a. Participating models

Table 1 briefly documents the 10 participating SCMs, including their stratiform cloud and boundary layer turbulence parameterizations and three measures of their operational vertical resolution: the number of model levels below 1200 m and in total, and the grid spacing Δz at the base of the inversion z_i = 840 m. The Naval Research Laboratory (NRL) SCM did not execute the simulations with operational resolution; thus no such information was provided. Only the European Centre for Medium-Range Weather Forecasts (ECMWF), National Center for Atmospheric Research Community Atmospheric Model (NCAR-CAM), Community Atmospheric Model with University of Washington PBL scheme (CAM-UW), and Canadian Centre for Climate Modelling and Analysis–University of British Columbia (CCCma-UBC) SCMs use a grid suitable for GCM simulation; the others have been especially configured for this particular simulation, with few levels above 1200 m. The ECMWF SCM uses several new physical parameterizations targeted for future introduction into their operational forecast model, including a new boundary layer parameterization and a new approach for predicting cloud fraction and condensate.

Six of the 10 SCMs [CCCma-UBC, Koninklijk Nederlands Meteorologisch Institute (KNMI), National Aeronautics and Space Administration Langley Research Center (NASA-LaRC), NCAR-CAM, CAM-UW, and Max Planck Institute (MPI)] determine cloud condensate prognostically, while the other four calculate cloud condensate diagnostically. All models diagnose subgrid cloud cover fraction based on either relative humidity (e.g., CCCma-UBC, NCAR-CAM) or probability density function (PDF) of conserved variables such as θ_l and q_t [e.g., Recherche en Prévision Numérique (RPN), NRL, and NASA-LaRC]. The ECMWF SCM prognoses the grid-mean and subgrid variance of total water specific humidity, from which cloud fraction and cloud water are diagnosed (Tompkins 2002).

The boundary layer turbulence schemes used by the SCMs are quite diverse. They can be classified according to (i) their formulation of mixing within the STBL, (ii) whether they account for (“moist”) or neglect (“dry”) moist thermodynamics in calculating stratification of partially or totally saturated layers, and (iii) whether cloud-top entrainment is explicitly computed or an implicit consequence of the model’s mixing scheme.

The U.K. Meteorological Office (METO), CCCma-UBC, and NCAR-CAM SCMs use bulk nonlocal K-profile schemes. The NCAR-CAM and CCCma-UBC scheme is dry and surface-driven, while the METO scheme is moist and accounts for turbulence production by cloud-top radiative and evaporative cooling (Lock et al. 2000). The remaining SCMs use turbulence closure models of various complexity levels. The RPN, KNMI, CAM-UW, and MPI SCMs determine turbulent diffusivity based on prognostic turbulent kinetic energy (TKE). The NRL (Golaz et al. 2002; Larson and Golaz 2005) and NASA-LaRC (Cheng et al. 2004) SCMs are both partially prognostic third-order closure schemes. They predict second-order moments of w, θ_l, q_t, and a limited number of third-order moments. NRL predicts only one third-order moment (), while NASA-LaRC predicts three third-order moments (, , and ). Both schemes diagnose the other high-order moments based on a PDF determined from the predicted moments. The new ECMWF model uses an eddy-diffusivity/mass-flux approach, which combines a K-profile diffusion term with a mass-flux term to describe nonlocal transports (Tompkins et al. 2004).

Four SCMs (METO, ECMWF, CAM-UW, and MPI) employ an explicit entrainment parameterization at the top of buoyancy-driven boundary layers. In the METO, ECMWF, and CAM-UW SCMs, the entrainment parameterization includes both evaporative and radiative cooling effects when the boundary layer is topped by clouds. In the METO and ECMWF models, the entrainment rate W_e is explicitly computed following Lock et al. (2000) as

View Expanded

where V is a diagnosed turbulence velocity scale, ΔB is the buoyancy jump across the inversion, ΔF is the cloud-top radiative flux divergence, and A₁ and A₂ are empirical coefficients. The CAM-UW model incorporates ΔF into the calculation of V, and takes A₂ = 0. The MPI entrainment parameterization only considers cloud-top radiative cooling (A₁ = 0). Cloud-top entrainment is not explicitly represented in the KNMI SCM but it is implicitly included in the moist turbulence scheme by adjusting the turbulent mixing length (1) via

View Expanded

where N is Brunt–Väisälä frequency, and the constant C_e is chosen to obtain reasonable entrainment rates through strong inversions (Lenderink and Holtslag 2000).

Models also differ in their operational vertical resolution and the organization of the physical parameterizations. Such a diversity, in a way, makes attribution of differences between model results harder to interpret, but on the other hand, it provides a great opportunity to look at the behavior of a similar type of parameterization in a different model framework. In addition to that, intercomparison also allows a close examination of different schemes in the same model framework; for example, the NCAR-CAM and CAM-UW simulations are particularly illuminating to compare since they only differ in their boundary layer and shallow convection schemes. A detailed comparison of SCM and global simulations between NCAR-CAM and CAM-UW will be presented in forthcoming papers by P. Zhu and C. S. Bretherton (2005, unpublished manuscript, hereafter ZB) and C. S. Bretherton and P. Zhu (2005, unpublished manuscript, hereafter BZ).

The companion LES intercomparison consists of 10 LES models differing mainly in their numerical algorithms and the treatment of subscale motions. A total of sixteen 4-h simulations were executed under the same initial conditions and external forcings on a grid mesh of 96 by 96 points in horizontal with a grid spacing of 35 m and a vertical grid spacing of 5 m or less in the vicinity of cloud top. Statistical analyses were then applied to individual runs to obtain a simulation ensemble. Detailed descriptions of models and analysis methods are provided by STE.

b. Numerical setup

To run an SCM, one needs to specify initial conditions, such as dynamic and thermodynamic vertical profiles, and external forcings, such as horizontal and vertical advections. The initial conditions and external forcings of the DYCOMS-II case for SCM simulations are specified based on the observations during RF01 and are the same as those for the companion LES intercomparison. Because of the nature of LES, which integrates from a nonturbulent state to a fully turbulent state, these conditions and forcings have been idealized in such a way that the important observed features of STBL can be best represented by LESs after reaching quasi steady state. STE provides more rationale for the exact case specifications, but these are also included here for reference.

Based on observations, the initial thermodynamic profiles are

View Expanded

View Expanded

where θ_l is the liquid water potential temperature; q_t, the total specific humidity; z, the height above the surface; and z_i = 840 m the base of the inversion, equivalent to the top of the cloud layer. These profiles compare well with the observed vertical thermodynamic structure shown in Fig. 1, which remained nearly constant throughout RF01 (Stevens et al. 2003b).

To obtain the observed boundary layer mean winds (U = 6 m s⁻¹, V = −4.25 m s⁻¹), vertically uniform geostrophic winds are specified to be U_g = 7 m s⁻¹ and V_g = −5.5 m s⁻¹, as are the initial winds. The surface pressure is taken as 1017.8 hPa.

Since observations showed a fairly uniform large-scale environment with a weak horizontal advection, the horizontal advective tendency of temperature and moisture are set to zero. The large-scale horizontal wind divergence is set to be D = 3.75 × 10⁻⁶ s⁻¹, which gives a large-scale subsidence rate of about −0.32 cm s⁻¹ at the cloud top. This subsidence rate was chosen to reconcile the thermal structure above the boundary layer with the deduced radiative fluxes and was in a good agreement with operational analyses for the study location and time period.

To minimize the different impact of radiative forcing arising from various radiative schemes, all participating SCMs and LESs used the following idealized formulation for calculating the net longwave radiative flux:

View Expanded

where

View Expanded

q_l is liquid water specific humidity, and ρ and C_p denote air density and isobaric specific heat of air. Following the specification of the LES intercomparison, the SCM boundary layer top z_i is defined at the height where q_t = 8 g kg⁻¹. The three terms on the left-hand side of Eq. (3) represent cloud-top radiative cooling, cloud-base radiative warming, and inversion-layer cooling effect, respectively. The parameters κ, F₀, and F₁ are chosen as 85 m² kg⁻¹, 70 W m⁻², and 22 W m⁻², respectively, to match radiative fluxes from a δ-four stream radiative code.

Table 2 summarizes the two principal types of simulation used for the SCM intercomparison. For both, all models were run with their shallow cumulus parameterization active. This is because turbulence schemes in several SCMs, such as RPN and NRL, include cumulus convection as a form of turbulence, so that it is not possible to switch off shallow convection in these models. The only exception was KNMI, which was run without its shallow convection scheme for all simulations.

Precipitation and deep convection schemes were switched off in models during the simulations. Again, there was one exception. In the ECMWF simulations, the deep convection scheme was turned on, but it was not triggered during these simulations; therefore, it basically does not have an effect on simulations. Both experiments A and B were nocturnal (no insolation). Experiment A was designed to compare the SCM simulations at high vertical resolution with LES results; thus, it closely follows the LES case specification of a 6-h simulation with a high vertical resolution of 10 m (and a time step of 5 s, unless otherwise noted), forced by a fixed surface sensible heat flux of 15 W m⁻² and a latent heat flux of 115 W m⁻². Experiment B was designed to test SCMs at their operational vertical resolution using a 48-h simulation. To reduce model drift, a fixed SST of 292.5 K is specified, so that the surface heat fluxes are determined interactively in experiment B rather than in experiment A.

The results of experiment B are also compared with an identically forced 48-h simulation using the University of California, Los Angeles (UCLA) LES. This LES is further described by STE. For the 48-h simulation, it was run with no subgrid-scale turbulence scheme applied to scalars and with a vertical grid spacing of 5 m between 775 and 1025 m, stretching to 10 m below this layer and continually stretched above, that is, the UCLA-0 configuration as described in STE. An experiment-A simulation with this configuration was quite consistent with observations (STE). Unfortunately, the inversion rose through the top of the refined grid around hour 40, leading to a spurious increase in entrainment and decrease in liquid water path and cloud cover in the next few hours. Hence, we will show UCLA-LES results only up through hour 36.

The sensitivity of the simulations to modeled precipitation processes and shallow convection was also explored by a number of groups.

Because of the limitation of model design, not every model can be configured to a specified high vertical resolution or to the requested sensitivity tests. Table 3 summarizes the numerical experiments executed by various SCMs. The NRL model was targeted for regional weather prediction and was not designed to run with an 1800-s time step. Therefore, it was only used for experiment A.

a. Experiment A

The time variation of the simulated total cloud cover fraction and vertically integrated LWP from experiment A are shown in Fig. 2, along with LES-predicted values and their standard deviations. STE lumped the LES simulations into LWP quartiles, and found that the highest-LWP quartile matched the observations much better than the other quartiles. Our measure of the LES simulation spread was chosen in part because the mean plus one standard deviation of the LWP for the entire group of LES simulations is a good proxy for the LES upper-quartile mean LWP. In these high-resolution simulations, all SCMs except KNMI and NASA-LaRC predicted 100% cloud cover, in agreement with observations. The KNMI and NASA-LaRC SCMs produced a cloud cover fraction as low as 80% within an hour. The rapid and unrealistic desiccation of the marine layer in the KNMI and NASA-LaRC simulations suggests overefficient entrainment in the presence of clouds in this case. Several LES models also predicted less than full cloud cover.

By the end of 6 simulation hours, the SCM-simulated LWPs span a quite large range, from about 12 to over 120 g m⁻². The large spread in LWP in part reflects the sensitivity of liquid water content to small changes in total humidity and temperature induced by the different turbulent transport and microphysics schemes employed in the SCMs. It should be pointed out that NCAR-CAM and CAM-UW use the same cloud microphysics, and yet produce significant difference in LWPs, indicating that different turbulent transport realized in models is a more important factor than microphysics for causing the large LWP spread in this case. The SCM-mean LWP, however, is comparable to the observational estimate of 60 g m⁻², which is slightly larger than the LES-mean LWP. Note that the range of LES-simulated LWPs is comparable to that of SCMs. Hence, as a group, LES models are still not accurate enough to usefully substitute for observations of MSC turbulence, entrainment, and cloud properties.

Figure 3 shows the profiles of θ_l, q_t, U, and V from experiment A averaged over the third and fourth simulation hours, the same period used for the LES intercomparison analysis. The θ_l and q_t profiles from the high-resolution SCM simulations and LESs agree well with each other and with the observed mixed-layer structure. The observed inversion-capped mixed-layer structure shown in Fig. 1 is well simulated by all the SCMs.

Compared to the LES mean, some SCMs tend to smooth the sharp jumps a little at the cloud top and have larger gradients in θ_l and q_t within the mixed layer. While these differences are subtle, they do affect LWP predicted by these SCMs for a given boundary layer mean height and thermodynamic properties. The SCM inversion structure depends on details of its turbulent mixing parameterization; some SCMs require an inversion transition layer several grid points deep, while others (e.g., with explicit entrainment parameterization) can produce inversions only a single grid layer thick.

One useful measure of internal thermodynamic gradients across the STBL, also discussed by STE, is the mixed-layer moisture difference δq_t defined as

View Expanded

Figure 4 shows a scatterplot of δq_t against LWP for hours 3–4 along with their mean and standard deviation from LES and SCM. Except for NCAR-CAM (which will be discussed later), large LWP is correlated to small δq_t (a particularly well mixed q_t profile), presumably because this promotes higher q_t within the cloud layer. Since all SCMs must pump the same surface moisture flux up through the boundary layer, eddy-diffusivity reasoning suggests that small δq_t (as observed) should be associated with vigorous turbulent mixing across the STBL.

STE showed that some LES models produced larger δq_t in association with marginal decoupling of the boundary layer turbulence into separate cloud and subcloud turbulent layers with a cloud-base minimum in vertical velocity variance. A similarly simple interpretation does not carry over to SCMs. Their diverse turbulence parameterizations allow δq_t to vary between models for many reasons, including formulations for turbulent length-scale, turbulence production in partly or fully cloudy conditions, turbulent transport, and nonlocal flux-gradient relations.

The horizontal wind profile shown in Fig. 3 evolves substantially in the boundary layer from its initial geostrophic specification and shows some spread across SCMs. Both SCM-mean U and V have a noticeable bias compared with the LES-mean wind profiles. This may reflect inadequate SCM treatment of momentum transport, as well as the non–steady state of boundary layer winds in SCM simulations. The wind profiles from all SCMs are still evolving substantially after 3 h, a time when most LES models have reached a quasi steady state. Note that the plotted mean U and V profiles do not include the CCCma-UBC results, which specified constant winds (equal to the geostrophic winds) throughout the simulation. Note that in this experiment, there is little potential impact of wind biases on thermodynamic fields since the surface fluxes are specified and the TKE production is dominated by buoyancy rather than shear.

Figure 5 shows the profiles of q_l averaged over 3–4 simulation hours, which vary substantially between models to create the LWP variations seen in Fig. 2. Like LWP, the q_l profile is sensitive to the small changes in θ_l and q_t caused by different turbulence schemes. In most SCMs, q_l increases with height up to a maximum very near cloud top similar to what is shown in observations (Fig. 1). One exception is the CCCma-UBC simulation, in which q_l reaches its maximum somewhere in the upper part of the cloud layer. This is associated with the technique used to calculate the nonlocal diffusivity, which employs relaxation toward a well-mixed reference state. Several SCMs predict a subadiabatic increase of q_l with height and a smaller maximum q_l than was observed. The NCAR-CAM model has a secondary q_l maximum near 400 m, which appears to be an artifact of its cloud microphysics scheme. In contrast, the ECMWF SCM produces a superadiabatic q_l gradient, generating the observed q_l maximum, but only half as thick as observations.

The simulated STBL evolution is heavily influenced by the boundary layer entrainment rate. For those models in which explicit entrainment parameterization is employed, the entrainment rate W_e is readily obtained, while for others, it must be inferred indirectly, for example, from the simulated inversion rise rate via

View Expanded

The simulated boundary layer top is diagnosed as the height at which q_t = 8 g kg⁻¹ based on a linearly interpolated q_t profile between grid levels. Its evolution in experiment A is plotted in Fig. 6.

Over simulation hours 2–6, SCMs predicted an inversion rise rate dz_i/dt in a range of 0.3–0.5 cm s⁻¹, which is very close to observations (Stevens et al. 2003b), and LESs (STE) near 0.4 cm s⁻¹. These small rise rates in most SCMs and LESs were consistent with the observed rough balance between entrainment and large-scale subsidence Dz_i ≈ 0.32 cm s⁻¹. Compared with observations and other SCMs, the ECMWF inversion rise rate is notably fast. This is most likely caused by its entrainment calculation, which is highly sensitive to numerical discretization around the inversion. This problem could be solved by using the method proposed by Lock (2001) and Grenier and Bretherton (2001).

Table 4 lists the resulting hours 2–6 mean entrainment rates deduced from (6), as well as the observational and LES estimates. The entrainment rates estimated from experiment B (to be discussed shortly) are also listed in the table. The consistency of W_e between most models and observations is comparable to that documented in Duynkerke et al. (2004) and is an improvement from the earlier GCSS study of Bechtold et al. (1996). This consistency in W_e may result from both careful case specification and parameterization improvements. The latter is evident from the consistency between the calculated W_e from Eq. (6) and the predicted rate from those models, such as METO, CAM-UW, and ECMWF, which employ the explicit entrainment parameterization (not shown here). In contrast to broadly consistent W_e, SCM-simulated LWPs show a large spread at hour 6. This partly reflects the rapid transient evolution of inversion structure and LWP during hour 1, when several models undergo up to a fourfold LWP decrease.

It may seem paradoxical that W_e varies much less between SCMs than LWP, given the central role of entrainment in the evolution of mixed-layer properties. One perspective considers the feedbacks between the entrainment rate, thermodynamic state of STBL, and cloud properties illustrated by Fig. 7. In each SCM simulation, the parameterized entrainment process is sensitive to cloud liquid water through its impact on turbulence via mixing-induced evaporative cooling and radiative fluxes. This is indicated in Fig. 7 as an arrow from LWP to W_e. Meanwhile, W_e slowly affects the evolution of the mixed-layer profiles of θ_l and q_t (left arrow in Fig. 7). Since even a small change in θ_l and q_t can result in a substantial variation in liquid water content (right arrow in Fig. 7), the LWP rapidly evolves toward an equilibrium value that sustains roughly enough entrainment to keep LWP at a nearly steady state. As discussed by Bretherton and Wyant (1997), this forces the STBL toward approximate moisture and energy balance between entrainment warming, radiative cooling, and surface fluxes. Since the surface fluxes are fixed, as is the radiative cooling as long as the cloud remains optically thick, all SCMs maintain similar entrainment warming, and hence similar equilibrium W_e, but with subtly different θ_l and q_t profiles that produce considerably different cloud properties. This energy balance and hence the equilibrium W_e can be changed if the cloud is excessively thinned (as in the KNMI model) or if there are large numerical errors in treating entrainment process at cloud top (Lenderink and Holtslag 2000), as appears to occur in the ECMWF SCM.

Some participating SCMs are capable of maintaining the persistent clouds observed in this case, while others are not, although the CTEI criterion is met in all simulations. For example, the KNMI and NASA-LaRC models simulate partial cloud breakup in response to a parameterized form of CTEI through inversion TKE– W_e feedback and the negative feedback via radiative forcing; that is, thinning clouds correspond to the decrease of longwave cooling and the reduction of buoyancy, which leads to a further desiccation of the marine layer.

b. Experiment B

The first focus of experiment B is the extent to which coarse operational resolution affects the simulation of bulk cloud layer properties such as the LWP, cloud cover fraction, and mean thermodynamic profiles. The second focus is the long-term evolution of the SCMs subject to constant forcings.

Figure 8 shows the time evolution of SCM-produced total cloud cover fraction and integrated LWP from experiment B, along with those predicted by UCLA-LES. Even in the first 6 h cloud fractions show a larger spread among SCMs than in experiment A. One might blame this on the interactive formulation of the surface fluxes in this experiment. However, the surface fluxes computed interactively agreed fairly well across SCMs (Fig. 9), though the SCM-predicted surface latent heat fluxes and sensible heat fluxes are generally larger and smaller, respectively, than those of UCLA-LES, and do drift over 48 h from the prescribed values of experiment A. Thus, most intermodel variability in cloud properties must be due to the differences between cloud and turbulence schemes and in particular their numerical implementations, aggravated by coarse resolution.

Some SCMs such as RPN, NASA-LaRC, and NCAR-CAM predicted a lower cloud cover fraction in this experiment. The RPN SCM, an extreme case, produced 100% cloud cover in the high-resolution simulation, but less than 40% within a hour of simulation at coarse resolution, perhaps a consequence of its statistical cloud scheme applied to a vertically underresolved cloud layer or its vertical cloud overlap assumptions. However, like UCLA-LES, the CCCma-UBC, CAM-UW, MPI, and METO SCMs are still able to produce approximately 100% cloud cover in the coarse-resolution simulations.

Most SCMs produced 3–6-h average LWPs in experiment B, broadly comparable to experiment A (Fig. 10). The most extreme exception is the NCAR-CAM simulation in which LWP drops fourfold compared to the high-resolution simulation. In section 3c, we show this results from the shallow convection scheme employed in the model. The NASA-LaRC and RPN models also have rather lower LWP in experiment B, consistent with their lower cloud cover fraction.

It is worth comparing LWPs in NCAR-CAM and CAM-UW simulations since the difference is solely resulted from turbulence and shallow convection schemes. Figures 2b and 8b show that CAM-UW generated an LWP close to 80 g m⁻² in the first 6-h simulations in both experiments A and B. Sensitivity tests shown in section 3c indicate that the shallow convection scheme (ZB; BZ) employed in CAM-UW has little effect on LWP. The insensitiveness of LWP to the model vertical resolution shown in experiments A and B implies the robustness of the CAM-UW moist turbulence scheme (Grenier and Bretherton 2001; BZ; ZB) in representing MSC cloud fields. In contrast, the NCAR-CAM simulations show a marked sensitivity to the model vertical resolution. LWPs drop more than sixfold from experiment A to experiment B. Sensitivity tests shown later further indicate that the NCAR-CAM shallow convection scheme (Hack 1994) plays a significant role in determining LWP, suggesting that the NCAR-CAM dry turbulence scheme (Holtslag and Boville 1993) alone is not sufficient in representing MSC cloud fields although it may predict a similar entrainment rate to that of CAM-UW.

Compared with Fig. 2b, the long-term evolution of LWPs from different SCMs shows even more variability, as seen in Fig. 8b. In particular, even though the SCM-mean LWP is similar between experiments A and B, in B the models split after 6 h into one group with consistently high LWP of 75–140 g m⁻² and another equally sized group with low LWP of 5–50 g m⁻². This might be caused by the setup of experiment B, which leads to long-term drifts of temperature above the inversion due to radiative-subsidence imbalance as evidenced in temperature profiles shown in Fig. 11. Therefore, SCMs cannot be run to a steady state under the setup of experiment B. It might be enlightening to formulate future SCM intercomparisons designed to focus on the predicted STBL properties in simulations run to a near-equilibrium state.

As in experiment A, the long-term variation in LWPs shown in experiment B (Fig. 8b) is barely correlated to the mean entrainment rates of hours 2–36 listed in Table 4. For example, in the low LWP group, NASA-LaRC predicted a larger W_e as one would expect, but RPN and KNMI produced a smaller W_e. Comparisons between hours 2–6 entrainment rates further indicate that there is little correlation between the W_e predicted by a high-resolution simulation and its operational-resolution counterpart. For instance, the ECMWF SCM has the lowest W_e at operational resolution despite having the highest W_e at high resolution. This corroborates the perspective of Fig. 7 that the LWP adjusts in a model- and resolution-dependent fashion to a strong model-independent energy-balance constraint on W_e.

Figure 11 shows the mean profiles of θ_l, q_t, and q_l at 3–4 and 35–36 h predicted by various SCMs from experiment B compared with those of UCLA-LES. At 3–4 h, all SCM simulations and LES still reflect the initial θ_l and q_t in the mixed layer. The inversion-layer structure and q_l profiles are no longer well resolved by SCMs, and differ between models due to the different model grid levels and the mixing parameterizations. The coarse-resolution models produce a lower peak of q_l compared with the high-resolution simulations, but generally spread q_l over a broader height range to achieve LWP comparable to the high-resolution simulations shown in Fig. 10. Overall, the mixed-layer thermodynamic structure predicted by the LES is also simulated by most SCMs, though with larger vertical q_t gradients.

Figures 11d–f show that after 35 h, the simulated profiles, especially q_l, differ more between SCMs. More surprisingly, the LES cloud top (1025 m) is actually above the range of the SCM predictions. This implies a slightly stronger entrainment rate in the LES than in the SCMs, although its LWP was above the SCM median value shown in Fig. 8. This LWP could be sustained in the LES despite the dry boundary layer associated with this stronger entrainment, since the LES generated a sharp-topped adiabatic cloud layer and a more precisely well mixed q_t profile than most SCMs.

Although all SCMs maintain a fairly well mixed boundary layer below 600 m, there is some spread in the mean mixed-layer moisture, which may be attributed partially to the different surface moisture fluxes shown in Fig. 9, but mainly to different rates of entrainment drying. This is evident from the averaged entrainment rates of hours 2–36 listed in Table 4, which have a larger range than those of hours 2–6. For example, the boundary layer deepened much faster and stayed much drier in the NASA-LaRC simulation than in the CCCma-UBC simulation despite the similar surface fluxes. The LES, with strong entrainment, produces the driest mixed-layer moisture despite its relatively high surface moisture flux. However, there is still little correlation between the nearly 36-h-averaged entrainment rate and the hours 35–36 LWP. A more vigorously entraining STBL will have a higher cloud base because it is drier, but also a higher cloud top. Instead, as in the high-resolution simulations, each SCM adopts a different cloud and inversion structure to maintain roughly the same energy-balancing W_e. For instance, the RPN simulation has a slightly lower than average entrainment rate, yet the smallest hours 35–36 q_l and LWP of all models. Figures 11d–f suggest that this is associated with a “rounded-off” inversion zone that fills two contiguous layers, in which cloud could otherwise form, coupled to a turbulence scheme that allows this STBL structure to entrain vigorously despite the dearth of clouds.

c. Sensitivity to precipitation and shallow convection

As marine stratocumulus thickens, drizzle may also become important in the water budgets and turbulent dynamics through latent heating in the cloud layer and subcloud evaporative cooling. The RF01 intercomparison was chosen in part because of the observed lack of drizzle, but it is nevertheless possible that in some SCMs drizzle may significantly affect the simulated LWP and STBL structure. Because of the difficulties in separating precipitation from other processes in some models, only NASA-LaRC, CCCma-UBC, NCAR-CAM, CAM-UW, KNMI, and ECMWF SCMs executed the precipitation sensitivity tests at both high and coarse resolutions, while the MPI SCM did only the coarse resolution sensitivity test. For all SCMs, in experiment A, the simulated thermodynamic structure of STBL and its associated cloud fields are barely affected by parameterized precipitation. In experiment B, no significant change in cloud cover in response to precipitation was simulated by any SCMs, but a weak influence of precipitation on LWP was found in some SCMs. Table 5 lists the LWP with and without parameterized precipitation microphysics averaged over 40–48 h from experiment B. The CCCma-UBC, NCAR-CAM, CAM-UW, MPI, and ECMWF simulations predicted small precipitation fluxes that reduced LWP between 6% and 28%, while in the NASA-LaRC and KNMI simulations, no precipitation was simulated at all.

The weak influence of simulated precipitation on MSC cloud properties shown in this case appears to conflict with other simulations (e.g., Wang et al. 1993) that have found larger effects of suppressing precipitation on MSC cloud cover. This is to be expected since our simulations are based on a nonprecipitating case. An upcoming GCSS WG1 intercomparison will focus on an STBL in which substantial precipitation was observed.

We also performed a sensitivity study of the effects of parameterized shallow cumulus convection on the STBL simulation. Because of the diverse parameterizations for convective triggering, the shallow cumulus convection scheme in a SCM may activate in regimes where there is no shallow convection in reality. Among all the participating SCMs, CCCma-UBC, METO, NCAR-CAM, CAM-UW, and ECMWF executed this sensitivity test. In both high and coarse resolution, the simulations by the METO, CAM-UW, and ECMWF SCMs show little sensitivity to their shallow convection schemes. The other two models, NCAR-CAM and CCCma-UBC, also show an insignificant influence of the shallow convection scheme on the STBL in the high-resolution simulations, but a substantial influence of the convection scheme is found in their coarse-resolution simulations, as shown in Fig. 12. The NCAR-CAM simulations show an enormous sensitivity to its shallow convection scheme. Without the shallow convection scheme, the coarse-resolution NCAR-CAM produces an unrealistically thick cloud layer almost touching the ground, indicating that the boundary layer scheme does not generate sufficient turbulence to transport moisture upward efficiently. We speculate that the turbulent mixing scheme used in NCAR-CAM tends to produce a colder than realistic cloud layer with larger LWP and smaller entrainment rate. Since this bias disappears at high resolution, it might be addressable by reformulating the NCAR-CAM boundary layer height diagnosis. In contrast, turning on the shallow convection scheme increases LWP about 50% in the CCCma-UBC simulations. One possible reason for this increase of LWP suggested by Lenderink and Holtslag (2004) is that the convection scheme tends to produce liquid water through its detrainment process at the top of clouds.

The appropriate coupling between the boundary layer scheme and shallow convection scheme is not just an issue for correctly simulating the STBL. It becomes more important in the transition from stratus deck regime to trade wind cumulus regime where the interaction between stratiform and convective clouds is the key to determine the thermodynamic structure, cloud cover, and air–sea fluxes in this region.