calculated from surface temperature, it will in general include a small number of sea ice surface temperature grid points at the very north of the domain.

Atmosphere-ocean linkages are assessed here using correlation coefficients between time series of monthly-mean ocean surface diagnostics (SST_SPG and HF_SPG) and JSI. Time series correlations are presented as r_ts = cor(JSI, sfc), where 'sfc' is either SST_SPG or HF_SPG. For example, correlations between JSI and SST_SPG for January are calculated from year-to-year time series of January-mean values. All time series are linearly de-trended before calculation of correlations. Although time series correlation is indicative of coupling, it is recognised that it does not quantify atmosphere-ocean interactions in a strict sense.

Correlation is also used to quantify linear associations in CMIP cross-model scatter plots between two variables, denoted r_cm. For example, cross-model relationships between climatological jet latitude and time series correlation diagnostics (r_ts) are examined to help establish whether atmosphere-ocean linkages are weaker or stronger in models with more equatorward climatological mean jet latitudes.

In this section we assess whether the systematic early-winter equatorward bias in jet latitude that exists in previous model generations is also present in CMIP6. Figure 1 shows October to March climatological mean jet latitude. It is evident that the early-winter equatorward bias does indeed persist in CMIP6, which is reflected in spatial maps of climatological low-level winds (figure S1 available online at stacks.iop.org/ERL/16/014025/mmedia). A shorter 1900–2005 period is shown in figure 1(b) to allow the inclusion of ERA20C and demonstrates agreement with 20CRv3. The reanalysis-derived climatological jet latitude exhibits an overall equatorward shift from ∼50° N in October to ∼46° N in March. In comparison, both CMIP ensembles exhibit a more rapid equatorward shift of the jet from October to November. As a result almost all CMIP model realizations exhibit equatorward climatological jet latitude biases in early winter, with ensemble mean biases of 3.0° (November) and 3.0° (December) for CMIP5 and 2.5° and 2.2° for CMIP6. There is therefore an overall improvement in CMIP6 compared to CMIP5, with proportional reductions in ensemble mean equatorward jet latitude bias of 18% and 26% in November and December respectively. In late winter the ensemble mean biases are smaller, but with a larger inter-model spread.

Zoom In Zoom Out Reset image size

To evaluate the early-winter jet latitude bias in more detail, the frequency distribution of daily jet latitude anomalies from ten day low-pass filtered daily data for November-December is shown in figure 2. Figure 2(a) shows the high latitude peak (at ∼58°–60° N) dominating in 20CRv3. At lower latitudes there are indications of the central and southern peaks of the tri-modal structure that is prominent in winter months (DJF) (Woollings et al 2010) (figure 2(b)), but this is less distinct in early winter. It is also apparent from figure 2 that in general this high latitude peak is not reproduced in CMIP historical simulations and that they exhibit an overly dominant central peak at ∼45° N. With regard to interactions with the SPG, this is highly relevant since the position of the northern peak (∼58°–60° N) coincides with latitudes of the SPG (45°–65° N). Consistent with the results for the mean jet latitude (figure 1), there is some improvement in CMIP6 compared to CMIP5 in terms of the relative strengths of the central and northern preferred jet locations. Note that a version of figure 1(b) for the shorter 1950–2005 period used in figure 2 is shown in supplemental material (figure S2) and reproduces the pronounced early-winter equatorward bias in jet latitude.

Zoom In Zoom Out Reset image size

Thus, despite improvements in CMIP6, the majority of CMIP6 models still exhibit an equatorward bias in early winter. The main problem seems to be in representing the high-latitude peak in jet occupancy (figure 2). As a starting point to explaining this early winter bias, we assess whether there are biases in climatological upstream lower-tropospheric meridional temperature gradient simulated by the models (also referred to here as baroclinicity) over the North American continent. This region has been identified as a key region for explaining seasonal climatological NA jet shifts (Hoskins and Hodges 2019a).

Figure 3 shows reanalysis-derived climatological 850 hPa temperature (figures 3(a), (c), (e) and its meridional gradient (figures 3(b), (d), (f)) for early, middle and late months of the extended winter season. A distinct region of strong meridional temperature gradient is evident at mid-latitudes over the eastern North American continent and this extends out over Gulf Stream region of the western NA and north across Newfoundland to the southern border of Labrador.

Zoom In Zoom Out Reset image size

We examined the baroclinic zone upstream of the NA jet, over eastern North America, and found that early-winter biases in meridional temperature gradient are largest on its poleward side, shown by box A in figure 3 (90°–70° W, 45°–55° N). Two key reasons that one might expect the region on the poleward flank of the baroclinic zone to be important for NA jet latitude are: (a) theoretical/idealised studies have shown that shifts in jet latitude are more strongly associated with shifts in the region of maximum baroclinicity than strengthening/weakening of an existing maximum (Baker et al 2017) and (b), consistent with this, (Hoskins and Hodges 2019b) found that summer baroclinicity is centred on this region along with a more poleward position of the NA storm track during summer months (see also figure S3). Here, we examine whether meridional temperature biases in this region in CMIP models might help to explain the early-winter latitude biases of the NA jet further downstream. For reference the box used define the SPG region is also shown in figure 3.

Figure 4 shows that both CMIP5 and CMIP6 models systematically exhibit too weak 850 hPa meridional temperature gradients in Region A that are most pronounced in early winter (November and December), though the biases are on average slightly smaller in CMIP6. The ensemble mean model biases gradually reduce to small values in March. This early-to-late winter evolution closely reflects the evolution in latitude biases of the NA jet further downstream (figure 1).

Zoom In Zoom Out Reset image size

The association across the CMIP models between meridional temperature biases in Region A and jet biases is investigated in figure 4(b). This shows only the models with the ten most poleward and ten most equatorward jet latitudes and demonstrates a clear pattern whereby models with stronger climatological baroclinicity (more negative gradients) in region A exhibit more poleward jet positions and vice versa.

Although these upstream biases in low-level baroclinicity are indicative of a causal link, more detailed investigations such as sensitivity experiments would be required make definitive conclusions. However, one key driver of baroclinicity is surface air temperature (TAS) gradients, which are influenced by surface processes and features including snow cover. Figures 4(c) and (d) confirm significant model biases in surface temperature gradients, with a similar winter evolution in meridional gradient bias to that seen at 850 hPa. Reanalysis-derived gradients are more clearly outside the range of the CMIP models in November and December (figure 4(c)). Since 20CRv3 does not assimilate temperature observations, results are also shown in figure 4(c) for the fully-comprehensive ERA5 reanalysis. There is strong agreement between ERA5 and 20CRv3, although ERA5 produces slightly stronger meridional gradients in TAS (figures 4(c) and (d)).

Overall, the more prominent biases in higher-latitude aspects of the NA jet further motivate a focus on the SPG region in terms of the representation of NA ocean-atmosphere interactions.

In this section the impacts of the early winter bias in jet latitude on jet-SST linkages over the SPG are assessed. In particular, we focus the question of whether models with mean jet positions further equatorward exhibit weaker jet-SST correlations than those with more poleward jet positions.

To set a baseline for comparison with CMIP data, figure 5 shows reanalysis-derived spatial maps of correlations between year-to-year time series of JSI and ocean surface variables at each grid point for each month of the extended winter from November through March. This shows that the previously-documented significant correlations between JSI and surface parameters over the SPG region (e.g. Woollings et al 2015, Ma et al 2020) extend to the early winter months in which climatological jet latitude biases are most pronounced. Correlations based on SST anomalies from one month later (e.g. time series of January JSI and February SST) show qualitatively similar results (figure S4). This implies that the correlations in year-to-year variability over the SPG region are dominated by the ocean responding to atmospheric variability. The heat loss and cooling over the SPG under strong jet conditions are associated with strong winds and cold air advection over central-eastern part of the climatological jet (figure 5), with the region of cold-air advection extending to the seas around Greenland (for details see Ma et al (2020)).

Zoom In Zoom Out Reset image size

Equivalent calculations were conducted for each available realization of historical CMIP model simulations, with results for November averaged across all models shown in figure 6 and December in figure 7. Focussing initially on SST over the SPG region, it is evident from figures 6(b) and 7(b) that the main region of negative correlations is located towards the southern boundary of the SPG box in the CMIP ensemble mean. In contrast, the reanalysis correlations are confined to the centre of the box and extend further north into the western Greenland Sea and Labrador Sea.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Selecting subsets of CMIP models with high and low latitude climatological jets demonstrates a link between climatological jet bias and the position of the spatial correlations (figures 6(c), (d) and 7(c), (d)). Models with the ten most equatorward jets (figures 6(d) and 7(d)) show negative correlations shifted further equatorward with the jet and vice versa for those with the ten most poleward jets. Correlations with one-month lagged SST show qualitatively similar results, but more pronounced in association with larger correlation magnitudes over the SPG region (see figures S4–S6). Indeed, the CMIP models with more poleward jets exhibit a closer similarity to reanalysis output, with the main area of negative correlations located centrally within SPG box and regions of negative correlation to the west of Greenland (figures 6(c) and 7(c)). Taking a broader perspective, it is perhaps unsurprising that the NA correlation pattern is in general located further south in the subset of models with more equatorward climatological jet latitudes. Correlations with one-month lagged SST again show qualitatively similar results, but with larger correlation magnitudes over the SPG region (see figures S4 and S5).

With regard to time series correlations between JSI and turbulent HF, the results are qualitatively similar (lower rows of figures 5–7). In particular, models with more equatorward jets have JSI-HF correlation patterns that are further south in the SPG, and in the models with more poleward (and less biased) jets these patterns look more like the reanalysis. This qualitative consistency, along with the stronger correlations for HF, provides further evidence for the extent of impacts of climatological jet latitude biases on atmosphere-ocean interaction over the SPG and more widely across the NA. Since we find that the CMIP6 models exhibit some improvement over CMIP5 in terms of early-winter jet latitude bias, the results in figures 6 and 7 suggest that this may have translated into a small improvement in representation of JSI-ocean linkages over the SPG.

To explore the relationship between climatological jet latitude and JSI-SPG time series correlations in more detail, scatter plots are shown in figure 8. These demonstrate that, compared to reanalysis data, most CMIP models exhibit weaker JSI-SPG time series correlations associated with the systematic early-winter equatorward jet biases. There is no clear difference in behaviour between the CMIP5 and CMIP6 models, although a larger proportion of CMIP5 models occupy the low-latitude/weak correlation quadrant of the scatter plots.

Zoom In Zoom Out Reset image size

Potential impacts of sampling uncertainty on the above results are evident in figures 8(b) and (d). Here the within-model spread (i.e. the spread across realizations for a specific model) in output from models with a large number of historical realizations shows that there is a non-negligible sampling uncertainty for individual realizations. This is mainly related to uncertainty in the time-series correlations (along the y-axes), with much smaller within-model spreads for jet latitude (along the x-axes). The implication is that associations between jet latitude bias and JSI-SPG time series correlation strength are potentially weaker than they would be if a large number of realizations were available from each model (i.e. reduced sampling uncertainty).

One way to reduce the impacts of sampling uncertainty is to consider the full winter season. Therefore figure 8 was repeated for extended winter (Nov–Mar) diagnostics (figure 9). The broad picture reflects that seen in early winter, with more poleward climatological winter jets generally exhibiting stronger time series correlations between SPG surface variables and JSI. The overall linear cross-model association is clearer than for early winter (r_cm = −0.70 and 0.80 for SST and HF respectively), which is consistent with the smaller sampling uncertainty of extended winter mean diagnostics (figures 9(b) and (d)). However, the cross-model relationships (r_cm) appear non-linear, with stronger sensitivities to latitude in the lower half of the latitude range. This is likely due in part to the regions of stronger JSI-surface correlations extending out of the southern boundary of SPG box in models with more equatorward jet latitudes (e.g. see the early-winter spatial maps in figures 6 and 7). The above scatter plots were repeated with one-month lagged SSTs (e.g. Nov–Dec JSI and Dec–Jan SST), which again show qualitatively similar results but with larger sampling uncertainty and slightly weaker cross-model relationships (figures S7 and S8).

Zoom In Zoom Out Reset image size

In addition to considering sampling uncertainty, the 80 ensemble members from 20CRv3 are shown individually in figures 8(b) and (d), which provides an measure of the uncertainty in the reanalysis-derived diagnostics. This shows that reanalysis uncertainty is small compared to the spread across different models and between historical realizations of the same model.