3.1. Characteristics of Clouds Observed From M-PACE
 Various types of clouds that often occur in the Arctic during its transition season were observed in the M-PACE field experiment. Figure 1a shows the time-pressure cross section of observed frequency of occurrence of clouds at Barrow by integrating measurements from the ARM cloud radar and other sensors using the ARSCL (Active Remotely Sensed Clouds Locations) algorithm [Clothiaux et al., 2000]. These data are originally at 10-s and 45-m time and height intervals. They are averaged to 3-h and 25 hPa intervals to better represent clouds over a large-scale general circulation model (GCM) grid box, which usually represents an area of 200 km × 200 km. One issue with the ARSCL clouds is that cloud radar tends to underestimate the cloud top heights for high-altitude clouds because it will not be able to detect cloud particles if they sufficiently small. Another issue is that cloud radar detected cloud base can be contaminated with ice precipitation. To reduce this problem, we use the ARM laser ceilometer and micropulse lidar measurements, which are usually insensitive to ice precipitation (if the concentration of precipitation particles is not sufficiently large) or clutter, to determine the cloud base. As indicated by Clothiaux et al. , the laser ceilometers and micropulse lidar can provide quite accurate cloud base measurements.
Figure 1. Time-height cross sections of (a) ARSCL cloud frequency and modeled cloud fraction (b) CAM3, (c) AM2, and (d) CAM3LIU at Barrow during M-PACE. The unit is %.
Download figure to PowerPoint
 Even with these uncertainties, the cloud radar and other remote sensors provide extremely valuable information about the vertical distribution of various types of clouds over the ARM barrow site. During M-PACE, the ARSCL data indicated that Barrow was covered with multilayered stratus clouds in the midlevels and low levels with the cloud top up to 550 hPa for 5–8 October, persistent single-layer boundary layer stratocumulus with the cloud top around 850 hPa during the period from 8 to 14 October, and deep prefrontal and frontal clouds (including cirrus) from 15 to 22 October.
 The observed cloud systems were largely controlled by the synoptic-scale circulation affecting that area during M-PACE. As described by Verlinde et al. , for the period from 5 to 15 October, the North Slope of Alaska (NSA) was dominated by a strong surface high-pressure system built over the pack ice to the northeast of the Alaska coast. Associated with the strong surface high, east-northeasterly flow prevailed at low levels. The low-level northeasterly flow combined with a midlevel low-pressure system drifted along the northern Alaska coast generated the complicated multilayer cloud structure over NSA from 5 to 7 October. The single-layer low-level clouds observed from 8 to 15 October originally formed over the ocean adjacent to the Alaskan coast as the low-level east-northeasterly flow brought cold near-surface air from the pack ice to the warm ocean and then advected to Barrow. Large surface turbulent fluxes are the major driver for the evolution of the boundary layer clouds. During this period, there was a substantial temperature decrease at altitudes below the 665 hPa pressure level and a sharp moisture decrease over the Barrow site. The range of cloud temperature was from −5°C to −20°C, indicating that the cloud condensate was mixed phase. After 14 October, the boundary layer clouds started to disappear as a warm front moved through the area on 15–16 October and a deep ridge moved over the NSA. Southwesterly flow prevailed in the entire troposphere except on late 19 October when there was an abrupt wind direction change from the southwest to the southeast associated with a strong warm frontal passage which brought in deep prefrontal and frontal clouds. Cirrus clouds were seen during this period.
 To obtain in situ and remote sensing measurements of microphysical properties of these cloud systems, the ARM millimeter cloud radar, micropulse lidars, laser ceilometers, and two instrumented aircraft were used in the experiment. For the single-layer boundary layer clouds, data collected from both the surface-based remote sensing instruments and the aircraft revealed the presence of a maximum liquid water layer near cloud top and liquid and irregular ice crystals within the cloud layer with precipitating ice beneath the liquid cloud base [McFarquhar et al., 2007]. This result is consistent with the findings from other arctic field campaigns [Pinto, 1998; Hobbs and Rangno, 1998; Curry et al., 2000]. The multilayered clouds had a more complicated structure than the single-layer clouds. Up to six liquid cloud layers were detected by the ARM narrow-band lidar and the depth of individual liquid cloud layers varied from 50 to 300 m. Combined radar and lidar data indicated the existence of precipitating ice with low ice water content between the layers. These characteristics are similar to those from the in situ measurements by the aircraft. A detailed summary of the observed clouds during M-PACE is given by Verlinde et al.  and McFarquhar et al. . In the following discussion, we examine how well CAM3 and AM2 capture these observed features in the arctic clouds.
3.2. Model-Simulated Clouds
 Figures 1b, 1c, and 1d show the model-produced cloud fraction at Barrow from CAM3, AM2, and the CAM3 with the new ice microphysics described in section 2 (hereafter CAM3LIU), respectively. It should be noted that model clouds represent a fraction of a model grid box occupied by clouds, which is different from the radar and lidar detected frequency of occurrence of clouds as shown in Figure 1a. Cloud fractions in CAM3 and AM2 have to be parameterized in terms of the large-scale variables such as grid mean relative humidity because they cannot be resolved by CAM3 and AM2 with their current spatial resolutions. There is always a concern about the comparison between the model clouds and the single point radar and lidar measurements. Averaging the ARSCL clouds from the 10-s and 45-m time and height intervals onto the 3-h and 25 hPa intervals improves the representation clouds over a GCM grid box, especially for the highly horizontally advective boundary layer clouds and frontal clouds observed during M-PACE. Nevertheless, it is still difficult to quantitatively compare model clouds with the ARSCL data. So the purpose here is to qualitatively evaluate the model clouds using the available ARM radar and lidar data and demonstrate intermodel differences in their simulated clouds.
 Figure 1 shows that all the models are able to qualitatively reproduce the cloud types observed during M-PACE, such as the multilayered clouds from 5 to 8 October, the boundary layer clouds from 8 to 14 October, and the frontal deep high clouds from 15 to 22 October. However, there are considerable differences in detailed structures of the clouds between the observations and the model simulations. For the period 5 to 14 October, the default CAM3 substantially underestimates the observed multilayered and single-layer boundary layer cloud fraction. In contrast, AM2 produces much more midlevel and low-level cloud fraction than CAM3. It is interesting to see that the CAM3 with the new ice microphysics produces more realistic single-layer boundary clouds than the default CAM3 while it generates too many midlevel and high-level clouds. The overestimation of midlevel and high-level clouds is partially related to the scheme's allowance of ice supersaturation. As discussed earlier, CAM3 uses a RH-based cloud scheme to diagnose stratiform cloud fraction (equation (2)) and its RH is determined by a combination of ice and water saturation. Given the same threshold RHmin, the new scheme would lead to more cloud fraction than the default CAM3 because of the allowance of ice supersaturation. We have found that the RH in CAM3LIU is often supersaturated with respect to ice in the midlevels and high levels where temperature is usually less than −20°C during M-PACE. One common problem for all the models is that the modeled cloud top and cloud base are lower than the observed for the period 8–15 October. The averaged cloud [top, base] pressures over this period for ARSCL, CAM3, CAM3LIU, and AM2 are [840, 939], [855, 985], [851, 991], and [865, 1006] (hPa), respectively. This may be partially related to the coarse vertical resolutions used in these models, which cannot well resolve the observed boundary layer structure. For example, CAM3 only has four model levels below 850 hPa, the level of the observed single-layer boundary layer cloud top. For the deep frontal clouds, the models tend to overestimate the clouds at high levels and underestimate them at midlevels and low levels. The problem with the midlevel and low-level clouds is particularly severe for the CAM models. In addition, the model-simulated frontal clouds tend to have a longer lifetime and weaker temporal variability than the observed. This is a common problem for most large-scale models in simulating frontal clouds [e.g., Klein and Jakob, 1999; Zhang et al., 2005; Xie et al., 2005]. The temporal variability in the observed frontal clouds is partially related to subgrid-scale dynamics which cannot be resolved in large-scale models. The difference in temporal variability between the models and observations may also be due to the fact that the ARM observations are from a point whereas the models are grid box averaged.
 Figure 2 compares the total cloud fraction from the models and the total cloud frequency from the observations at Barrow. The observed total cloud frequency is calculated from the ARSCL products assuming maximum cloud overlap over a 3-h interval. The observations typically showed a persistent almost 100% cloud cover during the period 5–14 October except on 7–8 and 11 October where the cloud cover decreased slightly. Consistent with earlier discussions, CAM3 considerably underestimates the total cloud cover for this period. This problem is significantly reduced in CAM3LIU when the new physically based ice microphysical scheme is used. AM2 also produces a much better cloud cover than the default CAM3. It is seen that the cloud fraction produced by the default CAM3 shows larger temporal variability than the observations, indicating the inability to produce clouds under the same conditions as nature as the conditions change. In contrast, CAM3LIU and AM2 have 100% cloud cover for most of the period 5–14 October, similar to the observations. For the deep frontal clouds, both CAM3LIU and AM2 largely overestimate the observed cloud fraction while CAM3 generally agrees well with the observation.
Figure 2. Time series of the total ARSCL cloud frequency and modeled cloud fraction. Black line with dots is for ARSCL, red line is for CAM3, green is for CAMLIU, and blue is for AM2. The unit is %.
Download figure to PowerPoint
 Figures 3a–3c show the grid box mean liquid water mixing ratio (LWC) produced from these models. The contour lines in Figure 3 are the model produced temperatures. All the models are able to produce two or more liquid cloud layers for the period 5–8 October even though the fine vertical structures of the observed multilayer clouds as shown by Verlinde et al.  are not well simulated because of the coarse model vertical resolution. In comparison with CAM3LIU, CAM3 predicts similar amount of cloud liquid water for the boundary clouds even though its cloud fraction is much lower. This is partially due to its temperature-dependent liquid/ice partitioning. For the range of temperature −5°C ∼ −20°C, the majority of cloud condensate produced in CAM3 will be liquid. Another noteworthy feature is that CAM3 has much more liquid in the midlevel and upper level clouds than both CAM3LIU and AM2, which leads to a considerable overestimation of the observed liquid water path in CAM3 during these periods. It is noted that AM2-produced clouds contain much less liquid than CAM3LIU for the mixed-phase boundary clouds although they produce comparable cloud fraction and include the Wegener-Bergeron-Findeisen microphysical process. This suggests a faster conversion rate of liquid to ice in AM2 than CAM3LIU, which should be partially related to the differences in specifying the vapor saturation and the cloud ice number concentration between these two models as discussed in section 2.2.
Figure 3. Time-height cross sections of model-produced liquid water mixing ratio (mg/kg). (a) CAM3, (b) AM2, and (c) CAM3LIU. The solid lines are model-simulated temperatures.
Download figure to PowerPoint
 Figure 4 is the same as Figure 3 except for ice water mixing ratio. Since there is no distinction between ice and snow inside the cloud for AM2 (i.e., AM2 ice includes snow inside the cloud) but for CAM3 there is, we add model snowfield to the ice water mixing ratio in CAM3 and CAM3LIU for a better comparison with AM2. For simplicity, we use “ice” to represent the sum of ice and snow in our following discussions. It should be noted that the snow in CAM3 and CAM3LIU has no impact on radiation while the snow inside cloud in AM2 affects model radiation since it is treated as ice. Compared to CAM3 and CAM3LIU, AM2 produces less ice for boundary layer clouds and near the surface partially because of the fact that the snow falling out of clouds is not included in Figure 4b while it produces significantly larger ice in the strong frontal clouds on 19 October. Generally, CAM3LIU generates more ice than the default CAM3, especially for the boundary layer mixed-phase clouds.
 Figures 5a and 5b show the observed and modeled cloud liquid water path (LWP) and cloud ice water path (IWP) at Barrow, respectively. Note that both observed and modeled IWPs include snow component since the observations cannot separate snow from ice. There are two sources for the observed LWP. Both are based on the ARM Microwave Radiometer (MWR) measurements but they are retrieved using different retrieval algorithms. One is based on the algorithm described by Turner et al.  and another one is derived using Wang . The observed IWP is derived from the ARM cloud radar and lidar measurements [Wang and Sassen, 2002]. The remote sensing retrieved IWP is currently available for the single-layer boundary layer mixed-phase clouds with an estimated error of about 50%. The uncertainty in the LWP retrieved using the algorithm from Turner et al.  is about 15–25 g m−2 for clouds of any liquid water path and it is about 6 g m−2 based on the uncertainty estimated in the Wang  retrieved LWP for clouds with LWP up to 40 g m−2 (there is no uncertainty estimate for Wang's data for the M-PACE period where cloud LWP is usually larger than 100 g m−2). It is seen that the LWPs from these two measurements agree with each other very well for the period when the radar and lidar retrievals are available. CAM3 reasonably reproduces the observed LWP for the single-layer mixed phase clouds even though its cloud amount is significantly smaller than the observations. This inconsistency between LWP and cloud fraction in CAM3 is due to the fact that CAM3 cloud fraction is determined by its large-scale relative humidity rather than its cloud condensate. One clear problem with the default CAM3 is that it largely overestimates the observed LWP for the midlevel and high-level clouds (e.g., 7, 16, and 18–20 October). This problem is significantly reduced with the use of the new ice microphysical scheme as shown in CAM3LIU, which also predicts a reasonable LWP for the boundary layer clouds. Consistent with earlier discussion, the LWP in AM2 is considerably smaller than the CAM models and the observations for the boundary layer clouds, suggesting the conversion rate of liquid to ice might be too fast in AM2. However, it is surprising to see that the single-layer boundary layer clouds produced by AM2 do not have much ice either. Further sensitivity tests with its microphysical scheme are needed to fully understand this inconsistency. For the frontal clouds occurring during 15–22 October, the IWPs simulated by CAM3 and CAM3LIU agree with each very well while AM2 produces significantly larger IWP than CAM3 for the strong deep frontal clouds on 19 October, which suggests more rapid glaciations occurred in AM2 than the CAM models for the deep frontal clouds.
Figure 5. Time series of the observed and model-produced (a) cloud liquid water path (g/m2) and (b) ice water path (g/m2) during M-PACE. The black solid line with dots is from Turner's retrievals, and the plus is from Wang's retrievals. Red lines are for CAM3, green lines are for CAMLIU, and blue lines are for AM2.
Download figure to PowerPoint
 To better understand the large differences in the simulated cloud fields among these models, we examine the model simulated surface turbulent fluxes, which are largely responsible for driving the evolution of the boundary layer clouds for the period 9–14 October. Since the clouds observed at Barrow originally advected from the ocean adjacent to the Alaskan coast, we examine the model results at an upwind model grid point. The location of this selected upwind model grid point is (152.5°W, 72°N) for CAM3 and CAM3LIU and (153.75°W, 73°N) for AM2. Over 9–14 October, AM2 has slightly larger sensible and latent heat fluxes than CAM3 and CAM3LIU. The average sensible heat flux over this period is 146 W m−2 for AM2, 123 W m−2 for CAM3, and 129 W m−2 for CAM3LIU and the average latent heat flux is 99 W m−2 for AM2, 87 W m−2 for CAM3, and 85 W m−2 for CAM3LIU. The slightly larger hear fluxes in AM2 might partially lead to a larger cloud fraction compared to the default CAM3. However, the large differences shown in the CAM3 and CAM3LIU produced cloud fraction and cloud properties cannot be easily explained by their surface turbulent fluxes, which are very similar for these two models. This suggests that the differences shown in the CAM3 and CAM3LIU simulated clouds are mainly due to the different microphysical parameterizations used in these two models. Over the period, AM2 also produces a slightly larger surface precipitation rates than CAM3 and CAMLIU. The average surface precipitation rates at Barrow are 0.7 mm d−1 for AM2, 0.43 mm d−1 for CAM3, and 0.42 mm d−1 for CAM3LIU.
3.3. Microphysical Properties in the Single-Layer Mixed-Phase Clouds: Model Versus Aircraft Data
 During M-PACE, there were four flights conducted on 9–12 October to obtain cloud properties in the single-layer boundary layer mixed-phase clouds. Each flight lasted 1 or 2 h with cloud data collected every 10 s. While these in situ aircraft data provided unique information to understand the microphysical properties in the mixed-phase clouds, it is difficult to use them to quantitatively compare with model results because the mismatches between them. For example, the model outputs are at a much lower temporal and spatial resolution (representing a mean over 3 h and an area of 200 km × 200 km) than the aircraft measurements (10 s, point measurements). So our purpose here is to see if these models can reproduce qualitatively well some important statistical features revealed by the aircraft data. In Figures 6 and 7, a cloud is defined when the total cloud condensate is larger than 0.001 g m−3 for both model results and in situ measurements. To improve statistics, the model data used in Figures 6 and 7 are for the entire period from 9 to 14 October when the single layer boundary layer clouds are generated in these models.
Figure 6. Liquid fraction as a function of cloud height. (a) UND citation data, (b) CAM3, (c) AM2, and (d) CAM3LIU. Different symbols in Figure 6a represent data collected from different flights. Note that the cloud altitude in the figure is normalized from 0 at cloud base to 1 at cloud top.
Download figure to PowerPoint
Figure 7. Liquid fraction as a function of temperature. (a) UND citation data, (b) CAM3, (c) AM2, and (d) CAM3LIU. Different symbols in Figure 7a represent data collected from different flights.
Download figure to PowerPoint
 Figure 6a displays the liquid fraction (fl) in the total cloud condensate as a function of height measured by the University of North Dakota (UND) Citation from the four flights conducted on 9–12 October for the single-layer mixed-phase clouds. The 10 s raw aircraft data were processed by McFarquhar et al. . The cloud altitude is normalized from 0 at liquid cloud base to 1 at cloud top. The aircraft data revealed the dominance of cloud liquid water in the boundary layer mixed-phase clouds with 79% of cases having fl > 90%. In general, fl increases with height and is larger than 80% near cloud top. It is important to notice that many data points with low fl are found in the lower half of the cloud, indicating the presence of significant amounts of ice. The strong liquid layer near cloud top leads to strong cloud top radiative cooling, which may play an important role in maintaining the persistence of mixed-phase boundary layer clouds [e.g., Pinto, 1998].
 Figures 6b–6d are the same as Figure 6a except for CAM3, AM2, and CAM3LIU, respectively. The snow component is added to the total cloud condensate when the modeled liquid fraction is calculated in order to be consistent with the aircraft measurements processed by McFarquhar et al. , which include all particles greater than 53 μm. This observed vertical distribution of fl is clearly not reproduced by CAM3 in which fl generally decreases with height because of its temperature dependence. The few points with low fl found at the cloud base in Figure 6b are due to the model-produced snow. In contrast, the observed variation of liquid water fraction with cloud height is reasonably captured by CAM3LIU. AM2 also shows a better agreement with the observations than CAM3. The lack of low fl points near the cloud base in AM2 is probably due to the fact that the snow falling out of the cloud is not included in the AM2 total cloud condensate when fl is calculated.
 Figure 7a shows the measured fl as a function of temperature from the same flights as Figure 6a. The measured cloud temperatures during these flights are about between −16°C and −9°C. It is seen that there is no clear relationship between fl and temperature in the observations. Significant amounts of liquid and ice coexist within this temperature range. It is obvious that any temperature based liquid/ice partitioning schemes will fail to reproduce the observed structure, such as the scheme used in CAM3 (see Figure 7b). Once again, AM2 and CAM3LIU reasonably reproduce the observed variation with temperature of fl by including the Wegener-Bergeron-Findeisen process (Figures 7c and 7d). This indicates that the Wegener-Bergeron-Findeisen process is critical for the models to correctly capture observed structure of cloud condensate in the mixed-phase clouds.
 Clouds have a large impact on surface radiation. However, it is difficult to evaluate model shortwave radiation (SW) with point measurements taken at a station located near the coast (e.g., Barrow). The closest CAM and AM2 model output grid points to the Barrow site cover both ocean and land areas, over which the surface characteristics are very different. For example, there is a very strong contrast in the surface albedo between ocean and land. During M-PACE, the ARM Barrow site was covered by snow with the surface albedo in a range of 0.7 to 0.9 [Xie et al., 2006] while its nearby ocean was open water, which had much smaller surface albedo (less than 0.2). The difference in the surface albedo between the models and the observations makes it difficult to interpret model-observation comparison since surface albedo has a large impact on both the surface upward and downward radiation, in addition to clouds. Thus, in this study we will focus our discussion on the surface downward longwave radiation and the top of the atmosphere (TOA) outgoing longwave radiation, which are more related to clouds and less dependent on surface conditions. Moreover, longwave radiative fluxes are the dominant terms in the surface and TOA energy budgets in the cold Arctic season.
 Figure 8a displays the observed and modeled downward longwave radiative fluxes (LW) at surface. The observed surface radiation data are obtained from the ARM Solar and Infrared Radiation Station. For the period 5–14 October, the observed surface downward LW shows a rather weak temporal variability due to the presence of persistent low-level clouds. The observed value is significantly underestimated by CAM3, due primarily to its underestimation of the low-level clouds as shown in Figures 1 and 2. In addition, CAM3 shows much larger temporal variation in the surface downward LW than the observations, consistent with the larger temporal variation in its produced cloud cover (Figure 2). These problems are largely reduced in CAM3LIU, which only slightly overestimates the observations for the period 10–14 October. The overestimation may be related to the lower cloud base altitude in CAM3LIU. AM2 also shows a better simulation of the surface downward LW than CAM3. Its surface downward LW agrees well with the observations for most of the period while it significantly underestimates the observations on days 9, 13, and 14 associated with the problem with its simulated cloud field. The averaged surface downward LW fluxes over the period 5–14 October are 284, 264, 291, and 278 (W/m2) for the observations, CAM3, CAM3LIU, and AM2, respectively. For the period 15–22 October, all the models generally overestimate the observed surface downward LW, partially because of the longer lifetime for the frontal clouds simulated by these models.
Figure 8. Time series of the observed and model-produced (a) surface downwelling longwave radiative fluxes (W/m2) and (b) TOA outgoing longwave radiative fluxes (W/m2). Black lines are observations. Red lines are for CAM3, green lines are for CAMLIU, and blue lines are for AM2.
Download figure to PowerPoint
 Figure 8b is the same as Figure 8a except for the outgoing longwave radiative fluxes (OLR) at top of the atmosphere. The observed TOA radiative fluxes are from the 1° × 1° analysis of the NASA Terra and NOAA 16 satellite measurements. All the models consistently overestimate the observed OLR in the presence of the single-layer boundary layer clouds (9–14 October). This is related to the underestimation of the cloud fraction and cloud liquid water path during this period as discussed earlier. The model underestimation of the low-level cloud top altitude may also contribute to this problem. Compared to CAM3, the overestimation is largely reduced in CAM3LIU. It is seen that CAM3LIU considerably underestimates the observed OLR on day 7 when the multilayered clouds occurred. This is manly because CAM3LIU clouds extend to much higher altitude (300 hPa) than the observed (∼550 hPa) (see Figure 1). For the deep frontal period, the smaller OLR produced by the models on 15–16 October and 17–18 October is consistent with the higher frontal cloud fraction generated by these models compared to the observations.