Assessing Prior Emergent Constraints on Surface Albedo Feedback in CMIP6

Chad W. Thackeray Department of Atmospheric and Oceanic Sciences, University of California, Los Angeles, Los Angeles, California

Search for other papers by Chad W. Thackeray in
Current site
Google Scholar
PubMed
Close
,
Alex Hall Department of Atmospheric and Oceanic Sciences, University of California, Los Angeles, Los Angeles, California

Search for other papers by Alex Hall in
Current site
Google Scholar
PubMed
Close
,
Mark D. Zelinka Atmospheric, Earth, and Energy Division, Lawrence Livermore National Laboratory, Livermore, California

Search for other papers by Mark D. Zelinka in
Current site
Google Scholar
PubMed
Close
, and
Christopher G. Fletcher Department of Geography and Environmental Management, University of Waterloo, Waterloo, Ontario, Canada

Search for other papers by Christopher G. Fletcher in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

An emergent constraint (EC) is a popular model evaluation technique, which offers the potential to reduce intermodel variability in projections of climate change. Two examples have previously been laid out for future surface albedo feedbacks (SAF) stemming from loss of Northern Hemisphere (NH) snow cover (SAFsnow) and sea ice (SAFice). These processes also have a modern-day analog that occurs each year as snow and sea ice retreat from their seasonal maxima, which is strongly correlated with future SAF across an ensemble of climate models. The newly released CMIP6 ensemble offers the chance to test prior constraints through out-of-sample verification, an important examination of EC robustness. Here, we show that the SAFsnow EC is equally strong in CMIP6 as it was in past generations, while the SAFice EC is also shown to exist in CMIP6, but with different, slightly weaker characteristics. We find that the CMIP6 mean NH SAF exhibits a global feedback of 0.25 ± 0.05 W m−2 K−1, or ~61% of the total global albedo feedback, largely in line with prior generations despite its increased climate sensitivity. The NH SAF can be broken down into similar contributions from snow and sea ice over the twenty-first century in CMIP6. Crucially, intermodel variability in seasonal SAFsnow and SAFice is largely unchanged from CMIP5 because of poor outlier simulations of snow cover, surface albedo, and sea ice thickness. These outliers act to mask the noted improvement from many models when it comes to SAFice, and to a lesser extent SAFsnow.

Supplemental information related to this paper is available at the Journals Online website: https://doi.org/10.1175/JCLI-D-20-0703.s1.

© 2021 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Chad W. Thackeray, cwthackeray@ucla.edu

Abstract

An emergent constraint (EC) is a popular model evaluation technique, which offers the potential to reduce intermodel variability in projections of climate change. Two examples have previously been laid out for future surface albedo feedbacks (SAF) stemming from loss of Northern Hemisphere (NH) snow cover (SAFsnow) and sea ice (SAFice). These processes also have a modern-day analog that occurs each year as snow and sea ice retreat from their seasonal maxima, which is strongly correlated with future SAF across an ensemble of climate models. The newly released CMIP6 ensemble offers the chance to test prior constraints through out-of-sample verification, an important examination of EC robustness. Here, we show that the SAFsnow EC is equally strong in CMIP6 as it was in past generations, while the SAFice EC is also shown to exist in CMIP6, but with different, slightly weaker characteristics. We find that the CMIP6 mean NH SAF exhibits a global feedback of 0.25 ± 0.05 W m−2 K−1, or ~61% of the total global albedo feedback, largely in line with prior generations despite its increased climate sensitivity. The NH SAF can be broken down into similar contributions from snow and sea ice over the twenty-first century in CMIP6. Crucially, intermodel variability in seasonal SAFsnow and SAFice is largely unchanged from CMIP5 because of poor outlier simulations of snow cover, surface albedo, and sea ice thickness. These outliers act to mask the noted improvement from many models when it comes to SAFice, and to a lesser extent SAFsnow.

Supplemental information related to this paper is available at the Journals Online website: https://doi.org/10.1175/JCLI-D-20-0703.s1.

© 2021 American Meteorological Society. For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Chad W. Thackeray, cwthackeray@ucla.edu

1. Introduction

The extent of spring snow cover and summer sea ice across the Northern Hemisphere has decreased substantially over recent decades in response to warming (Mudryk et al. 2017; Notz and Stroeve 2018; Stroeve and Notz 2018; Meredith et al. 2019; Thackeray et al. 2019), trends that are projected to continue (Massonnet et al. 2012; Stroeve et al. 2012; Notz et al. 2020; Thackeray et al. 2016; Mudryk et al. 2020). The loss of snow and ice reduces Earth’s surface reflectivity, which impacts the climate system through the surface albedo feedback (SAF), acting to enhance warming and promote additional melt (Hall 2004; Bony et al. 2006; Thackeray and Fletcher 2016). The widespread seasonal presence of snow and ice across the Northern Hemisphere middle to high latitudes makes SAF a key driver of regional climate change (Qu and Hall 2014; Screen et al. 2012; Graversen et al. 2014; Pithan and Mauritsen 2014). Global climate models (GCMs) can qualitatively capture SAF, but they exhibit large variability in the strength of this process. This intermodel spread has been a common feature of prior model generations (Cess et al. 1991, Winton 2006; Qu and Hall 2007; Fletcher et al. 2015), making it an important source of uncertainty in GCM projections of future change (Pithan and Mauritsen 2014; Qu and Hall 2014; Boeke and Taylor 2018).

A promising avenue for wide-scale GCM improvement pertaining to SAF is the “emergent constraint” (EC) technique. This widely used form of model evaluation seeks strong relationships between an observable climate parameter and a future climate metric that emerge from an ensemble of climate models (Klein and Hall 2015; Eyring et al. 2019; Hall et al. 2019; Brient 2020). Hall and Qu (2006) first introduced an emergent constraint on springtime snow albedo feedback (SAFsnow) under climate change using an observable version of SAFsnow associated with the retreat of snow cover during the current climate’s seasonal cycle. Subsequent analysis later demonstrated that the relationship stems from physical processes common to the models (Qu and Hall 2007; Fletcher et al. 2012). Since we can observe seasonal SAFsnow, taking steps to reduce its model spread should result in smaller future SAFsnow spread. Despite these efforts, the intermodel spread in SAFsnow did not decline between the third and fifth phases of the Coupled Model Intercomparison Project (CMIP) (Qu and Hall 2014; Fletcher et al. 2015). Therefore, it is imperative that we track the development of models between generations, including key structural and parametric changes behind improved or worsened model performance. By revealing how changes can be made to reduce model bias, the hope is that other modeling centers may follow suit. A first step toward this type of targeted model development can be found in Thackeray et al. (2018). Unfortunately, this analysis was published after much of the CMIP6 development cycle had finished.

More recently, Thackeray and Hall (2019) proposed an EC on future summertime sea ice albedo feedback (SAFice) across the Arctic Ocean. It is similar to the EC for SAFsnow in that the observable climate parameter is the albedo feedback stemming from seasonal sea ice retreat during the historical period. However, the SAFice emergent relationship only exists until midcentury because of the rapid projected loss of summer sea ice, signaling the transition to a new ice-free regime with no analogs in the current seasonal cycle. Given how quickly simulated sea ice changes in the twenty-first century in some models (Stroeve et al. 2014; Notz et al. 2020), there is also a question of model suitability for our purpose. Thackeray and Hall (2019) show that the emergent relationship strengthens after removing GCMs with unrealistically thin historical ice conditions.

The availability of a new generation of GCMs (CMIP6; Eyring et al. 2016) presents a chance to revisit prior findings and monitor collective progress being made across the ensemble in possible spread reduction and model development. Additionally, ECs that exist in multiple GCM ensembles are less likely to be a result of random chance (Caldwell et al. 2018). Thus, this type of “out-of-sample” verification serves as a major indicator of EC robustness (Hall et al. 2019). To date, several ECs founded on CMIP5 relating to climate and hydrologic sensitivity have been shown to either weaken substantially or become nonexistent in CMIP6 (Pendergrass 2020; Schlund et al. 2020). Moreover, there is great interest in understanding how climate feedbacks have changed in CMIP6 given the tendency for higher equilibrium climate sensitivities (ECS) within this ensemble (Zelinka et al. 2020; Flynn and Mauritsen 2020; Meehl et al. 2020). Therefore, the primary goal of this research is to provide an update on SAF in CMIP6, with an emphasis on what it means for ECs identified in previous GCM ensembles. We also seek to shed light on the model development steps that have translated to improved or worsened SAF. The hope is that this information will be useful for the model development community going forward.

2. Data and methods

Output from a collection of CMIP5 (Taylor et al. 2012) and CMIP6 (Eyring et al. 2016) models is used to assess SAF across the Northern Hemisphere (NH). Extensive land and sea ice model development has taken place between these two model generations, thus making changes in SAF between ensembles an intriguing case study. For consistency with the CMIP5 literature, we use data from each GCM’s historical and high-emissions pathway (RCP8.5, SSP5–8.5) experiments. Only the first realization available for each GCM is utilized throughout, as prior work has shown a limited role for internal variability in generating statistical uncertainty in long-term metrics like SAF (Thackeray and Hall 2019). All calculations are performed on the GCM’s native grid, while bilinear remapping is used to visualize ensemble mean characteristics on a common grid.

Various observational and reanalysis datasets are used in conjunction with GCMs. For consistency with prior literature, we use the same datasets as Qu and Hall (2014) for SAFsnow and Thackeray and Hall (2019) for SAFice. Two estimates of observed SAFsnow are derived using black- and white-sky surface albedo measurements from the Moderate Resolution Imaging Spectroradiometer (MODIS; Schaaf et al. 2002) and near-surface air temperature from the European Centre for Medium-Range Weather Forecasts (ERA-Interim; Dee et al. 2011) averaged from 2001 to 2012. The mean of these values is used throughout, with the difference between them representing observational uncertainty (Qu and Hall 2014). A more robust assessment of observational uncertainty in SAFsnow estimates shows a weak sensitivity to the choice of albedo and temperature datasets (Fletcher et al. 2015). Following Thackeray and Hall (2019), four observational estimates of SAFice are produced from two albedo datasets [the Extended Advanced Very High Resolution Radiometer (AVHRR) Polar Pathfinder (APP-x; Wang and Key 2005) and the Satellite Applications Facility on Climate Monitoring (CM SAF) Cloud, Albedo and Surface Radiation dataset from AVHRR data (CLARA-A2; Karlsson et al. 2017)] and two near-surface air temperature datasets [the National Centers for Environmental Prediction Reanalysis 2 (NCEP-II; Kanamitsu et al. 2002) and Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2; Gelaro et al. 2017)]. Each estimate is derived using climatological data averaged over the 1982–2012 period. The mean of these four values is used as the best estimate of SAFice, while observational uncertainty is shown by ±1 standard deviation. Note that SAF values shown here are slightly stronger than that of prior studies because of the choice of radiative kernels (explained in more detail below).

SAF strength is derived from the regionally averaged change in net top-of-atmosphere shortwave radiation (ΔQnet) due to surface albedo changes stemming from a 1°C perturbation in regionally averaged surface air temperature (ΔTs). Thus, by changing the regions over which we average, SAF can be broken down into its primary components (stemming from snow and sea ice). We first calculate SAF over the NH extratropics (30°–90°N), so that its units are W m−2 K−1 of NH extratropical warming. Similar calculations are then performed over NH extratropical land and the Arctic (70°–90°N) to isolate for snow- (SAFsnow) and sea ice–driven SAF (SAFice), respectively. This follows previously published methodologies from Qu and Hall (2014, hereafter QH14) and Thackeray and Hall (2019, hereafter TH19). SAF is quantified locally as the product of one term indicating how net shortwave radiation at the top of atmosphere (TOA) changes with surface albedo variations (∂Qnet/∂αs), and another indicating the monthly change in surface albedo stemming from a 1°C change in regionally averaged surface air temperature (Δαs/[ΔTs]) [Eq. (1)]. Square brackets denote an area-averaged quantity, and t represents the 12 months of the year. (Changes are calculated between two climatological periods: 1980–99 and 2080–2100.)
SAF(t)=Qnetαs(t)Δαs(t)[ΔTs](t).
Consistent with prior studies (e.g., Colman 2013; QH14, Schneider et al. 2018; Donohoe et al. 2020), the first term is derived using radiative kernel methods (Soden et al. 2008; Shell et al. 2008). A radiative kernel represents the direct TOA radiative response to a small (e.g., 1%) change in a surface or atmospheric property (e.g., albedo). It can be generated using a climate model’s radiative transfer code. QH14 had shown that using radiative kernels derived from two different GCMs had little impact on SAFsnow strength. However, since that time radiative kernels from several other GCMs have been produced. A more recent assessment of five different sets of radiative kernels shows that the two used in QH14 fall on the weaker side of the distribution when it comes to albedo (Donohoe et al. 2020). Thus, to get a representative sample of different model regimes we use a blend of kernels from GFDL AM2 (Soden et al. 2008), CAM3 (Shell et al. 2008), CAM5 (Pendergrass et al. 2018), ECHAM6 (Block and Mauritsen 2013), and HadGEM2 (Smith et al. 2018). This better resembles the real-world relationship between planetary and surface albedo (particularly across the Arctic; Donohoe et al. 2020). It should be noted that these GCM-derived estimates of SAF are ~15%–20% weaker than if a radiative kernel derived from ERA-Interim (Huang et al. 2017; Zelinka et al. 2020) is used. However, so long as it is applied consistently, the choice of kernels will not significantly impact the key questions we are interested in here, such as intermodel spread, changes from one ensemble to the next, and the robustness of ECs. Thus, it should be noted that identical radiative kernels are applied to all observational and model datasets throughout this study.
We also break down SAF into its primary components (stemming from snow and sea ice) using previously published methodologies (QH14; TH19). These feedbacks are calculated in both the seasonal cycle and climate change contexts. Climate change feedbacks are calculated for each month of the year using the climatological change in local monthly αs and area-averaged monthly Ts. When calculating seasonal or annual means, the monthly feedback values [Eq. (1)] are first rescaled to account for the differing warming rates across each month before averaging over a subset of months or all months [Eq. (2); consistent with QH14]. For SAFsnow, we evaluate the changes between 2080–99 and 1980–99 for consistency with QH14. (Changing the historical time frame to a more recent period such as 1980–2014 has little impact.) On the other hand, because of rapid sea ice loss in the twenty-first century, we calculate SAFice at various points in the twenty-first century with respect to a historical period of 1980–2014. (TH19 used 1980–2015.) Only the results from midcentury (2030–50) are shown here for brevity as this period was found to best resemble seasonal sea ice evolution (TH19).
[SAF]annual=t{[SAF](t)[ΔTs](t)}/t[ΔTs](t).
In contrast to the climate change feedback calculation, seasonal feedbacks are derived from the climatological change in monthly mean αs and Ts (over consecutive months for SAFsnow, and over a seasonal period for SAFice). Because the seasonal feedbacks are observable (meaning we can assess model bias in the seasonal feedbacks), their evolution is what we are most interested in evaluating from CMIP5 to CMIP6. For SAFsnow, prior research has shown that the vast majority of the intermodel spread can be explained by variability in ΔαsTs (QH14), so we also calculate ΔαsTs here following the approach of Fletcher et al. (2015). For consistency with previously published CMIP5 results, we calculate ΔαsTs over the MAMJ period and limit the analysis domain to latitudes north of roughly 45°N.

Last, when interpreting results of SAFice analysis we have previously relied on sea ice thickness (SIT) data (TH19). In CMIP5, SIT was a measure of sea ice volume per grid cell area (variable “sit”), but in CMIP6 the preferred output (“sithick”) refers to the actual thickness of ice floes. Thus, only a portion of CMIP6 models evaluated here (17/27) provide a measure of SIT (variable “sivol”) that is directly comparable to CMIP5. Six of the remaining GCMs only provide “sithick,” several of which happen to fall on the upper end of the CMIP6 distribution for this variable. Given that 17 GCMs provide both “sithick” and “sivol,” we use the ensemble mean monthly differences between these two quantities to rescale “sithick” in the other six GCMs. This generates values that are comparable to CMIP5 output. We show the ensemble mean SIT calculated across both the original 17 models and our expanded collection of 23 models (Fig. 10), but focus primarily on the larger sample when discussing SIT.

3. Results

a. Surface albedo feedback under climate change

The CMIP6 ensemble exhibits an annual mean NH SAF of 0.66 ± 0.10 W m−2 K−1 under twenty-first-century climate change. Hemispheric averages for each GCM are shown in Table 1, ranging from 0.45 to 0.84 W m−2 K−1. For comparison, the CMIP5 ensemble features a slightly weaker annual mean NH SAF of 0.61 ± 0.13 W m−2 K−1 over the same period. We break down the ensemble mean response into its seasonal and regional components in Fig. 1. The annual mean is characterized by local maxima across the Tibetan Plateau, Hudson Bay, and the Arctic Ocean with slightly weaker but still important contributions from western North America and the Okhotsk Sea (Fig. 1a). Elsewhere, feedback strength across most extratropical land areas is slightly weaker because seasonal snow cover is still present over these areas at the end of the twenty-first century (2080–2100) as opposed to the complete snow/ice loss experienced in the aforementioned areas with high SAF. Note that Δαs is also weaker in vegetated regions with low historical αs due to weaker influence of snow on αs values, and where the room for future decreases due to snow loss is correspondingly limited (Essery 2013; Fig. S1 in the online supplemental material). Individual model differences can be seen in the supplemental material (Fig. S2) and these will be discussed at length in later sections. The ensemble standard deviation reveals large GCM disparities in SAF strength across the Tibetan Plateau (Fig. 1b), and to a lesser extent in the aforementioned areas where the ensemble mean strength is large (e.g., Hudson Bay, western North America). The signal-to-noise ratio (not shown) can further highlight areas of model disagreement across eastern Siberia, boreal forests, and the high Arctic Ocean. These areas are characterized by greater model uncertainty because of increasing future snowfall despite warming (Räisänen 2008), complex snow–vegetation interactions (Qu and Hall 2007; QH14; Thackeray et al. 2014, 2015; Loranty et al. 2014), and highly variable sea ice thickness influencing ice-free timing (Stroeve et al. 2014; Massonnet et al. 2018; TH19), respectively.

Table 1.

List of CMIP6 models used in this study along with their annual mean Northern Hemisphere surface albedo feedback, global feedback equivalent (equates the prior column to a global feedback using the ratios for regional to global warming and area). Seasonal cycle feedbacks for SAFsnow and SAFice are also shown for comparison. All units are W m−2 K−1, but the area over which the denominator (temperature change) is measured differs in all cases (NH extratropics, global, NH extratropical land, Arctic only).

Table 1.
Fig. 1.
Fig. 1.

Spatial distribution of future surface albedo feedback across the Northern Hemisphere extratropics from the CMIP6 ensemble mean averaged over various months: (a) annual, (c) February–May, (d) June–September, and (e) October–January. (b) The ensemble standard deviation is also shown as an annual mean. Area-weighted regional averages are displayed to the bottom right of the maps.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

We can obtain a better understanding of future SAF in CMIP6 by breaking the annual mean into its seasonal components. (We divide the calendar year into three seasons based on the shared characteristics of SAF behavior within those seasons.) The ensemble mean climate change feedback strength is strongest during late winter/spring (February–May mean = 1.11 ± 0.16 W m−2 K−1), when high insolation levels coincide with future snow cover and sub-Arctic sea ice losses (Hall 2004; Flanner et al. 2011; Thackeray et al. 2016). Like the annual mean, we see local maxima across the Tibetan Plateau, western North America, Hudson Bay, and the Barents and Bering Seas (Fig. 1c). SAF is likely the largest positive climate feedback operating across the NH extratropics during these months (Zelinka and Hartmann 2012; QH14). During summer/early fall, SAF shifts to being of greater importance in the Arctic (Figs. 1c,d), with an ensemble mean of 0.76 ± 0.18 W m−2 K−1 during June–September. We find very strong SAF over the entire Arctic Ocean during summer because most CMIP6 models have reached ice-free conditions for several months of the year by the end of the century in SSP5–8.5. Last, we see a weaker, but still meaningful signal over land during late fall/early winter (October–January) stemming from delays in the onset of snow cover (mean = 0.21 ± 0.06 W m−2 K−1).

We further dissect NH SAF into contributions from land and ocean areas to better understand its spatial structure. This is done by averaging the aforementioned SAF results across ocean and land grid cells, respectively. We find that the CMIP6 mean annual SAF stemming from land is 0.70 ± 0.16 W m−2 K−1, while the SAF stemming from ocean areas is slightly weaker (0.62 ± 0.09 W m−2 K−1). Note that both values are normalized by NH extratropical warming—later analysis isolates the more traditional feedbacks where the denominator is either land or ocean warming (sections 3b and 3c). Weaker monthly NH extratropical warming rates in summer, when SAFice is strongest, act to reduce the impact of these months relative to winter during annual averaging. As expected, SAF over land dominates the NH average during early spring and the fall (Fig. 2), while SAF over ocean areas dominates during summer. The late spring is also the time when intermodel differences are greatest because of uncertainty around the resilience of late season snow cover and sub-Arctic sea ice under future climate change. The seasonality of regional SAF is also a consistent feature across both CMIP5 and CMIP6.

Fig. 2.
Fig. 2.

CMIP6 ensemble mean monthly mean climate change SAF averaged over land (brown) and ocean areas (blue). Error bars represent ±1 standard deviation from the CMIP6 ensemble. The CMIP5 ensemble means for the same quantities are shown as dashed lines for comparison.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

For better comparison with past literature we translate the regional SAF calculated thus far into its global-mean feedback equivalent. To do so, we multiply regional SAF from each GCM by the ratio of the area covered by the region of interest (e.g., NH extratropics) to that of the globe, and the ratio of regional to global warming (consistent with QH14). For the NH extratropics, these values are (on average) 0.25 and 1.51, respectively. This produces a CMIP6 ensemble mean NH contribution to global albedo feedback of 0.25 ± 0.05 W m−2 K−1, with individual GCMs exhibiting substantial variations (0.16 to 0.37 W m−2 K−1). These findings are very similar to a recent estimate of 0.27 W m−2 K−1 using the CMIP5 ensemble and different methodology (Donohoe et al. 2020). Moreover, since the NH extratropics are believed to contribute about two-thirds of the global albedo feedback (Hall 2004; Donohoe et al. 2020), this finding is also in line with studies that have estimated the CMIP multimodel mean global feedback (including the SH) to be between 0.37 and 0.45 W m−2 K−1 (Schneider et al. 2018; Donohoe et al. 2020; Zelinka et al. 2020). Note that the spread in these estimates largely stems from the choice of radiative kernel. For example, if we recalculate the Zelinka et al. (2020) global albedo feedback of 0.45 W m−2 K−1 using our set of radiative kernels, we find a CMIP6 multimodel mean global SAF of 0.39 W m−2 K−1, with the NH contributing approximately 61% of this total. This ratio of NH to global SAF varies slightly between radiative kernels (58%–65%; not shown). Moreover, using abrupt-4xCO2 simulations, Zelinka et al. (2020) finds little change in the global albedo feedback from CMIP5 to CMIP6. To better monitor changes between model generations, we will follow previously published methodologies to isolate contributions of SAFsnow and SAFice.

b. Snow albedo feedback

1) Climate change

Here, we calculate future SAFsnow largely following the methods of Qu and Hall (Qu and Hall 2007; QH14). (As mentioned in section 2, a more complete and updated set of radiative kernels produces a more realistic estimate of SAFsnow). The CMIP6 ensemble is found to exhibit an annual mean SAFsnow under climate change of 0.59 ± 0.13 W m−2 K−1. This is only slightly stronger than the ensemble mean of 23 CMIP5 models (0.56 ± 0.18 W m−2 K−1). Along with increased mean strength, the CMIP6 ensemble shows a smaller spread, tied to a feedback strengthening for the lower bound (from 0.26 to 0.29 W m−2 K−1) and a large weakening for the upper bound (from 1.01 to 0.85 W m−2 K−1). From prior work, we know that most of the annual mean SAFsnow in previous ensembles stems from the snowmelt season (averaging across the months of February, March, April, May, and June). This is confirmed for the CMIP6 ensemble (Fig. S3; R2 = 0.97). Given that the spring mean can explain most of the intermodel variability in the annual mean, we will focus on it from here on. When averaged only over this extended spring melt period (February–June), the CMIP6 mean is 1.06 ± 0.23 W m−2 K−1. We find that spring SAFsnow varies from 0.57 to 1.57 W m−2 K−1 in the CMIP6 ensemble, a roughly 24% reduction in spread from CMIP5 (0.46 to 1.78 W m−2 K−1).

2) Seasonal cycle

SAFsnow also operates each spring in the current climate as snow cover retreats from its winter maximum. In calculating seasonal SAFsnow, we first compute the feedback for each month-to-month transition (e.g., March to April, April to May) using climatological differences between consecutive months before averaging across a series of monthly pairs (from February to July, defined as FMAMJ in QH14). The CMIP6 ensemble exhibits a mean seasonal SAFsnow of 1.08 ± 0.25 W m−2 K−1. This is in good agreement with both observations (1.06 ± 0.07 W m−2 K−1) and the CMIP5 ensemble mean (1.06 ± 0.26 W m−2 K−1). It should be noted that the values reported here are substantially stronger (~20%) than prior estimates of observational and CMIP5 mean SAFsnow from QH14. We believe that the incorporation of newer radiative kernels here better samples GCM variability in ∂αp/∂αs and thus leads to more realistic estimates of SAFsnow. A change of this magnitude also illustrates the regional importance of radiative kernel choice to feedback strength, similar to what Donohoe et al. (2020) show over the Arctic. The spatial structure of ensemble mean seasonal SAFsnow resembles that of future SAFsnow, an indicator of similar underlying processes (Fig. 3). In both contexts, the strongest SAFsnow resides in snow-covered areas of high surface albedo in the historical climate (e.g., Tibetan Plateau, Arctic tundra; Fig. S1) (Thackeray et al. 2015).

Fig. 3.
Fig. 3.

Spatial distribution of the simulated NH snow albedo feedback in the (a) seasonal cycle and (b) climate change (averaged across the months of February–June) contexts. Showing the CMIP6 mean for both time contexts.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

Seasonal SAFsnow has been shown to strongly correlate with SAFsnow under climate change across the CMIP3 (r = 0.86) and CMIP5 (r = 0.88) ensembles (Hall and Qu 2006; QH14), forming the basis of an EC. Here, we find that this relationship is equally robust in CMIP6 (r = 0.87; Fig. 4). Although the ensembles are not entirely independent of one another, given the nature of model development (Sanderson et al. 2015; Knutti et al. 2013), this type of “out-of-sample” testing is an important indicator of emergent constraint robustness (Hall et al. 2019). The EC can also be used to produce a constrained estimate of future SAFsnow as in Bowman et al. (2018). Using the hierarchical statistical framework laid out in that study we find a conservative 95% prediction interval of 0.80–1.24 W m−2 K−1 when considering all available ensembles (shown by green shading on Fig. 4). This represents a nearly three-fifths reduction from the unconstrained 95% prediction interval across these ensembles (0.49–1.53 W m−2 K−1). Note that prediction intervals are calculated for each ensemble individually, then averaged together to produce one set of values. Furthermore, because seasonal SAFsnow can be quantified from observations, we can use it to assess how GCM biases have changed in CMIP6.

Fig. 4.
Fig. 4.

Scatterplot of snow albedo feedback in the seasonal cycle (mean of all month-to-month transitions during February–July) context and under climate change (February–June seasonal mean). Each CMIP6 model is shown by a letter (corresponding to Table 1), while red diamonds (23 in total) and blue circles (19 in total) correspond to CMIP5 and CMIP3 models, respectively. Note that while CMIP3 is not our focus in this paper it is included here because the original Hall and Qu (2006) snow albedo feedback EC was developed using that ensemble. Gray shading indicates an observational estimate using MODIS albedo and its uncertainty adapted from QH14. Green shading indicates the 95% prediction interval for spring SAFsnow under climate change using the hierarchical statistical framework of Bowman et al. (2018). Note that prediction intervals are calculated for each ensemble individually, then averaged together to produce one set of values.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

3) Tracking changes in seasonal SAFsnow

It is clear from Fig. 4 that a substantial intermodel spread still exists in CMIP6. The model characteristics driving SAFsnow variability are well documented (Qu and Hall 2007; QH14; Loranty et al. 2014; Fletcher et al. 2015), but less is known about the structural and parametric uncertainties surrounding SAFsnow (Thackeray et al. 2018). To better understand why we have not seen a greater reduction in spread, we see value in investigating SAFsnow changes on a model-by-model basis. In doing so, it is also useful to consider the relevant modeling developments made between generations and how they may be influencing SAFsnow. For tracking model progress, we will use a simple quantity, ΔαsTs, which is commonly used as a proxy for the seasonal SAFsnow strength (Hall and Qu 2006; Fletcher et al. 2012, 2015; Thackeray et al. 2018) as it explains most of the intermodel spread (not shown; QH14). In this case, ΔαsTs estimates are calculated following the methods of Fletcher et al. (2015), meaning that we isolate for areas north of 45°N and mask out Greenland. There is better model agreement on the presence of snow cover across this smaller domain (Thackeray et al. 2016; Mudryk et al. 2020), which allows for snow-covered surface albedo differences to be emphasized and evaluated. Thus, a given model’s position relative to the rest of the ensemble may be slightly different than in Fig. 4. Consistent with SAFsnow, we find that ensemble-mean seasonal spring ΔαsTs has increased in CMIP6 (−1.20% to −1.29% K−1; for reference an average of observed estimates is ~−1.22% K−1; Fletcher et al. 2015). Increases in ensemble mean ΔαsTs stem from both a slight reduction in ΔTs and a slight strengthening of Δαs relative to CMIP5 (not shown). We also see a roughly 10% reduction in its intermodel spread (0.99% to 0.89% K−1), a promising sign that outliers are migrating toward the observed values.

We can track this model improvement further back to CMIP3 using results from Fletcher et al. (2015). Figure 5 shows the seasonal ΔαsTs across the three most recent CMIP ensembles. We find substantial migration of the upper and lower bounds toward the observed range in CMIP6, along with greater aggregation of GCMs near the median (apparent through smaller interquartile ranges). Note that we deem MIROC6 to be an outlier within CMIP6 (black dot) because it falls well outside of the inferred maximum of the distribution formed by the other models. This clustering around the observed range is an encouraging feature of the CMIP6 ensemble. While some might suggest this is evidence of decreased independence in the model ensemble (e.g., due to shared components like NEMO becoming prevalent), we believe the primary source of spread reduction comes from correcting land surface parameters tied to vegetation and albedo.

Fig. 5.
Fig. 5.

Seasonal cycle ΔαsTs averaged across NH land (north of 45°) in three generations of GCMs. This proxy measure of seasonal SAFsnow is commonly used to assess GCM variability. Box plots show the ensemble median, interquartile range (IQR), minimum (Q1 − 1.5 × IQR), and maximum (Q3 + 1.5 × IQR). If a GCM falls outside 1.5 times the IQR, it is shown as an outlier (black dot). Results for CMIP3 and CMIP5 are from Fletcher et al. (2015) and Thackeray et al. (2018). Gray shading shows the full observational spread in ΔαsTs from Fletcher et al. (2015) derived from a combination of albedo (MODIS, APP-x), temperature (ERA-Interim, MERRA, NCEP), and snow-cover datasets (ERA-Interim, IMS, MODIS, NOAA; see their Table 3).

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

Last, we break down the evolution of ΔαsTs at the individual model level. Figure 6 shows how the ΔαsTs from each modeling center (one model or an average of their models if more than one is available) has changed from generation to generation. The CMIP3 to CMIP5 progressions were discussed in prior work (Thackeray et al. 2018), so here we focus on the CMIP5 to CMIP6 changes. Consistent with Thackeray et al. (2018), we classify the change as “similar” if the CMIP6 value is within 0.1% K−1 of the CMIP5 value. Changes of more than 0.1% K−1 are classified as either “better” if the move is toward the observed estimate or “worse” if it moves farther away. For reference, Fletcher et al. (2015) find a spread in observed ΔαsTs derived from 24 possible combinations of surface albedo, temperature, and snow-cover datasets of 0.1% K−1 (denoted by gray shading in Figs. 5 and 6). In general, we find that a couple of outliers moved substantially closer to observations, driving the reduction in spread seen in Fig. 5, while much of the remaining ensemble rearranged itself. Notably, there is worsened representation of SAF in 6 of 15 model families (much more than for CMIP3 to CMIP5), but a majority of these models were performing relatively well in CMIP5 (i.e., within 0.1% K−1 of observations) so they do not have the same impact on ensemble spread as the aforementioned improvements to outliers. Below we discuss some of the more interesting cases.

Fig. 6.
Fig. 6.

Illustration of the change in NH land albedo (north of 45°) per 1 K of land warming (ΔαsTs) from CMIP3 to CMIP5 to CMIP6. The direction of change between ensembles, with respect to observations, is shown by arrows. A model that has changed by more than 0.1% K−1 is classified as either “better” if it moves closer to the central observational estimate (−1.2) or “worse” if it moves farther away. A model is denoted as “similar” if the change is less than 0.1% K−1 or the model falls equally on either side of OBS in both ensembles. The observational range from Fletcher et al. (2015) is shown by gray shading (as in Fig. 5). Letters correspond to a model family. Updated from Thackeray et al. (2018). Note that EC-Earth (label H) did not have the necessary CMIP5 data available so only its CMIP6 value is shown here.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

One of the biggest SAFsnow outliers in the CMIP5 ensemble was IPSL-CM5A-LR and its family of variants (Fig. 6; label R). As alluded to in Thackeray et al. (2018), substantial model development efforts have gone into improving the land component (ORCHIDEE) of this model (Boucher et al. 2020; Cheruy et al. 2020). These efforts include a new three-layer snow scheme (previously one layer) and revised parameter values for snow-covered and snow-free surface albedo based on MODIS data. Such changes have likely contributed to a much more realistic representation of mean αs, and thus seasonal ΔαsTs in its CMIP6 version. Plots of climatological March surface albedo show substantial increases in albedo where snow cover is present (Fig. 7). Similarly, the BCC-CSM model (Fig. 6; label B) has also improved its prior underestimate of ΔαsTs. This can be directly tied to a new parameterization of how snow albedo changes with snowpack age and more sophisticated representation of fractional snow cover (Wu et al. 2019), both of which increase March surface albedo over snow-covered regions (Fig. 7).

Fig. 7.
Fig. 7.

Change in climatological (1980–2015) March surface albedo between the CMIP6 version of a model and the most relevant version of its predecessor from the CMIP5 ensemble. This is meant to show systematic changes in surface albedo (consistent decrease in albedo over Eurasia in INMCM, increase over sea ice in GISS).

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

Models can also have equally or slightly less realistic representations of SAF for defensible reasons. These are GCMs that have undergone major overhauls to individual surface processes shaping SAFsnow, but this development is masked by the fact that the overall feedback is not improved. For example, CNRM-CM6-1 (Fig. 6; label F) features a new snow scheme, which produces a better representation of (increase in) annual maximum snow cover (Voldoire et al. 2019). However, the snow-cover retreat during spring is also more rapid than in CMIP5 (mean snow-cover fraction poleward of 45°N decreases by 0.78 from March to June compared to 0.60 in the prior version). Additionally, the snow-free surface albedo was reduced to better match observations (Decharme et al. 2016, 2019), thus creating more potential for albedo decline during the melt season and contributing to a slight strengthening of ΔαsTs. Similarly, the land component of CESM2 (label D) features several SAF-relevant changes, including improved canopy snow processes (interception and offloading), updated fresh snow density, and a simple firn model (Lawrence et al. 2019). Corrections to snow interception and offloading result in substantially improved (increased) ΔαsTs across the expansive boreal forest, which drives an increase in the NH average (Fig. 7; see also Fig. S5).

Finally, we discuss the small number of CMIP6 models whose ΔαsTs is substantially degraded. For example. the GISS-E2-1-G model (label N) has a realistic representation in CMIP5, but develops a very weak ΔαsTs in CMIP6 (Fig. 6). This coincides with large decreases in snow-covered αs, particularly over midlatitude forests (Fig. 7), and slower snow-cover melt (not shown). However, model documentation does not mention any changes to relevant processes (Kelley et al. 2020). Notably, the decrease in terrestrial surface albedo is counteracted by increased surface albedo over the Arctic ice pack (Fig. 7), perhaps signaling some model tuning as the CMIP5 version of this model features low biases in Arctic summer sea ice extent and thickness (Stroeve et al. 2012; TH19). Decreased SAFsnow also occurs in the MPI-ESM series of models (label U) despite a lack of relevant changes according to model documentation (Mauritsen et al. 2019). In this case, decreased spring αs is particularly strong across North America (Fig. 7) and the change appears to be tied to a substantial underestimate of winter snow-cover extent (more than 15 million km2 less than observed; Mudryk et al. 2020). We speculate that the change in SAFsnow is thus driven by the atmospheric component of the model. On a related note, biases in climatological snow cover are also apparent for some models with the strongest SAFsnow. For example, the EC-Earth models (label H; Fig. 6) drastically overestimate historical maximum snow-cover extent (nearly 10 million km2 more than the next highest GCM; Mudryk et al. 2020) and thus feature the strongest seasonal SAFsnow. These models are also outliers in the climate change context (Fig. 4) as they have more/less potential for future snow-cover loss (Levis et al. 2007). This demonstrates the importance of properly simulating maximum snow-cover extent for SAFsnow. Therefore, despite widespread model development concerning land surface processes, the reduction is SAFsnow spread from CMIP5 to CMIP6 remains limited by poor simulation of snow-cover extent and snow-covered surface albedo.

c. Sea ice albedo feedback

1) Climate change

As in TH19, we calculate SAFice over several time periods in the twenty-first century. However, given that its EC was shown to be strongest in the coming decades, we choose to only discuss results for the 2030–50 period for brevity. It should be noted that the EC also begins to break down during the latter half of the twenty-first century in CMIP6 (as ice-free conditions become more prevalent; not shown). This is particularly apparent when low-ice months like August and September are included in the calculation of future SAFice (similar to Fig. 4a of TH19). The CMIP6 ensemble exhibits an annual mean SAFice strength under midcentury climate change of 1.46 ± 0.30 W m−2 K−1. This is slightly stronger than that of the CMIP5 ensemble (1.33 ± 0.31 W m−2 K−1). The intermodel spread is largely unchanged between generations. TH19 showed that most of the annual SAFice over the Arctic (70°–90°N) occurs during the summer melt period [May–September (MJJAS)] so SAFice calculated over these months is our primary metric going forward. Averaging SAFice over MJJAS results in ensemble means of 5.06 ± 1.13 W m−2 K−1 in CMIP6 and 4.82 ± 1.07 W m−2 K−1 in CMIP5. The Arctic average is dominated by large regional contributions spanning the Eurasian sector of the Arctic Ocean and extending into the Beaufort Sea (Fig. 8b), a pattern that is largely unchanged from CMIP5 (Fig. 8d). This shows that the response of sea ice to warming is generally consistent across these two GCM generations, despite large changes to climate sensitivity (Zelinka et al. 2020).

Fig. 8.
Fig. 8.

Spatial distribution of SAFice in the (a),(c) seasonal cycle and (b),(d) midcentury (2030–50) climate change contexts according to the ensemble means from (a),(b) CMIP6 and (c),(d) CMIP5. The climate change means are averaged across the months of May–September. Calculated following TH19.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

2) Seasonal cycle

Next, we derive seasonal SAFice from the climatological (1980–2014) monthly mean albedo and temperature for May and August as in TH19. We find that ensemble mean SAFice strength is largely unchanged from CMIP5 to CMIP6, both in its spatial distribution (Figs. 8a,c) and when taken as an Arctic average (both ~6.00 W m−2 K−1). This is a slight underestimate of our best observational estimate (6.45 ± 0.63 W m−2 K−1; updated from TH19 with different radiative kernels), but within the range of observational uncertainty. Notably, there is increased intermodel spread in CMIP6, with Arctic means ranging from 2.3 to 9.6 W m−2 K−1. (For comparison, the same measure in CMIP5 ranges from 2.9 to 8.7 W m−2 K−1.) This fourfold variability demonstrates the large model disagreement when it comes to simulating seasonal ice loss during the historical period. It is also possible that some of this increased spread could be tied to the use of a somewhat larger number of GCMs for CMIP6 here (27 vs 23).

3) Emergent constraint on SAFice

TH19 showed that there is a strong relationship in CMIP5 models between how Arctic sea ice responds to seasonal warming during the melt period and how it responds to climate change (limited to areas north of 70°N). We find that this relationship generally holds in CMIP6, with some differing details: In CMIP5, the emergent relationship was shown to be strongest in the near-future early melt period (May–June), gradually weakening as more and more models reach ice-free status in the future. When we reproduce this calculation using CMIP6 models, we find this SAFice relationship holds, but it is somewhat weaker (Pearson’s r decreases from 0.74 to 0.56; Fig. 9a). Once again, we can use the EC to produce a constrained estimate of midcentury SAFice (Bowman et al. 2018; TH19). We find a 95% prediction interval of 3.8–8.2 W m−2 K−1 when considering both ensembles (shown by green shading on Fig. 9a), which represents a nearly one-third reduction from the unconstrained 95% prediction interval. The weaker correlation in CMIP6 drives a wider prediction interval than that suggested by only CMIP5 models (4.4–7.3 W m−2 K−1). At the same time, we find an increase in the correlation strength relative to CMIP5 when the full melt period (MJJAS) is used to derive climate change SAFice (Fig. 9b). Thus, seasonal SAFice is now less predictive of the May–June climate change SAFice, but more predictive of the entire melt season’s climate change SAFice. These findings point to a shift in the portion of the future melt season that most resembles the current seasonal cycle in the latest ensemble.

Fig. 9.
Fig. 9.

Scatterplots of sea ice albedo feedback in the seasonal cycle (May–August) context and under midcentury climate change (2030–50) averaged over (a) May–June and (b) May–September. Each CMIP6 model is shown by a letter (corresponding to Table 1), while red diamonds correspond to a CMIP5 model. The gray shading indicates our best observational estimate of seasonal SAFice updated from TH19 plus and minus one standard deviation. Consistent with Fig. 4, green shading indicates the 95% prediction interval for SAFice under midcentury climate change using the hierarchical statistical framework of Bowman et al. (2018). Note that the NorCPM1 model (Y2) does not have ScenarioMIP output, so data from the 1pctto2xCO2 experiment (years 45–65 were found to best represent the warming seen by midcentury in the rest of the ensemble) are used instead to calculate climate change SAFice. This model’s inclusion is warranted because it has the weakest seasonal SAFice, which allows for us to best capture the full range of CMIP6 variability.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

To better understand the reasons behind this changing relationship, we assess the sea ice characteristics in both ensembles. We see little change in the ensemble mean ice-covered surface albedo at the start or end of the melt season (Fig. S1). Moreover, recent work has shown that annual maximum and minimum sea ice extent are largely consistent across ensembles (Notz et al. 2020). TH19 suggested that SIT plays an important role in this relationship because models with thicker historical ice will generally take longer to reach ice-free conditions, while models with thin ice have less potential for ice loss in the future because their ice extent is already relatively low. However, a comparison of SIT across generations is not necessarily straightforward due to changes in CMIP guidelines (see section 2). Using a collection of 23 GCMs providing SIT data (either “sivol” or “sithick”), we find a general tendency toward slightly thicker Arctic (70°–90°N) sea ice in CMIP6 during the 1980–2014 period. It should be noted that this appears to be in contrast with findings of decreased 1979–98 Arctic sea ice volume in CMIP6 reported by Notz et al. (2020). The main reason for this discrepancy is model selection, while differing time periods and study domains are secondary factors. The difference arises because we aim to limit the number of highly similar GCMs from the model subsets as the presence of several different model versions could artificially strengthen an emergent relationship. Using data from Tables S2 and S3 of Notz et al. (2020) we find that Arctic sea ice volume is slightly larger in our subset of CMIP6 models (15 500 km3) than our CMIP5 subset (15 300 km3). Figure 10 shows the seasonal cycle of historical (1980–2014) SIT derived from the CMIP6 and CMIP5 ensemble means along with and observational estimate from PIOMAS (from TH19). The CMIP6 mean also exhibits a slightly weaker change in Arctic average SIT from May to August (−0.74 m) than that of CMIP5 (−0.90 m, which was in exceptionally good agreement with PIOMAS). Furthermore, the CMIP6 ensemble mean August SIT across the Arctic (NH70) is 1.65 m, whereas the CMIP5 and PIOMAS values are 1.40 m. When the bias from each individual CMIP model is plotted, we can see a slight shift to models overestimating summer SIT in CMIP6 (Fig. 10b). Future analysis of snow on sea ice loss during the melt season may also help to explain some of these differences in SAFice. The fact that a thicker icepack typically exists later in the summer than it did for CMIP5 may hint as to why we see an increased correlation for CMIP6 in Fig. 9b. By having fewer GCMs with unrealistically thin historical summer SIT, the CMIP6 ensemble is likely less susceptible to degradation of the SAFice relationship caused by ice-free conditions (seen by TH19 to occur both later in the melt season and further into the future).

Fig. 10.
Fig. 10.

(a) Seasonal cycle of historical (1980–2014) sea ice thickness (SIT) across the Arctic (70°–90°N) from the CMIP6 and CMIP5 ensemble means along with PIOMAS (observations). The gray shading is meant to highlight the summer melt period, where we calculate SAFice. (b) Histograms of the historical SIT bias (relative to PIOMAS) during the months of May–September for each GCM in the CMIP6 and CMIP5 ensembles that provide necessary data. Negative y values are used for CMIP5 to better visualize the data. SIT is averaged over areas with greater than 10% sea ice concentration. The CMIP6 mean combines output from 17 GCMs providing a variable (“sivol,” also known as equivalent thickness) that is directly comparable to CMIP5 output (shown as a dashed line) and 6 models whose SIT is rescaled based on a newer ice thickness output variable (“sithick,” the actual thickness of simulated ice over the ice-covered part of a grid cell). The SIT threshold used by TH19 to classify GCMs as unrealistically thin is shown in red for reference. Error bars in (a) illustrate plus and minus one standard error.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

As previously mentioned, TH19 showed that the presence of ice-free conditions, found for example by looking further into the future or later into the melt season, significantly weakens the SAFice relationship in CMIP5. Thus, it may be useful to exclude some GCMs that clearly reach ice-free status too soon in the future from the ensemble in the present survey of CMIP6 models. TH19 removed GCMs based on their historical ice thickness, several of which exhibited unrealistically thin conditions (six GCMs were at least 50% thinner than observed estimates of Arctic SIT). However, in CMIP6, only two versions of the CNRM climate model and NESM3 satisfy this criterion. Removing these models does not alter the CMIP6 SAFice relationship. In fact, both CNRM models exhibit anomalous cooling over the North Atlantic well into the twenty-first century, promoting increases in sea ice cover and weakening their Arctic SAFice (not shown). Given the slightly thicker historical ice conditions across the CMIP6 ensemble, an alternative to the approach of removing the models with inappropriately thin ice might be to remove GCMs that first reach ice-free conditions (<1 million km2) during September in the very near future (by 2025). There are four such models (NESM3, CESM2, HadGEM3-GC31-LL, and EC-Earth3) but removing them only slightly strengthens the SAFice emergent relationship. Thus, consideration of model initial conditions appears to have been more important in CMIP5 because of its generally thinner historical ice pack.

4) Tracking changes in seasonal SAFice

Last, we briefly break down the progress in representing seasonal SAFice at the individual model level. In the case of sea ice, tracking model progress is slightly more difficult than snow over land because GCMs with identical ice models can produce more variability in SAFice than identical land models are likely to produce for SAFsnow. For example, the CESM2 models, each with the same ice module, exhibit considerable discrepancies in the simulation of historical ice cover (not shown) and thus SAFice (Table 1). Figure 11 shows how SAFice from each modeling center (one model or an average of their models if more than one is available) has changed from CMIP5 to CMIP6. In general, we find most GCMs feature improved or similar representation of SAFice in CMIP6 (11 of 15). However, a couple of GCMs now exhibit significantly deteriorated SAFice, thus driving increases in ensemble spread. Next, we discuss the factors contributing to unrealistic representation of SAFice in these outliers.

Fig. 11.
Fig. 11.

Illustration of the change in seasonal SAFice from CMIP5 to CMIP6. The direction of change between ensembles, with respect to observations, is shown by arrows. The observational range produced from four estimates of SAFice is shown by gray shading. A model that has changed by more than 0.6 W m−2 K−1 (roughly equal to one standard deviation derived from observational estimate of SAFice) is classified as either “better” if it moves closer to the central observational estimate or “worse” if it moves farther away. A model is denoted as “similar” if the change is less than 0.6 W m−2 K−1 or the model falls equally on either side of observed SAFice in both ensembles. Letters correspond to a model family listed in Table 1.

Citation: Journal of Climate 34, 10; 10.1175/JCLI-D-20-0703.1

First, the CESM models (e.g., CESM2 and CESM2-WACCM; label D) now feature unrealistically strong SAFice following implementation of the CICE5 model. CICE5 notably includes more layers for sea ice and snow on ice (see the CESM CICE5 users’ guide, available at https://buildmedia.readthedocs.org/media/pdf/cesmcice/latest/cesmcice.pdf), new mushy-layer thermodynamics (Turner and Hunke 2015), and a new melt pond scheme (Hunke et al. 2013). Hunke et al. (2013) note that their parameterization leads to enhanced albedo variations due to greater ice area exposure to ponding, so we suspect it is the last of these changes that seems most likely to drive amplified SAFice. The resultant decline in historical late-summer surface albedo from CMIP5 is clear (Fig. S5). MPI-ESM-1-2-HR (label U) also features strengthened SAFice, leading to slightly worsened performance relative to observations (Fig. 11). This is likely tied to the correction of a model bug that prevented interaction between melt pond evolution, ice albedo, and ice melt in the CMIP5 version (Mauritsen et al. 2019), resulting in lower late-summer albedo over the Arctic (Fig. S5). We can also see the influence of broader GCM development on SAFice in the case of INM-CM. Significantly weakened SAFice compared to observations in the CMIP6 version of this model (label P) appears to stem from less seasonal change in ice area than its predecessor (Volodin et al. 2017a) and a cold bias across the Arctic (Volodin et al. 2017b), rather than changes to parameterizations. In summary, we find these instances of deteriorated SAFice are the exception rather than the norm, and that most GCMs better represent seasonal SAFice in CMIP6 despite larger biases in Arctic ice thickness.

4. Summary and conclusions

This study provides an evaluation of Northern Hemisphere surface albedo feedbacks in CMIP6 through comparisons with both observational estimates and past model generations. We first break down the hemispheric feedback into its contributions from snow cover and sea ice to find that they are largely of equal importance to the NH total. Under twenty-first-century climate change, both feedbacks are slightly stronger in CMIP6 than they were in CMIP5, possibly contributing to the greater climate sensitivity in CMIP6 [although global mean SAF calculated using 4xCO2 experiments was unchanged in Zelinka et al. (2020)]. Both feedbacks also have a contemporary analog that occurs each year during seasonal retreat of snow and sea ice. Moreover, it is important to assess the change in the seasonal versions of these feedbacks (SAFsnow and SAFice) because their intermodel variability has been shown to be strongly correlated with their variability under climate change. We show that these previously documented emergent relationships still exist in CMIP6, but with differing levels of robustness. The SAFsnow constraint exhibits remarkable consistency across three model generations (Fig. 4), whereas the SAFice constraint weakens somewhat and exhibits seasonal differences in CMIP6 (Fig. 9). Methodological changes in this study (compared to QH14 and TH19), notably the use of a blend of radiative kernels, improve our quantification of feedbacks so that they better resemble the real-world influence of surface albedo changes on the TOA energy balance [as documented in Donohoe et al. (2020)]. We find NH SAF to have a contribution to the global albedo feedback of 0.25 ± 0.05 W m−2 K−1, thus accounting for roughly 61% of the total global albedo feedback in CMIP6 (0.39 W m−2 K−1). This is in good agreement with prior assessments for CMIP5 using different methods (Donohoe et al. 2020; Schneider et al. 2018). The improved choice of kernels has the effect of slightly strengthening prior estimates of SAFsnow, SAFice, and their global impact (QH14; TH19). However, the simulated global albedo feedback is still weaker than past observational assessments of the satellite record (0.48 W m−2 K−1; Flanner et al. 2011; Cao et al. 2015). We believe that much of this discrepancy between recent and future SAF stems from internal variability in observed Arctic sea ice trends (Swart et al. 2015; Ding et al. 2017, 2019) and changes in aerosol forcing during this period (increasing black carbon deposition and decreasing global sulfate concentrations) that do not resemble the twenty-first-century climate change signal (Gagné et al. 2015; Mueller et al. 2018).

The ensemble mean seasonal SAFsnow is largely unchanged from CMIP5 to CMIP6, thus remaining in good agreement with observational estimates. Moreover, the intermodel spread in SAFsnow is similar in both ensembles. Perhaps this is not surprising because the spread in climatological maximum snow-cover extent remains large (Mudryk et al. 2020). Thus, to better understand model development pertaining to snow-covered surface albedo, we also assess ΔαsTs (a proxy for SAFsnow) over a smaller domain where model agreement regarding the presence snow cover is much greater (NH land areas north of 45°). Notably, we find a 10% reduction in the intermodel spread for ΔαsTs in CMIP6. In general, this reduction is driven by a couple of outliers that moved substantially closer to observations. These improvements were spurred on by a variety of new or updated parameterizations for snow, vegetation, and surface albedo. However, there is also degraded representation of SAFsnow in 6 of 15 model families, although the magnitude of change is rather small in most cases. We previously provided an optimistic outlook on ΔαsTs spread derived from a small subset of preliminary CMIP6 output (Thackeray et al. 2018). Of this subset, the two models that were substantial outliers in CMIP5 (IPSL and INM-CM) both feature substantial improvements in CMIP6. However, the overall spread reduction that materialized is less than what we had anticipated. This is largely due to degraded SAFsnow in the GISS, MPI, and EC-Earth model families. These and other CMIP6 outliers can be directly attributed to large biases in either historical snow-cover extent or snow-covered surface albedo (often tied to vegetation biases; Loranty et al. 2014). Given that this large intermodel spread has persisted across several generations, despite widespread knowledge of the Hall and Qu (2006) EC, we provide some speculation as to why there may be difficulty in reducing it. In rare instances, models that were outliers in prior generations have simply not taken sufficient action (e.g., NH vegetation biases in MIROC driving unrealistically high snow-covered αs). Moreover, it seems increasingly likely that parameters relating to snow may be relied upon to reduce temperature biases across midlatitude continents in several models. The continued introduction of new GCMs also adds to the complexity of narrowing model uncertainty.

The ensemble mean seasonal SAFice is also very similar in both the CMIP5 and CMIP6 ensembles, and slightly weaker than observational estimates (updated from TH19). However, in this case we find a small increase in intermodel spread for CMIP6. Given the more recent publication of the constraint on SAFice (TH19), perhaps this is not surprising. Extensive sea ice model development in recent years has results in improved SAFice across most GCMs, although there are a couple of outliers with degraded performance that prevent a reduction in ensemble spread. The CMIP6 ensemble notably exhibits fewer models with unrealistically thin historical ice conditions, which may be driving changes to the character of the SAFice EC (Fig. 9).

Going forward we hope that this assessment will provide useful insights to modeling centers for the next cycle of development. The confirmation of emergent constraints on SAFsnow and SAFice means that continued steps to reduce the intermodel spread in the observable seasonal cycle feedbacks should result in better constrained future feedbacks. Modeling centers should seek to reduce biases in maximum annual snow-cover extent, snow-covered surface albedo, and sea ice thickness, as these all serve as good proxies for improving SAF. Future efforts should seek to incorporate more cryosphere-focused metrics into model benchmarking packages (e.g., ILAMB; Collier et al. 2018) so that analysis of this type is more readily accessible during model development.

Acknowledgments

CWT and AH acknowledge support from the National Science Foundation grant (1543268) titled “Reducing Uncertainty Surrounding Climate Change Using Emergent Constraints” and the Regional and Global Model Analysis Program for the Office of Science of the U.S. Department of Energy. MDZ’s work was performed under the auspices of the U.S. Department of Energy (DOE) by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 and was supported by the Regional and Global Model Analysis Program of the Office of Science at the DOE. We thank the World Climate Research Programme’s Working Group on Coupled Modelling and the individual modeling groups for their roles in making CMIP6 data available. For CMIP the U.S. Department of Energy’s Program for Climate Model Diagnosis and Intercomparison provided coordinating support and led the development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. We also thank three anonymous reviewers for their constructive feedback.

Data availability statement

The data that support the findings of this study are publicly available. The CMIP output is available from the Earth System Grid Federation (https://esgf-node.llnl.gov/). The ice thickness data can be downloaded from http://psc.apl.uw.edu/research/projects/arctic-sea-ice-volume-anomaly/data/. Satellite albedo data are available from https://lpdaac.usgs.gov/products/mcd43c3v006/, https://www.ncei.noaa.gov/data/avhrr-polar-pathfinder-extended/access/, and https://wui.cmsaf.eu/safira/action/viewDoiDetails?acronym=CLARA_AVHRR_V002. The temperature reanalysis datasets are available from https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era-interim, https://www.esrl.noaa.gov/psd/data/gridded/data.ncep.reanalysis2.html, and https://gmao.gsfc.nasa.gov/reanalysis/MERRA-2/data_access/.

REFERENCES

  • Block, K., and T. Mauritsen, 2013: Forcing and feedback in the MPI-ESM-LR coupled model under abruptly quadrupled CO2. J. Adv. Model. Earth Syst., 5, 676691, https://doi.org/10.1002/jame.20041.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Boeke, R. C., and P. C. Taylor, 2018: Seasonal energy exchange in sea ice retreat regions contributes to differences in projected Arctic warming. Nat. Commun., 9, 5017, https://doi.org/10.1038/s41467-018-07061-9.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bony, S., and Coauthors, 2006: How well do we understand and evaluate climate change feedback processes? J. Climate, 19, 34453482, https://doi.org/10.1175/JCLI3819.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Boucher, O., and Coauthors, 2020: Presentation and evaluation of the IPSL-CM6A-LR climate model. J. Adv. Model. Earth Syst., 12, e2019MS002010, https://doi.org/10.1029/2019MS002010.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bowman, K. W., N. Cressie, X. Qu, and A. Hall, 2018: A hierarchical statistical framework for emergent constraints: Application to snow-albedo feedback. Geophys. Res. Lett., 45, 13 05013 059, https://doi.org/10.1029/2018GL080082.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Brient, F., 2020: Reducing uncertainties in climate projections with emergent constraints: Concepts, examples and prospects. Adv. Atmos. Sci., 37, 115, https://doi.org/10.1007/s00376-019-9140-8.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Caldwell, P. M., M. D. Zelinka, and S. A. Klein, 2018: Evaluating emergent constraints on equilibrium climate sensitivity. J. Climate, 31, 39213942, https://doi.org/10.1175/JCLI-D-17-0631.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cao, Y., S. Liang, X. Chen, and T. He, 2015: Assessment of sea ice albedo radiative forcing and feedback over the Northern Hemisphere from 1982 to 2009 using satellite and reanalysis data. J. Climate, 28, 12481259, https://doi.org/10.1175/JCLI-D-14-00389.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cess, R. D., and Coauthors, 1991: Interpretation of snow–climate feedback as produced by 17 general circulation models. Science, 253, 888892, https://doi.org/10.1126/science.253.5022.888.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cheruy, F., and Coauthors, 2020: Improved near-surface continental climate in IPSL-CM6A-LR by combined evolutions of atmospheric and land surface physics. J. Adv. Model. Earth Syst., 12, e2019MS002005, https://doi.org/10.1029/2019MS002005.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Collier, N., F. M. Hoffman, D. M. Lawrence, G. Keppel-Aleks, C. D. Koven, W. J. Riley, M. Mu, and J. T. Randerson, 2018: The International Land Model Benchmarking (ILAMB) system: Design, theory, and implementation. J. Adv. Model. Earth Syst., 10, 27312754, https://doi.org/10.1029/2018MS001354.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Colman, R. A., 2013: Surface albedo feedbacks from climate variability and change. J. Geophys. Res. Atmos., 118, 28272834, https://doi.org/10.1002/jgrd.50230.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Decharme, B., E. Brun, A. Boone, C. Delire, P. Le Moigne, and S. Morin, 2016: Impacts of snow and organic soils parameterization on northern Eurasian soil temperature profiles simulated by the ISBA land surface model. Cryosphere, 10, 853877, https://doi.org/10.5194/tc-10-853-2016.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Decharme, B., and Coauthors, 2019: Recent changes in the ISBA-CTRIP land surface system for use in the CNRM-CM6 climate model and in global off-line hydrological applications. J. Adv. Model. Earth Syst., 11, 12071252, https://doi.org/10.1029/2018MS001545.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553597, https://doi.org/10.1002/qj.828.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ding, Q., and Coauthors, 2017: Influence of high-latitude atmospheric circulation changes on summertime Arctic sea ice. Nat. Climate Change, 7, 289295, https://doi.org/10.1038/nclimate3241.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Ding, Q., and Coauthors, 2019: Fingerprints of internal drivers of Arctic sea ice loss in observations and model simulations. Nat. Geosci., 12, 2833, https://doi.org/10.1038/s41561-018-0256-8.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Donohoe, A., E. Blanchard-Wrigglesworth, A. Schweiger, and P. J. Rasch, 2020: The effect of atmospheric transmissivity on model and observational estimates of the sea ice albedo feedback. J. Climate, 33, 57435765, https://doi.org/10.1175/JCLI-D-19-0674.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Essery, R., 2013: Large-scale simulations of snow albedo masking by forests. Geophys. Res. Lett., 40, 55215525, https://doi.org/10.1002/grl.51008.

  • Eyring, V., S. Bony, G. A. Meehl, C. A. Senior, B. Stevens, R. J. Stouffer, and K. E. Taylor, 2016: Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev., 9, 19371958, https://doi.org/10.5194/gmd-9-1937-2016.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Eyring, V., and Coauthors, 2019: Taking climate model evaluation to the next level. Nat. Climate Change, 9, 102110, https://doi.org/10.1038/s41558-018-0355-y.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Flanner, M. G., K. M. Shell, M. Barlage, D. K. Perovich, and M. A. Tschudi, 2011: Radiative forcing and albedo feedback from the Northern Hemisphere cryosphere between 1979 and 2008. Nat. Geosci., 4, 151155, https://doi.org/10.1038/ngeo1062.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Fletcher, C. G., H. Zhao, P. J. Kushner, and R. Fernandes, 2012: Using models and satellite observations to evaluate the strength of snow albedo feedback. J. Geophys. Res., 117, D11117, https://doi.org/10.1029/2012JD017724.

    • Search Google Scholar
    • Export Citation
  • Fletcher, C. G., C. W. Thackeray, and T. M. Burgers, 2015: Evaluating biases in simulated snow albedo feedback in two generations of climate models. J. Geophys. Res. Atmos., 120, 1226, https://doi.org/10.1002/2014JD022546.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Flynn, C. M., and T. Mauritsen, 2020: On the climate sensitivity and historical warming evolution in recent coupled model ensembles. Atmos. Chem. Phys., 20, 78297842, https://doi.org/10.5194/acp-20-7829-2020.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gagné, M., N. P. Gillett, and J. C. Fyfe, 2015: Impact of aerosol emission controls on future Arctic sea ice cover. Geophys. Res. Lett., 42, 84818488, https://doi.org/10.1002/2015GL065504.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gelaro, R., and Coauthors, 2017: The Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2). J. Climate, 30, 54195454, https://doi.org/10.1175/JCLI-D-16-0758.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Graversen, R. G., P. L. Langen, and T. Mauritsen, 2014: Polar amplification in CCSM4: Contributions from the lapse rate and surface albedo feedbacks. J. Climate, 27, 44334450, https://doi.org/10.1175/JCLI-D-13-00551.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hall, A., 2004: The role of surface albedo feedback in climate. J. Climate, 17, 15501568, https://doi.org/10.1175/1520-0442(2004)017<1550:TROSAF>2.0.CO;2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hall, A., and X. Qu, 2006: Using the current seasonal cycle to constrain snow albedo feedback in future climate change. Geophys. Res. Lett., 33, L03502, https://doi.org/10.1029/2005GL025127.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hall, A., P. Cox, C. Huntingford, and S. Klein, 2019: Progressing emergent constraints on future climate change. Nat. Climate Change, 9, 269278, https://doi.org/10.1038/s41558-019-0436-6.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Huang, Y., Y. Xia, and X. Tan, 2017: On the pattern of CO2 radiative forcing and poleward energy transport. J. Geophys. Res. Atmos., 122, 10 57810 593, https://doi.org/10.1002/2017JD027221.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hunke, E. C., D. A. Hebert, and O. Lecomte, 2013: Level-ice melt ponds in the Los Alamos Sea Ice Model, CICE. Ocean Modell., 71, 2642, https://doi.org/10.1016/j.ocemod.2012.11.008.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kanamitsu, M., W. Ebisuzaki, J. Woollen, S. K. Yang, J. J. Hnilo, M. Fiorino, and G. L. Potter, 2002: NCEP-DOE AMIP-II reanalysis (R-2). Bull. Amer. Meteor. Soc., 83, 16311644, https://doi.org/10.1175/BAMS-83-11-1631.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Karlsson, K. G., and Coauthors, 2017: CLARA-A2: The second edition of the CM SAF cloud and radiation data record from 34 years of global AVHRR data. Atmos. Chem. Phys., 17, 58095828, https://doi.org/10.5194/acp-17-5809-2017.

    • Search Google Scholar
    • Export Citation
  • Kelley, M., and Coauthors, 2020: GISS-E2.1: Configurations and climatology. J. Adv. Model. Earth Syst., 12, e2019MS002025, https://doi.org/10.1029/2019MS002025.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Klein, S. A., and A. Hall, 2015: Emergent constraints for cloud feedbacks. Curr. Climate Change Rep., 1, 276287, https://doi.org/10.1007/s40641-015-0027-1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Knutti, R., D. Masson, and A. Gettelman, 2013: Climate model genealogy: Generation CMIP5 and how we got there. Geophys. Res. Lett., 40, 11941199, https://doi.org/10.1002/grl.50256.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Lawrence, D. M., and Coauthors, 2019: The Community Land Model version 5: Description of new features, benchmarking, and impact of forcing uncertainty. J. Adv. Model. Earth Syst., 11, 42454287, https://doi.org/10.1029/2018MS001583.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Levis, S., G. B. Bonan, and P. J. Lawrence, 2007: Present-day springtime high-latitude surface albedo as a predictor of simulated climate sensitivity. Geophys. Res. Lett., 34, L17703, https://doi.org/10.1029/2007GL030775.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Loranty, M. M., L. T. Berner, S. J. Goetz, Y. Jin, and J. T. Randerson, 2014: Vegetation controls on northern high latitude snow-albedo feedback: Observations and CMIP5 model simulations. Global Change Biol., 20, 594606, https://doi.org/10.1111/gcb.12391.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • 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, 13831394, https://doi.org/10.5194/tc-6-1383-2012.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Massonnet, F., M. Vancoppenolle, H. Goosse, D. Docquier, T. Fichefet, and E. Blanchard-Wrigglesworth, 2018: Arctic sea-ice change tied to its mean state through thermodynamic processes. Nat. Climate Change, 8, 599603, https://doi.org/10.1038/s41558-018-0204-z.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mauritsen, T., and Coauthors, 2019: Developments in the MPI-M Earth system model version 1.2 (MPI-ESM1.2) and its response to increasing CO2. J. Adv. Model. Earth Syst., 11, 9981038, https://doi.org/10.1029/2018MS001400.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Meehl, G. A., C. A. Senior, V. Eyring, G. Flato, J. F. Lamarque, R. J. Stouffer, K. E. Taylor, and M. Schlund, 2020: Context for interpreting equilibrium climate sensitivity and transient climate response from the CMIP6 Earth system models. Sci. Adv., 6, eaba1981, https://doi.org/10.1126/sciadv.aba1981.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Meredith, M., and Coauthors, 2019: Polar regions. IPCC Special Report on the Ocean and Cryosphere in a Changing Climate, H.-O. Pörtner et al., Eds., IPCC, 203–320.

  • Mudryk, L., P. J. Kushner, C. Derksen, and C. Thackeray, 2017: Snow cover response to temperature in observational and climate model ensembles. Geophys. Res. Lett., 44, 919926, https://doi.org/10.1002/2016GL071789.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mudryk, L., M. Santolaria-Otín, G. Krinner, M. Ménégoz, C. Derksen, C. Brutel-Vuilmet, M. Brady, and R. Essery, 2020: Historical Northern Hemisphere snow cover trends and projected changes in the CMIP6 multi-model ensemble. Cryosphere, 14, 24952514, https://doi.org/10.5194/tc-14-2495-2020.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mueller, B. L., N. P. Gillett, A. H. Monahan, and F. W. Zwiers, 2018: Attribution of Arctic sea ice decline from 1953 to 2012 to influences from natural, greenhouse gas, and anthropogenic aerosol forcing. J. Climate, 31, 77717787, https://doi.org/10.1175/JCLI-D-17-0552.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Notz, D., and J. Stroeve, 2018: The trajectory towards a seasonally ice-free Arctic Ocean. Curr. Climate Change Rep., 4, 407416, https://doi.org/10.1007/s40641-018-0113-2.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Notz, D., and Coauthors, 2020: Arctic sea ice in CMIP6. Geophys. Res. Lett., 47, e2019GL086749, https://doi.org/10.1029/2019GL086749.

  • Pendergrass, A. G., 2020: The global-mean precipitation response to CO2-induced warming in CMIP6 models. Geophys. Res. Lett., 47, e2020GL089964, https://doi.org/10.1029/2020GL089964.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Pendergrass, A. G., A. Conley, and F. M. Vitt, 2018: Surface and top-of-atmosphere radiative feedback kernels for CESM-CAM5. Earth Syst. Sci. Data, 10, 317324, https://doi.org/10.5194/essd-10-317-2018.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Pithan, F., and T. Mauritsen, 2014: Arctic amplification dominated by temperature feedbacks in contemporary climate models. Nat. Geosci., 7, 181184, https://doi.org/10.1038/ngeo2071.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Qu, X., and A. Hall, 2007: What controls the strength of snow-albedo feedback? J. Climate, 20, 39713981, https://doi.org/10.1175/JCLI4186.1.

  • Qu, X., and A. Hall, 2014: On the persistent spread in snow-albedo feedback. Climate Dyn., 42, 6981, https://doi.org/10.1007/s00382-013-1774-0.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Räisänen, J., 2008: Warmer climate: Less or more snow? Climate Dyn., 30, 307319, https://doi.org/10.1007/s00382-007-0289-y.

  • Sanderson, B. M., R. Knutti, and P. Caldwell, 2015: A representative democracy to reduce interdependency in a multimodel ensemble. J. Climate, 28, 51715194, https://doi.org/10.1175/JCLI-D-14-00362.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schaaf, C., and Coauthors, 2002: First operational BRDF, albedo nadir reflectance products from MODIS. Remote Sens. Environ., 83, 135148, https://doi.org/10.1016/S0034-4257(02)00091-3.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schlund, M., A. Lauer, P. Gentine, S. C. Sherwood, and V. Eyring, 2020: Emergent constraints on equilibrium climate sensitivity in CMIP5: Do they hold for CMIP6? Earth Syst. Dyn., 11, 12331258, https://doi.org/10.5194/esd-2020-49.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Schneider, A., M. Flanner, and J. Perket, 2018: Multidecadal variability in surface albedo feedback across CMIP5 models. Geophys. Res. Lett., 45, 19721980, https://doi.org/10.1002/2017GL076293.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Screen, J. A., C. Deser, and I. Simmonds, 2012: Local and remote controls on observed Arctic warming. Geophys. Res. Lett., 39, L10709, https://doi.org/10.1029/2012GL051598.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Shell, K. M., J. T. Kiehl, and C. A. Shields, 2008: Using the radiative kernel technique to calculate climate feedbacks in NCAR’s Community Atmospheric Model. J. Climate, 21, 22692282, https://doi.org/10.1175/2007JCLI2044.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Smith, C. J., and Coauthors, 2018: Understanding rapid adjustments to diverse forcing agents. Geophys. Res. Lett., 45, 12 02312 031, https://doi.org/10.1029/2018GL079826.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Soden, B. J., I. M. Held, R. C. Colman, K. M. Shell, J. T. Kiehl, and C. A. Shields, 2008: Quantifying climate feedbacks using radiative kernels. J. Climate, 21, 35043520, https://doi.org/10.1175/2007JCLI2110.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stroeve, J., and D. Notz, 2018: Changing state of Arctic sea ice across all seasons. Environ. Res. Lett., 13, 103001, https://doi.org/10.1088/1748-9326/aade56.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stroeve, J., 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, https://doi.org/10.1029/2012GL052676.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Stroeve, J., A. Barrett, M. Serreze, and A. Schweiger, 2014: Using records from submarine, aircraft and satellites to evaluate climate model simulations of Arctic sea ice thickness. Cryosphere, 8, 18391854, https://doi.org/10.5194/tc-8-1839-2014.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Swart, N. C., J. C. Fyfe, E. Hawkins, J. E. Kay, and A. Jahn, 2015: Influence of internal variability on Arctic sea-ice trends. Nat. Climate Change, 5, 8689, https://doi.org/10.1038/nclimate2483.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Taylor, K. E., R. J. Stouffer, and G. A. Meehl, 2012: An overview of CMIP5 and the experiment design. Bull. Amer. Meteor. Soc., 3, 485498, https://doi.org/10.1175/BAMS-D-11-00094.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thackeray, C. W., and C. G. Fletcher, 2016: Snow albedo feedback: Current knowledge, importance, outstanding issues and future directions. Prog. Phys. Geogr., 40, 392408, https://doi.org/10.1177/0309133315620999.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thackeray, C. W., and A. Hall, 2019: An emergent constraint on future Arctic sea-ice albedo feedback. Nat. Climate Change, 9, 972978, https://doi.org/10.1038/s41558-019-0619-1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thackeray, C. W., C. G. Fletcher, and C. Derksen, 2014: The influence of canopy snow parameterizations on snow albedo feedback in boreal forest regions. J. Geophys. Res. Atmos., 119, 98109821, https://doi.org/10.1002/2014JD021858.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thackeray, C. W., C. G. Fletcher, and C. Derksen, 2015: Quantifying the skill of CMIP5 models in simulating seasonal albedo and snow cover evolution. J. Geophys. Res. Atmos., 120, 58315849, https://doi.org/10.1002/2015JD023325.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thackeray, C. W., C. G. Fletcher, L. R. Mudryk, and C. Derksen, 2016: Quantifying the uncertainty in historical and future simulations of Northern Hemisphere spring snow cover. J. Climate, 29, 86478663, https://doi.org/10.1175/JCLI-D-16-0341.1.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Thackeray, C. W., X. Qu