1. Introduction
The atmospheric boundary layer (ABL) is the lowest layer of the earth’s atmosphere through which exchanges of momentum, energy, and other chemical species with the surface are transported vertically, via turbulent motions, on a time scale that is rapid in comparison with exchanges within the free atmosphere above (e.g., Stull 1988). As such, the ABL plays a critical role in both the evolution of flow near the surface, and in the evolution of larger-scale weather phenomena as well. Accurate prediction of atmospheric flow and physical processes spanning many scales of motion, both near the surface and aloft, depends upon accurate modeling of the ABL and the turbulence within.
While all atmospheric flow models must contend with the turbulence closure problem, mesoscale models rely almost entirely upon parameterizations, as essentially all of the turbulence phenomena exist beneath their coarse resolutions. Such approaches, which are required for large computational domains, can unfortunately lead to significant errors in boundary layer flows. Simply increasing the resolution of such models will, in itself, not reliably improve the simulation of turbulence, because of the assumptions upon which such parameterizations are based. These turbulence parameterizations are appropriate for coarse resolutions, but become increasingly tenuous at higher resolutions (see e.g., Wyngaard 2004). When ABL simulations of the highest accuracy are desired, a different approach, the large-eddy simulation (LES), is the preferred methodology (Mason 1994).
The LES technique is predicated on explicit resolution of the largest, energy-producing turbulent motions within the flow. The smaller scales of motion, which function primarily to extract and dissipate energy from the larger scales, are removed from the solution using a low-pass filter. The effects of the filtered scales of motion on the resolved-scale flow are instead parameterized in a subfilter-scale (SFS) stress model. The SFS stress model used in LES provides temporally and spatially varying forcing for turbulence motions that are explicitly resolved, rather than representing the effects of the entire turbulence spectrum on a slowly varying averaged quantity, as large-scale parameterizations do.
Since the first application of the LES technique to the atmospheric boundary layer in the early 1970s (e.g., Deardorff 1970) the approach has been successfully employed in a variety of atmospheric and engineering applications (e.g., see Sullivan et al. 1994; Meneveau and Katz 2000). The complexity and variety of SFS stress models have likewise advanced considerably, resulting in a spectrum of models based upon different assumptions, representing different physical processes, and requiring a wide range of computational resources. Here we investigate the performance of a selection of turbulence models of varying levels of complexity and computational expense, in a widely used atmospheric large-eddy simulation model, the Weather Research and Forecasting (WRF) model, to assess the benefits and drawbacks of more elaborate approaches in simulations of neutral boundary layer flow over both a flat terrain and a symmetric Gaussian ridge. The need for more sophisticated closure options in models such as WRF is motivated by the increasing application of such codes to LES of the ABL (e.g., Moeng et al. 2007; Mirocha et al. 2010).
2. Large-eddy simulation
Here, 
3. SFS stress models
The SMAG model relates the deviatoric part of the SFS stress field τij to the resolved strain-rate tensor, 
The value of the parameter cs,Δ determined by Lilly (1967) has, in practice, been modified by LES practitioners (e.g., Deardorff 1970; Mason 1994; Ciafalo 1994) due to overprediction of the near-surface stress field. Commonly used values are 0.1 < cs,Δ < 0.25; cs,Δ = 0.25, the default WRF model value, was used in this study. The length scale is typically related to the size of the computational mesh as Δ = (ΔxΔyΔz)1/3.
The constant-coefficient SMAG and TKE models given by Eqs. (5) and (6) are well established in the literature and widely used (e.g., Cai 2000; Nakayama et al. 2008). These models, however, have been shown to be overdissipative, and to produce excessive shear in the near-wall regions for wall-bounded shear flows (e.g., Mason and Thomson 1992; Porté-Agel et al. 2000).
a. Lagrangian-averaged scale-dependent dynamic model
An alternative to the constant-coefficient SMAG and TKE models is the dynamic Smagorinsky model (DSM), which computes the constant cs,Δ in Eq. (5), using the resolved-scale stresses within the computed flow to provide information about the value of the model constant.
Application of a second test filter with a width of α2Δ provides a means for calculating the resolved stresses arising from a second, larger range of scales, such that the relative stresses arising from the larger (α2Δ) and smaller (αΔ) resolved scales can be obtained from expressions analogous to Eqs. (7)–(10) to determine how the stresses and the model coefficient vary across scales. This ratio β can then be used to extrapolate an appropriate value of the constant for use at the SFS scales, cs,Δ. The scale dependence of the coefficients under all stability conditions, as well as the implicit assumption of this approach that β itself does not vary with scale, were verified a priori using field experimental data by Bou-Zeid et al. (2008).
Both scale-dependent and scale-independent versions of the dynamic model can take locally small or even negative values, resulting in the potential for numerical instability. Dynamic models are often stabilized by averaging the coefficients in space, typically either within planes or in homogeneous directions of the flow, which is denoted by the angle brackets in Eq. (10). For many flows, however, particularly those over complex terrain, homogeneous directions may not exist. An alternative is to perform Lagrangian averaging along the fluid trajectories (Meneveau et al. 1996; Bou-Zeid et al. 2005). This approach extends the averaging required to stabilize the dynamic SFS stress models to flows over complex terrain. Bou-Zeid et al. (2005) recently combined a simplified implementation of the scale-dependent dynamic SFS model with Lagrangian-averaging to develop the Lagrangian-averaged scale-dependent (LASD) SFS stress model.
One problem with the eddy-viscosity SFS models, such as the constant and dynamic Smagorinsky and TKE-1.5 models described above, is that they assume a perfect alignment of the eigenvectors of the SFS stress and strain tensors. The invalidity of this assumption has been demonstrated, for example, by Higgins et al. (2003) and Bou-Zeid et al. (2010). Another drawback of these models is their inability to account for backscatter, the reverse flow of energy from smaller to larger scales, in a physically consistent manner (e.g., Kosović 1997). While the transfer of TKE and reproduction of low-order flow statistics remain the most important aspects of SFS model performance (Meneveau and Katz 2000) and are well represented by eddy-viscosity models, other aspects of SFS model performance, including, for example, alignment of the eigenvectors and energy backscatter, remain important as secondary considerations (Kosović et al. 2002).
b. Nonlinear backscatter and anisotropy model
The backscatter coefficient controls the partitioning of the stresses arising from the Smagorinsky term and the nonlinear terms in Eq. (12) and effectively reduces the overall dissipation rate (see Kosović 1997). The NBA model can also be formulated in terms of SFS TKE, with modifications to the model constants. The model parameters are chosen to ensure that the normal stress effects in sheared homogenous turbulence are captured, and have been shown to better recover the misalignment of the eigenvectors of the stress and strain-rate tensors, while also reducing the excessive dissipation of the constant-coefficient SMAG model (Kosović et al. 2002).
While the backscatter parameter in the NBA model represents an average value, the instantaneous backscatter is determined by the resolved flow structures. However, Chen et al. (2003, 2005), Chen and Tong (2006), and Chen et al. (2009) have argued that the conditional backscatter is more important for one-point statistics, such as mean and variance, and is also a necessary condition for the correct evolution of the resolved-scale velocity joint probability density function (JPDF), due to the direct influence of the conditional energy transfer rate on the JPDF transport equation. Chen and Tong (2006) showed that the NBA model improved predictions of both the conditional SFS stress and the conditional SFS stress production rate, as well as the level of anisotropy, relative to the Smagorinsky model, in simulations of the convective ABL. Both the magnitude of the conditional SGS stress and the level of anisotropy, however, were underpredicted relative to observations.
c. Dynamic reconstruction model
Another approach for modeling the SFS stresses that has appeared in the literature recently is the dynamic reconstruction model (DRM; Chow 2004; Chow et al. 2005). The DRM is a mixed model, utilizing both a dynamic eddy-viscosity component (similar to the DSM), and a scale-similarity term, computed using an explicit filtering and reconstruction approach.
The first group of terms on the right-hand side of Eq. (14) represents the stresses from scales not resolvable due to the combined effect of the computational mesh and discretization errors in the numerical solver (the “subgrid” stresses SGS). The second group of terms represents scales that are, in principle, resolvable, but have been removed by the effects of the explicit filter. As such, these are denoted as the resolvable subfilter-scale stresses (RSFS). The SGS term, which arises from scales not representable because of the grid and effects of the numerical solution procedure, can be modeled using a traditional eddy-viscosity model similar to those described above. The RSFS term, since it results from application of an explicit filter, can be computed using the process of explicit filtering and reconstruction.
The LASD model has been shown to produce sufficient model stresses near the surface when implemented into a pseudospectral code (Bou-Zeid et al. 2005). When implemented into the ARW dynamical core, however, the model stresses, as with the DRM, are underpredicted near the surface, resulting in poor agreement with the similarity wind speed profile. The disparate performance of the LASD model using different numerical solution procedures suggests that the increased numerical dissipation arising from the ARW’s finite-difference numerical solvers (fifth- and third-order options in the horizontal and vertical directions, respectively, were selected for this study), hinders representation of the smallest-scale structures resolvable upon the computational mesh. It is precisely these small scales upon which both of the dynamic SFS stress procedures (particularly the LASD model, which uses two levels of filtering) are based (see also Brasseur and Wei 2010). This underprediction of the near-wall stresses in the LASD model is likewise ameliorated by utilizing the near-wall stress model described above, albeit with slightly different parameters than the DRM (discussed below).
In the next section, we compare the Lagrangian-averaged scale-dependent SFS model of Bou-Zeid et al. (2005) and the dynamic reconstruction model of Chow et al. (2005) with the existing WRF model’s linear (SMAG) and nonlinear (NBA) SFS stress closures, and against similarity solutions, in simulations of neutrally stratified atmospheric boundary layers.
4. Simulation setup
This study uses the WRF model, a general atmospheric simulation model that can be applied to many scales of flow, from numerical weather prediction to large-eddy simulation. Specifically, the Advanced Research WRF (ARW-WRF) core is used. The ARW-WRF solves the compressible, nonhydrostatic Navier–Stokes equations (Skamarock et al. 2008). Its vertical coordinate is a terrain-following hydrostatic pressure coordinate. The ARW-WRF has an LES capability in the form of SFS stress closure models, idealized lateral boundary conditions, and the option to specify surface roughness and heat and moisture fluxes. Three of the SFS models described above, the Smagorinsky, TKE-1.5, and NBA models, are included in the current version of the WRF (version 3.2).
The new SFS stress models were implemented in a manner consistent with the WRF model’s native SFS stress models. The SFS stresses and eddy-viscosity coefficients are computed at the same locations within the staggered mesh, and use the same averaging stencils. Moeng et al. (2007) compared the calculation of the SFS eddy-viscosity coefficients (and TKE) at the gridcell interfaces (w levels), as opposed to the gridcell center (mass levels), as is standard in the WRF model, but did not observe any significant improvement in the results.
The performance of the new SFS models is evaluated in LES of ABL flows over a flat terrain, and over a transverse ridge, both under neutral stability. Periodic boundary conditions are used in the lateral directions. A geostrophic wind of 10 m s−1 is applied in the x direction, which is similar to many other LES investigations (e.g., Andren et al. 1994; Chow et al. 2005; Mirocha et al. 2010). The Coriolis parameter is fc = 10−4 s−1, corresponding to latitude of approximately 45°N. Fifth- and third-order spatial differencing is used for the horizontal and vertical advection, respectively.
The surface stresses 
Here CD = κ2{ln[(z1 + z0)/z0]}−2 is the coefficient of drag formulated for neutral flow, for which a roughness length of z0 = 0.1 m is used, subscript 1 denotes the first model grid point above the surface for which u and υ are computed, and κ is the von Kármán constant, which is 0.4.
The simulations were run for approximately 24 physical hours using the LASD SFS stress model until a nearly steady-state solution was obtained. The simulations were then run for an additional 2 h with each different SFS stress model activated to allow the flow to equilibrate to the new SFS stress model before calculating statistics for each model. A summary of the SFS models used in this study is given in Table 1. The statistics are obtained by averaging the solution over horizontal planes for a duration of 4 h for the flat-terrain case; for the ridge topography, only time and cross-stream averaging are performed.
The summary of the SFS models used in this study.
Each simulation was run on 64 processors of a Linux cluster using AMD Opteron CPU cores. The implementation of the new SFS models in WRF in parallel required no special treatment beyond the standard WRF model’s halo and periodic boundary condition specification protocols in the registry. We have not conducted a scaling test to very large numbers of processors, but do not anticipate any issues related to parallelism as the models rely on local filtering and differentiation operators. All of the turbulence models were implemented in version 3.1 of the WRF model, which is used to perform the simulations.
5. Results
a. Neutral ABL flow simulations over flat terrain
The neutral ABL is characterized by zero surface heat flux and can occur during the morning and evening transitions, or when the surface and the air above have the same temperature. The computational domain used for the flat-terrain simulations has a size of 2.048 km in the x and y directions and 1.024 km in the z direction. The performance of the SMAG and TKE-1.5 models during these idealized, neutrally buoyant simulations was nearly identical (as also observed by Mirocha et al. 2010); therefore, only the SMAG model will be discussed. For the DRM, we use level-2 reconstruction to calculate the RSFS velocity field, balancing the improvement in the solution and the computational cost of a higher-level reconstruction (Chow et al. 2005).
We first discuss a case where the first z level above the lower boundary (Δz) is located at approximately 8 m above the surface and the horizontal grid resolution (Δh) is 32 m, which yields an aspect ratio Δh/Δz = 4 near the surface. The vertical grid spacing is stretched by a factor of approximately 1.05 per nodal index above the surface. This resolution and aspect ratio were also studied by Mirocha et al. (2010), in their WRF simulations of neutral ABL flow over flat terrain. They showed that aspect ratios of 3 and 4 produced vertical wind speed profiles in close agreement with the similarity solution, with larger deviations from similarity at larger or smaller aspect ratios (Mirocha et al. 2010). We analyze this case to see if the dynamic models perform similarly to the other models.
Figure 1 shows the vertical profiles of mean total τ13 stress for all the SFS models. The magnitudes of the surface stresses predicted by each model are in good agreement, which is not surprising, since the wall stress must balance the geostrophic forcing applied. The partitioning of the total stresses into resolved and SFS components away from the wall, however, are quite different. The NBA and LASD models resolve a higher fraction of the total stresses in the lower portion of the boundary layer compared with the other SFS models. The magnitude of the resolved stresses is smallest for the DRM model as a result of the partitioning of the SFS stresses into SGS and RSFS components (i.e., more of the stress is placed in the subfilter scales). For both of the dynamic models, the magnitude of the near-surface stress term is larger than the modeled stress term for z/H < 0.05, accounting for approximately 60% of the total stress at the first horizontal velocity grid point.
Vertical profiles of time- and plane- averaged τ13 stress component using the (a) SMAG, (b) NBA, (c) LASD, and (d) DRM models.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
Figure 2 portrays the vertical profiles of mean wind speed for all of the SFS models. The thin black line corresponds to the theoretical log-law profile given by Monin–Obukhov similarity theory. The SMAG model underpredicts the theoretical profile near the surface (z/H < 0.03, H being the height of the ABL) and overpredicts it at higher elevations. This poor performance is a result of its overdissipative nature in the near-surface region. The NBA model produces a similar bias, underprediction near the surface followed by overprediction farther aloft; however, the magnitudes of the biases are reduced compared with the SMAG model. The dynamic models provide further improvement in agreement with the log-law profile, especially for z/H > 0.02, compared with the other models. This agreement is due, in part, to the additional near-wall stress term, which possesses several parameters that can be tuned to compensate for errors not only in the SFS model physics, but in other aspects of model performance as well (e.g., surface stress parameterization and numerical errors). The parameters used in the near-wall stress parameterizations for the simulations in this study [see Eq. (17)] are 
Vertical profiles of time- and plane-averaged wind speed scaled by friction velocity using all four SFS models.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
Mirocha et al. (2010) and Brasseur and Wei (2010) recently reported that the near-surface resolution Δz and aspect ratio are important in obtaining the correct log-law profile for large-eddy simulation of atmospheric boundary layer flows. As such, once Δz is chosen, only certain range of aspect ratios give the correct log-law profile. Here, we investigate the dependence of the mean wind speed profiles on Δz and aspect ratio for the SFS models used in this study. This is also important for WRF’s nested domain simulations where Δz is held constant but the aspect ratio is refined or coarsened (e.g., by a factor of ~2–4 or greater) across nest interfaces. We use the same near-wall stress parameterizations (coefficients listed above) to evaluate the behavior of the dynamic models. Figure 3 shows the vertical profiles of the mean wind speed for two aspect ratios and different values of Δz. The profiles from Δh = 16 m case (high-resolution case) are shown on the left (Figs. 3a,c,e,g) and the ones from Δh = 32 m case (low-resolution case) are shown on the right (Figs. 3b,d,f,h). The first observation is that changing the aspect ratio has a similar effect on the profiles for both resolutions and that the aspect ratio of 4 minimizes the departures from the log-law solution (the gray line). When the SFS models are compared with each other, however, some differences are observed. Reducing the aspect ratio for the NBA and SMAG models increases velocities near the surface, whereas for the dynamic models, the opposite occurs. The behavior of the dynamic models might seem counterintuitive at first but since the near-wall stress parameterizations are optimized for an aspect ratio of 4; different parameters would optimize the solution for an aspect ratio of 2. In fact, we tested a smaller near-surface model coefficient for the aspect ratio of 2, and obtained good results (not shown here). We therefore conclude that aspect ratio is more important than near-wall resolution to obtain the correct near-surface similarity for all of the SFS models. Similar results were obtained by Chow et al. (2005). We will now continue analyzing (Δh, Δz) = (32, 8) m case to compare the SFS models with each other.
Vertical profiles of time- and plane-averaged wind speed for (a),(b) SMAG; (c),(d) NBA; (e),(f) LASD; and (g),(h) DRM model using aspect ratios (a.r.) of 2 and 4 and horizontal grid resolutions of (left) Δh = 16 m and (right) Δh = 32 m.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
Profiles of the nondimensional mean velocity gradient [ΦM(z/L) = (κz/u*)(∂U/∂z)] from simulations using each of the SFS models are shown in Fig. 4, using Δh = 32 m and an aspect ratio of 4. The ΦM should take a value of 1 for neutral ABLs in the framework of the Monin–Obukhov similarity theory. The SMAG model produces the largest departures from the expected value, while the NBA reduces the overpredicted ΦM value in the z/H < 0.15 region (consistent with Fig. 2). The dynamic SFS models further improve the profiles in this region, with the DRM model predicting the best agreement overall. The agreement of the dynamic model simulations with the similarity profile is sensitive to the near-wall stress term, the value and profile shape of which are chosen to match the expected similarity behavior near the surface. It should be noted that the performance of the LASD model is somewhat poorer than its performance in other codes (e.g., Bou-Zeid et al. 2005; Kleissl et al. 2006), underlining the importance of the interaction between the numerical solution method and SFS stress modeling in LES.
Vertical profiles of time- and plane-averaged nondimensional shear [ΦM = (κz/u*)(∂U/∂z)] using all four SFS stress models.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
Figure 5 shows vertical profiles of the resolved velocity variances that, in addition to being a measure of turbulence model performance, are important for particle transport and diffusion as well (e.g., Michioka and Chow 2008). The peak of the variance is located furthest from the surface using the SMAG model, while the other models predict maxima closer to the surface. The magnitudes of the variances are highest for the LASD model. Between the dynamic models, the DRM model predicts lower variances compared with the LASD model due to the RSFS stresses, which reduce the magnitude of resolved stresses, therefore reducing the resolved velocity variances. The maximum values of the nondimensional u, υ, and w variances are in good agreement with previous neutral ABL studies. For example, the maximum velocity variances reported by Grant (1992, and references therein) are of the magnitude: 
Vertical profiles of time- and plane-averaged resolved velocity variances for (a) u-, (b) υ-, and (c) w-velocity components using all four SFS stress models.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
The performance of the SFS models in generating turbulent structures with different length scales can be quantified by examining the spectra of the velocity field. Figure 6 shows the time and horizontal plane averaged energy spectra of streamwise velocity at different heights for the SFS models used in this study. The differences among the spectra arising from the use of the different SFS models will be greatest near the surface, due to the decrease of the sizes of the eddies approaching the surface, which increases the role of the SFS stress models there. The first observation is that the SMAG model predicts a steeper slope (< − 1) in the production range (the low wavenumber range, i.e., k1z < 1) near the surface compared with the NBA and dynamic models. This clearly portrays the overdissipative behavior of the SMAG model. The energy levels extend well into smaller wavenumbers in the production range (k1z < 1) for the dynamic models, which are less dissipative. The spectral energy in the inertial range (k1z > 1) is higher for the NBA and dynamic models compared with the SMAG model. At higher elevations, all SFS models predict a 
Time- and plane-averaged spectra of u-velocity computed in the x-direction vs k1z at heights (from left to right) z/H = 0.012, 0.03, 0.058, 0.12, 0.23, 0.5 for the (a) SMAG model, (b) NBA model, (c) DRM model with reconstruction level 2, and (d) LASD model.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
Time and plane-averaged spectra of u-velocity computed in the x-direction vs k1z using each of the four SFS stress models at (a) z/H = 30 m and (b) z/H = 120 m.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
Figure 8 shows contours of instantaneous u velocity at approximately 50 m above the surface for the 4 SFS models used in this study. Various laboratory experiments and field observations have reported that the flow field in the bottom 10% of the ABL contains elongated streaky structures with lengths 15δ–20δ (δ is the depth of the surface layer, ~0.1H) that are aligned parallel to the mean wind direction (e.g., Hutchins and Marusic 2007). The SMAG model predicts very long streaky structures with relatively fewer small-scale structures (Fig. 8a). This behavior likely stems from the excessive dissipation of the SMAG model, reducing the energy of smaller structures and permitting the maintenance of larger coherent structures. The average length of the streaks generated by the NBA and dynamic models are much shorter (around 1000 m), which is consistent with the observations (Hutchins and Marusic 2007) and previous model comparisons (Ludwig et al. 2009).
Contours of instantaneous u-velocity in the x–y plane at z = 50 m for (a) SMAG, (b) NBA, (c) DRM, and (d) LASD models, showing turbulence structures embedded in the flow. Flow is forced with a geostrophic wind from left to right.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
b. Neutral ABL flow simulations over a symmetric transverse ridge
Simulations of neutral flow over a two-dimensional Gaussian ridge are performed to assess the performance of the SFS models over nonuniform terrain. The domain is 1024 m in each horizontal direction, and its height is the same as the flat-terrain simulations (1024 m). The ridge is symmetric in the x direction, and uniform in the y direction, with a height of approximately 50 m. The maximum slope (the ratio of the change in height to the horizontal distance between model grid points) is 0.3. The forcing parameters used for these simulations are identical to those used for the flat-terrain simulation.
Figure 9 shows instantaneous contours of u velocity in the x–y plane at a height of approximately 10 m above the surface from simulated flow over the ridge using each of the SFS models. The flow is oriented in the positive x direction, and contours of negative zonal velocity are depicted by dotted lines. As with the flat terrain simulations, a significant difference among each of the flows over the ridge is the scale of the streamwise-oriented correlated streaks. The SMAG simulation again produces streaks with the greatest coherence, while the other models reduce the size of these streaky structures, with the dynamic models again producing the smallest structures overall. The excessive streakiness in the SMAG simulations is again likely related to the overdissipation of the SMAG model. The generation of small structures in the lee of the ridge axis occurs farther upstream for the DRM and LASD models than for the others.
Contours of instantaneous u-velocity in x–y plane at approximately 10 m above the surface from simulated flow over a symmetric transverse ridge using each of the 4 SFS models with a mesh resolution of Δh, Δz = 8, 2 m. Flow is forced with a geostrophic wind in the x direction. Dotted contours indicate negative velocities, showing flow separation and reversal in the lee of the ridge.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
The differences between the model predictions for the ridge case can also be seen in Fig. 10, where the vertical profiles of time and y-averaged wind speed are plotted at six points at the hill apex and downstream. The locations of those points are shown in Fig. 10g. At each of the points shown in Fig. 10, significant differences are observed among the models, with a slow convergence to one another at large downstream distances. At points b, c, and d (Figs. 10b–d), each of the models produces negative velocities at one or more of the locations, with the SMAG model predicting the smallest negative velocities overall, and the NBA model producing the largest. The NBA model also produces the deepest layer of negative velocities. Without comparison to experimental results, we only note that the differences in the profiles obtained using different SFS models can be significant, even for this simple flow over a shallow, two-dimensional Gaussian ridge.
Vertical profiles of time and spanwise (y direction) averaged wind speed profiles at (a)–(f) six points downstream of the hill apex. (g) The locations of the points are shown.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
Because of the importance of flow separation in applications involving hilly terrains (e.g., contaminant dispersion and wind turbine siting), the relative occurrences of flow reversal in the lee of the hill were selected for qualitative comparison. Figure 11 shows the relative frequencies of the magnitude of the u-velocity component at four heights above the surface downstream from the hill apex from the higher-resolution simulations (Δh, Δz = 8, 2 m) using each of the four SFS models. Following Mirocha et al. (2010), histograms were computed from the flow at 160 m downstream of the ridge apex at 4 heights specified in Fig. 11. The velocities were binned into 160 intervals spanning [− 4, 10] m s−1. The interpolation and binning were repeated across the entire domain in the y direction. The velocity distributions are depicted in the same panel for each for the four SFS stress models. The solid gray vertical lines indicate a value of zero, while the dotted gray lines indicate the mean values from each distribution at each height.
Histograms of u-velocity from 160 m downstream of the hill apex at 4 heights above the surface from simulations using mesh resolutions of Δh, Δz = 8, 2 m, using each of the 4 SFS models. Solid gray lines indicate zero, and dotted gray lines indicate the mean values from each of the distributions.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
Figure 11 indicates that the near-surface velocities in the lee of the hill contain many negative values, transitioning to positive velocities farther aloft. The NBA model predicts the greatest frequency of negative velocities close to the surface. Farther aloft the differences between the models are reduced with all models showing similar mean values and distribution.
While the simulations using different SFS stress models are not identical, we can still assume that overall the simulations conducted using smaller mesh spacing are more accurate than those using large mesh spacing. From this perspective, following Mirocha et al. (2010), we compare simulations performed on the finest computational mesh to simulations performed on meshes coarsened by factor of 1.5 (Δh = 12, Δz = 3 m) and 2 (Δh = 16, Δz = 4). The distributions are shown in Figs. 12 and 13, respectively. Again the NBA model produces relatively larger proportions of negative velocities at two heights above the surface, heights at which each of the higher-resolution simulations produced significant proportions of the negative velocities. The DRM produces the next greatest frequency of negative velocities at two heights above the surface compared with the SMAG and LASD models in the coarsest resolution simulation. The DRM model predicts the smallest proportion of negative velocities in the high-resolution case, which makes it the least sensitive model to the mesh resolution. The dynamic models and the SMAG model produce somewhat similar histograms at the four heights for all the three mesh resolutions, although the coherent structures they produce in the flow over the ridge (Fig. 9) are quite different. We note that dynamic models, especially the LASD, are expected to become resolution sensitive at very coarse grid resolution since they involved test filtering at a larger scale. We also note that the performance of the dynamic models in flow over the ridge is expected to be sensitive to the near-wall stress model parameters, which were calibrated for the flat-terrain applications. Different parameter values might alter the performance of the models in complex-terrain applications.
As in Fig. 11, but at 156 m downstream of the hill apex, using mesh resolutions of Δh, Δz = 12, 3 m.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
As in Fig. 11, but using mesh resolutions of Δh, Δz = 16, 4 m.
Citation: Monthly Weather Review 140, 1; 10.1175/MWR-D-11-00037.1
In general, finer grid spacing results in a greater proportion of negative velocities near the surface. Among the models, the NBA and DRM models produce relatively more negative velocities near the surface and a broader range of velocities across the vertical profile, which is more consistent with the higher-resolution simulations.
6. Summary and conclusions
The WRF model’s LES capability has become a powerful tool to study ABL flows in recent years. Accurate LES of the ABL requires both sufficiently high mesh resolution to capture the energy-producing scales of turbulence, and robust SFS stress models to provide the proper forcing for those resolved scales. In this study, two recently developed dynamic SFS stress models, the dynamic reconstruction model (DRM) and the Lagrangian-averaged scale-dependent (LASD) models, are implemented into the WRF model and compared with the WRF model’s constant-coefficient linear eddy-viscosity Smagorinsky model (SMAG) and the nonlinear backscatter and anisotropy (NBA) SFS stress closures. Large-eddy simulations of neutrally stratified ABL flow are performed over both flat terrain, and a symmetric transverse ridge. The two dynamic models (DRM and LASD) produce similar flow properties and statistics, including very good agreement with the Monin–Obukhov similarity theory as well as the production and, to a lesser extent, the inertial range scaling of spectra. At the smallest resolved scales the dynamic models produce spectra that decay more rapidly than the 
For the flat-terrain case, the two dynamic SFS models and the nonlinear model improve the vertical profiles of mean wind speed and nondimensional mean velocity gradients relative to the constant-coefficient Smagorinsky closure. The dynamic models show the greatest agreement with the expected similarity profiles of wind speed and spectral scaling. Both dynamic models require a separate near-surface stress model to augment the stresses near the surface that are underpedicted by the dynamic procedure.
Our WRF simulations also portrayed that all of the SFS models require a correct combination of the near-surface resolution and aspect ratio to produce the correct Monin–Obukhov similarity for the neutral conditions over a flat terrain case. This relationship also depends on the surface roughness (not shown here).
Among the models, the LASD model generates the greatest amount of turbulent fluctuations and therefore greater velocity variances near the ground. The SMAG model variances peak farther aloft than the other models, likely due to the SMAG model’s excessive dissipation. The large u-velocity variance produced by the SMAG model is due to the elongated streaks oriented primarily in the x direction. The DRM model produces the smallest variances overall, due to the partitioning of the subfilter-scale stresses into RSFS and SGS components, which increases the SFS stress contribution while reducing the resolved component, such that approximately the same total stresses are maintained.
For the transverse ridge case, the significant differences are observed between the models in predicting the u-velocity field and the vertical profiles of wind speed both at the hill apex and in the lee region. The low-resolution NBA and DRM simulations are able to generate distributions of u velocity that agree most closely with the high-resolution simulations. We expect the distribution of near-surface velocities, especially in the lee of the ridge to have some dependence on the parameters of the near-wall stress model, motivating further evaluation. In both the flat-terrain and ridge simulations, the SMAG model produces excessively elongated streaky structures in the streamwise direction, while the nonlinear and dynamic models predict shorter streaky structures in the lowest 10% of the ABL, with the dynamic models producing the least streakiness overall.
The improved performance afforded by the more sophisticated SFS stress models comes with an increase in complexity, and therefore an increase in computational expense, both in terms of floating-point operations per model time step, and memory requirements for additional variables. Table 2 shows the CPU time used by the various SFS models for a flow simulation over a flat terrain, compared to that required by the SMAG model. The NBA model increases the CPU time by 13%, compared with the SMAG model. The more complex LASD and DRM-R2 models increase the CPU time by 28%. The cost of the DRM model also depends on the level of the reconstruction used for the RSFS velocities. We observe a 3% difference in the CPU time between level-0 and level-2 reconstruction. While the NBA model requires only three additional three-dimensional variables (the three unique elements of the rotation-rate tensor) beyond those variables required by the SMAG model, the dynamic models increase the memory requirements significantly, as many more variables must be saved in memory for the filtering operations. It is possible that the code can be further optimized to reduce some of these costs. Updating the dynamic coefficient less frequently (i.e., every fifth time step instead of every time step), for example, can reduce the total CPU time used by the dynamic models. Further evaluation of the new SFS models in more complex flow situations, such as those involving complex terrain and stability effects, is required to better determine their performance improvements relative to their increased computational requirements.
The CPU time required by the SFS models. The CPU time is normalized by the SMAG model’s CPU time.
Acknowledgments
We are grateful to the editor and anonymous reviewers for their constructive comments. This work is performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52–07NA27344. Staff effort at LLNL was supported by the LLNL Laboratory Directed Research and Development (LDRD) Program, Project 09-ERD-038. Computations at LLNL’s Livermore Computing were also supported by LDRD. EBZ is supported by NSF under CBET-1058027 and FKC is supported by ATM-0645784.
REFERENCES
Andren, A., A. R. Brown, J. Graf, P. J. Mason, C.-H. Moeng, F. T. M. Nieuwstadt, and U. Schumann, 1994: Large eddy simulation of a neutrally stratified boundary layer: A comparison of four codes. Quart. J. Roy. Meteor. Soc., 120, 1457–1484.
Bou-Zeid, E., C. Meneveau, and M. B. Parlange, 2005: A scale-dependent Lagrangian dynamic model for large eddy simulation of complex turbulent flows. Phys. Fluids, 17, 025105, doi:10.1063/1.1839152.
Bou-Zeid, E., N. Vercauteren, M. B. Parlange, and C. Meneveau, 2008: Scale dependence of subgrid-scale model coefficients: An a priori study. Phys. Fluids, 20 (11), 115106, doi:10.1063/1.2992192.
Bou-Zeid, E., C. Higgins, H. Huwald, M. B. Parlange, and C. Meneveau, 2010: Field study of the dynamics and modelling of subgrid scale turbulence in a stable atmospheric surface layer over a glacier. J. Fluid Mech., 665, 480–515.
Brasseur, J. G., and T. Wei, 2010: Designing large-eddy simulation of the turbulent boundary layer to capture law-of-the-wall scaling. Phys. Fluids, 22, 021303, doi:10.1063/1.3319073.
Brown, A. R., J. M. Hobson, and N. Wood, 2001: Large-eddy simulation of neutral turbulent flow over rough sinusoidal ridges. Bound.-Layer Meteor., 98, 411–441.
Cai, X.-M., 2000: Dispersion of a passive plume in an idealized urban convective layer: A large-eddy simulation. Atmos. Environ., 34, 61–72.
Carati, D., G. S. Winckelmans, and H. Jeanmart, 2001: On the modeling of the subgrid-scale and filtered-scale stress tensors in large-eddy simulation. J. Fluid Mech., 441, 119–138.
Chen, Q., and C. Tong, 2006: Investigation of the subgrid-scale stress and its production rate in a convective atmospheric boundary layer using measurement data. J. Fluid Mech., 547, 6–104.
Chen, Q., H. Zhang, D. Wang, and C. Tong, 2003: Subgrid-scale stress and its production rate: Conditions for the resolvable-scale velocity probability density function. J. Turbul., 4, doi:10.1088/1468-5248/4/1/027.
Chen, Q., D. Wang, H. Zhang, and C. Tong, 2005: Effects of subgrid-scale turbulence on resolvable-scale velocity-scalar statistics. J. Turbul., 6, doi:10.1080/14685240500455824.
Chen, Q., M. J. Otte, P. P. Sullivan, and C. Tong, 2009: A posteriori subgrid-scale model tests based on the conditional means of subgrid-scale stress and its production rate. J. Fluid Mech., 626, 149–181.
Chow, F. K., 2004: Subfilter-scale turbulence modeling for large-eddy simulation of the atmospheric boundary layer over complex terrain. Ph.D. dissertation, Stanford University, 339 pp.
Chow, F. K., R. L. Street, M. Xue, and J. H. Ferziger, 2005: Explicit filtering and reconstruction turbulence modeling for large-eddy simulation of neutral boundary layer flow. J. Atmos. Sci., 62, 2058–2077.
Ciofalo, M., 1994: Large-eddy simulation: A critical survey of models and applications. Advances in Heat Transfer, J. P. Hartnett and T. F. Irvine Jr., Eds., Vol. 25, Academic Press, 321–419.
Deardorff, J. W., 1970: Preliminary results from numerical investigations of the unstable planetary boundary layer. J. Atmos. Sci., 27, 1209–1211.
Deardorff, J. W., 1980: Stratocumulus-capped mixed layers derived from a three-dimensional model. Bound.-Layer Meteor., 7, 199–226.
Drobinski, P., P. Carlotti, R. K. Newsom, R. Banta, R. Foster, and J.-L. Redelsperger, 2004: The structure of the near-neutral atmospheric surface layer. J. Atmos. Sci., 61, 699–714.
Germano, M., U. Piomelli, P. Moin, and W. H. Cabot, 1991: A dynamic subgrid-scale eddy viscosity model. Phys. Fluids, 3A, 1760–1765.
Grant, A. L. M., 1992: The structure of turbulence in the near-neutral atmospheric boundary layer. J. Atmos. Sci., 49, 226–239.
Higgins, C., M. B. Parlange, and C. Meneveau, 2003: Alignment trends of velocity gradients and subgrid scale fluxes in the turbulent atmospheric boundary layer. Bound.-Layer Meteor., 109, 59–83.
Hutchins, N., and I. Marusic, 2007: Evidence of very long meandering features in the logarithmic region of turbulent boundary layers. J. Fluid Mech., 579, 1–28.
Kleissl, J., C. Meneveau, and M. B. Parlange, 2003: On the magnitude and variability of subgrid-scale eddy-diffusion coefficients in the atmospheric surface layer. J. Atmos. Sci., 60, 2372–2388.
Kleissl, J., V. Kumar, C. Meneveau, and M. B. Parlange, 2006: Numerical study of dynamic Smagorinsky models in large-eddy simulation of the atmospheric boundary layer: Validation in stable and unstable conditions. Water Resour. Res., 42, W06D10, doi:10.1029/2005WR004685.
Kosović, B., 1997: Subgrid-scale modeling for the large-eddy simulation of high-Reynolds-number boundary layers. J. Fluid Mech., 336, 151–182.
Kosović, B., D. I. Pullin, and R. Samtaney, 2002: Subgrid-scale modeling for large-eddy simulations of compressible turbulence. Phys. Fluids, 14, 1511–1522.
Lilly, D. K., 1967: The representation of small-scale turbulence in numerical simulation experiments. Proc. IBM Scientific Computing Symp. on Environmental Sciences, Yorktown Heights, NY, IBM Data Processing Division, 195–210.
Lilly, D. K., 1992: A proposed modification of the Germano subgrid-scale closure method. Phys. Fluids, 4, 633–635.
Ludwig, F. L., F. K. Chow, and R. L. Street, 2009: Effect of turbulence models and spatial resolution on resolved velocity structure and momentum fluxes in large-eddy simulations of neutral boundary layer flow. J. Appl. Meteor. Climatol., 48, 1161–1180.
Mason, P., 1994: Large-eddy simulation: A critical review of the technique. Quart. J. Roy. Meteor. Soc., 120, 1–26.
Mason, P., and D. J. Thomson, 1992: Stochastic backscatter in large-eddy simulations of boundary layers. J. Fluid Mech., 242, 51–78.
Meneveau, C., and J. Katz, 2000: Scale-invariance and turbulence models for large-eddy simulation. Annu. Rev. Fluid Mech., 32, 1–32.
Meneveau, C., T. Lund, and W. Cabot, 1996: A Lagrangian dynamic subgrid-scale model of turbulence. J. Fluid Mech., 319, 353–385.
Michioka, T., and F. K. Chow, 2008: High-resolution large-eddy simulations of scalar transport in atmospheric boundary layer flow over complex terrain. J. Appl. Meteor. Climatol., 47, 3150–3169.
Mirocha, J. D., J. K. Lundquist, and B. Kosovic, 2010: Implementation of a nonlinear subfilter turbulence stress model for large-eddy simulation in the Advanced Research WRF model. Mon. Wea. Rev., 138, 4212–4228.
Moeng, C.-H., J. Dudhia, J. Klemp and P. Sullivan, 2007: Examining two-way grid nesting for large eddy simulation of the PBL using the WRF model. Mon. Wea. Rev., 135, 2295–2311.
Moin, P., K. Squires, W. Cabot, and S. Lee, 1991: A dynamic subgrid-scale model for compressible turbulence and scalar transport. Phys. Fluids, 3A, 2746–2757.
Nakayama, H., T. Tamura, and S. Abe, 2008: LES on plume dispersion in the convective boundary layer capped by a temperature inversion. J. Fluid Sci. Technol., 3, 519–532.
Porté-Agel, F., C. Meneveau, and M. B. Parlange, 2000: A scale dependent dynamic model for large eddy simulation: Application to a neutral atmospheric boundary layer. J. Fluid Mech., 415, 261–284.
Skamarock, W., and Coauthors, 2008: A description of the advanced research WRF version 3. Tech Rep. NCAR/TN-475+STR, National Center for Atmospheric Research, Boulder, CO, 113 pp.
Smagorinsky, J., 1963: General circulation experiments with the primitive equations. I. The basic experiment. Mon. Wea. Rev., 91, 99–164.
Stolz, S., and N. A. Adams, 1999: An approximate deconvolution procedure for large-eddy simulation. Phys. Fluids, 11, 1699–1701.
Stull, R. B., 1988: An Introduction to Boundary Layer Meteorology. Springer, 683 pp.
Sullivan, P. P., J. C. McWilliams, and C.-H. Moeng, 1994: A sub-grid scale-model for large-eddy simulation of planetary boundary-layer flows. Bound.-Layer Meteor., 71, 247–276.
Wong, V. C., and D. K. Lilly, 1994: A comparison of two dynamic subgrid closure methods for turbulent thermal-convection. Phys. Fluids, 6, 1016–1023.
Wyngaard, J. C., 2004: Toward numerical modeling in the “terra incognita.” J. Atmos. Sci., 61, 1816–1826.
