The standard and polar versions 3.1.1 of the Weather Research and Forecasting (WRF) model, both initialized by the 40-yr ECMWF Re-Analysis (ERA-40), were run in Antarctica for July 1998. Four different boundary layer–surface layer–radiation scheme combinations were used in the standard WRF. The model results were validated against observations of the 2-m temperature, surface pressure, and 10-m wind speed at 9 coastal and 2 inland stations. The best choice for boundary layer and radiation parameterizations of the standard WRF turned out to be the Yonsei University boundary layer scheme in conjunction with the fifth-generation Pennsylvania State University–National Center for Atmospheric Research Mesoscale Model (MM5) surface layer scheme and the Rapid Radiative Transfer Model for longwave radiation. The respective temperature bias was on the order of 3°C less than the biases obtained with the other combinations. Increasing the minimum value for eddy diffusivity did, however, improve the performance of the asymmetric convective scheme by 0.8°C. Averaged over the 11 stations, the error growths in 24-h forecasts were almost identical for the standard and Polar WRF, but in 9-day forecasts Polar WRF gave a smaller 2-m temperature bias. For the Vostok station, however, the standard WRF gave a less positively biased 24-h temperature forecast. On average, the polar version gave the least biased surface pressure simulation. The wind speed simulation was characterized by low correlation values, especially under weak winds and for stations surrounded by complex topography.
Despite the considerable progress made in the field of numeric weather prediction, some serious problems remain, one of them being the parameterization of the stable atmospheric boundary layer. There are numerous reasons for this difficulty. First of all, the size of the turbulent eddies in a typical stable boundary layer is small compared to the grid size used in most models (Beare et al. 2006). Second, parameterizing surface fluxes is not straightforward. Many of the assumptions made therein (usually based on the Monin–Obukhov similarity theory) are not valid when the stability increases drastically. There are also many physical processes that are often not well presented in numerical models, such as gravity waves, dominant shear generation of turbulence on top of the boundary layer, and cold air drainage (Mahrt 1998). In this light, the Antarctic region in winter, characterized by a prevailing surface-based temperature inversion and katabatic flows, would be a demanding domain for any NWP model. In addition to problems related to model physics, difficulties in the Antarctic coastal zone also arise from the resolution of essential topographic structures. Moreover, the role of the continent as a huge barrier requires major adjustments in wind and pressure fields (Parish and Cassano 2001).
To meet the requirements of extremely cold regimes such as Antarctica, polar versions have been developed for a few major NWP systems. These include the Polar fifth-generation Pennsylvania State University–National Center for Atmospheric Research (PSU–NCAR) Mesoscale Model (MM5), the Antarctic version of Australian Bureau of Meteorology’s Limited Area Predictive System (LAPS), and the Polar Weather Research and Forecasting model (WRF). The modified Antarctic version of the Australian Bureau of Meteorology’s LAPS was run over East Antarctica with a horizontal resolution of 0.25° by Adams (2004). The emphasis was to perform a single-station verification study for the coastal station Casey and to examine the gale storm events common to the area in question. Bromwich et al. (2005) utilized the Antarctic Mesoscale Prediction System (AMPS), which employed the Polar MM5 with four different domains. They considered the behavior of temperature, air pressure, and wind in local and large-scale situations. The grid lengths in this case were 90, 30, 10, and 3.3 km. Powers and Manning (2007) tested the Polar WRF and the standard version of the same model for 2-week periods in summer and fall conditions in Antarctica with grid sizes ranging from 60 to 2.2 km.
In this paper we concentrate exclusively on the modeling of the Antarctic winter conditions. Various model experiments of the standard and polar versions of the WRF are carried out for July 1998. The results of these simulations are compared to the synoptic data obtained from 11 weather stations around the continent. This paper has three objectives: to examine the sensitivity of the standard WRF V3.1.1 to boundary layer and radiation parameterizations, to compare this version of the model to the Polar WRF V3.1.1 in very stable conditions, and to briefly investigate the effect of finer horizontal resolution on the results. In its comparative manner this study should thus help to distinguish the magnitude of the simulation enhancement achieved with various model configurations.
2. WRF model
a. Basic information
The WRF model is a numerical weather prediction system for both research and operational forecasting purposes. The model has been developed in collaboration among several U.S. institutions. The WRF Software Framework (WSF) includes dynamics solvers, physics packages (linked to the solvers via a standard physics interface), initialization programs, and the WRF variational data assimilation system (WRF-Var; Skamarock et al. 2007). The Advanced Research WRF (ARW-WRF) dynamics solver was used for the model experiments in this study. The full description of this solver is presented by Skamarock and Klemp (2007).
b. Boundary layer schemes
The WRF physics packages include various boundary layer schemes to choose from. Most schemes use dry mixing, but the inclusion of saturation effects in the vertical stability is not ruled out. Once a boundary layer scheme is activated, it takes care of the vertical diffusion in the entire atmospheric column. Horizontal diffusion is managed by selecting either a diffusion choice based on horizontal deformation or one with a constant turbulent exchange coefficient Kh. In the latter case, horizontal and vertical mixing are treated independently. A boundary layer scheme is used together with a surface layer scheme. The main task of a scheme of this type is to calculate friction velocities and turbulent exchange coefficients that enable the determination of the surface stress by the boundary layer scheme, and the surface moisture and heat fluxes by the land surface model (Skamarock et al. 2007). The following three boundary layer schemes were included in our model experiments.
The Yonsei University (YSU) boundary layer scheme is a nonlocal first-order closure scheme that assumes that the turbulent flux of a quantity moves down the mean gradient of the quantity with a rate of the movement proportional to eddy diffusivity. The YSU scheme was developed on the basis of the medium-range forecast (MRF) PBL scheme. The major difference between the two is the explicit treatment of entrainment processes at the top of the boundary layer, which is included in the YSU scheme (Hong and Noh 2006). The surface layer scheme used with this alternative is called the MM5 scheme. It employs the Monin–Obukhov similarity theory (e.g., Skamarock et al. 2007).
The Asymmetric Convective Scheme (ACM2) also employs a first-order closure scheme. The ACM exhibits fast upward air transport by buoyancy plumes and slower downward movement by the compensating subsidence. Some local mixing can be present at all levels, which results in more continuous profiles in lower layers. In the model runs the ACM scheme is accompanied by the Pleim–Xiu (PX) surface layer scheme (Pleim 2007a).
The formulation of the Mellor–Yamada–Janjic (MYJ) 2.5-order closure scheme is based on a prognostic equation for the turbulent kinetic energy (TKE), which is solved iteratively. The scheme applies an upper limit for the turbulent master length scale in order to keep the denominators in the stability functions from declining to zero. This limit depends on buoyancy and TKE (Nakanishi and Niino 2006). If conditions are stable, the upper limit is defined using the requirement that the ratio of the variance of the vertical velocity deviation and the TKE must not be smaller than that corresponding to the regime of disappearing turbulence. The MYJ boundary layer scheme is run in conjunction with the Eta surface layer scheme, which is also based on the Monin–Obukhov theory (e.g., Skamarock et al. 2007).
c. Radiation schemes
The schemes for longwave and shortwave radiation tested in the V3.1.1 model experiments are the Rapid Radiative Transfer Model (RRTM), the Dudhia shortwave scheme, and the Community Atmospheric Model (CAM) schemes for longwave and shortwave radiation. The RRTM and the Dudhia scheme both originate from the MM5 model. The former uses preset tables to describe longwave processes caused by water vapor, ozone, carbon dioxide, and trace gases (Mlawer et al. 1997). The Dudhia scheme features a simple downward integration of solar flux, accounting for water vapor absorption, clear-air scattering, and cloud albedo and absorption (Dudhia 1989). The CAM schemes for longwave and shortwave radiation are from the CAM 3 climate model. The longwave radiative transfer is based on an absorptivity/emissivity formulation by Ramanathan and Downey (1986), whereas the shortwave solution uses the δ-Eddington approximation for the effects of multiple scattering (Briegleb 1992).
d. Polar WRF
The Polar WRF has been developed by the Ohio State University. Compared to the standard version of the model the polar modification uses the following adjustments described by Bromwich et al. (2009). The subgrid-scale cumulus parameterization is run by the Grell and Devenyi ensemble scheme (Grell and Devenyi 2002) and cloud physics is run by the WRF single moment five-class microphysics scheme (Hong et al. 2004). For longwave radiation the RRTM is used (Mlawer et al. 1997). Shortwave radiation is described by the Goddard scheme (Chou and Suarez 1994). The planetary boundary layer is represented by the MYJ scheme in conjunction with the Eta surface layer scheme. The land/ice surface is simulated applying the unified Noah land surface model (Chen and Dudhia 2001) with several modifications to its standard version (Hines and Bromwich 2008). These changes made for simulations on the Greenland Ice Sheet include the use of latent heat of sublimation for calculations of latent heat fluxes over ice surfaces, an adjustment of thermal diffusivity and snow heat capacity for the subsurface layer, an increase in the value of snow albedo and emissivity value for snow, an adjustment of snow density, the assumption of ice saturation in surface saturation mixing ratio calculations over ice, and a modification of the skin temperature calculation. Moreover, the Polar WRF takes into account fractional sea ice coverage in grid cells whereas the standard WRF V3.1.1 considers cells fully covered or fully open as a default setting (Powers and Manning 2007). The main differences between the standard WRF and its polar modification used in our study are listed in Table 1.
The version of the WRF used for the ARW experiments in our study was 3.1.1, released in July 2009. There were two domains under consideration: the larger of which was centered at the South Pole and had horizontal dimensions of 5900 km × 5400 km. The horizontal grid size was 100 km. The smaller one was centered approximately at (74.00°S, 8.00°W) and had a resolution of 20 km. An illustration of the domains is presented in Fig. 1. In the vertical regime, 30 σ levels were used in all experiments. The vertical coordinate σ is defined as σ = (ph − pht)/(phs − pht) where ph is the hydrostatic component of pressure, and phs and pht stand for the pressure values along the surface and top boundaries, respectively (Laprise 1992). The initialization of the WRF was carried out by the 40-yr European Centre of Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40; Uppala et al. 2005) with a horizontal resolution of 1.125°. We selected ERA-40 for this purpose because Monaghan et al. (2003) had found it superior over a few other models because of its high resolution near the South Pole and its high-quality initial conditions. The ERA-40 data were also used as lateral boundary conditions for the WRF, updated with 6-h intervals. A time step of 60 s was applied in the model. The WRF Preprocessing System (WPS) was utilized to ensure that all the variables were horizontally interpolated to the correct gridpoint staggering. The correct wind rotation to the WRF model map projection was also accomplished by the WPS (Wang et al. 2009). The terrain elevations on Antarctica were described by the Global 30 Arc-Second Elevation Dataset (GTOPO30) from the U.S. Geological Survey (USGS) and the land use by 24-category data of the USGS.
For model validation, we used observations from 11 Antarctic SYNOP weather stations: Halley (75.35°S, 26.36°W), Casey (66.28°S, 110.52°E), Vostok (78.27°S, 106.52°E), Dumont d’Urville (66.67°S, 140.02°E), Rothera (67.57°S, 68.13°W), Mawson (67.36°S, 62.52°E), Syowa (69.00°S, 39.35°E), Neumayer (70.62°S, 8.37°W), Davis (68.35°S, 77.58°E), Vernadsky (65.25°S, 64.27°W), and Aboa (73.05°S, 13.42°W). The elevations of these stations were taken from Turner and Pendlebury (2004). The geographical positioning of the stations is shown in Fig. 1.
The highest σ level was set to be approximately at 10 hPa for both the standard and polar version of the model. This high level for the model top enables better treatment of upward-propagating gravity waves generated by topography (Guo et al. 2003). When employing 30 σ levels the lowest 10 levels were situated approximately at 10, 20, 40, 70, 110, 160, 220, 290, 370, and 460 m from the local surface. Four different radiation–surface layer–boundary layer scheme combinations were used for the experiments testing the sensitivity of the standard version of the model to boundary layer and radiation parameterizations. The model experiments are summarized in Table 2.
Two 9-day experiments, one for the standard WRF and one for the Polar WRF, were carried out in the purpose of finding out how errors develop during the simulation. Subsequently, five 30-day simulations consisting of successive model runs of the selected length (24 h) were completed. Four of these simulations featured different radiation–surface layer–boundary layer scheme combination and one of them had a boundary layer scheme (ACM2) with a modified value for the minimum eddy diffusivity. Having completed this, a Polar WRF simulation of 30 days was also implemented. Finally, we experimented the influence of increased horizontal resolution on the 10-m wind speed simulation in two experiments (one for the standard version of the model and one for the Polar WRF) employing two-way nesting. The second domain with a horizontal resolution of 20 km was placed in the Dronning Maud Land (Fig. 1). The simulation length in this case was 5 days, from 17 to 21 July 1998. In the case of the second domain we applied a time step of 15 s. This domain took the boundary information from the outer domain and was run for four time steps before feeding information back to the parent domain on the coincident interior points.
For the 30-day experiments we present the monthly mean biases, RMS errors, and correlation coefficients (Table 3), and time series of the validation results for selected stations in Figs. 2 and 3, so that they include examples of stations with good and poor model performance. The 30-day time series obtained for the stations depicted in Fig. 1 had 240 SYNOP weather observations and the corresponding databased on the simulations. The exceptions were Vostok and Vernadsky where the scarcity of the synoptic data available led to the usage of 120 and 60 observations, respectively.
It had to be taken into account that the nearest model grid point did not geographically coincide with the station itself. Therefore, the difference in height between these two locations was determined (Table 4) and subsequently, any necessary corrections for temperature and air pressure were carried out. Although the vertical temperature profile during the Antarctic winter often includes a surface-based inversion, the correction should be based on the typical temperature profiles along a sloping surface. Hence, we applied an adiabatic lapse rate of 0.010°C m−1 and the hypsometric equation for pressure. The absolute values of the temperature corrections were on the order of 0°–3°C and the corrections for surface pressure ranged between 0 and 51 hPa. However, the temperature profiles along a sloping surface are not always adiabatic in the Antarctic winter conditions: in the interior parts of the continent, surface temperatures decrease superadiabatically along the slope because of the wind conditions whereas near the coast the decrease is less than adiabatic (King and Turner 1997, 82–83). The temperature corrections did not hence completely remove the elevation-induced errors.
Before any comparisons between the observational data and model results are considered, a few important aspects have to be taken into account. As the horizontal resolution of the model runs for the domain is as coarse as 100 km, it means that in the worst case the nearest grid point is approximately 70 km away from the station itself. Even though the adiabatic temperature and pressure corrections were made, other influences of the significant local topographic variations were left untouched. This is especially important in the steep coastal areas where the nearest grid point may be on sea ice or high up in the outskirts of the glacier. Such spatial differences affect above all the wind comparisons as the wind field over Antarctica is highly dependent on topography (Parish and Bromwich 1998). Temperatures can be biased as well, especially if the sea ice cover is broken by coastal polynyas or leads, in which case the heat flux from the open water starts playing an important role in the air temperature distribution near the surface (e.g., Valkonen et al. 2008).
a. Standard WRF
1) Sensitivity to boundary layer parameterizations
(i) Air temperature
The 2-m temperature simulation reveals clear differences in the performances of different boundary layer–surface layer scheme combinations. The best temperature simulation is given by the YSU–MM5 (experiment 2) with a mean bias of −0.1°C (Table 3). The differences in the RMS errors are also clearly distinguished in favor of the YSU–MM5. The superiority of the MYJ–Eta (experiment 3) over the ACM2–PX (experiment 4) is 0.6°C. However, increasing the minimum value of eddy diffusivity with one order of magnitude in the ACM2 boundary layer scheme (experiment 6) improves the temperature simulation: the mean bias is 0.8°C smaller than in experiment 4 employing the unmodified ACM2 scheme.
On the coldest station (Vostok; Fig. 2), the MYJ–Eta and ACM2–PX parameterizations produce better results than the YSU–MM5. This superiority is on the order of 1°–2°C. On warmer stations such as Rothera and Casey this setting is reversed with the YSU–MM5 giving mean temperature biases superior to the ones obtained with the other two boundary layer parameterizations; the temperatures obtained with MYJ–Eta and ACM2–PX are up to 2°–3°C inferior.
(ii) Atmospheric pressure
On 6 stations, the surface pressure is well simulated with RMS errors around 4 hPa or less and the absolute value of the mean bias around 1 hPa. Three of the remaining stations (Vostok, Casey, and Mawson) have a pressure bias around 5 hPa or greater. At Neumayer and Aboa, the simulation yields an approximate bias of −3 hPa (Table 5). The pressure correlation is clear on all stations.
The pressure simulations do not show significant sensitivity to boundary layer parameterizations (Table 3). The mean pressure biases are close to 2 hPa and the RMS errors around 4–6 hPa in all cases. The correlation values are almost identical.
(iii) Wind speed
The severest problem concerning the wind speed simulation is not a large bias but often a devastatingly bad correlation (Tables 3 and 5). Vernadsky, Rothera, and Casey stand out as pinnacles not even having a weak correlation for any of the three boundary layer–surface layer scheme combinations. The simulations for other stations do not have good results either, excluding Neumayer and Davis, which both show a clear correlation. Figure 4 reveals that the correlation values have the tendency of climbing in the right direction as the mean observed wind speed increases, although the relationship between the two quantities is not unambiguous. The ACM2–PX (experiments 4 and 6) provides wind speed RMS error and bias slightly better than the other two parameterizations, but this superiority is not observable in correlation coefficients (Table 3).
2) Sensitivity to radiation parameterizations
As the best results temperature-wise are given by the YSU–MM5 boundary layer–surface layer scheme combination, we use this configuration together with two different longwave–shortwave scheme combinations to distinguish the sensitivity of the model to radiation parameterizations. The results given by experiment 5 (CAM schemes) are inferior to the ones given by experiment 2 (RRTM and Dudhia schemes): the mean 2-m temperature bias and RMS error are −1.3° and 5.1°C in the former whereas the latter features the values of −0.1° and 4.7°C for the corresponding quantities. The temperature correlation values do not show any distinctive difference between the two experiments. Apart from a 1-hPa improvement in the RMS error of surface pressure when changing from RRTM–Dudhia to CAM, no significant changes in the errors related to surface pressure and 10-m wind speed are observed.
b. Polar WRF
Experiments 1 and 8 (9-day simulations with the standard WRF and the Polar WRF; Fig. 5) show that the temperature bias given by the standard version is, on the whole, slightly more negative than that of the Polar WRF. However, during the first 24 h the polar version gives more negatively biased values. The air pressure biases given by the two versions are both, for most parts, positive and stay relatively close to one another up to day 5. In the case of the 10-m wind speed, the polar version tends to yield slightly higher wind speed bias than the standard version after day 2.
The standard WRF is observed to give a pronounced negative temperature bias at Rothera and a clear warm bias at Neumayer. It is therefore relevant to compare the performance of the two model versions on these stations to see if the Polar WRF can produce better results. The improvement in the temperature fit at Rothera is distinguishable, but at Neumayer the polar version is even more positively biased than the standard WRF (Fig. 2). On average, the polar version gives a mean temperature bias 0.1°C inferior to the best result obtained with the standard WRF (employing the YSU–MM5 boundary layer parameterization). The mean absolute temperature biases for the standard and polar versions are 2.9° and 3.4°C, respectively. Moreover, the temperature time series for Vostok (Fig. 2) shows a clear warm bias related to the polar version of the model.
The effects of the Polar WRF on the surface pressure and 10-m wind speed are depicted in Fig. 3. The polar version does produce a less-biased pressure simulation, but the respective value of RMS error is larger than that achieved with the standard WRF (Table 3). In the wind speed simulation no notable differences between the two model versions are observed.
c. Sensitivity to horizontal resolution
The 10-m wind speed simulation at Halley and Neumayer undergoes a clear improvement when changing the horizontal grid length from 100 to 20 km. The results presented in Table 6 and Fig. 6 show that the errors in wind speed decrease for these two stations. The Polar WRF with a 20-km resolution gives the best results, even yielding correlation values as good as 0.9 for the 5-day period. The coarser horizontal resolution leads to overestimating the wind speed in Halley and underestimating it in Neumayer. Both the standard and polar version with the 20-km resolution capture well the strong wind event at Neumayer on 20 July even though the initial values given by ERA-40 are erroneous. The situation is quite different at Aboa, surrounded by more complex topography: the increase in horizontal resolution does not yield in any improvements in the wind speed biases and RMS errors observed in experiments 2 and 9 (Table 6).
5. Discussion and conclusions
The results pointed out that by selecting the YSU–MM5 boundary layer parameterization the temperature simulation (24-h forecasts) of the standard WRF gives practically as good results as the Polar WRF for the reference stations in the very stable Antarctic winter conditions. Results of this kind are surprising as the Polar WRF has been specifically designed for polar applications (Bromwich et al. 2009). However, the polar version did yield better results in the 9-day experiment and with the 20-km horizontal resolution.
The temperature simulations in the case of both model versions were still far from perfect. On the coldest station, Vostok, the Polar WRF suffered a clear defeat as the standard WRF gave a better, less positively biased temperature simulation. On the other hand, at Rothera the standard WRF suffered from a strong negative bias. The Polar WRF was also compared to the standard version in Antarctica by Powers and Manning (2007). Their simulations for Dome C (1000 km from Vostok) on 6–30 April 2007 with a model run length of 24 h showed a difference of 2.02°C in the temperature biases in favor of the Polar WRF. The corresponding result obtained for Vostok in our study was an approximate superiority of 5°C of the standard WRF (with YSU–MM5) over the polar version. The drastic difference between the two studies is explained by the fact that Powers and Manning (2007) applied an older standard WRF version. Their temperature simulation for April featured a distinct positive bias for all three stations considered, whereas in this study the version 3.1.1 produced negative mean temperature biases on 6 stations out of 11.
One of the notable results of this study was the superiority of the YSU–MM5 boundary layer–surface layer scheme combination in the 2-m temperature simulation over the other two combinations considered. Bromwich et al. (2009) did a sensitivity test for the Polar WRF over the Arctic Ocean. The difference in the mean bias of the 2-m temperature between the YSU and MYJ schemes was only 0.3°C whereas the corresponding value obtained in our study was 2.9°C. Bromwich et al. (2009) used the Polar WRF. This, together with less stable stratification and lack of orography, may have contributed to more accurate results: the RMS error was 2.1°C for the YSU scheme and 2.3°C for the MYJ scheme. The corresponding numbers in our study were 4.7° and 5.9°C.
Gilliam et al. (2007) investigated the differences in the performances of the YSU and ACM schemes on 900 different sites in the United States. This study done in summer conditions revealed that, in general, the 2-m temperature simulation carried out by the ACM parameterization did not succeed as well as the one with the YSU scheme, the RMS error being 2.33°C in the former case and 2.19°C in the latter. In our very stable winter conditions both the errors and differences between the results given by different schemes became much more pronounced: the mean RMS error for the YSU–MM5 obtained in this study was 4.7°C whereas the ACM2–PX combination produced the value of 6.3°C.
A noteworthy characteristic of the results of this study was the distinctive negative temperature bias produced by each and every parameterization on many coastal stations. The Polar MM5 has also showed this type of behavior. Bromwich et al. (2005) employed the AMPS Polar MM5 with a 30-km horizontal resolution for the Antarctic region. They concluded that the results calculated for the 24-month period (September 2001–August 2003) showed slightly positive biases for high elevations and negative for most coastal stations.
The reasons leading to the negative temperature bias may be numerous. One of the simplest ones is incorrect station elevations discussed earlier. The differences between the modeled and real elevations could not be properly dealt with because of the uncertainties in the adiabatic temperature correction method. We also tried applying the surface temperature equations by King and Turner (1997, p. 83) instead of the adiabatic method, but this did not lead to any improvements in the results. It is noteworthy that the temperature RMS error and bias did not systematically depend on the difference between the elevations of the station and the model grid point. This difference is of the same order for Davis, Rothera, Casey, and Vernadsky (Table 4), but the temperature biases are clearly negative for Rothera and Casey whereas other two stations have biases close to zero. On the basis of this, the importance of the adiabatic-correction-related error to the difference between modeled and observed temperatures can be questioned. Instead, subsidence heating due to local downslope winds not resolved by our simulations may well have increased the magnitude of the negative bias at Rothera and Casey.
Adams (2004) carried out a 3-month temperature simulation for the coastal station of Casey in 2001 with the Australian Global Assimilation and Prediction System. He utilized 48-h model runs with a two-step data assimilation process carried out in the beginning of each run. The obtained mean bias for the temperature at Casey on the lowest model level was slightly positive, 0.3°C, whereas the corresponding bias for Casey obtained in this study ranged from −1.4° to −4.4°C depending on the boundary layer parameterization used. According to Pleim (2007b) the positive temperature bias often related to the ACM scheme, especially in stable conditions, may be caused by too large a minimum value for eddy diffusivity. This applies to other PBL schemes as well: Steenveld et al. (2006) found that the MM5 model overpredicted temperature using the MYJ scheme during very stable nights. In the case of our domains, it could be too low a value for eddy diffusivity, which reduces the amount of turbulence too much and is, at least partly, responsible for the cold bias in the 2-m temperature. The results obtained in experiment 6 point in this direction: increasing the minimum value for eddy diffusivity by 1 order of magnitude had an improving effect of 0.8°C on the average 2-m temperature bias. It is apparent that in order to avoid the overestimated surface cooling rates caused by the turbulent fluxes approaching zero, such minimum values for turbulent diffusion are needed (Vihma et al. 2005; Savijärvi 2009), but it is difficult to find minimum values suitable for all conditions. This is partly a result of the variable contribution of the unresolved flow features.
At the coldest station Vostok, the Polar WRF yielded a distinct positive temperature bias. According to Otieno et al. (2009), one error source might be a bias in the ERA-40 snow temperatures. Indeed, when comparing the averaged snow temperatures 0–255 cm below the surface measured at the South Pole in July 1992 (Brandt and Warren 1997) to the corresponding values given by ERA-40 for the same period, we detected a warm bias in the ERA-40 values ranging from 4° to 8°C.
The sensitivity test on radiation schemes (experiments 2 and 5) had the same outcome as the study of Borge et al. (2008) for the Iberian Peninsula in southern Europe. The use of the RRTM longwave scheme yielded approximately 1°C better results in 2-m temperature than the CAM scheme, which gave a slightly negative bias in both cases. This suggests that the sensitivity of the model to radiation parameterizations does not alter much with changing stability.
Relatively poor correlation values were the most intrinsic feature of the wind speed simulation: the average values close to 0.5 were produced by all three boundary parameterizations of the standard WRF and also by the polar version. The correlation coefficient strongly depended on the local environment with the lowest values of 0.2–0.3 for the stations surrounded by complex topography (Rothera and Vernadsky). The topography is not quite as complex around Casey and Syowa, but large mesoscale spatial variations in the wind field are common there (Turner and Pendlebury 2004). In the case of Halley and Vostok, the decrease of the model performance was probably due to the very stable stratification, whereas in Dumont d’Urville the probable reason is the strong katabatic wind. Applying a model resolution of 30 km, Bromwich et al. (2005) observed mean correlation values close to ours. However, experiments 7 and 10 in our study revealed that decreasing the horizontal grid length even further to 20 km does result in an improved wind speed simulation at Halley and Neumayer. These stations are, yet, situated on flat ice shelves. To achieve such an improvement for a station with complex topography, the resolution should be much higher. Powers (2007) used the ARW to simulate a severe wind event that struck McMurdo in May 2004. Despite the horizontal resolution of 3.3 km, the results achieved without data assimilation featured wind speeds half of those observed. Figure 6 depicts the same type of situation at Aboa: during the most extreme winds the modeled wind speed values were even in the best case 15 m s−1 lower than the observed ones, which were affected by channeling along the slope of a nunatak. In this specific region, a horizontal resolution of approximately 1 km is needed for a reasonable wind simulation (Valkonen et al. 2010).
In our study, the YSU–MM5 and MYJ–Eta parameterizations were found to yield approximately the same mean wind speed biases. This is contradictory to the results obtained by Bromwich et al. (2005). In their study for the Arctic region the MYJ scheme had a 0.4 m s−1 smaller mean bias than the YSU scheme and even the correlation of the former was a bit better for the wintertime conditions. The ACM2–PX parameterization gave, however, the least-biased wind speed simulation in our study, outplaying the MYJ–Eta by 0.6 m s−1 Yet, somewhat contradicting results were obtained by Krieger et al. (2009) for the Arctic Beaufort Sea region: the ACM produced a bias 0.4 m s−1 larger than the MYJ. A uniform conception of the best possible boundary layer parameterization option for the wind speed is thus difficult to find.
The good results produced by the air pressure simulation were not surprising; success in such work for the Antarctic region has been reached before (e.g., Bromwich et al. 2005; Powers and Manning 2007). However, the results of the pressure simulation did feature two stations (Casey and Vostok) with a distinct positive pressure bias on the order of 7–8 hPa. Positive pressure anomalies for Casey have also been observed by Adams (2004) though the bias in his study was half smaller. A probable reason for the biases at Casey is mesoscale low pressure systems not being resolved by numerical models. According to Turner and Pendlebury (2004) these lows in the vicinity of Casey are almost impossible to forecast.
On the whole, the results suggested that in Antarctic winter simulations the standard WRF with the YSU boundary layer parameterization and the Polar WRF perform equally well. No validation was, however, made over the sea ice zone where the Polar WRF probably benefits from the more sophisticated treatment of leads and polynyas. Further research should be carried out with the boundary layer, cloud, and snow parameterizations and boundary conditions so that they would succeed better in very stable conditions.
We thank David Bromwich for providing us with the Polar WRF model code and Hannu Savijärvi for fruitful discussions. The study was supported by the Academy of Finland (Contracts 128533 and 128799). The computing resources were provided by CSC–IT Center for Science, Espoo, Finland.