We use 17 LGM coupled climate model simulations from PMIP2 and PMIP3/CMIP5 Table ; and their corresponding pre-industrial (PI) control simulations as a reference. LGM boundary conditions follow the PMIP2 and PMIP3 protocols and include reduced greenhouse-gas concentrations, changed astronomical parameters, prescribed continental ice sheets and a lower global sea level. Nearly half (7 out of 17) of these simulations include dynamic vegetation, while the remaining use prescribed PI vegetation (Table 1). See https://pmip2.lsce.ipsl.fr (last access: 14 February 2020) and https://pmip3.lsce.ipsl.fr (last access: 14 February 2020) for further details and references. The analysis of PMIP model output is based on climatological means, and all output was regridded to a common 0.9∘×1.25∘ horizontal resolution. In order to compare the sea-level pressure results from different models and between PI and LGM, we removed the respective global mean before calculating the anomalies. For the analysis of geopotential height fields, only 16 instead of 17 PMIP models are used because PMIP2 LGM geopotential height from ECHAM5-MPIOM was not available to us.

To study the simulated LGM temperatures in the Siberian region in more detail and to isolate individual mechanisms, we analyzed a number of sensitivity experiments performed with the state-of-the-science coupled climate model CESM version 1.2;. The model includes the Community Atmosphere Model (CAM), Community Land Model (CLM4.0), the Parallel Ocean Program (POP2) and the Community Ice Code (CICE4). In all our CESM experiments, we use a horizontal resolution of 1.9∘×2.5∘ in the atmosphere (finite volume core) and land, and a nominal 1∘ resolution of the ocean (60 levels in the vertical) and sea-ice models with a displaced North Pole.

For the CESM LGM simulations, we followed the most recent PMIP protocol PMIP4;, including greenhouse-gas concentrations (190 ppm CO2, 357 ppb CH4 and 200 ppb N2O), orbital parameters (eccentricity of 0.019, obliquity of 22.949∘ and perihelion – 180∘ of 114.42∘) and changes in the land–sea distribution and altitude due to lower sea-level . In this study, we used as default the GLAC-1D LGM ice-sheet reconstruction . Note that the PMIP4 CH4 concentration of 375 ppb is slightly higher than the one used here.

In the first set of sensitivity experiments, we altered the imposed LGM ice-sheet boundary conditions. Within the framework of PMIP4, two LGM ice-sheet reconstructions are suggested as boundary conditions for the LGM experiments , namely GLAC-1D and ICE-6G . When comparing these two ice-sheet reconstructions, we find substantial differences, especially an overall increase of the height of the North American ice sheets in ICE-6G compared to GLAC-1D and a lowering of the Eurasian ice sheet Fig. a; both differences are on the order of 10 % of the total ice-sheet height; for more details, see. Changes in surface roughness resulting from the ice-sheet changes are highly uncertain and have not been taken into account. We performed a set of experiments to investigate the impact of these two different ice-sheet reconstructions on simulated Siberian LGM temperatures (see “continental ice sheets” set of experiments in Table ).

In the second set of sensitivity experiments, we used two different versions of the atmosphere model, CAM4 and CAM5, to investigate the importance of the atmospheric model physics (see “atmospheric model physics” set of experiments in Table ). CAM5 differs from its predecessor because it simulates indirect aerosol radiative effects by including full aerosol–cloud interactions. Furthermore, it includes improved schemes for moist turbulence, shallow convection and cloud micro- and macrophysics. Finally, while CAM4’s grid has 26 vertical levels, in CAM5, four levels were added near the surface for a better representation of boundary layer processes. See Neale et al. (2010) for a more detailed description of the atmospheric models used in CESM.

Furthermore, the CLM4.0 land model includes the possibility to use a representation of the carbon–nitrogen cycle and to calculate the resulting changes in leaf area index, stem area index and vegetation heights per plant functional type . These changes in the biophysical properties of the vegetation cover impact, for instance, evapotranspiration and surface albedo. Note that the spatial distribution of plant functional types is prescribed in CLM4.0, which is why the model is sometimes described as a semi-dynamic vegetation model. Nonetheless, for simplicity, we will refer to simulations that include carbon–nitrogen dynamics as “interactive vegetation” simulations in the remainder of this paper. To study the interdependency of interactive vegetation and atmospheric model physics, we performed a total of four experiments, with either CAM4 or CAM5 and including or excluding interactive vegetation, that are referred to as the “interactive vegetation” set of experiments (Table ).

All LGM experiments performed with CESM start from a previous LGM simulation and are run for at least 200 years to obtain a new surface climate equilibrium. Carbon pools in the litter and soils take centuries to equilibrate. However, we find that the trends are sufficiently small after 200 years to perform a robust analysis of the surface climate. Changes in Siberian (global) vegetation carbon pools amount to less than 2 % (0.6 %) of the total PI–LGM change for the model years 150–200. Top-of-the-atmosphere imbalances in the simulations including the carbon–nitrogen cycle are -0.1 and -0.185 Wm-2 using the CAM4 and CAM5 atmospheric models, respectively. Climatologies are calculated based on the last 30 years of the simulations. For the sensitivity experiments focusing on interactive vegetation and atmospheric model physics, we also performed corresponding PI simulations (Table ) to enable a proper analysis. Our five CESM LGM experiments are jointly referred to as the CESM LGM ensemble.

Throughout this paper, we focus on boreal summer (June–July–August; JJA) near-surface air temperatures and simply referred to them as “JJA temperatures” in the remainder of this paper. Moreover, when calculating LGM anomalies, we refer to the difference between an LGM simulation and the corresponding PMIP or CESM PI experiment (Table ). It is in turn differences between these CESM LGM anomalies that we use to highlight mechanisms behind the susceptibility of Siberian summer temperatures (Sect. ).

The combined PMIP2 and PMIP3 LGM experiments reveal the particularity of LGM JJA temperatures in Siberia. Of all continental areas that were not covered by large ice sheets, Siberia shows the largest intermodel spread of LGM anomalies (standard deviation; Fig. b). Another striking feature of the Siberian region is that it is one of the few regions where the PMIP multi-model mean temperature anomaly is close to or even above zero in some areas, indicating that LGM summers were potentially as warm as they are at present (Fig. a). Taken together, PMIP simulations show LGM JJA temperatures in Siberia ranging from warmer to substantially colder than they are at present. If we define a target region for Siberia based on the area where the PMIP multi-model spread is larger than 7 ∘C (green contours in Fig. ; referred to as the “Siberian target region” in the remainder of the paper and located roughly between 120–180∘ E and 70–75∘ N), we see that JJA temperature anomalies averaged over the target region for the individual models range between -12 and +12 ∘C (Fig.  and Table ). The spread in simulated LGM temperatures in the Siberian target region increases compared to PI in all seasons; however, JJA really stands out (top row of Fig. ).

Disentangling the causes of the particularity of the Siberian LGM summer temperatures based on PMIP results is not straightforward because of multiple possible underlying causes; nonetheless, some aspects can be identified. Whereas the simulated temperature changes are quite different among PMIP models, a robust decrease in precipitation on the order of 20 %–30 % is simulated (Fig. c and d). As a consequence, the (Pearson) correlation between temperature change and precipitation change in the target region is insignificant at the 0.05 significance level (R=0.36; Fig. a; note that, throughout the paper, correlation refers to intermodel correlation). A significant correlation is found between temperature and snow cover, with higher temperatures corresponding to a lower snow cover (R=-0.60; p<0.05; Fig. e). There are similarities between the spatial patterns of the PMIP multi-model spread in temperature anomalies and cloud cover anomalies (Fig. b and f); however, within the Siberian target region, local JJA temperature anomalies and cloud cover anomalies are not correlated at the 0.05 significance level (R=-0.45; Fig. b), arguing against a leading role of local cloud dynamics to explain the large intermodel spread in Siberian temperatures. As in , we find that a weakening of the North Pacific high during JJA is a consistent feature of PMIP LGM simulations (Fig. g). Moreover, a strong anticorrelation is found in the PMIP LGM simulations between JJA temperature and sea-level pressure anomalies over the Siberian target region (R=-0.72; p<0.05; Fig. c): a more positive temperature anomaly locally creates a thermal low and hence corresponds to a less pronounced sea-level pressure anomaly. Concurrently, higher sea-level pressure anomalies correspond to more positive cloud cover anomalies (R=0.50; p<0.05; Fig. d). found in their model that higher pressure is associated with lower cloud cover that in turn leads to an increase in JJA temperatures, but our results suggest that this is not the leading mechanism in the majority of PMIP LGM results. Inspecting the PI and LGM seasonal cycles for cloud and snow cover, we find that also for these variables the changes in intermodel spread in Siberia are most pronounced in summer. In contrast, the intermodel spread in precipitation does not change much between PI and LGM (Fig. ).

The strong negative correlation between JJA temperature and sea-level pressure anomalies suggests that the sea-level pressure changes could be a consequence of local temperature changes. Indeed, another reason for the negative correlation could be a remote forcing through anomalous heat advection into the Siberian target region. We find evidence for such a remote forcing of the temperature variations in the Siberian target region in the significant correlation with the large-scale mid- to high-latitude stationary wave pattern, resembling a wavenumber 2 structure (Fig. ). Increased Siberian JJA temperatures correspond to a lowering (increasing) of the JJA 500 hPa geopotential height to the southwest (southeast) of the region. The remote forcing of Siberian temperatures can thus be the result of an increase in northward-flowing relatively warm air masses over the eastern part of the Asian continent into the region of interest.

A deeper understanding of the large multi-model spread in PMIP LGM JJA temperatures over Siberia and of the mechanisms proposed above is hampered by a multitude of differences between PMIP simulations: different model formulations, different parts of the climate system that are included and different boundary conditions including the uncertainty in the reconstructed LGM ice sheet and continental outlines. Moreover, certain key climate variables are not available for a sufficiently large number of the PMIP models. In the following, we will therefore investigate a purpose-built CESM-based ensemble of LGM simulations with clearly defined differences between the individual sets of sensitivity experiments.

We construct three sets of LGM sensitivity experiments performed with the CESM climate model in order to investigate in more detail the impact of changes in boundary conditions (continental ice sheets), model formulations (atmospheric model physics) and including different components of the climate system (interactive vegetation; Table ).

Despite the fact that our total CESM LGM ensemble is smaller than the PMIP ensemble (n=5 instead of n=17) and that it was not designed to mimic the PMIP ensemble, we find that the spread in the CESM LGM temperature anomalies is surprisingly similar to the PMIP multi-model spread, both in terms of spatial distribution as well as magnitude (Fig. b). This gives us confidence that investigating the causes of the sensitivity of northeastern Siberian temperatures in the CESM ensemble can provide insights into the PMIP intermodel differences. JJA temperatures in the Siberian target region for the individual CESM experiments are listed in Table .

First, we analyze the first set of experiments (“continental ice sheets”), differing only in the imposed ice-sheet boundary conditions, namely LGM experiments forced by the GLAC-1D (LGM_CAM5_noVeg) or ICE-6G (LGM_CAM5_noVeg_ice6g) ice-sheet reconstructions (Table ). On a large scale, using the GLAC-1D ice-sheet reconstruction leads to a smaller LGM JJA temperature anomaly in the Northern Hemisphere (-6.4 ∘C) than the simulation that includes the ICE-6G ice-sheet reconstruction (-7.2 ∘C; Fig. a). Especially in the northeastern Siberian target region, the LGM simulation using GLAC-1D ice sheets is substantially warmer (9.0 ∘C) compared to the simulation using ICE-6G (6.0 ∘C; Table ). This can only be caused by changes in the large-scale atmospheric circulation since the simulations are identical apart from the ice sheets over North America and Eurasia. In line with the PMIP simulations, we find that higher JJA temperatures in the Siberian target region correspond to specific changes in the 500 hPa geopotential height field, with negative anomalies to the southwest and positive anomalies to the southeast (Fig. b), and that this stationary wave pattern results in anomalous 500 hPa southerly winds into the target region and a corresponding anomalous northward heat transport almost all the way from 30∘ N to the North Pole (Fig. c). We thus find a high sensitivity of Siberian JJA temperatures with respect to relatively minor changes in the continental ice-sheet geometries, which in turn induce changes in the circumpolar stationary wave pattern and anomalous northward heat transport in CESM. The similarity of the associated temperature and geopotential height anomaly patterns (wavenumber 2 structure; Fig. ) with the PMIP-based response (Figs.  and ) suggests that this mechanism could also explain part of the spread in PMIP simulations. The anomalous northward heat transport we see in the stationary waves contributes to reinforce the (climatological) thermal low over Siberia and explains the negative relationship between JJA temperature and JJA surface pressure anomalies in the Siberian target region, both in the CESM “continental ice sheets” set of experiments (Table ) as well as in the PMIP results (Fig. c).

The second set of CESM LGM simulations (“atmospheric model physics”) is comprised of simulations in which different versions of the atmospheric model were used (CAM4 or CAM5; Table ). Between the LGM_CAM4_noVeg and LGM_CAM5_noVeg simulations, we find changes in the large-scale atmospheric circulation, in particular the stationary waves, and northward heat transport into Siberia (Fig. ) that are broadly similar to the response to different ice sheets as described above. Similar to the analysis of the PMIP models (Fig. ) and the CESM “continental ice sheets” set of experiments (Fig. ), we find that using different atmospheric model physics can lead to JJA warming (cooling) in the Siberian target region in response to enhanced (decreased) meridional heat transport into northeastern Siberia. Interestingly, if we look in more detail, we find that the resulting surface temperature changes in Siberia are more complex in the “atmospheric model physics” set of experiments than they are for the experiments described previously. There is warming in some parts of the region but also cooling in other parts (Fig. a), and there are differences in the stationary wave pattern and in meridional heat transport. This is possibly related to slight shifts in the centers of action in the geopotential height anomalies and resulting changes in the air masses that enter the Siberian target region. This highlights the complexity of comparing simulations with different atmospheric model versions that not only differ in their response of the large-scale atmospheric circulation to LGM boundary conditions but also exhibit different local feedbacks with changes in cloud cover, humidity and pressure, which are directly influenced by, for instance, differences in cloud parameterizations and radiative properties of the atmosphere. This point is further exemplified by the substantial differences between CAM4 and CAM5 in Siberian JJA temperatures and snow cover under PI conditions (Table ). The models in the PMIP ensemble all differ in the included atmospheric physics and dynamics; thus, the described mechanism in this CESM “atmospheric model physics” set of experiments could as well explain (part of) the spread within the PMIP ensemble.

An important element in the high-latitude climate system is the vegetation–climate feedback. In the PMIP ensemble, 7 out of 17 models include the vegetation–climate feedback (Table ). However, a systematic difference in simulated JJA LGM temperature anomalies for the Siberian region could not be found when comparing models with vegetation feedback with those that did not include this additional feedback. This does not come as a surprise if one considers the relatively small sample size with respect to all the intermodel differences that impact the simulated LGM JJA temperatures. We performed PI and LGM simulations with CESM including and excluding interactive vegetation (the “interactive vegetation” set; Table ) to investigate its importance for Siberian temperatures. We find that the vegetation–climate feedback leads to a large LGM JJA cooling over Siberia, which is even more pronounced when using the CAM5 atmospheric model instead of CAM4 (Fig.  and Table ). If vegetation is allowed to respond to the changing climate through carbon–nitrogen dynamics, the tree and shrub limits shift south by several degrees of latitude as shown by the leaf area index (Fig. a and c). In CESM, the presence of vegetation, its height as well as its density have a large impact on the surface albedo through the vegetation–albedo feedback: vegetation that protrudes through the snowpack lowers the surface albedo that in turn leads to a positive feedback loop with increasing temperatures, more snowmelt, more vegetation growth and an even lower surface albedo. Accordingly, the situation in the CESM simulations including interactive vegetation is such that the cold and snow-covered landscape limits vegetation growth and leads to a southward migration of the tree and shrub limits. This relationship between vegetation and snow cover also determines the resulting LGM JJA temperature changes (compare Fig. a and c; Table ). Previous studies also found an important role of vegetation feedbacks in defining LGM Arctic temperatures . The impact of interactive vegetation in CESM is also clearly seen in the PI simulations, resulting in a substantial decrease in the leaf area index with respect to the prescribed values (Fig. a and b) and is in line with the cold bias in modeled Siberian surface temperatures described by (see also Table ).

Looking at all the experiments in the third set of experiments (“interactive vegetation”; Table ), using different atmospheric model physics (Fig. ) with or without interactive vegetation (Fig. ), we find that the strong cooling in Siberia in the simulation that combines both the different atmospheric model physics and interactive vegetation (LGM_CAM5_Veg; Fig. b) is not readily explained as a linear combination of the two individual effects. This is true for Siberian JJA temperatures but also for other key climate variables (Table ). It should be noted that the simulations with the lowest JJA LGM temperatures in Table  are in fact the ones with the highest precipitation rates (not only in JJA but also in the annual mean; not shown). This all shows the complexity of the response to a combination of factors, in this case changes in large-scale atmospheric circulation, local atmospheric processes and local land-surface processes. It is to be expected that the response of individual PMIP simulations is similarly complex.