This paper describes the performance of the Community Atmosphere Model (CAM) versions 4 and 5 in simulating near-surface parameters. CAM is the atmospheric component of the Community Earth System Model (CESM). Most of the parameterizations in the two versions are substantially different, and that is also true for the boundary layer scheme: CAM4 employs a nonlocal K-profile scheme, whereas CAM5 uses a turbulent kinetic energy (TKE) scheme. The evaluation focuses on the diurnal cycle and global observational and reanalysis datasets are used together with multiyear observations from 35 flux tower sites, providing high-frequency measurements in a range of different climate zones. It is found that both model versions capture the timing of the diurnal cycle but considerably overestimate the diurnal amplitude of net radiation, temperature, wind, and turbulent heat fluxes. The seasonal temperature range at mid- and high latitudes is also overestimated with too warm summer temperatures and too cold winter temperatures. The diagnosed boundary layer is deeper in CAM5 over ocean in regions with low-level marine clouds as a result of the turbulence generated by cloud-top cooling. Elsewhere, the boundary layer is in general shallower in CAM5. The two model versions differ substantially in their representation of near-surface wind speeds over land. The low-level wind speed in CAM5 is about half as strong as in CAM4, and the difference is even larger in areas where the subgrid-scale terrain is significant. The reason is the turbulent mountain stress parameterization, only applied in CAM5, which acts to increase the surface stress and thereby reduce the wind speed.
To understand and predict future climate, we rely on using numerical models of the climate system. Previous generations of global climate models (GCMs) only described the physical–dynamical climate system; whereas the current generation, Earth system models (ESMs), expand to interactively couple key biogeochemical cycles, such as the carbon and nitrogen cycles, into the physical–dynamical GCMs. However, the reliability of simulations of present and future climate is dependent not only on the complexity of the model system but also, for example, on the performance of the physical parameterization schemes used. As the ESMs become more complex by adding components such as emissions of aerosols and biogeochemical cycles (including dynamic vegetation), biases due to deficiencies in the physical–dynamical part of the model may be enhanced since the coupling of the additional components allow for many more feedback mechanisms. The coupling between the physical and the biogeochemical systems are mainly through near-surface variables. Thus, it is of interest to evaluate them in order to assess biases and possible model deficiencies.
Few studies that evaluate the performance of planetary boundary layer (PBL) parameters in GCMs are found in the literature. Some very early studies include Boer et al. (1991) and Randall et al. (1992). At the time of these intercomparisons, most models did not resolve the diurnal variation in solar insolation and only had a few vertical grid levels in the entire PBL (Garratt 1993). As the computational resources were increasing more attention was given to the PBL and land surface parameterizations and some substantial developments were made during the 1990s (e.g., Holtslag and Boville 1993; Martin et al. 2000), but some fundamental problems still remain: for example, difficulties modeling the stably stratified boundary layer (Holtslag 2006; Cuxart et al. 2006; Svensson et al. 2011).
Commonly, evaluations of GCMs focus on mean fields at monthly, seasonal, and annual time scales. As pointed out by Lin et al. (2000), there are apparent limitations to this approach. It is possible for GCMs to produce realistic climate states for the wrong reasons: for instance, the mean near-surface temperature could be simulated correctly without any diurnal variation. Most evaluations of the diurnal cycle in global models consider precipitation (e.g., Lin et al. 2000; Betts and Jakob 2002) and they usually contain some of the turbulent parameters: for example, the surface heat fluxes. A more comprehensive study from the PBL point of view is reported in Garratt et al. (2002), where they compared 5 yr of GCM results with detailed boundary layer observations at six locations, two over the ocean and four over land. The observational data was limited to hourly data for a month or two at each site. They found overall good agreement between the observations and the model except for some unrealistic model mixed-layer temperature profiles over land in clear skies, which they related to the use of a simple local first-order turbulence closure. The diurnal cycle in an earlier version of the National Center for Atmospheric Research (NCAR) Community Climate System Model, version 2 (CCSM2) has been evaluated by Dai and Trenberth (2004). They found that the diurnal amplitude of surface air temperature over the continents was in good agreement with surface synoptic observations (SYNOP) station data but that the amplitude was too small over the ocean.
More recently, the Global Energy and Water Cycle Experiment (GEWEX) Atmospheric Boundary Layer Study (GABLS) has coordinated several model intercomparison studies with focus on PBL parameterizations in numerical weather prediction and climate models. The first intercomparison clearly illuminated problems in simulating weakly stably stratified conditions in terms of too deep boundary layers with not enough wind turning compared to large-eddy simulation results (Cuxart et al. 2006; Svensson and Holtslag 2009). The second GABLS study (Svensson et al. 2011) concluded that many models have difficulties representing the diurnal variation in the wind speed and that no significant difference in performance could be seen based on the type of closure. First-order models did not perform worse than turbulent kinetic energy (TKE) types of closures.
The most recent release of the NCAR Community Earth System Model version 1 (CESM1) contains the Community Atmosphere Model version 5 (CAM5; Neale et al. 2010b), but CESM may also be run with CAM4 (Neale et al. 2010a). Both versions are operational and both are used in the Coupled Model Intercomparison Project phase 5 (CMIP5) experiments. CAM4 and CAM5 contain substantially different parameterizations of the PBL. This gives an opportunity to study two very different atmospheric model versions that both use the same land model and therefore to compare two fundamentally different PBL schemes in the same modeling framework.
In this paper, we examine the performance of the PBL parameterizations in CESM, by evaluating a number of near-surface parameters, including turbulent fluxes of heat and momentum. Here, 5-yr simulations with hourly output of key variables for the two versions of CAM have been performed and results are compared with each other, with global reanalysis datasets and high-frequency data from a network of observational sites, and 35 flux-tower sites with multiyear observations in different climate zones have been selected for the comparison. The focus in the present paper is on the diurnal cycle of the near-surface parameters.
a. Observations from flux tower sites
For turbulent fluxes there are no global observational datasets, the ones that exist are reanalysis products. However, the 1990s saw an increase in the establishment of flux tower sites. FLUXNET (Baldocchi et al. 2001) is a network consisting of over 500 sites of which many are providing long-term measurements of turbulent fluxes with high-frequency data.
The sites in this study are chosen from various parts of the world, covering a range of different climate zones and ecosystems (see Table 1 for details and Fig. 1 for locations). When selecting the sites, the aim was to find horizontally homogenous sites with long-term measurements and avoid sites located too close to the ocean, since this study concerns boundary layers over land. Most flux tower sites are temperate and situated at the midlatitudes in Europe or North America, whereas tropical, Arctic, and arid sites are more scarce and consequently such sites that fulfill our requirements have been harder to find. In total, 35 micrometeorological flux tower sites provide the observational data used in this study. All sites included in the study have at least 2 yr and up to 16 yr of eddy-correlation measurements.
Seven sites are located in the tropics, and four of them are tropical rain forest sites: Santarem, Tapajos, Palangkaraya, and Sakaerat. Virginia Park and Howard Springs are savanna sites with open eucalyptus forests and wet and dry seasons, and Tchizalamou is a humid grassland site. There are 17 midlatitude sites in our study out of which 9 are covered with forest. The other eight consists of seven grassland sites and one cropland site. In the polar areas we have included three Alaskan tundra sites, Barrow, Atqasuk, and Ivotuk, as well as one Finnish wetland site, Kaamanen, which is without permafrost. Hyytiälä, Sodankylä, Boreas, and Southern Khentei Taiga are all boreal forest sites. Even though Southern Khentei Taiga is in the midlatitudes, it is located on a high altitude and therefore qualifies as a boreal climate site (Li et al. 2005a). There are three hot, semiarid sites used in our study. One is Maun, a semiarid shrubland site in Botswana. On the Northern Hemisphere we have included two semi-arid sites: Audubon Grasslands and Santa Rita Mesquite both located in Arizona.
At the flux tower sites, turbulent fluxes are measured in 30- or 60-min intervals using the eddy-covariance technique (e.g., Aubinet et al. 1999; Baldocchi et al. 1988). In addition to the fluxes, the sites maintain measurements of standard meteorological variables, such as temperature, wind speed, and pressure.
The datasets are provided by CarboEurope (http://www.carboeurope.org), Fluxnet Canada (http://fluxnet.ccrp.ec.gc.ca), AsiaFlux (http://asiaflux.net), CarboAfrica (http://www.carboafrica.net), Ozflux (http://www.cmar.csiro.au/ozflux/), and Ameriflux (http://public.ornl.gov/ameriflux), which are all part of FLUXNET, as well as by the Coordinated Energy and Water Cycle Observation Project (CEOP) archived by the NCAR Earth Observing Laboratory (EOL; http://data.eol.ucar.edu).
None of the data used for this study was gap filled. Quality control and flux corrections of the data were left to the individual principal investigators (PIs) that supplied the data. We have chosen not to omit any data, except for what has already been done by the individual working groups. All of the data included are most likely not entirely reliable, but the way we have chosen to present the data, using median values and percentiles, means that outliers will have a minor influence on our results. The level-2 datasets were used for the Ameriflux and CarboEurope sites. We are aware of that there are data of higher levels that have gone through a more rigorous quality control. However, the level-4 data are a gap-filled product and, since we are not in need of a continuous series of measurements, it is preferable to use unmanipulated data. The level-3 data contain the same values as level 2 but have quality flags assigned to some of the variables (friction velocity and radiation are the only flagged variables that are examined in this study). There is virtually no difference in the median values that we are presenting in this study between using all data and excluding values that have been flagged as unreliable. Therefore, we have chosen to prioritize using the longer series of measurements provided by the level-2 data, and we trust that we have not lost any data quality in doing so.
In this study, we examine CAM4 (Neale et al. 2010a) and CAM5 (Neale et al. 2010b), two versions of the atmospheric component of CESM1. CAM5 has been substantially modified as compared to CAM4 and contains a range of new parameterizations. Improvements include updated schemes for cloud microphysics, radiative transfer, macrophysics, aerosol formations, ice clouds, and shallow convection and a new moist turbulence parameterization. In our study, the focus is on the boundary layer, so, even though all of these updates play important roles, we expect the turbulence parameterization scheme to be of particular significance. The two different schemes used in CAM4 and CAM5 are therefore briefly described below.
Both CAM4 and CAM5 are using the same land model: Community Land Model version 4 (CLM4; Oleson et al. 2010; Lawrence et al. 2011). The surface fluxes are calculated by the land model using Monin–Obukhov similarity theory with the stability functions from Zeng et al. (1998). They are similar to the ones in CAM4 (Holtslag and Boville 1993) but differ substantially from those in CAM5 (Bretherton and Park 2009).
The PBL schemes in CAM4 and CAM5 have both been tested using the first and second GABLS case. The GABLS1 case is an idealized, weakly stable case without moisture and radiative cooling (Cuxart et al. 2006). The University of Washington (UW) scheme, used in CAM5, performs quite well, whereas the Holtslag and Boville (HB) scheme, from CAM4, overmixes at both high and low resolution (Bretherton and Park 2009).
In GABLS2, the participating single column models were forced with a prescribed surface temperature, constant geostrophic wind, and large-scale divergence (Svensson et al. 2011). The performance of the two CAM schemes is very similar and both simulate the largest diurnal amplitude in near-surface air temperature, regardless of version. When it comes to sensible heat flux, they both underestimate the daytime flux but exhibit too much negative flux at nighttime. Both schemes exhibit near-surface wind speeds that are higher than those in the other models and with rather strange diurnal cycles that show a sudden drop to very low wind speeds in the middle of the day, something that is not seen in the GCM simulations presented in this study (Svensson et al. 2011; Svensson and Lindvall 2012).
1) University of Washington scheme (CAM5)
The UW scheme (Bretherton and Park 2009; Park and Bretherton 2009) in CAM5 is a first-order moist turbulence scheme in which the diagnostic TKE is explicitly calculated. It was developed with focus on improving the representation of stratocumulus-topped boundary layers (Grenier and Bretherton 2001). In the UW scheme, convective layers (CLs) and stably stratified turbulent layers (STLs) are diagnosed by calculating the bulk moist Richardson number Ri using moist-conserved variables. Interfaces with Ri > 0.19 are considered nonturbulent. Interfaces with Ri < 0.19 are turbulent and convective when Ri < 0. When several stably stratified turbulent or convective interfaces are adjacent, they constitute a stably stratified turbulent layer or a convective layer, respectively. The scheme allows for several decoupled turbulent layers in an atmospheric column, and all turbulent layers are treated in the same manner independent of height. The stability functions, however, do not match the surface layer theory used by the land model (Oleson et al. 2010).
In each turbulent layer, a downgradient mixing scheme is applied, in which the eddy diffusivities are calculated using the diagnosed TKE. TKE storage is neglected in all layers and so is the TKE transport in STLs. The external interfaces at the top and bottom of a CL are entrainment interfaces, where an explicit entrainment closure is used to calculate an “entrainment diffusivity.”
The planetary boundary layer height (PBLH) in CAM5 is defined using discrete levels. If a surface-based convective layer exists, the top interface of that layer is defined as the boundary layer height. If no such layer exists, PBLH is set as the model level height just above the highest interface of the surface-based stably stratified turbulent layer. The surface interface is always considered to be turbulent, so, in cases where the lowest model level is considered nonturbulent (Ri > 0.19), PBLH is set to the height of the lowest model level, giving a minimum boundary layer height of around 60 m.
2) Holtslag and Boville scheme (CAM4)
CAM4 employs the HB boundary layer scheme based on Holtslag and Boville (1993), but with an updated formulation of the boundary layer height from Vogelezang and Holtslag (1996). It is a nonlocal diffusivity K-profile scheme developed for dry convective conditions. The PBL height is explicitly calculated using a bulk dry Richardson number and the height, together with a turbulent velocity scale, determines the eddy diffusivity profile. Consequently, the HB scheme assumes that boundary layer turbulence is forced from the surface. It also accounts for nonlocal transport of heat, specific humidity, and scalars using a countergradient term for the convective PBL (Holtslag and Boville 1993). Above the PBL a local first-order closure scheme is being used. The stability functions correspond to those employed in the land model. As opposed to the Bretherton and Park scheme, there is no cutoff Richardson number and thus there always exists some background turbulence.
The PBLH is defined using bulk Richardson methods based on Troen and Mahrt (1986), where the PBLH is determined iteratively as the height at which the bulk Richardson number exceeds the critical value, but with an updated formulation from Vogelezang and Holtslag (1996), where shear in the outer boundary layer as well as surface friction is taken into account. The minimum PBLH is defined 50 m above the top interface of the lowest model level, resulting in approximately 180 m.
c. Experimental setup
For the presented simulations, both atmospheric model physics packages (CAM4 and CAM5) were run with a prognostic land model (CLM4; Lawrence et al. 2011) and with climatological sea surface temperature and sea ice boundary conditions based on the Hurrell et al. (2008) dataset averaged over the years 1982–2001. We used 5 yr of model simulations with hourly output and a horizontal resolution of 0.9° latitude × 1.25° longitude. There are 26 vertical levels in CAM4 and 30 in CAM5. The lowest levels are located at ~60, 250, 600, 1200, and 2000 m in CAM4 and at ~60, 200, 370, 550, 770, 1000, 1250, 1650, and 2250 m in CAM5. Thus, the four extra model levels are all located in the lower troposphere, almost doubling the vertical resolution below 700 hPa. These extra levels improve the performance of the UW scheme described above. Notice that using the same four extra levels in CAM4 would degrade the simulations because of biases that arise from the shallow convection formulation with increased vertical resolution (Williamson 2012).
It should be noted that in CAM5 there is a turbulent mountain stress (TMS) parameterization to take into account the effect of unresolved orography. As the effects of topography are not fully captured at the model resolution, the unresolved orography is parameterized as a turbulent surface drag using an effective roughness length z0oro derived from the standard deviation of orography in the grid cell (Richter et al. 2010). TMS is applied using a function of the Richardson number f(Ri): f(Ri) = 1 if Ri < 0, f(Ri) = 0 if Ri > 1, and f(Ri) = 1 − Ri if 0 ≤ Ri ≤ 1. It is only employed in the atmospheric component and not in the land model, where the vegetation roughness length is used instead. The momentum transfer from the atmosphere is therefore directly affected, whereas the heat and moisture transfer as well as the surface stress calculated by the land model only feels the impact of the reduced wind speeds but not of the larger roughness lengths.
TMS was added to remove enough momentum from the atmosphere in order to improve the general circulation. The effect is profound on the surface pressure fields and the Icelandic and Aleutian low pressure patterns are, for instance, substantially improved. The TMS has a significant impact on near-surface variables and in particular the near-surface wind speed and friction velocity. We suspect that the effect of TMS might overshadow some of the other differences between the two model versions, and therefore a second simulation with CAM5 with TMS turned off (CAM5noTMS) has been included in our study.
The comparison of the models to the flux observations is done by choosing the closest model land grid point for each study site. The observational data differ from site to site in both which years and the number of years that are included. Therefore, when examining the diurnal cycle, the comparison to observations is done using “climatological” diurnal cycles instead of comparing specific years. Hourly model output from the 5-yr period is used to derive median diurnal cycles for each month and season, treating the data as climatological data and thus providing 12 monthly median diurnal cycles and four seasonal median diurnal cycles from the model. The observed 2-m temperature and moisture are computed from the temperature and moisture at measurement height using Monin–Obukhov similarity theory with the same stability functions as in CLM4 for stable and unstable conditions (Oleson et al. 2010).
The model simulations are also evaluated against global observation and reanalysis datasets. Two datasets for near-surface temperature are included in this study: the Willmott and Matsuura (2001) dataset version 3.02 for 1950–99 and Climate Research Unit (CRU) high-resolution climate data, version 2.1 (Mitchell and Jones 2005) for the years 1961–90. Three reanalysis datasets are used. The European Centre for Medium-Range Weather Forecasts (ECMWF) Interim Re-Analysis (ERA-Interim; Simmons et al. 2007) for 1989–2010 has the horizontal resolution T255 and 60 vertical levels. NCEP/Department of Energy Global Reanalysis 2 (NCEP-2; Kanamitsu et al. 2002) covers the years 1979–2010 and uses a T62 grid with 28 levels. The 25-yr Japanese Reanalysis (JRA25; Onogi et al. 2007) covers the years 1979–2004, and the model has a spectral resolution of T106 with 40 vertical levels. The lowest model level is at ~10 m in ERA-Interim and ~40 m in NCEP-2 and JRA25. There are no reliable global observations of turbulent heat fluxes, so as a reference we only show reanalysis datasets for these variables. Boundary layer height is only available in the ERA-Interim dataset.
The 2-m temperature pattern over land in CAM4 and CAM5 (not shown) is rather similar in the two models with an annual-mean warm bias over Europe and the United States and a cold bias over Africa, southern Asia, and South America. CAM5 has a warmer annual mean than CAM4, but CAM5 is colder at high latitudes while being warmer than CAM4 in the subtropics.
Figures 2a–c shows the annual-mean wind speed at 10 m in the two versions of CAM and, although the wind speed in a reanalysis contain biases and should not be taken as the truth, we have chosen to include the ERA-Interim dataset as a reference. The differences between the models and between model and reanalysis are also shown (Figs. 2d–f). Considering the large changes in model physics between CAM4 and CAM5, the wind speeds over the ocean are surprisingly similar and in both models higher than in ERA-Interim. Over land, however, the models differ. The spatial variability is low in CAM5 and the wind speed is generally between 1 and 4 m s−1, with lower wind speeds than ERA-Interim in areas where the wind speed is relatively high. CAM4 simulates a higher wind speed than both CAM5 and ERA-Interim, especially in mountain regions. This can likely be explained by the fact that CAM5 and ERA-Interim employ orographic drag parameterizations. The TMS, used in CAM5, is briefly described in section 2c. ERA-Interim has a subgrid-scale orographic drag scheme, which handles gravity wave drag as well as blocked flow drag (Lott and Miller 1997). Subgrid horizontal scales less than 5000 m are accounted for in the turbulent orographic form drag scheme, where the stress is parameterized for sinusoidal hills and the vertical distribution is dependent on the horizontal scale of the topography (Beljaars et al. 2004). Although different, both the TMS in CAM5 and the parameterizations in ERA-Interim act to increase the drag and thus reduce the wind speed, particularly in mountain regions.
The PBLH in ERA-interim is based on Troen and Mahrt (1986) and just as in CAM4 defined using bulk Richardson methods. Although not the same, the methods are still similar. CAM5, on the other hand, defines PBLH differently, using discrete model levels. Over the oceans, both models underestimate the PBLH compared to ERA-Interim. Over land, the spatial variation of the annual-mean PBLH, shown in Fig. 3, is much larger in the ERA-Interim reanalysis than in either of the CAM versions.
CAM5 simulates in general shallower boundary layers than CAM4 with a difference of over 200 m in many areas (Figs. 3a,b,f). The notable exception is the stratocumulus regions, where CAM5 has PBLHs that are 300–400 m deeper than in CAM4 and in much better agreement with ERA-Interim (Figs. 3d,e). This is a result of the ability of the UW turbulence scheme in CAM5 to account for turbulence caused by cloud-top radiative cooling, which directly influences the entrainment rate of dry air at the top of the cloud (Bretherton and Park 2009), a feature lacking in CAM4 (Holtslag and Boville 1993).
a. The effect of TMS on 10-m wind speed, friction velocity, and PBLH
As mentioned in section 2c, the TMS parameterization in CAM5 calculates an additional surface stress used in the atmospheric component. Figures 4a,b show the effect of the TMS parameterization in CAM5 on the 10-m wind speed and the PBLH. Figure 4c shows the orographic roughness length z0oro used in the TMS parameterization. The z0oro is rarely less than 1 m and is sometimes as high as 40 m. These values should be compared with the vegetation roughness length in the land model, which varies between 0.06 and 2.6 m.
The near-surface wind speeds are 2 m s−1 higher in CAM5noTMS than in CAM5 over vast land areas, with even larger model differences in regions with a high z0oro. It is apparent that the effect of TMS on wind speed is enhanced at higher latitudes with more frequently stably stratified boundary layers.
The largest model differences in PBLH are found in the Northern Hemisphere in the dry subtropical and tropical regions. Here, where the turbulence is mainly buoyancy driven, the larger roughness length in CAM5 increases the mixing and the PBLH is several hundred meters higher than in CAM5noTMS. At higher latitudes, where the boundary layer is more stable and the turbulence is often shear driven, the TMS parameterization has two opposing effects. A larger roughness length enhances mixing, while a lower wind speed reduces mixing, making the net effect on PBLH relatively small. The impact of TMS is small in forest regions and mostly close to zero over ocean, where the TMS parameterization is not employed.
The effect of TMS on the wind speed and friction velocity is clearly seen by looking at an example from the Atmospheric Radiation Measurement Program (ARM) Southern Great Plains (SGP) site in Oklahoma. Figure 5 shows a scatterplot of the wind speed at the lowest model level versus the friction velocity. Using the logarithmic wind law, theoretical lines for the relation between wind speed and friction velocity in near-neutral conditions are also drawn in the figure. For neutral conditions, the surface layer wind profile is given by
where U is the mean wind speed, u* is the friction velocity, k is the von Kármán constant, z is the height, and z0 is the roughness length.
While the observational, CAM4, and CAM5noTMS data follow the theoretical lines derived from the neutral drag law above and the vegetation roughness length from the land model (where the spread is due to different stability conditions), the CAM5 data do not. Instead, part of it clusters along the theoretical line calculated using the orographic roughness length z0oro and part of it, those cases where Ri > 1 and TMS is not applied, lies in the range of CAM4 data. There are also data in between, since the drag coefficient in the TMS parameterization is a linear function of Ri when 0 < Ri < 1. The observed data are from a rather flat region; still the standard deviation of subgrid-scale elevation is 18.5 m, translating to a z0oro of 1.4 m. This is a large value considering that the z0 due to vegetation, used in the land model and in CAM4, at this site is only 0.06 m. It is clear from Fig. 5 that the impact is substantial.
b. Seasonal variations in three climate zones
Figure 6 shows the seasonal variability over land of 2-m temperature, 10-m wind speed, sensible heat flux, latent heat flux, and net radiation from CAM4, CAM5, CAM5noTMS, observational, and reanalysis datasets. Tropical rain forest areas in Asia and South America, European and North American midlatitudes, and polar areas north of 60° are shown separately (see Fig. 1). The regions are chosen to cover the three climate zones in which most of the eddy-correlation sites are located.
The reanalysis datasets are not in agreement with each other or with the observational datasets but still show similar features. In most instances, the three versions of CAM are closer to each other than to the observations and reanalysis products, except for the wind speed.
In the polar and midlatitude regions (Figs. 6a,b), CAM5 underpredicts the wintertime temperatures up to 4 K compared to the CRU dataset (Mitchell and Jones 2005), whereas CAM4 near-surface temperature is closer to observations. The difference between CAM4 and CAM5 is as large as 4 K in December in the polar regions. The reanalysis datasets, on the other hand, generally display a warm bias in winter. In summer, the models overestimate the temperatures in the polar and midlatitude regions, indicating a too large annual cycle (Figs. 6a,b). CAM5noTMS simulates slightly higher temperatures than CAM5 throughout the year in the polar regions. A possible cause could be a larger poleward energy transport as a response to lower friction. In the tropical rain forests, the models are in general agreement with each other and just slightly below what is observed, approximately 0.5° lower than the Willmott and Matsuura (2001) dataset (Fig. 6c).
The wind speed reveals large differences both between the models and between the reanalysis datasets (Figs. 6d–f). The ERA-Interim wind speed is at times 3 m s−1 (300%) higher than the near-surface wind speed in JRA25, and they do not agree on the annual cycle in the midlatitudes. CAM4 overestimates the wind speed in the midlatitude and polar regions compared to all three reanalyses. CAM5 has values similar to NCEP-2, which in magnitude is between ERA-Interim and JRA25. In the tropics, all CAM versions lie within the range of the reanalyses, but with smaller amplitudes in the seasonal cycle (Fig. 6f). The decrease in wind speed due to the TMS is apparent in all regions, and CAM4 and CAM5noTMS match each other quite well. The shape of the annual cycle in CAM4 agrees fairly well with the reanalyses in the polar and tropical regions but only with ERA-Interim in the midlatitudes (Figs. 6d,e). The inclusion of TMS lessens the seasonal variation in all regions and in CAM5 the seasonal cycle is almost completely absent. A likely cause is that, in shallower boundary layers, the effect of having an increased surface friction will be more strongly felt by the near-surface wind. The largest reduction in near-surface wind speed in the polar regions and the midlatitudes will thus be seen in the cold season, when the wind speeds otherwise are higher. In the tropics, however, the highest wind speeds co-occur with the deepest PBLHs, rendering this explanation invalid.
The spread in surface heat fluxes between the reanalyses is large, whereas the three versions of CAM produce fairly similar results (Figs. 6g–l). During the polar winter, the models are all roughly within the range of the reanalyses, but with a more negative sensible heat flux in CAM5noTMS than in CAM5 and even more negative in CAM4 (Fig. 6g). In the midlatitudes the annual mean is similar in all three model versions, but CAM4 has a larger seasonal cycle (Fig. 6h). In terms of the latent heat flux, all three versions of CAM are similar in the midlatitude and polar regions, except summer when CAM4 exhibits larger fluxes (Figs. 6j–l). There are significant differences between the three reanalysis datasets in the tropical regions, whereas the three CAM versions, in comparison, are quite close to each other. All models underestimate the seasonal variations compared to the reanalyses.
When it comes to surface net radiation, all CAM versions are in closer agreement with each other than with the reanalysis datasets (Figs. 6m–o). Most of the year in the polar and midlatitude sites, the net radiation (downward) is smaller in CAM4 than in CAM5, and this difference is due to the longwave component (not shown). The opposite is seen in July at the polar sites and from May to September in the midlatitudes, caused by a larger net shortwave radiation in CAM4. The difference between ERA-Interim and the CAM models in the polar spring stems mainly from the shortwave component. In the tropics, both the net outgoing longwave and the net incoming shortwave is larger in CAM4 than in CAM5, with the shortwave dominating, giving a larger downward net radiation in CAM4 throughout the year. Adding the turbulent heat fluxes together with the net radiation to calculate a residual (that includes ground heat flux) shows very small differences between the models and generally the model differences in net radiation are compensated by differences in the turbulent heat fluxes.
c. Flux sites
The median diurnal cycles (see section 2c) are used to derive the Taylor diagram statistics (Taylor 2001) in Fig. 7. The diagram shows the performance of the model versions at all flux sites included in the study for six near-surface parameters. Since median diurnal cycles are used, the standard deviation is a measure of the diurnal amplitude and the correlation reflects the timing of the modeled diurnal cycle compared to the observations. Also shown is the mean root-mean-square error (RMSE) and bias for all flux sites.
It is apparent that all model versions simulate too large diurnal cycles, especially when it comes to wind speed, where the standard deviation is approximately 2.5 times larger in the models than what is observed. The lowest model level is at ~60 m in both models and it could be expected that the diurnal cycle would be damped by having such a thick lowest model layer. However, the cause is likely the overestimated diurnal amplitude of net radiation driving the diurnal cycle. The correlation is high for the surface heat fluxes, temperature and net radiation (above 0.90), slightly worse for friction velocity (0.81–0.83) and low for wind speed (0.59–0.64). The RMSE in temperature (3°–4°C), turbulent heat fluxes (17–23 W m−2), and net radiation (40–48 W m−2) are similar for all model versions. The bias is small in temperature (0.3–0.4 K), sensible heat flux (between −1.8 and −0.5 W m−2) and net radiation (−2 to 1.1 W m−2), but all model versions substantially overestimate the latent heat flux with biases in the order of 10 W m−2. The model differences in bias and RMSE in temperature and surface heat fluxes are minor, but there are large differences between the model versions in wind speed and friction velocity due to the TMS parameterization. CAM4 and CAM5noTMS show a 5 times larger bias in wind speed than CAM5, but their results for the friction velocity are much closer to what is observed by the local measurements, while the friction velocity in CAM5 is much larger because of the use of an orographic roughness length.
For individual sites (not shown), the site-to-site spread in annual-mean latent heat flux, sensible heat flux, and temperature is similar between the models and relatively consistent with observations. This is not the case for wind speed and friction velocity in which the site-to-site variation is too low in CAM4 and even slightly lower in CAM5. The higher (lower) wind speeds and lower (higher) friction velocity that are observed at the unforested (forested) sites are only barely seen in the models. The models tend to simulate the forested and the unforested sites similarly. This can partly be attributed to the fact that we have point observations, whereas the grid boxes represent larger areas in which several vegetation types typically are included. The models will therefore have a larger roughness length than the observations at the unforested sites and vice versa at the forest sites. However, the feature is stronger in CAM5 than in CAM4 and CAM5noTMS because of TMS. The roughness length depends only on the vegetation type in CAM4 and CAM5noTMS. The inclusion of the TMS parameterization in CAM5 implies an increased roughness length, due to orography, is added in the model. By doing so, the differences due to vegetation will decrease.
Figure 8 shows the diurnal cycles of temperature, turbulent heat fluxes, 10-m wind speed, and friction velocity for groups of sites, divided by climate zone and by vegetation. Although the sites included in each group share many similarities, the mean wind speeds and the mean temperatures differ. For a fair comparison of the diurnal cycles, the temperature and the wind speed are shown with the observational daily mean at each site removed. The observational daily mean is chosen instead of the modeled daily mean in order to have the possibility of examining the model observation bias. Note that, in cases where the diurnal range is large, the scale is increased around zero for the surface heat fluxes to allow model and observation differences to be visible even when the fluxes are small. For net surface radiation, the observations are shown in its upper panel and the model results are shown in its lower panel as a difference to the observed diurnal cycle. CAM5noTMS is omitted in this section, since the effect of TMS has already been discussed and here we focus on differences between the two released model versions.
Figures 8a–f display several common features. While the timing of the diurnal cycles is captured rather well, the diurnal amplitude of the parameters is in most instances larger in the models than in the observations, consistent with the findings in Fig. 7. This is true for all regions but especially for the forest sites (Figs. 8b,d,e). Since the near-surface variables are coupled to each other, a large amplitude in one variable is likely to be reflected in the others.
The summer daytime temperatures are substantially overestimated in all regions and the model observation difference is around 5°C at the midlatitude, boreal forest, and arid sites. Both models are too cold in winter at the polar sites (Figs. 8c,d). CAM4 and CAM5 show a cold bias of 4° and 5°C, respectively, at the Arctic tundra and wetland sites. At the boreal forest sites, the modeled diurnal wintertime amplitude is far too large, resulting in a 7°C cold bias during the night in CAM5. At the midlatitude and polar sites (Figs. 8a–d), CAM5 is colder than CAM4 in winter with a model difference going up to 4°C during the boreal forest night. The underestimated temperatures at the high-latitude wintertime, which seem to be especially pronounced at nighttime, are connected with too much radiative cooling. Unfortunately, many of the sites lack data of the radiative flux components, and thus it is hard to attribute the cause of this excess cooling. However, since the simulated temperatures are lower than those observed and the shortwave component is small, it is likely that it is due to too little incoming longwave radiation. This cannot explain the differences between the models since CAM4 show a more negative bias in the net radiation. It seems as if in CAM4 it is partly compensated by a larger negative bias in the sensible heat flux, whereas in CAM5 the sensible heat flux bias is smaller. Considering the lower near-surface wind speeds in CAM5, larger temperature gradients could still be needed to sustain this amount of heat flux. In CAM5 the UW scheme also shuts off turbulence when the Richardson number exceeds 0.19 and thus prevents downward mixing of warm air. At the tropical sites (Fig. 8f), CAM4 and CAM5 are more similar, both simulating slightly too warm day temperatures and 2°–3°C too cold nights.
In terms of sensible heat flux, the difference between the two models during the night is persistent in all climate zones. The turbulence is shear driven when the boundary layer is stably stratified and the overestimated winds in CAM4 results in too large fluxes, whereas CAM5 matches the observations better. That turbulence is shut off in CAM5 when Ri > 0.19 could also contribute to the less negative sensible heat flux in CAM5. However, since the surface energy budget is required to be in balance, the surface fluxes in CAM5 are in general of the same magnitude as in CAM4, despite the large differences in wind speed. As mentioned above, to maintain similar sizes of the fluxes, larger temperature gradients are needed in CAM5. Examining the stability as measured by the Obukhov length and the Richardson number (not shown), CAM5 is much more stably stratified during stable conditions and also slightly more unstable under unstably stratified conditions than both CAM4 and CAM5noTMS.
The largest discrepancies between models and observations in daytime fluxes are found at the midlatitude and boreal forest sites in winter, where the modeled latent heat flux is up to 300% too high. In summer, the arid sites stand out with a too high sensible heat flux and a latent heat flux that appears to be capped and is only one-third of what is observed. The cap is caused by a restriction of the stomatal resistance (which controls the latent heat flux) of grass at low relative humidity in order to prevent numerical instability in the iterative stomatal resistance calculation in the land model (Oleson et al. 2010).
The phase and shape of the diurnal cycles of wind speed and friction velocity are similar in CAM4 and CAM5. The differences in magnitudes between the models due to TMS are apparent, and they are smaller at the forest sites, particularly in the tropics, just as was seen in Fig. 4. The overestimated wind speeds in CAM4 are seen in all regions, whereas CAM5 matches the observations better. Both models overestimate the diurnal amplitude of wind speed at the forest sites with simulated amplitudes resembling those at the unforested sites. The friction velocity is in general overestimated at the unforested sites and closer to the observations at the forested sites. This is partly caused by the problematic comparison of a point observation with a grid box that contains more than one vegetation type.
At the high and midlatitude sites, the net radiation (downward) is too negative in winter, especially at nighttime. In summer it is still too low throughout the day at the tundra sites but overestimated at the boreal forest and midlatitude sites. At the midlatitude sites, the bias stems from the longwave radiation (not shown) at those sites where the separate components are available, and at the forest sites there is also a discrepancy in the outgoing shortwave radiation (not shown), probably because the albedo at the observation site is different from the mean albedo in the grid box. At the polar sites, the data are lacking to determine in which radiative components there is a bias. At the tropical sites, the net radiation is underestimated in the mornings and overestimated in the afternoons. The arid sites show a too large daytime net radiation. This is especially true in winter because of too little outgoing shortwave (not shown), despite a too large outgoing longwave component.
However, it should be noted that flux sites in general have a closure problem in which the components of the surface energy balance do not add up. Caution should therefore be used when trying to deduce the surface energy balance from these observations.
The simulated 2-m specific humidity (not shown) shows larger diurnal amplitudes than the observations. The exception is in winter at the polar sites, where there is no clear diurnal cycle at all in the observations. Both model versions contain morning and evening maxima that are either completely absent or much smaller in the observations. A possible explanation could be the increase (decrease) in latent heat flux preceding (succeeding) the increase (decrease) in boundary layer height in the morning (evening) (not shown), resulting in a flux of moisture into a shallow layer.
The PBL parameterization gives the turbulent vertical mixing in the model through the eddy diffusivities for heat, momentum, and scalars. As is mentioned in section 2b, the models calculate the vertical diffusion diffusivities in very different manners. CAM4 uses a K-profile scheme and has a minimum value for the diffusivities, keeping the atmosphere constantly turbulent. It should also be noted that, in CAM4, part of the mixing in the convective boundary layer is done by the countergradient term. CAM5, on the other hand, determines the diffusivity at each height, using the diagnosed TKE, and the diffusivities go all the way down to zero. As in all models, the numerical schemes also introduce some numerical diffusion.
The probability distribution function (PDF) of the diffusivity for heat KH at each model level for the flux sites is presented in Fig. 9. The eddy diffusivity for momentum KM is not shown but gives a similar picture only with in general slightly lower values, because of the turbulent Prandtl number used. The turbulent Prandtl number (not shown) is at all times slightly less than unity in CAM5. In CAM4, it shows a larger variability, but remains in most instances below unity. The left panels in Fig. 9 show the fraction of KH > 0.01 m2 s−1 (the minimum value in CAM4). For CAM5 and CAM5noTMS, KH is equal to zero in 99% of the cases where KH < 0.01 m2 s−1, and thus the fraction shown is virtually the same as the ratio of turbulent cases.
The cutoff Richardson number applied in CAM5 results in this fraction being much lower than in CAM4, and it is even less in CAM5noTMS because of its smaller roughness lengths. This is clearly seen near the surface. At the lowest model level in CAM4, the ratio of diffusivities that exceeds the background value is 92%, whereas the lowest model level in CAM5 is only turbulent 74% of the time at the flux sites.
On occasions when turbulence is strong, KH reaches higher values in CAM5 and CAM5noTMS than in CAM4. This results in larger spread of diffusivities in the CAM5 versions than in CAM4. The highest diffusivities are found at the height of approximately 1–2 km, with values above 103 m2 s−1 in CAM5.
At higher altitudes, above 3 km, KH is in CAM4 at most times equal to the background value, 0.01 m2 s−1. The only exception is when there is enough shear to generate turbulence through the free-atmosphere parameterization (see section 2b), which occurs in less than 15% of the cases. The relative occurrence of turbulence above 3 km in CAM5 and CAM5noTMS is similar to what it is CAM4. However, turbulence here is mainly caused by cloud-top cooling in the CAM5 versions, giving rise to much larger free-atmospheric values with KH around 1 m2 s−1. When no clouds are present, the free atmosphere in CAM5 and CAM5noTMS is almost always nonturbulent with values of KH above zero in less than 4% of the cases (not shown).
The disparities in both occurrence and magnitude of the eddy diffusivities in the two model versions are substantial and are expected to give rise to rather large differences in the vertical transport. The necessity to extract more momentum through the TMS parameterization in CAM5 is likely to originate from this difference. Observations of eddy diffusivities are rare but indicate that the atmosphere appears to always be turbulent (e.g., Clayson and Kantha 2008). However, the effect on the global circulation and large-scale transport of heat and scalars are outside the scope of this paper.
In this study, we have evaluated the performance of near-surface parameters in the two versions of the atmospheric component of CESM: namely, CAM4 and CAM5. There are substantial differences in the parameterization schemes between the two versions and most parameterizations in the atmosphere and the land surface model have impact on the near-surface parameters. While we study the mean variables such as temperature and wind speed, we emphasize the comparison with observed turbulent fluxes and net radiation when we compare the model simulations with, in total, over 260 yr of flux station data gathered at 35 sites in different climate zones. From this evaluation, we find the following:
Both CAM versions capture the timing of the diurnal cycles reasonably well but substantially overestimate the diurnal amplitude in general. This is particularly true for temperature and wind speed, which consistently have too large simulated amplitudes, but also in most cases for the turbulent fluxes. The cause is likely the too large diurnal amplitude of net radiation, driving the diurnal cycle.
The summer temperatures are too high in CAM4 and CAM5 compared to the global reanalysis and observational datasets as well as to the flux site observations. The flux tower observations show that the models are too warm primarily during daytime, with biases on the order of 5°C at the midlatitude, boreal forest, and arid sites. It is likely connected to the overestimated net surface radiation, although at the arid sites the limited evaporation seems to play an important role.
Both models are too cold in winter at high latitudes, and CAM5 underestimates the temperature more than CAM4. At the boreal forest sites, where the observations show a diurnal cycle even in wintertime, the largest differences between the models and between models and observations occur at nighttime. The low temperature bias is connected with too much radiative cooling in both models. Lower wind speeds and a turbulence scheme that shuts off mixing when Ri > 0.19 lead to less mixing in CAM5 than in CAM4 and could account for the model dissimilarities. The sensible heat flux has also a larger negative bias in CAM4, compensating slightly for the net radiation bias. However, there might also be other explanations, such as difference in cloudiness, which we have not investigated.
The sensible heat flux is too negative in CAM4 when the turbulence is shear driven: that is, at nighttime and in wintertime at high latitudes. This is could be due to the overestimated wind speeds in CAM4 or a too diffusive PBL scheme.
Both CAM versions simulate similar 10-m wind speeds over ocean, while the wind speed is substantially lower in CAM5 than in CAM4 over land. CAM5 matches the flux observations and the NCEP-2 dataset better than CAM4 does, but the spatial variation is underestimated in CAM5 compared to ERA-Interim. CAM4, on the other hand, simulates a higher the wind speed, particularly in mountain regions. The model differences are explained by the TMS parameterization applied in CAM5, which decreases the wind speed. Its impact is large in unforested areas, where the roughness length otherwise would be low, and in mountain regions. The effect of TMS is enforced when the boundary layer is stably stratified (e.g., at high latitudes).
The PBLH is generally lower in CAM5 than in CAM4. The exception is the stratocumulus regions, where CAM5 simulates about 200–400 m higher PBLHs than CAM4 and is in closer agreement with ERA-Interim. This can be attributed to the ability of the UW turbulence scheme in CAM5 to account for turbulence caused by cloud-top cooling. TMS increases the PBLH in the convective areas of the dry subtropics, regions where CAM5 still simulates a shallower PBLH than ERA-Interim.
It is apparent that the TMS parameterization in CAM5, by reducing the magnitude of the wind speed as well as increasing the PBLH in convective areas, has positive effects on the model simulations. Nevertheless, the implementation could be improved. In the current formulation of TMS, the roughness length is substantially increased even in relatively flat regions, causing a too low spatial variation of near-surface wind speed. Also, the fact that TMS is only applied in the atmospheric component and not in the land component gives rise to inconsistencies in the model.
The UW turbulence scheme in CAM5 is in many ways an improvement compared to the HB scheme in CAM4, in particular the ability to account for turbulence caused by cloud-top radiative cooling. However, the fact that mixing is completely shut off when Ri > 0.19 is not consistent either with observations, which indicate a constantly turbulent atmosphere, or with the surface layer scheme in the land model.
Depending on which of the two versions one would use for applications (e.g., wind energy potential or natural aerosol emissions), the disparities in the near-surface wind climate would give rise to substantial differences. For natural emissions of aerosols, the difference would be small for sea-salt aerosols but much larger for dust emissions. Furthermore, the concentrations would also disagree since the parameterizations of vertical turbulent mixing and the resulting boundary layer height differ between the model versions.
This study documents the performance of near-surface parameters in two versions of CAM and demonstrates the utility of using flux station observations for model evaluation and to assess the diurnal cycle. It serves as a benchmark to measure improvements in parameterization development (e.g., improvement of the TMS parameterization).
We thank the individual PIs at the flux sites and their teams for the data collection and preparation. We acknowledge the Ameriflux, AsiaFlux, CarboAfrica, CarboEurope IP, Fluxnet-Canada, LBA, and Ozflux projects, all parts of FLUXNET, for coordinating and providing data. Data from the Coordinated Energy and Water Cycle Observation Project (CEOP) were provided by NCAR EOL (http://data.eol.ucar.edu/) under sponsorship of the National Science Foundation. The CESM project is supported by the National Science Foundation and the Office of Science (BER) of the U.S. Department of Energy. NCAR is sponsored by the National Science Foundation. Computing resources were provided by the Climate Simulation Laboratory at NCAR’s Computational and Information Systems Laboratory (CISL), sponsored by the National Science Foundation and other agencies. We also thank Bert Holtslag for valuable input and discussions.
This article is included in the CESM1 Special Collection.