The data used in this study include measured temperature in four boreholes in Antarctica. We refer to them as West Antarctic Ice Sheet – “WAIS”, “Larissa”, “Mill Island”, and “Styx”. Figure 1 and Table 1 provide their locations and corresponding references. The borehole temperature profiles were sampled in January 2008 and January 2009 (WAIS), December–February 2009/10 (Larissa), the summer of 2009/10 (Mill Island), and the summer of 2014/15 (Styx). As shown in Fig. 1 (in red rectangles), the borehole temperature is affected by the seasonal cycle in the upper 15 m (Bodri and Cermak, 2011, chap. 1), which is not adequate for the reconstruction of annual mean surface temperature. Consequently, only depth under 15 m is used to reconstruct the surface temperature history and to compare with simulated borehole temperature profiles.

The temperature reconstructions and their uncertainty estimates for the four boreholes are shown in Fig. 1. For WAIS, Mill Island, and Styx, the reconstructed surface temperature series (Fig. 1a, c, d) are computed using a generalized least-squares algorithm (e.g., Orsi et al., 2012). For Larissa, the surface temperature is recovered by the Tikhonov regularization algorithm (Zagorodnov et al., 2012). This method has been proven to be valid for inverse problems such as the reconstructions based on borehole temperature observations, and the details of this method are explained in Nagornov et al. (2001, 2006). Since the temperature reconstructions are sensitive to the technique used, when we drive the borehole temperature model selected in this study by the published reconstructed temperature histories and compare them to the observed borehole temperature, differences are found. They are likely attributed to the different methodology and hypothesis. However, they are relatively small (Fig. S6 in the Supplement), suggesting that they do not have a major impact on the final conclusions.

The simulated surface air temperature used in this study (Table 2) is extracted from general climate model (GCM) simulations covering the past millennium performed in the framework of the third phase of the Past Model Intercomparison Project (PMIP3; Otto-Bliesner et al., 2009) and the fifth phase of the Coupled Model Intercomparison Project (CMIP5; Taylor et al., 2012). These simulations cover the period 850–1850,CE (referred to as the past1000 experiment in CMIP/PMIP nomenclature) and the years 1850–2005 CE (historical period). For the majority of the models, the simulations start thus in 850 CE and finish in 2005 CE. However, for two of the models, CESM1-CAM5 (Otto-Bliesner et al. 2016) and MPI-ESM-P (Stevens et al. 2013), the historical simulations covering 1851–2005 CE were performed independently of the simulations covering 850–1850 CE. In order to obtain results over the full millennium, we adopt the approach from Klein and Goosse (2018) and merge the first ensemble members (r1i1p1) of the past1000 experiment with the corresponding ensemble members of the historical experiment. Although not continuous, there is no large discrepancy in 1850 CE between the two merged simulations (e.g., Klein and Goosse, 2018).

These simulations are driven by natural (orbital, solar irradiance, volcanic) and the anthropogenic (well-mixed greenhouse gases, ozone, aerosols, land use/land cover) forcings (Schmidt et al., 2011, 2012). Note that BCC-CSM1-1 and IPSL-CM5A-LR ignore the impact of land use/land cover, and IPSL-CM5A-LR does not consider any variations in aerosols and tropospheric ozone. Further description of the simulations and the forcing can be found, for instance in Klein et al. (2016). For CESM1-CAM5, it produces 12 different simulations with the same physics and same input forcings but slightly different initial conditions in the model. Therefore, the differences between ensemble members attributable to the process internal to climate system provide an estimate of the internal variability. For CCSM4, GISSE2-R, IPSL-CM5A-LR, MPI-ESM-P, and BCC-CSM1-1, there is only one simulation available. In addition, although we can obtain the simulated surface mass balance (SMB) from these models (e.g., Dalaiden et al., 2020), we do not use it here and keep the observed accumulation rate in the forward model, since biases in the simulation of SMB may affect our conclusions, and the focus here is on the simulated temperature evolution.

The forward model used herein to simulate the propagation of the signal coming from the surface temperature history into the subsurface is based on the one-dimensional heat and ice flow equation (Alley and Koci, 1990): 1ρcp∂T∂t=∂∂zk∂T∂z-ρcpw∂T∂z+Q,

where T is the temperature, t is the time, cp is the heat capacity, ρ is the density of firn/ice, z is the depth, w is the downward velocity of the firn/ice, and Q is the heat production term. In the Eq. (1), the term on the left side represents the change in heat content. On the right side, the first term corresponds to the rate of temperature change due to conduction based on Fourier's law. Ice moving vertically (z direction) with downward velocity, w, conveys a heat flux ρcpwT across a plane of unit area, oriented perpendicular to z, which is accounted for in the heat transfer by advection shown as the second term. The third term, Q, consists of two parts: (1) ice deformation (Cuffey and Paterson, 2010, chap. 9, Eq. 9.30) and (2) firn compaction (Cuffey and Paterson, 2010, chap. 9, Eq. 9.33). Important model parameters are obtained from the references given in the Table 1, and they are summarized in Table 3. A detailed description of the model is available in the Supplement.

According to the original studies describing the records and the surface temperature reconstructions, the various parameters in the forward model have effects of different magnitude on the results for the different sites. Consequently, in order to assess the uncertainty in the model–data comparison related to the parameters of the forward model, we perform a series of sensitivity experiments on the parameters which have been shown to have the largest effects on each of the borehole profiles shown in the Fig. 2.

At WAIS, the spread of the sensitivity tests is lower than the spread in the simulated borehole profiles driven by different climate model results (solid lines in color in Fig. 2a and b). However, the initial temperature derived from a steady-state profile has an influence on the slope of the profile in the deeper part and on the depth of the temperature minimum, contributing to the uncertainty in the intensity of the pre-1900 cooling trend and the timing of the temperature minimum.

At Larissa, the effect of the bottom boundary conditions is important in setting up the temperature gradient from the bottom to 300 m, and therefore, we will not interpret that segment of the data in terms of climate. It is also evident in Fig. 2c that the different temperature histories produce a very similar depth profile over that interval.

At Mill Island, the borehole profile is shallow and covers only a fraction of the full thickness of the ice sheet. At sites with such a deep ice sheet and with a high accumulation rate, the optimal surface temperature history was found to be essentially independent of the location of the imposed bottom boundary condition for depths in excess of 180 m below the surface (Roberts et al., 2013). Consequently, here we modeled the temperature by assuming a zero-heat-flux bottom boundary. Although the initial temperature has an influence on the slope of the profile deeper than 120 m, this sensitivity is weak in depths shallower than 80 m, and the borehole profile is dominated by the surface temperature history.

At Styx, the bottom boundary condition is adjusted to reproduce the slope of the temperature profile in the deeper part (100–200 m). The simulated borehole profiles driven by GCMs (solid lines in the Fig. 4e) show the large deviation in the top 100 m compared to stationary temperature profile, which suggests that there is climate information stored in the upper part of the profile. Meanwhile, at depths shallower than 50 m, the effect of boundary conditions is weaker than the differences in the temperature histories from the different models, which means the borehole temperature data can be used to discriminate between temperature histories provided by the different models.

The internal variability also has a significant impact on the shape of the simulated borehole profiles. At these four sites, the range of simulations driven by CESM ensemble is much larger than the range of the different sensitivity tests in the top 50 m (shown as the shaded area in Fig. 4b, d, f, and h). This confirms that internal variability is a dominant source of uncertainty in a model–data comparison, at least in the top 50 m. For the deeper part of WAIS, as the shape of subsurface temperature profiles is influenced by the parameters of the forward model, the evaluation of the long-term cooling trend is more uncertain.

Figure 3 displays the comparisons between climate model results and temperature reconstructions from the boreholes. The simulated temperatures displayed in Fig. 3 come directly from the surface temperature calculated by the climate model, based on its own dynamics and the forcing applied as discussed in Sect. 2.2. In order to ensure that the climate model results have the same mean over the reference period (which is the whole period derived from the reconstruction) as the reconstruction, we applied a very simple, constant correction to remove the mean bias of the climate model results as shown in the Fig. 3. Due to the nature of physical diffusion, the heat propagation acts similarly to a low-pass filter. The reconstructions thus suffer from an attenuation of high-frequency temperature variability that becomes stronger as time goes back (Beltrami et al., 2006; Harris and Gosnold, 1999). For instance, in the reconstructed surface temperature of Styx, the point corresponding to 1800 CE in the curve may represent an average temperature between around 1600 and 1900 CE, while in 1900 CE it corresponds to an average over around 200 years. This characteristic complicates the model–data comparison. Therefore, in order to facilitate the comparison between the reconstruction and climate model results, we use variable smoothings to mimic the characteristic as much as possible. Since the reconstructions have much wider ranges than those ones from the climate model results, the basic compatibility between model and data will not be changed due to various smoothing. Nevertheless, Fig. 3 must be interpreted carefully because of this inhomogeneous smoothing.

Because of the internal variability in the system, a single simulation without error bound is not expected to reproduce well all the characteristics of the observed variations. The difference can be large, in particular at the local level (e.g., Goosse et al., 2005), but the observations should correspond to a credible member of an ensemble of simulations. Ensuring this compatibility can be achieved using various techniques, but the first step is to simply check if the reconstruction is within the range provided by the ensemble (e.g., PAGES2k-PMIP group, 2015).

Considering the large uncertainty range in these reconstructions, the climate models are visually able to reproduce the general characteristics of reconstructed temperature variability, particularly the long-term cooling during the last millennium and the recent warming (Figs. 3 and 4). Nevertheless, disagreements have also been identified.

The first major feature in the data is this long-term cooling trend, visible at the WAIS and Larissa sites. At Larissa, the borehole temperature reconstruction gives a cooling trend of -0.94±0.12 ∘C per century from 1825 to 1925 CE (Zagorodnov et al., 2012). None of the models are able to reproduce this observation, and instead, they all show a warming trend of comparable magnitude (Figs. 3c and 4c). At WAIS, the borehole temperature inversion also shows a long-term cooling trend, from 1000 CE to about 1600 CE, with a magnitude of -0.10±0.07 ∘C per century (Fig. 3a). The large uncertainty in the long-term trend is principally due to the uncertainty in the initial surface temperature (Fig. 2a; Orsi et al., 2012, their Fig. 3). The quantitative comparison between the trend of reconstructions and climate model outputs (Fig. 4a) indicates that the simulations generally show a cooling trend over 1000–1600 CE, in agreement with previous studies (e.g., Goosse et al., 2012; Abram et al., 2016; Klein et al., 2019). The amplitude of the trend is lower, particularly for GISS (-0.01 ∘C per century) and IPSL (-0.03 ∘C per century) models, but most remain within the lower end of the reconstructed uncertainty range. This long-term cooling trend is a feature of the Antarctic climate that is visible in many other ice core records (Stenni et al., 2017). A recent compilation of PAGES Antarctica2k datasets calculated a trend of -0.26 to -0.4 ∘C per 1000 years for the period 0–1900 CE for the Antarctic continental average (Stenni et al., 2017). In the high latitudes of the Southern Hemisphere, the origin of this millennial-scale cooling is currently not well understood, but an intermediate complexity model has shown multimillennial cooling in summer because of a delayed response to the decrease in local spring insolation (Renssen et al., 2005) with an additional potential influence of volcanic forcing (Goosse et al., 2012; Abram et al., 2016; Stenni et al., 2017).

A second feature of the data is a warming trend in the 20th century, which started at different times in the different records. Styx shows an early warming trend from 1900 to 1980 CE and a general stabilization of the temperature afterwards (Fig. 3h). This signal is consistent with the data from weather stations and ice core isotope-derived records (Yang et al., 2018). Models tend to show the opposite timing, with nearly no trend from 1900 to 1960 CE and a late warming trend that differs in amplitude between models. Overall, the simulated warming of the 20th century is about half of what is observed (Fig. 4f), with BCC (1.63 ∘C per century) and CCSM4 (1.23 ∘C per century) having the largest trends, closest to the observations (1.81 ∘C per century).

Larissa shows a temperature minimum in 1940s, followed by a steady warming trend until around 1995 CE. The magnitude of the 20th-century trend is 1.99 ∘C per century. Most models reproduce the timing of the warming reasonably well, with the exception of MPI, which shows an early warming, but no trend in 1940–2005 CE, and GISS, which has a very muted trend. If the trend present in the other models is too low, it seems rather due to a lack of cooling in the preceding century than because of errors in the latest decades.

Mill Island shows a late warming trend starting in the 1980s. Models tend to overestimate this trend (Fig. 4e), in particular IPSL, BCC, and CCSM4. Similarly to Mill Island, WAIS also shows a positive trend over the period 1900–present that intensifies after 1980 CE. The amplitude of the 20th century warming (0.53 ∘C per century) is well simulated, but the start of the trend occurs sometimes too early, with the exception of CESM, BCC, and IPSL, which show a late warming trend (Fig. 3b).

Overall, for WAIS (Fig. 4b) and Larissa (Fig. 4d), the reconstructed trends lie in the CESM ensemble range, suggesting many apparent model disagreements for those sites can be due to internal variability. For Styx (Fig. 4f) and Mill Island (Fig. 4e), the reconstructed trends are larger than the spread of the CESM ensemble, which means the disagreements are not only due to internal climate variability but also related to a systematic climate model bias in this region.

However, as stated above, borehole temperature reconstructions are “underdetermined”, which means that there are many possible temperature histories that can fit the data (more detailed explanation of underdetermined is given in the Introduction). The next step is to determine if the differences between simulated and reconstructed time series can be discriminated when analyzing observed and simulated temperature profile.

The simulated subsurface temperature profile is the results of the superposition of two components: (1) the initial temperature profile that incorporates the effects basal heat flux and vertical advection due to ice accumulation and (2) the subsurface temperature deviations arising from the surface temperature variability. Since the initial temperature profile for each borehole is obtained by driving the forward model with the optimal parameters obtained from the original publications (see Sect. 2.4), the differences among the simulated borehole profiles for each location are caused only by the changes in the upper boundary, i.e., in the climate model outputs. The simulated subsurface temperature profiles for each borehole are displayed in Fig. 5.

As previous studies have shown (e.g., Bodri and Cermak, 2011, chap. 2), a “U”-shaped subsurface temperature profile is a direct evidence for the past climate change with a minimum that separates the deeper warming trend due to geothermal heating and shallower warming trend related to a recent temperature increase (Orsi et al., 2012; Stevens et al., 2008). Among these four sites, WAIS and Larissa have such characteristics of a U shape. For Mill Island, this is less clear, but a significant breaking point in each simulated subsurface temperature profile reflects the surface temperature warming over recent decades. For Styx, such a break does not seem to be present at all, and the slope does increase with depth.

Aided by these key properties, we can identify a link between the interpretation in the depth domain and in the time domain. The analysis of the simulated and observed temperature profile confirms the main conclusion obtained in Sect. 3.1, in particular the agreement between model and data on the general tendencies, characterized by a long-term cooling trend over last millennium and the recent warming. For the deeper part of the profile, the simulated temperature profiles driven by MPI, IPSL, and GISS at WAIS almost coincide with the corresponding observation, but they fail to reproduce the depth of the temperature minimum around 120 m in the data. This is consistent with the fact that IPSL and MPI are at the edge of the reconstructed cooling trend of the last millennium, and GISS presents a significant underestimation of this trend (Fig. 4a). However, the CESM ensemble follows the borehole temperature profile (shaded area in Fig. 5a) and can also reproduce the magnitude of the cooling trend for some of the members (Fig. 4a). Specifically, the minima in the simulated profiles driven by MPI, IPSL, GISS, and CESM are -30.06, -30.06, -30.07∘, and a range of -30.8 to -30.17∘, respectively, which is very close to the minimum of -30.08∘ in the observation.

At Larissa, the bottom (270–450 m) of the profile is controlled by boundary conditions (Fig. 2c) and contains no climate information, as demonstrated by the fact that all curves are on top of each other in Fig. 5c. Additionally, no simulation has a pronounced inflection point around 170 m as shown in the observation. These characteristics are perfectly consistent with the lack of a cooling trend from mid ∼ 19th century to the early 20th century in the simulations (Fig. 3c). We conclude from this that the cooling trend of 1825–1925 CE is a robust feature in the data that can be used to benchmark climate models.

For the recent warming, we see some significant discrepancies among the simulated subsurface temperature profiles driven by different climate models at the four boreholes in the depth domain that are consistent with the signal analyzed in the time domain. For WAIS, in the uppermost part, the simulated subsurface temperature profiles driven by GISS, CCSM4, and BCC display significantly higher temperature than those in the observations, while IPSL and MPI-simulated profiles are close to the observations (Fig. 5b). This is in perfect agreement with the too high temperatures in models compared to the reconstructions in the second half of the 20th century (Fig. 3b).

For Larissa, all simulated profiles display an increasing temperature toward the surface as in observations but with different magnitude and shape (Fig. 5c). The temperature in the simulation driven by MPI displays a relatively rapid increase until around 100 m and then is constant, which is consistent with the near-constant temperature from 1940 to 2005 CE (Fig. 3d). For the ones driven by CCSM4 and BCC, they are warmer than the observation between depths of 15 and 50 m, which reflects the consistently warmer temperature shown in Fig. 3d. The IPSL-simulated subsurface temperature profile displays the largest similarity to the observations, while the simulations performed with CESM can cover almost all the observation in the shallow zone.

For Mill Island, the simulated subsurface temperature profiles are warmer than observations above 50 m, confirming the too large a warming trend deduced from the analysis of surface temperature. In particular, the IPSL model has the largest warming trend (Fig. 3e and f) and also has the warmest temperature profile (Fig. 5e and f), followed by MPI. For Styx, the main discrepancies occur over the shallow depths, between of 15 to 60 m, where all the simulations depict colder conditions compared with observations (Fig. 5g and h), as for the surface temperature over the recent decades in Fig. 3.

We also find in the depth domain some signals that are not obvious in the time domain. In particular, for WAIS, one of the CESM runs matches the warming trend of the top 100 m, while in time domain the CESM ensemble is significantly colder than reconstruction over recent decades. The CESM outputs generally follow the data in the deeper part of the profile (200–300 m) and have an even steeper slope between 100 and 200 m (Fig. 5), while in the time domain, the cooling trend was underestimated (Fig. 4a). In addition, for WAIS, the simulated subsurface temperature profiles driven by CCSM4 and BCC over the deeper part of the profile are colder than observations, but the warming trend starts deeper, at about 200 m compared to 120 m in the observations. This seems puzzling because, in the time domain, the cooling trend continues until 1800 CE for CCSM4 (Fig. 2a, yellow). However, the larger warming in the last 100 years is probably shifting the temperature minimum downwards. This example shows that it is difficult to pinpoint the date corresponding to a temperature minimum in the depth profile, because it depends on the respective speed of warming and cooling before and after. At Mill Island, in the deeper part (around from 140 to 100 m) of the profile, the simulated subsurface temperature profile driven by IPSL is very different from the other ones, with a slightly decreasing temperature and a colder climate than observations. However, in the time domain, the difference compared to other time series for IPSL was much less clear (Fig. 3e), but the consistency between these two domains still exists, and especially the temperature minimum in 1980 CE might correspond to the deeper part (around 100 m) in the depth domain.

The comparison between the analyses in the two domains appears thus complementary and instructive as it illustrates that the interpretation may be easier in one case or the other. It also shows that the observations can help evaluate the models by comparing different borehole temperature profiles driven by the different climate model results with the corresponding observation. In particular, the analysis of the simulated temperature profiles confirms that CESM ensemble can reproduce the multidecadal and centennial climate variability at WAIS.

Of the four records presented here, WAIS has the longest retrievable history. We propose here to use the temperature trend of the period of 1000 to 1600 CE as a metric, with the magnitude of -0.102±0.07 ∘C per century (Fig. 4a). The end of the cooling trend is not clearly defined by the data, due to the complex time-varying smoothing of the borehole temperature record, but 1600 CE seems to be safely in the cold interval (See Orsi et al., 2012, for details). The start of the period is more open, and we chose 1000 CE to be compatible with last-millennium simulations. External evidence from a compilation of water isotope records indicates that the cooling trend extended likely from 0 to 1900 CE in many parts of Antarctica (Stenni et al., 2017). It is a robust feature of the Antarctic climate of the last 2 kyr, and the WAIS record is unique in providing a clear quantification of the temperature trend.

In Fig. 6, we show the 1000 to 1600 CE surface temperature trend at WAIS and at other sites in Antarctica from the models output. Visually, for most simulations, the cooling at the grid cell of WAIS is similar to the one obtained at many locations in West Antarctica. Only the first member of CESM shows a small warming trend in West Antarctica. The large spatial coherence of the trend indicates that, although we are making a single-point comparison, it represents a signal common to a large part of the continent. It is also important to estimate the magnitude of the trend at WAIS compared to other regions. To do so, we calculate the ratio of the trend of surface temperature from 1000 to 1600 CE at any location with the one at WAIS (Fig. 7). Except the first member of CESM, if the value is greater than 1 (shown in red tones), it means the trend at the grid cell is larger than that at WAIS; if the value lies between 0 and 1 (shown in blue tones), it means the trend at the grid cell is less than that observed at WAIS. Negative values (i.e., a trend of a different sign compared to WAIS) are not shown, and the corresponding region is blank. Since the goal of Fig. 7 is to show the intensity of cooling at WAIS compared with other points in Antarctica, the first member of CESM 1, which shows a warming trend close to zero at WAIS, is not very meaningful, but it is still included for completeness. Seventy-five percent of models show that WAIS displays larger cooling from 1000 to 1600 CE than other locations in Antarctica (shown in blue) but with magnitude similar to other grid cells in West Antarctica. This is consistent with the reconstruction of Stenni et al. (2017) that shows the largest cooling in this region over the period 0–1900 CE. The spatial patterns of the trends (Fig. 7) are different not only between models but also within the CESM ensembles, showing that the changes in Antarctica are strongly influenced by internal variability, even at the century timescale. Future work including more sites or using water isotopes and the Antarctica-2K database will help constrain the spatial pattern of this trend.

The second metric is the surface temperature trend over the period from 1825 to 1925 CE at Larissa, with the magnitude of -0.94±0.12 ∘C per century. Figure 8 shows the spatial correlation between the temperature from 1825 to 1925 CE at Larissa and other grid cells for each climate model. As there are no significant differences between each member in CESM ensemble (see in the Fig. S1), only one member of CESM1 is presented in the Fig. 8. Despite the correlation coefficient decreasing with the distance from the Larissa, the values are higher than 0.6, at least around Larissa, showing that this metric is representative of part of the AP region and not extremely site specific.

None of models are able to capture the observed temperature trend from 1825 to 1925 CE (Fig. 9). Overall, models are showing a warming trend (largest for CCSM4, MPI, and BCC), contradicting the observations, as highlighted already in Fig. 4c. Only four members of CESM (CESM1, 7, 8, and 9) show a cooling trend over AP, but their magnitudes are still less than the observed one.

The 19th century is a time period when the Northern Hemisphere has started warming, whereas Southern Hemisphere records (Neukom et al., 2014), specifically Antarctica, show no general warming trend (Stenni et al., 2017). Models tend to overestimate the interhemispheric synchronicity (Neukom et al., 2014) and show a warming trend also in Antarctica, possibly in response to the anthropogenic forcing. This metric is thus an important tool for future research to evaluate whether the model data mismatch is due to internal variability (which will be investigated with more ensembles of the same model) or to an overestimated sensitivity to the anthropogenic forcing.

The warming trend of the last 50 years is one of the clearest features of the observations. Here we propose a metric of the warming trend from 1950 to 2005 CE at each of the four sites, to investigate whether model can reproduce these features.

First we look at the spatial correlation of the temperature between each site and other grid cells for all GCMs (Fig. 10). Only one member of CESM1 is presented in the Fig. 10 since no significant difference is observed between each member in the CESM ensemble (see Figs. S2–S5). The correlation is calculated from annual data for 1950 to 2005 CE. It is clear that each of our borehole temperature sites gives information about different sectors of Antarctica. Generally speaking, WAIS is representative of the West-Antarctic continent, with a more pronounced dipole between WAIS and the Weddell Sea sector in MPI and, to a lesser extent, CESM and GISS. Larissa is representative of the AP as a whole, and from this resolution of climate model results, there is no evidence of a dipole between both sides of the Transantarctic Mountains. Similar to WAIS, MPI has the strongest expression of a dipole between the AP and East Antarctica, a feature that is weaker but also present in GISS. Mill Island is generally representative of the Wilkes Land sector of East Antarctica, with the largest spatial homogeneity for BCC and IPSL (Fig. 10c). Finally, for Styx, the models with the largest spatial homogeneity (BCC and IPSL) show a strong correlation between Victoria Land and the rest of East Antarctica.

Figure 11 shows the surface temperature trend from 1950 to 2005 CE. The strong warming trend at Larissa is underestimated in all the models except the CESM ensemble (Fig. 11b). Additionally, 3 out of 12 CESM simulations indicate cooling in West Antarctica, which is coherent with the hypothesis that the part of the observed warming is due to unforced variability and that models are not expected to match this trend perfectly. The warming at Mill Island is relatively well reproduced. However, none of the models can reproduce the muted recent warming seen at Styx. The lower spatial representativity of this site (Fig. 10) leads us to interpret this as local processes missing in low-resolution GCMs, such as the influence of topography on katabatic wind forcing, rather than a large-scale failure of models to represent reality.