Evaluating and understanding top of the atmosphere cloud radiative effects in Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report (AR5) Coupled Model Intercomparison Project Phase 5 (CMIP5) models using satellite observations
Climate Science Branch, NASA Langley Research Center, Hampton, Virginia, USA
Science Systems and Applications, Inc., Hampton, Virginia, USA
Corresponding author: H. Wang, Climate Science Branch, NASA Langley Research Center, 1 Enterprise Pkwy., Suite 200, Hampton, VA 23666, USA. (email@example.com)
 In this study, the annual mean climatology of top of the atmosphere (TOA) shortwave and longwave cloud radiative effects in 12 Atmospheric Model Intercomparison Project (AMIP)-type simulations participating in the Coupled Model Intercomparison Project Phase 5 (CMIP5) is evaluated and investigated using satellite-based observations, with a focus on the tropics. Results show that the CMIP5 AMIPs simulate large-scale regional mean TOA radiative fluxes and cloud radiative forcings (CRFs) well but produce considerably less cloud amount, particularly in the middle and lower troposphere. The good model simulations in tropical means, with multimodel mean biases of -3.6 W/m2 for shortwave CRF and -1.0 W/m2 for longwave CRF, are, however, a result of compensating errors over different dynamical regimes. Over the Maritime Continent, most of the models simulate moderately less high-cloud fraction, leading to weaker shortwave cooling and longwave warming and a larger net cooling. Over subtropical strong subsidence regimes, most of the CMIP5 models strongly underestimate stratocumulus cloud amount and show considerably weaker local shortwave CRF. Over the transitional trade cumulus regimes, a notable feature is that while at varying amplitudes, most of the CMIP5 models consistently simulate a deeper and drier boundary layer, more moist free troposphere, and more high clouds and, consequently, overestimate shortwave cooling and longwave warming effects there. While most of the CMIP5 models show the same sign as the multimodel mean, there are substantial model spreads, particularly over the tropical deep convective and subtropical strong subsidence regimes. Representing clouds and their TOA radiative effects remains a challenge in the CMIP5 models.
 Clouds exert substantial influence on radiative energy balance within the Earth-atmosphere system. Clouds reflect incoming solar radiation back to space and cause a cooling influence on the Earth's climate (cloud albedo effect), they also trap the infrared radiation emitted from the Earth's surface and warm the Earth's climate (cloud greenhouse effect). The net radiative effects of clouds depend on their macrophysical and microphysical properties, including cloud amount, height, thickness and characteristics of particles that form the cloud. High, thin clouds are highly transparent to incident shortwave solar yet trap infrared radiation emitted from the Earth's surface, thereby warming the surface. Low, thick clouds mainly reflect incident solar radiation and cool the Earth's surface. The tropical thick deep convective clouds exert substantial cloud albedo effect and cloud greenhouse effect with the former dominating, and have a net weak cooling effect. The effects of clouds on the Earth's radiation budget are often quantified using cloud radiative forcing (CRF) defined as the top of the atmosphere (TOA) radiative flux differences between clear-sky and all-sky conditions [e.g., Charlock and Ramanathan, 1985]. Satellite measurements show that globally the shortwave cooling effect of clouds dominates their longwave warming effect, resulting in a net cooling effect of the clouds [Ramanathan et al., 1989; Cess et al., 1997; Webb et al., 2001; Bony et al., 2004; Su et al., 2010; Allan, 2011].
 Coupled general circulation models (GCMs) are the major tool to predict future climate change, yet cloud-climate feedback constitutes the largest source of uncertainty in these modeled future climate projections [e.g., Intergovernmental Panel on Climate Change (IPCC), 2007]. In the GCMs, since cloud processes often occur on spatial scales smaller than can be resolved by the models, clouds are parameterized using large-scale model variables. Despite significant progress in improving the cloud parameterization schemes in the GCMs, the knowledge of cloud processes is limited and the GCMs are challenged in simulating many aspects of the clouds and cloud processes [Stephens, 2005; Zhang et al., 2005; IPCC, 2007]. The differences in cloud parameterizations of the GCMs are a major contributor to the differences between the GCM simulations.
 Our confidence in the future climate change projections by the coupled GCMs to a large extent depends on how well these models simulate the observed present-day distribution of clouds and their associated radiative fluxes. There have been numerous studies evaluating and understanding clouds and TOA CRFs in atmospheric and coupled GCMs. Potter and Cess  showed that most of the nineteen Atmospheric GCMs (AGCMs) they analyzed produce realistic regionally averaged longwave and shortwave CRFs over the western Pacific and southeastern Pacific. The good agreement in the southeastern Pacific however is a result of compensating errors in either cloud vertical structure, cloud optical depth or cloud fraction. Stowasser and Hamilton  investigated the relationships between shortwave CRF and local meteorological variables over tropical and subtropical oceans in the Phase 3 of Coupled Model Intercomparison Project (CMIP3) coupled model simulations. They showed that the largest model errors are in the rather moist western tropical warm pool area and the dry subtropical subsidence regions. By comparison, the coupled models perform better over regimes with weak vertical motions. Ichikawa et al.  used the International Satellite Cloud Climatology Project (ISCCP) and the Earth Radiation Budget Experiment (ERBE) observations to evaluate the CRFs over tropical convective regions in the CMIP3 coupled model simulations, and showed that most of the CMIP3 models systematically overestimate shortwave CRF and underestimate longwave CRF over regions with weak vertical motion. They also found that over these regions most of the CMIP3 models have lower high-cloud amount and yet stronger shortwave cooling. Using the ISCCP and ERBE observations, and postprocessing model layer clouds with the assumptions that the ISCCP uses to define low-, middle-, and high-level clouds, Karlsson et al.  showed that the CMIP3 coupled models simulate less low-level clouds and yet excessive cloud radiative cooling over marine subtropics, and pointed out there had been no major progress in simulating the marine subtropical clouds. Several studies [Siebesma et al., 2004; Karlsson et al., 2010; Teixeira et al., 2011] investigated cloud regimes and regime transitions in tropics and subtropics in Atmospheric Model Intercomparison Project (AMIP) simulations from a number of weather and climate prediction models by focusing on a transect in the Pacific Ocean during summer months. The transect covers the transition from the stratocumulus regions off the west coast of California, across the shallow convection dominated trade winds, to the deep convection regions of the ITCZ. Siebesma et al.  showed that almost all the AGCMs they examined strongly underestimate stratocumulus clouds and local shortwave CRF; meanwhile most of these models overestimate clouds over trade wind regions and the ITCZ and thereby shortwave CRF there. Teixeira et al.  illustrated the diverse nature of clouds, boundary layer, and cloud parameterizations in the participating weather and climate prediction models, and showed that lack of stratocumulus clouds continues to be a significant problem in these models.
 The objectives of this study are to evaluate and understand the TOA CRFs in the current generation GCMs using satellite-based observations. The output from 12 AMIP type simulations participating in the Coupled Model Intercomparison Project Phase 5 (CMIP5) for the Fifth Assessment Report (AR5) of the IPCC are analyzed. In particular, we investigate any systematic errors that are common to these CMIP5 models, and address model uncertainty by studying fields and regions where these models diverge from each other the most. These CMIP5 AMIP type simulations are forced with observed monthly sea surface temperature (SST) and ice fraction, and thus have the same time signature as observations. The comparison with observations therefore excludes model biases due to SST differences as seen in the coupled climate model simulations and facilitates our investigation of model deficiencies. In addition, compared to earlier versions of AMIP simulations, a distinct advantage of the CMIP5 AMIP simulations is that some of the AGCMs provide satellite simulator output. This allows the direct comparison of cloudiness between the CMIP5 models and satellite-based observations such as the GCM-Oriented Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) Cloud Product (CALIPSO-GOCCP) [Chepfer et al., 2010]. The satellite-based observations that we use to evaluate these CMIP5 AMIPs include the Clouds and Earth's Radiant Energy System (CERES) Energy Balanced and Filled (EBAF) TOA shortwave and longwave radiative fluxes [Loeb et al., 2009], the CALIPSO-GOCCP, and the Atmospheric Infrared Sounder (AIRS) relative humidity [Aumann et al., 2003; Chahine et al., 2006]. Note that since we only have radiative fluxes and cloud fraction data available from the CMIP5 data portal (http://pcmdi3.llnl.gov/esgcet), we mainly investigate TOA CRF and their relationships with cloud fraction at different vertical levels. The data on other cloud properties (e.g., cloud type, cloud optical thickness) are not available to us.
 This paper is organized as follows. Following the Introduction, we describe in section 2 the data and methods used in this study. Section 3 evaluates the TOA shortwave and longwave radiative fluxes and CRFs in the CMIP5 AMIP simulations using the CERES EBAF data. The tropical CRFs in the CMIP5 AMIPs are then further investigated by examining their distribution in different dynamical regimes and linking the model CRFs with the model simulations of relative humidity and cloud fraction at high, middle, and low levels. The conclusion and discussion are given in section 4.
2 Data and Methods
2.1 CMIP5 AMIP Simulations
 In this study, the AMIP type simulations from 12 CMIP5 AGCMs are analyzed (Table 1). These AMIP simulations are of varying spatial resolution and country of origin. Among the 12 models, the GFDL-HIRAM-C360 model and the MRI-AGCM3-2S model provide AMIP simulations at one-quarter degree or finer resolution; four models provide AMIP simulations at about 1 degree resolution, and six models provide AMIP simulations at about 2 degree resolution. All models except the CNRM-CM5 are forced with the combinations of most, if not all, of the following external forcing agents: well-mixed greenhouse gases, black carbon, organic carbon, sea salt, mineral dust, anthropogenic sulfate aerosol direct and indirect effects, tropospheric and stratospheric ozone, land use change, solar irradiance and volcanic aerosols. The CNRM-CM5 AMIP is forced with greenhouse gases and volcanic aerosol only. Six of the 12 models (CanAM4, CNRM-CM5, HadGEM2-A, IPSL-CM5A-LR, MIROC5, MPI-ESM-LR) provide CALIPSO satellite simulator output, including total, low-, middle-, and high-cloud fraction data. These AMIP simulations are available monthly over the period January 1979 – December 2008.
Table 1. List of the 12 Atmospheric GCMs and Their Relevant Information
aCCCMA, Canadian Centre for Climate Modelling and Analysis; CNRM, Centre National de Recherches Météorologique; GFDL, Geophysical Fluid Dynamics Laboratory; GISS, Goddard Institute for Space Studies; INM, Institute of Numerical Mathematics; IPSL, Institut Pierre-Simon Laplace; MRI, Meteorological Research Institute.
 The CERES EBAF Edition 2.6r data [Loeb et al., 2009] are used to evaluate the TOA radiative fluxes and CRFs in the CMIP5 AMIP simulations. The data include TOA incident solar flux, and TOA upwelling shortwave and longwave radiative fluxes under clear-sky and all-sky conditions. The incident solar radiative flux is based on daily varying Solar Radiation and Climate Experiment (SORCE) TOA solar incoming irradiances, and has a long-term mean of about 1361 W/m2. The grid box mean of clear-sky TOA fluxes are determined using an area-weighted average of CERES/Terra broadband fluxes from completely cloud-free CERES footprints (20 km equivalent diameter at nadir), and the Moderate Resolution Imaging Spectrometer (MODIS)/Terra-derived “broadband” clear-sky fluxes (1 km) estimated from the cloud-free portions of partly and mostly cloudy CERES footprints. The use of high-resolution (1 km) MODIS data in producing the CERES EBAF clear-sky fluxes increases the sampling of clear-sky scenes over regions that are frequently cloudy, and thereby lowers the sampling uncertainty. With an objective constrainment algorithm, the all-sky shortwave and longwave fluxes are adjusted within their range of uncertainty to remove the inconsistency between global net TOA flux and heat storage in the Earth-atmosphere system. These radiative flux data are available at 1° longitude by 1° latitude resolution, and monthly over the period March 2000 – June 2011.
 To better understand the TOA CRFs in the AMIP simulations, we use the GOCCP data [Chepfer et al., 2010] to evaluate cloudiness in the AGCMs that provide CALIPSO satellite simulator output. Processed at the Institute Pierre Simon Laplace (IPSL) Laboratoire de Météorologie Dynamique (LMD), the GOCCP data are designed to evaluate the cloudiness simulated by the GCMs. They were obtained by processing the Cloud-Aerosol Lidar with Orthogonal Polarization Level 1 cloud product [Winker et al., 2007] using the same steps as in a lidar simulator. To produce the GOCCP data, Chepfer et al.  first computed instantaneous profiles of the lidar scattering ratio (SR) at the highest horizontal resolution of the data but at the vertical resolution typical of current GCMs. They then infer cloud diagnostics from the vertical distribution of cloud fraction, horizontal distribution of low-, middle-, high-, and total cloud fractions, instantaneous SR profiles, and SR histograms as a function of height. The GOCCP clouds are diagnosed using the same spatial resolution and cloud detection threshold in Cloud Feedback Model Intercomparison Program (CFMIP) Observation Simulator Package (COSP) [Bodas-Salcedo et al., 2011] used for the CMIP5 models, and therefore can be directly compared with the lidar simulator outputs from the CMIP5 simulations. The GOCCP data are available at 2° longitudes by 2° latitudes resolution, and monthly over the period June 2006 to December 2010. For this study, the total, high-, middle-, and low-cloud fraction data from the GOCCP are used. In addition, the relative humidity data from the AIRS Level 3 Version 5 standard products [Aumann et al., 2003; Chahine et al., 2006] are used to evaluate the model simulations of relative humidity in the tropics. Recent studies showed that the sampling bias of the AIRS climatology is overall small at low latitudes [Pierce et al., 2006; Hearty et al., 2011]. Pierce et al.  used an older version of the AIRS data (Level 3, Version 4) data to evaluate the CMIP3 coupled model simulations of air temperature and specific humidity profiles. They showed that the AIRS-model differences are considerably larger than the sampling biases of the AIRS measurements, and thus are due to systematic model errors. The AIRS relative humidity data are at 1° longitude by 1° latitude resolution and are at 12 standard pressure levels from 1000 hPa to 100 hPa. The data used in this study are monthly over the period January 2003 to December 2008. We note that there are only four pressure levels in the lower troposphere (1000, 925, 850, 700 hPa), thus the AIRS data provides a relatively coarse indication of the boundary layer.
2.3 Analysis Methods
 Following past studies [e.g., Charlock and Ramanathan, 1985], we use shortwave and longwave CRFs at the TOA to measure the effects of clouds on the Earth's radiation budget. The CRF is defined as the difference of TOA radiative fluxes between clear-sky and all-sky conditions:
 We focus our evaluation on the annual mean climatology with concurrent AMIP output and satellite observations, i.e., the period 2001–2008 for TOA radiative fluxes and CRFs, the period 2007–2008 for cloud fraction, and the period 2003–2008 for relative humidity. Since the focus is on annual mean climatology, only one AMIP ensemble member from each of the models is used. To facilitate the comparison, all the AMIP simulations are regridded onto the spatial grids of satellite observations via bilinear interpolation which has minimal impact on regional mean radiative budget.
 The model simulations are evaluated against the satellite-based observations in terms of regional means and spatial pattern correlation over specified spatial domains. When investigating the spatial distribution of model simulations, we use the difference between the multimodel mean and satellite observations to represent the mean bias of the models, and the standard deviation of the models relative to their multimodel mean to indicate the intermodel differences. In addition, to compare the CMIP5 model simulations with satellite observations in different circulation regimes, we focus on the tropical region (30°S–30°N) and stratify TOA CRFs, cloud fraction and relative humidity using the vertical velocity in pressure coordinate (ω) at 500 hPa (ω500), following Bony et al. . The ω500 serves as a proxy for tropical dynamical regimes, as the tropical atmospheric circulation is fundamentally diabatically forced. The ω500 from the ERA-Interim reanalysis (Dee et al. 2011) is used for satellite observations, and ω500 from each of the CMIP5 AMIP simulations is used for fields in that model.
3.1 Comparisons of TOA Radiative Fluxes and CRFs
 The global mean TOA radiative fluxes from the 12 CMIP5 AMIP simulations and the CERES EBAF are shown in Figure 1 and Table 2. Compared to the CERES EBAF, all the 12 models consistently have stronger incoming solar radiation at the top of the atmosphere (Figure 1), with the multimodel mean higher by 1.1 W/m2 (Table 2). The above differences are due to the different solar irradiances used in the CERES EBAF and the CMIP5 models. The EBAF Edition 2.6r uses daily varying SORCE TOA solar incoming irradiances that have a long-term mean of about 1361 W/m2 [Loeb et al., 2009], whereas most of the CMIP5 models use solar irradiances (see http://sparcsolaris.gfz-potsdam.de/cmip5.php for the CMIP5 solar forcing data) that have a long-term mean of about 1365 W/m2.
Table 2. Comparison of Global Mean TOA Radiative Fluxes and CRF Between the 12 CMIP5 AMIP Simulations and the CERES EBAF
12-Model Mean-CERES EBAF (/CERES EBAF)
σ (12 Models)
All-Sky TOA Radiative Fluxes
Clear-Sky TOA Radiative Fluxes
 Under clear-sky condition, 10 of the 12 models produce higher global mean TOA upwelling shortwave radiative flux than the CERES EBAF, and the 12-model mean bias is 1.7 W/m2. Five models (CanAM4, GFDL-HIRAM-C360, GISS-E2-R, IPSL-CM5A-LR and MIROC5) perform better and have less than 1 W/m2 biases. Nine of the 12 models underestimate global mean clear-sky outgoing longwave radiation (OLR), among which the GISS-E2-R model has the largest underestimation of 9 W/m2. The global mean OLR from the 12-model mean is smaller than that from the CERES EBAF by 2.7 W/m2. The CanAM4, HadGEM2-A, and MRI-AGCM3-2S perform better and have about or less than 1 W/m2 biases. The net downward clear-sky radiative flux in the 12-model mean is 2.1 W/m2 stronger than that of the CERES EBAF (Table 2). The overestimation of TOA clear-sky upwelling shortwave flux in most of the CMIP5 models may result from underestimation of atmospheric water vapor and ozone absorption in these models, a typical problem in older generation GCMs [Cagnazzo et al., 2007; Wild, 2008]. The underestimation of global mean clear-sky OLR in most of the CMIP5 models is partly because these GCMs compute the clear-sky radiative fluxes using methodology different from that for the CERES EBAF [e.g., Allan and Ringer, 2003]. Compared to the CERES EBAF in which the clear-sky TOA fluxes are strictly based on clear-sky sampling of the CERES and MODIS measurements, all GCMs compute clear-sky radiative fluxes by simply setting cloud fraction to zero and repeating the radiative transfer calculations. The atmosphere thus could be more moist than that under true clear-sky condition. Since clear-sky OLR in the models is determined by not only the specific model code but also the vertical profiles of air temperature and humidity, the more moist model environment contributes to a negative clear-sky OLR bias [Zhang et al., 1994]. In terms of spatial pattern, these 12 models agree with the CERES EBAF very well for both the TOA clear-sky upwelling shortwave flux (0.9–0.97) and the clear-sky OLR (~0.99).
 Under all-sky condition, nine of the 12 models produce stronger TOA upwelling shortwave radiative flux than the CERES EBAF, and the multimodel mean bias is 2.2 W/m2. Eight of the 12 models underestimate OLR, with the multimodel mean being weaker by 0.9 W/m2. The net downward radiative flux at the TOA in the 12-model mean (0.6 W/m2) is slightly smaller than that of the CERES EBAF (0.8 W/m2).
 The contrast between the clear-sky and all-sky fluxes shows the substantial effects of clouds on the TOA radiative fluxes in these CMIP5 models. There is considerable model spread for both the TOA upwelling shortwave flux and OLR. It is notable that the all-sky OLR biases in most of the CMIP5 models are smaller than the corresponding clear-sky OLR biases. This is due to the aforementioned deficiency in the methodology used by the models to compute the clear-sky radiative fluxes, as well as the fact that the all-sky radiative fluxes in the models have been tuned against satellite estimates on a global mean basis. In terms of spatial pattern correlation, the inclusion of cloud effects considerably deteriorates the spatial pattern for TOA upwelling shortwave flux (reduces from 0.9–0.95 for clear sky to 0.4–0.9 for all sky); clouds only slightly worsen the spatial pattern for OLR (reduces from 0.99 for clear sky to about 0.96 for all sky).
 The comparison of TOA shortwave CRF, longwave CRF and net CRF between the 12 AMIP simulations and the CERES EBAF in terms of global mean and global spatial pattern correlation are shown in Figure 2. Here we use a modified format of the Taylor diagram [Taylor, 2001], referred to as “modified Taylor diagram” hereafter. Instead of using the radius to show ratio of spatial variance of two fields as defined in the standard Taylor diagram, we use the radius to show the ratio of regional means of two fields, i.e., the ratio of regional mean over a specified spatial domain for a model to that for the CERES EBAF. This is because regional mean is a more straightforward and commonly used measure than spatial variance. We use the cosine of angle to show the spatial pattern correlation between an AMIP simulation and the CERES EBAF over the specified spatial domain, as in the standard Taylor diagram. Figure 2 and Table 2 show that the 12-model mean value of global mean shortwave CRF is only slightly strong than that of the CERES EBAF (by 0.4 W/m2), suggesting the CMIP5 models provide a good simulation of global mean shortwave CRF. For longwave CRF, however, the 12-model mean value is smaller than that from the CERES EBAF by 1.9 W/m2. The magnitude and the spatial pattern correlation for the net CRF are notably worse than the shortwave and longwave CRF. The net CRF difference between the CERES EBAF and the 12-model mean is dominated by the LW CRF difference, and on a global mean basis the 12-model mean overestimates the TOA cooling effect by 2.3 W/m2 (Table 2).
 For all the TOA radiative fluxes as well as the CRFs except for the incoming solar radiation, the multimodel mean bias is considerably smaller than the intermodel differences (Table 2); the CERES EBAF falls within the spread of the 12 models. Overall, the CMIP5 models provide a fairly good simulation of globally averaged CRFs at TOA. Such good model simulations are partially due to artificial model tuning against satellite estimates in terms of global means [e.g., Wild, 2008].
 The comparison between the CMIP5 models and the CERES EBAF for clear-sky and all-sky shortwave and longwave radiative fluxes in the tropics (30°S–30°N) is shown in Figure 3 and Table 3. Similar to the global mean (Figure 1), all the CMIP5 models consistently have higher TOA incoming solar radiation, with the multimodel mean greater than the CERES EBAF by 1.1 W/m2. The overall model simulations in the tropics (Figure 3) are not as good as those for the global domain (Figure 1), which is not surprising as these models are tuned against satellite estimates based on global means rather than regional means. Under clear-sky conditions, nine of the 12 models overestimate the upwelling shortwave radiation, with a multimodel mean bias of 1.7 W/m2. The spatial correlations for these 12 models range from 0.86 for the CNRM-CM5 to 0.97 for the HadGEM2-A. Nine of the 12 models underestimate the tropical mean clear-sky OLR, and the multimodel mean bias is -2.0 W/m2. The spatial pattern correlations range from 0.76 for INMCM4 to 0.92 for MRI-AGCM3-2S, not as good as those for the shortwave flux. Similar to the clear-sky fluxes, most of the CMIP5 models overestimate the TOA all-sky upwelling shortwave and underestimate the OLR, with the multimodel mean biases being 5.2 W/m2 and -1.0 W/m2 respectively. The contrast between clear-sky and all-sky fluxes shows that the inclusion of clouds deteriorates model simulation of shortwave radiative flux, yet improves model simulations of OLR. The poorer model simulations of clear-sky OLR relative to all-sky OLR again reflect the aforementioned deficiency in the approach that most GCMs use to compute the clear-sky OLR.
Table 3. Same as Table 2, but for Tropical Mean Values
12-Model Mean CERES EBAF (/CERES EBAF)
σ (12 Models)
All-Sky TOA Radiative Fluxes
Clear-Sky TOA Radiative Fluxes
 The modified Taylor diagrams evaluating the AMIP simulations of TOA shortwave CRF, longwave CRF and net CRF in the tropics are shown in Figure 4. Over half of the 12 models overestimate the shortwave CRF and underestimate the longwave CRF. The net result is a distinct overestimation of the cooling effect of clouds in the 12 models, with the IPSL-CM5A-LR model being the exception which shows a net CRF only 50% of that of the CERES EBAF. The multimodel mean overestimates the shortwave CRF by 3.6 W/m2 and underestimates the longwave CRF by 1.0 W/m2, resulting in an overall overestimation of net cooling effect of the clouds by 4.6 W/m2. In terms of spatial correlation, the model simulations for the longwave CRF are considerably better than those for the shortwave CRF. This is not surprising, as the longwave CRF is mainly impacted by high clouds whereas the shortwave CRF is affected by clouds at all altitudes.
3.1.3 Spatial Distribution of TOA CRFs
 The TOA CRF differences between the CERES EBAF and the 12 models are further investigated by examining their spatial distribution in Figure 5. The shortwave cooling effect (Figure 5a, left) is mainly present over tropical deep convective regions, the Intertropical Convergence Zone (ITCZ), subtropical marine stratocumulus regions and midlatitude storm track regions. The longwave warming (Figure 5a, middle) is present over tropical deep convection regions where clouds can reach rather high altitude and have cold cloud tops. While adding the shortwave cooling and the longwave warming results in a large amount of cancellation, the shortwave cooling dominates resulting in a net cooling effect (Figure 5a, right), particularly over the subtropical stratocumulus regions and extratropical storm track regions.
 While the 12-model mean (Figure 5b) has a good agreement with the CERES EBAF in large-scale features (Figure 5a), there are considerable differences in regional details (Figure 5c). Compared to the CERES EBAF, the 12-model mean considerably overestimates the shortwave cooling effect over the subtropical central Pacific, the ITCZ, western tropical Indian and Atlantic ocean, and strongly underestimates the shortwave cooling over the eastern side of subtropical Pacific and Atlantic, southeastern Indian Ocean off the southwest coast of Australia, and the Southern Hemisphere (SH) extratropical oceans. Interestingly, the tropical and subtropical regions that the multimodel mean overestimates the shortwave cooling, i.e., the subtropical central Pacific, western tropical Indian and Atlantic Oceans, also show stronger longwave warming effect. The above suggests that the CMIP5 models may simulate stronger tropical convection and more high clouds over these regions. This is supported by Liu et al. , who found more precipitation in the CMIP5 AMIP simulations compared with observational estimates over these regions. We will further investigate this through cloud fraction comparison with the GOCCP observations in section 3.4. Over the Maritime Continent and nearby area, the multimodel mean shortwave CRF and longwave CRF are smaller than those from the CERES EBAF, suggesting weaker model tropical convection there. We also note that there are considerable intermodel differences over this region (Figure 5d), likely reflecting the diversity of convective parameterizations used in these models. The net CRF difference between the multimodel mean and the CERES EBAF is dominated by that of shortwave CRF, and is characterized by stronger cooling effects over the majority of tropical and subtropical regions, and weaker cooling effects over the eastern subtropical Pacific, southeastern subtropical Atlantic, southeastern Indian Ocean, as well as the extratropical SH ocean. Figure 5d shows that there are considerable intermodel differences in shortwave CRF over the majority of the tropics, including the tropical convection region, ITCZ, regions to the northeast of SPCZ, and subtropical eastern oceans where low clouds prevail. The intermodel differences for the longwave CRF are considerably smaller than those for the shortwave CRF, and are mainly over tropical regions where the multimodel mean overestimates the longwave warming. The intermodel differences for the net CRF are dominated by those for the shortwave CRF.
3.2 Comparison of Tropical CRFs Composited by Dynamical Regimes
 To further investigate the model simulations of TOA CRFs and how they differ from those in the CERES EBAF in different tropical circulation regimes, we next stratify the TOA shortwave CRF, longwave CRF and net CRF in the tropics using ω500. To ensure enough data samples for different dynamical regimes, the monthly data over the period January 2001 to December 2008 are used. Since the majority of the features to our interest are over the oceans, we show in Figures 6 and 7 the results based on tropical ocean only. We note that the results for the entire tropics do not differ notably from those for the tropical ocean.
 We first examine how well the CMIP5 models simulate the Probability Distribution Function (PDF) of ω500 (Figure 6). The PDF of ω500 in the ERA-Interim reanalysis shows that the tropics is dominated by weak vertical motion (-10 to 25 hPa/day), with ω500 at 20 hPa/day accounting for the largest areal fraction (Figure 6b). The PDF's of ω500 in most of the 12 models (Figure 6b) are largely similar to that of the ERA-Interim reanalysis except over subsidence domain where the model simulations tend to place slightly stronger weight onto larger ω500. The CNRM-CM5, GISS-E2-R and INMCM4 models are the exceptions that tend to place less weight onto regions with weak descent and more weight onto regions over weak ascent. Figure 6a shows that regions with weak vertical motion are mainly at the transitional zones between the tropical deep convective regions over western tropical Pacific and Atlantic warm pools and eastern side of subtropical ocean basins where there is strong subsidence. Because of their huge statistical weight in fractional area, any radiative flux changes over these weak vertical motion regions could impose a large impact on the tropical mean radiative budget. It is interesting to note that the main differences between the 12-model mean and the CERES EBAF, particularly the CMIP5 model overestimation of the shortwave CRF (Figure 5c), are just over these weak ascent and descent regions.
 The composite results of CRFs for the CERES EBAF (Figure 7) show that the amplitudes of both the shortwave CRF and longwave CRF decrease with ω500 from strong convective regime to weak subsidence regime, yet are independent of ω500 over regions with strong subsidence. For all the tropical dynamical regimes, the shortwave cooling dominates the longwave warming resulting in a net cooling effect. The ratio of shortwave cooling to longwave warming is particularly large over regions with weak and moderate subsidence, peaking at a value of 4 when ω500 is about 30 hPa/day.
 Most of the CMIP5 models reproduce the dependence of shortwave CRF on ω500 over deep convective and weak subsidence regimes (Figure 7a), though with different slopes. Over strong subsidence regimes, in contrast to the CERES EBAF shortwave CRF which is independent of ω500, most of the CMIP5 models continue to show a decrease with ω500 and are more dynamically controlled. This is likely due to the lack of subtropical low stratocumulus clouds in these models which will be investigated using the GOCCP data in section 3.4. The GFDL-HIRAM-C360 is the only model that agrees with the CERES EBAF very well for all dynamical regimes including the independence of ω500 over the strong subsidence regimes. By contrast, the MIROC5 model deviates from the CERES EBAF the most and provides remarkably stronger shortwave CRF for all regimes. Comparing to the CERES EBAF, the 12-model mean shows slightly stronger shortwave CRF over tropical convective regions and considerably weaker shortwave CRF over strong subsidence regions. Over regions with weak subsidence which accounts for rather large tropical fractional area, the shortwave CRF from the majority of the 12 models and their multimodel mean is distinctly greater than that from the CERES EBAF. Thus, the relatively good AMIP simulations of tropical mean shortwave CRF in Figures 4 are essentially a result of compensating errors in different tropical dynamical regimes. The slight overestimation of tropical mean shortwave CRF in the multimodel mean is a competing result of overestimation over the weak subsidence and deep convective regimes and underestimation over the strong subsidence regimes.
 As for the longwave CRF, most of the models simulate the dependence on ω500 well for all tropical dynamical regimes. There is a considerable spread in the tropical convective regions though, an indication of various representations of tropical convective processes in these models. The NorESM1-M and GISS-E2-R models are outliners which respectively simulate the strongest and weakest longwave warming. The 12-model mean considerably underestimates longwave CRF over regions with ascent and strong subsidence, and is comparable to the CERES EBAF over the transitional regions where there is weak subsidence. This results in the underestimation of tropical mean longwave CRF in Figure 4. The underestimation of longwave CRF over strong ascent regimes by most models is again associated with their use of different sampling methodologies from the CERES EBAF [Allan and Ringer, 2003].
 In terms of net CRF, the multimodel mean indicates a stronger cooling effect than the CERES EBAF over convective and weak vertical motion regions, but notably weaker cooling effect over the strong subsidence region. The former is mainly contributed by the underestimation of longwave CRF over tropical convective regions and overestimation of shortwave CRF over weak subsidence regions, whereas the latter is due to the stronger underestimation of shortwave CRF than that of the longwave CRF over the strong subsidence regions. There is substantial spread among the models. Due to its rather strong overestimation of shortwave CRF, the MIROC5 model again stands out and shows the strongest overestimation comparing to the CERES EBAF.
 The ratio of shortwave CRF to longwave CRF compares the magnitudes of these two CRFs, and has important implications on the net CRF. It is also invariant to cloud fraction and removes cloud fraction effect [Cess et al., 2001]. Figure 7d shows that the ratio of shortwave to longwave cloud forcing is about 1.4 in strong ascent regimes. It gradually increases to a peak of 4.0 in the moderate descent regime (ω500 = 30 hPa/day), and then reduces to about 3.0 in strong subsidence regimes. For all the dynamical regimes, the ratio in the multimodel mean is consistently larger than that in the CERES EBAF. The typical ratios for the multimodel mean in regions with strong ascent, moderate descent and strong descent are respectively 1.6, 4.3, and 3.4. Most of the models resemble the multimodel mean. There are a few outliners. Related to its strong underestimation of longwave warming effect, the GISS-E2-R model strongly overestimates the ratio for all dynamical regimes. Owing to its strong underestimation of shortwave CRF in strong subsidence regimes, the IPSL-CM5A-LR model shows notably smaller ratio than the CERES EBAF there. We will show in section 3.4 that the above CRF bias in the IPSL-CM5A-LR model results from its lack of low-level clouds.
3.3 Comparison of Relative Humidity
 The relative humidity is both temperature and moisture dependent, and provides information on boundary layer, cloud properties and dryness of the subtropical free troposphere. To better understand the CRFs in the CMIP5 AMIPs, we next examine composites of the vertical distribution of relative humidity in different tropical circulation regimes defined from ω500 in the model simulations, and compare with those from the AIRS composited by ω500 from the ERA-Interim reanalysis (Figure 8). Monthly data over the period January 2003 to December 2008 are used for the analysis.
 The relative humidity in the AIRS (Figure 8a) shows the transition from the rather shallow and moist boundary layer in subtropical strong subsidence regime (>20 hPa/day), to deepened boundary layer in the transitional trade cumulus regime (-10 to 20 hPa/day), and to the rather moist deep convective regime (<-40 hPa/day) where the relative humidity exceeds 90% in the upper (150 hPa) and lower (925 hPa) troposphere. The 12-model mean (Figure 8b) captures the general characteristics of the AIRS relative humidity-ω500 distribution (Figure 8a), but shows distinct quantitative differences. In the deep convective regimes, compared to the AIRS, the multimodel mean has smaller relative humidity and lacks the sharp vertical gradient in both upper and lower troposphere. For the subtropical stratocumulus and trade cumulus regimes, the multimodel mean has a considerably deeper boundary layer and weaker vertical relative humidity gradient in the lower troposphere. In the subtropical free troposphere, the multimodel mean is not as dry as the AIRS. The difference between the 12-model mean and the AIRS in Figure 8c shows that for all dynamical regimes, the 12-model mean has notably higher relative humidity from about 900 hPa to 200 hPa. Rather large positive biases (>15%) occur over strong ascent regimes at about 250 hPa, as well as over descent regimes (10 to 60 hPa/day) at about 850 hPa which is associated with the deeper boundary layer in the CMIP5 models than that indicated in the AIRS data. There are distinct dry biases near surface (below 900 hPa) for all circulation regimes. By comparing vertical profiles of air emperature and specific humidity from the AIRS with those from the CMIP3 coupled model simulations, past studies [Pierce et al., 2006; John and Soden, 2007] found that most of the CMIP3 models also simulate a large cold and moist bias in the free troposphere but a dry bias in the boundary layer. Above 200 hPa, the 12-model mean has distinct negative bias over ascent and weak descent regimes that peak at 150 hPa, and moderate positive bias over strong descent regions. Given the relatively coarse resolution of relative humidity in the lower troposphere, we roughly define the boundary layer height as the 50% relative humidity level for the stratocumulus and trade cumulus regions [Siebesma et al., 2004]. Figure 8d shows that over regions with strong descent, most of the models and the 12-model mean have deeper boundary layer than the AIRS. Over the weak vertical motion regimes, a notable feature is that all of the CMIP5 models consistently have higher boundary layer caps than the AIRS. Furthermore, the CMIP5 models consistently have higher relative humidity than the AIRS throughout the entire troposphere except near 150 hPa. Observations show that while the transitional regimes are dominated by shallow convection and the rather dry air above the moist low boundary layer inhibits development of deep convection, there are occurrences of deep convection and high clouds over this region [Takayabu et al., 2010; Ichikawa et al., 2012]. The deeper low boundary layer and more moist free troposphere in the transitional weak vertical motion regimes in most of the CMIP5 models could provide an environment that facilitates more frequent occurrence of deep convection and high clouds. This is further investigated in section 3.4 in which we evaluate the model cloud fraction using the GOCCP data.
3.4 Comparison of Cloud Fraction
 To better understand the CRF differences between the CMIP5 models and the CERES EBAF, we utilize the GOCCP data and satellite simulator output of the six CMIP5 models and evaluate the model simulations of cloudiness. We note that the multimodel mean bias and intermodel differences for the CRFs based on the six models (not shown) are largely similar to those based on the 12 models (Figure 5).
3.4.1 High-Cloud Fraction
 Figure 9 compares the six models and the GOCCP in terms of annual mean high-cloud fraction climatology over the period 2007–2008. The high clouds in the GOCCP (Figure 9a) are mainly over deep convective regions, including the tropical Pacific warm pool, tropical South America, tropical Africa, ITCZ, South Pacific Convergence Zone (SPCZ), as well as parts of the Northern Hemisphere (NH) extratropical oceans. The geographic distribution of high clouds in the six-model mean is in good agreement with that of the GOCCP, with however considerably weaker amplitude. Their difference map (Figure 9c) clearly shows that while the six-model mean has less tropical mean high cloud than the GOCCP, it has more high cloud over subtropical central Pacific, western tropical Indian Ocean and Atlantic Oceans. The larger high-cloud fraction over these regions reflects more shortwave radiation back to the space as well as traps more longwave radiation from beneath, leading to model overestimations in both shortwave CRF and longwave CRF there (Figure 5). Over the Maritime Continent, the six-model mean has less high clouds than the GOCCP, which explains the model underestimation of shortwave and longwave CRF there (Figure 5). Figure 9d shows that the intermodel differences are mainly over western tropical Pacific and Atlantic. Interestingly, there are rather weak intermodel differences over the central subtropical Pacific where the six-model mean shows more high clouds (Figure 9c), suggesting that the overestimation of high-cloud fraction there (Figure 9c) is largely a common model bias. This is confirmed by Figure 10 which shows the high-level cloud fraction differences between each of the six models and the GOCCP. While these six models differ from each other greatly in total tropical cloudiness, all of them show more high cloud over the central subtropical Pacific, and indications of less high cloud over the Maritime Continent. In addition, most of these six models show more high cloud over western tropical Indian Ocean and Atlantic Ocean. This is associated with the CMIP5 model overestimation of TOA shortwave cooling and longwave warming over these regions (Figure 5).
3.4.2 Low Cloud Fraction
 Figure 11 compares the six models and the GOCCP in low-cloud fraction. The GOCCP (Figure 11a) shows that the low clouds are prevalent over the eastern part of subtropical Pacific, SH subtropical Atlantic, and southeastern Indian Oceans off southwest coast of Australia, as well as extratropical oceans. By comparison, the low clouds in the six-model mean (Figure 11b) are mainly over extratropical oceans with considerably smaller amount. There are essentially no low clouds over eastern subtropical oceans in the six-model mean. The difference map (Figure 11c) further highlights the lack of low clouds over the majority of global oceans, particularly the subtropical stratocumulus regions. The lack of low clouds over eastern subtropical oceanic basins in the CMIP5 models allow more incident shortwave radiation to reach the Earth's surface leading to considerably weaker shortwave cooling effects over these regions (Figure 5). Figure 11d shows that these six models exhibit large intermodel differences over these subtropical stratocumulus regions as well. The above suggests that similar to earlier generation AGCMs, the CMIP5 models continue to be challenged in simulating low-level subtropical stratocumulus clouds and their associated CRFs.
3.4.3 Cloud Fraction Composited by Dynamical Regimes
 To better understand the cloud differences between the six CMIP5 models and the GOCCP, we further stratify their total, high-, middle-, and low-cloud fraction using the ω500 (Figure 12). The monthly data over the period January 2007 to December 2008 are used to provide more samples. Comparing to the GOCCP, the CMIP5 models have notably less clouds, particularly at the middle and low levels. All the six models underestimate the total cloud cover in the tropics, particularly over subsidence regions, with the tropical mean of the six-model mean being about 75% of that of the GOCCP (Figure 12a). The IPSL-CM5A-LR simulates the smallest total cloud fraction. In terms of high cloud (Figure 12b), compared to the GOCCP, most of the six models simulate less high clouds in all circulation regimes. Among the six models, the MIROC5 model has the least tropical mean high cloud which is about 70% of that of the GOCCP. The MPI-ESM-LR is the only model that simulates more high cloud in most of the tropical dynamical regimes, and the CanAM4 model simulates more high cloud than the GOCCP only over deep convective regions. For the middle cloud (Figure 12c), all of the six models show considerably less cloud amount than the GOCCP, and the tropical mean of the multimodel mean is about 50% of that of the GOCCP. The IPSL-CM5A-LR model simulates the least amount of midlevel cloud; its tropical mean is only 25% of that of the GOCCP. For the low cloud (Figure 12d), all the six models except the MIROC5 model underestimate tropical mean. The IPSL-CM5A-LR model again stands out and simulates the least amount of low-level cloud, with its tropical mean being only 30% of that of the GOCCP. The lack of middle and low clouds in the IPSL-CM5A-LR contribute to its rather weak total cloud fraction (Figure 12a). While the MIROC5 model has larger tropical mean low-cloud fraction, these low, thick clouds are erroneously placed over tropical convection regions; the subtropical stratocumulus clouds are still notably weaker in this model. Despite its weak high cloud, the rather extensive low-level thick cloud over the majority of the tropical convection regions in the MIROC5 contributes to its strong overestimation of the shortwave CRF at the TOA (Figure 3). By comparison, the CanAM4 and HadGEM2-A models provide better performance (Figure 12d), with their tropical mean low cloud weaker than that of the GOCCP by 13% and 37% respectively (not shown).
5 Conclusion and Discussion
 In this study, the annual mean climatology of TOA cloud radiative effects in 12 CMIP5 AMIP simulations are evaluated and investigated, with the focus on the tropics (30°S–30°N). The satellite-derived observations used for the model evaluation include the CERES EBAF TOA radiative fluxes, the GOCCP cloud fraction and the AIRS relative humidity. Results show that the CMIP5 models provide fairly good simulations of large-scale regional mean TOA radiative fluxes and CRFs. Compared to the CERES EBAF, most of the CMIP5 models show small global mean and tropical mean biases, with the CERES EBAF falling within the spread of the 12 models. The small regional mean biases in most of the CMIP5 models and the multimodel mean show the model overestimation of the shortwave CRF, underestimation of the longwave CRF, and a stronger net cooling effect by clouds.
 Further analysis on the tropical region shows that the good model simulations of tropical mean CRFs are a result of compensating errors over different dynamical regimes. Most of the CMIP5 models considerably overestimate the shortwave CRF over regimes with weak vertical motions, yet strongly underestimate the shortwave CRF over the subtropical stratocumulus region. The former dominates resulting in the overall moderate overestimation of tropical mean shortwave CRF in most of the models and the multimodel mean. The underestimation of tropical mean longwave CRF in most of the CMIP5 models and the multimodel mean is contributed by the moderate underestimation over tropical convective region and strong underestimation over the strong subsidence region. Therefore, these CMIP5 models are challenged in simulating regional details of CRFs.
 The above model simulations of CRFs are closely related to their simulations of cloudiness and associated processes. The CMIP5 models produce considerably less cloud amount compared to the GOCCP, particularly in the lower and middle troposphere. Over the Maritime Continent, most of the models simulate moderately less high-cloud fraction than the GOCCP, possibly due to weaker deep convection in the models. This leads to smaller shortwave cooling and longwave warming and a larger net cooling than the CERES EBAF. Over subtropical strong subsidence regimes, consistent with many early generations of GCMs, most of the CMIP5 models strongly underestimate stratocumulus cloud amount, resulting in their strong underestimation of local shortwave CRF. Over the transitional trade cumulus regimes, a notable feature is that while at varying amplitudes, most of the CMIP5 models consistently simulate deeper boundary layer, more moist free troposphere (higher relative humidity), and more high clouds over the subtropical central Pacific, western tropical Indian Ocean and western tropical Atlantic Ocean, and consequently overestimate shortwave cooling and longwave warming effects there. While most of the CMIP5 models show the same sign as the multimodel mean, there are substantial model spreads, particularly over the tropical deep convective and subtropical strong subsidence regimes.
 Our results suggest that several major issues in the CMIP3 models and other earlier versions of GCMs continue to exist in the CMIP5 models, though the amplitude of the biases may have been reduced. Over the tropical deep convective regions, the CMIP5 models continue to exhibit substantial intermodel differences likely due to the different convective parameterizations they use. Over subtropical stratocumulus regions, similar to earlier versions of GCMs [e.g., Karlsson et al., 2008; Teixeira et al., 2011], while exhibiting considerable intermodel differences the CMIP5 models show lack of low clouds and correspondingly strongly underestimate the shortwave CRF there. Over regions with weak vertical motions, the CMIP5 AMIPs agree with the CMIP3 coupled model simulations [Ichikawa et al., 2012] in that most of them overestimate the shortwave CRF. In addition, similar to the coupled CMIP3 model simulations [Pierce et al., 2006; John and Soden, 2007], the CMIP5 AMIPs simulate more moist free troposphere yet drier low boundary layer in the tropics. Different from the CMIP3 coupled simulations [Stowasser and Hamilton, 2006; Ichikawa et al., 2012], however, most of the CMIP5 AMIPs consistently simulate deeper boundary layer, wetter free troposphere and more high clouds with subsequent biases in shortwave and longwave CRFs in the transitional regime. The above differences are partly due to the different SSTs they use. The CMIP5 AMIPs are forced with the same observed SST, whereas the CMIP3 coupled simulations predict SST from their ocean models and thus their CRFs are subjected to biases in modeled SST climatology. While the CMIP3 coupled model simulations of SST climatology is fairly realistic (less than 2 °C zonally averaged SST error in the multimodel mean for all latitudes), most of the coupled models have a large warm bias in the eastern parts of the tropical ocean basins, near the continental boundaries. Such SST bias is associated with insufficient model resolution, which leads to problems in the simulation of the local wind stress, oceanic upwelling and underestimation of the low-cloud amounts (IPCC AR4 [IPCC, 2007]). Relative to the AMIP simulations which are free of SST biases, the SST biases in the coupled models considerably contribute to the biases of the coupled model simulations as well as their intermodel differences. The common model bias in the CMIP5 AMIPs in the weak vertical motion regimes is rather intriguing and is a subject of our future research. Overall, representing clouds and their TOA radiative effects continues to be a challenge in the CMIP5 models.
 This study is supported by the NASA CLOUDSAT and CALIPSO Science Team Recompete program (NNH09ZDA001N-CCST). We acknowledge the World Climate Research Programme's Working Group on Coupled Modeling, which is responsible for CMIP, and we thank the climate modeling groups (listed in Table 1 of this paper) for producing and making available their model output. For CMIP the U.S. Department of Energy's Program for Climate Model Diagnosis and Intercomparison provides coordinating support and led development of software infrastructure in partnership with the Global Organization for Earth System Science Portals. The CERES EBAF edition 2.6r data were obtained from the NASA Langley Research Center CERES ordering tool at http://ceres.larc.nasa.gov/cmip5_data.php/. The CALIPSO-GOCCP data were obtained from the GOCCP Web site (http://climserv.ipsl.polytechnique.fr/cfmip-obs/). The modified Taylor diagrams were made using GrADS scripts from Bin Guan's GrADS Script Library (http://www.atmos.umd.edu/~bguan/grads/GrADS_Scripts.html). We thank two anonymous reviewers, whose comments and suggestions have significantly improved this paper.