Abstract

Century-scale global near-surface temperature trends in response to rising greenhouse gas concentrations in climate models vary by almost a factor of 2, with greatest intermodel spread in the Arctic region where sea ice is a key climate component. Three factors contribute to the intermodel spread: 1) model formulation, 2) control climate state, and 3) internal climate variability. This study focuses on the influence of Arctic sea ice in the control climate on the intermodel spread in warming, using idealized 1% yr−1 CO2 increase simulations of 33 state-of-the-art global climate models, and combining sea ice–temperature relations on local to large spatial scales. On the Arctic mean scale, the spread in temperature trends is only weakly related to ice volume or area in the control climate, and is probably not dominated by internal variability. This suggests that other processes, such as ocean heat transport and meteorological conditions, play a more important role in the spread of long-term Arctic warming than control sea ice conditions. However, on a local scale, sea ice–warming relations show that in regions with more sea ice, models generally simulate more warming in winter and less warming in summer. The local winter warming is clearly related to control sea ice and universal among models, whereas summer sea ice–warming relations are more diverse, and are probably dominated by differences in model formulation. To obtain a more realistic representation of Arctic warming, it is recommended to simulate control sea ice conditions in climate models so that the spatial pattern is correct.

1. Introduction

High northern latitudes exhibit more warming in a climate with increasing greenhouse gas concentrations than other regions, indicating an amplified climate response in the Arctic. The amplified Arctic warming is visible in both observations and climate model simulations (Johannessen et al. 2004; Manabe and Stouffer 1980; Miller et al. 2010; Serreze and Barry 2011). In coupled models, the rate of Arctic annual mean warming (based on 1% yr−1 increase in atmospheric CO2 concentration) is projected to be roughly two to four times larger than the global mean response (Holland and Bitz 2003; Winton 2006) (Table 2). The intermodel spread in projected warming is also greatest in the Arctic, indicating a large uncertainty in the simulated climate response in the Arctic compared to any other region of the globe. Shortwave feedbacks (Hall 2004; Winton 2006), longwave feedbacks (Boé et al. 2009; Bintanja et al. 2011), and northward heat transport (Holland and Bitz 2003; Cai 2006; Ridley et al. 2007; Mahlstein and Knutti 2011) have all been shown to contribute to Arctic warming, but there is no consensus on their relative magnitude (Kay et al. 2012). The physical processes that contribute most to Arctic warming are not necessarily the same that cause the intermodel spread. In high northern latitudes, sea ice is thought to play a crucial role in amplifying local temperature changes (Screen and Simmonds 2010). Global climate models underestimate the trend in the annual minimum sea ice extent in the Arctic over the past decades (Stroeve et al. 2007; Winton 2011), although the newest generation of climate models seems to perform somewhat better on this point (Stroeve et al. 2012). Mahlstein and Knutti (2012) found that Arctic temperature and sea ice area are nearly linearly related, which can be used to estimate at what global temperature increase the Arctic will be ice-free in September. It is therefore imperative to elucidate the reasons behind the spread in both temperature and sea ice trends in the Arctic.

Three important aspects determine the climate pathway, and thus temperature and sea ice trends, of a single model simulation in response to (changes in) radiative forcing: 1) model formulation of physics and dynamics (including parameterizations), 2) control climate state, and 3) internal climate variability (Kattsov et al. 2010). If we want to understand the cause of the spread in climate response, the uncertainty that is related to all of these three aspects should be analyzed.

Uncertainty related to the model formulation of physics and dynamics is largest at high latitudes (Hawkins and Sutton 2009). Being a key component of the Arctic climate system, sea ice is expected to be an important factor in the temperature response in the Arctic through various feedback mechanisms. Intermodel differences in representations of the various sea ice–related feedbacks in climate models can significantly contribute to divergence of the projected temperature responses. Model features responsible for different climate responses are, however, not the focus of this study.

The second major factor that governs the response of climate models is the mean climate state from which a climate simulation is started, termed the control climate. The control climate is usually obtained from a reference model integration in which the forcing is kept constant over hundreds of years or longer. Each global climate model will simulate its own unique control climate conditions due to a combination of internal model physics and tuning. These conditions include the state of the atmosphere, ocean, sea ice, and land. Even though climate models are mostly tuned so that control conditions best match observational constraints, fundamental laws of nature (e.g., conservation of energy, mass, and momentum) must be obeyed. As a consequence, control climate conditions of Arctic sea ice vary considerably among global climate models (Flato 2004; Arzel et al. 2006). If the climate system is very sensitive to these control conditions, then projections of climate will exhibit an uncertainty associated with the varying control state.

The third source of spread in climate response is internal climate variability. Deser et al. (2012) show that, besides variability on interannual to decadal time scales, in some regions even trends over 50 years are subject to substantial uncertainty related to internal variability. Therefore, the influence of internal variability should be evaluated when externally forced climate change signals are assessed (Tebaldi et al. 2011). This applies especially for regions were natural variability is high, such as the Arctic. However, the relative importance of internal variability compared to model physics and dynamics in the spread of climate projections decreases over longer time scales (Hawkins and Sutton 2009).

As the largest intermodel spread in temperature response is in the Arctic, where sea ice is the key component, one would expect that sea ice is at least partially responsible for the modeled spread in warming and Arctic amplification. The role of the mean control climate state of Northern Hemispheric sea ice on the spread in simulated temperature response was examined in various modeling studies. For instance, Rind et al. (1995) found that global climate sensitivity is greater for more extensive or thinner sea ice cover. This finding was based on one particular model with a mixed layer ocean coupled to an atmospheric general circulation model. In a subsequent study, Rind et al. (1997) found that Northern Hemisphere sea ice thickness affects global climate sensitivity more than its extent. Holland and Bitz (2003) expanded upon this by making an intermodel comparison between atmosphere–ocean general circulation models, and focused on Arctic temperature amplification instead of global warming. They found enhanced Arctic amplification in models with thinner sea ice, which typically exhibit larger ice extent loss, suggesting a stronger ice-albedo feedback. Flato (2004) confirmed that models with thick ice in their control climate tend to produce less Arctic warming than models with thin ice. Mahlstein and Knutti (2011) also found an intermodel relation between Arctic warming and sea ice extent, linked by ocean heat transport differences between models. However, all of the above studies dealt mainly with the large-scale mean state of Arctic sea ice: its mean thickness and extent. The large-scale state of sea ice, however, does not give information about its spatial distribution, which may be relevant to explain different patterns of simulated warming, through local sea ice feedbacks. Also, differences in model physics might be more pronounced on smaller scales because these might be overwhelmed by large-scale processes if averaged over a large geographic region.

Our aim is to obtain a better understanding of the relation between control climate sea ice conditions and surface air temperature trends, by expanding upon previous studies with a multiscale analysis: varying from a classification based on local sea ice conditions to the entire Arctic. For the former, the sea ice cover is spatially divided into several sea ice classes based on its local sea ice thickness or concentration. On this scale, correlation analyses are applied to each model separately and provide more insight into local model physical processes that are related to spatial differences in near-surface warming. On the Arctic scale the focus is on the influence of differences in control climate sea ice on the intermodel spread in surface air temperature trends. On this spatial scale, the hemispherically integrated quantities of thickness and concentration, that is, volume and area, are studied. Correlation analysis is applied to the multimodel ensemble to get a better understanding of the cause of intermodel differences in Arctic near-surface warming.

2. Data and methods

Arctic surface air warming (at 2 m) and its relation to the control climate mean sea ice state and evolution are assessed, based on simulations of a multimodel ensemble of 33 atmosphere–ocean general circulation models (AOGCMs) and earth system models (ESMs). All of the included simulations contribute to phase 5 of the Coupled Model Intercomparison Project (CMIP5) (Taylor et al. 2012). The CMIP5 models and modeling groups that provided the data are listed in Table 1. For each of the models a set of two simulations with different forcing conditions was used: a control simulation with constant preindustrial (1850) CO2 concentration (piControl) and an idealized simulation with increasing CO2 concentration (1pctCO2). In the 1pctCO2 scenario, CO2 in the atmosphere was increased at a rate of 1% yr−1 from piControl CO2 concentration to about four times its control value at 140 yr. Only the first ensemble member of each model was included in the analysis. The 1pctCO2 scenario with a well-defined and large external forcing (“academic climate sensitivity experiment”) was chosen over more realistic future scenarios involving various forcing agents, as our aim is to better understand why climate models differ in their temperature response.

Table 1.

Model name, modeling center, and institution of the models used in this study.

Model name, modeling center, and institution of the models used in this study.
Model name, modeling center, and institution of the models used in this study.

Each CMIP5 model provides output on its native grid with its original land–sea mask, which quantifies the portion of the surface covered by land and by sea. To facilitate the grid pointwise intercomparison of models with output on a different grid, prior to analysis, all datasets were interpolated to a common 1° latitude × 1° longitude grid, using bilinear interpolation. By interpolating to a typical grid resolution of the ocean grid (on which sea ice output is stored) of CMIP5 models, 1° (Taylor et al. 2012), we try to minimize uncertainties that are associated with the interpolation. We did not apply a common land–sea mask for sea ice computations because this has a potentially large effect on the total computed sea ice volume or area (Massonnet et al. 2012).

We consider four characteristics of Arctic sea ice cover: thickness, concentration, volume, and area. Ice thickness is equal to the total ice volume in a grid cell distributed equally over the cell area, and Arctic ice volume is the integrated sea ice thickness across all grid cells in the Northern Hemisphere. Similarly, sea ice concentration is the fraction of the grid cell covered by sea ice, and sea ice area is the integrated sea ice concentration across the Northern Hemisphere. Mean thickness or concentration of a geographical region is computed by excluding grid cells that consist of only open ocean. We use sea ice area instead of extent because sea ice extent is generally based on a binary term that marks each grid cell as either ice covered or not ice covered, defined by a cutoff concentration to mark the ice edge (typically 15%), and does not give information about the real surface area that is actually covered with ice (Chevallier and Salas-Mélia 2012). Arguably, in studies that focus on sea ice–related warming feedbacks, total area is more relevant than extent (Blanchard-Wrigglesworth et al. 2011). Arctic surface air temperature is defined as the average surface air temperature over the region north of 70°N.

The piControl mean climate is defined as the 80-yr mean value of the piControl simulation. From the 1pctCO2 simulation, linear trends of the four sea ice components and surface air temperature are computed by ordinary least squares (OLS) regression. The linear regression over 140 years reveals the long-term systematic climate trend and reduces the effect of low-frequency natural climate fluctuations over (multi)decadal time scales (Vinnikov et al. 1999). Before computing the trends, each dataset was corrected for drift using the piControl simulation (Table 1).

3. Temperature trends and control sea ice state

In this section, we examine the surface air temperature trends and sea ice state of the set of models used in our analyses. First, we examine the near-surface temperature trends in the 1pctCO2 simulation, to get a clear picture of model similarities and differences therein. Next, the piControl Arctic sea ice cover is explored, both its multimodel mean state and individual model spatial characteristics.

a. Surface air temperature trends

Figure 1a shows the zonal annual-mean temperature trend near the surface in the 1pctCO2 simulation for the various CMIP5 models. The temperature trend in models GFDL-ESM2G and GFDL-ESM2M is small relative to other models because CO2 concentration was held constant after doubling its piControl value (Caldeira and Myhrvold 2013). Globally averaged, the annual-mean warming to a transient CO2 quadrupling ranges from 3.15 to 5.71 K across the remaining 31 global climate models used in this study, with a multimodel mean (±1 standard deviation) of 4.50 ± 0.73 K, and the magnitude of annual-mean Arctic warming is 11.69 ± 2.48 K (Table 2). The Arctic warming is amplified relative to the global warming by a factor ranging from 1.94 to 3.40 with a mean value of 2.59 ± 0.38 (Table 2). The spread (defined as the standard deviation) in the annual-mean Arctic temperature trend is more than three times larger than the global mean value.

Fig. 1.

(a) Zonal mean surface air temperature trends over 140 years in the 1pctCO2 simulation for the 33 global climate models, and the multimodel mean (black line). (b) Seasonal cycle of Arctic (70°–90°N) surface air temperature trends. Models GFDL-ESM2G and GFDL-ESM2M did not continue with increasing CO2 after reaching twice the piControl concentration (Caldeira and Myhrvold 2013) and are not included in the multimodel mean.

Fig. 1.

(a) Zonal mean surface air temperature trends over 140 years in the 1pctCO2 simulation for the 33 global climate models, and the multimodel mean (black line). (b) Seasonal cycle of Arctic (70°–90°N) surface air temperature trends. Models GFDL-ESM2G and GFDL-ESM2M did not continue with increasing CO2 after reaching twice the piControl concentration (Caldeira and Myhrvold 2013) and are not included in the multimodel mean.

Table 2.

Annual-mean global warming (ΔTglobal) and Arctic warming (ΔTarctic) to a transient CO2 quadrupling (at 140 yr), as computed from the temperature trends in the 1pctCO2 simulations. Arctic amplification (AA) is estimated by OLS linear regression of annual-mean Arctic temperature on global mean temperature, where the magnitude of Arctic amplification is then equal to the regression slope.

Annual-mean global warming (ΔTglobal) and Arctic warming (ΔTarctic) to a transient CO2 quadrupling (at 140 yr), as computed from the temperature trends in the 1pctCO2 simulations. Arctic amplification (AA) is estimated by OLS linear regression of annual-mean Arctic temperature on global mean temperature, where the magnitude of Arctic amplification is then equal to the regression slope.
Annual-mean global warming (ΔTglobal) and Arctic warming (ΔTarctic) to a transient CO2 quadrupling (at 140 yr), as computed from the temperature trends in the 1pctCO2 simulations. Arctic amplification (AA) is estimated by OLS linear regression of annual-mean Arctic temperature on global mean temperature, where the magnitude of Arctic amplification is then equal to the regression slope.

We use a 36-member multimodel ensemble of the 95-yr representative concentration pathway 8.5 (RCP8.5) scenario (Taylor et al. 2012) to estimate the effect of internal variability on intermodel spread in century-scale surface air temperature trends. Note that this is a different model ensemble than the one used in the rest of the paper. RCP8.5 is chosen over 1pctCO2 because it includes more members per individual model than 1pctCO2 and because the surface air temperature trend in RCP8.5 is of similar magnitude as in 1pctCO2. Spread in a multimodel ensemble is due to model formulation, control climate, and internal variability, whereas in an ensemble of an individual model spread is due to internal variability alone. By computing the ratio of the spread in an individual model’s ensemble to the spread in a multimodel ensemble with an equal number of members, an estimation of the contribution of internal variability relative to model formulation-plus-control climate is obtained. If the ratio is less than 0.5, internal variability has a smaller contribution to the intermodel spread than model formulation and control climate. The ensemble spread in four individual model ensembles with 5–10 members is compared with the mean spread of 1000 randomly chosen multimodel ensembles of the same size, with members selected from the 36-member multimodel ensemble. In this comparison the ratio of the four sets of model ensembles varies between 0.11 and 0.32, suggesting that intermodel variability does not dominate the intermodel spread in temperature trends. By taking a large multimodel ensemble (33 models) for our analyses, and computing trends over 140 years, we expect that the influence of the internal variability of an individual model on our results is reduced.

Multimodel mean warming is larger over the ocean than over land and is strongest near Franz Josef Land (Fig. 2). The intermodel spread is also larger over the ocean than over land, and the largest intermodel spread is found over the Barents Sea. This is an indication that sea ice presumably plays a role in the spread of near-surface temperature trends. The intermodel spread is greater in particular regions because the spatial pattern of annual mean surface air temperature trends in the Arctic varies considerably from model to model (Fig. 3). All models agree on strongest warming being located over the ocean. Although the majority of the models show peak warming concentrated near Franz Josef Land or over the Barents Sea, their exact location of maximum warming varies. Other models show peak warming on the opposite side of the Arctic near the Chukchi and/or Beaufort Sea (ACCESS1.0, ACCESS1.3, BNU-ESM, CanESM2, and FGOALS-s2). A few models typically have more northerly located maximum temperature trends over the central Arctic Ocean [CESM1(CAM5), FGOALS-g2, GFDL CM3, and HadGEM2-ES], whereas others have more southerly located maximum warming over the Norwegian or Greenland Sea (IPSL-CM5B-LR and MIROC-ESM). The seasonal cycle of surface air temperature trends (Fig. 1b) shows that annual mean warming is dominated by the winter season because the temperature trends in winter are larger than in summer. The intermodel spread in surface air warming is also largest in winter. In summer, all models simulate greatest warming over land, with minor warming over the central Arctic Ocean. Averaged over the Arctic, we found no intermodel relation between annual mean control climate temperature and temperature trends (r = 0.03, p = 0.87), either for winter or summer. This indicates that the intermodel spread in century-scale Arctic surface air temperature trends is not sensitive to differences in the Arctic mean temperature of the control climate, and the feedbacks related to rising CO2 concentrations govern the differences in the response.

Fig. 2.

Multimodel mean of surface air temperature trends [K (10 yr)−1] in the 1pctCO2 simulation, and its standard deviation. Note the differences in color scale. Models GFDL-ESM2G and GFDL-ESM2M are excluded.

Fig. 2.

Multimodel mean of surface air temperature trends [K (10 yr)−1] in the 1pctCO2 simulation, and its standard deviation. Note the differences in color scale. Models GFDL-ESM2G and GFDL-ESM2M are excluded.

Fig. 3.

Maps of surface air temperature trends [K (10 yr)−1] in the 1pctCO2 simulation, for each model separately. Models GFDL-ESM2G and GFDL-ESM2M did not continue with increasing CO2 after reaching twice the piControl concentration (Caldeira and Myhrvold 2013).

Fig. 3.

Maps of surface air temperature trends [K (10 yr)−1] in the 1pctCO2 simulation, for each model separately. Models GFDL-ESM2G and GFDL-ESM2M did not continue with increasing CO2 after reaching twice the piControl concentration (Caldeira and Myhrvold 2013).

b. Sea ice control state

Next we examine the sea ice state in the piControl climate. The multimodel mean sea ice state is characterized by high sea ice concentrations over the central Arctic Ocean (Fig. 4) and thick ice near the coasts of Greenland and the Canadian Archipelago (Fig. 5). The intermodel spread in sea ice concentration is largest in the Barents Sea and is related to the varying location of the ice edge in different models (Fig. 6). In some models the Barents Sea is annually averaged nearly ice free (e.g., ACCESS1.3, INM-CM4, and MIROC-ESM), whereas in other models sea ice extends into the Norwegian and/or Greenland sea [e.g., BCC-CSM1.1, BCC-CSM1.1(m), GISS-E2-H, GISS-E2-R, IPSL-CM5A-LR, IPSL-CM5B-LR, and MRI CGCM3]. The models generally agree more on the spatial distribution of sea ice concentration than on the distribution of sea ice thickness (Fig. 7). The spread in thickness is largest in the regions with thick ice and in the East Siberian Sea. The latter intermodel spread is caused by several models that simulate a second sea ice thickness maximum in the East Siberian Sea near the New Siberian Islands [e.g., BNU-ESM, CCSM4, CESM1(BGC), CESM1(CAM5), CNRM-CM5, CSIRO Mk3.6, MIROC5, NorESM1-M, and NorESM1-ME], whereas others only simulate thin ice at this location. In many models thick ice extends from the Canadian Archipelago into the Arctic Ocean north of the Beaufort Sea (e.g., ACCESS1.0, ACCESS1.3, HadGEM2-ES, INM-CM4, IPSL-CM5A-MR, MPI-ESM-LR, MPI-ESM-MR, and MPI-ESM-P).

Fig. 4.

Multimodel mean of sea ice concentration (%) in the piControl climate (80-yr mean value), and its standard deviation. Note the differences in color scale.

Fig. 4.

Multimodel mean of sea ice concentration (%) in the piControl climate (80-yr mean value), and its standard deviation. Note the differences in color scale.

Fig. 5.

As in Fig. 4, but of sea ice thickness (m).

Fig. 5.

As in Fig. 4, but of sea ice thickness (m).

Fig. 6.

Maps of mean sea ice concentration (%) in the piControl climate (80-yr mean value), for each model separately. Note that model BNU-ESM has in two quadrants of the Arctic almost no sea ice output.

Fig. 6.

Maps of mean sea ice concentration (%) in the piControl climate (80-yr mean value), for each model separately. Note that model BNU-ESM has in two quadrants of the Arctic almost no sea ice output.

Fig. 7.

As in Fig. 6, but of mean sea ice thickness (m).

Fig. 7.

As in Fig. 6, but of mean sea ice thickness (m).

Figure 8 shows the normalized intermodel spread in Northern Hemisphere control sea ice conditions for volume, thickness, area, and concentration, across the 33 global climate models. The average sea ice state in the piControl climate varies widely between models, as was also found for the historical simulation of CMIP5 models (Stroeve et al. 2012). Not surprisingly, models with a warmer Arctic control climate typically have a smaller sea ice area (r = −0.40, p = 0.02) and volume (r = −0.54, p = 0.01). A larger range in thickness and volume is found, compared to area and concentration. The ensemble-mean Northern Hemisphere sea ice volume is 25.2 ± 8.1 × 103 km3, and the mean sea ice area is 10.4 ± 1.4 × 106 km2. The relative magnitude of the intermodel standard deviation compared to the mean value (normalized standard deviation s) confirms that models tend to agree more on control surface area covered with sea ice (s = 13%) than on ice volume (s = 32%). This might be partly due to the land-locked Arctic Ocean, for which the adjacent continents act as a natural boundary to the extent of sea ice cover beyond which sea ice cannot expand (Bitz et al. 2005). The comparable large spread in Arctic mean sea ice thickness and volume illustrates that intermodel differences in piControl sea ice volume are governed by ice thickness rather than by ice area or concentration. Indeed, the intermodel correlation between volume and thickness is strong (r = 0.80, p < 0.01), whereas volume and area are not notably related (r = 0.12, p = 0.52).

Fig. 8.

The normalized intermodel spread of the Arctic sea ice state across the 33 global climate models in the piControl climate, averaged over 80 years. Arctic mean sea ice volume, thickness, area, and concentration are depicted, normalized by their multimodel mean value. The box-and-whisker plots show the maximum and minimum values among the models as well as the 25th and 75th percentile and the median. Models that deviate more than two intermodel standard deviations from the multimodel mean are depicted as outliers (open circles). The thickness outlier is EC-EARTH (thick ice). The area outliers are CSIRO Mk3.6 (large area) and MRI CGCM3 (large area). The concentration outliers are MIROC5 (high concentration), BNU-ESM (low concentration), and INM-CM4 (low concentration).

Fig. 8.

The normalized intermodel spread of the Arctic sea ice state across the 33 global climate models in the piControl climate, averaged over 80 years. Arctic mean sea ice volume, thickness, area, and concentration are depicted, normalized by their multimodel mean value. The box-and-whisker plots show the maximum and minimum values among the models as well as the 25th and 75th percentile and the median. Models that deviate more than two intermodel standard deviations from the multimodel mean are depicted as outliers (open circles). The thickness outlier is EC-EARTH (thick ice). The area outliers are CSIRO Mk3.6 (large area) and MRI CGCM3 (large area). The concentration outliers are MIROC5 (high concentration), BNU-ESM (low concentration), and INM-CM4 (low concentration).

Figure 9 depicts the seasonal cycle in control ice area and ice volume for each model. Most models agree on the month of maximum and minimum sea ice conditions. Generally, ice area reaches its maximum in March, and ice volume reaches its maximum in April. Minimum sea ice conditions are in August–September. The intermodel spread in control sea ice conditions is largest in winter for sea ice area. For sea ice volume, the intermodel spread is equally distributed throughout the year.

Fig. 9.

Seasonal cycle of (a) sea ice area and (b) sea ice volume in the piControl climate for the 33 global climate models, averaged over 80 yr. The colors for the models correspond with Fig. 1. The black line represents the multimodel mean.

Fig. 9.

Seasonal cycle of (a) sea ice area and (b) sea ice volume in the piControl climate for the 33 global climate models, averaged over 80 yr. The colors for the models correspond with Fig. 1. The black line represents the multimodel mean.

4. Local sea ice–temperature response

Now that we have established that the intermodel spread in surface air temperature trend is largest over the Arctic Ocean and piControl sea ice conditions vary widely between models, we study the local temperature response over ice-covered regions in the Arctic. We choose to study a summer and winter season separately because the Arctic has a large seasonal cycle in piControl sea ice cover (Fig. 9) and in 1pctCO2 temperature trends (Fig. 1). The summer season is defined as June–September (JJAS) and the winter season as December–March (DJFM) so that the extreme control sea ice conditions (Fig. 9) are included in the analysis.

To study the sea ice conditions and temperature trends locally, for each model, control sea ice cover was divided into 10 thickness or concentration classes. Figure 10 shows the distributions of sea ice concentration and thickness in the CMIP5 models for maximal sea ice conditions in March and minimal sea ice conditions in September. There is a much larger proportion of area with concentration lower than 10% and higher than 90% than in the other concentration classes. There is also a much larger proportion of area in the thinnest ice bin (0–0.5 m) compared to thicker ice classes. Over each class and for each month, the field mean surface air temperature trend was computed over 140 years.

Fig. 10.

Distribution of sea ice grid area over 10 sea ice bins across the 33 global climate models. The plots show the proportion of sea ice grid area per concentration class in (a) March and (b) September and per thickness class in (c) March and (d) September. We only consider the area of grid cells with a mean concentration or thickness greater than zero. The last thickness bin contains all thicknesses of 4.5 m and more, without an upper thickness limit. In bins with ice thicker than 2.0 m fewer models are considered since not all models simulate ice in all 10 thickness classes.

Fig. 10.

Distribution of sea ice grid area over 10 sea ice bins across the 33 global climate models. The plots show the proportion of sea ice grid area per concentration class in (a) March and (b) September and per thickness class in (c) March and (d) September. We only consider the area of grid cells with a mean concentration or thickness greater than zero. The last thickness bin contains all thicknesses of 4.5 m and more, without an upper thickness limit. In bins with ice thicker than 2.0 m fewer models are considered since not all models simulate ice in all 10 thickness classes.

a. Concentration–warming relation

First, we study the local relation between concentration and warming. Figures 11a and 11b depict for each model the field mean surface air temperature trend over each concentration class, for winter and summer. The annual mean correlation coefficient for the linear regressions between the concentration bins (5%, 15%, 25%, etc.) and field mean temperature trends over these bins averaged over all models is 0.97 ± 0.02 (p < 0.01), indicating that fields with a high ice concentration in the control climate typically warm more. The strength of the annual correlation between control concentration and warming is dominated by the winter season (r = 0.97 ± 0.04, p < 0.01) because the temperature trends during winter are much larger than during summer. The signs of the summer correlations are on average opposite to the winter correlations (r = −0.26 ± 0.61, p = 0.12 ± 0.20), with smaller temperature trends over bins with higher ice concentration. This is in agreement with largest warming being found over the ocean in winter, where ice concentration is higher, and over the subpolar seas and land in summer, where ice concentration is lower or absent. The large intermodel spread in summer is caused by oppositely signed relations of different models, with a distinction between models having a positive concentration-warming correlation (r = 0.67, p = 0.11 ± 0.20, 9 models) and models having a negative correlation (r = −0.61, p = 0.13 ± 0.21, 24 models).

Fig. 11.

Relation between surface air temperature trends in the 1pctCO2 simulation and piControl climate ice concentration bins averaged over (a) winter and (b) summer, and ice thickness bins averaged over (c) winter and (d) summer, for the 33 global climate models. The colors for the models correspond with Fig. 1. The black line represents the multimodel mean. The peak in the 40%–50% concentration bin belongs to the HadGEM2-ES model.

Fig. 11.

Relation between surface air temperature trends in the 1pctCO2 simulation and piControl climate ice concentration bins averaged over (a) winter and (b) summer, and ice thickness bins averaged over (c) winter and (d) summer, for the 33 global climate models. The colors for the models correspond with Fig. 1. The black line represents the multimodel mean. The peak in the 40%–50% concentration bin belongs to the HadGEM2-ES model.

We checked whether the unequal bin areas (Fig. 10) affected the spread in temperature trends around the bin averages of each model, that is, the spread between different grid cells belonging to a particular ice class. The multimodel mean spread around the bin average was not larger in bins with more grid area than in other ice bins, from which we conclude that bin area does not have a large effect on the bin spread of near-surface temperature trends.

Figure 12a shows the multimodel mean surface air temperature trends per concentration bin, split by month. Temperature trends are approximately linearly related to concentration classes, with increasing trends over higher concentration in winter and decreasing trends over higher concentration in summer months. The largest seasonal amplitude in surface air temperature trends is thus over ice with the highest concentration.

Fig. 12.

Relation between surface air temperature trends in the 1pctCO2 simulation and piControl climate sea ice bins for each month for (a) concentration and (b) thickness, averaged over all models.

Fig. 12.

Relation between surface air temperature trends in the 1pctCO2 simulation and piControl climate sea ice bins for each month for (a) concentration and (b) thickness, averaged over all models.

Now that we have an indication that control sea ice concentration relates to local near-surface temperature trends, we examine how sea ice concentration in a given class varies with near-surface temperature during the entire 140-yr simulation. Figure 13a shows the relation between 2-m temperature and sea ice concentration over time series of 140 yr for winter and summer, averaged over all models. Each line (10 per season) represents the average of all time series in one ice class, starting with the control mean ice concentration in one given bin at the start of the 140-yr model run (black crosses). As time evolves, grid points are defined to stay in the class from which they started, even though their ice concentration changes. The concentration–temperature curves of the different bins overlap and have similar shapes, showing that ice concentration relates nonlinearly with surface air temperature during climate change in a generic way. The strength of the local interaction between surface air temperature and concentration might be represented by the tangent line to each curve, which is basically the change in sea ice per degree Arctic warming. It appears that in both winter and summer, ice concentration is most sensitive to temperature perturbations between 50% and 60% ice concentration, suggesting that ice adjusts most to a certain temperature perturbation in this ice concentration range. Above the melting point of sea ice, ice concentration and temperature are almost not related because there is almost no ice left (<10%). For high ice concentrations (>80%), the sensitivity of sea ice concentration to temperature also becomes weaker, but is still present. Probably the relatively thick ice in high ice concentration classes delays the loss of concentration.

Fig. 13.

(a) Multimodel mean relations between surface air temperature and sea ice concentration during the 140-yr 1pctCO2 simulation. Each line represents the multimodel average of all 33 time series starting in one given bin (10 in total). Winter (blue lines) and summer (red lines) are plotted separately. The starting point of each ice class is labeled by a black cross. (b) As in (a), but for sea ice thickness.

Fig. 13.

(a) Multimodel mean relations between surface air temperature and sea ice concentration during the 140-yr 1pctCO2 simulation. Each line represents the multimodel average of all 33 time series starting in one given bin (10 in total). Winter (blue lines) and summer (red lines) are plotted separately. The starting point of each ice class is labeled by a black cross. (b) As in (a), but for sea ice thickness.

In summer months, we found that regions with a higher control climate ice concentration undergo less warming (Figs. 11a and 12a). Melting of sea ice in summer keeps the near-surface temperature near the melting point temperature of sea ice. This is clearly visible in Fig. 13a (red lines), which depicts the steepest slopes in ice concentration in summer near the melting point. As a consequence, in summer, surface air temperatures over melting sea ice hardly increase as long as sea ice is present, which is most effective in high concentration bins in the control climate, where sea ice takes longer to disappear. Temperatures stagnate near the melting point temperature while the ice melts and increase at a much faster rate after most ice has disappeared. In low control ice concentration bins, melting leads faster to open water areas that absorb more energy and cause more near-surface warming. In higher ice concentration classes, the summer sea ice does not disappear after 140 years (Fig. 13a) so that the temperatures do not exceed the ice melting point, and temperature trends are therefore small. This might explain why there is less warming in higher concentration classes. This is illustrated in Fig. 13a by the flattening of the concentration–temperature curve above the sea ice melting point: at low ice concentration hardly any additional ice can melt to keep the surface air temperature near the melting point, so surface air temperature goes up in each ice class where ice disappears. An additional effect that can explain the relation between control concentration and warming in summer is advection of warm surface air from the surrounding warmer land to the ocean. In summer, warming is larger over land, and this land-to-sea advection might be responsible for the peak warming near the ice periphery, where most models simulate a low control ice concentration.

In contrast to summer, in winter months there is more warming over higher ice concentration fields (Figs. 11a and 12a). Although one could think that the positive control concentration–warming relations only reflect latitude-dependent warming, and thus visualize indirectly Arctic amplification, the overlap between the concentration–temperature curves is an indication that the relation between concentration and warming is more than just a latitudinal dependence of both concentration and warming; the temperature–concentration pathway is almost independent of the control concentration. There will be more transfer of heat from the ocean to the atmosphere if more open ocean is exposed because sea ice acts as an insulator between ocean and atmosphere. We thus expect that in areas where more ice disappears, more ocean heat will be released to the atmosphere, thereby creating a larger surface air temperature trend. Figure 13a (blue lines) illustrates that high ice concentration classes have a larger decrease in ice concentration compared to lower concentration classes, and thus a larger increase in open ocean area. In addition, sea ice fields starting in a higher concentration class start with a lower near-surface temperature and keep a lower surface air temperature than lower concentration classes during the entire 140-yr simulation, but they also have a larger change in surface air temperature and thus a larger temperature trend (Fig. 13a, blue lines). Most warming in winter thus occurs over regions with higher ice concentration, which can be explained by a larger decrease in ice concentration and thus a larger change in the amount of heat that can be transferred to the atmosphere, warming the surface air. The potential warming by this mechanism, if all ice were to disappear in those regions, would be even larger because high ice concentration classes have more ice to melt.

In summary, it appears that sea ice concentration and surface air temperature interact during climate warming, but the type of interaction does not depend on the control ice concentration class from which the curve started because nearly all curves overlap and have a similar shape, which is nonlinear. The interaction does depend on season because there is an opposite relation between concentration and warming for summer and winter. An explanation for the negative concentration–warming relation in summer months is that melting prevents the near-surface air from warming, a mechanism that is effective longer over higher ice concentration classes because they exhibit more years of melting. In winter, the positive concentration–warming relation might be related to the insulating role of sea ice for the heat fluxes between ocean and atmosphere. However, both these hypotheses should be tested in further research before a definitive conclusion can be drawn.

Despite that the intermodel spread in Arctic mean sea ice concentration is small (Fig. 8), there are large intermodel differences in the temperature response within each single ice concentration class (Fig. 11). In both summer and winter, models with a relatively high temperature trend over low ice concentration typically also have a relatively high temperature trend over high ice concentration. In winter, the intermodel spread in the temperature trend increases with larger control climate ice concentration (Fig. 11a), whereas in summer the spread is about equal for all ice concentration classes (Fig. 11b). The greater spread in winter over higher ice concentrations means that most uncertainty in winter is over the central Arctic. In summer, the equally distributed intermodel spread in temperature response over each ice concentration class suggests that there is a common cause.

We compared the spread around the bin average with the intermodel spread to check if a regression through the bin averages is meaningful. This comparison shows that the multimodel mean spread around the bin average is of about the same size as the intermodel spread between the bin averages of the models, which implies that the relation between control concentration bins and surface air temperature trends of each individual model is not very robust. However, in winter, all models agree on the sign of the control concentration–warming relations, which is a strong indication that the control concentration–warming relations are more than just a coincidence. In addition, the clear seasonal cycle in control concentration–warming relations (Fig. 12a) corroborates these results.

b. Thickness–warming relation

The same analysis is repeated replacing concentration with thickness. Figures 11c and 11d depict the spread in the local temperature trend for 10 thickness classes in 0.5-m bins, for winter and summer. Averaged over all models, the correlation coefficient for the linear regressions between the centers of the thickness bins and the field mean surface air temperature trend is 0.58 ± 0.26 (p = 0.19 ± 0.27) (0.59 ± 0.27 when the eight models that do not have ice in all thickness bins are excluded). As for sea ice concentration, the relation between sea ice thickness and temperature trends varies by season. In winter, a clear distinction is visible in the temperature trends between regions with thin (<1.0 m) and thick (>1.0 m) sea ice, with smaller trends in temperature in the two thin ice classes (Fig. 11c). This result agrees well with our finding that regions with low ice concentration in winter have a relatively small temperature response because regions with thinner ice generally have a lower ice concentration. The winter mean correlation between the thickness classes and field mean temperature trends is 0.42 ± 0.37 (p = 0.27 ± 0.30), and positive for 28 models out of 33. In summer, as for concentration, temperature trends are almost independent of ice class (Fig. 11d). The summer mean correlation between thickness classes and field mean temperature is −0.13 ± 0.60 (p = 0.23 ± 0.29) and more or less equally distributed between positive and negative values with 18 out of 33 models having a negative correlation. Figure 12b depicts the monthly dependence of the control thickness–warming relation. For ice thinner than 1 m temperature trends do not exhibit a seasonal cycle; temperature trends of different months converge to one value. For ice thicker than 1 m trends do have a monthly dependence, but there is hardly any difference in the seasonal cycle of temperature trends for ice classes thicker than 1 m.

The relation between near-surface temperature and thickness during the 140-yr simulation is visible in Fig. 13b. As for ice concentration, all thickness–temperature curves overlap and have a similar shape. In classes with thick ice in the control climate, more ice melt takes place for a certain increase in temperature than over thin ice. This relation is termed the growth–thickness feedback by Bitz and Roe (2004). Thin ice need not thin much to adjust to a certain temperature perturbation, whereas thicker ice, which is generally over the central Arctic (Fig. 7), must adjust more.

Although the monthly control thickness–warming curves differ from the control concentration–warming relations (Fig. 12), their general behavior is the same. They exhibit more warming in winter over higher ice concentration/thicker ice and slightly less warming in summer over higher ice concentration/thicker ice. Also, they have a stronger seasonal dependence over thick ice classes/classes with a higher concentration. So the thickness results correspond roughly with the results for ice concentration, and the possible explanation is the same.

5. Arctic mean sea ice–temperature response

So far the analysis has focused on local sea ice–temperature relations at the sea ice–class scale. Now we analyze if relations at the sea ice–class scale are also recognizable at the Arctic mean scale. At the Arctic mean scale, spatial variations in concentration and thickness are averaged out, and their integrated quantities, area and volume, remain. We start by looking at the relation between control sea ice and near-surface temperature trends on the Arctic mean scale. Then the trends in volume and area and their relation with surface air temperature trends are assessed. In addition, we examine whether connections between the sea ice–class scale and Arctic mean scale exist.

In this section models are defined as outliers and excluded from the analyses when they deviate more than two intermodel standard deviations from the multimodel mean.

a. Control climate volume and area

First we examine the large-scale relation between control sea ice and Arctic warming (Fig. 14). The intermodel correlation between Northern Hemisphere control sea ice volume and Arctic warming is −0.27 (p = 0.15) (Fig. 14a), showing that there is no clear relation between ice volume in the control climate and Arctic warming. On the other hand, models with a larger control climate ice area tend to have smaller trends in Arctic surface air temperature (r = −0.36, p = 0.06) (Fig. 14b). On the local scale, warming is larger over ice with a higher ice concentration. Based on the local-scale results we would expect that, if sea ice physical processes play a dominant role in intermodel differences of Arctic warming, models with a larger sea ice area should warm more. The opposite seems to be the case, but the correlation is quite small. Also, this indicates that other processes, such as poleward ocean heat transport, play a more prominent role in intermodel differences of Arctic warming, while the physical processes of sea ice account for spatial differences within each single model. For example, enhanced ocean heat transport can lead to less sea ice, and therefore more energy can be absorbed during summer and radiated back during winter.

Fig. 14.

Arctic near-surface temperature trends in the 1pctCO2 simulation vs (a) ice volume and (b) ice area in the piControl climate for the models, excluding models GFDL-ESM2G and GFDL-ESM2M, and models that deviate more than two intermodel standard deviations from the multimodel mean (see Fig. 8). The numbers for the models correspond with Table 1.

Fig. 14.

Arctic near-surface temperature trends in the 1pctCO2 simulation vs (a) ice volume and (b) ice area in the piControl climate for the models, excluding models GFDL-ESM2G and GFDL-ESM2M, and models that deviate more than two intermodel standard deviations from the multimodel mean (see Fig. 8). The numbers for the models correspond with Table 1.

We find essentially no correlation between relative Arctic warming (Arctic amplification) and either control climate ice volume (r = 0.01, p = 0.94), or ice area (r = 0.04, p = 0.83). Holland and Bitz (2003) found a somewhat stronger relation between control climate ice thickness and Arctic amplification.

These weak correlations suggest that differences in the mean control sea ice state do not drive the intermodel spread in the Arctic warming or amplification signal. For winter and summer, none of the correlations of the control climate sea ice state with the Arctic temperature trend, or with Arctic amplification, is much different than its annual mean value.

b. Trends in volume and area

The relations between trends in sea ice and control climate sea ice conditions might provide more insight in the feedback mechanisms underlying the relation between control sea ice conditions and temperature trends. The intermodel correlation between control climate ice volume and trends in ice volume is strong and negative (r = −0.82, p < 0.01) (Fig. 15a), showing that models with more ice volume in piControl conditions exhibit a larger decline in ice volume, which is not surprising because a model cannot lose any additional ice in summer if the ice disappears earlier in one simulation compared to another. At the end of the 140-yr simulation, the Arctic Ocean is nearly free of summer sea ice in 20 models out of 32 (less than 1 M km2, Table 1). This is related to the growth–thickness feedback of sea ice discussed earlier; most ice disappears over initially thick ice, which results in most thinning in bins with thick ice, and hence in models with a larger control climate ice volume.

Fig. 15.

Sea ice volume trends in the 1pctCO2 simulation vs (a) ice volume and (b) ice area in the piControl climate for the models, excluding models GFDL-ESM2G and GFDL-ESM2M, and models that deviate more than two intermodel standard deviations from the multimodel mean (see Fig. 8). The numbers for the models correspond with Table 1.

Fig. 15.

Sea ice volume trends in the 1pctCO2 simulation vs (a) ice volume and (b) ice area in the piControl climate for the models, excluding models GFDL-ESM2G and GFDL-ESM2M, and models that deviate more than two intermodel standard deviations from the multimodel mean (see Fig. 8). The numbers for the models correspond with Table 1.

The intermodel correlation of control climate ice area and trends in ice area is negligibly small (r = 0.01, p = 0.97) (Fig. 15b), and comparably small for the winter and summer season, implying that there is no tendency for models with a larger control area to decline faster in area. We find, however, a strong positive correlation between control climate ice volume and trends in ice area (r = 0.53, p < 0.01), indicating that models with more piControl ice volume have smaller ice area trends. This relation probably shows that sea ice in models with more ice volume first has to thin before its concentration, and thus its area, decreases.

As a next step the intermodel relation between trends in sea ice and temperature trends is explored. Annually averaged Arctic surface air temperature trend is not related to ice volume trends (r = −0.07, p = 0.73) (Fig. 16a) or, hence, to mean thickness trends. The correlation is very weak for both summer and winter.

Fig. 16.

(a) Sea ice volume trends and (b) sea ice area trends, vs Arctic near-temperature trends, in the 1pctCO2 simulation for the models, excluding models GFDL-ESM2G and GFDL-ESM2M, and models that deviate more than two intermodel standard deviations from the multimodel mean (see Fig. 8). The numbers for the models correspond with Table 1.

Fig. 16.

(a) Sea ice volume trends and (b) sea ice area trends, vs Arctic near-temperature trends, in the 1pctCO2 simulation for the models, excluding models GFDL-ESM2G and GFDL-ESM2M, and models that deviate more than two intermodel standard deviations from the multimodel mean (see Fig. 8). The numbers for the models correspond with Table 1.

Arctic warming is strongly negatively related to ice area change (r = −0.79, p < 0.01) (Fig. 16b), indicating that models with more near-surface warming tend to have a larger decline in ice area. In those models with a larger decline in ice area, it is likely that more heat is absorbed during summer and radiated back during winter, thereby warming the surface air. This correlation is only slightly stronger in winter than in summer.

6. Discussion and conclusions

The aim of this study is to elucidate whether the projected ranges in century-scale Arctic temperature trends in transient climate change simulations are influenced by sea ice conditions in the control climate from which the simulations were initiated. Arctic warming is part of global warming and, although the Arctic region covers only a small part of the globe, it has a relatively large influence on the global-mean temperature response, as the magnitude of Arctic warming is about three times as large as global warming. This study uses the 1% yr−1 CO2 increase simulation of an ensemble of 33 state-of-the-art global climate models, all part of CMIP5 (Taylor et al. 2012). The greatest intermodel spread in the temperature response to increasing CO2 levels for the global climate models is located in the Arctic, in accordance with Holland and Bitz (2003) and Winton (2006). Understanding this spread is important for more reliable climate change projections. The main explanation for the spread is probably related to model formulation of physics and dynamics, or control climate conditions, because we estimated that natural variability is probably not the dominant cause of intermodel spread in century-scale projections. However, we recommend that the influence of natural variability on climate change projections should be analyzed more extensively in further research. We search for processes that explain geographical patterns of long-term Arctic warming in transient climate model simulations, and which of these processes can explain the intermodel spread on the Arctic mean scale, keeping in mind that the main cause of model-to-model differences in projections of Arctic warming can in principle be different from the main cause of Arctic warming itself. For this purpose, we combine physical understanding of surface air temperature trends on local and large scales into a consistent picture, using correlation analyses. Correlations may occur by chance, or may reflect common model deficiencies, but by attempting to explain them in terms of known physical processes they can give us more insight into underlying mechanisms related to Arctic warming.

To study the local sea ice–temperature response, for each model, control sea ice cover was divided into 10 thickness or concentration classes. On the sea ice class scale, for each model, the temperature trends and ice loss patterns depend on the control climate ice thickness and concentration distribution. Generally speaking, annually averaged, in high-ice regions there is more ice loss and more warming. Bitz and Roe (2004) gave an explanation for the larger ice loss in high-ice regions in terms of the negative growth–thickness feedback, which states that the growth rate of ice is inversely proportional to ice thickness. We found that the spatial concentration/thickness–warming relations are dominated by the winter season because winter warming exceeds summer warming, and summer has an oppositely signed concentration/thickness–warming relation with a much larger intermodel spread. A physical explanation for the enhanced warming in high-ice regions in winter is related to the insulating capacity of sea ice. When sea ice retreats, the relatively warm ocean releases heat to the atmosphere, thereby warming the surface air. This mechanism is expected to be more effective in regions with more ice loss. Also, inversion strength can strengthen the positive spatial relations between control climate sea ice concentration and near-surface warming in winter. In a spatial comparison, Pavelsky et al. (2011) show that wintertime sea ice concentration and inversion strength are positively correlated. Bintanja et al. (2011) show that the wintertime temperature inversion in the Arctic amplifies near-surface warming. Combining these two findings supports that inversion strength can enhance local amplification of near-surface warming over high ice concentrations in winter.

In summer, melting of sea ice keeps the near-surface temperature close to its melting point temperature, which prevents the surface air from warming as long as the ocean is ice covered. This might explain why summer warming is smaller over high-ice regions, whereas regions near the ice periphery exhibit more warming. Atmospheric advection of warm air from the surrounding land could also be the main cause of more warming over thinner ice (which is generally located near the ice edge) because summer temperature trends over the continents are larger than over the ocean.

The spatial distribution of the control climate ice concentration/thickness thus relates to the distribution of the local temperature response within each individual model in a seasonally dependent way. However, models do not agree on the magnitude of the temperature response for one particular control sea ice concentration or thickness class. In summer, a large part of the substantial spread in local sea ice–warming relations can likely be attributed to the formulation of the individual sea ice models. Although we did not look into details of sea ice models, this should be the focus of future studies. In winter, the relations are more universal among models, suggesting that the distribution of sea ice in the control climate plays a role in century-scale surface air temperature trends. However, it is complicated to separate the influence of control conditions from the influence of model formulations within a multimodel ensemble.

Remarkably, in contrast to control sea ice, control climate near-surface temperature has essentially no correlation with temperature trends in transient climate change. One should thus be careful with using model results in the present climate to inform future climate projections.

The largest intermodel spread in Arctic mean sea ice conditions in piControl climate is in sea ice thickness, which mainly determines sea ice volume, whereas models tend to agree better with respect to control sea ice area and concentration. Sea ice volume trends are strongly linked to ice volume in the control climate state, which can be understood from the spatial thickness–warming relation on local scales. This result is also found by Gregory et al. (2002), who showed that sea ice declined more rapidly in a global climate model when it was initially thicker and more extensive, and by Bitz (2008), who showed that intermodel spread in sea ice thickness decline in a CMIP3 model ensemble can be explained by spread in control climate thickness. A large intermodel spread in control climate sea ice volume can result in large intermodel differences in ice volume trends. To improve projections of sea ice volume loss, it is thus necessary to reduce spread in simulated sea ice volume in the control climate.

The intermodel spread in Arctic mean temperature trends is only weakly related to ice volume or area in the control climate. The weak correlation suggests that on the large scale other processes, such as ocean heat transport and meteorological conditions, play a more important role in the spread of century-scale warming than local sea ice processes. In addition, the intermodel spread in Arctic amplification is not correlated at all with control climate ice volume or ice area. Related to this, Holland and Bitz (2003) also found no significant correlation between control climate ice extent and Arctic amplification.

In conclusion, although there is no clear relation between control climate sea ice and temperature trends at the Arctic mean scale, there is a subtle dependency of surface air temperature trends on control sea ice conditions at the sea ice class scale. This means that it is important to simulate control sea ice conditions in a realistic way, not only at the Arctic mean scale, but also at the spatial scale, because this can lead to constraints for the climate response.

Acknowledgments

We acknowledge the World Climate Research Programme’s Working Group on Coupled Modelling, which is responsible for CMIP, and we thank the climate modeling groups (listed in Table 1) for producing and making available their model output.

REFERENCES

REFERENCES
Arzel
,
O.
,
T.
Fichefet
, and
H.
Goosse
,
2006
:
Sea ice evolution over the 20th and 21st centuries as simulated by current AOGCMs
.
Ocean Modell.
,
12
,
401
415
,
doi:10.1016/j.ocemod.2005.08.002
.
Bintanja
,
R.
,
R.
Graversen
, and
W.
Hazeleger
,
2011
:
Arctic winter warming amplified by the thermal inversion and consequent low infrared cooling to space
.
Nat. Geosci.
,
4
,
758
761
,
doi:10.1038/ngeo1285
.
Bitz
,
C. M.
,
2008
: Some aspects of uncertainty in predicting sea ice thinning. Arctic Sea Ice Decline: Observations, Projections, Mechanisms, and Implications, Geophys. Monogr., Vol. 180, Amer. Geophys. Union, 63–76.
Bitz
,
C. M.
, and
G.
Roe
,
2004
:
A mechanism for the high rate of sea ice thinning in the Arctic Ocean
.
J. Climate
,
17
,
3623
3632
,
doi:10.1175/1520-0442(2004)017<3623:AMFTHR>2.0.CO;2
.
Bitz
,
C. M.
,
M.
Holland
,
E.
Hunke
, and
R.
Moritz
,
2005
:
Maintenance of the sea-ice edge
.
J. Climate
,
18
,
2903
2921
,
doi:10.1175/JCLI3428.1
.
Blanchard-Wrigglesworth
,
E.
,
K. C.
Armour
,
C. M.
Bitz
, and
E.
DeWeaver
,
2011
:
Persistence and inherent predictability of arctic sea ice in a GCM ensemble and observations
.
J. Climate
,
24
,
231
250
,
doi:10.1175/2010JCLI3775.1
.
Boé
,
J.
,
A.
Hall
, and
X.
Qu
,
2009
:
Current GCMs’ unrealistic negative feedback in the arctic
.
J. Climate
,
22
,
4682
4695
,
doi:10.1175/2009JCLI2885.1
.
Cai
,
M.
,
2006
:
Dynamical greenhouse-plus feedback and polar warming amplification. Part I: A dry radiative-transportive climate model
.
Climate Dyn.
,
26
,
661
675
,
doi:10.1007/s00382-005-0104-6
.
Caldeira
,
K.
and
N.
Myhrvold
,
2013
: Projections of the pace of warming following an abrupt increase in atmospheric carbon dioxide concentration. Environ. Res. Lett.,8, 034039, doi:10.1088/1748-9326/8/3/034039.
Chevallier
,
M.
, and
D.
Salas-Mélia
,
2012
:
The role of sea ice thickness distribution in the Arctic sea ice potential predictability: A diagnostic approach with a coupled GCM
.
J. Climate
,
25
,
3025
3038
,
doi:10.1175/JCLI-D-11-00209.1
.
Deser
,
C.
,
R.
Knutti
,
S.
Solomon
, and
A. S.
Phillips
,
2012
:
Communication of the role of natural variability in future North American climate
.
Nat. Climate Change
,
2
,
775
779
,
doi:10.1038/nclimate1562
.
Flato
,
G.
,
2004
:
Sea-ice and its response to CO2 forcing as simulated by global climate models
.
Climate Dyn.
,
23
,
229
241
,
doi:10.1007/s00382-004-0436-7
.
Gregory
,
J.
,
P.
Stott
,
D.
Cresswell
,
N.
Rayner
,
C.
Gordon
, and
D.
Sexton
,
2002
:
Recent and future changes in Arctic sea ice simulated by the HadCM3 AOGCM
.
Geophys. Res. Lett.
,
29
, 2175,
doi:10.1029/2001GL014575
.
Hall
,
A.
,
2004
:
The role of surface albedo feedback in climate
.
J. Climate
,
17
,
1550
1568
,
doi:10.1175/1520-0442(2004)017<1550:TROSAF>2.0.CO;2
.
Hawkins
,
E.
, and
R.
Sutton
,
2009
:
The potential to narrow uncertainty in regional climate predictions
.
Bull. Amer. Meteor. Soc.
,
90
,
1095
1107
,
doi:10.1175/2009BAMS2607.1
.
Holland
,
M. M.
, and
C. M.
Bitz
,
2003
:
Polar amplification of climate change in coupled models
.
Climate Dyn.
,
21
,
221
232
,
doi:10.1007/s00382-003-0332-6
.
Johannessen
,
O. M.
, and Coauthors
,
2004
:
Arctic climate change: Observed and modelled temperature and sea-ice variability
.
Tellus
,
56A
,
328
341
,
doi:10.1111/j.1600-0870.2004.00060.x
.
Kattsov
,
V. M.
,
V. E.
Ryabinin
,
J. E.
Overland
,
M. C.
Serreze
,
M.
Visbeck
,
J. E.
Walsh
,
W.
Meier
, and
X.
Zhang
,
2010
:
Arctic sea-ice change: A grand challenge of climate science
.
J. Glaciol.
,
56
,
1115
1121
,
doi:10.3189/002214311796406176
.
Kay
,
J. E.
,
M. M.
Holland
,
C. M.
Bitz
,
E.
Blanchard-Wrigglesworth
,
A.
Gettelman
,
A.
Conley
, and
D.
Bailey
,
2012
:
The influence of local feedbacks and northward heat transport on the equilibrium Arctic climate response to increased greenhouse gas forcing
.
J. Climate
,
25
,
5433
5450
,
doi:10.1175/JCLI-D-11-00622.1
.
Mahlstein
,
I.
, and
R.
Knutti
,
2011
:
Ocean heat transport as a cause for model uncertainty in projected arctic warming
.
J. Climate
,
24
,
1451
1460
,
doi:10.1175/2010JCLI3713.1
.
Mahlstein
,
I.
and
R.
Knutti
,
2012
: September Arctic sea ice predicted to disappear near 2°C global warming above present. J. Geophys. Res.,117, D06104, doi:10.1029/2011JD016709.
Manabe
,
S.
, and
R. J.
Stouffer
,
1980
: Sensitivity of a global climate model to an increase of CO2 concentration in the atmosphere. J. Geophys. Res., 85 (C10), 5529–5554, doi:10.1029/JC085iC10p05529.
Massonnet
,
F.
,
T.
Fichefet
,
H.
Goosse
,
C. M.
Bitz
,
G.
Philippon-Berthier
,
M. M.
Holland
, and
P.-Y.
Barriat
,
2012
:
Constraining projections of summer Arctic sea ice
.
Cryosphere
,
6
,
1383
1394
,
doi:10.5194/tc-6-1383-2012
.
Miller
,
G. H.
,
R. B.
Alley
,
J.
Brigham-Grette
,
J. J.
Fitzpatrick
,
L.
Polyak
,
M. C.
Serreze
, and
J. W.
White
,
2010
:
Arctic amplification: Can the past constrain the future?
Quat. Sci. Rev.
,
29
,
1779
1790
,
doi:10.1016/j.quascirev.2010.02.008
.
Pavelsky
,
T. M.
,
J.
Boé
,
A.
Hall
, and
E. J.
Fetzer
,
2011
:
Atmospheric inversion strength over polar oceans in winter regulated by sea ice
.
Climate Dyn.
,
36
,
945
955
,
doi:10.1007/s00382-010-0756-8
.
Ridley
,
J.
,
J.
Lowe
,
C.
Brierley
, and
G.
Harris
,
2007
: Uncertainty in the sensitivity of Arctic sea ice to global warming in a perturbed parameter climate model ensemble. Geophys. Res. Lett.,34, L19704, doi:10.1029/2007GL031209.
Rind
,
D.
,
R.
Healy
,
C.
Parkinson
, and
D.
Martinson
,
1995
:
The role of sea ice in 2 × CO2 climate model sensitivity. Part I: The total influence of sea ice thickness and extent
.
J. Climate
,
8
,
449
463
,
doi:10.1175/1520-0442(1995)008<0449:TROSII>2.0.CO;2
.
Rind
,
D.
,
R.
Healy
,
C.
Parkinson
, and
D.
Martinson
,
1997
:
The role of sea ice in 2 × CO2 climate model sensitivity. Part II: Hemispheric dependencies
.
Geophys. Res. Lett.
,
24
,
1491
1494
,
doi:10.1029/97GL01433
.
Screen
,
J. A.
, and
I.
Simmonds
,
2010
:
The central role of diminishing sea ice in recent Arctic temperature amplification
.
Nature
,
464
,
1334
1337
,
doi:10.1038/nature09051
.
Serreze
,
M. C.
, and
R. G.
Barry
,
2011
:
Processes and impacts of Arctic amplification: A research synthesis
.
Global Planet. Change
,
77
,
85
96
,
doi:10.1016/j.gloplacha.2011.03.004
.
Stroeve
,
J. C.
,
M. M.
Holland
,
W.
Meier
,
T.
Scambos
, and
M.
Serreze
,
2007
:
Arctic sea ice decline: Faster than forecast
.
Geophys. Res. Lett.
,
34
, L09501,
doi:10.1029/2007GL029703
.
Stroeve
,
J. C.
,
V.
Kattsov
,
A.
Barrett
,
M.
Serreze
,
T.
Pavlova
,
M.
Holland
, and
W. N.
Meier
,
2012
:
Trends in Arctic sea ice extent from CMIP5, CMIP3 and observations
.
Geophys. Res. Lett.
,
39
, L16502,
doi:10.1029/2012GL052676
.
Taylor
,
K. E.
,
R. J.
Stouffer
, and
G. A.
Meehl
,
2012
:
An overview of CMIP5 and the experiment design
.
Bull. Amer. Meteor. Soc.
,
93
,
485
498
,
doi:10.1175/BAMS-D-11-00094.1
.
Tebaldi
,
C.
,
J. M.
Arblaster
, and
R.
Knutti
,
2011
:
Mapping model agreement on future climate projections
.
Geophys. Res. Lett.
,
38
, L23701,
doi:10.1029/2011GL049863
.
Vinnikov
,
K. Y.
, and Coauthors
,
1999
:
Global warming and Northern Hemisphere sea ice extent
.
Science
,
286
,
1934
1937
,
doi:10.1126/science.286.5446.1934
.
Winton
,
M.
,
2006
: Amplified Arctic climate change: What does surface albedo feedback have to do with it? Geophys. Res. Lett.,33, L03701, doi:10.1029/2005GL025244.
Winton
,
M.
,
2011
:
Do climate models underestimate the sensitivity of Northern Hemisphere sea ice cover?
J. Climate
,
24
,
3924
3934
,
doi:10.1175/2011JCLI4146.1
.