1. Introduction
Temperature inversions in the lower troposphere are a common feature of the Arctic climate (Serreze et al. 1992; Liu et al. 2006; Zhang et al. 2011), and these surface-based inversions are more frequent and more stable during the winter months compared to summer months (Tjernstrom and Graversen 2009; Zhang et al. 2011). Regional and global climate models (GCMs) have difficulties representing Arctic inversions (Dethloff et al. 2001; Tjernstrom et al. 2008; Boé et al. 2009; Kay et al. 2011; Medeiros et al. 2011; Pithan et al. 2014).
Boé et al. (2009) found that many of the models in the phase 3 of the Coupled Model Intercomparison Project (CMIP3) have an overly stable lower troposphere when compared to reanalysis data. In further evaluating the lower tropospheric stability (LTS) in the CMIP3 models, Medeiros et al. (2011) determined that the Arctic LTS of 21 GCMs and reanalysis data are bimodal. There was a stable mode over the Arctic Ocean and adjacent continents, and an unstable mode over the North Atlantic Ocean. Medeiros et al. (2011) found that about half the climate models examined had an Arctic averaged overly stable LTS because of the bias in the partitioning of the two modes and the other half had an Arctic averaged overly stable LTS because of the bias in LTS in the stable mode. Pithan et al. (2014) examined models from phase 5 of the Coupled Model Intercomparison Project (CMIP5; Taylor et al. 2012) and found an overly stable lower troposphere occurred in many of these updated GCMs, and theorized that the overly stable Arctic winter is due to a misrepresentation of low Arctic mixed-phase clouds, which controlled the surface net longwave flux.
Representing the Arctic LTS in climate models is important because the mean LTS may be related to the amount of Arctic climate change in an enhanced CO2 world (Boé et al. 2009; Bintanja et al. 2011, 2012). Boé et al. (2009) found a linear relationship between GCMs’ mean Arctic LTS and the Arctic longwave (LW) feedback in the CMIP3 archive, and suggested that GCMs have an unrealistic Arctic negative LW feedback because GCMs have a too stable Arctic lower troposphere. Bintanja et al. (2011) examined the relationship between the mean LTS state and Arctic climate change in an enhanced CO2 world in the EC-Earth Consortium GCM (EC-EARTH) by altering the stable boundary layer mixing parameterization. As expected, lower mixing values resulted in a more stable mean Arctic LTS, whereas more mixing resulted in a less stable Arctic LTS in the GCM. When increasing CO2 in EC-EARTH with these different values of boundary layer mixing, the run with a more stable LTS in the mean state had a greater amount of Arctic surface warming and a greater amount of sea ice loss. Bintanja et al. (2011) theorized that the model with a more stable LTS allowed less Arctic low-level heat to be released into the free troposphere, and hence greater surface temperature change at the surface.
The above studies analyzed Arctic LTS in models, but particular points of uncertainty in GCM LTS biases remain. In particular, this paper explores three questions:
Do GCMs have biases in Arctic LTS during all periods of the year? Because of the frequency of Arctic inversions during periods of no insolation, many Arctic LTS studies have focused on winter months (Boé et al. 2009; Medeiros et al. 2011; Pithan et al. 2014). In addition, Kay et al. (2011) found biases in Arctic LTS in July examining the National Center for Atmospheric Research’s (NCAR’s) Community Atmospheric Model version 4 (CAM4). A study analyzing Arctic LTS biases throughout the year has not been published.
What atmospheric level or levels may be causing an Arctic bias in LTS? LTS is defined as the temperature or potential temperature difference between a level above the inversion and the surface. The GCM bias may be due a temperature bias above the inversion, at the surface, or of a combination of biases at both levels. Many studies have found that GCMs produce Arctic surface temperatures that are lower than observations (Chapman and Walsh 2007; Liu et al. 2011; Xie et al. 2013), which are caused by surface longwave radiation and cloud biases (Svensson and Karlsson 2011; Cesana et al. 2012; Pithan et al. 2014). Pithan et al. (2014) demonstrated that the net longwave radiation is very much related to the representations of Artic LTS in climate models, but did not examine biases above the inversion. Is the known GCM surface temperature biases the main cause of the Arctic LTS bias or is there an additional bias above the inversion?
What are possible causes of the Arctic LTS bias? As mentioned in Bintanja et al. (2011), the stability is controlled by the boundary layer mixing in GCMs, and difficulties in the parameterization of boundary layer mixing may lead to Arctic LTS biases in GCMs. With many representations of the mixing in the stable boundary layer, mixing tends to shut off as stability increases, which may cause the surface to decouple from the troposphere, leading to runaway cooling of the surface (Derbyshire 1999; Steeneveld et al. 2006; Holtslag et al. 2013), and overly stable conditions may occur. Some operational models may arbitrarily increase the amount of stable mixing to avoid surface temperature biases under stable conditions (Sandu et al. 2013), but such tunings may not be present in all models.
An alternate but not exclusive explanation of why Arctic LTS is incorrect in GCMs, which we investigate in this paper, is that there is a misrepresentation in the radiative driving of the surface temperature, particularly a misrepresentation of surface downwelling radiation, which may be impacted by clouds. Arctic clouds greatly regulate surface temperature in the Arctic (Shupe and Intrieri 2004), and models routinely have a difficult time representing Arctic clouds correctly (Prenni et al. 2007; Tjernstrom et al. 2008; Klein et al. 2009; Morrison et al. 2009; Pithan et al. 2014). In particular, many models produce too little liquid in low-level Arctic clouds, which creates a cold surface temperature compared to observations (Liu et al. 2011; Xie et al. 2013; de Boer et al. 2014) and can lead to an overly stable atmosphere. If errors in clouds contribute to significant errors in the downward longwave radiation, then we need to determine which characteristics of clouds are most responsible. In fact, Pithan et al. (2014) suggested that a misrepresentation of mixed-phase clouds during the Arctic winter relates to errors in the CMIP5 models’ LTS. How much is the cloud error found in previous research related to Arctic LTS biases?
This study differs from many previous studies because we use GCM hindcast runs (Phillips et al. 2004) and free-running climate models to analyze Arctic LTS throughout the year in a subset of CMIP5 models. Because hindcasts are initialized with operational weather analyses, the analysis of hindcast runs allows for a determination of initial error in climate models. Multiple feedbacks may lead a GCM bias to worsen, but these feedbacks may also lead to compensating errors in which biases are reduced; either way, hindcast runs aid in isolating the factors contributing to GCM biases. Hindcast runs also allow for a comparison with observations at specific time steps and locations, something that is not as easily done with the typical GCM diagnosis that generally relies on comparisons of the statistics from decades of model output with observations or reanalysis. For some fields, such as surface temperature and cloud properties, model biases are evident in shorter time frames of about one year in a typical GCM diagnosis. We analyze high-frequency output from the hindcast simulations, which allows for a more process-driven approach in determining biases. This is particularly helpful in this study since the Arctic surface is dominated by a bimodal distribution of net surface LW radiation apparent in hourly data (Persson et al. 1999; Stramler et al. 2011; Morrison et al. 2012; Engström et al. 2014). This bimodal distribution is masked by monthly averages, but may be important for simulating the Arctic climate.
In this paper, we first describe the hindcast modeling runs (section 2a). We also examine the same GCMs in a free-running mode to relate to the hindcast modeling runs, and a description of these models is found in section 2b. Next, a description of operational analysis, reanalysis, and in situ data is presented (section3), and we discuss methods and variable definitions (section 4). We first present results of the hindcast runs for areas poleward of 70°N (section 5a) and at Barrow, Alaska (section 5b). Then we present the results of free-running climate integrations of the same models (section 5c). Last, we discuss and reiterate the main conclusions of this paper (section 6).
2. Model runs
a. Hindcast runs
Hindcast runs from the Transpose Atmospheric Model Intercomparison Project phase II (T-AMIP) experiment and internal runs at Lawrence Livermore National Laboratory (LLNL) are used in this paper. T-AMIP is a model intercomparison experiment for GCMs run in hindcast mode (Williams et al. 2013) with a goal to understand how biases grow in climate models from a well-initialized state. The hindcast method emphasizes model errors in the moist processes because the analyses used to initialize models constrain the large-scale dynamics to be close to the observed analysis (Phillips et al. 2004; Xie et al. 2012).
The models we use in the T-AMIP experiment include HadGEM2-A, IPSL-CM5A-LR, CNRM-CM5, and MIROC5 (see Table 1, which includes expansions of model names). The T-AMIP seasons started on 15 October 2008, 15 January 2009, 15 April 2009, and 15 July 2009 at 0000 UTC, and 16 hindcast runs were performed. Each run was started 30 h apart from each other, produced 5-day forecasts, and 3-h mean output was archived. These runs were initialized by the European Centre for Medium-Range Weather Forecasts (ECMWF) Year of Tropical Convection (ECMWF-Y) analysis data, and sea ice concentrations were from monthly observed values. The model output from T-AMIP is available from the Earth System Grid Federation (ESGF), and more detailed information is presented in Williams et al. (2013).
Model reference and expansion for each GCM used in this study.
In addition, the Community Atmospheric Model versions 4 and 5 (CAM4 and CAM5) from NCAR were run at LLNL in hindcast mode. These runs were initialized at 0000 UTC every day from May 2008 to March 2010 and produced 5-day forecasts. These are the same runs used in Barton et al. (2012), Xie et al. (2012), and Ma et al. (2013). The initialization data for CAM4 and CAM5 are the same as the T-AMIP initialization data (i.e., the ECMWF-Y analysis), and 10-day averages were used for observed sea ice. To be consistent in the data analysis, CAM4 and CAM5 runs are only analyzed during the seasons laid out by the T-AMIP runs. There are more data points in the CAM4 and CAM5 output because these models were initialized every day instead of once every 30 h. Official T-AMIP CAM4 model runs will be released by the ESGF in the future, but these runs were not available at the time of this study.
b. AMIP runs
To compare with the hindcast results, we analyze free-running or Atmospheric Model Intercomparison Project (AMIP) runs for the models listed in Table 1. AMIP runs have monthly-observed sea surface temperatures and sea ice concentrations as boundary conditions. We analyze daily output from the AMIP runs in which atmospheric composition emulates the twentieth century (i.e., twentieth-century historical AMIP runs), and the output from these runs is available at ESGF under the CMIP5 (Taylor et al. 2012) archive.
3. Operational analysis, reanalysis, and in situ data
To evaluate the hindcasts at a large-scale, we compare models to the ECMWF operational analyses for the Year of Tropical Convection (Waliser et al. 2012). These are the same data used to initialize the hindcast runs. We use the Climate Modeling Best Estimate (CMBE) data (Xie et al. 2010) at Barrow [e.g., the North Slope of Alaska (NSA)] to compare these runs against in situ observations. The data from Barrow were collected from a long-term measurement site of the U.S. Department of Energy’s Atmospheric Radiation Measurement (ARM) program. Lastly, to evaluate the AMIP simulations at larger temporal averages than the ~2 yr of ECMWF-Y, we use the Interim ECMWF Re-Analysis (ERA-Interim, hereafter ERA-I) output (Dee et al. 2011).
Because we evaluate the GCMs against the ECMWF-Y analysis, the question arises as to how well the ECMWF-Y analysis represents the components of LTS. Bias, mean absolute bias (MAE), and root-mean-square error (RMSE) between ECMWF-Y and the NSA LTS (
4. Methods
a. Temporal period of analysis
Conclusions from the hindcast runs are similar with the T-AMIP defined seasons of autumn–winter and spring–summer. This is because the Arctic T-AMIP results in autumn–winter are during the Arctic polar night and the spring–summer results are during the Arctic polar day. Insolation is of first order when analyzing the temperature structure of the lower Arctic atmosphere and other seasonal characteristics, such as albedo and polar heat transport, are of second order in the hindcast runs. For this reason we analyze the hindcast runs based on whether the sun was above or below the horizon. For each time step, data averaged poleward of 70°N of the hindcast simulation are labeled either as polar night or polar day based upon an average solar zenith angle (SZA) calculated for all modeled longitudes at a latitude of 80°N. All output were analyzed on the models’ given grid and not interpolated to a common grid. For the comparison with the CMBE-NSA data, the SZA, and hence polar night and polar day, are calculated at each time step using the latitude and longitude at Barrow, Alaska.
The AMIP runs are analyzed during the same temporal periods that overlap with the ERA-I product (January 1979–December 2008), and polar night and day are defined the same way as in the T-AMIP runs, but using daily output instead of 3-hourly output.
b. Lower tropospheric stability
Lower tropospheric stability is defined as the temperature or potential temperature difference at a height above the boundary layer and at a layer near the surface. The surface is usually defined at 1000 hPa for the lower latitude oceans (Klein and Hartmann 1993). In the Arctic, the surface temperature is a more ideal indicator of LTS than 1000 hPa because the surface temperature can be lower than the temperature at 1000 hPa. In this study, we define the surface as the 2-m air temperature, which is a standard output for GCMs, operational analysis, reanalysis products, and in situ observations, and use the surface pressure in the potential temperature equation instead of 1000 hPa. The main conclusions would not change if surface skin temperature were used in place of 2-m air temperature because these variables almost have a 1 to 1 relationship in these models.
c. Longwave cloud radiative effect
While 
Defining CRELW with only downward (instead of net) fluxes follows previous research (Shupe and Intrieri 2004; Barton and Veron 2012) and is a measure of the instantaneous effect of clouds on the surface energy budget. The CRELW values are generally positive because the 
d. Cloud properties
If CRELW significantly relates to Arctic LTS biases, it is important to understand which cloud properties contribute to the relationship. Cloud properties that affect averaged CRELW, which we examine in this paper, include cloud cover amount (both frequency and averaged amount), and cloud liquid water path (LWP) and cloud ice water path (IWP). Because of the lack of available model data, the effects of cloud particle size on longwave radiation (Garrett and Zhao 2006) are not considered. In the Arctic, the bimodal temporal distribution of hourly surface LW radiation relates to differences in humidity and may relate to differences in clouds (Persson et al. 1999; Stramler et al. 2011; Morrison et al. 2012; Engström et al. 2014); hence, the temporal distribution (i.e., frequency) of cloud properties are quantified for a comparison with surface temperature.
We compute the fraction of grid boxes that LWP and IWP in the 3-hourly (or daily for the AMIP runs) output exceed specified thresholds, to roughly represent the fraction of time that the atmosphere contains an opaque cloud. The LWP and IWP thresholds are determined by the LWP and IWP amount that affects surface 
The IWP threshold for opaqueness is suggested by Ebert and Curry (1992). They modeled cloud emissivity as a function of cloud IWP and ice effective radius and, through their analysis, we choose an IWP threshold for 25 g m−2 of cloud opaqueness caused by ice. In Ebert and Curry (1992), clouds with IWPs at 25 g m−2 had emissivities ranging from 0.5 to >0.9 with effective radii ranging from 20 to 90 μm. The T-AMIP model output does not have ice cloud effective radii data to compare with the modeling results of Ebert and Curry (1992), but changing the IWP threshold to 40 g m−2 does not change the conclusions of this paper. In addition, CMBE-NSA data do not have IWP observations so measurement uncertainty is not discussed. In addition, we analyze the frequency that either the LWPs are above 20 g m−2 or IWPs are greater than 25 g m−2 to determine a full opaqueness of the atmosphere due to clouds. In passing, we note that we are using the averaged 3-hourly gridbox mean LWP or IWP output instead of computing whether an in-cloud value of LWP or IWP exceeds the threshold and then multiplying by the cloud fraction for the 3-hourly period. The difference in frequency of opaque cloud calculated with this alternate approach is small.
5. Results
a. Hindcast biases poleward of 70°N
1) Biases in LTS, potential temperature at 850 hPa, and near-Surface temperature
For each 3-h temporal period, variables were averaged poleward of 70°N, and then averaged for polar night and polar day periods. Averaged polar night and polar day biases of LTS, potential temperature at 850 hPa, and surface temperature are shown in Fig. 1. Biases are more prominent during the polar night compared to the polar day for all models except CNRM-CM5, which has small biases in both the polar night and polar day. During the polar night, CAM4, CAM5, HadGEM2-A, and IPSL-CM5A-LR all have a more stable Arctic lower troposphere compared to the ECMWF-Y analysis, in agreement with the previous findings that climate models tend to have overly stable conditions in polar regions (Boé et al. 2009; Medeiros et al. 2011; Pithan et al. 2014). In contrast, the MIROC5 model exhibits very different behavior by drifting toward a less stable lower troposphere and a LTS bias of −2.7 K during the polar day and −6.2 K during the polar night on hindcast day 5. Day 1 is defined by hindcast hours 3–24. In most models, the bias nearly saturates by the hindcast day 2 and the biases are not much larger in during days 3–5. Two exceptions are CAM5 and MIROC5, in which biases continue to grow with forecast day. Relatively small biases in the day-1 hindcast in CAM4, CAM5, HadGEM2-A, and CNRM-CM5 suggest that the initialization is successful setting LTS to the analysis value.
Hindcast model biases of Arctic region: (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature averaged for periods of (left) polar day and (right) polar night. The Arctic is defined as poleward of 70°N. The biases are calculated from the ECMWF-Y analysis. The x axis displays the bias for each day since the start of the hindcast and the y axis displays the bias for each model. Hours 3–24 of each hindcast represent day 1. For the HadGEM2-A, IPSL-CM5-LR, CNRM-CM5, and MIROC5 models, 16 hindcast simulations starting 30 h apart from each other were performed starting on 15 October 2008 (autumn), 15 January 2009 (winter), 15 April 2009 (spring), and 15 July 2009 (summer) at 0000 UTC. CAM4 and CAM5 runs started each day at 0000 UTC for the T-AMIP seasons. The autumn and winter runs largely occurred during the Arctic polar night and the spring and summer runs occurred during the Arctic polar day, as defined by SZA.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Hindcast model biases of Arctic region: (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature averaged for periods of (left) polar day and (right) polar night. The Arctic is defined as poleward of 70°N. The biases are calculated from the ECMWF-Y analysis. The x axis displays the bias for each day since the start of the hindcast and the y axis displays the bias for each model. Hours 3–24 of each hindcast represent day 1. For the HadGEM2-A, IPSL-CM5-LR, CNRM-CM5, and MIROC5 models, 16 hindcast simulations starting 30 h apart from each other were performed starting on 15 October 2008 (autumn), 15 January 2009 (winter), 15 April 2009 (spring), and 15 July 2009 (summer) at 0000 UTC. CAM4 and CAM5 runs started each day at 0000 UTC for the T-AMIP seasons. The autumn and winter runs largely occurred during the Arctic polar night and the spring and summer runs occurred during the Arctic polar day, as defined by SZA.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Hindcast model biases of Arctic region: (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature averaged for periods of (left) polar day and (right) polar night. The Arctic is defined as poleward of 70°N. The biases are calculated from the ECMWF-Y analysis. The x axis displays the bias for each day since the start of the hindcast and the y axis displays the bias for each model. Hours 3–24 of each hindcast represent day 1. For the HadGEM2-A, IPSL-CM5-LR, CNRM-CM5, and MIROC5 models, 16 hindcast simulations starting 30 h apart from each other were performed starting on 15 October 2008 (autumn), 15 January 2009 (winter), 15 April 2009 (spring), and 15 July 2009 (summer) at 0000 UTC. CAM4 and CAM5 runs started each day at 0000 UTC for the T-AMIP seasons. The autumn and winter runs largely occurred during the Arctic polar night and the spring and summer runs occurred during the Arctic polar day, as defined by SZA.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
The biases in 
Medeiros et al. (2011) stated that Arctic LTS is a bimodal distribution with higher LTS values occurring over the Arctic sea ice and land, and lower LTS values occurring over the Arctic open ocean. In addition, Pavelsky et al. (2011) found that mean annual sea ice concentrations correlate well with the December–February mean Arctic LTS (r = 0.88). Because the mean Arctic LTS state is dependent on sea ice, are the biases also dependent on surface type? In Fig. 2, the Arctic LTS, potential temperature at 850 hPa, and near-surface temperature biases are displayed averaged over regions of sea ice, land, and open water for regions poleward of 70°N during the polar night period. The polar night period is only shown because of the relatively large biases compared to the polar day. The largest biases occurred over sea ice–covered regions compared to regions of land and open water. The relatively small biases over open water are expected because these models are forced with observed SSTs.
Hindcast model biases of Arctic region (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature averaged for polar night periods averaged over (left) sea ice, (center) land, and (right) water. The biases are calculated from the ECMWF-Y analysis. The x axis displays the bias for different hindcast days and the y axis displays the bias for each model. Sea ice regions are defined by concentrations >15%.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Hindcast model biases of Arctic region (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature averaged for polar night periods averaged over (left) sea ice, (center) land, and (right) water. The biases are calculated from the ECMWF-Y analysis. The x axis displays the bias for different hindcast days and the y axis displays the bias for each model. Sea ice regions are defined by concentrations >15%.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Hindcast model biases of Arctic region (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature averaged for polar night periods averaged over (left) sea ice, (center) land, and (right) water. The biases are calculated from the ECMWF-Y analysis. The x axis displays the bias for different hindcast days and the y axis displays the bias for each model. Sea ice regions are defined by concentrations >15%.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Even though the LTS biases over sea ice are larger than over land, these biases are generally of the same sign in a given model. The IPSL-CM5A-LR model is an exception in which over sea ice the LTS drift is toward a more stable lower troposphere and there is little LTS bias over land. The relatively small amount of grid cells over land makes the land results difficult to interpret, but this suggests that the biases are not caused by processes in the sea ice or land. As displayed in Fig. 1, the LTS biases over sea ice are related to the near-surface temperature bias. Hereafter we focus on polar night over sea ice when examining the hindcast modeling runs because of the relative larger biases.
2) Model spread in near-surface temperatures
(i) Comparison with LW radiation
Polar night biases in LTS over sea ice relate to biases in the near-surface temperature. The correlation coefficient between these area-averaged variables for each model and hindcast day is −0.993. Because of the connection between Arctic LTS and surface temperature bias and the relatively small biases at 850 hPa, we focus on understanding the spread in surface temperatures across the models and hindcast days.
The spread in surface temperature correlates to the spread in surface downwelling longwave radiation 
Scatterplots between surface temperature and (a) downwelling longwave radiation at the surface (
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Scatterplots between surface temperature and (a) downwelling longwave radiation at the surface (
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Scatterplots between surface temperature and (a) downwelling longwave radiation at the surface (
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Is the intermodel spread in surface temperature, and hence 
(ii) Comparison of clear-sky LW radiation using RRTMG
Because these models were initialized with the same temperature and moisture profiles one would expect the spread in clear-sky 
This spread may be due to differences in temperature and humidity that develop during the first 3 h of the hindcast, the radiation code, and/or other factors that affect LW radiation, such as aerosols, trace gases, or even the vertical resolution of the model. To test how much of the initial difference is due to differences in the temperature and humidity profiles present even on the first day, we ran RRTMG using temperature and humidity profiles averaged poleward of 70°N for each model and compared clear-sky 
As in Fig. 3, but for clear-sky longwave down output from each model that is on the y axis and clear-sky longwave down output from RRTMG that is on the x axis; and hindcast hours 3–21 are displayed instead daily averages. The solid black line is the one-to-one line. The gray shading represents the clear-sky longwave downwelling radiation from RRTMG using the ECMWF-Y analysis. The error bars represent the 95% confidence interval. The short black bars along the axes represent the range of the hour-3 output.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, but for clear-sky longwave down output from each model that is on the y axis and clear-sky longwave down output from RRTMG that is on the x axis; and hindcast hours 3–21 are displayed instead daily averages. The solid black line is the one-to-one line. The gray shading represents the clear-sky longwave downwelling radiation from RRTMG using the ECMWF-Y analysis. The error bars represent the 95% confidence interval. The short black bars along the axes represent the range of the hour-3 output.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, but for clear-sky longwave down output from each model that is on the y axis and clear-sky longwave down output from RRTMG that is on the x axis; and hindcast hours 3–21 are displayed instead daily averages. The solid black line is the one-to-one line. The gray shading represents the clear-sky longwave downwelling radiation from RRTMG using the ECMWF-Y analysis. The error bars represent the 95% confidence interval. The short black bars along the axes represent the range of the hour-3 output.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
In the RRTMG calculations, clouds were set to zero, ozone and other gases were the same for each model, and the vertical grids of each model were first interpolated to the ECMWF-Y grid. We also ran RRTMG using each model’s native vertical grid, but the main conclusions did not change. For reference, we also include the result of RRTMG calculations using the temperature and water vapor fields from ECMWF-Y analysis in Fig. 4. At forecast hour 3, these RRTMG calculations had a clear-sky 
We note that most of the intermodel spread of 4 W m−2 in clear-sky 
(iii) Comparison with cloud properties
In Fig. 5, we analyze the spread in total cloud cover, LWP, IWP, frequency of clouds with LWPs greater than 20 g m−2, frequency of clouds with IWPs greater than 25 g m−2, and the frequency of clouds with LWPs greater than 20 g m−2 or IWPs greater than 25 g m−2 in relation to the model’s surface temperature. All data are averaged over sea ice regions during polar night. The linear relationships between surface temperature and total cloud cover or IWP are not statistically significant at the 0.05 level. CAM5 has relatively high values of total cloud cover but low values of surface temperature, whereas MIROC5 is near the median amount of total cloud cover but has high surface temperatures. The IPSL-CM5A-LR, HadGEM2-A, and MIROC5 models have relatively high values of IWPs, but the range in surface temperatures among the models is close to 10 K.
As in Fig. 3, but for (a) total cloud cover, (b) all-sky liquid water path, (c) all-sky ice water path, (d) the frequency of the LWPs > 20 g m−2, (e) the frequency of IWPs > 25 g m−2, and (f) the frequency that LWPs were >20 g m−2 or IWPs were >25 g m−2.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, but for (a) total cloud cover, (b) all-sky liquid water path, (c) all-sky ice water path, (d) the frequency of the LWPs > 20 g m−2, (e) the frequency of IWPs > 25 g m−2, and (f) the frequency that LWPs were >20 g m−2 or IWPs were >25 g m−2.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, but for (a) total cloud cover, (b) all-sky liquid water path, (c) all-sky ice water path, (d) the frequency of the LWPs > 20 g m−2, (e) the frequency of IWPs > 25 g m−2, and (f) the frequency that LWPs were >20 g m−2 or IWPs were >25 g m−2.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
The linear trend in the averaged LWP and the surface temperature is statistically significant with an r value of 0.579 and a p value less than 0.05. Larger LWPs occur in the models with higher surface temperatures. One exception is CAM4, which has higher mean LWPs compared to the other models, but lower surface temperatures. There are two models (CAM5 and HadGEM2-A) that have averaged LWPs close to zero in the polar night. Arctic observations show liquid in clouds occur throughout the polar night over sea ice (Shupe et al. 2011). The models’ LWPs are not correlated with the models’ IWPs. For example, IPSL-CM5A-LR and HadGEM2-A have high IWP values but low LWP values during this period of analysis. In addition, CAM4 has higher values of LWP compared to other models, and a median value of IWP.
For CAM4, we note that there is a disconnect between mean LWPs, 
The regression between surface temperature and the frequency of clouds with LWPs greater than 20 g m−2 is statistically significant, and the r value of 0.776 is close to the r value between surface temperature and CRELW (r = 0.821) and greater than the r value between surface temperature and mean LWP (r = 0.579). This suggests that the surface radiative effect over sea ice of clouds in these hindcast simulations is largely controlled by the frequency that a liquid cloud occurs during the polar night. CAM4 is still an outlier in the analysis with a relatively high frequency of clouds with LWPs greater than 20 g m−2 but low surface temperatures. CAM5, HadGEM2-A, and IPSL-CM5A-LR have a frequency of clouds with LWPs greater than 20 g m−2 that is less than 20%. The CNRM-CM5 model has a frequency near 40%–50% and surface temperatures near the ECMWF-Y analysis. The MIROC5 model has frequencies slightly above 50%, which relates to a high mean surface temperature compared the other models and ECMWF-Y analysis.
The regression between surface temperature and the frequency of clouds with IWPs greater than 25 g m−2 is not statistically significant; as with the mean IWP comparison, the models with a high frequency of clouds with IWP greater than 25 g m−2 are not the same models with a high frequency of clouds with LWPs greater than 20 g m−2. A lack of relationship of CRELW with the frequency of IWP greater than 25 g m−2 may reflect the fact that high-altitude ice clouds may less efficiently change the surface 
b. Hindcast biases at Barrow
1) Biases in LTS, potential temperature at 850 hPa, and near-surface temperature
A benefit of hindcast climate model simulations is that these runs can be compared to in situ observations at specific times. This has been done in previous Arctic research (Xie et al. 2008; Kay et al. 2011; Liu et al. 2011; Barton et al. 2012), in which biases have been observed and possible solutions to the problems put forth. In addition, we did not examine 
Variables for each model are examined at the grid point closest to the North Slope of Alaska site at Barrow and compared to the Climate Modeling Best Estimate data. Compared to the analysis conducted poleward of 70°N with the ECMWF-Y analysis, the biases for the climate models are not as large, but are comparable to the polar night land biases for five of the six models with all models having the same sign except HadGEM2-A (Fig. 6). Similar biases are not necessarily expected at the NSA because the analysis is at local region compared to averages poleward of 70°N, the CMBE-NSA data are located over land, and there are differences in model horizontal resolution.
As in Fig. 3, but showing the scatter between LTS (y axis) and near-surface temperature (x axis) at the North Slope of Alaska (NSA).
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, but showing the scatter between LTS (y axis) and near-surface temperature (x axis) at the North Slope of Alaska (NSA).
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, but showing the scatter between LTS (y axis) and near-surface temperature (x axis) at the North Slope of Alaska (NSA).
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
The analysis at the NSA shows the similar linear relationship between LTS and surface temperature in the polar night (Fig. 6). The linear regression of the models’ hindcast value for each day between LTS and near surface temperature is −0.979. Similar to the analysis poleward of 70°N, the MIROC5 model has higher surface temperatures and a less stable lower troposphere compared to CMBE-NSA data. CAM4, CAM5, and IPSL-CM5A-LR have LTS values slightly larger than the mean LTS value from the CMBE-NSA data similar to the analysis poleward of 70°N, but many of these models’ hindcast days lie in the 95% confidence interval of the CMBE-NSA data. HadGEM2-A has the opposite sign in the CMBE-NSA results compared to the analysis performed poleward of 70°N (i.e., Fig. 2). The linear relationship between LTS and surface temperature leads to the conclusion that an analysis of the relationship between the surface temperature biases and other variables at NSA would be relevant to the analysis for the land and sea ice–covered Arctic.
2) Model spread in near-surface temperatures
(i) Comparison with LW radiation
At the NSA, there are observed values of 
As in Fig. 3, but the analysis is performed at the North Slope of Alaska (NSA). Note that the (a) 
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, but the analysis is performed at the North Slope of Alaska (NSA). Note that the (a)
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, but the analysis is performed at the North Slope of Alaska (NSA). Note that the (a)
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
(ii) Comparison with cloud properties
We further analyze which cloud properties relate to the spread in surface temperatures in the hindcast models (Fig. 8). Similar to the analysis poleward of 70°N, total cloud cover and mean IWPs do not significantly relate to surface temperature at the NSA. CAM5 has the highest cloud cover amount, but the surface temperatures are relatively low. MIROC5 and IPSL-CM5A-LR have similar cloud cover amounts, but surface temperatures differ near 8 K.
As in Fig. 5, except the analysis is performed at the North Slope of Alaska (NSA). There are no ice water path (IWP) data at the NSA for this time period. The confidence interval for the NSA frequency of LWPs > 20 g m−2 was computed using bootstrapping by removing a data point and recalculating the frequency 1000 times.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 5, except the analysis is performed at the North Slope of Alaska (NSA). There are no ice water path (IWP) data at the NSA for this time period. The confidence interval for the NSA frequency of LWPs > 20 g m−2 was computed using bootstrapping by removing a data point and recalculating the frequency 1000 times.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 5, except the analysis is performed at the North Slope of Alaska (NSA). There are no ice water path (IWP) data at the NSA for this time period. The confidence interval for the NSA frequency of LWPs > 20 g m−2 was computed using bootstrapping by removing a data point and recalculating the frequency 1000 times.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Unlike the analysis poleward of 70°N, the linear relationship between near-surface temperature and mean LWP during the polar night is not statistically significant (Fig. 7b; r = 0.085, p = 0.654). Again, CAM4 is an outlier in the analysis with relatively large LWPs but relatively low surface temperatures compared to the other models. MIROC5 has the next largest values of cloud LWP (cf. CAM4), and MIROC5 has the largest values of surface temperatures. CAM5 and IPSL-CM5A-LR both have mean LWP values lower than the CMBE-NSA data and lower surface temperatures. Many CNRM-CM5 values of LWP and surface temperature lie in the uncertainty range of the CMBE-NSA data. The HadGEM2-A model has relatively low values of LWP, but relatively large surface temperatures. In addition, the HadGEM2-A model has relatively high values of IWP compared to the other models and the large amount of this cloud ice is occurring at altitudes less than 2 km (not shown). Low-level ice clouds aid in the relatively high temperature values with relatively low cloud liquid amount.
When examining the frequency of clouds with a LWP threshold of 20 g m−2, the relationship with surface temperature is more linear compared to the mean LWP values, but the results are not as statically significant as the values obtained from the analysis poleward of 70°N (NSA, r = 0.359, p = 0.052; poleward of 70°N, r = 0.776, p = 0.000). The CMBE-NSA observations have a frequency of LWP clouds in excess of the threshold of 30%–40%, and the CNRM-CM5 model has a similar frequency and a similar surface temperature even though CNRM-CM5 and ECMWF use different microphysics (CNRM-CM5: Ricard and Royer 1993; ECMWF: Forbes et al. 2011). MIROC5 has a frequency of opaque clouds greater than the CMBE-NSA data and surface temperatures greater than the CMBE-NSA data. CAM5 and HadGEM2-A have a frequency of LWP greater than 20 g m−2 less than 15%, but CAM5 has surface temperatures lower the CMBE-NSA observations whereas HadGEM2-A has surface temperatures greater than the observations.
When examining the frequency of clouds with IWPs greater than 25 g m−2, there is not a statistically significant relationship between this frequency and surface temperature, but some insights about the models are discovered. For example, the HadGEM2-A model has a relatively high frequency of IWPs greater than 25 g m−2, which explains, in addition to cloud ice height, why the surface temperature is relatively high even though the LWP threshold frequency is low. When examining the frequency of clouds with LWPs greater than 20 g m−2 or IWPs greater than 25 g m−2, the linear regression line has a higher r value (r = 0.435) than either the LWP or IWP threshold regressions.
Because of the contemporaneous nature of LWP and surface temperature data from the observations and models at NSA, one can stratify temperature biases according to the signs of the LWP bias in order to obtain a better sense of contribution of LWP biases to surface temperature biases at the NSA. Figure 9 examines the surface temperature bias during four distinct periods of Artic liquid cloud production compared to observations: 1) periods in which the models produce clouds with LWPs greater than 20 g m−2 and the observations have clouds with LWPs less than 20 g m−2, 2) periods in which the models and observations both have LWPs greater than 20 g m−2, 3) periods in which the models and observations both have LWPs less than 20 g m−2, and 4) periods in which the models have LWPs less than 20 g m−2 and the observations have LWPs greater than 20 g m−2 (Fig. 9). Considering the results in Fig. 8, analyzing the IWPs would have been beneficial, but IWP data at the NSA do not exist for this time period. Only day-2 hindcast results are shown in Fig. 9 for clarity.
Day-2 temperature biases for four periods of modeled cloud production compared to observations at the NSA: (a) periods in which the models produce clouds with LWPs > 20 g m−2 and the observations have clouds with LWPs < 20 g m−2, (b) periods in which the models and observations have clouds with LWPs > 20 g m−2, (c) periods in which the models and observations have clouds with LWPs < 20 g m−2, and (d) periods in which the models produced clouds with LWPs < 20 g m−2 and the observations have clouds with LWPs > 20 g m−2. The temperature bias is on the left y axis and the percent of occurrence that each model occurs in these regimes is on the x axis. The colored symbols represent biases for the regime that have p values < 0.05 defined by a two-tailed Student’s t test. The right y axis is the weighted bias and is calculated by the bias multiplied by the percent of occurrence.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Day-2 temperature biases for four periods of modeled cloud production compared to observations at the NSA: (a) periods in which the models produce clouds with LWPs > 20 g m−2 and the observations have clouds with LWPs < 20 g m−2, (b) periods in which the models and observations have clouds with LWPs > 20 g m−2, (c) periods in which the models and observations have clouds with LWPs < 20 g m−2, and (d) periods in which the models produced clouds with LWPs < 20 g m−2 and the observations have clouds with LWPs > 20 g m−2. The temperature bias is on the left y axis and the percent of occurrence that each model occurs in these regimes is on the x axis. The colored symbols represent biases for the regime that have p values < 0.05 defined by a two-tailed Student’s t test. The right y axis is the weighted bias and is calculated by the bias multiplied by the percent of occurrence.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
Day-2 temperature biases for four periods of modeled cloud production compared to observations at the NSA: (a) periods in which the models produce clouds with LWPs > 20 g m−2 and the observations have clouds with LWPs < 20 g m−2, (b) periods in which the models and observations have clouds with LWPs > 20 g m−2, (c) periods in which the models and observations have clouds with LWPs < 20 g m−2, and (d) periods in which the models produced clouds with LWPs < 20 g m−2 and the observations have clouds with LWPs > 20 g m−2. The temperature bias is on the left y axis and the percent of occurrence that each model occurs in these regimes is on the x axis. The colored symbols represent biases for the regime that have p values < 0.05 defined by a two-tailed Student’s t test. The right y axis is the weighted bias and is calculated by the bias multiplied by the percent of occurrence.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
MIROC5 has a mean surface temperature that is higher than the CMBE-NSA observations (e.g., Fig. 6), and these higher values occur during periods in which MIROC5 has LWPs greater than 20 g m−2 and the observations have LWPs less than 20 g m−2. In addition, higher surface temperatures compared to CMBE-NSA data occur when both MIROC5 and CMBE-NSA data have LWPs less than 20 g m−2. To analyze the relative contribution of each of these errors to the time-mean bias in surface temperature, we compute a weighted bias by multiplying the percentage of time that model occurs in the regime by the temperature bias in that regime. For MIROC5, the weighted bias during the period in which the model has LWPs greater than 20 g m−2 and the observations have LWPs less than 20 g m−2 is greater than the regime when both the MIROC5 and the CMBE-NSA data have LWPs less than 20 g m−2.
The surface temperatures in CAM4, CAM5, and IPSL-CM5A-LR are all lower than the CMBE-NSA observations (Fig. 6). The largest absolute values of the weighted bias for the models occur when the observations have LWPs greater than 20 g m−2 and the models have LWPs less than the threshold, and surface temperature biases are statistically different than zero for all of these models in this period (Fig. 9). The IPSL-CM5A-LR model also has a statistically significant difference when the model and CMBE-NSA observations both have LWPs greater than 20 g m−2, but the weighted difference is not as large. These results for CAM4, CAM5, IPSL-CM5A-LR, CNRM-CM5, and MIROC5 show that the surface temperature biases largely relate to when the models fail to produce the liquid cloud state when it was observed.
The HadGEM2-A model is an exception at the NSA. When compared to the CMBE-NSA data, the HadGEM2-A model has a positive surface temperature bias. The largest weighted bias in HadGEM2-A occurs during periods in which both HadGEM2-A and the observations have LWPs less than 20 g m−2. As shown in the IWP and IWP frequency scatterplots, the HadGEM2-A has relatively larger values and more frequent periods of large IWPs. During the period in which the HadGEM2-A and CMBE-NSA observations had LWPs less than 20 g m−2, HadGEM2-A IWPs were 43.4 g m−2, while every other model had IWPs less than 23 g m−2. Still considering how the warm bias in surface at NSA is not representative of the biases over larger areas in HadGEM2-A (Fig. 2), it is not clear how much emphasis should be placed on the results for this model at NSA. The NSA results using in situ data confirm the analysis over the Arctic domain.
c. Free-running model (AMIP) biases in Arctic LTS
Biases in LTS, potential temperature at 850 hPa, and near-surface temperature
How do these hindcast biases relate to biases of the same model run in AMIP mode? Xie et al. (2012), Williams et al. (2013), and Ma et al. (2014) determined that many fast physics AMIP modeling errors occur in hindcast simulations, but Arctic LTS was not specifically examined.
For the AMIP analysis, biases in each month are examined because data for all months are available. There are some similarities and differences between the biases in AMIP (Fig. 10) and hindcast (Fig. 1) mode. Similar to the hindcast runs, AMIP runs of CAM4, CAM5, and HadGEM2-A have larger LTS values compared to the ERA-I reanalysis during periods of nearly zero insolation. These LTS biases in CAM4, CAM5, and HadGEM2-A are largely due to the surface temperature being lower than ERA-I. As with the hindcast modeling runs, difficulties in representing Arctic LTS during the winter is related to difficulties in representing the surface temperature. CNRM-CM5 has similar LTS and surface temperature compared to the ERA-I reanalysis during the winter months, which is similar to the hindcast modeling results.
AMIP model biases for (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature. The biases are calculated from the ERA-I reanalysis. The vertical axis represents the individual model and the horizontal axis is the month of the bias.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
AMIP model biases for (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature. The biases are calculated from the ERA-I reanalysis. The vertical axis represents the individual model and the horizontal axis is the month of the bias.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
AMIP model biases for (top) lower tropospheric stability, (middle) potential temperature at 850 hPa, and (bottom) surface temperature. The biases are calculated from the ERA-I reanalysis. The vertical axis represents the individual model and the horizontal axis is the month of the bias.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
However, there are differences between the AMIP results and hindcast results including an enhanced LTS error during the summer months in CAM4, CAM5, HadGEM2-A, and CNRM-CM5 in the AMIP simulations, and the IPSL-CM5-LR and MIROC5 AMIP simulations having similar LTS and surface temperature values compared to ERA-I, whereas large biases occurred in the hindcast runs. In CAM4, CAM5, and HadGEM2-A, the summer LTS bias is a combination of the biases at 850 hPa and the surface. In these models, the surface is slightly colder than the ERA-I reanalysis and the temperature at 850 hPa is slightly warmer. In the CNRM-CM5 model, the summer LTS bias is largely due to the potential temperature at 850 hPa.
In four of the six models (CAM4, CAM5, HadGEM2-A, and CNRM-CM5), the winter–polar night LTS bias was similar in the AMIP runs compared to the hindcast runs. MIROC5 and IPSL-CM5A-LR have dissimilar biases between AMIP and hindcast during the winter period. The main conclusion of Xie et al. (2012), which states that fast physics (e.g., clouds, radiation) hindcast errors are well related to those errors manifested in AMIP simulations, seems only partially confirmed when examining the LTS errors in the Arctic region, but a larger sample size of different GCMs would be beneficial.
Comparison with LW radiation
To further compare the AMIP runs to the hindcast runs, we focus on polar night biases defined by daily data, and only examine output over sea ice because the largest LTS biases occurred over sea ice in the hindcast simulations. Surface temperature biases are compared to 
As in Fig. 3, except the analysis is performed on AMIP simulations from 1979 to 2008 using daily output poleward of 70°N and over sea ice regions. The gray shading represents the 95% confidence interval for the ERA-I reanalysis.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, except the analysis is performed on AMIP simulations from 1979 to 2008 using daily output poleward of 70°N and over sea ice regions. The gray shading represents the 95% confidence interval for the ERA-I reanalysis.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
As in Fig. 3, except the analysis is performed on AMIP simulations from 1979 to 2008 using daily output poleward of 70°N and over sea ice regions. The gray shading represents the 95% confidence interval for the ERA-I reanalysis.
Citation: Journal of Climate 27, 19; 10.1175/JCLI-D-14-00126.1
The differences between the AMIP and hindcast relationships are not surprising because of the multiple feedbacks that occur in AMIP simulations. For example, Song and Mapes (2012) suggested that day 30 and greater hindcast Arctic errors in the Climate Forecasting Model are related to errors in the thermal wind, which relates to the advection of heat and moisture into the Arctic.
6. Conclusions
This study examined global climate models (GCMs) run in hindcast and free-running mode to analyze errors in the Artic lower tropospheric stability. Because Arctic lower tropospheric stability may be related Arctic climate change (Boé et al. 2009; Bintanja et al. 2011, 2012), it is important to have a better understanding of lower tropospheric stability errors and hypothesize why these errors exist. The results related to the questions posed in the introduction are summarized below.
Do GCMs have biases in Arctic LTS during all periods of the year? We found that the Arctic lower tropospheric stability bias predominantly occurs during the polar night over sea ice regions in the hindcast simulations. AMIP simulations do not show clear seasonality in lower tropospheric stability biases as the hindcast simulations do, but similar biases occur during the winter months. Unlike Kay et al. (2011), who found a summer LTS bias over the Arctic ocean, we did not fully analyze the summer months because the largest bias occurred during the winter months.
What atmospheric level or levels may be causing an Arctic bias in LTS? The polar night Arctic lower tropospheric stability bias is related to biases in Arctic surface temperature. Very small biases occurred at the 850-hPa level. Artic surface temperatures lower than observations have been shown in previous research (Chapman and Walsh 2007; Liu et al. 2011; Xie et al. 2013; de Boer et al. 2014). Our result shows a strong linearity between Artic stability and surface temperature biases in hindcast model and AMIP output during the winter months and clearly connects these biases. AMIP summer biases relate to biases at the surface as well as temperature biases above the inversion layer.
What are possible causes of the Arctic LTS bias? Downwelling longwave radiation significantly relates to the models’ spread in surface temperature during the polar night in the hindcast simulations, which supports previous studies that have suggested a connection between surface longwave radiation and lower tropospheric stability (Pithan et al. 2014; de Boer et al. 2014). Clear-sky longwave radiation and the cloud radiative effect both relate to the spread in surface temperature in the hindcast models, and the production of cloud liquid water is shown to significantly relate to surface temperatures over the Arctic region. Previous studies examining Artic surface temperature biases similarly found that cloud liquid is a major driver of Arctic surface temperatures (Prenni et al. 2007; Tjernstrom et al. 2008; Svensson and Karlsson 2011; Liu et al. 2011; Pithan et al. 2014; de Boer et al. 2014). Ice water in Arctic clouds is important in the HadGEM2-A model when analyzing hindcast simulations at the North Slope of Alaska, but the lack of similarity to the Arctic domain analysis suggest that ice water amount in clouds is a secondary effect.
Conclusions from this study that were not proposed in the initial questions include the following:
Variables other than temperature, moisture, and clouds must be aiding in the spread of the models’ downwelling clear-sky longwave radiation. The relatively large spread of clear-sky downward longwave radiation was not expected because the hindcast models were forced by the same temperature and moisture profiles. We analyzed this spread by using the same radiative transfer code to calculate clear-sky downwelling longwave radiation from the models’ temperature and moisture profiles, and found that the models’ output of clear-sky longwave radiation had a much larger spread than the clear-sky longwave radiation from the same radiative transfer model.
The AMIP results during the polar night largely corresponded with the hindcast conclusions, but there were larger LTS errors during periods of insolation in the AMIP runs. Other variables than longwave radiation and liquid clouds must be affecting the model spread in AMIP–free-running simulations. In fact, Song and Mapes (2012) suggested that circulation errors contribute to Arctic surface temperature errors in 30-day hindcasts of a fully coupled atmosphere–ocean–land model, which is well beyond the up-to-5-day hindcasts examined here. Although beyond the scope of the T-AMIP modeling runs, it would be beneficial to determine if the same models that have Arctic temperature errors in the 5-day hindcast have similar errors at time periods in which the dynamics have a greater chance to feed back onto the system.
The analysis shows strong covariability between clear-sky longwave radiation, the cloud radiative effect, largely due to differences in liquid clouds, and surface temperature during the polar night periods. Models that produced larger clear-sky LW values also produced larger CRELW values and a higher frequency of clouds with LWPs greater than 20 g m−2. There is interplay between the clear sky and clouds that affect the Arctic polar night surface temperature in these T-AMIP models. The connection between clear-sky temperature, moisture, and clouds greatly affects the Arctic surface radiative budget (Francis et al. 2005; Stramler et al. 2011; Morrison et al. 2012), but it is not clear if clear-sky periods or cloudy periods drive the relationship. Determining whether the clear-sky or cloudy-sky errors drive the surface temperature bias is important and may be best suited for idealized standalone modeling studies.
It is well known that models have difficulties in reproducing Arctic clouds (Tjernstrom et al. 2008; Klein et al. 2009; Morrison et al. 2009). Pithan et al. (2014) suggested that the misrepresentations of mixed-phase clouds relate to biases in Arctic LTS in the CMIP5 archive. We add that in addition to cloud biases, clear sky longwave biases are also important when examining lower tropospheric stability errors. Improved simulation of the processes controlling downward longwave radiation, such as cloud microphysics, boundary layer mixing, and radiative transfer modeling, will lead to an improved simulation of Arctic climate. Improved downward longwave radiation will require an attention to mixed-phase clouds, frequency of liquid that creates opaque Arctic clouds, and processes controlling clear-sky longwave radiation.
Acknowledgments
The contribution of N. P. Barton, S. A. Klein, and J. S. Boyle to this work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. Support for N. P. Barton, S. A. Klein, and J. S. Boyle was provided by the Regional and Global Climate and Earth System Modeling Programs of the Office of Science at the U. S. Department of Energy. We acknowledge the Working Group on Numerical Experimentation (WGNE) and the Working Group on Coupled Modeling (WGCM), who are responsible for Transpose-AMIP II, and we thank the climate modeling groups (listed in Table 1 of this paper) for producing and making available their model output. For Transpose-AMIP II the U.S. Department of Energy’s Program for Climate Model Diagnosis and Intercomparison (PCMDI) provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals.
REFERENCES
Barton, N. P., and D. E. Veron, 2012: Response of clouds and surface energy fluxes to changes in sea-ice cover over the Laptev Sea (Arctic Ocean). Climate Res., 54, 69–84, doi:10.3354/cr01101.
Barton, N. P., S. A. Klein, J. S. Boyle, and Y. Y. Y. Zhang, 2012: Arctic synoptic regimes: Comparing domain-wide Arctic cloud observations with CAM4 and CAM5 during similar dynamics. J. Geophys. Res., 117, D15205, doi:10.1029/2012JD017589.
Bintanja, R., R. G. Graversen, and W. Hazeleger, 2011: Arctic winter warming amplified by the thermal inversion and consequent low infrared cooling to space. Nat. Geosci., 4, 758–761, doi:10.1038/ngeo1285.
Bintanja, R., E. C. van der Linden, and W. Hazeleger, 2012: Boundary layer stability and Arctic climate change: A feedback study using EC-Earth. Climate Dyn., 39, 2659–2673, doi:10.1007/s00382-011-1272-1.
Boé, J., A. Hall, and X. Qu, 2009: Current GCMs’ unrealistic negative feedback in the Arctic. J. Climate, 22, 4682–4695, doi:10.1175/2009JCLI2885.1.
Cesana, G., J. E. Kay, H. Chepfer, J. M. English, and G. de Boer, 2012: Ubiquitous low-level liquid-containing Arctic clouds: New observations and climate model constraints from CALIPSO-GOCCP. Geophys. Res. Lett., 39, L20804, doi:10.1029/2012GL053385.
Chapman, W. L., and J. E. Walsh, 2007: Simulations of Arctic temperature and pressure by global coupled models. J. Climate, 20, 609–632, doi:10.1175/JCLI4026.1.
Chen, Y. H., F. Aires, J. A. Francis, and J. R. Miller, 2006: Observed relationships between Arctic longwave cloud forcing and cloud parameters using a neural network. J. Climate, 19, 4087–4104, doi:10.1175/JCLI3839.1.
de Boer, G., and Coauthors, 2014: Near-surface meteorology during the Arctic Summer Cloud Ocean Study (ASCOS): Evaluation of reanalyses and global climate models. Atmos. Chem. Phys., 14, 427–445, doi:10.5194/acp-14-427-2014.
Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553–597, doi:10.1002/qj.828.
Derbyshire, S. H., 1999: Stable boundary-layer modelling: Established approaches and beyond. Bound.-Layer Meteor., 90, 423–446, doi:10.1023/A:1001749007836.
Dethloff, K., C. Abegg, A. Rinke, I. Hebestadt, and V. F. Romanov, 2001: Sensitivity of Arctic climate simulations to different boundary-layer parameterizations in a regional climate model. Tellus, 53, 1–26, doi:10.1034/j.1600-0870.2001.01073.x.
Dufresne, J. L., and Coauthors, 2013: Climate change projections using the IPSL-CM5 Earth System Model: From CMIP3 to CMIP5. Climate Dyn., 40, 2123–2165, doi:10.1007/s00382-012-1636-1.
Ebert, E. E., and J. A. Curry, 1992: A parameterization of ice-cloud optical properties for climate models. J. Geophys. Res., 97, 3831–3836, doi:10.1029/91JD02472.
Engström, A., J. Karlsson, and G. Svensson, 2014: The importance of representing mixed-phase clouds for simulating distinctive atmospheric states in the Arctic. J. Climate, 27, 265–272, doi:10.1175/JCLI-D-13-00271.1.
Forbes, R., A. Tompkins, and A. Untch, 2011: A new prognostic bulk microphsyics scheme for the IFS. ECMWF Tech. Memo. 649, 20 pp.
Francis, J. A., E. Hunter, J. R. Key, and X. J. Wang, 2005: Clues to variability in Arctic minimum sea ice extent. Geophys. Res. Lett., 32, L21501, doi:10.1029/2005GL024376.
Garrett, T. J., and C. F. Zhao, 2006: Increased Arctic cloud longwave emissivity associated with pollution from mid-latitudes. Nature, 440, 787–789, doi:10.1038/nature04636.
Gent, P. R., and Coauthors, 2011: The Community Climate System Model version 4. J. Climate, 24, 4973–4991, doi:10.1175/2011JCLI4083.1.
Holtslag, A. A. M., and Coauthors, 2013: Stable atmospheric boundary layers and diurnal cycles: Challenges for weather and climate models. Bull. Amer. Meteor. Soc., 94, 1691–1706, doi:10.1175/BAMS-D-11-00187.1.
Kay, J. E., and A. Gettelman, 2009: Cloud influence on and response to seasonal Arctic sea ice loss. J. Geophys. Res., 114, D18204, doi:10.1029/2009JD011773.
Kay, J. E., K. Raeder, A. Gettelman, and J. Anderson, 2011: The boundary layer response to recent Arctic sea ice loss and implications for high-latitude climate feedbacks. J. Climate, 24, 428–447, doi:10.1175/2010JCLI3651.1.
Klein, S. A., and D. L. Hartmann, 1993: The seasonal cycle of low stratiform clouds. J. Climate, 6, 1587–1606, doi:10.1175/1520-0442(1993)006<1587:TSCOLS>2.0.CO;2.
Klein, S. A., and Coauthors, 2009: Intercomparison of model simulations of mixed-phase clouds observed during the ARM Mixed-Phase Arctic Cloud Experiment. I: Single-layer cloud. Quart. J. Roy. Meteor. Soc., 135, 979–1002, doi:10.1002/qj.416.
Liu, X., and Coauthors, 2011: Testing cloud microphysics parameterizations in NCAR CAM5 with ISDAC and M-PACE observations. J. Geophys. Res., 116, D00T11, doi:10.1029/2011JD015889.
Liu, Y., J. R. Key, A. Schweiger, and J. Francis, 2006: Characteristics of satellite-derived clear-sky atmospheric temperature inversion strength in the Arctic, 1980–96. J. Climate, 19, 4902–4913, doi:10.1175/JCLI3915.1.
Ma, H.-Y., S. Xie, J. S. Boyle, S. A. Klein, and Y. Zhang, 2013: Metrics and diagnostics for precipitation-related processes in climate model short-range hindcasts. J. Climate, 26, 1516–1534, doi:10.1175/JCLI-D-12-00235.1.
Ma, H.-Y., and Coauthors, 2014: On the correspondence between mean forecast errors and climate errors in CMIP5 models. J. Climate, 27, 1781–1798, doi:10.1175/JCLI-D-13-00474.1.
Martin, G. M., and Coauthors, 2011: The HadGEM2 family of Met Office Unified Model climate configurations. Geosci. Model Dev., 4, 723–757, doi:10.5194/gmd-4-723-2011.
Medeiros, B., C. Deser, R. A. Tomas, and J. E. Kay, 2011: Arctic inversion strength in climate models. J. Climate, 24, 4733–4740, doi:10.1175/2011JCLI3968.1.
Mlawer, E. J., S. J. Taubman, P. D. Brown, M. J. Iacono, and S. A. Clough, 1997: Radiative transfer for inhomogeneous atmospheres: RRTM, a validated correlated-k model for the longwave. J. Geophys. Res., 102, 16 663–16 682, doi:10.1029/97JD00237.
Morrison, H., and Coauthors, 2009: Intercomparison of model simulations of mixed-phase clouds observed during the ARM Mixed-Phase Arctic Cloud Experiment. II: Multilayer cloud. Quart. J. Roy. Meteor. Soc., 135, 1003–1019, doi:10.1002/qj.415.
Morrison, H., G. de Boer, G. Feingold, J. Harrington, M. D. Shupe, and K. Sulia, 2012: Resilience of persistent Arctic mixed-phase clouds. Nat. Geosci., 5, 11–17, doi:10.1038/ngeo1332.
Neale, R. B., and Coauthors, 2010: Description of the NCAR Community Atmosphere Model (CAM5.0). NCAR Tech. Note NCAR/TN-486+STR, 268 pp. [Available online at www.cesm.ucar.edu/models/cesm1.1/cam/docs/description/cam5_desc.pdf.]
Oreopoulos, L., and Coauthors, 2012: The Continual Intercomparison of Radiation Codes: Results from Phase I. J. Geophys. Res., 117, D06118, doi:10.1029/2011JD016821.
Pavelsky, T. M., J. Boé, A. Hall, and E. J. Fetzer, 2011: Atmospheric inversion strength over polar oceans in winter regulated by sea ice. Climate Dyn., 36, 945–955, doi:10.1007/s00382-010-0756-8.
Persson, P. O. G., T. Uttal, J. Intrieri, W. Fairall, E. Andreas, and P. Guest, 1999: Observations of large thermal transitions during the Arctic night from a suite of sensors at SHEBA. Third Symp. on Integrated Observing Systems, Dallas, TX, Amer. Meteor. Soc., J5.3. [Available online at https://ams.confex.com/ams/99annual/abstracts/477.htm.]
Phillips, T. J., and Coauthors, 2004: Evaluating parameterizations in general circulation models—Climate simulation meets weather prediction. Bull. Amer. Meteor. Soc., 85, 1903–1915, doi:10.1175/BAMS-85-12-1903.
Pithan, F., B. Medeiros, and T. Mauritsen, 2014: Mixed-phase clouds cause climate model biases in Arctic wintertime temperature inversions. Climate Dyn., 43, 289–303, doi:10.1007/s00382-013-1964-9.
Prenni, A. J., and Coauthors, 2007: Can ice-nucleating aerosols affect Arctic seasonal climate? Bull. Amer. Meteor. Soc., 88, 541–550, doi:10.1175/BAMS-88-4-541.
Ramanathan, V., R. D. Cess, E. F. Harrison, P. Minnis, B. R. Barkstrom, E. Ahmad, and D. Hartmann, 1989: Cloud-radiative forcing and climate: Results from the Earth Radiation Budget Experiment. Science, 243, 57–63, doi:10.1126/science.243.4887.57.
Ricard, J. L., and J. F. Royer, 1993: A statistical cloud scheme for use in an AGCM. Ann. Geophys., 11, 1095–1115.
Sandu, I., A. Beljaars, P. Bechtold, T. Mauritsen, and G. Balsamo, 2013: Why is it so difficult to represent stably stratified conditions in numerical weather prediction (NWP) models? J. Adv. Model. Earth Syst., 5, 117–133, doi:10.1002/jame.20013.
Serreze, M. C., J. D. Kahl, and R. C. Schnell, 1992: Low-level temperature inversions of the Eurasian Arctic and comparisons with Soviet drifting station data. J. Climate, 5, 615–629, doi:10.1175/1520-0442(1992)005<0615:LLTIOT>2.0.CO;2.
Shupe, M. D., and J. M. Intrieri, 2004: Cloud radiative forcing of the Arctic surface: The influence of cloud properties, surface albedo, and solar zenith angle. J. Climate, 17, 616–628, doi:10.1175/1520-0442(2004)017<0616:CRFOTA>2.0.CO;2.
Shupe, M. D., V. P. Walden, E. Eloranta, T. Uttal, J. R. Campbell, S. M. Starkweather, and M. Shiobara, 2011: Clouds at Arctic atmospheric observatories. Part I: Occurrence and macrophysical properties. J. Appl. Meteor. Climatol., 50, 626–644, doi:10.1175/2010JAMC2467.1.
Song, S. W., and B. Mapes, 2012: Interpretations of systematic errors in the NCEP Climate Forecast System at lead times of 2, 4, 8, ..., 256 days. J. Adv. Model. Earth Syst., 4, M09002, doi:10.1029/2011MS000094.
Steeneveld, G. J., B. J. H. Van de Wiel, and A. A. M. Holtslag, 2006: Modelling the Arctic stable boundary layer and its coupling to the surface. Bound.-Layer Meteor., 118, 357–378, doi:10.1007/s10546-005-7771-z.
Stramler, K., A. D. Del Genio, and W. B. Rossow, 2011: Synoptically driven Arctic winter states. J. Climate, 24, 1747–1762, doi:10.1175/2010JCLI3817.1.
Svensson, G., and J. Karlsson, 2011: On the Arctic wintertime climate in global climate models. J. Climate, 24, 5757–5771, doi:10.1175/2011JCLI4012.1.
Taylor, K. E., R. J. Stouffer, and G. A. Meehl, 2012: An overview of CMIP5 and the experiment design. Bull. Amer. Meteor. Soc., 93, 485–498, doi:10.1175/BAMS-D-11-00094.1.
Tjernstrom, M., and R. G. Graversen, 2009: The vertical structure of the lower Arctic troposphere analysed from observations and the ERA-40 reanalysis. Quart. J. Roy. Meteor. Soc., 135, 431–443, doi:10.1002/qj.380.
Tjernstrom, M., J. Sedlar, and M. D. Shupe, 2008: How well do regional climate models reproduce radiation and clouds in the Arctic? An evaluation of ARCMIP simulations. J. Appl. Meteor. Climatol., 47, 2405–2422, doi:10.1175/2008JAMC1845.1.
Voldoire, A., and Coauthors, 2013: The CNRM-CM5.1 global climate model: Description and basic evaluation. Climate Dyn., 40, 2091–2121, doi:10.1007/s00382-011-1259-y.
Waliser, D. E., and Coauthors, 2012: The “Year” of tropical convection (May 2008–April 2010) climate variability and weather highlights. Bull. Amer. Meteor. Soc., 93, 1189–1218, doi:10.1175/2011BAMS3095.1.
Watanabe, M., and Coauthors, 2010: Improved climate simulation by MIROC5: Mean states, variability, and climate sensitivity. J. Climate, 23, 6312–6335, doi:10.1175/2010JCLI3679.1.
Westwater, E. R., Y. Han, M. D. Shupe, and S. Y. Matrosov, 2001: Analysis of integrated cloud liquid and precipitable water vapor retrievals from microwave radiometers during the Surface Heat Budget of the Arctic Ocean project. J. Geophys. Res., 106, 32 019–32 030, doi:10.1029/2000JD000055.
Williams, K. D., and Coauthors, 2013: The Transpose-AMIP II experiment and its application to the understanding of Southern Ocean cloud biases in climate models. J. Climate, 26, 3258–3274, doi:10.1175/JCLI-D-12-00429.1.
Xie, S., J. Boyle, S. A. Klein, X. Liu, and S. Ghan, 2008: Simulations of Arctic mixed-phase clouds in forecasts with CAM3 and AM2 for M-PACE. J. Geophys. Res., 113, D04211, doi:10.1029/2007JD009225.
Xie, S., and Coauthors, 2010: Clouds and more: ARM Climate Modeling Best Estimate data. Bull. Amer. Meteor. Soc., 91, 13–20, doi:10.1175/2009BAMS2891.1.
Xie, S., H.-Y. Ma, J. S. Boyle, S. A. Klein, and Y. Zhang, 2012: On the correspondence between short- and long-time-scale systematic errors in CAM4/CAM5 for the Year of Tropical Convection. J. Climate, 25, 7937–7955, doi:10.1175/JCLI-D-12-00134.1.
Xie, S., X. Liu, C. Zhao, and Y. Zhang, 2013: Sensitivity of CAM5-simulated Arctic clouds and radiation to ice nucleation parameterization. J. Climate, 26, 5981–5999, doi:10.1175/JCLI-D-12-00517.1.
Zhang, Y., D. J. Seidel, J.-C. Golaz, C. Deser, and R. A. Tomas, 2011: Climatological characteristics of Arctic and Antarctic surface-based inversions. J. Climate, 24, 5167–5186, doi:10.1175/2011JCLI4004.1.
