Introduction

Since 1950, Arctic surface air temperatures have increased at a rate two to four times the planetary average1,2,3,4. This, observed2,5,6 and projected7,8,9, Arctic amplification owes to feedbacks related to sea ice5,10,11,12, temperature13,14,15, water vapor and clouds16,17 as well as remote factors related to poleward heat and moisture transport14,17,18. Arctic amplification is most pronounced during winter and close to the surface but largely absent during summer and above 850 hPa (~1.5 km height)14,19. This distinct spatiotemporal warming pattern is largely due to the strong stratification of the Arctic lower troposphere associated with temperature inversions during winter and linked with the lapse-rate feedback13,14,15. At the same time, the warming pattern implies that the stratification during winter weakens, and with it, the amplitude of the seasonal cycle of lower-tropospheric stability.

Weaker stratification enables stronger turbulence and thereby enhances vertical mixing of air and the properties it carries. This includes the transfer of horizontal momentum from the free troposphere and upper boundary layer towards the surface, thereby affecting the vertical wind profile20,21,22. The wind profile can feedback on atmospheric stability by affecting turbulence and mixing, in particular in regions of strong stability where wind shear is the primary source of turbulence23. This complex interplay is difficult to disentangle, but in any case, a thermally forced destabilization by surface-amplified warming is expected to strengthen the vertical momentum transfer. A strong influence of lower-tropospheric stability on near-surface winds in the lowest portion of the boundary layer, including at the standard observing height of 10m, has long been described for very different parts of the globe. For example, near-surface winds are relatively decoupled from the strong flow aloft above the equatorial Pacific cold tongue24. Moreover, a negative relation between stratification and the ratio between surface stress and geostrophic wind has been observed in the Arctic25,26.

Consistently, in a high-emission scenario, climate models contributing to the Coupled Model Intercomparison Project phase 5 (CMIP5)27 simulate a pronounced increase in near-surface wind speeds by up to 23% in winter over parts of the Arctic Ocean until the end of the 21st century28. This could have major implications for the evolution of sea ice29,30,31 and ocean waves32, and thereby on Arctic marine navigation and coastal erosion33.

Apart from reduced atmospheric stability, two more causal factors have been discussed, namely reduced surface roughness due to sea-ice decline and changes in atmospheric horizontal temperature and pressure gradients that may influence the storm tracks and wind patterns more generally12,28,34,35,36,37,38,39. It has been suggested that the latter plays a subordinate role, and the relative importance of atmospheric stratification versus surface roughness is still debated28,39.

The separation of these two factors is challenging because both are strongly affected by sea ice. First, sea ice is typically rougher than the open ocean26,28,34,35, although the relation between sea-ice concentration, effective roughness, and ocean stress is complex and exhibits a maximum at intermediate sea-ice concentrations40,41. Second, sea ice and the snow on top of it hinder warming of the near-surface air by latent cooling and reflection of solar radiation in summer, and by insulating the cold air from the warmer ocean in winter12. Correspondingly, strong negative correlations in time and space between sea-ice concentration and near-surface wind speed have been documented36,42,43, but two recent analyses trying to disentangle the contributions from surface roughness and atmospheric stability remain inconclusive.

One study based on CMIP5 projections concludes that “some combination of weakened atmospheric stability and reduced surface roughness” leads to the strengthening of surface winds28. Finding that one pair of similar models exhibits decreasing near-surface wind speeds despite decreasing atmospheric stability, it is argued that “surface roughness may be critical for explaining the sign and magnitude of future wind changes over the marine Arctic”28. Another study approaches the separation of these two mechanisms more directly with dedicated sensitivity experiments where the sea-ice surface roughness is modified such that it resembles the smoother open ocean in a coupled climate model39. With surface roughness trends thereby removed, it is found that the wind-speed increase until 2100 in a high-emission scenario is reduced over the Arctic Ocean on average by ~40% in autumn but only by ~10% in winter39. In contrast, much weaker but still positive near-surface wind-speed trends in spring and summer are turned into slightly negative trends by the elimination of the roughness effect—consistent with the finding that the atmosphere tends to become more stable during these times of the year39, when the remaining ice keeps the near-surface air temperature close to the freezing point while increasingly warm air is advected aloft from lower latitudes.

Here, changes in Arctic Ocean near-surface winds and their link to sea-ice decline and atmospheric stability are revisited based on ERA5 reanalysis44,45 as well as historical and high-emission scenario data of the CMIP646. It is shown that a coherent increase in winter near-surface wind speed, absent in reanalysis data over the satellite era since 197929,43, emerges in the Central Arctic when data since 1950 are included, despite negligible sea-ice concentration trends. While pre-1979 reanalysis data are more uncertain and trends in reanalyses must be regarded with caution47, the ERA5 historical trends are consistent with CMIP6 data. A tight relation between stratification trends and wind-speed profile trends across seasons and data sets is demonstrated, suggesting a dominant role of lower tropospheric stability in past and future Arctic wind trends.

Results

Near-surface patterns of trends

According to ERA5, Arctic winter (January, February, March; JFM) 2-meter temperature exhibits an ubiquitous surface warming over the Arctic Ocean (Fig. 1d), with two hot spots in the East Greenland and Barents Seas. These coincide with the areas of highest sea-ice concentration reduction (Fig. 1a). However, a clear near-surface warming trend also prevails beyond these seas, where the sea-ice concentration in winter has not yet declined; averaged over the inner Arctic Ocean (delineated in Fig. 1e and termed Arctic Ocean hereafter), the linear warming trend 1950–2020 is +0.8 K per decade, five times the global average trend exhibited by ERA5 for the same period (Supplementary Fig. 1). The CMIP6 multi-model mean exhibits similar trend patterns over the historical (including 2015–2020 scenario) period (Fig. 1b, e), although only the Barents Sea appears as a hot spot which is spatially more diluted, largely due to model differences in the ice-edge location (Supplementary Figs. 2 and 3); the Arctic Ocean average 2-meter temperature trend is +0.6 K per decade (range [+0.3, +1.3] K per decade; Fig. 2a). The trend steepens in the high-emission scenario and reaches +1.4 K per decade (range [+0.7, +2.6] K per decade) averaged over the extended period 1950–2100 (Fig. 2b). Some sea-ice decline now also reaches the central Arctic (Fig. 1c). The marginal seas still exhibit the strongest sea-ice trends, but they do not appear as warming hot spots anymore relative to the strong warming across the Arctic Ocean (Fig. 1f).

Fig. 1: Arctic winter (JFM) mean trends in sea-ice concentration, 2-meter temperature, and 10-meter wind speed.
figure 1

Arctic winter mean trends in sea-ice concentration (a–c), 2-meter temperature (d–f), and 10-meter wind speed (g–i). The trends are based on ERA5 and the CMIP6 multi-model mean for the periods 1950–2020 and 1950–2100, as indicated in the bottom right side of each map. The mask for spatial averages is delineated in e. Statistical significance of local CMIP6 multi-model mean trends is based on a two-sided one-sample t test for difference from zero; stippling indicates non-significant trends (p > 0.05).

Full size image
Fig. 2: Vertical profiles of Arctic Ocean winter (JFM) mean trends in air temperature, vertical temperature gradient, and wind speed.
figure 2

Vertical profiles of Arctic Ocean winter mean trends in air temperature (a, b), vertical temperature gradient (c, d), and wind speed (d, e). The period 1950–2020 (a, c, e) includes both ERA5 (black solid line) and CMIP6 models (colored lines), while the extended period 1950–2100 (b, d, f) includes only CMIP6 models. Insets (g–i) show the corresponding mean profiles for 1950–2020. The gray horizontal line marks the 850 hPa level at ~1.3 km altitude in the lower free troposphere. Statistical significance of the CMIP6 multi-model mean trend (dashed black/gray curve) is based on a two-sided one-sample t test for difference from zero.

Full size image

While there is no clear positive trend in near-surface wind speed across the Arctic Ocean in reanalysis data since 1979 (Supplementary Fig. 1; see also refs. 29,43), an ubiquitous speed-up, on average by 1% per decade (0.06 m s−1 per decade), emerges when ERA5 data since 1950 are included (Fig. 1g). This is consistent with CMIP6, although the historical CMIP6 mean trend is weaker (0.4% per decade; Fig. 1h). There is substantial model spread, with some models, such as AWI-CM-1-1-MR48, capturing the amplitude and pattern of the ERA5 wind-speed trend much better than others (Supplementary Fig. 4). The AWI-CM-1-1-MR ensemble exemplifies that the near-surface wind-speed trend pattern is fairly robust among the ensemble members, whereas the historical period seems to be too short to derive reliable trends for wind vectors.

As for the warming, the near-surface wind-speed trend steepens when the whole 21st century is included (1.15% per decade; Fig. 1i). Importantly, the wind-speed trend patterns resemble more closely the near-surface warming than the sea-ice concentration trends, in particular over the historical period (Fig. 1g, h). This provides the first evidence that near-surface warming, through its influence on atmospheric stability, may be more important than the sea-ice concentration decrease per se, through its influence on roughness.

Vertical structure of trends

To elucidate trends in atmospheric stability and vertical momentum transfer further, it is expedient to relate the near-surface trends to those further aloft. Consistent with earlier works14,19, both ERA5 and CMIP6 model data exhibit about three times faster warming close to the surface compared to the free troposphere, at and above the 850 hPa level, over the Arctic Ocean in winter (Fig. 2a, b). This implies a reduction of the inversion strength (Fig. 2c, d, g, h) and thus a weakening of the atmospheric stratification. The CMIP6 spread is large, but all models exhibit a negative temperature gradient trend that becomes even steeper when including the whole 21st century (Fig. 2d).

At the same time, relative wind speed trends exhibit a local maximum at the surface in ERA5 and in all models for both periods (Fig. 2e, f). This is consistent with enhanced vertical momentum transfer towards the surface, although it could also be caused by reduced surface roughness (further discussed below). Three models with a negative wind speed trend at 850 hPa exhibit a slightly negative trend also at the surface over the historical period, but the remaining 16 models exhibit a positive near-surface trend, consistent with ERA5’s winds strengthening by 1% per decade at the surface compared to 0.2% per decade at 850 hPa. Over the extended period, all models simulate increasing near-surface wind speeds, combined with a near-neutral to moderately negative wind speed trend at 850 hPa (Fig. 2f).

The largely negative wind speed trends in the free troposphere may at least partly have the same cause as the near-surface wind acceleration, namely the increased downward momentum transfer. However, the slow-down aloft may also be due partly to a weakening of the Arctic large-scale circulation, possibly in response to the reduced lower-tropospheric meridional temperature gradient. To filter out possible effects from large-scale circulation changes and to isolate the relation between atmospheric stability and momentum transfer, the following analysis relates trends in the temperature difference between the lower free troposphere and the near-surface to trends in the wind-speed ratio between the near-surface and the lower free troposphere, similar as in43. The wind speed ratio essentially measures how strongly the near-surface winds are coupled to the free-tropospheric winds. It is influenced, on the one hand, by the efficiency of downward momentum transfer across the whole boundary layer and, on the other hand, by how strongly the near-surface wind is attenuated by near-surface turbulence due to surface roughness. Thus both a weakened thermal stratification and a reduced surface roughness can cause an increased wind-speed ratio, whereas changing free-tropospheric winds, at first order, do not affect the wind-speed ratio.

The top of the Arctic boundary layer is typically a few hundred meters up to one kilometer above the surface23,36,49. This implies that the 850 hPa pressure level at ~1.3 km altitude (marked by a gray line in Fig. 2) is just above the maximum boundary layer height and can be considered representative of the lowest part of the free troposphere. The 850 hPa level, as well as the 10 m wind and 2 m temperature levels, are standard output and thus available from both ERA5 and CMIP6 data, so we use this to derive the wind speed ratio.

Relation between stratification and wind-speed ratio

Model differences regarding trends in stratification and trends in the wind-speed ratio in winter are highly and significantly anti-correlated and proportional, for both the historical (R= −0.84, slope −1.1% per K) and the extended (R = −0.95, slope −1.5% per K) period (Fig. 3a, b), precisely matching the ERA5 relation in the first case. The winter CMIP6 multi-model mean exhibits a very similar slope of −1.35% per K and an almost exactly linear progression of subsequent decadal means from 1950 to 2100 (Fig. 4b). The weaker inter-model correlation for the shorter historical period (Fig. 3a) as well as the less stringent progression of ERA5 winter decadal means (Fig. 4a) may simply be due to a lower signal-to-noise ratio caused by internal variability. Indeed, the ensemble spread (indicated by the crosses in Fig. 3a, b) of most individual models is about as large as the average signal over the historical period but not over the longer period. The overall tight relation between stratification and wind-speed ratio suggests a dominant role of trends in atmospheric stability for the intensification of Arctic Ocean near-surface winds in winter.

Fig. 3: Linear trends in wind speed ratio, vertical temperature gradient, and sea ice concentration.
figure 3

Linear trend in wind-speed ratio versus (a–d) linear trend in temperature difference between near-surface and 850 hPa levels and versus (e–h) linear trend in sea ice concentration. Winter, JFM, and summer, JAS, seasons are included. The period 1950–2020 includes both ERA5 (black square) and CMIP6 models (colored symbols), while the extended period 1950–2100 includes only CMIP6 models. The black dots show the multi-model mean trends. Lines denote linear least-square fits. For each model with more than one ensemble run included in the study, the ensemble spread is shown.

Full size image
Fig. 4: Decadal means of wind speed ratio and vertical temperature gradient.
figure 4

Decadal means of wind-speed ratio versus temperature difference between near-surface and 850 hPa levels in winter (JFM, circles) and summer (JAS, squares), for ERA5 1950–2020 (a) and the CMIP6 multi-model mean 1950–2100 (b).

Full size image

In summer (July, August, September; JAS), the sea-ice, temperature, and wind speed mean states and trends are markedly different (Supplementary Figs. 5 and 6). The warming over the Arctic Ocean is on average quite homogeneous throughout the troposphere, although some models simulate a near-surface dampening of the warming and thus a stabilization, consistent with39, whereas others simulate a near-surface amplification and thus a destabilization of the atmosphere (Supplementary Fig. 6a–d). Consistently, wind-speed ratios are not changing uniformly (Supplementary Fig. 6e, f), but again there is a highly significant anti-correlation (R = −0.61, slope −1.8% per K and R = −0.85, slope −1.9% per K) between the trend in stratification and the trend in wind-speed ratio (Fig. 3c, d).

ERA5 data exhibit a slight stabilization and decreasing wind-speed ratio in summer so that the summer and winter conditions approach each other in terms of stratification and wind-speed ratio (Fig. 4a), resulting in an attenuated amplitude of the seasonal cycle. Decreasing seasonality occurs also in the CMIP6 multi-model mean (Fig. 4b), but only because of the changes in winter. In summer, the CMIP6 multi-model mean stratification and wind-speed ratio remain unchanged over the historical period, but the wind-speed ratio gradually increases over the course of the high-emission scenario despite constant stratification.

Contributions from sea-ice decline and large-scale winds

Another factor may contribute to the projected increase of the wind-speed ratio over the Arctic Ocean in summer, namely the decreasing surface roughness resulting from the profound sea-ice decline, consistent with28,39. While other factors can affect ocean surface roughness, in particular sea state, here we consider only sea-ice concentration. Indeed, 15 out of the 19 models exhibit an area-averaged decrease of the summer sea-ice concentration beyond 40% (−2.67% per decade over 1950–2100), and 12 of these exhibit an increasing wind-speed ratio (Fig. 3h). Given this link, one may ask again more generally which factor is primarily driving the response of the wind-speed ratio: the thermal stratification or the surface roughness?

Relating changes in the wind-speed ratio to changes in sea-ice concentration across CMIP6 models for the long period 1950–2100 in winter reveals that the two are also closely anti-correlated (R = −0.97, slope −0.39% per %; Fig. 3f). Given the close link between the surface warming and the sea-ice decline in this case, the high anti-correlation with both sea-ice decline and temperature gradient reduction provides limited evidence to determine which factor is actually driving the response of the wind-speed ratio. More prominent differences arise when considering the situation in summer, where the long-term anti-correlation with the sea-ice concentration trend drops considerably (R = −0.5, slope −0.13% per %; Fig. 3h), whereas the anti-correlation with the temperature gradient remains much higher (R = −0.77; Fig. 3d). An even stronger contrast is evident for the historical period in both winter and summer (Fig. 3e, g): the correlations with the sea-ice concentration trend is weak and barely statistically significant in winter (R = −0.39, slope −0.3% per %) and summer (R = −0.45, slope −0.15% per %), in contrast to the medium to high correlations with the temperature gradient trend (R = −0.84 and −0.61; Fig. 3a, c). Combined with the better-matching spatial trend patterns (Fig. 1), the more robust link between the vertical temperature gradient and the wind-speed ratio across seasons and periods suggests that changes in the thermal stratification are dominating the response of the wind speed ratio rather than changes in surface roughness.

Finally, the near-surface wind speed trend can be influenced not only by changes in thermal stratification and surface roughness, which affect the wind speed ratio, but also by possible wind-speed changes aloft. The slow-down of the free-tropospheric winds in winter (Fig. 2e, f) can be interpreted as a direct consequence of the increased downward momentum transfer. However, some of the deceleration aloft may also be driven by changes in the large-scale circulation. This could explain why some models simulate only a weak winterly near-surface wind intensification over the extended period despite a pronounced weakening of the atmospheric stratification. In contrast, free-tropospheric summerly wind speeds are increasing in ERA5 and in most models in concert with the near-surface wind intensification (Supplementary Fig. 6e, f). While model differences between the near-surface and free-tropospheric relative wind-speed trends are related to the change in stratification also in summer (Fig. 3c, d), the vertically more homogeneous trends suggest that the increased wind speeds in summer can not be linked to enhanced vertical mixing or decreased surface roughness alone, but that they are mostly due to large-scale drivers.

Discussion

The ERA5 and CMIP6 data analyzed here suggest that the past and future near-surface wind intensification over the inner Arctic Ocean in winter is largely driven by an increasing downward momentum transfer due to a weakening atmospheric stratification, possibly damped by decelerating large-scale winds aloft. In contrast, the near-surface wind intensification in summer is largely driven by accelerating winds aloft, amplified in a high-emission future by decreasing surface roughness due to sea-ice decline. In both seasons, differences in wind-speed ratio trends and, thus, near-surface wind-speed trends between data sets, including models as well as ERA5, are closely linked to differences in atmospheric stability trends.

These results are qualitatively consistent with earlier work28,39. However, while we do not quantify the contributions with exact estimates but rather infer them from pattern similarity, the concluded relative importance of the factors contributing to near-surface wind-speed trends over the Arctic Ocean is different. Here, it is concluded that diminishing surface roughness due to sea-ice decline contributes notably only in summer over the course of the next decades, but not over the historical period, and in winter scarcely at all. A high anti-correlation between long-term simulated winter trends in wind-speed ratio and sea-ice concentration seems to be largely due to the close physical link between surface warming and sea-ice decline. Instead, our results suggest that trends in atmospheric stability play a dominant role and also largely explain differences between models. Moreover, large-scale wind-speed trends aloft may modify the near-surface wind speed trends, with a rather decelerating effect in winter and an accelerating effect in summer.

A clear limitation is that the validity of these results depends on the fidelity of the models to simulate the relevant processes. This pertains not only to CMIP6, but also to ERA5 where the polar boundary layers may be among the regions that are least well constrained by observations and thus depend most on the model formulation50. Another relevant caveat of ERA5 is that the sea-ice thickness is assumed to be constant, with no snow on top, leading to a surface warm bias in winter51 and to an inability to represent a contribution to the winter surface warming from sea-ice and snow thinning52. Moreover, trends in reanalyzes must be regarded with caution because they can be affected by discontinuous availability of assimilated observations47. Yet, for the same reason that reanalyzes are weakly constrained in the Arctic boundary layer, multi-decadal trends across the inner Arctic Ocean can not be evaluated by actual observations: because these are spatio-temporally too scarce50.

Bearing this in mind, the CMIP6 simulations suggest that, until the end of this century, in a high-emission scenario, the lower troposphere may become as unstable in winter as it is and remains in summer. The efficiency of downward momentum transfer would align seasonally likewise. This implies a fundamental regime shift of the Arctic winter near-surface winds and boundary layer, more generally, likely with major repercussions on the sea ice and ocean below, affecting ocean waves, marine navigation, and coastal erosion.

Methods

Models and data

All climate model data used here are from the CMIP6 dataset46,53. The period 1950–2020, referred to as the historical period here, is composed of the CMIP6 historical simulations (1950–2014) and the first years of the SSP370 scenario simulations (2015–2020); the period 2021—2100 is based on the remainder of the SSP370 scenario simulations. The model selection was determined by the availability of data for the required variables at the required temporal frequency, i.e., daily resolution for wind velocity components and monthly resolution for both air temperature and sea-ice concentration. ERA5 reanalysis data, covering the period 1979–202044 and the preliminary backward extension 1950–197845, have been used as reference for both air temperature and wind speed fields.

Wind speed is not an archived variable for CMIP6 models, except for the surface level in a few cases. To obtain a consistent dataset, monthly wind speed was computed by averaging daily wind speed derived from daily wind velocity components. Using monthly sampled velocity components, and thus not capturing sub-monthly variations in wind direction and speed, would have implied an underestimation of average wind speeds by about a factor of two54 and could have distorted the obtained trends.

Apart from 10-meter wind speed, 2-meter air temperature, and sea-ice concentration, slightly different pressure levels have been used for different data sets and variables due to data availability constraints. For CMIP6, 5 pressure levels (850, 700, 500, 250, 100 hPa) have been used for wind and 10 (925, 850, 700, 600, 500, 400, 300, 250, 200 hPa) for air temperature (original pressure levels without interpolation). For ERA5, 19 pressure levels from 950 hPa to 50 hPa, with intervals of 50 hPa, have been used for both variables.

For most of the models, all ensemble members with the required variables available in the aforementioned frequency and for the full time period considered, have been included in the study, see Table 1. For the few models with more than 10 ensemble members only the first 10 have been used. All the Figures showing results for individual models show ensemble means. The only exception is the AWI-CM-1-1-MR model in Supplementary Figs. 2–4, where all five ensemble members are shown individually to exemplify the robustness of the derived trend patterns given internal variability. While the exact magnitude of internal variability varies between data sets, including models as well as reality and how it is captured by ERA5, the same order of magnitude can be expected55.

Table 1 CMIP6 climate models included in the study
Full size table

Data analysis

Spatial trends (Fig. 1 and Supplementary Fig. 5) are based on the original horizontal grid for ERA5, whereas CMIP6 data have first been remapped to a common 1.25° × 1.85° latitude-longitude grid. Multi-model means have been computed afterwards. From four of the models considered (CESM2-WACCM, CESM2, FGOALS-g3, and NorESM2-LM), only the near-surface wind speed and no velocity components were available on the ESGF platform (see Supplementary Fig. 4). These have been included in the multi-model wind speed mean, but not in the multi-model wind vector field mean. Results for individual models (Figs. 2 and 3 and Supplementary Figs. 2–4, 6) have been derived separately based on the original horizontal model grids.

The Arctic Ocean domain used for spatial averages is the area north of 68°N for longitudes east of 103°E and west of 124°W, and the area north of 79°N at all other longitudes (Fig. 1e). Grid points on land or within 150 km from any coastline have been omitted to avoid the influence of land and orography, similarly to54.

Wind speed trends in Fig. 2 and Supplementary Fig. 6 are relative trends, obtained from division by the mean wind speed of the period 1950–2020 for the same model, height, and season. This normalization enables a more direct interpretation of the wind speed trend profiles regarding the efficiency of vertical momentum transfer. Temperature and temperature gradient trends in the same two figures are absolute trends as it is for all the trends shown in Figs. 1 and Supplementary Fig. 5.

Two-sided one-sample t tests have been used to test if CMIP6 multi-model mean trends are significantly different from zero. In Fig. 2 and Supplementary Fig. 6, the results of the t tests are displayed by coloring the multi-model mean curve either black (p < 0.05) or gray (p ≥ 0.05); here, the test has been applied to the area-averaged trends for each pressure level separately. In Fig. 1 and Supplementary Fig. 5, in the CMIP6 mean panels, grid points where p ≥ 0.05 have been stippled; here, the test has been applied locally for each grid point after performing the remapping.