Some primary physical relationships related to the surface climate and atmospheric boundary layer were examined over East Antarctica and evaluated in the regional climate model HIRHAM for 2005–2008. For stable conditions, the observation-derived relationship between wind-scaled sensible heat flux and air-surface temperature difference distinctively differs between different surface flux parameterizations. Some of them decrease the heat transfer coefficient CH for strongly stable conditions, while others, such as the Louis scheme, do not. However, HIRHAM's application of the Louis parameterization produces small CH for strongly stable conditions similar to observations and other schemes, likely because a surface roughness much larger than observed is used and the bulk Richardson number differs. For Zhongshan, the observed radiation-cloud, temperature-cloud, and temperature-wind relationships are reproduced in the model, though quantitative differences are evident. An observed longwave warming effect of clouds is larger in the model, while the reduction of downwelling shortwave radiation by clouds is twice as large in the model. The model partially reproduces an observed weak wind regime associated with atmospheric decoupling, but fails to reproduce increasing temperatures with increasing winds. The quantitative differences in the radiation-cloud relationship suggest that errors in cloud characteristics produce a significant deficiency in downwelling net radiation for clear and cloudy conditions. This deficiency is the likely cause of HIRHAM's strong cold bias in the surface temperature and positive bias in near-surface stability. The sensible heat flux analyses and a sensitivity test suggest that errors in the sensible heat flux relationship are not the primary cause.
 The East Antarctic ice sheet is the largest on Earth, has an average elevation of about 2500 m above sea level (a.s.l.), and is subject to harsh environmental conditions. The atmosphere is characterized by two unique meteorological phenomena: quasi-permanent surface-based temperature inversions and persistent katabatic winds. The difficulty in collecting data in this remote area of extreme climate conditions is one factor limiting understanding of the structure and variability of the near-surface climate and atmospheric boundary layer (ABL) over Antarctica. The installation of a large number of Automatic Weather Stations (AWS) contributed to enhanced meteorological observations. Particularly important are those on the plateau at Dome-A [Xiao et al., 2008; Ma et al., 2010] and Dome-C [e.g., Argentini et al., 2005; Genthon et al., 2010] which contribute to the few station measurements over the continent. The success of the Antarctic Mesoscale Prediction System (AMPS) is associated with the assimilation of all the different observation types and optimized model physics [Bromwich et al., 2005]. Further advances in Antarctic modeling will be based on the analysis of physical processes in comparison with observations. Complementary to the global models, regional climate models (RCMs) have the advantage of a higher horizontal resolution, and can be used as testbeds for improved parameterizations of important subgrid-scale processes. Recent studies of pan-Antarctic RCMs [Bromwich et al., 2005; Reijmer et al., 2005; Valkonen et al., 2008; Gallée and Gorodetskaya, 2010; Tastula and Vihma, 2011] demonstrate the current model skills. However, they also indicate the sensitivity of the results to the model configuration, emphasizing that careful model evaluations are a prerequisite for understanding how to improve the models.
 The aim of the present paper is to identify model biases and error structures related to physical parameterizations of the regional climate model HIRHAM, which has often been applied to study polar climate (most recently, e.g., Dethloff et al.  and Rinke et al. 2012]) and which will be used in the future. Special emphasis is on the shortcomings caused by ABL parameterizations similar to those often used in other climate models. Due to the analysis technique described later, conclusions for further model improvement and development are not only relevant for HIRHAM but also for other models using the same parameterizations, and especially for the present state-of-the-art version of the global climate model ECHAM, since it uses the same parameterization package for the ABL.
 Earlier, van Lipzig et al.  evaluated the same RCM in Antarctica utilizing observations from the Swedish research station Svea located in Dronning Maud Land during a short summer period and reported on the model's representation of surface heat exchange processes. They found a slight overestimation of the near-surface vertical temperature gradients combined with a slight underestimation of turbulent heat and moisture transport near the surface. Another result was an overestimation of the modeled cloud cover.
 Unlike previous model assessment work, the evaluation is here done through the use of diagnostics, proposed by Persson et al.  and Chen et al. , that address a model's ability to represent the observed physical relationships and processes. This approach allows a direct process evaluation circumventing the problem of validating energy fluxes with potentially incorrect environmental conditions in the models, and has been applied for evaluating the ABL-cloud-surface system in Arctic RCMs [Tjernström et al., 2005; Wyser et al., 2008]. Focus in the current paper is on the evaluation of primary physical processes and relationships related to the surface climate and ABL over Antarctica in the HIRHAM model. The evaluation is conducted specifically over East Antarctica along a traverse from the coast (Zhongshan) to the plateau (Dome-A) for the time period 2005–2008.
 In section 2, the available surface observations at three stations along the traverse and the applied RCM HIRHAM are described. Section 3 presents the evaluation results. Section 4 summarizes the results and concludes with suggestions of needed model improvements.
2. Data and Simulations
 Data from 3 stations (Zhongshan, Eagle, Dome-A) along a traverse from the coast to the plateau in East Antarctica from 4 years (2005 to 2008) were analyzed. The locations of the 3 sites are shown in Figure 1.
 The Chinese Zhongshan station is located at 68.58°S, 77.95°E, at an elevation of 22 m. Its climate conditions are influenced by a complex interplay of katabatic wind, topographic structures, and oceanic influences. Standard synoptic measurements are made at this station and we analyzed the data of 6 hourly 2 m air temperature and 10 m wind speed, shortwave and longwave radiation, cloud fraction (manually observed), and surface pressure. The radiative components were measured by the pyranometer/pyrgeometer pairs on an unventilated Kipp & Zonen CNR1 net radiometer. The lack of ventilation was not considered a problem because of the dry and windy conditions at Zhongshan so that icing/frosting on the sensors is not occurring.
 Turbulent sensible heat and momentum fluxes were measured with a WindMaster Pro (Gill, UK) sonic anemometer located at 4 m height, providing measurements of the three wind components and virtual temperature at a frequency of 10 Hz (see details in Ma ). These data were collected on the ice sheet approximately 6 km southwest of Zhongshan during the period from 2 January to 18 February 2008. The turbulent fluxes were calculated from the 30 min perturbation data, and these 30 min fluxes were averaged to produce 6 hourly eddy covariance (EC) turbulent fluxes used in the subsequent analyses. When calculating the turbulent fluxes, the triple coordinate rotation method has been applied to eliminate the impacts of tilt of sensors or terrain on the observation. Also a strict data quality control has been conducted, such as deleting the abnormal values, removing the noise by standard deviation method, and testing the dimensionless variance similarity.
Ma  and Lin et al.  compared the EC turbulent fluxes with bulk turbulent fluxes calculated from AWS data by means of different surface layer parameterization schemes. The required surface temperature was estimated from longwave radiation measurement and extrapolated from 10 cm and 20 cm snow temperatures. The authors identified the Louis scheme [Louis, 1979; Louis et al., 1982] as best suited for simulating the turbulent fluxes under the predominantly weakly stable conditions. The comparison between the EC-observed friction velocity u* and the Louis-derived u* showed a good agreement (correlation coefficient of 0.98, mean bias of −0.019 m/s, and the differences are within ±0.1 m/s in 97% of all cases). The agreement in sensible heat flux (SHF) was also reasonable (correlation coefficient of 0.75, mean bias of −4 W/m2, and differences within ±5 W/m2 / ±10 W/m2 in 41% / 85% of all cases). (Mean values of the EC observations are: u* = 0.217 m/s, SHF = −8.7 W/m2).
 Because no direct turbulent flux measurements are available at Dome-A and Eagle, bulk fluxes were derived from the AWS data at these sites. The bulk EC flux comparisons at Zhongshan provide some confidence in applying these bulk fluxes. The probability density of the dry bulk Richardson number RiB, calculated from observations at Zhongshan and Dome-A (Figure 2), show that the summertime surface layer is only slightly more stable at Dome-A compared to Zhongshan, that large RiB (>0.5) rarely occur at either station, and that RiB is frequently close to zero at both sites. Unstable surface layers occur more frequently at Zhongshan, hinting at the larger impact of cold air advection. Because summertime stability conditions are similar, the bulk methods validated at Zhongshan, can be used also at the inland stations.
 By the joint efforts of Chinese and Australian weather agencies, AWS were installed at the two stations Eagle and Dome-A. Eagle is located 800 km inland at 76.42°S, 77.02°E, 2852 m a.s.l., and experiences weak katabatic wind because it is on the gentle slope in the interior. Dome-A is over 1200 km inland on top of the Antarctic Plateau at 80.37°S, 77.37°E, and 4093 m a.s.l. For a detailed description of the AWS, the reader is referred to Ma et al.  and Xiao et al. . Cold temperatures and solar radiation make air temperature AWS measurements challenging even using passive ventilated radiation shields, but bias correction formulae are not easily derived [Genthon et al., 2011]. Here, we analyze the measured temperature and wind speed at three heights above ground (1 m, 2 m, and 4 m), and the surface pressure. Neither the surface temperature nor the turbulent fluxes were measured. Small inconsistencies in the AWS observations of temperature (e.g., due to radiation impact or riming effects) did not allow us to determine the surface temperature by a logarithmic extrapolation which would be consistent with surface layer theory. For this reason, we decided to estimate the surface temperature by linear extrapolation (first based on the temperature observations in 1 m and 2 m, and then alternatively based on a least square fit of the observed 1 m, 2 m, and 4 m values).
 The surface turbulent sensible heat and momentum fluxes were calculated from 2 m temperature and 4 m wind using bulk formulation parameterizations. Three different surface layer parameterizations [Louis, 1979; Louis et al., 1982; Holtslag and de Bruin, 1988; Grachev et al., 2007] were applied to calculate the surface turbulent fluxes. The Louis method (in the following referred to as L79) is widely used in climate models, including the HIRHAM model, and has been tested in the analysis at Zhongshan (see above). Andreas  compared six schemes for the snow and ice surface turbulent fluxes and recommended the Holtslag and de Bruin (in the following referred to as HB88) stability functions because of their best skill for strongly stable stratification cases. Grachev (in the following referred to as G07) introduced a new parameterization for universal functions under stable conditions and showed their best fit for the Surface Heat Budget of the Arctic Ocean Project (SHEBA) data. Our approach of using a set of three different parameterizations accounts for the uncertainty in the derived turbulent fluxes, which is primarily due to the uncertainty in the heat transfer and drag coefficients. Figure 3 illustrates the different stability dependence of the heat transfer coefficient according to the L79, HB88, G07 parameterizations.
 In this paper, we use 6 hourly averaged data for 2005–2008. The sign convention for the sensible heat flux is defined to be positive upward and negative downward (directed toward the surface). Observed low wind speeds u (u < 0.3 m/s) may experience extreme biases and were therefore not analyzed. The 4 m model wind considered in some figures was calculated using the 10 m wind and Louis functions as used in the model; the 2 m model temperature is a direct diagnostic output.
 The atmospheric RCM employed in this study is the HIRHAM model, which was recently applied over Antarctica [Glushak, 2008; Xin et al., 2010; Dethloff et al., 2010]. It is a hydrostatic primitive equation model, and contains the physical parameterization package of ECHAM4 [Roeckner et al., 1996]. Although recently the next generation (ECHAM5-based) HIRHAM has been applied over Greenland [Lucas-Picher et al., 2012], its circumpolar version still needs further adjustments and optimizations before it can operationally be used over this domain. However, the ABL parameterization has not been changed in ECHAM5 [Roeckner et al., 2003]. Therefore, our findings are still relevant to understand the ABL error structures and their causes. An extensive description of HIRHAM has been given in the above mentioned papers and the references therein, and can also be found in Christensen et al. . Here, only the relevant features are briefly described.
 The surface turbulent fluxes are calculated from bulk transfer relationships following Monin-Obukhov similarity theory. In the applied L79 scheme the transfer coefficients for heat (CH) and momentum (CM) depend on the roughness length z0 and bulk Richardson number RiB, such that CH,M = CDNFH,M with neutral coefficient CDN = k2/[ln(z/z0)]2 and stability functions FH,M, empirically specified for the different conditions. For stable cases (Ri > 0) it is FH = 1/[1 + 3bRiBsqrt(1 + dRiB)] and FM = 1/[1 + 2bRiB/sqrt(1 + dRiB)], with b = 5, d = 5. The roughness length z0 over land is a function of subgrid-scale orography and vegetation [Claussen et al., 1994]. z0 has a value of 7.39 10−4 m at the 3 station grid points of Zhongshan, Eagle, Dome-A. The roughness length for momentum is taken equal to that of heat, but different stability functions for momentum and heat are applied. Above the surface layer a higher-order closure scheme is used to parameterize the turbulence, and the exchange coefficients are calculated as functions of turbulent kinetic energy [Brinkop and Roeckner, 1995].
 The subsurface temperature is calculated via the heat diffusion equation using 5 layers to a depth of 10 m. The equations are solved with these characteristics of snow: volumetric heat capacity of 0.7 106 J/m3/K, thermal conductivity of 0.22 J/s/m/K, which follows the suggestion by van Lipzig et al. . The temperature profile in the snow is initialized based on the climatological annual cycle of surface temperature. Over ice sheets and glaciers, snow processes are neglected in ECHAM4; i.e., snow at the surface is zero. However, a melting term is diagnosed to include the cooling due to snowmelt on the surface temperature, which itself is obtained from the surface energy balance equation without considering snowmelt. During the heat budget calculation over the ice sheet, the surface albedo a is parameterized as a linear function of the surface temperature Tsrfc according to: a = amax − (amax − amin)[(Tsrfc − T0)/(Tm − T0)] with Tm = 273.15 K (melting snow/ice), T0 = 263.15 K (cold surfaces), and amin = 0.6, amax = 0.8 are minimum and maximum values representing melting and dry ice/snow, respectively.
 The stratiform cloud scheme is according to Sundqvist et al.  and includes fractional cloud cover (b) in a prognostic scheme. The governing equations for water vapor and cloud water (qw) contain all transport as well as cloud-microphysical terms. Such that, e.g., ∂qw/∂t = R(qw) + bCc + (1 − b)C0 − bPc, where R(qw) represents the transport terms of qw; the subscripts c and 0 refer to the cloudy and cloud-free part of a model grid box. The cloud-microphysical terms are condensation of water vapor in the cloudy part (Cc > 0), evaporation of cloud water (Cc < 0), evaporation of cloud water transported into the cloud-free part (C0 < 0), and precipitation (Pc). The fractional cloud cover (b) is parameterized in terms of grid-mean relative humidity (r): b = 1 − sqrt(1 − b0), where b0 = (r − r0)/(1 − r0) and r0 is a condensation threshold specified as a function of height.
 The radiation code is adopted from the ECMWF model with a few modifications [see Roeckner et al., 1996]. Shortwave radiation is treated by the two-stream method of Fouquart and Bonnell , and longwave radiation by the method of Morcrette . In the longwave radiation scheme, the treatment of clouds follows the method of Washington and Williamson . The fluxes for clear and overcast skies are calculated separately and than combined proportional to cloud amount with some cloud overlap assumption. For contiguous clouds, maximum overlap is assumed and random overlap otherwise. In the shortwave radiation calculation, the upward and downward fluxes are obtained from the reflectances R and transmittances Tr of the layers, accounting also for the presence of clouds in the layers according to Rtop = bRc+(1−b)R0 and Trbottom = bTrc + (1 − b)Tr0 where Rtop is the reflectance at the top and Trbottom is the transmittance at the bottom of a layer. The cloud optical depth, single scattering albedo and asymmetry factor enter the calculation of Rc and Trc.
 The model was configured for a circum–Antarctic domain (Figure 1) at 50 km horizontal resolution. As proposed, e.g., by van Lipzig et al. , a relatively high vertical resolution (25 vertical levels with the lowest level at ca. 12 m and 10 levels in the lowest 1 km) close to the surface is used to describe the generally shallow ABL in the Antarctic. A time step of 120 s has been used, and a full radiation calculation is carried out every 2 h. Simulations were carried out for the 4 year period 2005–2008 using ECMWF operational analyses for the lateral and lower boundary forcing and for the (re)initialization. During the time the simulations were conducted, the ERA-Interim data were not available; therefore the operational analyses have been used to force the model integrations. The lateral boundaries are updated 6 hourly, and at the ocean lower boundary, the sea surface temperatures and sea ice fractions are updated daily. For sea ice, fractional coverage was considered, ice thickness was prescribed to 2 m, and the sea ice surface temperature is calculated prognostically via a heat balance equation. The model was run in the so-called forecast mode, in which it was reinitialized every 12 h from the ECMWF analyses, with the aim to force the models to stay close to the observed weather situations. The cloud properties (cloud fraction, cloud water) are set to zero on initialization and subsequently adopted from the preceding simulation, as also done for the snow/ice prognostic variables. Additional analysis of time series for cloud cover and turbulent fluxes (not shown) made sure that model results were not affected by spin-up effects and neither spurious gravity waves and associated vertical motions nor artificial turbulent fluxes were generated by the initialization procedure.
 For the comparison of the simulations with the station data, the results from the station-nearest model grid point were taken. The analysis was made separately for summer (November–February; NDJF) and winter (June–August; JJA).
3.1. Comparison of Surface Pressure Variability and of Key Surface Parameters
 Observed seasonal and annual surface pressure variations at the sites along the East Antarctic traverse were discussed by Ma et al. . It is very important that the model captures the observed pressure systems and their variability. Otherwise, the pressure systems will contribute to, for example, biased model winds and turbulent fluxes. For that purpose, the simulated variability of the surface pressure on the different time scales were analyzed and compared with observations. First, the 1 hourly data have been investigated. The detected variance peaks are associated with baroclinic pressure systems (ca. 3–6 day period) and transient systems (ca. 7–10 day period), which agree in simulation and observation (not shown). Figure 4 shows the results of a power spectral analysis of the observed and simulated surface pressure at the 3 stations, based on daily average data from 2005 to 2008. The dominant feature is the large variance at intraseasonal (30–60 days) time scale. Such a predominant peak in the intraseasonal time scale was earlier reported for East Antarctica for wind speed and temperature [Yasunari and Kodama, 1993], and surface stratification and 500 hPa geopotential [Zhou et al., 2009]. The consistency of the spectral peaks in the observation and model suggests that the variations of pressure conditions are well reproduced by the simulations.
 Before we present the analysis of different physical relationships in the next section, the corresponding meteorological variables are evaluated by comparing observed and simulated time series in the traditional approach. Table 1 presents the calculated error statistics (correlation coefficient: r, root mean square error: rmse, mean bias: bias) for 6 parameters (near-surface wind and temperature, surface temperature and pressure, surface sensible heat flux SHF and friction velocity u*). The error statistics for SHF and u* have been calculated for all three different observation-derived estimates (based on L79, HB88, G07 parameterizations), but it is not sensitive to those (not shown).
Table 1. Error Statistics of the Differences Between Observations and HIRHAM Simulation at Dome-A, Eagle, and Zhongshan for Near-Surface Wind (u (m/s)) and Temperature (Tair (K)), Surface Temperature (Tsrfc (K)) and Pressure (psrfc (hPa)), Surface Sensible Heat Flux (SHF (W/m2)), and Friction Velocity (u* (m/s)), Calculated From 6 Hourly Data 2005-2008a
Measurements of SHF and u* at Zhongshan are only available for January and February 2008. Observation-derived SHF and u* at Dome-A and Eagle are based on the G07 parameterization. For Zhongshan, the numbers are shown for the two nearest model grid points: (102,66) and (103,66). The surface height values of the used station-nearest model grid points are given. Given are the linear correlation coefficient (r), the root mean square error (rmse), and the mean bias (bias; “model minus observation”). The total number of used samples is indicated by n. The statistics for summer (NDJF) only are given separately in parentheses.
Dome-A (80,60); 4029 m
Eagle (88,62); 2802 m
Zhongshan (102,66); 34 m
Zhongshan (103,66); 0 m
 At the inland stations (Eagle, Dome-A), the simulated time series of surface pressure and temperatures are quite consistent with the observations, which is expressed in high correlation coefficients (r > 0.8). The model skill for surface pressure is good (bias and rmse of 4–5 hPa), but the temperatures are too cold in the model. The Tair bias is approximately −4 K at both stations, while the Tsrfc bias is up to −7 K. Considering all terms of the surface energy budget, the cold bias suggests that in the model (1) the turbulent heat flux produces too much cooling or insufficient warming of the surface, (2) the radiative cooling is too large, and/or (3) the ground heat flux is too small. As concerns the first point, the cold bias could be related to an inappropriate stability function for stable ABL conditions. Arctic observations [e.g., Grachev et al., 2005] show that when the ABL becomes strongly stable, turbulence is strongly reduced and cannot maintain the downward (warming) heat flux any longer. Tsrfc is then determined by a balance between radiative cooling and conductive heat flux in the ground. The used surface layer parameterization could explain the cold bias if for a given stratification it would produce less mixing than in nature. The performance of the model heat flux parameterization is evaluated in section 126.96.36.199. Excessive radiative cooling could occur if either the net longwave or shortwave fluxes were too small. The former can be too small if either the longwave downwelling (LWD) is too small or the outgoing long wave is too large. The LWD can be too small if the cloud cover is too small, the optical depth (emissivity) of the clouds is too small, or if the cloud temperature is too low. Also, a too small emissivity of the clear sky would result in a too low LWD at the surface. Van de Berg et al.  found an underestimation of LWD in the model by comparing it with observations in Dronning Maud Land and related it to the cloud parameterization and longwave radiation scheme. The upwelling long wave can be too large if the surface temperature is too large; however, Table 1 shows that this is not the case for HIRHAM. The net shortwave radiation can be too small if the cloud cover is too large, if the optical depth or the albedo of the clouds are too large, or if the surface albedo is too large. Van As and van Den Broeke  tested the Antarctic ABL sensitivity to various parameters, and concluded that the surface albedo is most important for ABL temperature. However, the HIRHAM albedo of 0.6–0.8 is unlikely to be too large, as albedo measurements of snow in polar environments has a range of 0.65–0.90 [Perovich et al., 2002; Persson et al., 2002; Pirazzini, 2004], with values below 0.75 only occurring for melting snow. Therefore, evaluations of the radiation and clouds are important to help understand the model's cold bias and are performed in section 3.2.2. An evaluation of the ground heat flux is unfortunately not possible because necessary snow and ice measurements are not available.
 The model simulates higher wind speeds at both Dome-A and Eagle compared to the observations. However, the wind simulation is much better at Dome-A (bias = 1 m/s, rmse = 2 m/s) than at Eagle (bias = 4 m/s, rmse = 5 m/s). One possible reason is that Eagle is characterized by a more complex topography, but another and probably more likely reason is that the simulated downslope flow and cold air advection are overestimated because the modeled surface temperatures are too cold over the plateau (see discussion above and in section 3.2.1). Note that the wind speed performance at both stations is better in summer.
 The model evaluation at Zhongshan is hindered by its vicinity to the coast, so that the skill depends on the selection of the model grid point (see Table 1, and also section 3.2.2). To illustrate this impact on the comparison, the results of the two station-nearest model grid points are shown. The table shows a smaller bias and rmse for wind and pressure for the station-nearest land model grid point (102,66), compared to the ocean grid point (103,66). But the near-surface temperature is better represented by the latter. For all four station-surrounding grid points, the observed and simulated temperature time series are highly correlated (r = 0.9), but the bias ranges from −1 K to −13 K.
 Generally, the modeled u* agrees better with observations than SHF, which was also found in an evaluation during SHEBA over the Arctic [Tjernström et al., 2005]. The u* error statistics (biases are in the order of 50–100%) and the SHF rmse (12–40 W/m2) presented here are similar to those calculated for the Arctic, while the SHF bias is larger and correlation is smaller here. The results in Table 1 indicate very large biases in the modeled SHF relative to the small absolute values (SHF biases are in the order of up to 200%) occurring in the mostly stable ABL over Antarctica. This result differs from a study of van Lipzig et al.  who evaluated the same model over another location (Dronning Maud Land) and found for summertime conditions that the surface fluxes are well reproduced but that the fluxes at higher levels above 200 m were underestimated by the model.
3.2. Covariability of Variables
 This evaluation aims at investigating functional behaviors between different atmospheric variables and comparing those from observations with the model.
3.2.1. Turbulent Flux Relationships
 The focus is here on process relationships which have been discussed to be useful for evaluating the ABL surface system [Persson et al., 1999; Chen et al., 2003; Tjernström et al., 2005]. The investigation is limited to surface turbulent heat and momentum fluxes, and does not include humidity fluxes. The near-surface wind is an important variable affecting the ABL due to its influence on the turbulent momentum and heat transfer near the surface and the dissipation of turbulent kinetic energy. The 4 m wind, 2 m temperature, and the calculated bulk fluxes using these data are used in the following sections and figures, from both the AWS-derived data and the model.
188.8.131.52. Friction Velocity (u*) and Wind Speed (u)
 Both observations and model indicate a strong correlation between u* and u (r > 0.95) at the 3 stations. According to the bulk flux framework, τ/ρ = u*2 = CD u2 where τ is the stress and ρ is air density. The drag coefficient CD is calculated in the model as a function of surface roughness and stability using the momentum stability function as given in section 2.2. A larger roughness produces a larger CD and hence a larger momentum transport; a greater magnitude of stability function also produces a larger CD and hence a larger transport. From the u*-u relationship plotted in Figure 5, two results can be derived. The first is that on average momentum transports (u*) are very similar in observations and model provided that the model predicts the same wind speed as observed. We will see below that this is, however, not the case so that momentum transports will differ in general. The second result is related to the drag coefficients. The square root of the CD can be obtained from the u*/u ratio or from the slope of the line passing through that point and through (0,0).
 The observation-derived data sets and the EC data show linear relationships with a similar slope of 0.04 m/s/(m/s) (and hence similar CD). While the linear slopes through the model data points are similar (0.05 m/s/(m/s)), the u*/u ratios (and therefore CD) are a bit lower for small u (ca. u < 3 m/s) (because the line these points represent does not pass through the origin at (0,0)) compared to L79. These lower CD suggest that the model produces, under weak wind conditions, too small momentum transport to the surface for a given wind speed, thereby leading to stronger winds. The modeled CD values are getting closer to the observed ones as u increases. Under strong wind conditions (ca. u > 5 m/s), the model arrives at slightly higher CD values than observed. It is interesting to note that, while the model CD values are too low for weak wind conditions, the surface roughness length z0 appears to be several times larger than observed (Figure 6), which should produce a larger CD. This suggests that the stability correction is too small for weak wind cases. Table 2 compares the averaged observed and modeled drag coefficients, so that the stability effect becomes clearer. The model systematically overestimates CD at all sites, which is in accordance with what is expected from the higher z0.
Table 2. Comparison of the Observed and Modeled Drag Coefficient CD = (u*/u)2 and the Neutral Drag Coefficient CDN = [0.4/ln(z/z0)]2 (z = 4 m) for the Different Sites, Based on NDJF 2005–2008 Dataa
For LGB69, the numbers are based only on NDJF2007 and JF2008 due to serious missing values.
 Observations show that mean near-surface wind speeds increase from the coast to the escarpment region just inland, with strongest winds at katabatic sites, and then decrease rapidly toward the interior plateau (Figure 6), as also discussed by Yang et al.  and Ma et al. . The figure shows that the model reproduces this structure in general terms. At a site near the escarpment, LGB69 (70.84°S, 77.08°E, 1850 m a.s.l.), easterlies prevail year-round as a result of typical strong katabatic winds, and the observed annual mean wind speed is 8.5 m/s [Chen et al., 2007] which agrees with the simulation. Thus, the model does show a maximum wind just inland from the coast. The model does not accurately represent the rapid decrease of winds at Eagle just inland from the escarpment. However, the simulated higher wind speed at Eagle is in line with the AMPS-simulated annual wind speeds of 8–12 m/s [Parish and Bromwich, 2007]. The physical basis for these stronger winds is that the site is located on the slope near the edge of the ice sheet, though it is not on the steep portion of the slope. The model's overestimated annual winds at Eagle are primarily due to stronger simulated winds around winter, where only few (and uncertain) wind measurements are available. The model reproduces the weaker winds at the coast (Zhongshan) and on the flat area on the top of the ice sheet (Dome-A).
184.108.40.206. Wind-Scaled Sensible Heat Flux (SHF/u) and Air-Surface Temperature Difference
 The sensible heat flux (SHF) is an important component of the surface energy balance in Antarctica. It is frequently, especially in winter, directed downward (negative sign in our convention), thus transporting heat toward the surface and compensating for the radiative cooling. Turbulence and hence SHF in the stably stratified wintertime Antarctic atmosphere is mainly induced by the vertical wind shear. According to the bulk flux framework, SHF is proportional to the product of wind speed (u) and the air-surface temperature difference (Tair–Tsrfc; ΔT) such that SHF = −ρ cp u CH (Tair–Tsrfc) where ρ is air density, cp is the heat capacity at constant pressure, and CH is the heat transfer coefficient. Accordingly, slopes between the origin and points in a plot of SHF/u as a function of ΔT (T2m–Tsrfc, Figure 7) are proportional to CH and allow a direct check of the properly scaled flux response to the given near-surface temperature gradients.
 First, Figure 7 shows that the observation-derived relationships using different ABL parameterizations (L79, HB88, G07) are quite similar for unstable and near-neutral conditions, while they differ for strongly stable conditions. The figure indicates that the observation L79–derived scaled fluxes show a nearly linear dependence on the temperature gradient for all 3 stations for all conditions, with only slightly smaller slopes (in magnitude) for stable conditions than unstable ones. The stable condition L79 slopes range from −2.4 to −2.8 W/m2/m/s/°C at the three sites, corresponding to average CH values of ∼2.0–2.2 × 10−3. A less steep slope is found for stable conditions for the HB88- and G07-derived data (−1.5 to −1.9 W/m2/m/s/°C), indicating an average CH value of ∼1.2–1.5 × 10−3. The EC measurements at Zhongshan are in reasonable agreement with the HB88- and G07-derived data (slope of −1.4 W/m2/m/s/°C), indicating an average CH value of ∼1.1 × 10−3), and Arctic EC observations [e.g., Tjernström et al., 2005] also show the smaller CH for large ΔT. Hence, the smaller CH estimates from the HB88- and G07-derived data are consistent with those expected for stable conditions (also see Figure 3 and Grachev et al. [2005, 2007]) and are considered more realistic than those from the L79-derived data.
 Because the Zhongshan ABL is mostly characterized by weakly stable or unstable conditions (|ΔT| < 2 K), an ABL regime with (nearly) constant CH, i.e., with a linear dependency of the scaled SHF flux on ΔT, dominates. Generally, the CH, derived from the different parameterizations, match for |ΔT| smaller than 1 K or 2 K (depending on the station). For larger ΔT (stable conditions; ΔT > 2 K), the CH differ and the HB88- and G07-derived data show a decrease of CH. This indicates that, for those cases, the reduction in turbulence (decrease in CH) dominates over the increase in the temperature gradient, thereby reducing the surface warming due to the SHF. Figure 7 shows that this is also true for conditions of ΔT > 2 K at Zhongshan, based on direct EC measurements. This behavior has also been reported for Arctic observations [Tjernström et al., 2005].
 The HIRHAM modeled slopes for stable conditions range between −0.6 W/m2/m/s/°C and −1.0 W/m2/m/s/°C at the three sites, and are in much better agreement with the observation HB88– and G07-derived data than with the observation L79–derived data. The reason for the differences between the modeled and observation L79–derived slopes is in the calculation of CH which depends not only on ΔT but also on the roughness length z0 and bulk Richardson number RiB (see section 2.2). Both the model and the L79 observation–derived method use the same stability functions, but the model z0 is 4–6 times as large as the observed one (see Figure 6), and different RiB occurs because it depends on the near-surface wind and air temperature. Both have biases in the model (Table 1) which influences CH, and hence the modeled slope.
 A striking difference in the frequency of occurrence of different stability conditions is found between the observation-derived and HIRHAM model data. The temperature observations clearly show a significant number of unstable atmospheric events (ΔT < 0 K) in summer at both Dome-A and Eagle (Figure 7). The positive (upward) scaled heat flux indicates surface cooling associated with periods of unstable stratification. In contrast, Zhou et al.  reported that summertime temperature gradients are nearly always positive at Dome-A, indicating the existence of a persistent surface inversion layer. They argue that insufficient solar insolation and only small SHF make it hard to break up the inversion, although a shallow unstable layer may develop for short periods in summer. Their results are based on the temperature gradient T4m–T2m, also presented in Figure 8. However, analysis of temperature measurements below this layer shows that the temperature gradient decreases to the snow/ice surface significantly, with a highly nonlinear temperature profile in the near-surface layer below 4 m (Figure 8). Therefore, mainly shallow unstable cases are found in the lowest layer (below 2 m) during summer. Unstable cases, i.e., negative temperature differences of T4m–Tsrfc and T2m–Tsrfc, account for 56% and 64% of the total (1 hourly) samples during summer (NDJF, 2005–2008) at Dome-A. Thus, unstable (and near-neutral) boundary layers are prevalent and regularly appear over Antarctica during summer, as also reported by others [e.g., Argentini et al., 2005; Neff et al., 2008; Town and Walden, 2009; Vihma et al., 2009]. However, as discussed in section 2.1, the estimate of the observation-derived Tsrfc is uncertain. Figure 8 indicates that the least square fit method strongly reduces the unstable stratification compared to the simpler linear extrapolation.
 In contrast to the observations, the model produces a very persistent stable stratification in summer, based on the measure of T2m–Tsrfc (Figures 7 and 9), likely due to the bias for Tsrfc being colder than that for Tair (Table 1 and section 3.1). Hence, whatever is producing the cold bias at the surface is producing a bias in surface layer stability as well. This high-stability bias can be seen at both Dome-A and Eagle in winter and summer (Figure 9). Although the least squares estimate of observation-derived Tsrfc leads to fewer observed unstable cases, the observations still suggest a model deficiency, since the model results do not show any unstable cases (Figure 10a).
 The strongly varying temperature gradient in the near-surface layer suggests that having the lowest model level at 12 m (see section 2.2) may be inappropriate to represent such extreme conditions. Therefore, a sensitivity study for summer 2008 has been conducted in which additional layers near the surface have been included so that the 3 lowest atmospheric layers become located at heights of 2.5 m, 9 m, and 17.5 m. However, the results indicate no improvement in the simulation of near-surface stratification. The calculated changes in the simulated probability distribution of ΔT show only minor changes in the order of 0–0.05, compared to the control run (not shown).
 As already discussed in section 3.1, another reason for too cold simulated Tsrfc could be the use of inappropriate stability functions. Viterbo et al.  suggested different stability functions above the surface layer for stable conditions: FH = 1/[1 + 2bRiBsqrt(1 + dRiB)] and FM = 1/[1 + 2bRiB/sqrt(1 + dRiB)], with b = 5, d = 1, which produces CH values between those for HB88 and L79 and similar to those from G07 (see Figure 3). Although the idea of Viterbo et al.  was to enhance mixing above the surface layer by these functions we tested the sensitivity of the model results on these functions when they are applied to the surface layer since the used functions in ECHAM above the surface layer allow already efficient mixing. This modified formulation was tested in a sensitivity study again for summer 2008. But, similarly to Glushak  we found that the sensitivity was only small and did not improve the simulations (Figure 10b). A probable reason is that RiB was mostly smaller than 1 (as related to 4 m height) most of the time in our simulation (see also Figure 2), so changes in FH and associated surface heat fluxes are very small. If RiB were larger, the model might be more sensitive to changes produced by the Viterbo et al.  formulation.
 Hence, the above analysis and sensitivity test suggest that the turbulent heat flux parameterization is not likely the reason for the cold bias in Tsrfc (and for the high-stability bias). Analyses in section 3.3.2 suggest that the bias may be due to poor representation of cloud properties, especially in summer. It is also possible that these biases are related to a poor representation of the snow-atmosphere interface. In this regard, more advanced RCMs apply multilayer surface snow models or implement improved surface albedo schemes [e.g., Kuipers Munneke et al., 2011]. This is an aim of our future model development.
3.2.2. Physical Relationships With Radiation, Clouds, Temperature, and Wind
 This section compares the covariability of cloud-radiation-temperature-related variables against the covariability of the corresponding variables from direct observations to investigate how well the model captures these basic physical relationships. Shortwave and longwave radiation and cloud fraction data are only available at Zhongshan, and the former only for 2008, therefore the analysis of those is limited.
220.127.116.11. Downward Radiation and Cloud Fraction
 The ABL and cloud-radiation processes are related, and the latter play a key role in regulating the surface energy budget. The influence of clouds on the Antarctic surface energy balance in summer was examined, e.g., for Dronning Maud Land by Van Den Broeke et al. . Clear relationships between cloud fraction and surface shortwave and longwave downward radiation (SWD, LWD) are expected; i.e., as cloud fraction decreases, SWD should increase and LWD should decrease (e.g., as shown for the Adelie coast by Wendler et al. ). While these general relationships are confirmed here for both the observations and the model, the relationships have quantitative differences that are significant.
 These relationships are confirmed for Zhongshan, where all the relationships show a distinct nonlinearity (Figure 11). The analysis of SWD is only meaningful for polar day and when the sun rises sufficiently high above the horizon (because cloud sides or even the base become illuminated when the sun is very low). Therefore, the analysis has been limited to those cases in summer (NDJF) when the solar zenith angle is smaller than 65°. We also consider the SWD normalized by the cosine of the solar zenith angle to account for the geometric increase in optical thickness with higher solar zenith angles. This also accounts for the varying zenith angle with time over the months of the data.
 The observations show a decrease in SWD with increasing cloud fraction (slope of −3.7 W/m2/%), although the scatter is large (Figure 11a). The model qualitatively reproduces this observed relationship, but shows a stronger decrease of SWD with increasing cloud fraction (slope of −7.7 W/m2/%). The differences between the observed and model relationships are clearer when bin averages are used (bold dots). Both relationships are well correlated (r = −0.6 in observation, r = −0.8 in model). The too high sensitivity of the modeled SWD to an increase in cloudiness has been also identified in Arctic simulations by Wyser et al. , and can produce errors in SWD of a few hundred W/m2 for cloudy conditions, leading to deficiencies of many tens of W/m2 in net solar radiation. This error may be produced by an excessive modeled cloud albedo, as suggested by Wyser et al. , or by a too large optical depth produced by excessive cloud liquid water. Either mechanism would reduce the transmitted shortwave radiation. A too high cloud albedo could be produced by a biased separation of the cloud water into liquid and ice fractions (overestimation of liquid part) and/or an underestimation of the median ice effective radius. The former is supported by the finding of underestimated ice and overestimated liquid water content in ECHAM4 cloud scheme [Lohmann et al., 2007]. In addition to these liquid water path causes, an overestimation in the total absorption of shortwave radiation in the cloud layer, most likely related to the cloud optical properties, has to be considered and was discussed for the ECHAM4 radiation scheme by Bretherton et al. .
 A well-defined relationship between LWD and cloud fraction is expected, which is confirmed by the gradual increase in LWD as cloud fraction increases (Figures 11b and 11c). The model captures the observed strong correlation between both variables (r = 0.8 for both observation and simulation in summer and winter). The model also reproduces the observed relationship: The slopes for the observation/model relations are both 0.8 W/m2/% in summer and 0.8/0.9 W/m2/% in winter for all data. The respective slopes of the curves using cloud fraction bin averaging are 0.6/0.8 W/m2/% in summer and 0.7/0.8 W/m2/% in winter. Based on a 10 year data set (1994–2003) of monthly means at the South Pole, Town et al.  calculated a similar linear relationship between the LWD cloud radiative forcing and cloud fraction, with slopes of 0.7 W/m2/% in summer and 0.6 W/m2/% in winter. The results further show that the HIRHAM model produces too little LWD radiation for partly cloudy and clear skies, especially for winter, and approximately the right amount for overcast skies, especially in summer (Figures 11b and 11c). The excess cloud liquid water suggested to exist for large cloud fraction by the SWD analysis may compensate, through larger longwave emissivity, for other cloud/longwave radiative errors evident for small cloud fraction and winter conditions. Thereby, the LWD is approximately correct for cloudy/summer conditions and is underestimated in the model for clear/winter conditions. The lack of LWD in the model in winter for partly cloudy and clear conditions suggests a too small emissivity possibly as a result of too little liquid water and/or a low bias to the temperatures of the clouds and atmosphere, or due to underestimated humidity. Differences in (subcloud) humidity between observation and model might affect the displayed LWD-cloud relationship. Hence, the combination of errors affecting the SWD and LWD leads to a significant deficiency in downwelling radiation for both clear and cloudy conditions and in both summer and winter.
 An issue in Figure 11c (winter) is the large number of high-LWD values for low (zero) cloud fraction in the observations. It seems to be clear that these represent riming on the radiometer domes and the radiation data isn't useful. However, this could not be conclusively verified and therefore the data are kept in the figure but not further discussed.
Figures 11a–11c show the biases in the modeled response of radiative fluxes to cloud fraction, while the LWD and SWD biases may be also due to model errors in cloud fraction as well. Therefore, Figure 11d presents a comparison of the visually observed and modeled cloud fraction (cc) by means of a histogram, using the data of the whole year. The figure illustrates that the observations are much more bimodal (cc = 0 or cc = 1) than the model. Furthermore, the skies are less frequently overcast (cc = 1) in the model than in the observations, though they are also less frequently clear. For the summer, the underestimation of cloudy skies results in increased SWD but decreased LWD, which end up approximately balancing each other with a surface albedo of 80%. The underestimation of clear skies in the summer likewise also leads to compensating effects, especially since both the SWD and LWD relationships to cc is fairly insensitive to cc changes for low cc. However, in winter when SWD is insignificant, an underestimation of cc = 1 (cloudy conditions) will produce an erroneous cooling of the surface that is not compensated by increased SWD.
 Similar to our results, Hines et al.  reported on optically too thick clouds in Antarctica GCM simulations. Fogt and Bromwich  recently discussed that a smaller cloud emissivity (smaller absorption coefficient) greatly improved their modeled cloud fraction. By giving more than twice as much weight to the cloud ice water path (relative to the cloud liquid water path) the negative cloud bias (underpredicted cloudiness) over McMurdo and South Pole could be removed in their AMPS model.
 We cannot evaluate here the simulated relationships between radiative fluxes and cloud water path (CWP) because observations of CWP are not available at Zhongshan. However, the simulated LWD-CWP relationship (not shown) is in agreement with that derived from Arctic observations [Inoue et al., 2006; Shupe and Intrieri, 2004]: strong increase of LWD with increasing CWP until CWP is ca. 0.1 kg/m2; above that threshold, LWD is insensitive to CWP. The simulated SWD-CWP relationship shows a decrease of SWD with increasing CWP (not shown) and is in accordance with the Arctic results of Wyser et al. , who reported that HIRHAM simulates a steeper decrease than that seen in the observations, which gives the indication that cloud transmissivity decreases too rapidly with increasing CWP in the model.
18.104.22.168. Near-Surface Air Temperature (Tair) and Atmospheric Conditions (Clouds, Wind)
 In polar regions, the near-surface air temperature is sensitive to changes in atmospheric conditions (e.g., cloud fraction and near-surface wind speed) due to increased synoptic-scale activity and presence of a surface-based temperature inversion, especially in winter. The lowest temperatures occur on calm, clear days with a strong temperature inversion. In cases of inversion break (e.g., due to cloud forming) or strong wind (mixing the warmer air down to the surface), the temperature can rise. Thus generally, overcast conditions are associated with depression activity, and cause an increase in temperature, specific humidity and wind speed. These relationships can be confirmed for Zhongshan.
22.214.171.124. Tair and Cloud Fraction
 Generally it is expected that a low-level cloud fraction is associated with warming in winter and cooling in summer. For high-surface albedo surfaces like Antarctica, however, it has been shown that clouds warm the surface in all seasons [e.g., Pavolonis and Key, 2003; Town et al., 2007]. Recently, Town et al.  quantified for South Pole that the near-surface temperature rises by 0.5–1 K in summer and 3–4 K in winter under cloudy skies. For Dronning Maud Land stations, Van Den Broeke et al.  discussed that clouds are associated with higher surface temperatures in summer. Also, clouds often occur simultaneously with warm air advection and stronger winds, both of which also contribute to the near-surface warming [e.g., Vihma and Pirazzini, 2005].
 The observational analysis for Zhongshan shows no obvious cloudiness-temperature relationship in summer, but confirms a positive correlation and the relationship of increasing near-surface temperature with increasing cloud fraction in winter (Figure 12). The strength is of similar or larger magnitude in the model compared to the observations, depending on the model grid point compared. The slopes (for the curves using cloud fraction bin averaging) are 0.06 °C/% for the observation, and 0.07 °C/% for the (102,66) model grid point and 0.14 °C/% for the (103,66) grid point, respectively. Also, Tair and cloudiness are similarly correlated in the observations (r = 0.4) and simulation (r = 0.6). The simulated rate of surface warming due to clouds is higher for overcast conditions (cloud fraction > 70%) (0.25 °C/% for the (102,66) grid point, 0.34 °C/% for the (103,66) grid point), which is larger than observed (0.16 °C/%). Hence, the modeled Tair has a slightly higher sensitivity to cloud fraction than in the observations.
126.96.36.199. Tair and Wind Speed (u)
 Turbulent fluxes have a strong impact on the surface energy budget especially during winter, when shortwave radiation is missing. This in turn influences the near-surface conditions of wind and temperature. Thus a close-to-reality representation of the heat and momentum fluxes requires also that the wind and temperature regimes are well modeled, e.g., in the sense that the functional dependency of modeled wind and temperature agrees with the observed relationship. In general, temperature depends on many variables, but it has been shown by Lüpkes et al.  that over Arctic sea ice–covered regions during clear-sky conditions the lowest (quasi-equilibrium) temperatures depend on the wind speed and that two different regimes are separated by a critical wind speed. While the mechanism producing this dual-regime environment isn't totally clear, Lüpkes et al.  noted this dual-regime characteristic of Arctic environments. We hypothesize that it may be explained by the combined effect of wind-dependent downward sensible heat fluxes and longwave radiative cooling/heating of the surface and near-surface air. Below a certain threshold of wind, air temperature decreases with increasing wind because of increasing downward heat flux to the surface which is not large enough to compensate its radiative cooling under clear-sky conditions. In the strong wind regime, mixing is strong enough to generate a wind speed-dependent increase of the ABL mixed layer height h. Lüpkes et al.  show that for larger h the cooling effect on the ABL by downward heat fluxes is smaller, which explains the increase of temperature with increasing wind in the strong wind regime. Steinhoff et al.  reported the existence of two such wind regimes over the Ross ice shelf.
 A relationship of increasing minimum temperature with increasing wind can be expected especially during winter, when enhanced wind speeds increase the entrainment heat fluxes from the near-surface capping inversions, which are especially pronounced during cold temperatures. The observations at Zhongshan clearly indicate this functional dependency between the low temperatures and wind speed (Figure 13). The observation-derived slope of the temperature-wind curve is 0.8°C/m/s. The modeled slope (1.1°C/m/s for the (102,66) grid point and 1.4°C/m/s for the (103,66) grid point) is consistent with the observations. The figure further indicates that the model reproduces also the variability of temperature for a given wind speed at Zhongshan. This suggests that the katabatic and synoptic-scale forcings are well represented there. As might be expected due to probably weaker near-surface capping inversions, there is no clear relationship between near-surface wind and temperature in summer at Zhongshan (not shown).
 Another interesting result from Figure 13 (Zhongshan) and Figure 14 (Dome-A and Eagle) is the existence of two wind regimes in the observations at all three sites, which is similar to the above mentioned findings of Lüpkes et al.  and Steinhoff et al. . To account for the lowest temperatures (which represent clear skies), the fifth percentiles of Tair have been calculated for each wind speed class (using u steps of 0.5 m/s). Based on these curves it becomes obvious that the observed value ucrit is 2–4 m/s depending on the location and season (ucrit = 2 m/s at Zhongshan in winter, ucrit = 3.5 m/s at Dome-A in summer, ucrit = 4 m/s at Eagle in summer). Similar magnitudes have been reported for other regions (ucrit = 6 m/s over Ross ice shelf [Steinhoff et al., 2009] and ucrit = 4 m/s over Arctic sea ice [Lüpkes et al., 2008]). In the strong wind regime (u > ucrit) Tair increases with increasing u. Turbulence increases with increasing u, transporting warm air toward the surface and heating it and the near-surface air. In the weak wind regime (u < ucrit) the Tair-u relationship is different, since Tair decreases or is constant with increasing u. During weak winds, turbulence is damped which contributes to a decoupling of the atmosphere from the underlying surface at stable stratification. This phenomenon is often observed in polar regions [e.g., Grachev et al., 2005]. During these weak wind conditions the turbulence is absent so the surface temperature is governed by the net radiation and any conductive flux from the snow/ice. The figures show that the model has some difficulties in representing the complex wind regime behavior, particularly for Dome-A and Eagle. At Zhongshan, the model fairly well represents the typical winter conditions. At the two inland stations, the model also reasonably represents the typical weak wind regime (simulated ucrit = 2 m/s, Figure 14). However, the model fails to simulate the strong wind regime (increase of Tair with increasing u), but instead produces a constant Tair independent of u. The reasons remain unclear, though one may speculate that it is related to the model's bias toward high-stability conditions and cold Tsrfc. Further, Lüpkes et al.  obtained their results with a 1D model whose vertical resolution was much higher than the present version of HIRHAM which also might explain its inability to reproduce the correct behavior.
4. Summary and Conclusions
 This paper evaluated some primary physical relationships related to the surface climate and ABL over East Antarctica in observations and the HIRHAM model. The applied model evaluation approach allowed a direct process evaluation despite some detected incorrect environmental conditions in the simulation (like too cold temperatures and too strong winds over the plateau). The analysis showed that the model is able to represent most observed physical relationships, with the possible exception of some involving the sensible heat flux (SHF). However, significant quantitative aspects of the relationships were frequently in error. One of the most striking differences between model results and observations is the underestimation of the surface and near-surface air temperatures with the former being even more underestimated so that the temperature difference between air and surface is larger than the observed one indicating a stronger stability. This might explain the bias in the fluxes of sensible heat, which are smaller than in nature. This hints to a decoupling between the modeled ABL and the surface due to a large near-surface stability suppressing turbulence. This would explain also the positive bias in wind speed, since weaker turbulence would result in decreasing momentum loss to the surface.
 For the coastal station Zhongshan, which is particularly difficult to evaluate even in a 50 km resolution model with traditional techniques due to its vicinity to the coastline, it has been shown that the applied evaluation approach is helpful. Here, despite possible temperature biases (depending on which model grid point is selected to be evaluated), the observed cloud-radiation and temperature-atmospheric condition (cloud, wind) relationships are clearly reproduced in the model, indicating that these basic physical mechanisms are well described in the model.
 The evaluation of the turbulent flux relationships at the three stations showed that the observed link between the near-surface wind (u) and atmospheric turbulence (friction velocity) is fairly well represented in the model. The observation-derived relationship between SHF/u and air-surface temperature difference (ΔT) for different surface layer parameterizations are quite similar for unstable, near-neutral and weakly stable conditions, while they distinctively differ for strongly stable conditions. For strong stability, HB88- and G07-derived data show a decrease in the magnitude of the heat transfer coefficient (CH) for larger ΔT (stable conditions, ΔT > 2 K) at Dome-A and Eagle. This indicates that, for those cases, the reduction in wind shear generated turbulence dominates over the increase in the temperature gradient. This ABL behavior has been also reported for South Pole [Town and Walden, 2009] and the Arctic [Tjernström et al., 2005]. In contrast, the L79-derived wind-scaled SHF show a linear dependency on the temperature gradient for all ΔT at all three stations, indicating a (nearly) single-valued CH. Due to the fact that the model applies the L79 parameterization, the simulations also produce CH that appears to be independent of ΔT, though the slopes of SHF/u as a function of ΔT from model output indicate smaller CH than in the observation L79–derived data. This appears to be due to the wind, surface temperature and stability biases affecting CH through the bulk Richardson number calculation. However, errors in the model parameterization of SHF are not likely the cause of the cold bias in Tsrfc, as the response of SHF/u to changes in ΔT generally agrees with the observations. This means that the model would produce correct fluxes for the correct external forcing. Biases related to the fluxes are therefore most likely a consequence of biases in the forcing.
 While it would be preferable to have direct eddy covariance flux observations at both inland stations Dome-A and Eagle to properly validate the model flux process relationships at these sites, the accomplished application of three frequently used and/or relevant bulk parameterizations and the observed meteorological variables is able to show that the model is within the range of likely relationships. This result is not self-evident although both the modeled and measurement-based fluxes use a bulk approach. The slopes of the discussed flux relationships are functions, e.g., of wind speed, temperature and humidity so that differences between the observed and modeled range of these variables can cause large effects.
 Clear relationships between surface shortwave downward radiation (SWD) and cloud fraction as well as longwave downward radiation (LWD) and cloud fraction are shown for Zhongshan. Both the observations and model show a decrease in SWD and increase in LWD with increasing cloud fraction. However, for overcast conditions, the model simulates too little SWD reaching the surface, which has been discussed to be related to clouds that are optically too thick. The rate of increase of LWD with cloud cover is slightly overestimated in the model for both summer and winter, which could be consistent with an excess of liquid water and hence thermal emissivity for large cloud fractions. However, in winter, the LWD radiation is too low for all cloud fractions, perhaps reflecting a low bias in the temperature of the clouds and/or atmosphere and/or (subcloud) humidity. Compensations between those errors in the model appear to produce the good agreement in LWD for cloudy conditions between the model and Zhongshan observations. The relationships between SWD and LWD with the cloud fraction are clearly nonlinear, with steeper slopes for larger cloud fractions. Linear estimates of these slopes are consistent with those presented elsewhere in the literature. Errors in cloud characteristics in the HIRHAM model produce an obvious deficiency in combined SWD and LWD for clear and cloudy conditions and both summer and winter. The combined errors in downwelling radiation can produce deficiencies of a few to several tens of W/m2 in the net surface radiation. This is the likely cause of the cold Tsrfc bias and the positive bias in near-surface stability in HIRHAM.
 A relationship of increasing near-surface temperature Tair with increasing wind (u) is observed and simulated at Zhongshan in winter. Based on the lowest temperatures, two wind regimes in the observations at all three sites have been found. In the strong/weak wind regime (u > ucrit / u < ucrit) Tair increases/decreases with increasing u. A threshold value ucrit of about 2–6 m/s, which vary with site and season, divides the air-ice interaction process into weak and strong wind regimes. Such separation was recently reported for other polar sites [Lüpkes et al., 2008; Steinhoff et al., 2009]. The weak wind regime describes the decoupling of the atmosphere from the underlying surface at stable stratification, and this effect is only partly reproduced by the model. The model fails to simulate the strong wind regime at the inland stations Dome-A and Eagle.
 In future model improvements, the ECHAM5 physics will be incorporated. The main relevant advances there are associated with radiation and clouds, which, based on the above conclusions, should improve our ability to simulate the Antarctic surface environment. Therefore, we plan to continue our assessment work related to the cloud-radiation relationships. Also, based on the above results, a better simulation of the (near-)surface temperature due to a better snow treatment (e.g., considering a surface snow model or an improved snow albedo) and an improved description of the exchange coefficients CD,H should be given priority in future model improvements.
 This research was supported by the German Bosch Foundation via its program “Science Bridge: Asia” and by the international cooperation program of Chinese Arctic and Antarctic Administration. A.R. thanks the Chinese Academy of Meteorological Sciences, which supported her stay as a visiting scientist in Beijing. The efforts of P.O.G.P. for this work were supported by grants ARC0612428 and ARC1023366 from the National Science Foundation of the United States. We thank I. Hebestadt for her programming support. The authors are grateful to the Australian Antarctic Division and the members involved in the CHINAREN traverse route program for the AWS data collection. We are thankful for the comments of T. Vihma and two anonymous reviewers, which helped to improve the manuscript.