Notice: Wiley Online Library will be unavailable on Saturday 27th February from 09:00-14:00 GMT / 04:00-09:00 EST / 17:00-22:00 SGT for essential maintenance. Apologies for the inconvenience.
 This study uses a regime sorting technique to explore the relationships that ACCESS1.3 clouds have with the large-scale environment. Satellite simulator output is used to demonstrate that the modeled clouds have similar sensitivity to the large-scale dynamic and thermodynamic conditions as shown by CloudSat and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO). The high cloud cover and longwave cloud radiative effect is represented very well in the model across all regimes. The cloud types that the model simulates the most poorly are stratocumulus over cool sea surface temperatures (SSTs) and the deep convective regimes associated with strong upward midtropospheric vertical velocity and weak lower tropospheric stabilities. The reflectance of the deep convective regimes shows a stronger sensitivity to SST and less dependence on the large-scale dynamics than the observations. Many of the model errors identified occur across all regimes, such as the underestimate of clouds with large scattering ratios (SR) and the too frequent occurrence of drizzle and rain. A sensitivity test in which a different warm rain scheme was used shows that the modelled frequency of occurrence of nonprecipitating low cloud is quite sensitive to the autoconversion parameterization. The new scheme produced more cloud with large SR and higher cloud tops in better agreement with the observations. The thermodynamic regime analysis shows that the transition of shallow to deeper convection in the model requires a warmer SST and weaker LTS than the observations. The significant underestimate of cumulus congestus is likely to contribute to this delay due to the role these clouds have in preconditioning the midtroposphere for the onset of deep convection.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 The redistribution of energy from the tropics to the poles drives global weather and climate. Clouds play a key role in this redistribution through their ability to modulate the energy budget, and, therefore, their accurate representation in models is important for correctly simulating both the tropical and global climate. The ability to represent cloud processes in numerical models of the atmosphere is notoriously difficult even though their importance in climate science is well recognized [e.g., Arakawa, 1975]. The reasons for this include the vast range of spatial and temporal scales that are involved with cloud processes, as well as their interactions with dynamics, turbulence, radiative transfer, chemical, and surface processes. It is due to these complexities that cloud feedbacks remain the largest source of uncertainty regarding the spread in climate sensitivity [Randall et al., 2007]. Therefore, evaluating the sensitivity of modeled clouds to changes in the large-scale dynamics and thermodynamics is an important undertaking in order to identify model inadequacies, improve atmospheric physical parameterizations, and reduce uncertainty in climate projections.
 In the tropics, the large-scale atmospheric circulation is intimately tied to the sea surface temperatures (SSTs) and the cloud regimes are closely coupled to these forcings. The upward motion of the large-scale circulation exists over warm SSTs, and it is here that deep convective cloud systems are found, while in the subsiding branch of the circulation, extensive boundary layer cumulus and stratocumulus occur over cooler SSTs [e.g., Stevens, 2005]. A methodology that enables a way to understand more clearly the relationships between clouds and their environment is a regime-sorting approach, whereby tropical cloud properties are sorted into dynamic and thermodynamic regimes [e.g., Bony et al., 2004; Bony and Dufresne, 2005; Williams et al., 2003; Ringer and Allan, 2004; Wyant et al., 2006; Su et al., 2008]. Bony et al.  used the 500 hPa vertical motion (ω500) as a proxy for the large-scale circulation regimes to examine the cloud response to climate perturbations. Williams et al.  used both ω500 and SST to composite cloud properties into dynamic and thermodynamic regimes and illustrated the benefit of including SST to delineate more clearly between the boundary layer cloud types. Su et al.  compared the regime-sorted cloud water content observed by CloudSat to models using a number of different dynamic and thermodynamic quantities. Using the lower tropospheric stability (LTS, defined by Klein and Hartmann  as the difference in potential temperature between 1000 and 700 hPa), they showed the advantage of compositing with respect to LTS in addition to SST, in order to distinguish further the low cloud regime transitions. The strength of the regime-sorting methodology is the ability to evaluate the model's representation of the important cloud regimes in the tropics and the transitions between them. Given the importance of tropical and subtropical marine boundary layer clouds to the spread in climate sensitivity [Bony and Dufresne, 2005], understanding the representation and large-scale controls on the transitions between cloud regimes is essential for improving the understanding of modeled cloud-climate feedbacks.
 In this work, we composite cloud properties from satellite simulators running online in the Australian Community Climate and Earth System Simulator (ACCESS1.3) using both dynamic and thermodynamic regimes in the tropics to evaluate the performance of the model in simulating tropical oceanic clouds and the transitions between them. The transition from stratocumulus to trade cumulus is recognized to be an important component in determining the tropical cloud-climate feedback due to the significant change in the cloud radiative forcing between the strong shortwave cloud forcing of the stratocumulus clouds to the much smaller cloud fraction and weaker cloud forcing associated with trade cumulus [Bony and Dufresne, 2005]. In this study, we take advantage of the vertical cloud information available from CloudSat and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) to not only evaluate the model, but also to use the observations and novel techniques to provide insight into the cloud properties that are observed in these regimes.
 A description of the model, observations, and method used in this study is given in section 2. Section 3 evaluates separately the dynamic and thermodynamic regime sorted cloud properties, and section 4 examines the vertical structure of the composited cloud properties from the radar and lidar simulators and observations. The impact of a using different warm rain microphysics scheme in the model is explored in section 5, and the conclusions of this study are summarized in section 6.
2 Description of Model, Observations, and Method
 The Australian Community Climate and Earth System Simulator, ACCESS, is a new modeling system that is used for numerical weather prediction and studies of the climate system. In this study, we evaluate the tropical cloud properties of ACCESS1.3, a model version that has contributed to the Coupled Model Intercomparison Project CMIP5. The atmospheric component of ACCESS1.3 is the UK Met Office Unified Model (UM) Global Atmosphere (GA) 1.0 [Hewitt et al., 2011], with modifications to the cloud [Boutle and Morcrette, 2010; Franklin et al., 2012] and radiation schemes [Shonk et al., 2010] and the inclusion of the CSIRO Atmosphere Biosphere Land Exchange model [Wang et al., 2011]. A description of the ACCESS1.3 model configuration and atmospheric parameterizations is given in Bi et al.  and summarized in Franklin et al. . The model has a horizontal resolution of 1.25° latitude and 1.875° longitude and 38 vertical levels, with eight levels below 1.1 km and the model top at 39 km. Details on the implementation of the Cloud Feedback Model Intercomparison Project Observation Simulator Package [COSP; Bodas-Salcedo et al., 2011] in ACCESS1.3 and the associated uncertainties are covered in the companion paper Franklin et al. . The model results examined in this study are from a 2 year simulation (Dec 2006 – Nov 2008) using observed fields of SST and sea ice and include atmospheric gases and aerosols. COSP outputs are produced by the simulator package running online in the model and the monthly mean simulator fields are produced from daily means of 3 hourly calculations.
 In this study model outputs are analyzed from radar and lidar simulators, the details of which are given in Bodas-Salcedo et al. . The Polarization and Anisotropy of Reflectances for Atmospheric Science coupled with Observations from a Lidar (PARASOL) simulator described in Konsta et al. [Evaluation of clouds simulated by the LMDZ5 global climate model (GCM) using A-train satellite observations (CALIPSO-PARASOL-CERES), Submitted to Climate Dyn, 2012] is also used to evaluate the monodirectional reflectance of the modeled clouds. The CloudSat observations used are the 2B-GEOPROF product [Marchand et al., 2008]. To reduce the effect of false detections that occur close to the limit of the CloudSat sensitivity, a threshold of −25 dBZ rather than −30 dBZ has been imposed on the CloudSat observations and radar simulator output [Marchand et al., 2009]. The lidar observations are from the GCM-Oriented CALIPSO Cloud Product [GOCCP; Chepfer et al., 2010] where a lidar scattering ratio (SR) of 5 is the threshold used for detection of cloud. More details about the radar and lidar observations used and their limitations are given in Franklin et al. . Observations of cloud reflectance are from the level 1 PARASOL product for the 865 nm channel with a horizontal resolution of 6 × 6 km2 [Konsta et al., submitted]. The reflectance observations give information on the location and optical thickness of clouds that occur during the day time. The observations of the monthly averaged top-of-atmosphere cloud radiative effects (CREs) are from the Clouds and the Earth's Radiant Energy System (CERES) EBAF-TOA Ed6.2r [Loeb et al., 2008] that are energy balanced to the ocean heat storage term and the clear-sky fluxes have been spatially filled.
 To evaluate the regime-based cloud properties from the simulation, the tropical circulation (30°N – 30°S) is decomposed following Bony et al.  using the monthly mean pressure velocity at 500 hPa as a proxy for the large-scale dynamics and the SST and LTS are used to sort the clouds into thermodynamic regimes. ERA-Interim [Dee et al., 2011] 500 hPa pressure velocity (ω500) and temperature at 700 and 1000 hPa, and the Hadley Centre sea ice and SST dataset version 1 [HadISST1; Rayner et al., 2003] are used to decompose the satellite observations into regimes. The SSTs used to force the model are those described by Hurrell et al. , which use a merged product based on HadISST1 and the National Oceanic and Atmospheric Administration Optimum Interpolation analysis version 2 [Reynolds et al., 2002]. Monthly mean data are used for the period December 2006 – November 2008, and all data are interpolated onto the same 2° × 2° grid. As in previous conditional sampling studies [e.g., Bony et al., 1997; Williams et al., 2003; Su et al., 2011], the use of monthly means is suitable in the tropics as the variation of SST across a month is relatively small. The bin sizes used for the conditional sampling are 10 hPa day−1 for ω500 and 1°C for SST and LTS. Figure 1 shows the probability density functions of ω500, SST, and LTS over the 2 year period for the reanalyses/observations and ACCESS1.3. The distributions across the tropical oceans are very similar with the model having more occurrences in the peak of the ω500 distribution in the weak subsidence regimes and fewer occurrences across the convective regimes. The SST distributions show slightly less occurrences in the model at the peak of the distribution around 28°C, and this is balanced by a small increase in the occurrences over the warmest SSTs. The modeled LTS distribution has fewer occurrences between 14 and 17°, and more occurrences for smaller values of LTS. These differences in distributions are not expected to significantly impact the compositing results.
3 Regime Sorted Cloud Properties
 Sorting cloud fields using ω500 as a proxy for the large-scale atmospheric circulation allows for the interpretation of the cloud properties as a function of the dynamical regime. Figure 2a shows the results of the compositing for the GOCCP observations of total, high (50 – 440 hPa), middle (440 – 680 hPa), and low level (>680 hPa) cloud fractions and the corresponding lidar simulator output from the model. The high and low cloud amounts have opposite behaviors, with the high cloud reducing as the strength of the upward motion reduces and the low cloud increasing as the subsidence increases. The model captures these behaviors, and the result for the ACCESS1.3 high cloud fraction is in excellent agreement with GOCCP with errors less than 0.05. The low cloud fraction from the model is underestimated by up to 0.15, with the error increasing as the strength of the subsidence increases. Similar to the high cloud, the midlevel cloud fraction increases as the upward motion strengthens in both the observations and the model. However, the midlevel cloud amount in the model does not increase as much as the observations, with the largest underestimate of 0.15 occurring for the strongest upward motion. Lack of midlevel cloud is a well-known problem with GCMs and has been attributed to lack of detrainment from the cumulus parameterization [e.g., Bodas-Salcedo et al., 2008]. However, even when convection is explicitly resolved, models still show a significant underestimate of midlevel cloud associated with deep convection [Marchand et al., 2009; Kodama et al., 2012]. This suggests that to rectify this model error is more complicated than purely moving to a superparameterized or higher-resolution model.
 Compositing with respect to SST enables an evaluation of the thermodynamical controls on cloud regimes and allows a clear separation of the types of boundary layer clouds that are found in the tropical region of 30°N – 30°S [e.g., Williams et al., 2003]. Figure 2b shows the GOCCP and lidar simulator cloud fractions sorted by SST. The model low cloud fraction is underestimated when composited using SST; however, the underestimate is less than that shown over the subsidence regions in Figure 2a. Over the cooler SSTs the modeled low cloud cover is underestimated by less than 0.05, increasing to an error of 0.1 at 24°C. The largest error in the model total cloud cover occurs for the SST range of 25 – 28°C where the total cloud fraction is underestimated by up to 0.2. It is this region where the transition occurs from the low cloud amount dominating the total cloud fraction to the high amount contributing the most. At 26°C there is a sharp increase in the high cloud fraction in both the observations and the model, which is consistent with previous studies [e.g., Waliser and Graham, 1993]; however, the minimum total cloud amount shown here in the observations occurs over warmer SST in the model. This suggests that the SST threshold and associated reduced vertical stability [Betts and Ridgway, 1989] required for the transition from shallow to deeper convection is higher in ACCESS1.3. The observations show a reduction in high and total cloud fraction for SST >29.5°C, and the reduction of convection at this threshold SST was described by Waliser and Graham  to be caused by atmospheric variability uncoupled to the local SST, such as subsidence from remote convection, that warms and dries the boundary layer inhibiting deep convection and cloud and allowing the SST to increase [Lau et al., 1997]. The model results show an increase in high and total cloud fraction for SSTs greater than 29.5°C due to these results being from an atmosphere only simulation, highlighting the importance of the coupling between the surface heat and moisture fluxes between the atmosphere and ocean in simulating cloud properties over SSTs warmer than 29.5°C [Waliser and Graham, 1993].
 The transition between low cloud types has been shown to be particularly important in simulating low cloud feedbacks [Bony and Dufresne, 2005]. Stratocumulus and trade cumulus have distinctive cloud properties, and a different balance between physical parameterizations and the stratification using LTS allows insight into the model's ability to simulate the transitions between these cloud regimes. Figure 2c shows that the cloud fraction of the stratocumulus clouds associated with the strongest LTS is underestimated in the model by 0.08. The underestimate reduces as the stability weakens and the cloud type transitions to trade cumulus, and this result agrees with that of Konsta et al. [submitted], who showed that the LMDZ5A climate model significantly underestimates subtropical low clouds compared to GOCCP. The observations show that the onset of high and midlevel clouds occurs for LTS values of about 16°, indicative of the transition between shallow and deeper convection in the tropics where the gradient in the total cloud amount changes. The model shows the same behaviors; however, the LTS at which the high and total cloud amounts increase is reduced to 14-15° in agreement with the warmer SST required in the model to produce significantly more high cloud. This result will be explored further in the following section.
 COSP includes the PARASOL simulator [Konsta et al., submitted], which is a useful tool to evaluate the reflectance, a proxy for cloud optical depth, and as such the properties of the clouds that contribute to the shortwave CRE (SWCRE). Sorting with respect to ω500 demonstrates that the largest error in the modeled PARASOL reflectance occurs in the regions of strongest upward motion at 500 hPa (Figure 2d), where the reflectance is underestimated by 0.1. The model underestimates the reflectance for all of the convective regimes, with the error increasing as the strength of the vertical velocity increases. The failure of the model to simulate the large increase in reflectance with increasing upward vertical velocity suggests that the convective cloud properties that contribute to the reflectance, such as the vertical distribution of liquid/ice water paths are not as sensitive to the dynamics in the model as is shown by the observations. Su et al.  noted in their analysis of CloudSat and modeled cloud water content, that the high ice water in models they examined was well correlated with warm SST. This led them to suggest that the parameterization of deep convection was quite sensitive to SST. We find the same result for the model reflectance of the convective regimes that shows a stronger sensitivity to SST >28° than to increasing vertical velocity.
 There is an underestimate of the reflectance for the strongest subsidence regimes, and Figure 2e shows that this occurs over the SST ranges of 16 – 22°C. Part of the reason for the underestimate in reflectance in the model is due to not enough cloud cover; however, even when the cloud fraction is well modeled, for instance when SST <18°C, the error in the reflectance can be 0.05 and must be due to the underestimate of cloud condensate and/or overestimate of the effective radius that both determine the optical depth and reflectance. It is interesting to see the difference in the modeled reflectance when composited using LTS as compared to SST, which is due to the different LTS distributions of the observations and model as shown in Figure 1. Figure 2f shows that for large LTS, the reflectance compares well between the model and observations even though here the cloud fraction is underestimated. These results will be explored in the following sections where the radar reflectivity and lidar SR from the model are compared to observations for the dynamical and thermodynamical regimes.
 Ultimately, it is the effect of clouds on the radiative forcing that is important for driving coupled climate models. To understand how the errors determined by the COSP analysis manifest in the climate system forcing, the top-of-atmosphere CRE has been evaluated by compositing with respect to ω500, SST, and LTS. CRE is the difference between the all-sky and clear-sky fluxes and measures the effect of clouds on the radiative fluxes at the top of atmosphere. The underestimate of the modeled reflectance has a clear impact on the SWCRE as shown in Figures 2g, h, and i. The model underestimates the strength of the SWCRE across all of the convective regimes with negative values of 500 hPa pressure velocity, low LTS, and also over the coolest and warmest SSTs. The SWCRE error increases as the strength of the upward motion increases and the largest error when sorted with respect to ω500 is approximately 10 Wm−2. The LWCRE is represented very well in ACCESS1.3, which reflects the good simulation of high cloud cover and cloud top heights [Franklin et al., 2013]. The regimes where the LWCRE bias is larger than 5 W m−2 are for the LTS values around 14 and 24°. There is also a large error at the weakest LTS of 10°; however, as there are very few occurrences for this regime (see Figure 1), the results are not statistically robust. As discussed previously, LTS values around 14° is where the model and the observations differ in the transition to deeper convection, and it is the lack of high cloud from convection that causes the weaker LWCRE in the model for these regimes.
 To understand more about the model inadequacies and identify parameterization deficiencies, the macro and microscale cloud properties as a function of height will be examined by regime. The vertical distribution of CloudSat cloud fraction composited as a function of the large-scale dynamical regime using ω500 shows that the maximum cloud fraction occurs in the upper troposphere of the strongest convective regimes (Figure 3). The cloud fraction profile for the strongest upward motion is top heavy, and this is in contrast to the model radar simulator cloud fraction in these regimes that shows the maximum cloud/hydrometeor fraction occurring below the melting level with the upper level cloud fraction underestimated by 0.1. Note that the lowest 1 km of the CloudSat observations are contaminated by ground clutter and that the radar is sensitive to both cloud and precipitation sized particles. For the weak convective regimes with ω500 between −20 and 0 hPa day−1, the model cloud fraction profile closely resembles the observed profile except for the underestimate of cloud fraction of 0.05 in the midlevels. In ACCESS1.3 the prognostic cloud scheme PC2 [prognostic cloud, prognostic condensate; Wilson et al., 2008] represents the cloud fraction of both convective and stratiform cloud. This is achieved through the detrainment of condensate and cloud fraction from the convective plumes into the large-scale cloud variables. Wilson et al.  found that using a prognostic cloud scheme led to further reduced midlevel cloud compared to using a separate diagnostic cloud fraction for convective and stratiform cloud. It was their belief, however, that the underestimate of midlevel cloud was due to a lack of convective activity rather than the choice of diagnostic or prognostic cloud scheme parameterization. The subsidence regimes show that the model has a good representation of low level cloud top height at around 3 km; however, throughout all ω500 regimes, the model has too many occurrences of hydrometeors between 1 and 2 km.
 Comparing the radar detected cloud fraction with the lidar detected cloud fraction allows conclusions to be drawn regarding the modeled cloud properties across regimes. Given that the model underestimates the radar detected high cloud fraction in the deep convective regimes but overestimates the lidar detected high cloud fraction for the regimes with ω500 < −40 hPa day−1 suggests that the ice particle sizes of these clouds in the model are smaller than the observations and undetectable by the radar. Zhang et al.  evaluated tropical cloud regimes in CAM3 and showed the opposite high cloud error, with the model not able to simulate enough thin cloud that was only detected by the lidar. The lidar becomes attenuated in rain but is able to detect small cloud droplets in the low levels that CloudSat cannot. The lowest levels detectable by CloudSat show that the model overestimates the hydrometeor frequency of occurrence; however, the comparison with GOCCP shows the opposite with too little occurrence of cloud. This implies that in the lowest levels, the model has too many occurrences of rain and drizzle and not enough cloud. Stephens et al.  compared the precipitation characteristics from CloudSat to a number of models and found that the models tended to precipitate about twice as frequently as the observations. The results in Figure 3b show a similar order of magnitude overestimate in the frequency of occurrence of radar reflectivities in the low levels. As well as conventional GCMs, Stephens et al.  included models with higher resolution and explicit convection in their study, and these models produced a better frequency of precipitation in the tropics. However, Marchand et al.  showed that the Multiscale Modeling Framework superparmeterized model with explicit convection produced precipitation reflectivities too frequently in the trade cumulus regions, suggesting that this problem is multifaceted and likely due to the complex interactions across a wide range of scales that affect precipitation development.
 Comparing the radar and lidar observed and modeled cloud fractions composited as a function of SST and LTS shows more clearly the boundary layer cloud regimes and the transitions between them. SST and LTS are not independent variables, and as the results for the cloud fractions show similar results we only show the LTS composites (Figure 4). The radar results show that the model is producing cloud/hydrometeors about twice as often in the lowest levels compared to the observations. Given that the comparison with the cloud fraction from GOCCP and the lidar simulator compares reasonably well for LTS >22°C, this implies that for the trade cumulus region, the model precipitates twice as often as the CloudSat observations in agreement with previous studies. The cloud fraction from the lidar simulator is underestimated in the region of LTS between 20 and 22°, but agrees with the observations for the clouds associated with the strongest stability. Compared with CloudSat, the radar simulator shows cloud/hydrometeor occurrence between 2 and 3.5 km, which is not present in the observations. Given that the lidar simulator does not show cloud here, this suggests that in the model these are hydrometeors that have fallen from the overlying cirrus cloud that have not evaporated, and points to the air above the boundary layer in these strong LTS regions being too moist in the model.
 The magnitude of the low cloud fraction change as the thermodynamic cloud regime transitions across decreasing LTS is captured well by the model. However, the cloud top heights do not increase as much in the model as they do in the observations (maximum heights of 2 km in the model compared to 3.5 km in the observations), indicating that the depth of the boundary layer may be too shallow. Wyant et al.  showed in a multimodel evaluation study over the southeast Pacific that most models tend to underestimate the boundary layer depth of the stratus clouds and this is not purely due to resolution. The boundary layer scheme in ACCESS1.3 is the Lock et al.  scheme, that uses a nonlocal mixing scheme in unstable conditions. Beljaars and Betts  found that without explicit cloud top entrainment, boundary layers tended to be too shallow, cold, and moist. In the Lock et al.  scheme when the boundary layer is topped with stratocumulus, there is an explicit calculation of cloud top entrainment based on the cloud top radiative divergence [Lock, 2001]. This should allow the growth of the boundary layer into the drier free troposphere, and in the next section it will be shown that the model cloud top heights do increase with decreasing LTS; however, the deeper boundary layer clouds are less frequently produced by the model as compared to GOCCP.
4 Regime Sorted Radar Reflectivity and Lidar SR Histograms
 As well as cloud/hydrometeor occurrence observations, CloudSat and GOCCP provide radar reflectivity and lidar SR observations that allow for a more detailed evaluation of the modeled cloud properties. When errors in the modeled cloud amount are not present, we can compare CloudSat and GOCCP and the associated simulator output to elucidate how well the model simulates particle sizes and water contents. This is because CloudSat reflectivity increases with particle size and water content, whereas GOCCP SR increase with water content but become attenuated in the presence of large particles. However, even when the modeled cloud amount is erroneous, examining the reflectivities and SR provides useful information for model development. For example, in the case of cumulus congestus where the modeled cloud amounts are significantly underestimated, understanding whether the model is capable of producing this cloud with large SR is helpful for knowing whether to focus development efforts purely on increasing the frequency of occurrence of this cloud type or whether the cloud properties also need to be improved. Figure 5 shows the results of compositing the CloudSat radar reflectivity as a function of height with respect to the 500 hPa pressure velocity for three categories of reflectivity: −25 – −15 dBZ, −15 – 0 dBZ, and 0 – 20 dBZ. Separating the reflectivities into these categories provides insight into the observed physical processes when composited by large-scale dynamical or thermodynamical forcing and more clearly illustrates some of the errors and inadequacies in the model parameterizations.
 The high level cloud observed by CloudSat over the convective regimes has the largest frequency of occurrence for the lowest reflectivities between −25 and −15 dBZ, and this occurs over 11 – 14 km (Figure 5a). As the ice particles in the high clouds grow through deposition and aggregation, the maximum occurrence for the medium reflectivity range of −15 – 0 dBZ, which is indicative of larger particles, is at a lower altitude centered at 11 km. The highest frequency of occurrence for the largest observed reflectivities of 0 – 20 dBZ is more confined to the strongest 500 hPa upward motion, and these reflectivities occur mostly from 4 to 8 km. The modeled cloud top heights for the lowest reflectivity category show good agreement with the observations across all dynamical regimes including the subsiding motions. The frequency of occurrence of modeled reflectivities between −25 and −15 dBZ for the high clouds in the convective regimes shows the same features as the observations; however, the height of the corresponding frequencies occurs over a deeper layer. The smaller ice particle sizes produced by the model that was revealed in the previous section is highlighted by this analysis using different categories of radar reflectivity. The frequency of occurrence of the medium reflectivity category of −15 – 0 dBZ is underestimated from the model above the melting level for all convective regimes. The most striking difference between the model and the observations occurs for the largest reflectivity category of 0 – 20 dBZ, which indicates the presence of large and/or very numerous particles. The model simulates no occurrences of these reflectivities above 10 km for all dynamical regimes suggesting that the model produces ice cloud with smaller particle sizes and less ice water content than observed by CloudSat. Analyzing the results of sensitivity tests where the contribution from each of the convective and large-scale ice and water contents was excluded from the radar simulator showed that the large-scale rain accounts for the vast majority of the large reflectivities below 5 km (not shown). The convective precipitation contributes minimally to occurrences of reflectivities between 0 and 20 dBZ below 5 km due to attenuation. The convective ice is what is responsible for the reflectivities greater than 0 dBZ above 5 km, with the large-scale ice not contributing at all to the large reflectivity category, which partly reflects the inability of GCMs to simulate mesoscale convective complexes with organized convection and thick anvil located close to convective cores.
 The water contents of the convective clouds from the simulators may be underestimated due to the way in which the subgrid scale distribution of water content is partitioned across the gridbox subcolumns. The Subgrid Cloud Overlap Profile Sampler [SCOPS; Webb et al., 2001] is used in COSP to generate a distribution of cloud and precipitation across a number of subcolumns in each GCM grid box; 100 subcolumns are used in the simulation. The subcolumns are assigned a cloud fraction of 0 or 1 such that the total cloud fraction of the subcolumns is equal to the gridbox mean value and consistent with the model vertical overlap assumption (maximum-random in the case of ACCESS1.3). The gridbox mean water content is then spread equally across the subcolumns. In the analysis of deep convective clouds in the Modern-Era Retrospective Analysis for Research and Application, Posselt et al.  tested the impact of changing the subgrid distribution from being equally distributed to being exponentially distributed, such that the deepest clouds were assigned more condensate and the shallower clouds less. The result was a better comparison with the observations for cloud optical depth and total water path and a worse result for outgoing longwave radiation. This tells us that we could potentially produce more occurrences of 0 – 20 dBZ for the high convective regime clouds in the model by changing the subgrid distribution of condensate. However, this approach could degrade the comparison with the lidar simulator and GOCCP since the model is lacking in occurrences with large SR characterized by small effective radii (Figure 6).
 Due to CloudSat being sensitive to both cloud and precipitation sized particles, the comparison of observed and modeled radar reflectivity categories composited with respect to ω500 enable errors in the modeled precipitation to be identified for particular cloud regimes. Focusing on the levels between 1 and 2 km, Figure 5 shows that the observations in the convective regimes occur predominantly with reflectivities between −15 and 0 dBZ, typical of drizzle sized particles. Care needs to be taken in interpreting the results of Figure 5, particularly for the deep convective regimes and large reflectivities, due to attenuation and the lack of multiple scattering effects in the radar simulator, which is expected to be significant for rain rates greater than 3 mm h−1 [Marchand et al., 2009]. For the nonprecipitating sized particles in the reflectivity category of −25 – −15 dBZ, the observations and model show the maximum frequency of occurrence for the low clouds occurs for the strongest subsidence regime typical of stratocumulus. This cloud regime in the model also has more occurrences of reflectivities greater than 0 dBZ that are associated with rain, showing that the modeled stratocumulus has more drizzle and rain sized particles than the observations.
 SCOPS determines which subcolumns used in COSP contain precipitation using five options that are described in detail in Zhang et al. . The majority of precipitation occurrences are assigned to the subcolumns that have cloud at the current level or precipitation in the level above. The precipitation amount is then distributed equally among these subcolumns and converted to a mixing ratio using the drop size distribution (DSD) for snow and rain used in the model. Di Michele et al.  tested the sensitivity of their radar simulator in the ECMWF model to a different subgrid distribution of precipitation. They changed the precipitation overlap to be the same as that used in the model to calculate the evaporation of precipitation, which reduces the precipitation fraction in a gridbox compared to the SCOPS approach. Using the SCOPS distribution resulted in smaller reflectivities in the lower levels due to the precipitation being spread over a larger number of subcolumns. Therefore, changing to a different distribution of precipitation used in the simulators that reduces the precipitation fraction would likely increase the error in ACCESS1.3 of having too many occurrences of high reflectivities in the lowest levels of the subsidence regimes. The results of Di Michele et al.  showed a greater sensitivity in the low level reflectivity to a change in the rain DSD. To test the impact of this on our results we conducted a sensitivity experiment where the rain DSD was changed to the Marshall and Palmer  DSD with fixed intercept parameter N0 = 8 × 106 m−3 m−1. The DSD used in ACCESS1.3 is also an exponential distribution but predicts the intercept parameter as a function of the slope of the distribution (λ), where N0 = 26.2 λ1.57. The idea behind having the intercept parameter dependent on the rain water mixing ratio is to allow the distribution to change from being characterized by large numbers of drizzle sized drops for low rain water contents, to that of a distribution with a smaller number of larger drops for higher rain water contents. Using the Marshall-Palmer rain DSD produces more occurrences of reflectivities between 0 and 20 dBZ in the low levels across all regimes and an increase (decrease) in the occurrences of −15 – 0 dbZ for the convective (subsidence) regimes (not shown). The differences in frequency of occurrence are small, all less than 0.01 and typically less than 0.005, and the tendency for the model to produce significantly greater occurrences of drizzle and rain for all cloud regimes compared to CloudSat is unchanged.
 Figure 6 shows the lidar SR composited with respect to ω500 for three categories of SR that reflect increasing optical depths: 5 – 15, 15 – 30, and 30 – >80. For the lowest SR category, the high cloud of the convective regimes occurs more often in the model than in the observations, and while the observations show a symmetric frequency distribution with height, the model favors a bottom heavy profile with more occurrences at the lower altitudes of the high cloud at around 11 – 12 km. The same model error is seen in the other SR categories and may be due to the lack of parameterized turbulence in the cirrus cloud that would act to recirculate ice and maintain a more symmetric distribution. The bottom heavy profile in the model might indicate that particles are falling out of the cloud too rapidly, evaporating and contributing to the underestimate of ice water. This idea is supported by the results of a sensitivity test in Franklin et al.  where the fall speed of the larger ice category was reduced. For the high cloud in the Tropical Warm Pool region, the experiment with reduced fall speeds produced a shift in the lidar SR to larger values in better agreement with GOCCP. This error and improvement is the opposite of the tropical cirrus cloud error in NICAM presented by Kodama et al. . Their results for a 14 km horizontal resolution model with explicit convection showed too much high cloud and overestimated cloud top heights. By adding a fall speed parameterization for ice and increasing the rate of autoconversion from ice to snow, they improved their model error by reducing the amount of high cloud. In the Tropical Warm Pool region Nam and Quaas  showed that ECHAM5 also has the opposite error to ACCESS1.3 and overestimates SR >30 above 9 km. They found better agreement with GOCCP by doubling the effective radii of ice crystals. In the case of ACCESS1.3, the frequency of occurrence of high altitude cloud with medium and large SR is underestimated by the model across all regimes, indicating that the larger cloud fraction in the model shown in Figure 3 tends to be composed of smaller ice water contents and/or too large effective radii. Given that at around 14 km the radar reflectivities in the model are all underestimated, it is highly likely that the lack of ice water content is the primary reason for the tendency of the high cloud to have smaller SR. One way to increase the ice water content is to increase the amount of condensate detrained from convective plumes [Wilson et al., 2008]. Detraining more condensate will reduce the amount that is converted into convective precipitation, helping to balance the proportion of convective verses stratiform precipitation that models tend to skew towards too much convective rain [e.g., Dai, 2006]. However, this change would be expected to alter the thermodynamic structure of the atmosphere through radiative and latent heating effects, and the impact of those changes is likely to require additional modifications. The model does simulate large condensate amounts with SR >30 in the levels between 5 and 8 km where we would expect to see cumulus congestus type clouds; however, it does so less than half as frequently as the observations.
 Examining the cloud regimes stratified by LTS gives a better representation of the boundary layer clouds over the tropics and the transitions between them. Figure 7 shows categories of radar reflectivity as a function of height composited with respect to LTS. In the model the boundary layer cloud regimes over strong stability show the same frequency of occurrence for the nonprecipitating and the drizzle reflectivity categories. This is in contrast to the CloudSat observations that show more occurrences of nonprecipitating cloud. In the following section, the results of a sensitivity test using the warm rain parameterizations of Franklin  show that by changing the autoconversion scheme the model is able to simulate more occurrences of nonprecipitating cloud (Figure 10). Above the boundary layer, in each reflectivity category there are more occurrences of hydrometeors in the model as compared to CloudSat. As discussed previously, this might indicate that the model has a moist bias above the strong inversion associated with the LTS regimes >18° and does not evaporate particles falling from the overlying cirrus.
 The lidar observations composited as a function of LTS show that there are many occurrences of low clouds with SR <15 (Figure 8). The model does not simulate enough of these clouds with low liquid water contents and instead simulates clouds with medium and large SR and in a more confined vertical layer than the observations. The larger occurrence of modeled clouds with higher SR is the reason for the good reflectance shown in Figure 2 even though the cloud fraction is underestimated. GOCCP shows that clouds extend more vertically as the LTS reduces, demonstrating that the depth of the cloudy boundary layer increases as the cloud regime transitions from stratocumulus to trade cumulus. The model does capture this deepening of the cloudy boundary layer and more vertically extensive clouds as the stability weakens; however, there is a clear underestimation of cloud top heights in the model. Comparing the reduction in cloud occurrence as the stability weakens and the cloud regime transitions from the large stratocumulus cloud fractions to the lower cumulus cloud fractions shows that the model does not reduce the cloud frequency as much as GOCCP for LTS <18°.
 The occurrence of midlevel cloud with large SR >30 is clear in the observations across all regimes, with the frequency of occurrence increasing as the stability weakens. The model is able to simulate these clouds with large SR; however, it does so only half as often as is observed by GOCCP. Figure 2 shows that the transition from shallow to deeper convection in the model occurs with weaker LTS than in the observations, and the results in Figure 8 show that this may be due to the lack of modeled cumulus congestus and the associated transport of moisture into the midtroposphere that is important in the transition from shallow to deep convection [Waite and Khouider, 2010]. Franklin et al.  analyzed single column model simulations of the Tropical Warm Pool – International Cloud Experiment and showed that the prognostic cloud scheme used in ACCESS1.3 simulates significantly larger in-cloud water contents in the midlevels associated with cumulus congestus type cloud than the lidar/radar-derived observations and the model with the diagnostic cloud scheme. The reason for the in-cloud water contents being larger in the prognostic cloud scheme simulations was due to: the direct detrainment of condensate from the convective plumes into the large scale, rather than the case in the diagnostic scheme where the convective condensate is detrained into the environment, allowed to evaporate and is not picked up by the large-scale variables until the environment reaches saturation; the formulation of the detrainment of convective cloud fraction into the large-scale results in small cloud fractions, which add to the excessive in-cloud water contents. The result of these large in-cloud water contents is the production of large particles with fall speeds typically 25% faster than those observed, which do not evaporate enough contributing to a dry bias in the midlevels. This may contribute to the underestimated frequency of occurrence of clouds in the midlevels.
 Figure 2 showed that for the SST range of 16 – 18°C, the low cloud fraction is modeled well, but the reflectance is significantly underestimated. Figure 9 shows that GOCCP has large occurrences of cloud below 2 km with SR >30 and clouds at these levels are underestimated in the model, which is the reason for the lower reflectance over the coolest SSTs. The result from the lidar simulator shows that the model represents the transition of cloud types across the thermodynamic regimes associated with SST fairly well in terms of the overall reduction in frequency of occurrence. At 26°C, the observations show a transition from shallow to deeper convective cloud types (Figure 9e). The model maintains the same frequency of occurrence of low cloud from 24 to 27°C and requires a 1° warmer SST before this transition occurs, which is also apparent in Figure 2. This agrees with the LTS results that showed the model required a LTS 1° less than the observations before the onset of deep convection occurs.
5 Impact of Warm Rain Parameterizations
 Analyzing the results for a sensitivity experiment in which the contribution of the large-scale rainfall was eliminated from the simulators confirmed the result of Bodas-Salcedo et al.  where they showed that the dominant contribution to the drizzle reflectivities was due to large-scale rain. In warm clouds it is the autoconversion and accretion processes that govern the rate at which the cloud water is converted into rain, and we have tested the impact of using the warm rain scheme of Franklin . Using collision kernels that included/excluded the effects of turbulence on cloud droplet collision rates, two schemes were developed by Franklin  to cover a broad range of cloud types and conditions. Here we only test the nonturbulent scheme, as the implementation of the turbulent scheme into a GCM requires a representation of the distribution of subgrid scale turbulence within clouds. In the control version of the model, the autoconversion parameterization follows Manton and Cotton 
where qcl is the cloud water mixing ratio (kg kg−1), g is gravitational acceleration (m s−2 ), Ec is the collision efficiency set to 0.55, ρ and ρliq are the densities of air and water (kg m−3), μ is the dynamic viscosity due to air (kg m−1 s), and Nc is the number concentration of cloud droplets (m−3). In the sensitivity test, we use the autoconversion parameterization from Franklin  developed for nonturbulent conditions and given by
 We also tested the impact of a different accretion parameterization that represents the collection of cloud water by raindrops. The default parameterization is given in Wilson and Ballard  and is a function of qcl and the rain drop fall speed, where it is assumed that all collisions result in collection. The accretion parameterization from Franklin  is a function of both the cloud and rain water mixing ratios and does not depend on rain drop fall speeds. For the results presented in this section, the sensitivity test includes both the autoconversion and accretion parameterizations from Franklin . However, performing experiments with only the different autoconversion or accretion parameterization demonstrated that the model has little sensitivity to the accretion parameterization, and it is the representation of autoconversion that has more significant effects on the modeled cloud properties.
 Comparing the radar simulator results from the sensitivity test in Figure 10 with the CloudSat observations and control model results composited as a function of LTS in Figure 7 shows that the Franklin scheme produces more nonprecipitating boundary layer cloud with reflectivities between −25 and −15 dBZ above 1.5 km. The frequency of occurrence of nonprecipitating cloud is now greater than the observations with better agreement for the LTS range of 17 – 22°C and worse for the stronger LTS. The sensitivity experiment also produces more drizzle above 1.5 km and a small reduction of rain and large reflectivities below 1 km. The sensitivity experiment did not retune some of the parameters used in the autoconversion parameterization, in particular the collision efficiency, and the results could be improved by adjusting various parameters. The results demonstrate that the frequency of occurrence of nonprecipitating cloud can be significantly impacted by the choice of autoconversion scheme.
 To examine the effects of changing the warm rain parameterizations on the cloud water contents, the lidar simulator results composited with respect to LTS are shown in Figure 10. When comparing with Figure 8, the most significant change in the sensitivity test for the low clouds associated with stronger LTS is the increase in the frequency of occurrence of clouds above 1.5 km in the largest SR category. The control experiment under predicted the occurrence of clouds with SR >30 between 1.5 and 2 km by a factor of 4. The sensitivity experiment still underestimates the occurrence of these clouds but by less than a factor of 2, suggesting an improvement in the representation of the boundary layer depth. The increase in cloud amount will produce a stronger CRE, which will drive more cloud top entrainment and enhanced growth into the overlying warmer and drier free troposphere. The overall result of the nonturbulent warm rain parameterizations is to reduce the rate at which the cloud water is converted to rain water in order to increase the frequency of occurrence of nonprecipitating clouds and cloud occurrence above 1.5 km for the LTS range of 16 – 25°. There is also an impact on the high ice clouds in the sensitivity test with a small increase in the frequency of occurrence of SR >30 in better agreement with the observations. Changing the precipitation and the vertical distribution of cloud water is expected to change the latent heating within the cloud and subcloud layers, the boundary layer circulations, and surface fluxes. A more detailed examination of the sensitivity experiment will be conducted in the future to understand the full range of impacts of changing the warm rain parameterizations in ACCESS1.3.
 Cloud properties from CloudSat, GOCCP and PARASOL, and the corresponding satellite simulator output from ACCESS1.3 have been composited into the dynamical and thermodynamical regimes of the tropical oceans using ω500, SST, and LTS. The sensitivity of the cloud properties to the large-scale tropical dynamics and thermodynamics is generally well represented by the model, with the high cloud fraction and LWCRE having maximum errors less than 0.05 and 5 W m−2, respectively. Over the tropical oceans, the two cloud types that ACCESS1.3 simulates the reflectance and CREs most poorly is the deep convection associated with strong ascent and low LTS and the stratocumulus clouds that occur over the coolest SSTs. For these cloud regimes the model does not simulate enough SWCRE, with the maximum error close to 20 W m−2. This implies that in coupled ocean-atmosphere simulations, ACCESS1.3 will be simulating the wrong interactions between these clouds and the ocean surface. For example, where the model simulates stratocumulus the ocean surface will be receiving too much solar radiation, which could warm the SST and impact the structure of the boundary layer, leading to reduced cloud fraction and an even larger SWCRE error in these regions. The SWCRE errors in the coupled model are between 10 – 30 W m−2 in the Tropical Warm Pool region and 10 – 20 W m−2 in most of the stratocumulus regions [Bi et al., 2013], with the maximum errors comparable between the coupled and uncoupled simulations, suggesting that coupling with the ocean model causes some cancellation of the stratocumulus error that keeps the SST from warming due to too little SWCRE.
 The reason for the underestimate in the strength of the SWCRE was shown to be due to the model simulating high clouds in the convective regimes that have too little ice water content and too small particle sizes. The presence of large particles above 10 km that produce radar reflectivities of 0 – 20 dBZ in these regimes is never produced by the model and instead the modeled high clouds tend to occur too often in the optically thinnest category of lidar SR and too infrequently in the optically medium and thick categories. The lack of midlevel cloud across the convective regimes and over SSTs >29°C contributes to the underestimate of the SWCRE, with the magnitude of the SWCRE error increasing as the SST and upward midtropospheric vertical velocity increases. The PARASOL reflectance shows a stronger sensitivity to SST in the model and does not depend as much as the observations do on the 500 hPa vertical velocity. This suggests that the vertical distribution of the convective cloud water paths in the model that are important in determining the reflectance is too dependent on the thermodynamic forcing as determined by the SST, which is likely to impact the cloud-climate feedbacks in perturbed climate simulations.
 While the low level cloud fractions in the model are biased low by about 0.1 across most regimes, the transition from stratocumulus to trade cumulus with smaller cloud fractions is represented quite well by the model showing similar sensitivity to the large-scale dynamics and thermodynamic conditions as seen in the observations. This transition occurs as the subsidence weakens, the SST warms and the LTS reduces, and the boundary layer structure changes from being well mixed to decoupled where there exists a stably stratified layer near cloud base, and finally to a cumulus capped layer with stronger vertical transports of heat and moisture [Wood and Bretherton, 2004]. The boundary layer scheme used in the UM represents distinct regimes [Lock et al., 2000], and the regime-sorting results showed that this scheme does well at modeling the transitions of the low clouds across the tropical oceans.
 The comparison of the lidar observations and simulator composited with respect to LTS and SST show that the model requires a 1° weaker LTS and warmer SST than the observations in transitioning from shallow to deeper convection. This may be due to cumulus congestus occurring less than half as frequently in the model. Congestus is important for transporting moisture into the midtroposphere and preconditioning the environment for deep convection, which as suggested by Waite and Khouider  is more important than the buildup of convectively available potential energy in controlling the transition from shallow to deep convection. Increasing the entrainment rates to make the cloud tops more sensitive to environmental relative humidity would help with this [Derbyshire et al., 2004]; however, this will reduce the amount of high cloud and the depth of convection, which will degrade the good LWCRE result that ACCESS1.3 currently has and also make the lack of ice water content in the upper troposphere more severe.
 By using a compositing technique, this study has explored the relationships that ACCESS1.3 clouds have with the large-scale environment. The modeled clouds show similar sensitivity to the large-scale dynamic and thermodynamic conditions as CloudSat and GOCCP and many of the model errors identified occur across all regimes, such as the underestimate of optically thick clouds and the too frequent occurrence of drizzle and light rain. However, this analysis has also revealed regimes specific errors such as: trade cumulus clouds that tend to have lower cloud top heights compared to what is observed by GOCCP, and; not enough stratocumulus clouds with low SR typical of low water contents. Testing a different warm rain microphysics scheme [Franklin, 2008] showed that the frequency of occurrence of nonprecipitating cloud in the model is quite sensitive to the representation of the autoconversion process. The new scheme increased the occurrence of nonprecipitating cloud above 1.5 km and resulted in more low cloud with SR >30 in better agreement with GOCCP. Including a double moment microphysics scheme with prognostic number concentrations of cloud particles is likely to improve the modeled distributions of radar reflectivity and lidar SR for all low cloud regimes. Given the importance of low clouds to cloud feedback [Bony and Dufresne, 2005] and the significant underestimate of the stratocumulus SWCRE, this suggests that improving these cloud regimes would improve the predictive capability of the coupled ACCESS model for both present and perturbed climate simulations.
 This work has been undertaken as part of the Australian Climate Change Science Program, funded jointly by the Department of Climate Change and Energy Efficiency, the Bureau of Meteorology, and CSIRO. This work was supported by the NCI National Facility at the ANU. A. Bodas-Salcedo was supported by the Joint DECC/Defra Met Office Hadley Centre Climate Programme (GA01101). We thank and acknowledge the constructive comments from the reviewers.