The patterns of sea ice retreat in the Arctic Ocean are investigated using two global climate models (GCMs) that have profound differences in their large-scale mean winter atmospheric circulation and sea ice drift patterns. The Community Earth System Model Large Ensemble (CESM-LE) presents a mean sea level pressure pattern that is in general agreement with observations for the late twentieth century. The Community Climate System Model, version 4 (CCSM4), exhibits a low bias in its mean sea level pressure over the Arctic region with a deeper Icelandic low. A dynamical mechanism is presented in which large-scale mean winter atmospheric circulation has significant effect on the following September sea ice extent anomaly by influencing ice divergence in specific areas. A Lagrangian model is used to backtrack the 80°N line from the approximate time of the melt onset to its prior positions throughout the previous winter and quantify the divergence across the Pacific and Eurasian sectors of the Arctic. It is found that CCSM4 simulates more sea ice divergence in the Beaufort and Chukchi Seas and less divergence in the Eurasian seas when compared to CESM-LE, leading to a Pacific-centric sea ice retreat. On the other hand, CESM-LE shows a more symmetrical retreat between the Pacific, Eurasian, and Atlantic sectors of the Arctic. Given that a positive trend in the Arctic Oscillation (AO) index, associated with low sea level pressure anomalies in the Arctic, is a robust feature of GCMs participating in phase 5 of the Coupled Model Intercomparison Project (CMIP5), these results suggest that the sea ice retreat in the Pacific sector could be amplified during the transition to a seasonal ice cover.
Significant changes in the Arctic sea ice cover have been observed in recent years, particularly at the end of summer (Stroeve et al. 2007). The decline in minimum sea ice extent has accelerated from approximately −2.2% decade−1 between 1978 and 1996 to −10.1% decade−1 between 1996 and 2007 (Comiso et al. 2008; Serreze and Stroeve 2015). In addition to the loss of area of perennial ice (Tucker et al. 2001), there has been an overall thinning of the Arctic ice cover (Kwok and Rothrock 2009), with a significant decrease in the oldest and thickest ice within the multiyear ice pack (Pfirman et al. 2004; Maslanik et al. 2007). A thin ice cover is more vulnerable to strong summer retreat under anomalous atmospheric forcing or increasing ocean heat transport to the Arctic (Maslanik et al. 1996; Holland et al. 2006; Stroeve et al. 2012b). Moreover, the decrease in areal extent promotes a positive ice–albedo feedback (Perovich et al. 2007a).
The spatial variability of sea ice retreat may also impact global weather patterns. Several studies have associated Arctic sea ice loss with increased storm surge activity and cold winter extremes in North America, Europe, and northern Asia (Maslanik et al. 1996; Petoukhov and Semenov 2010; Francis and Vavrus 2012; Inoue et al. 2012; Tang et al. 2013; Vermaire et al. 2013; Vihma 2014; Francis and Vavrus 2015; Gervais et al. 2016), although the exact nature of this link is still an active area of debate (Barnes 2013; Barnes et al. 2014). While model simulations from phase 5 of the Coupled Model Intercomparison Project (CMIP5) offer some evidence that future Arctic warming and sea ice loss may have an effect on the midlatitude jet stream, the net circulation response is unlikely to be explained primarily by Arctic amplification (Barnes and Screen 2015; Barnes and Polvani 2015).
As the summer sea ice retreats, industrial and commercial interests are evolving. Travel time between Europe and the northern ports of Asia can be reduced by 50% when the Northern Sea Route is free of ice (Peters et al. 2011). The length of the navigation season of the Northern Sea Route and Northwest Passage is projected to increase by 173% and 156%, respectively, by the end of the twenty-first century (Khon et al. 2010; Smith and Stephenson 2013). Although oil and gas activity is currently at a relative low, the Arctic holds a large fraction of undiscovered energy resources, which will be increasingly accessible as sea ice retreats and thins (Gautier et al. 2009; Blanken et al. 2015, 2016, manuscript submitted to Mar. Pollut. Bull.). The rate and patterns of change of the Arctic ice cover will be of high importance for planning and development, resource management, species preservation, national security, and international environmental monitoring programs (Campbell et al. 2007; Byers 2010).
Many studies have investigated the thermodynamic and dynamic mechanisms driving the observed sea ice retreat (Serreze et al. 2007). For instance, prior to the record sea ice extent minimum of 2007, the oceanic heat fluxes through the Bering Strait reached about twice the average of the prior 7 years, enough to melt one-third of the 2007 estimated total sea ice loss (Steele et al. 2010; Woodgate et al. 2010). In 2007, the Arctic dipole anomaly (with a high-pressure system over the arctic regions of North America) persisted throughout the summer. It led to a strengthening of the Transpolar Drift Stream and ice export through the Fram Strait, leaving open water off the shores of Alaska and Siberia, and to cloud-free skies, allowing for more surface solar heating of the ocean (L’Heureux et al. 2008; Kay et al. 2008; Schweiger et al. 2008; Zhang et al. 2008; Wang et al. 2009; Screen et al. 2011; Overland et al. 2012). Nonetheless, Lindsay et al. (2009) found that the 2007 ice mass loss followed the trend, as the area of thin ice at the beginning of the melt season and the total volume of ice in the summer have been steadily decreasing since 1987.
The large-scale pattern of sea level pressure in the Arctic can be characterized, to first order, by the Arctic Oscillation (AO) (Thompson and Wallace 1998), the first EOF of Northern Hemisphere sea level pressure, or by the northern annular mode (NAM) (Thompson and Wallace 2001), a closely related high-latitude pattern. Rigor et al. (2002) showed that the AO explains 52% of the variance of sea level pressure over the Arctic Ocean during winter. They also suggest that the summer sea ice concentration is correlated with the AO index of the previous winter, reflecting the dynamical influence of the wintertime sea level pressure on thickness distribution. Likewise, Rigor and Wallace (2004) found that more than half of the variance in summer sea ice extent can be explained by the age of the ice pack, a proxy for ice thickness. More recently, Williams et al. (2016) linked the phase of the mean winter (December–March) AO index and the divergence of sea ice away from the Eurasian and Alaskan coastlines . Coastal divergence in turn leads to the formation of thin first-year ice that melts more readily in early summer and a smaller September sea ice extent (Williams et al. 2016). These results reinforce the idea of potential predictability of the September sea ice extent based on the previous winter mean sea ice circulation (Holland and Stroeve 2011).
Seasonal predictions of Arctic sea ice have been based on a variety of modeling, statistical, and heuristic approaches (Guemas et al. 2014). Stroeve et al. (2014b) found that sea ice concentration is the most important predictor of September sea ice extent between 2008 and 2013 for lead times of 2 months or less, but for longer lead times ocean temperature and sea ice thickness come into play as reemergence mechanisms (Blanchard-Wrigglesworth et al. 2011a; Chevallier and Salas-Mélia 2012; Chevallier et al. 2013). Atmospheric variability also provides an inherent limit on sea ice predictability (Guemas et al. 2014; Serreze and Stroeve 2015), making interannual departures from the long-term trend hard to predict (Tietsche et al. 2013; Stroeve et al. 2014b). Even with a “perfect model” assumption, under which ensemble integrations are initialized from a reference model, initial-value predictability of sea ice area is only significant for up to 1–2 years (Blanchard-Wrigglesworth et al. 2011b; Tietsche et al. 2013). Beyond 3 years, predictability is dominated by climate forcing rather than initial conditions (Blanchard-Wrigglesworth et al. 2011b). In addition, the long-term predictive capability of the September minimum sea ice extent decreases as the ice pack gets thinner (Blanchard-Wrigglesworth et al. 2011a; Holland et al. 2011; Tietsche et al. 2013). Nonetheless, Holland and Stroeve (2011) looked at climate model projections through the end of the twenty-first century and found that the correlation between September sea ice extent anomalies and winter–spring predictors such as the area of the Arctic Basin covered by thin ice increases as the pack ice transitions from perennial to seasonal ice cover. At the same time, they also found that the variance of modeled sea ice extent explained by summer large-scale atmospheric circulation over the Arctic Ocean decreases from the late twentieth century to the middle of the twenty-first century.
A positive AO (or NAM) index is characterized by low sea level pressure anomalies over the Arctic, a small Beaufort Gyre, and increased coastal divergence in the East Siberian and Laptev Seas (Dickson et al. 2000; Rigor et al. 2002). Moreover, it can lead to a higher number of storms entering the Arctic through the Nordic seas, which can favor the inflow of warm Atlantic waters and enhance the winter sea ice retreat north of Scandinavia (Bengtsson et al. 2004; Sorteberg et al. 2005; Sorteberg and Kvingedal 2006). GCMs simulate a steady positive trend in the NAM (Fyfe et al. 1999; Rauthe and Paeth 2004; Miller et al. 2006), suggesting either that the observed return to a negative phase since the nineties is a manifestation of internal variability or results from a mechanism not resolved by models (Gillett and Fyfe 2013; Jin-Qing et al. 2013). In particular, Gillett and Fyfe (2013) show that CMIP5 models on average simulate an increase in the NAM in every season by 2100, with the largest increases in autumn and winter.
GCMs that have a reasonable late-twentieth-century Arctic climate predict ice-free summers before 2100 for a business-as-usual emission scenario (Holland et al. 2006; Stroeve et al. 2012a). Even though these models agree on the decline of sea ice extent and the likelihood of a seasonal Arctic sea ice cover, the pattern of the predicted sea ice loss varies widely (Bitz et al. 2005). Recently, Stroeve et al. (2012a) showed that predicted trends from the models participating in CMIP5 are consistent with observations, but they nevertheless exhibit declines that are smaller than the observed value. Although the mean thickness distributions from GCMs are in good agreement with observations, there are large differences in the spatial patterns of sea ice thickness in CMIP3 and CMIP5 models (Bitz et al. 2002; Stroeve et al. 2014a). Such disparities can be the result of relatively small biases in atmospheric circulation patterns that drive the movement of sea ice. Nevertheless, all models agree on the dominant role of sea ice melt in the simulated volume loss (Holland et al. 2010).
In this paper, we expand on the work of Williams et al. (2016) and link projected decadal trends of regional (Pacific and Eurasian sectors) summer sea ice extent loss in the Arctic to trends in sea ice drift patterns from the previous winter. We look at two GCMs that have important differences in their late-twentieth-century large-scale mean winter atmospheric circulation: one that is similar to observations [Community Earth System Model Large Ensemble (CESM-LE)] and one that exhibits a positive AO bias [Community Climate System Model, version 4 (CCSM4)]. As a result, CCSM4 simulates larger divergence in the sea ice field in the Beaufort and Chukchi Seas compared to CESM-LE, leading to a Pacific-centric sea ice retreat. On the other hand, CESM-LE has a mean winter atmospheric circulation in the twentieth- and twenty-first century that is similar to the observed late twentieth century, leading to a more symmetrical retreat between the Pacific, Eurasian, and Atlantic sectors of the Arctic. Given that a positive trend in the AO index is a robust feature of GCMs participating in CMIP5, our results suggest that sea ice retreat in the Pacific sector could be amplified during the transition to a seasonally ice-free Arctic.
We use the National Snow and Ice Data Center (NSIDC) Polar Pathfinder Daily 25-km Equal-Area Scalable Earth Grid (EASE-Grid) Sea Ice Motion Vectors, version 3 (Tschudi et al. 2016), for the period of January 1990 to May 2015. This dataset is constructed using an optimal interpolation of sea ice motion vectors from the Scanning Multichannel Microwave Radiometer (SMMR), Special Sensor Microwave Imager (SSM/I), Advanced Microwave Scanning Radiometer-EOS (AMSR-E), Advanced Very High Resolution Radiometer (AVHRR), drifting buoys from the International Arctic Buoy Program (IABP), and free drift estimates derived from the 10-m winds of the NCEP–NCAR reanalysis data. For this analysis, we do not consider the satellite-derived velocities from SMMR (1979–87) because of a reported low bias in sea ice velocity before the change from SMMR to SSM/I in July 1987 (NSIDC 2016). In this dataset, IABP data are considered as truth and are used with a weight of unity that decreases to zero outside a radius of influence of approximately 200 km. Sea ice drift vectors have been derived using a maximum cross-correlation method applied to brightness temperature data (Emery et al. 1997; Meier et al. 2000; Fowler et al. 2004; Meier and Dai 2006). The process results in a sparse field of ice drift vectors, where features can be tracked from one image to the next. Satellite-derived daily ice motion vector fields were interpolated on the 25-km EASE-Grid (Brodzik et al. 2012) and averaged over a one-week and one-month time period, resulting in low root-mean-square errors and making long-term Lagrangian tracking possible (Meier and Maslanik 2001; Sumata et al. 2015b).
We use sea ice concentration data derived from passive microwave brightness temperatures between January 1990 and December 2014 from the National Oceanic and Atmospheric Administration (NOAA)/NSIDC Climate Data Record (CDR), version 2 (Meier et al. 2015). The sea ice concentration is defined as the fraction of a gridcell area covered by sea ice. CDR concentration estimates are derived using the NASA Team (NT) algorithm (Cavalieri et al. 1984) and the Bootstrap (BT) algorithm (Comiso 1986). The CDR algorithm then blends the NT and BT output concentrations by selecting, for each grid cell, the higher concentration value (Peng et al. 2013; Meier et al. 2014). The sea ice concentrations were interpolated onto the same 25-km EASE-Grid.
Monthly mean sea level pressure is obtained from ERA-Interim, a product of the European Centre for Medium-Range Weather Forecasts (ECMWF) (Dee et al. 2011), for the period of January 1979 to December 2015, as well as from the National Centers for Environmental Prediction and the National Center for Atmospheric Research (NCEP–NCAR) between January 1948 and December 2014. Sea level pressures were interpolated onto the same 25-km EASE-Grid.
We also use the NSIDC IABP drifting buoy data product that includes 12-h latitude/longitude positions from 1988 to 2011 (Rigor 2002).
b. Global climate models
This study utilizes monthly mean fields of the u and υ components of sea ice velocity, sea ice concentration, and sea level pressure between 1900 and 2100 (historical and RCP8.5 simulations) from the six ensemble members of the CCSM4 provided by NCAR (Gent et al. 2011).
We also use monthly mean u and υ components of sea ice velocity, sea ice concentration, and sea level pressure between 1920 and 2100 from the 40 ensemble members of the CESM-LE with historical and RCP8.5 forcings (Kay et al. 2014). CESM-LE uses the latest version of the Community Atmosphere Model (CAM5), which has undergone substantial improvements in the representation of physical processes in the atmosphere (e.g., parameterizations of diabatic processes, interactive carbon–nitrogen cycling, and treatment of water substances and aerosols) (Hurrell et al. 2013). Of importance to the Arctic are the resulting improvements in the representation of the total cloud percentage (Barton et al. 2012; Kay et al. 2012).
Although the land, ocean, and ice components, as well as the spatial resolution are the same for both models, CESM-LE tends to produce thicker ice, which may be a result of the reduced twentieth-century warming in CAM5 (Hurrell et al. 2013). CESM-LE also has stronger climate feedbacks, leading to a higher climate sensitivity compared to CCSM4 and greater warming in the twenty-first century (Meehl et al. 2013).
Our goal is to assess the spatial patterns of future sea ice retreat on decadal time scales given the projected changes in large-scale atmospheric circulation patterns over the Arctic Ocean. To this end, we quantify the divergence of sea ice in different regions of the Arctic using a Lagrangian model to backtrack a selected set of points (virtual ice floes) from the approximate time of melt onset to their prior positions throughout the previous winter. A similar approach has been used successfully to track ice age over several years (Fowler et al. 2004; Pfirman et al. 2004; Rigor and Wallace 2004; Maslanik et al. 2007), and results compared well with ice thickness data, a proxy for ice age (Maslanik et al. 2007).
In this analysis, we use the ice drift vectors to backtrack an imaginary line at 80°N latitude and quantify sea ice divergence in two separate regions: the Pacific sector (Beaufort and Chukchi Seas) and the Eurasian sector (East Siberian, Laptev, and Kara Seas) of the Arctic. All Lagrangian trajectories are initialized on 1 June, the approximate start of the melt onset. At the beginning of the integration, each selected grid cell is considered an independent Lagrangian particle and is advected using weekly averaged sea ice motion vectors. After each time step, the gridded ice velocity is interpolated at the new particle location. This procedure is done repeatedly until 1 November of the previous year, when the Arctic Basin is still fully ice covered. We then calculate the fluxes of ice along the 80°N line in both regions (see example in Fig. 1). We do not include the Barents Sea in the analysis of the Eurasian sector because the ice edge in this region is strongly affected by ocean heat fluxes (Bitz et al. 2005). In section 4, we present an error analysis to quantify the biases in sea ice drift associated with the use of the Lagrangian model and the increased temporal resolution of the observational data (weekly) when compared with model data (monthly).
The advection of individual ice parcels is done using a first-order finite-difference approximation of the ice velocity:
where is the mean sea ice velocity at time , is the location of the particle at time t, and is equal to one week and one month for the observations and models, respectively.
To calculate the mean divergence over a given region, we use the two-dimensional divergence theorem:
This can be rewritten in terms of the rate of change in area provided by the Lagrangian backtrajectory model [see p. 6 of Vallis (2006)]:
For instance, the sea ice divergence from the Pacific and Eurasian sectors can be written as
where and are the original areas (before backtracking) of the Pacific and Eurasian sectors, respectively, and , , , , and are the areas of regions A, B, C, D, and E enclosed between the original position of the 80°N line (black solid line) and its backtracked position (black dashed line) (Fig. 1).
We have opted to calculate the divergence as a rate of change in area instead of spatial derivatives of the velocity because it gives an intuitive way of looking at the same process and highlights aspects of the divergent field that are not readily obvious when using ice motion vectors. In addition, calculating spatial and temporal averages of the Lagrangian trajectories decreases the errors associated with the vector field (Sumata et al. 2015b).
4. Error analysis
To quantify the error in drift distances from the Lagrangian model, we compare simulated sea ice drift forced with weekly and monthly mean sea ice velocities from the Polar Pathfinder dataset with the actual buoy drift trajectories from the IABP buoy data between 1988 and 2011 (Meier and Maslanik 2001). To this end, we track each individual IABP buoy from its initial position to its last recorded position for a maximum of 1 year. For each simulated trajectory, the error in modeled location compared to the actual buoy position was calculated at each time step until the final position of the buoy. The procedure allows us to estimate model errors as well as to quantify the impact of temporal resolution on the estimated errors. Recall that the observed IABP drift velocities are used to construct the Polar Pathfinder sea ice motion vectors, making a precise error estimate of the drift trajectories more strenuous. Averaging daily sea ice drift into longer time periods introduces an error in the initial velocity of the simulated trajectory. This error will grow in time as the distance between the simulated position and the actual buoy position increases.
Figure 2 shows the median, the interquartile range, and the number of trajectories used to calculate the error in drift distance using weekly and monthly time resolutions for trajectories of up to 1 year. The median is closer to the bottom of the interquartile range because the distribution is positively skewed, indicating that most buoys do not travel long distances and stay within the same area. Larger errors will arise when trying to simulate fast-drifting buoys, which represent a small portion of the data (Sumata et al. 2015a). As expected, the error growth is faster for pairs of points separated by more than 200 km, the radius of influence used for IABP data in the NSIDC interpolation scheme. Reducing the temporal resolution from weekly to monthly leads to an increase of the error in drift distance from about 83 to 128 km after a year of tracking. Given an observed mean sea ice drift speed in the Arctic Ocean of about 3–5 cm s−1, the total mean drift after a year is approximately 1200 km. The median errors in the trajectories using weekly and monthly time resolutions are therefore 7% and 11%, and the upper-quartile error is about 16% and 26%, respectively. Note that weekly averaged sea ice velocities are used throughout this study for observational data.
5. Results and discussion
The winter (November–May) mean Arctic sea level pressure (above 70°N) from NCEP–NCAR, ERA-Interim, CCSM4, and CESM-LE is shown in Fig. 3. CESM-LE mean wintertime sea level pressure is similar to observations. On the other hand, CCSM4 has a bias of around −7 mb over the Arctic region when compared to observations (De Boer et al. 2012; Jahn et al. 2012), as well as larger internal variability than observed (Vavrus et al. 2012). This is indicative of a mean state with a very positive AO index. Note that the years with highest mean sea level pressure in CCSM4 are comparable to years of lowest observed values. The decreasing trend in winter mean sea level pressure in CESM-LE and CCSM4 (Fig. 3) starts in the first and second halves of the twenty-first century, respectively, associated with a later transition to a seasonal sea ice cover in CCSM4 when compared to CESM-LE (Fig. 4) (Meehl et al. 2013).
The winter mean Arctic atmospheric circulation is characterized by a semipermanent high pressure system (Arctic or Beaufort high), flanked by two semipermanent climatological low pressure systems in the Atlantic (Icelandic low) and Pacific sectors (Aleutian low). The resulting observed winter mean sea ice drift pattern is the Beaufort Gyre, an anticyclonic circulation that redistributes thick multiyear ice north of the Canadian Arctic Archipelago (CAA) within the central Arctic, and the Transpolar Drift Stream that carries ice from the Eurasian coastline toward the North Pole and out through the Fram Strait (Figs. 5a,b).
The last decade of the twentieth century in CESM-LE is characterized by a broad Beaufort Gyre with faster sea ice drift speed when compared to observations and a smaller Transpolar Drift Stream that advects sea ice from the Kara Sea to the Fram Strait and out of the Arctic Ocean (Figs. 5c,d). There is also a bias in the location of the center of the Beaufort Gyre toward the Eurasian coastline, redistributing thin ice from the East Siberian Sea and part of the Laptev Sea to the central Arctic. In contrast, there is no clear evidence of the presence of the Beaufort Gyre circulation in the winter mean sea ice velocities averaged over 1990–99 in CCSM4 (Figs. 5e,f). In this model, the Icelandic low is deeper than observed, and it penetrates farther in the eastern Arctic, resulting in a broader Transpolar Drift Stream carrying thick ice north of the CAA out through the Fram Strait (De Boer et al. 2012). This atmospheric pattern is a typical manifestation of an extremely positive AO index.
The years with high sea level pressure anomalies in CESM-LE (Fig. 6c) are characterized by a strengthening of the Beaufort Gyre circulation and a weakening of the Transpolar Drift Stream, along with a larger bias in the location of the center of the Beaufort Gyre toward the Eurasian coast. Conversely, the years with low sea level pressure anomalies see a weakening of the Beaufort Gyre circulation and a strengthening of the Transpolar Drift Stream (Fig. 6d). Note that the same features in the sea ice drift pattern are present in the observations (Figs. 6a,b), although the observed departures from the mean are of smaller amplitude compared to CESM-LE. The large internal variability of the sea level pressure in CCSM4 accounts for important interannual variability in the large-scale mean winter atmospheric circulation (Vavrus et al. 2012). In fact, the sea ice circulation shown in Fig. 5e with the absence of a clear Beaufort Gyre is the result of the average of two very different regimes. In years with high sea level pressure anomalies (Fig. 6e), the winter sea ice velocities are very similar to the mean observed sea ice drift field (Fig. 5a). In years with low sea level pressure anomalies (Fig. 6f), the winter sea ice velocities follow a cyclonic circulation over the whole Arctic Basin, driven by an extremely deep Icelandic low, and a Transpolar Drift Stream that carries ice from the Beaufort Sea directly into the Fram Strait via the Lincoln Sea (Rigor et al. 2002). This circulation is characteristic of an extremely positive AO phase.
Future projections of sea ice extent from CESM-LE present a sea ice retreat that is relatively symmetric around the Lincoln Sea (north of Ellesmere Island) (Figs. 7a,c,e), and thick ice north of the CAA is maintained by the presence of the Beaufort Gyre throughout the twenty-first century (results not shown). In CCSM4, the simulated patterns of sea ice retreat are drastically different. The positive AO state that characterizes the large-scale mean winter atmospheric circulation results in a sea ice retreat preferentially in the Beaufort Sea (Figs. 7b,d,f). The region north of the CAA becomes ice free in September before the central Arctic (around the North Pole), with sea ice remaining in the region north of Greenland and in the narrow channels of the CAA (Jahn and Holland 2013).
We quantify the sea ice divergence from the Pacific and Eurasian sectors of the Arctic Ocean from November to June using the Lagrangian model (see example on Fig. 1). Figure 8 shows the net divergence from both sectors for CCSM4 and CESM-LE from the early twentieth century to 2100. In CESM-LE, the presence of a strong Beaufort Gyre circulation leads to a net convergence of ice in the Pacific sector. On the other hand, the extremely positive AO phase in CCSM4 prevents the redistribution of thick ice north of the CAA into the Beaufort and Chukchi Seas so that the net divergence of sea ice is approximately zero (Fig. 8b). Conversely, divergence of sea ice is somewhat more important in CESM-LE than in CCSM4 in the Eurasian sector since the strong anticyclonic circulation of the Beaufort Gyre in CESM-LE carries ice away from the Eurasian coastlines toward the North Pole (Fig. 8a). This highlights the importance of the different large-scale atmospheric circulation patterns on projected Arctic sea ice conditions (Liu et al. 2013; Meehl et al. 2013).
In the observations, the first decade of the twenty-first century has seen successive record low sea ice extent minima (Stroeve et al. 2012b), most dramatically visible in the sea ice retreat in the Beaufort, Chukchi, East Siberian, and Laptev Seas (Fig. 9b). The general thinning of the ice pack has led to an overall increase in sea ice velocities across the Arctic Ocean and a more mobile ice pack, except for the area north of the CAA (Fig. 9a). The increase in ice drift speed along the Alaskan coastline has strengthened the anticyclonic circulation in the Pacific sector. We again use our Lagrangian model to evaluate the observed divergence from the Pacific and the Eurasian sectors. We calculate the net sea ice divergence from the first week of February to the first week of June (approximate beginning of the melt season). Several studies have shown this period to have the highest correlation between the September sea ice extent anomaly and sea ice thickness distribution (Nikolaeva and Sesterikov 1970; Chevallier and Salas-Mélia 2012; Williams et al. 2016).
Since the early nineties, we observe a significant positive trend in the net divergence in the Eurasian sector [in accord with the trend found by Alexandrov et al. (2000)], but no significant trend in the Pacific sector, where the retreat has been greatest (Fig. 10). Clearly, ice advection is not the only factor driving the sea ice retreat in this sector. Note that the trend in sea ice divergence in each individual sea within a given sector is the same as that of the sector itself (results not shown). This result puts into question the idea of a potential predictability of the sea ice retreat on the regional scale. For instance, Rigor and Wallace (2004) found that the AO index can explain as much as 64% of the variance in summer sea ice extent in the Eurasian sector, while it explains less than 20% of the variance in the region north of Alaska. The Pacific region is greatly influenced by oceanic heat inflow through the Bering Strait (Bitz et al. 2005; Steele et al. 2010; Woodgate et al. 2010). Southerly winds associated with the Arctic dipole can also indirectly lead to enhanced surface ocean warming by pushing the ice out of the Arctic Ocean, which lowers the albedo and allows for more shortwave radiation input (Wang et al. 2009; Overland et al. 2012). The increase in sea ice velocity along the Alaskan coast favors the upwelling of warm summer Bering Sea waters due to Ekman pumping divergence (Zhang et al. 1998; Yang 2006). These factors can be amplified by positive feedbacks, such as the ice–albedo feedback (Holland et al. 2006; Maslanik et al. 2007; Perovich et al. 2007a) and the radiative feedback associated with the presence of melt ponds at the surface of first-year ice (Perovich et al. 2007b). Finally, changes in multiyear ice concentration explain over 50% of the variance in sea ice extent in the Beaufort/Chukchi region (Rigor and Wallace 2004). Multiyear ice, compared to thin first-year ice, is preferentially advected into the Pacific region when there is a strong Beaufort Gyre circulation. Since the intensity of the Beaufort high decreases with a more positive AO index, it is possible that, even though the correlation with sea ice divergence has recently been weak in this sector, the predictive value of the AO may nonetheless be significant, operating through a divergence of mass rather than area alone. However, this is beyond the scope of this study. Nevertheless, our results show no significant trend in the divergence for any of the model’s ensemble members during the twenty-first century (Fig. 8) that could relate to the predicted negative trend in the overall thickness of the Arctic ice cover (Holland et al. 2010), suggesting that the signal primarily comes from wind stress and not as much from the thickness distribution. Furthermore, since the early nineties, there has been no significant trend in the AO in the observational record (Overland and Wang 2005), and considering that the departures from the mean observed large-scale atmospheric circulation are of smaller amplitude than the ones from the GCMs previously analyzed, we expect a more subtle relationship between observed large-scale wind pattern and local divergence.
In this study, we analyzed the winter mean sea ice motion in order to gain insight on future patterns of sea ice retreat in the Arctic Ocean. The wintertime sea level pressure is directly linked to the large-scale mean atmospheric circulation, the main driver of sea ice motion. Two different GCMs were used in this analysis: CCSM4 has an important low bias (−7 mb) in the wintertime sea level pressure over the Arctic, characteristic of an extremely positive AO index, and CESM-LE presents a strong Beaufort Gyre circulation, with the center of the Beaufort high shifted toward the Eurasian coast, a pattern typical of a low-to-neutral AO index.
To quantify the potential for sea ice retreat, we looked at sea ice divergence from the Pacific and Eurasian sectors using our Lagrangian backtracking model. Sea ice divergence is related to formation of first-year ice, which is more likely to melt in the summer (Rigor et al. 2002; Rigor and Wallace 2004; Williams et al. 2016). The absence of a Beaufort Gyre circulation in CCSM4 leads to a net divergence of approximately zero in the Pacific sector, compared to convergence in CESM-LE. Divergence in the Eurasian sector is slightly more important in CESM-LE, along with a broad Beaufort high pushing the ice north of Siberia toward the North Pole. The positive trend in the AO index predicted by CMIP5 models (Gillett and Fyfe 2013) suggests that the Beaufort and Chukchi Seas could continue to lead in future sea ice retreat.
We do not claim that the mechanism applied here to two sets of climate output offers a full explanation of the dynamics underlying the future patterns of sea ice retreat. The determinants of sea ice extent are complex, and factors other than the large-scale atmospheric circulation have an important influence on determining its details, such as ocean heat inflow from the subarctic, small-scale atmospheric anomalies, vertical heat flux through Ekman pumping, and the interaction of ice growth/melt anomalies with the ice–albedo feedback. These mechanisms may also play an important role during the transition to a seasonally ice-free Arctic.
However, from related work (Rigor and Wallace 2004; Williams et al. 2016), the mechanism applied here to climate model output is known to be correlated with approximately 50% of the interannual variance in sea ice extent. Moreover, it rests on a straightforward and well-studied physical mechanism (Rigor et al. 2002), which is a strong argument for causation. Future work will include the study of mechanisms responsible for sea ice retreat that operate more locally, within particular subareas of the Arctic system. Nevertheless, this work, which links sea ice retreat to robust large-scale atmospheric patterns, constitutes an advancement toward better predictability of the future Arctic climate.
This work was funded by the Office of Naval Research (N000141110977). Patricia DeRepentigny is grateful to the Natural Sciences and Engineering Council of Canada (NSERC) and Fonds de Recherche du Québec–Nature et Technologies (FRQNT) for their financial support through research scholarships, as well as ArcticNet, Québec-Océan, and ArcTrain Canada. Bruno Tremblay is grateful for the financial support of ArcticNet, the NSERC Discovery Program, Environment Canada Grants and Contribution program, and the Canadian Sea Ice and Snow Evolution (CanSISE) Network funded by the NSERC Climate Change and Atmospheric Research program.
Additional affiliation: Lamont-Doherty Earth Observatory, Columbia University, Palisades, New York.