Evaluation of Polar WRF from Modeling the Atmospheric Boundary Layer over Antarctic Sea Ice in Autumn and Winter

Transcription

1 DECEMBER 2012 T A S T U L A E T A L Evaluation of Polar WRF from Modeling the Atmospheric Boundary Layer over Antarctic Sea Ice in Autumn and Winter ESA-MATTI TASTULA University of Helsinki, Helsinki, Finland TIMO VIHMA Finnish Meteorological Institute, Helsinki, Finland EDGAR LANDREAS NorthWest Research Associates, Lebanon, New Hampshire (Manuscript received 12 January 2012, in final form 15 May 2012) ABSTRACT Regional simulations of the atmospheric boundary layer over Antarctic sea ice that have been adequately validated are rare. To address this gap, the authors use the doubly nested Polar Weather Research and Forecasting (Polar WRF) mesoscale model to simulate conditions during Ice Station Weddell (ISW) in the austral autumn and winter of The WRF simulations test two boundary layer schemes: Mellor Yamada Janjic and the Asymmetric Convective Model. Validation is against surface-layer and sounding observations from ISW. Simulated latent and sensible heat fluxes for both boundary layer schemes had poor correlation with the observed fluxes. Simulated surface temperature had better correlation with the observations, with a typical bias of 0 2 K and a root-mean-square error of 6 7 K. For surface temperature and wind speed, the Polar WRF yielded better results than the ECMWF Re-Analysis Interim (ERA-Interim). A more challenging test of the simulations is to reproduce features of the low-level jet and the temperature inversion, which were observed, respectively, in 80% and 96% of the ISW radiosoundings. Both boundary layer schemes produce only about half as many jets as were observed. Moreover, the simulated jet coincided with an observed jet only about 30% of the time. The number of temperature inversions and the height at the inversion base were better reproduced, although this was not the case with the depth of the inversion layer. Simulations of the temperature inversion improved when forecasts of cloud fraction agreed to within 0.3 with observations. The modeled inversions were strongest when the incoming longwave radiation was smallest, but this relationship was not observed at ISW. 1. Introduction The atmospheric boundary layer (ABL) over sea ice is predominantly stably stratified for most of the year and is near neutral during summer (Andreas et al. 2000; Persson et al. 2002; Grachev et al. 2005). Physical processes in a stable boundary layer (SBL) are quantitatively not well known and include complicated interactions among turbulence, waves, and longwave radiation. Over sea ice, which is usually broken by cracks, leads, and Corresponding author address: Esa-Matti Tastula, University of Helsinki, PL 64 (Gustaf Hällströmin katu 2a), Helsinki 00014, Finland. tastulae@mail.usf.edu polynyas, the situation is further complicated by localized convection over the areas of open water (e.g., Andreas et al. 1979; Lüpkes et al. 2008). Temperature and humidity inversions often characterize the SBL over sea ice. In winter, temperature inversions are often surface based, whereas elevated inversions prevail in summer (Serreze et al. 1992). The mechanisms generating temperature inversions include radiative cooling of the surface and the air, sensible heat flux from the air to the surface, warm-air advection over a cold surface, and subsidence. Specific humidity inversions often coincide with temperature inversions (Serreze et al. 1995). Horizontal advection of moist air, subsidence, condensation, gravitational fallout of the condensate, deposition of hoar frost at the surface, and DOI: /MWR-D Ó 2012 American Meteorological Society

2 3920 M O N T H L Y W E A T H E R R E V I E W VOLUME 140 turbulent transport of moisture contribute to the generation and maintenance of humidity inversions (Curry 1983). In the Antarctic, several studies have addressed the ABL over coastal polynyas (e.g., Bromwich and Kurtz 1984; Kottmeier and Engelbart 1992; Renfrew and King 2000; Renfrew et al. 2002; Cassano et al. 2010), but relatively few meteorological field experiments have been carried out in the inner Antarctic sea ice zone. The first observations originate from vessels stuck in sea ice: Deutschland in 1911 and Endurance in 1915 (Meinardus 1938). The first wintertime crossing of the Antarctic sea ice zone by a ship took place in 1981 the Weddell Polynya Expedition and yielded results on air sea exchange processes, ABL structure, and the effects of warm-air advection (Andreas 1985; Andreas et al. 1984; Andreas and Makshtas 1985). Kottmeier and Hartig (1990) later applied ship-based radiosonde soundings to study baroclinic zones in the troposphere over the icecovered Weddell Sea. Year-round observations have been collected by drifting buoys, but these data are restricted to the lowest meters of the atmosphere (Kottmeier and Hartig 1990; Launiainen and Vihma 1994; Vihma et al. 1996, 2002). Only three drifting ice stations with dedicated ABL observations have been organized in the Antarctic sea ice zone. Two of them lasted for only a relatively short period: Ice Station Polarstern lasted for 5 weeks in December January 2004/05 (Bareiss and Görgen 2008; Vihma et al. 2009), and the Sea Ice Mass Balance in the Antarctic (SIMBA) for two weeks in September October 2007 (Vancoppenolle et al. 2011). The only seasonal ice station has been the U.S. Russian Ice Station Weddell (ISW) in February June 1992 in the western Weddell Sea. It included observations on the atmospheric surface layer (Andreas 1995; Andreas and Claffey 1995; Andreas et al. 2004, 2005) as well as the whole ABL (Andreas et al. 2000). Several meteorological modeling studies over the Antarctic sea ice zone have been carried out without concrete validation of the results (Simmonds and Budd 1991; Vihma 1995; Watkins and Simmonds 1995; Dare and Atkinson 1999; Birnbaum 2003). Most of these addressed the importance of leads and polynyas. Vihma et al. (2002) validated the operational analyses of the European Centre for Medium-Range Weather Forecasts (ECMWF) and the reanalysis of the National Centers for Environmental Prediction National Center for Atmospheric Research (NCEP NCAR) against buoy data from the Weddell Sea. Large errors were found in near-surface air temperature and the turbulent fluxes of sensible and latent heat. The errors were partly related to an oversimplified representation of sea ice in the models. The Polar fifth-generation Pennsylvania State University NCAR Mesoscale Model (MM5) model and buoy data were applied by Valkonen et al. (2008) in a late-autumn case study over the Weddell Sea. During a cold-air outbreak, the modeled 2-m air temperature was very sensitive to the sea ice concentration data applied as a lower boundary condition. In the western Weddell Sea, the region of the present study, ice conditions resemble those in the Arctic, with thick, deformed ice and a high ice concentration. In most other parts of the Antarctic, the sea ice cover is thinner and the ice concentration is lower (Hibler and Ackley 1982; Gloersen et al. 1992). The climate in the western Weddell Sea is some 7 K colder than at the same latitudes on the western side of the Antarctic Peninsula (Vaughan et al. 2003). The wind field is often affected by a low over the eastern Weddell Sea and barrier winds close to the Antarctic Peninsula (Parish 1983). In this study, we apply the Polar version of the Weather Research and Forecasting (WRF) model (Polar WRF; Hines and Bromwich 2008), the successor of Polar MM5, to simulate the ISW period with detailed validation against the observations of Andreas et al. (2000, 2004, 2005). In addition to near-surface variables, our focus is on the properties of temperature and humidity inversions and low-level jets, which strongly affect the atmosphere ice ocean exchange processes. Experiences from the Arctic have demonstrated that inversions and low-level jets are particularly challenging to simulate (Kilpeläinen et al. 2012), even with the help of data assimilation (Lüpkes et al. 2010; Tjernström and Graversen 2009). We aim for a better understanding of the main challenges in ABL modeling over the Antarctic sea ice. Our specific objectives are 1) to identify the meteorological conditions, variables, and processes most liable for large model errors; 2) to compare the performance of two ABL parameterizations in the Polar WRF; and 3) to compare the performance of Polar WRF and the ECMWF Re-Analysis-Interim (ERA-Interim) (Dee et al. 2011). In addition to presenting model skill scores, we study how well the Polar WRF can reproduce the observed relationships among variables. Validation of the Polar WRF in the Antarctic is particularly important because the cornerstone for operational weather prediction in the Antarctic, the Antarctic Mesoscale Prediction System (Powers et al. 2012; Bromwich et al. 2005; Fogt and Bromwich 2008), is based on the Polar WRF; however, previous validation of the Polar WRF has mostly been restricted to the Antarctic continent and ice shelves. We summarize the meteorological observations at ISW in section 2 and follow with a presentation of the Polar WRF model (section 3) and our simulation strategy (section 4). The results are presented in section 5

3 DECEMBER 2012 T A S T U L A E T A L and the discussion in section 6. Finally, conclusions are given in section Measurements at Ice Station Weddell Ice Station Weddell was established with a broad objective of investigating the interactions among air, sea, and ice in the western Weddell Sea. The turbulent exchange of momentum, heat, and moisture between seaiceandatmospherewasstudiedonthebasisofnearsurface observations (Andreas and Claffey 1995; Andreas et al. 2004, 2005) and radiosoundings through the atmospheric boundary layer (Claffey et al. 1994; Andreas et al. 2000). A 5-m instrument tower measured near-surface quantities. Temperature and dewpoint temperature were recorded at a height of 4.65 m (hereafter referred to as 5-m temperature and dewpoint temperature). The fluctuations in temperature and in the longitudinal, transverse, and vertical components of the wind vector were obtained with a three-axis K-type sonic anemometer/ thermometer from Applied Technologies Inc., also at 4.65 m. A Lyman-alpha hygrometer measured humidity fluctuations near the sonic. For more detailed description of the instrumentation, see Andreas et al. (2004, 2005). The first sounding was made at 0000 UTC 21 February 1992 and the last at 0000 UTC 4 June There were 164 soundings in total, 103 of which were carried out with Tethersondes (instruments carried by a tethered, aerodynamically shaped balloon), and the rest with Airsondes (on free-flying helium balloons), both from Atmospheric Instrumentation Research. The Tethersondes had a 10-s sampling interval that gave a typical vertical resolution of m. The sondes typically reached elevations from 500 to 600 m. These devices were equipped with sensors providing pressure, temperature, dewpoint, wind speed, and wind direction. Airsondes had a finer sampling interval of 5 s, a vertical resolution of m, and typically reached heights between 6000 and m before the radio signal was lost. The Airsondes did not measure wind speed nor wind direction (Claffey et al. 1994). Andreas et al. (2000) used the protocol of Kahl (1990) to identify inversions. According to this approach, if the temperature at a level increased, the level immediately below it was called the base of the inversion. Moving upward, if the temperature at a subsequent level started to decrease again, the level immediately below it was named the top of the inversion (Fig. 1). Andreas et al. (2000) ignored thin layers with slightly negative lapse rates that were embedded within a deeper inversion. When calculating the inversion statistics based on the model results, we thus omitted such layers thinner than 90 m with a lapse rate larger than K (25 m) 21. For defining a low-level jet (LLJ), Andreas et al. (2000) adapted the definition by Stull (1988) with a deviation concerning the possible positioning of the jet: an LLJ could not occur at the surface. The approach was as follows: if the wind speed profile showed a local maximum that was 2 m s 21 higher than the speed both above and below, the maximum was called a jet (Fig. 1). 3. Polar WRF The high sensitivity of the polar regions to climate change and many challenging aspects related to numerical modeling in this specific region have contributed to the need for developing polar versions of major numerical weather prediction systems. The WRF model, developed as a collaboration among many U.S. institutions (Skamarock and Klemp 2007), had its first polar version introduced by Hines and Bromwich (2008). This Polar WRF incorporates many modifications to the standard version of the WRF. These adjustments are described by Bromwich et al. (2009), for example, and include, especially, adjustments to the surface parameterizations. The changes made in the Noah land surface model (LSM; Chen and Dudhia 2001) include using the latent heat of sublimation for calculating latent heat fluxes over ice surfaces, increasing the snow albedo and the emissivity value for snow, adjusting snow density, modifying thermal diffusivity and snow heat capacity for the subsurface layer, changing the calculation of skin temperature, and assuming ice saturation in calculating the surface saturation mixing ratio over ice (Powers and Manning 2007). In addition to the Noah LSM, the Polar WRF version used in our model experiments has the following physical parameterizations: microphysics by the Morrison double-moment scheme (Morrison et al. 2005), cumulus parameterization by the Grell 3D ensemble cumulus scheme (Grell and Devenyi 2002), longwave radiation by the Rapid Radiative Transfer Model (Mlawer et al. 1997), shortwave radiation by the Goddard scheme (Chou and Suarez 1994), and the planetary boundary layer by the Mellor Yamada Janjic (MYJ) scheme with prognostic turbulent kinetic energy, in conjunction with the Eta surface layer scheme (Nakanishi and Niino 2006). In our study, we also carried out a sensitivity test using the Asymmetric Convective Model (ACM2) boundary layer scheme (Pleim 2007) instead of the MYJ scheme. The ACM2 scheme has been specifically designed for convective boundary layers with a representation for both supergrid- and subgrid-scale components for turbulent transport. It is therefore interesting to assess its ability to cope with very stable conditions.

4 3922 M O N T H L Y W E A T H E R R E V I E W VOLUME 140 FIG. 1. (left) Parameters of the temperature inversion: z b is the height of the base of the inversion; T b is the temperature there; z t is the height at the top of the inversion, T t is the temperature there. Adapted from Kahl (1990) by Andreas et al. (2000). (right) Parameters of low-level jet: z j is the height of the jet core; U j is the wind speed in the core. Adapted from Stull (1988) by Andreas et al. (2000). For the SBL, the main difference between the two schemes lies in the way they calculate the eddy viscosities for heat, momentum, and moisture. MYJ determines the spatially varying viscosities based on turbulent kinetic energy, local gradients, and a diagnosed length scale. ACM2 combines nonlocal and local approaches to calculate viscosities. The nonlocal formulation is based on ABL height and friction velocity whereas the local one uses the local Richardson number and wind shear. Whichever gives a larger value for viscosity is used for further calculations in ACM2. Both schemes, however, include a critical Richardson number, the use of which is questionable in the light of recent studies (Grachev et al. 2012, manuscript submitted to Bound.-Layer Meteor.; Galperin et al. 2007). The Polar WRF can handle fractional sea ice cover (Powers and Manning 2007), but the thickness of sea ice is kept constant. The ACM2 boundary layer scheme, however, does not allow fractional sea ice coverage; in it, the sea ice concentration (SIC) is either 0% or 100%. The SIC applied in the Polar WRF was based on the ERA-Interim reanalysis (Dee et al. 2011). Because the ERA-Interim SIC in the western Weddell Sea was 85% 100%, (mean 97%, standard deviation 3%), the simulations applying ACM2 always had a 100% SIC, whereas values ranging from 85% to 100% were used in the experiments applying the MYJ scheme. The true SIC within 200 km of ISW was, however, typically 98% 99%. There was a period from early February to mid- March when the true SIC reached values as low as 85% (Geiger and Drinkwater 2005). Yet, this period does not coincide with the appearance of 85% SIC in ERA- Interim,whichhappenedattheendofAprilandin early May. 4. Simulation strategy The period for the model experiments covered three and a half months in the austral autumn and winter of 1992, starting on 20 February and ending on 4 June. The model results were validated against the near-surface observations and soundings made at Ice Station Weddell. The ERA-Interim reanalysis was used for the model initialization and lateral boundary conditions during the simulation. The boundary conditions were given by the 6-hourly analyses.

5 DECEMBER 2012 T A S T U L A E T A L FIG. 2. (a) The larger (d01) and smaller (d02) computational domains; the heavy curve is the drift track of ISW. (b) Details of the ISW track. The circles are the indicated Julian day in In (b), the locations of the first and last radiosoundings are shown. Although (a) shows the ISW drift track starting outside domain d02, the first radiosounding was within d02 in (b). Because of our emphasis on the fine structure of the atmospheric boundary layer, we used a very high vertical resolution near the surface: a total of 55 eta levels, 33 of which were below 400 m. To enable better treatment of upward propagating gravity waves, we set the highest level at 10 hpa (Guo et al. 2003). Thirty-four simulations, each 72 h long, were completed using two Lambert conformal domains with horizontal resolutions of 30 km and 10 km. The larger domain had grid points in horizontal directions, whereas the smaller one incorporated grid points. The geographical positioning of the domains is shown in Fig. 2. The model used two-way nesting between the domains. The inner domain took boundary information from the outer domain and ran four time steps for each parent time step before giving back information on the coincident interior points of the outer domain. The selected time steps were 5 and 20 s. Each simulation began at 0000 UTC, and the first output values for model validation were taken 12 h later. In late May and early June, soundings were made more often, and there were seven cases in which there was also a sounding at 0600 UTC and the first output was hence taken 6 h after the simulation started. We used longitude and latitude provided by Claffey et al. (1994) to select coordinates and times from the model output to obtain vertical profiles of temperature and wind speed corresponding to those observed in the radiosoundings. The model grid point closest to the sounding site was used for model validation. With the 10-km horizontal resolution in the inner domain, the distance between these two points never exceeded 7.1 km. The time interval between two successive model outputs was 1 h. For each sounding time, a corresponding model output could be found. The high vertical resolution in the model experiments made it unnecessary to use any interpolation for the model data in the vertical regime. 5. Results a. Near-surface variables and comparison with ERA- Interim We present model validation against the hourly observed surface-layer data from ISW for 1500 UTC 25 February 2200 UTC 29 May (Table 1). Although the correlation coefficients between the observed and modeled fluxes of sensible and latent heat at the surface are all insignificant, the biases for both quantities are small most of the time (Fig. 3, Table 1). The bias and RMSE values for H and LE did not depend much on the ABL scheme, but in the case of T s, the ACM2 scheme performed better than the MYJ scheme. If we recalculate the error statistics given in Table 1 for cases in which both the observed and modeled N is less than or equal to 0.2 ( cases depending on the number of missing values for different variables), interesting results emerge: the negative biases for ACM2 grow considerably (24.48C for T s and 26.98C for T 5m ). These changes demonstrate the problems typical for the ACM2 scheme in cold conditions: the observed average surface temperature for the cases with low cloud cover

6 3924 M O N T H L Y W E A T H E R R E V I E W VOLUME 140 TABLE 1. Comparisons of measured and modeled hourly sensible heat flux (H), latent heat flux (LE), surface temperature (T s ), 5-m temperature (T 5m ), 5-m wind speed (V 5m ), and cloud cover fraction (N) in the Polar WRF experiments employing the MYJ and ACM2 boundary layer schemes. BIAS is the average of x model 2 x obs. RMSE stands for the root-mean-square error and r for the correlation coefficient. The numbers in square brackets indicate the number of data pairs used for calculations. The validation period is 1500 UTC 25 Feb 2200 UTC 29 May. H [1153] LE [1079] T s [1835] T 5m [2013] V 5m [1944] N [2096] MYJ RMSE 17.8 W m W m C 5.48C 3.2 m s Bias 22.2 W m W m C 0.78C 20.6 m s r ACM2 RMSE 16.2 W m W m C 6.98C 3.2 m s Bias 1.3 W m W m C 22.18C 20.4 m s r fraction is C whereas the corresponding value for the entire ISW period is C. To address the influence of fractional sea ice cover on the modeled near-surface variables, we compared the results of the whole simulation period against those from a period of 100% ERA-Interim SIC (2 March 22 April). The largest effect was found in the RMSEs: when SIC was 100%, the RMSE values compared to the whole period (Table 1) decreased by 6.5 W m 22 (H), 2.9 W m 22 (LE), and 2.38C (T s ). The biases, however, were affected FIG. 3. Observed and modeled sensible and latent heat fluxes and surface temperature at ISW run through a 24-h boxcar filter. Abbreviations MYJ and ACM2 refer to the boundary layer schemes used in the Polar WRF.

7 DECEMBER 2012 TASTULA ET AL. to a lesser degree (2 denotes decrease and 1 increase): 11.0 W m22 (H), 21.0 W m22 (LE), and 20.18C (Ts). Yet, considering the high mean value of SIC (97%) and small standard deviation (3%), we cannot conclude that the larger errors in the modeled surface variables when SIC was not always 100% are exclusively due to the fractional sea ice cover. Synoptic-scale conditions are likely to affect the validation statistics as well. For instance, the strong fluctuations in the modeled surface temperatures (days in Fig. 3) fall outside the 50day 100% SIC period. We further calculated the normalized sensible heat flux H/V5m as a function of temperature difference Ts 2 T5m for both observations and model results (Fig. 4). Neither one of the boundary layer parameterizations succeed in reproducing the observations. When observed jts 2 T5mj is large (.6 K), the majority of observed H/ V5m values are found between 21 and 1 J m21, whereas the model results indicate a clear downward H/V5m. For large jts 2 T5mj values, the MYJ scheme gives a downward flux on the order of 8 13 J m21, whereas the corresponding values for the ACM2 scheme are 2 10 J m21. Fractional sea ice did not have a distinguishable effect on the values given above. When we plotted (not shown) H/V5m versus Ts 2 T5m for the 50-day time period, which features 100% sea ice for MYJ, the downward flux values for the cases of large jts 2 T5mj remained close to the values given above. Accordingly, the ACM2 scheme succeeds better in producing the observed decoupling of air and surface in conditions of very stable stratification, and its superiority over MYJ in this respect is not due to the fractional sea concentration used in MYJ. Bromwich et al. (2009) included a figure similar to our Fig. 4 in their study to evaluate the Polar WRF over the Arctic Ocean. There are some interesting similarities and differences between the two figures. First, the amount of scatter in our study is notably larger both in the model and observations. Second, the range of normalized sensible heat fluxes is greater in our case (228 to 26 kg m21 s22 in the observations and 229 to 39 kg m21 s22 in the model) than in Bromwich et al. (2009), who reported 24 to 4 kg m21 s22 in the observations and 29 to 14 kg m21 s22 in the model. However, if we compare the errors obtained for H in these two studies, the RMSEs are almost identical; the average bias of H in our case, however, is considerably smaller (22.2 W m22; MYJ) than in their study (29.6 W m22). The choice between the ABL schemes had a large effect on the simulated Ts and T5m. For Ts, the smallest bias (0.028C) was obtained with ACM2, but for T5m, with MYJ. The correlation coefficients between the observations and model results were as high as 0.7 for Ts and T5m FIG. 4. Normalized sensible heat flux as a function of Ts 2 T5m for ISW observations and model results. Abbreviations MYJ and ACM2 refer to different boundary layer schemes used in the Polar WRF.

8 3926 M O N T H L Y W E A T H E R R E V I E W VOLUME 140 TABLE 2. Comparison of the error statistics obtained from ERA-Interim (ERA-I) analyses and Polar WRF (PWRF) 72-h forecasts (with the MYJ boundary layer scheme) for the observed surface temperature (T s ), mean sea level pressure (MSLP), and 10-m wind speed (V 10m ) at sounding times. BIAS, RMSE, and r are defined for Table 1. In total, 163 data pairs were used. T s MSLP V 10m ERA-I PWRF (MYJ) ERA-I PWRF (MYJ) ERA-I PWRF (MYJ) RMSE 6.78C 6.78C 1.8 hpa 2.5 hpa 1.9 m s m s 21 Bias 4.68C 2.48C 21.2 hpa 21.7 hpa 0.9 m s m s 21 r The simulations of 5-m wind speed feature RMSEs of the order of 3 m s 21, biases around 20.5 m s 21, and correlation coefficients of 0.2. The results were not sensitive to the ABL scheme. Cloud cover fraction (N), ranging from 0 to 1, is not a standard output variable in the Polar WRF. To perform model validation for N, we use a formula given by Fogt and Bromwich (2008): N 5 0:075åCLWP 1 0:170å CIWP, (1) where CLWP and CIWP denote the cloud liquid water path and cloud ice water path (both in meters), respectively, calculated based on cloud water and cloud ice mixing ratios. The summation is from the lowest model half level to the highest. We find no agreement, however, between the observed and modeled cloudiness. We also carried out multiple regression analysis to assess the possible dependencies between near-surface model variables. We try to explain the 5-m temperature bias (BT 5m ) using the modeled 5-m temperature T 5m (8C) and the ERA-Interim SIC (0 1). The RMSE of the regression is 4.34 (8C) and r The equation takes the following form: BT 5m 5 0:44T 5m 2 23:5SIC 1 32:7. (2) Analysis of the factors affecting the bias of relative humidity yields a strong dependence between the bias in 5-m relative humidity and the observed relative humidity at the same level: BRH 5m 5 0:79RH 5m (3) In this case, the RMSE value is 7% and r It is also our goal to carry out comparisons between the surface quantities produced by ERA-Interim and the Polar WRF, though keep in mind that the former is an analysis and the latter is a 72-h downscaling run. The validation results for surface temperature, mean sea level pressure, and 10-m wind speed at sounding times are displayed in Table 2. Because no 10-m wind observations were made at ISW, we carry out the validation using the observed 5-m wind instead. The difference in wind speed correlation between Tables 1 and 2 is caused by poorly captured synoptic-scale events during time periods of no soundings. The RMSE values given by ERA-Interim are clearly superior for atmospheric pressure and close to equal for T s and V 5m compared to the ones obtained with the Polar WRF with the MYJ boundary layer scheme. However, in the case of the surface temperature and wind speed biases, the Polar WRF succeeds much better. Especially notable is the near-zero mean surface temperature bias given by the ACM2 scheme (see Table 1) compared to 4.68C by ERA-Interim. b. Low-level jets First, we analyzed how the model errors were affected by the length of the time span from the beginning of a simulation to the model output time coinciding with the sounding time. This comparison is needed because the sounding times occurred anywhere from 6 to 72 h after the start of each simulation. It turned out that the time that passed since the beginning of the model experiment did not have any statistically significant impact on the errors (same for LLJs as well as temperature and humidity inversions, addressed in sections 5c and 5d). At ISW, approximately 80% of the Tethersonde soundings featured an LLJ. Based on the simulation results obtained for the three-and-a-half-month sounding period, the Polar WRF with MYJ gives an LLJ in 42 cases out of 103 (41%). The corresponding number for the ACM2 scheme is 37 (36%). If we take into account only the cases in which both the observations and the model show an LLJ, the numbers are reduced to 31 (MYJ) and 32 (ACM2). The Polar WRF gives a jet core height approximately the same (650 m) as observed in only six (ACM2) to nine (MYJ) jet cases. To better understand if the errors in LLJs are due to errors in the wind profile higher up, we analyzed the results separately for three groups according to the error of the modeled wind speed in the m layer above sea level: cases with a relative error of wind speed (i) #5%, (ii) between 5% and 50%, and (iii) $50%. These groups consisted of cases when both the observations

9 DECEMBER 2012 T A S T U L A E T A L TABLE 3. Errors related to the simulated jet core height (z j ) and wind speed at the jet core (U j ) for the cases in which both the model and observations yield a jet. The values in parentheses refer to the errors calculated for the cases in which the relative error of the wind speed in the m layer is less than or equal to 5%. The number of data pairs for the values without parentheses is 31 (MYJ) and 32 (ACM2). The values in parentheses consist of 10 (MYJ) and 7 (ACM2) pairs. z j MYJ RMSE 78 (138) m 2.2 (4.9) m s 21 Bias 26 (26) m 0.4 (20.3) m s 21 r 20.1 (20.1) 20.2 (20.4) ACM2 RMSE 84 (236) m 2.1 (2.9) m s 21 Bias 13 (17) m 0.45 (21.3) m s 21 r 20.2 (20.3) 0.2 (0.6) U j and model results included a LLJ. Unlike what one would expect, the jet-related errors do not get smaller when we include only cases for which the wind speed in the m layer is well simulated (Table 3). It is also noteworthy that a significant number of simulated jet cases (16 17 depending on the ABL scheme) belong to the category of poorly simulated wind speed between altitudes of 300 and 500 m. Figure 5 illustrates the problems in wind simulations: the mean modeled vertical wind speed profiles exhibit a significant positive bias in the entire layer considered (0 650 m). One noteworthy aspect is to compare the distributions of jet core height, wind speed at the jet core, and wind direction at jet core in the model and observations. The histograms in Fig. 6 show that, even though both model results and observations form a frequency peak between 25 and 125 m, modeled jets rarely occur above 125 m. The modeled wind speed at the jet core is on average higher than in the observations. In particular, weak jets are almost absent in the model results. The most dominant wind directions at the cores of the observed jets were between 2108 and 2408 (Andreas et al. 2000); however, in the model results, a large portion, 15% 20%, of winds blew from the opposite direction (northeast), independent of the ABL scheme. For both ABL schemes, westerly jets are the most common ones (frequency of 25% in both cases). Accordingly, the differences between the ABL schemes do not affect the wind direction but affect the cases when LLJs occur. c. Temperature inversions For the Polar WRF, there are only 7 (MYJ) or 8 (ACM2) cases (out of 164) for which both the upper and lower boundary of the inversion layer given by the model deviate by 50 m or less from the observed values. FIG. 5. Observed and modeled mean vertical profiles of (top) temperature and (bottom) wind speed at ISW (i.e., averaged over all radiosoundings and simultaneous model profiles). (middle) The temperature profiles in the lowest 35 m. Abbreviations MYJ and ACM2 refer to the boundary layer schemes used in the Polar WRF. To get the mean observed temperature and wind profiles, we linearly interpolated the sounding-based temperature and wind observations to the 38 lowest model half levels.

10 3928 M O N T H L Y W E A T H E R R E V I E W VOLUME 140 FIG. 6. Observed and modeled distribution of (left) jet core height and (right) wind speed at the jet core at ISW. Abbreviations MYJ and ACM2 refer to the boundary layer schemes used in the Polar WRF. On average, the Polar WRF tends to yield inversions that are too thick: the bias of the inversion depth is positive for both schemes (Table 4). If we look at only the cases in which the difference between the observed and modeled cloud cover fraction is less than 0.3, the results generally get better (Table 4); however, the model s average RMSE for the height of the inversion base gets worse for the MYJ simulations, suggesting that compensating errors occur when more clouds are present. As shown on the bottom of Table 4, carrying out the corresponding analysis for ERA-Interim reanalysis yields considerably larger RMSEs and biases for the inversion base height and the inversion depth than in the case of the Polar WRF. At least a part of these errors is due to much coarser vertical resolution of ERA-Interim in the boundary layer. For temperature inversion strength (TIS), however, ERA-Interim performs better than the Polar WRF with the ACM2 scheme. The WRF-based mean temperature has a distinctive positive bias above 50 m (Fig. 5). Close to the surface (the first sounding observation was typically at the height of 5 10 m), ACM2 yields mean temperature that is too low by 2.18C whereas the value predicted by MYJ is too high by 1.08C. These results are well in agreement with the validation against the tower data, which yielded biases of and 0.78C for the 5-m temperature, respectively (Table 1). Comparing the inversions given by the two model configurations reveals two interesting features: the average inversion depth given by the MYJ scheme (496 m) is almost 100 m larger than that of the ACM2 scheme. In addition, the mean inversion strength provided by MYJ is 2.48C less than what is given by ACM2. There are 68 surface-based inversions observed in the 164 soundings, whereas the model gives 89 with MYJ and 60 with ACM2. For these 68 observed cases, the Polar WRF yielded simultaneous surface-based inversions in 50 (MYJ) and 32 (ACM2) cases. Table 5 lists correlation values between three inversion properties and different quantities possibly controlling them in the model and the observations. The modeled TIS (5T t 2 T b, t referring to top and b to bottom) shows a clear negative correlation with incoming longwave radiation (lwr) and surface temperature (T s ); but in the observations, surprisingly, such a dependence is not found. Furthermore, we checked if the lack of TABLE 4. Errors related to the simulated height of the base of the inversion (z b ), inversion depth (d 5 z t 2 z b ), and inversion strength (TIS 5 T t 2 T b ). Here, z and T refer to height and temperature, respectively, and subscripts t and b stand for inversion top and bottom, respectively. The values in parentheses refer to cases in which the difference between observed and modeled cloud cover fraction is less than 0.3. The number of data pairs for the values without parentheses is 112 (MYJ) and 111 (ACM2, ERA-Interim). The values in parentheses consist of 29 pairs. z b d TIS MYJ RMSE 149 (102) m 293 (239) m 4.68C (3.88C) Bias 235 (238) m 88 (44) m 0.48C (20.28C) r 0.4 (0.5) 0.2 (0.4) 0.4 (0.5) ACM2 RMSE 157 (157) m 273 (238) m 7.18C (4.18C) Bias 213 (212) m 49 (51) m 2.18C (1.28C) r 0.3 (0.3) 0.2 (0.1) 0.3 (0.4) ERA-Interim RMSE 172 m 404 m 4.98C Bias 295 m 284 m 22.38C r

11 DECEMBER 2012 T A S T U L A E T A L TABLE 5. Variables related to the temperature inversions in the model (PWRF) (with the MYJ scheme) and the observations: correlation coefficients between TIS (temperature difference between inversion top and base), d (inversion depth), and z b (height of the inversion base), and 1) incoming longwave radiation at the surface (lwr), 2) mean sea level pressure (mslp), 3) wind direction at the surface (wdir), 4) surface temperature (T s ), 5) temperature at 300 m (T 300m ), 6) cloud cover fraction (N), 7) relative humidity at 300 m (rh 300m ), and 8) geopotential at 850 hpa (z 850 ). The missing values are due to the fact that soundings rarely reached the 850-hPa pressure level. Significant correlations (p, 0.01) are marked in boldface. TIS d z b Obs PWRF Obs PWRF Obs PWRF lwr mslp wdir T s T 300m N rh 300m z correlation was due to the large variability in the maximum heights reached by the soundings, but the correlation between TIS and lwr calculated for the m layer was almost just as weak (20.17) as in Table 5 (20.14). The only observed quantity listed in Table 5 that shows a weak correlation (20.43) with the observed TIS is the observed relative humidity (RH) at 300 m, close to the corresponding value for the modeled TIS and RH at 300 m (20.37). In the case of inversion base height z b, a positive correlation is present for longwave radiation and T s, this time both in the model (r ) and in the observations (r ). There is correlation (r ) between z b and cloud cover in the observations but not in the model. The factors related to the TIS in the model are further examined with multiple regression analysis. Among the several explaining variables, the highest r 2 value 0.79 was obtained by an equation based on the surface temperature (T s ) and the height of the 850-hPa pressure level (z 850 ): TIS 520:67T s 1 0:123z (4) Here TIS and T s are in degrees Celsius, and z 850 is in meters. Equation (4) suggests that the colder the surface temperature and the higher the pressure, the stronger is the temperature inversion. Equally high correlation (r ) can be obtained by using lwr as the explaining variable for TIS: TIS 5 20:232lwr 1 76:2. (5) A regression based on T s alone yields (r ) FIG. 7. Observed and modeled distributions of (top) inversion base height and (bottom) inversion depth at ISW. Abbreviations MYJ and ACM2 refer to the boundary layer schemes used in the Polar WRF. TIS 520:32T s 1 1:6. (6) As with the LLJs, it is worth comparing the distribution of inversion properties in the model and observations. Figure 7 shows that the observed inversion depth is most likely between m, whereas the model (for both the MYJ and ACM2 schemes) yields a dualfrequency peak. Very thin inversions (0 50 m thick) or relatively thick inversions ( m) are most common. The distributions of the height of the inversion base behave in the same way in both the model and observations: a vast majority of the inversions are surface based (Fig. 7). The inversion strength (i.e., the temperature change through the inversion) has also quite similar distributions in the model and the observations (Fig. 8). A strength of the order of 08 58C is predominant in all cases. The MYJ scheme yields a second frequency peak for the range C, whereas the ACM2 scheme produces more than 10 cases in which the inversion strength is at least 208C. Such inversions are not found in the observations, and the MYJ scheme produces only one of them. For the temperature at the inversion base, the MYJ scheme gives a better fit to the observed temperature distribution: both the observations and MYJ show a frequency peak in the range extending from 2258 to

12 3930 M O N T H L Y W E A T H E R R E V I E W VOLUME 140 FIG. 8. Observed and modeled distributions of temperature at (left) inversion base and (right) inversion strength (T t 2 T b ) at ISW. Subscripts t and b refer to inversion top and bottom, respectively. Abbreviations MYJ and ACM2 refer to the boundary layer schemes used in the Polar WRF. 2208C (Fig. 8). The ACM2 scheme gives a quite flat distribution, with a large number of cases in which the temperature at the inversion base is lower than 2358C. Such inversions seldom appear in the observations or in the model results with the MYJ scheme. On the other hand, all model simulations produce over 20 inversions in which temperature at the base is higher than 258C. Only one such case is found in the observations. One possible reason for the low near-surface temperatures given by ACM2 is a small value for eddy diffusivity that limits the turbulent mixing (Tastula and Vihma 2011). The more complex prognostic treatment for TKE in MYJ appears to work better in stable conditions as the modeled temperatures at the inversion base are closer to the observed values. d. Specific humidity inversions The average specific humidity inversion depth given by the model (432 m: MYJ, 378 m: ACM2) is smaller than the average modeled temperature inversion depth (496 m: MYJ, 378 m: ACM2). The MYJ scheme provides nearly the same average base height (62 68 m) and top height ( m) for both temperature and specific humidity inversions, whereas the ACM2 scheme places temperature inversions up higher than specific humidity inversions. For this scheme, the average base and top heights for temperature inversion are 85 and 489 m, the corresponding numbers for specific humidity inversion being 59 and 436 m. The mean specific humidity inversion strength is similar for both schemes (0.47 g kg 21 : MYJ, 0.43 g kg 21 : ACM2). Although a thorough comparison between the observed and modeled specific humidity in the boundary layer is not possible due to many gaps in the humidity observations, it is still worthwhile to investigate what factors control the specific humidity inversions in the Polar WRF and what is the relationship between the humidity and temperature inversions. The only significant (albeit weak) correlations between specific humidity inversion properties and different meteorological variables are found between the inversion base height and (i) incoming longwave radiation (0.40), (ii) surface temperature (0.43), and (iii) air temperature at 850 hpa (0.35). The correlation between base heights in specific humidity and temperature inversions is 0.82, whereas the correlation between temperature and specific humidity at inversion base even reaches 0.92, which is related to the strong dependence of saturation specific humidity on temperature. On the other hand, the correlations between specific humidity and temperature inversion strength and depth yield weak and statistically insignificant values, r and r , respectively. e. Seasonal evolution of LLJs and inversions Because the ISW measurements extend from February all the way to the beginning of June, the influence of the season on the model results is an interesting factor to consider. Table 6 presents the occurrence and strength of LLJs and temperature and specific humidity inversions divided into three categories covering 1) February March, 2) April, and 3) May June. Both in the model results and observations, the temperature inversions in April are on average 48 58C stronger than those in the February March time span. The MYJ scheme and observations feature a decrease in the temperature inversion strength from April to the May June category, whereas the ACM2 scheme yields an increase. This is probably due to the northward drift

13 DECEMBER 2012 T A S T U L A E T A L TABLE 6. Seasonal dependence of the modeled and observed low-level jets and of the temperature and specific humidity inversions. of ISW in May June to areas of lower SIC where temperature inversions became weaker, except in the case of ACM2, which does not allow fractional sea ice. LLJs are the strongest in April; model experiments and observations agree on this. In addition, humidity inversion strength also reaches its maximum in April. The April maxima are probably due to the northward drift in ISW; in May June, the northerly location compensated the effects of the seasonal change. 6. Discussion Occurrence (%) of all soundings Strength [(m s 21 ), (8C), (kg kg 21 )] Time period MYJ ACM2 Obs MYJ ACM2 Obs Low-level jets Feb Mar Apr May Jun Temperature inversions Feb Mar Apr May Jun Specific humidity inversions Feb Mar Apr May Jun The Polar WRF underestimated the frequency of occurrence of LLJs by roughly 50%, and the modeled jet characteristics did not correlate with the observed ones. These results demonstrate that a very high vertical resolution does not guarantee successful LLJ simulations. Moreover, we found that the ABL parameterizations succeed in reproducing the LLJs only in a minority of the cases (in 38% 39% of the observed LLJs), and LLJs that were also predicted at the correct altitude were even more scarce (4% 5% of the observed LLJs). These results agree with the findings from single-column benchmark studies such as the Global Energy and Water Cycle Experiment (GEWEX) Atmospheric Boundary Layer Study 1 (GABSL1; Beare et al. 2006; Cuxart et al. 2006), which demonstrated the difficulty models have in reproducing low-level jets. Based on their LLJ simulations, Storm et al. (2009) concluded that WRF overestimates jet core height. In our model results, the jet core height bias given by the MYJ scheme was slightly negative whereas the ACM2 scheme overestimated jet core height. Moreover, Storm et al. (2009) found out that all six model configurations underestimated wind speed at the jet core. In our study, both the MYJ and ACM2 schemes also gave a slightly negative bias for the wind speed at the jet core, but only for the cases in which the wind speed in the m layer was well simulated (Table 3). Taking into account all soundings, the average wind speed bias at the jet core was slightly positive. However, when comparing our study with Storm et al. (2009), realize that Storm et al. (2009) considered only 2 LLJ cases whereas we addressed 103 cases. To address the possible connection between the relatively coarse horizontal resolution and errors in the simulated wind and LLJs, we recalculated the error statistics presented in Table 1 for cases in which the modeled wind was blowing from the sector and the modeled wind speed was.5 ms 21. Most of the cases meeting these criteria should represent a flow over the mountains of the Antarctic Peninsula, the accurate modeling of which depends on appropriate terrain resolution. Based on the new calculations, the only significant change was a slight increase in the biases of T s, T 5m, and V 5m. When we, however, calculated the same statistics for all cases of modeled wind speed exceeding 5ms 21, the bias values remained the same. Therefore, we conclude that the increase in these biases for T s, T 5m, and V 5m compared to the values shown in Table 1 is related to the increase in wind speed, not to the selection of a specific wind direction. The rather coarse horizontal resolution of 30 km used in the simulation is thus not likely the reason for the errors in wind speed and simulations of the LLJ. If resolution were the cause, winds from the orographically complex Antarctic Peninsula would result in larger errors in the modeled wind on the east side of the Peninsula. Using a grid size of 30 km in this study is also supported by Stössel et al. (2011), according to whom the ECMWF operational analyses with 25-km resolution generate realistic wind forcing on the western Weddell Sea. One interesting aspect revealed in the aforementioned analyses (wind speed.5 ms 21 ) is that the correlation between the observed and modeled N depends on the wind direction. When we included only winds from the sector , correlation stayed at 0.1 (both PBL schemes). However, when we removed the restriction on wind direction, the correlation jumped to 0.6 (MYJ) and 0.5 (ACM2). Based on this result, we suggest that when the wind speed is higher and weather is more likely to be dominated by a synoptic-scale system, the model succeeds considerably better in simulating cloud cover, but not in cases when the wind is from the peninsula. This is, however, not surprising; clouds related to synoptic-scale lows (winds.5 ms 21 ) are typically thicker than clouds that occur during weak winds or

14 3932 M O N T H L Y W E A T H E R R E V I E W VOLUME 140 during westerly flow over the peninsula, which may often be associated with föhn and formation of thin, high clouds, also related to mountain lee waves. Thin clouds are more liable to modeling errors when the validation is simply based on the cloud fraction. Moreover, modeling cloud cover in the vicinity of the Antarctic Peninsula is problematic. Microphysics schemes like Morrison et al. (2005) are likely to overestimate cloud ice concentration in this region. This is due to a strong marine influence on the aerosol composition: a large component of the aerosol present is likely to be sea salt, which does not nucleate ice (Bromwich et al. 2012). The simulations succeeded much better with the occurrence of temperature inversions: in the 164 soundings, 68 surface-based inversions were observed and, depending on the scheme, 60 to 89 were modeled. The modeled temperature inversions were, however, too thick with both the base too low and the top too high; the temperatures of both base and top were too high. The inversion thickness seems to be a problematic variable to model. As in our results, Zhang et al. (2011) obtained temperature inversions that were too thick for Antarctic coastal stations and the South Pole, whereas the WRF experiments of Kilpeläinen et al. (2012) underestimated temperature inversion thickness but overestimated humidity inversion thickness over Svalbard fjords in the Arctic. According to Zhang et al. (2011), the strength of surface-based temperature inversions in the Antarctic coastal regions is typically C, which is close to ISW observations. Zhang et al. (2011) found that the Geophysical Fluid Dynamics Laboratory (GFDL) Atmospheric Model, version 3 (AM3) and the Community Atmosphere Model, version 3 (CAM3) climate models and the 40-yr ECMWF Re-Analysis (ERA-40) underestimated the surface-based temperature inversion strength along the Weddell Sea coast; but in our experiments, the Polar WRF slightly overestimated inversion strength for both surface-based inversions and inversions in general. Vihma et al. (2011) observed that the temperature inversion strength over Svalbard fjords was mostly controlled by variables at the 850-hPa level, with relative humidity as the most important one. In our Polar WRF results, the 300-m relative humidity was the second-most important factor after downward longwave radiation (if we exclude the noncausal T s ),andinour observations 300-m relative humidity was the only variable that correlated with TIS. In the Polar WRF, TIS was related to T s and downward longwave radiation. The relationship with T s [Eq. (6)] is surprisingly similar to the one observed by Jouzel and Merlivat (1984) for the annual mean TIS and T s over the Antarctic continent: TIS T s The higher constant in our modeling study (cf. 1.6 and 21.2) may be interpreted as a result of stronger warm-air advection from the open ocean to the sea ice zone. To our knowledge, the properties of humidity inversions over the Antarctic sea ice zone have not been addressed in previous studies. In the Polar WRF, the humidity inversions were generated by a complex interaction of various processes, seen as the fact that no single variable explained the humidity inversion depth, strength, and base height. On the contrary, Vihma et al. (2011) observed over Svalbard fjords that the humidity inversion strength correlated with several variables, above all with those at the 850-hPa level. We applied two ABL schemes, but the results did not exclusively prove the superiority of one scheme over the other. The typical model errors in the lowermost 600 m were a warm bias of 28 38C and a wind bias of 1 2 m s 21. These were not sensitive to the ABL scheme, except for the lowermost tens of meters (Fig. 5). The two schemes were equally unsuccessful when predicting the occurrence of LLJs. On average, the ACM2 scheme placed jet cores higher than the MYJ scheme. The surface and 5-m air temperatures were sensitive to theablschemes;theacm2schemeresultedinthe smallest bias for T s, and MYJ had the smallest bias for T 5m. ACM2 provided less mixing in the near-surface air than MYJ. The small bias of the ACM2 scheme was, however, associated with cases of far too low and high T s. 7. Conclusions We applied the Polar WRF model to simulate the period of Ice Station Weddell in February to June 1992 the longest period of ABL observations over the Antarctic sea ice zone. The model experiments consisted of 72-h-long downscaling runs, with initial and boundary conditions from the ERA-Interim reanalysis. With respect to the surface pressure, ERA-Interim was more accurate than the Polar WRF; this is partly due to assimilation of atmospheric pressure data from drifting buoys nearby ISW (Vihma et al. 1996), which was not done for the Polar WRF. Moreover, the success of ERA-Interim with respect to surface pressure is not surprising; in assessment studies of different reanalysis products in the Antarctic by Bromwich et al. (2011) and Nicolas and Bromwich (2011), ERA-Interim was found to give the most realistic depiction for surface pressure of the six reanalyses considered. For the surface temperature, the Polar WRF gave, however, clearly smaller biases, ranging from to 2.48C, depending on the ABL scheme, compared to the 4.68C for ERA-Interim. Errors in the modeled surface and 5-m temperatures as well as turbulent fluxes were, in part, related to the poorly simulated cloud cover. In polar regions, this is