This study proposes a systematic approach to investigate cloud-radiative feedbacks to climate change induced by an increase of CO2 concentrations in global climate models (GCMs). Based on two versions of the Model for Interdisciplinary Research on Climate (MIROC), which have opposite signs for cloud–shortwave feedback (ΔSWcld) and hence different equilibrium climate sensitivities (ECSs), hybrid models are constructed by replacing one or more parameterization schemes for cumulus convection, cloud, and turbulence between them. An ensemble of climate change simulations using a suite of eight models, called a multiphysics ensemble (MPE), is generated. The MPE provides a range of ECS as wide as the Coupled Model Intercomparison Project phase 3 (CMIP3) multimodel ensemble and reveals a different magnitude and sign of ΔSWcld over the tropics, which is crucial for determining ECS.
It is found that no single process controls ΔSWcld, but that the coupling of two processes does. Namely, changing the cloud and turbulence schemes greatly alters the mean and the response of low clouds, whereas replacing the convection and cloud schemes affects low and middle clouds over the convective region. For each of the circulation regimes, ΔSWcld and cloud changes in the MPE have a nonlinear, but systematic, relationship with the mean cloud amount, which can be constrained from satellite estimates. The analysis suggests a positive feedback over the subsidence regime and a near-neutral or weak negative ΔSWcld over the convective regime in these model configurations, which, however, may not be carried into other models.
The general circulation model, or the global climate model (GCM), is a unique tool for physically based simulations of the earth’s climate. GCMs have been improved during the past three phases of the Coupled Model Intercomparison Project (CMIP) (Reichler and Kim 2008) and extensively used in the Intergovernmental Panel on Climate Change (IPCC) Assessment Reports (ARs; Solomon et al. 2007). While many aspects of the climate simulated in GCMs, such as temperature and wind fields, are much more realistic than in the past, the representation of clouds remains one of their largest limitations. Indeed, the current IPCC-class models show a substantial divergence in terms of sign and magnitude of the cloud–radiative feedback in response to an increase in atmospheric CO2 concentration (e.g., Bony and Dufresne 2005; Soden and Held 2006; Webb et al. 2006).
Since uncertainty in the dynamical core will be small despite diversity in its configuration in the current generation GCMs, the diversity of the cloud feedback is recognized to arise mostly from different parameterization schemes for unresolved physical processes in the atmosphere. While the CMIP phase 3 (CMIP3) provides experimental data from 23 GCMs and enables various analyses, the multimodel ensemble (MME) alone is insufficient to understand the source of cloud feedback diversity because the models are structurally different from one another.
There are three alternatives for dealing with the diversity in the cloud feedback and equilibrium climate sensitivity (ECS) in GCMs. Given that the CMIP3 models show cloud feedback with increasing CO2 levels, varying both in magnitude and sign, each of these approaches first selects one or two particular models. When we perturb model parameters without changing the model code and perform CO2 doubling (or equivalent) runs with each set of parameters, the model ensemble helps quantify the range of uncertainty in the feedback processes due to limited knowledge of certain parameter values (e.g., Murphy et al. 2004; Stainforth et al. 2005). This type of ensemble, called a perturbed physics ensemble (PPE), has been generated by several modeling groups (Collins et al. 2010; Yokohata et al. 2010; Sanderson et al. 2010; Klocke et al. 2011), all of whom show that PPE is useful in quantifying climate change uncertainties due to model parameters. However, PPE is not necessarily suitable for exploring the feedback mechanism in the base model, on which the ensemble property crucially depends (Yokohata et al. 2010).
The second approach is to simplify the model’s configuration from a realistic GCM to an idealized aqua planet, and to a single column, sharing the parameterization schemes. The use of such a hierarchy of models is relevant for understanding mechanisms of cloud feedback in a chosen GCM, as long as the simplified models reproduce the cloud and cloud–radiative properties in a full model (Zhang and Bretherton 2008; Medeiros et al. 2008). Brient and Bony (2012, hereafter BB) showed that a single-column model (SCM) based on L’Institut Pierre-Simon Laplace Coupled Model, version 5a (IPSL CM5A) GCM can reproduce the vertical profile of cloud fraction over the subsidence regime in the GCM. They then clarified the mechanism of the decrease of low clouds in the global warming simulation found therein. While the dominant process controlling cloud feedback in their GCM may not be operating in others (Wyant et al. 2009), the hierarchical modeling provides a process-based understanding of the cloud feedback.
The third approach, adopted in the present work, maintains the same level of complexity in the model configuration, but attempts to trace the source of different behavior between two GCMs. This is accomplished by replacing one or more parameterization schemes in the two models, and then evaluating the cloud feedback from each of the hybrid models. This ensemble, called a multiphysics ensemble (MPE) throughout this paper, directly encompasses the structural difference of the models, and is therefore conceptually different from a PPE. The MPE would be particularly helpful when we have models coded in a similar manner (e.g., different versions of a GCM, but exhibiting very different cloud feedback). While some studies have applied MPE to numerical weather prediction (Houtekamer et al. 1996; Stensrud et al. 2000), few have investigated diversity in the climate feedback using MPE because of the cost of constructing the hybrid models. Recent work by Gettelman et al. (2012) is an exception: they swapped cloud macro–microphysics, radiation, aerosol, turbulence, and shallow convection schemes between two versions of the National Center for Atmospheric Research (NCAR) Community Atmosphere Model (CAM) to find the reason for their different climate sensitivities. They found that the newer version of CAM5 has a higher ECS associated with the positive cloud–shortwave feedback (ΔSWcld) over the trade cumulus region and the midlatitude storm tracks, which is mainly due to the updated shallow convection scheme.
With the aim of contributing to the CMIP phase 5 (CMIP5) and the IPCC Fifth Assessment Report, we have continuously developed our GCM, called the Model for Interdisciplinary Research on Climate (MIROC). In a new version of MIROC5, many climate aspects have been improved by not only increasing the resolution but also updating the parameterization schemes (Watanabe et al. 2010). Of particular interest is that MIROC5 has a lower ECS (2.6 K) than the previous version, MIROC3.2 (3.6 K), and this is attributed to the difference in the cloud–shortwave feedback (Watanabe et al. 2011, 2012). In the present study, the MPE is constructed on the basis of these two models to understand crucial processes controlling the cloud–shortwave feedback, and hence ECS. We have also made PPEs using both MIROC3.2 and MIROC5 (Yokohata et al. 2010; Shiogama et al. 2012), enabling us to analyze them together in some parts of the paper, which will demonstrate the efficacy of the MPE.
The present paper is organized as follows. In section 2, ensembles based on the two versions of MIROC and the experimental designs are described. In section 3, we evaluate the range of ECS among PPEs and the MPE, and demonstrate that the cloud–shortwave feedback over the tropics is a major factor for different ECSs between the PPEs. The mean cloud fields and their response to surface warming are then analyzed in section 4. Composites sorted by the circulation regime indicate that a different set of the coupled physical processes has a dominant role in modulating cloud response at the convective and subsidence regimes. In section 5, a nonlinear relationship is identified between the cloud–shortwave feedback and the mean cloud amount in each regime, which is used to discuss the relative credibility of the cloud feedback mechanisms in the two versions of MIROC. Section 6 gives the discussion and conclusions.
2. Model ensembles
a. MIROC3 PPE
MIROC3.2, which was used for the CMIP3, has been jointly developed at the Centre for Climate System Research (CCSR),1 the University of Tokyo, the National Institute for Environmental Studies (NIES), and the Japan Agency for Marine-Earth Science and Technology (JAMSTEC) (K-1 Model Developers 2004). When generating the PPE, the ocean component model has been replaced by a 50-m-deep slab ocean, and the horizontal resolution of the atmospheric component has been reduced from T42 to T21 to save computational effort. To ensure a realistic mean climatology, the flux adjustment was applied to sea surface temperature (SST) and sea ice distributions.
Among various techniques to perturb the system, the MIROC3.2 PPE (referred to as MIROC3 PPE-S in this study, where S stands for the slab ocean GCM) is generated following the methods of Annan et al. (2005) and Hargreaves et al. (2007) in the Japan Uncertainty Modeling Project (JUMP). Specifically, the model mean states were constrained by assimilating observations when determining optimal sets of perturbations for 13 parameters in the atmospheric component. Further details are described in Yokohata et al. (2010). We use 32 members, each of which consists of a 70-yr control and a 2 × CO2 runs.
b. MIROC5 PPEs
A PPE using MIROC5 was recently generated to evaluate the parametric uncertainty of the ECS obtained from the official version of MIROC5. This PPE can also be used for comparison with the MIROC3 PPE-S and PPEs from other GCMs in future work. As in the MIROC3 PPE-S, the horizontal resolution of the atmosphere model was reduced from the standard configuration (from T85 to T42) after we confirmed that the essential property of the feedback and ECS is unchanged by this reduction. Unlike the previous PPE, however, we attempt to use the full coupled model to avoid any artificial influence of the flux adjustment on the ECS (Jackson et al. 2011). A thorough description of the method for generating perturbations is given by Shiogama et al. (2012), and is briefly explained below.
The procedure is divided into two parts. First, an ensemble of the atmosphere model (MIROC5 PPE-A, where A stands for the atmosphere model) was generated by varying a single parameter to its maximum and minimum values, as determined by experts’ judgment. This was repeated for 20 preselected parameters and one logical switch to yield a 42-member ensemble (including the standard setting), each consisting of a 6-yr-long integration of the control, 4 × CO2, and warmed SST runs. The length of the integration is relatively short, but we have confirmed with the standard setting that the results are not significantly altered when the integration is extended to 30 yr.
In the former two runs, the SST and sea ice concentration are prescribed by the control climatology of the full MIROC5, whereas the SST run is driven using the monthly climatology of SST obtained from an abrupt 4 × CO2 coupled model experiment (years 11–20). Time-mean differences in the top-of-atmosphere (TOA) radiative budgets for the last 5 yr between the control and 4 × CO2 runs define the radiative forcing, and similarly the differences between the control and SST runs scaled by the global-mean surface air temperature (SAT) difference give the feedback in this ensemble (cf. section 3).
Another set of the ensemble is made with the full GCM (MIROC5 PPE-C, where C stands for the coupled model), in which the atmospheric component is the same as that used for constructing PPE-A. With a reduced set of 10 parameters, 5000 perturbation samples are generated using the Latin hypercube sampling technique. To avoid climate drift, the N samples with the smallest radiative imbalance at TOA, estimated using a linear emulator of the MIROC5 PPE-A, are selected (cf. Shiogama et al. 2012). The above requirement ensures that the global-mean SAT is not significantly different from the standard experiment without observational constraints unlike MIROC3 PPE-S. We set N = 35, and the 30-yr control integration (initial 10 yr are the spinup period and are excluded from the analysis) and 20-yr abrupt 4 × CO2 runs are carried out for each member. The radiative forcing and feedback are then calculated using the difference in annual-mean fields between the two runs, following Gregory et al. (2004).
c. MIROC5 MPE
In generating the MPE, we use the T42L40 atmosphere component of MIROC5 as a base model. The parameterization schemes for three processes in MIROC5—cumulus convection, large-scale condensation (LSC) and cloud microphysics (treated altogether), and turbulence—are reverted to those in MIROC3.2. References and major properties of the schemes are summarized in Table 1. Briefly, each of the schemes implemented in MIROC5 has a greater number of degrees of freedom (e.g., time-dependent entrainment profile in cumuli, explicit treatment of cloud liquid and ice, prognostic turbulent kinetic energy). Unlike some GCMs, we have not implemented a specific scheme for shallow cumulus clouds, but the new convection scheme in MIROC5 is expected to represent these to some extent (Chikira and Sugiyama 2010). The atmosphere model of MIROC5 is different from that of MIROC3.2 in respect to some other physical processes. For example, an updated radiation code calculates the radiative heating more accurately, and the aerosol module was upgraded to include a prognostic scheme for determining the cloud droplet and ice crystal number concentrations, which are important for the indirect aerosol effect (see Watanabe et al. 2010 for details). However, we restrict our attention to the cloud–radiative interaction in this study, and thus do not change these schemes.
The resulting ensemble is called MIROC5 MPE-A, and consists of eight slightly different models (see Table 2), including the standard MIROC5 (STD). The abbreviations CLD, CNV, and VDF indicate that the cloud (LSC and microphysics), cumulus, and turbulence schemes are respectively replaced by the corresponding old routines. When two of these schemes are replaced, the model is denoted as CLD+CNV, CNV+VDF, or CLD+VDF. The model CLD+CNV+VDF, in which all three schemes have been reverted, is the closest to MIROC3.2 in terms of the representation of cloud-related processes. We have chosen MIROC5 as a base model of the MPE-A for practical reasons, so that changes from STD to other models are inevitably retrospective. To avoid a large radiative imbalance (allowable imbalance is ±2 W m−2) at the TOA, we retuned each model by slightly modifying a few parameters among the 13 and 10 control parameters in MIROC3 PPE-S and MIROC5 PPE-C, respectively. The radiative forcing and feedback are evaluated in a similar manner to MIROC5 PPE-A, using the atmosphere-only integration.
3. Climate sensitivity
As a prelude to a more in-depth look at the MIROC5 MPE-A, the ECS of various ensembles is compared in this section. Recall that the definition of ECS () is given by the global-mean energy equation,
where denotes the change in the net TOA radiation when doubling the CO2 level (positive value means heating the system) and −α is the total feedback parameter in W m−2 K−1. In atmospheric GCM (AGCM) experiments such as our MPE, these quantities are obtained from the time-mean differences in F between the control and CO2 runs and between the control and SST runs divided by the global-mean SAT difference. ECS is then estimated using (1).
ECS in the standard version of MIROC3.2 is relatively high in the range obtained from the CMIP3 ensemble [2.5–6.3 K, following Gregory and Webb (2008)]. This results in MIROC3 PPE-S yielding an ECS range of 4.5–9.6 K (Table 2). In contrast, MIROC5 PPE-A and PPE-C have smaller ECS spreads: 2.3–3.1 and 2.2–2.7 K, respectively. Even though these produce a spread of radiative forcing and feedback comparable to MIROC3 PPE-S, the ECS is proportional to −α−1 and hence the MIROC5 PPEs based on a low-sensitivity base model generate the reduced diversity in ECS. The fact that the ECS ranges of the MIROC3 PPE-S and MIROC5 PPEs do not overlap (cf. Table 2) indicates that the structural differences between the two base models are greater than the uncertainty range due to model parameters. The MPE was constructed precisely to span this gap. The ECSs of the MIROC5 MPE-A range over 2.3–5.9 K. The value of ECS in STD is the lowest and within a range of ECSs in MIROC5 PPE-A, while the ECS of CLD+CNV+VDF is the highest, at 5.9 K, which is above the range of the MIROC5 PPEs but within that of MIROC3 PPE-S. The values from other models are 2.5 K in CLD, 2.4 K in CNV, 2.3 K in VDF, 3 K in CLD+CNV, 2.9 K in CNV+VDF, and 4.2 K in CLD+VDF.
ECSs of the various ensembles are plotted as a function of the radiative forcing and the total climate feedback in Fig. 1, on which isolines of for a given and α are also imposed. The values in MIROC3 PPE-S are at the high end while those in MIROC5 PPEs are at the low end (green and red symbols, respectively) compared with the scattering of the CMIP3 MME. Unlike the CMIP3 MME showing substantial differences in both and α (see also Gregory and Webb 2008), the difference in our PPEs is attributable to a different magnitude of the total feedback, but not the radiative forcing. The spreads in and α in MIROC3 PPE-S and MIROC5 PPE-C are narrower than those in CMIP3 MME, but wider than in MIROC5 PPE-A. As expected, the MPE-A fills the gap between MIROC3 PPE-S and the MIROC5 PPEs, despite the fact that and α are somewhat different from one another. It is intriguing that and α are negatively correlated within the respective ensembles of MIROC3 PPE-S and MIROC5 PPE-C, which has been pointed out by Shiogama et al. (2012).
Forcing and feedback associated with individual components of the radiative fluxes (not shown) reveal that the primary component responsible for the different ECS in MIROC5 MPE-A is the cloud–shortwave feedback, ΔSWcld. This is clearly seen from a scatter diagram of ECS and ΔSWcld for eight models (Fig. 2, marked by ×). The correlation between the two quantities reaches 0.85, with the highest value of ΔSWcld in CLD+CNV+VDF. When ΔSWcld is decomposed into tropical (30°S–30°N) and extratropical (30°–90°S and 30°–90°N) components, the latter is always negative and does not differ much among the models, whereas the former is highly correlated with ECS as in the global mean, suggesting the dominant role of the tropical cloud response. In contrast to the negative ΔSWcld in STD (−0.31 W m−2), two models (CLD+VDF and CLD+CNV) show positive values of ΔSWcld (0.53 and 0.20 W m−2, respectively) close to that of the CLD+CNV+VDF model (0.56 W m−2). Since ΔSWcld is weakly positive in CLD and CNV but nearly neutral in VDF, the effect of coupling between two processes is nonlinear. It is thus likely that the major source of diversity in ΔSWcld is a coupling between subgrid-scale processes, rather than one single process. This argument is consistent with the conclusion of Zhang and Bretherton (2008) who used a multiphysics SCM to examine the cloud feedback in CAM.
The horizontal distribution of ΔSWcld is compared among eight models in Fig. 3, from which we identify the following differences. The old cloud scheme in CLD does not modify the overall pattern of ΔSWcld, but strengthens the positive feedback over the subtropical oceans and tropical continents (Fig. 3b). The replacement of the convection scheme changes the sign of ΔSWcld (from negative to positive) over the convective regions, whereas the different turbulence scheme in VDF appears to have little effect on ΔSWcld (Figs. 3c,d). The effects of these individual schemes persist when we couple them (Figs. 3e–h). However, the coupling effect of these processes does not work additively, and hence suppresses or amplifies the regional change in ΔSWcld.
Before conducting a thorough analysis of the dependence of ΔSWcld on the circulation regime in the next section, we can look at the relative contribution of ΔSWcld in different regimes (Fig. 4). By referring to the 500-hPa vertical p velocity (ω500), every grid over 30°S–30°N is classified into one of four regimes: strong subsidence (ω500 > 30 hPa day−1), weak subsidence (0 < ω500 < 30), weak ascent (−40 < ω500 < 0), and strong ascent (ω500 < −40 hPa day−1). It is shown in Fig. 4 that ΔSWcld is positive in the subsidence regime of all the models, despite the different magnitudes. However, ΔSWcld in the convective (i.e., ascent) regimes is positive in some models and negative in others. Because the weak subsidence regime occurs the most frequently (Fig. 4b), the positive ΔSWcld in this regime will dominate, as emphasized in previous studies (e.g., Bony and Dufresne 2005). Nevertheless, the sign of ΔSWcld in this regime does not change in the MIROC5 MPE-A. This is consistent with the finding of Watanabe et al. (2012), who showed that low clouds decrease for the 4 × CO2 case in both MIROC3.2 and MIROC5. The different sign of the tropical-mean ΔSWcld in the MPE-A (Fig. 2) thus comes from diversity in the weak ascent regime, in addition to the different magnitude of positive ΔSWcld in subsidence regimes.
4. Regime analysis for the cloud and cloud–shortwave feedback
a. Regime dependence of cloud response and ΔSWcld
It has been shown that ΔSWcld in many GCMs is primarily due to changes in cloud fraction at a given circulation regime but not changes in the circulation regime as represented by the probability density function (PDF) of ω500 (Bony et al. 2004; Bony and Dufresne 2005). Diversity in ΔSWcld associated with changes in clouds is therefore grasped by calculating the local regression of ΔSWcld on changes in high, mid-, and low cloud (ΔCh, ΔCm, and ΔCl) in MPE-A (Fig. 5). The cloud amounts were derived from the International Satellite Cloud Climatology Project (ISCCP) simulator with the definition following Rossow and Schiffer (1999). More negative ΔSWcld occurs in a model and a greater increase of cloud is observed, and vice versa, depending on the region. Changes in low clouds (ΔCl) control ΔSWcld over the subtropical cool oceans while ΔCm mostly affects the convective areas (ΔCh has the least connection to ΔSWcld).
For a deeper look at the differences in ΔSWcld across the models, we made composites of ΔSWcld and the associated ΔCh, ΔCm, and ΔCl with respect to ω500 over 30°S–30°N. The width of the ω bin is 5 hPa day−1, and the composites for the eight models in MIROC5 MPE-A are plotted together in Fig. 6. The spread of ΔSWcld is large for both the convective (ω500 < 0) and subsidence (ω500 > 0) regimes (Fig. 6a). All of the models show a positive ΔSWcld over the subsidence regions, whereas it is both positive and negative over the convective regions. There is a contrast in the diversity of cloud amounts between the middle/high level and low level—the former has a large spread over the convective regime and the latter varies more over the subsidence regime (Figs. 6b–d). A careful comparison of the ΔSWcld composite and the cloud changes reveals that the cloud–shortwave feedback in different circulation regimes is associated with the cloud change at different altitudes. For example, ΔSWcld is strongly positive in CLD+VDF over the subsidence regime, where ΔCl is remarkably negative (red lines in Figs. 6a,d). Also, two models (CLD+CNV and CLD+CNV+VDF) show a positive ΔSWcld over the convective regime accompanied by a reduction in Cm (Figs. 6a,c). In CLD+CNV+VDF, ΔCh is also negative, which might contribute to an amplification of ΔSWcld.
The cloud–shortwave feedback to ocean surface warming related to an increase in atmospheric CO2 occurs not only over the subtropical cool oceans, but also over the entire tropics. Thus, the variety in ΔSWcld cannot be explained solely by the change in low-level clouds. Indeed, Figs. 5, 6 illustrate that ΔSWcld is associated with the change in midlevel clouds over the convective regime. In the next section, we extend our regime analysis to examine changes in the vertical structure of clouds and their mechanisms.
b. Analysis using the saturation excess
Observations show that the spatial pattern and seasonal cycle of the subtropical Cl are closely related to the inversion strength above the planetary boundary layer (PBL), as measured by lower-tropospheric stability (LTS) or its variant of the estimated inversion strength (Klein and Hartmann 1993; Wood and Bretherton 2006). Previous studies applied this empirical relationship to interpret the response of Cl to global warming in GCMs (e.g., Wyant et al. 2009; Medeiros and Stevens 2011; Watanabe et al. 2012). However, LTS cannot be used to diagnose the change in the vertical profile of clouds. We therefore use a different in situ variable, which is more directly related to warm-phase cloud generation and dissipation in the model.
In many LSC schemes, a PDF with respect to subgrid-scale liquid temperature (Tl) and total water (qt) is used to calculate cloud fraction (C) based on the “fast condensation” assumption (see Watanabe et al. 2009, and references therein). By referring to the grid-scale saturation excess (Qc) and the higher PDF moments (μi), C is expressed as
where f is a nonlinear function depending on the base distribution of the PDF. The saturation excess is defined as
where p is pressure, qt is the grid-scale total water, qs is the saturation specific humidity, L is the condensational heat, and cp is the specific heat for air.
A joint PDF on the Qc– C plane demonstrates that the saturation excess can measure the cloud fraction well (Fig. 7). The relationship is clearer in PBL at η = 0.9 (η is the model’s hybrid σ–p coordinate and the value roughly corresponding to 900 hPa) and less so above. As C also depends on the PDF moments (cf. Watanabe et al. 2009) and is modified by other microphysical processes during a model time step, the PDF is not exactly fitted by the theoretical curve expected from the LSC scheme. Yet, Fig. 7 shows that Qc provides a good measure for C, not only in terms of the spatial pattern but for temporal variability (not shown). The gradient of differs somewhat between the LSC schemes used in MIROC3 and 5, but this does not seriously affect the present analysis. A similar relationship is found between C and relative humidity, but with a lesser degree of the co-occurrence (not shown).
Particularly large ΔSWcld occurs in the two models of CLD+VDF and CLD+CNV, with respect to STD (Figs. 3, 6), so regime composites of C and Qc from these two models are compared with those from STD, focusing on their vertical profiles (Fig. 8). The composite of the cloud fraction in the respective control runs generally shows middle and high clouds over the convective regime and low clouds over the subsidence regime (Figs. 8a,d,g). A salient feature in STD is the coexistence of the three types of cloud at ω500 < 0: high cloud at around η = 0.2, middle cloud at η = 0.6–0.7, and low cloud above η = 0.8 (Fig. 8a). These are consistent with the observed trimodal structure of cumulonimbus, cumulus congestus, and shallow cumulus (Johnson et al. 1999), although Cl is somewhat overrepresented. In CLD+VDF, high clouds are exaggerated and low clouds form near the surface, particularly in the subsidence regime (Fig. 8d). These changes from STD are caused by the old turbulence scheme, which tends to simulate shallow PBL, and the old cloud scheme, which does not implement cold rain microphysics and thereby overestimates ice clouds (Watanabe et al. 2010). In CLD+CNV, the overall cloud structure is similar to that in CLD+VDF, but lacks a sharp cumulus congestus peak and shallow cumulus clouds in the convective regime (Fig. 8g). It has been confirmed that the standard cumulus scheme in MIROC5 generates more congestus and shallow cumulus clouds than the old scheme (Chikira and Sugiyama 2010).
The change in the cloud fraction (ΔC) in STD is shown in Fig. 8b (shading) imposed on the contour of the mean cloud fraction. As commonly found in global warming experiments, deep cumulus clouds (both congestus and cumulonimbus) are shifted to higher altitudes because of the increased moist adiabatic temperature profile (cf. Hartmann and Larson 2002). The shift results in an increase and decrease of cloud above and below the mean position, respectively, without large changes in Ch and Cm because of their cancellation (cf. Figs. 6b,c). However, low cloud is increased over the convective regime. The changes in C are well measured by the change in Qc, except for the ice cloud at η < 0.2 (Fig. 8c). The mechanism of the change in Qc is examined further in section 4c.
The composite of ΔC in CLD+VDF is similar to that in STD, except for a stronger contrast between positive and negative changes and an opposite sign at low levels (Fig. 8e). This is reasonable, because ΔSWcld is different between the two models due to the opposite sign of ΔCl (Fig. 6). It is worth noting that the composites of ΔQc from STD and CLD+VDF have a similar structure, even at low levels (Fig. 8f). In STD, the positive low-level ΔQc occurs where clouds exist in the control run, and therefore serves to increase C. In CLD+VDF, however, clouds form in the control run beneath the level of the positive ΔQc. Clouds are insensitive to Qc in the unsaturated condition (Fig. 7), so that ΔQc cannot act to amplify the low clouds at η = 0.8–0.9.
Similarly, the variation in ΔC where ω500 > 0 can be explained by the difference in the peak altitude of the mean clouds. The negative ΔQc in the lower troposphere at ω500 > 0 is not uniform, being larger near the surface and where η < 0.8 (Figs. 8c,f). The mean cloud (i.e., climatology in the control run) in CLD+VDF is generated at η > 0.9, where ΔQc is large. This causes the cloud reduction to occur. The difference in altitude of the mean cloud in STD and CLD+VDF is associated with the different PBL depth, which tends to be thinner when the lower-order turbulence closure is used. In summary, our comparison of ΔC and ΔQc between STD and CLD+VDF indicates that the mean cloud structure, and whether it is formed at the height where ΔQc works effectively, is the key to the opposite ΔCl (and hence ΔSWcld) behavior.
In CLD+CNV, high clouds are shifted upward and low clouds over the convective regime change little in response to the positive Qc (Figs. 8h,i). A major difference from STD (and also CLD+VDF) is the change in the middle cloud, which shows a marked decrease at η = 0.6–0.7. The middle cloud tends to decline at the peak altitude of the mean cloud in all the models. The middle clouds are much broader in CLD+CNV than in the other two models, and their decrease is not compensated for by an increase in the upper levels at around η = 0.45. The ΔCm takes its largest negative value in this model (Fig. 8h), which explains the strong positive ΔSWcld over the convective regime. It is thus likely that the response of the cumulus congestus is different between models adopting different convection schemes, which is another crucial factor in the diversity of ΔSWcld.
c. Common change in thermodynamic condition
where H denotes relative humidity and ql is the cloud water. The overbar and Δ indicate values from the control run and the difference between the control and SST runs, respectively. Following the decomposition in (4), ΔQc may be explained by four effects, corresponding to each term in the rhs of the second equation: a temperature effect, a relative humidity (RH) effect, a condensate effect, and the Clausius–Clapeyron (CC) effect. The CC effect works through a reduction in αL with increasing Tl, and vice versa.
To examine the reasons for positive ΔQc where ω500 < 0 and negative ΔQc where ω500 > 0 commonly found in the lower troposphere (Figs. 8c,f,i), regime composites of the above terms are calculated. Figure 9 presents three of these terms, and their sum, from STD. The condensate effect is found to be negligible, and hence is not shown. In response to the ocean surface warming, radiative cooling in the free troposphere is known to strengthen over the subsidence region (Zhang and Bretherton 2008; Wyant et al. 2009; Brient and Bony 2012). The enhanced clear-sky longwave cooling is associated with the tropospheric warming, which works to reduce Qc (Fig. 9a). The negative temperature effect is similarly found near the surface because of increased sensible heat. The CC effect, which comes from nonlinearity in the Clausius–Clapeyron relationship, is roughly the opposite of the temperature effect (Fig. 9b), although weaker in magnitude. The sum of the above two effects is dominated by the temperature effect, and shows a uniform negative contribution to Qc in the PBL (not shown).
Unlike the other terms, the RH effect has positive and negative values in the PBL at ω500 < 0 and ω500 > 0, respectively (Fig. 9c). This contrast is reflected in the sum of the three terms (Fig. 9d), which reproduces the actual structure of ΔQc well (cf. Fig. 8c). Considering that ΔH can be decomposed into , the positive contribution indicates that the moisture increase is greater than the increase in saturation humidity, and vice versa for the negative contribution. The RH effect seen in the middle level corresponds to the upward shift of clouds, in addition to an enhanced cloud re-evaporation at η = 0.55, which has been confirmed from the composites of the water vapor and cloud water tendency terms (not shown). There is a theoretical study that supports roughly constant free-tropospheric RH in global warming (Sherwood and Meyer 2006), but the RH change responsible for Fig. 8c is small: about 2% in the PBL and 8% in the middle troposphere. These changes are allowable in theory, but may be sufficient to change the cloud properties.
5. Possible constraint to the cloud–shortwave feedback
The results of the analysis presented in the previous section highlight the active role played by the changes in Cl and Cm over the convective regime, and the change in Cl over the subsidence regime, in understanding ΔSWcld. In this section, we again use the models available in the MPE to identify a more generalized relationship in the ensemble.
Figure 10 shows a scatter diagram of ΔSWcld against both ΔCl + ΔCm and ΔCl, according to the circulation regime. The condition of weak ω500 may be a mixture between the convective and subsidence regimes, but we simply use a threshold of ω500 = 0 to partition the two regimes. The composite is also plotted for MIROC3 PPE-S and MIROC5 PPE-A, but not for MIROC5 PPE-C because the regional change in this ensemble is contaminated by the natural variability arising from a full atmosphere–ocean coupling.
Overall, ΔSWcld is negatively correlated with the cloud changes in both regimes; this is not surprising because more cloud will reflect more shortwave radiation and vice versa. This negative relationship is even found in each ensemble, with the exception of the MIROC5 PPE-A over the subsidence regime, where ΔSWcld and ΔCl remain almost unchanged (red circles in Fig. 10b). With the MIROC5 MPE-A, we obtain estimates of a −1.4 and −1.2 W m−2 decrease of SWcld per 1% increase of Cl + Cm and Cl over the convective and subsidence regimes, respectively. These estimates do not change much when we use two PPEs together. It is interesting to note that ΔSWcld is positive in most cases, indicating an additional process that contributes to solar insolation without changes in the cloud amount. A possible explanation for this is the so-called cloud masking effect (Soden et al. 2004), which occurs because of the different sensitivities of clear- and cloudy-sky shortwave radiation to changes in water vapor and albedo, but not cloud fraction.
Given a crude dependence of ΔC upon the vertical structure of the mean cloud fraction (cf. section 4b), we intuitively assume that there is a systematic relationship between the mean cloud states and their change, and hence ΔSWcld (e.g., Yokohata et al. 2010 ). To confirm this idea, scatter diagrams are produced of the regime-sorted ΔSWcld against mean cloud amounts, and , over the respective regime obtained from the annual-mean climatology in the control runs (Fig. 11). The mean cloud amounts can be compared with the ISCCP data (Rossow and Schiffer 1999), which are represented Fig. 11 by thick vertical lines with gray shading to indicate the range of the interannual variability for 1984–2007.
Over the convective regime, MIROC5 MPE-A fills the gap between two PPEs, one showing less and larger ΔSWcld (MIROC3 PPE-S) and the other with more and near-neutral ΔSWcld (MIROC5 PPE-A) (Fig. 11a). Possible reasons for the nonmonotonic dependence of ΔSWcld on , as approximated by a second-order polynomial (black curve in Fig. 11a), are as follows: the sign and magnitude of ΔCm are related to the sharpness of , which is proportional to itself (cf. Fig. 8). The negative slope for > 30% is therefore interpreted to imply that less mean cloud (i.e., a less sharp vertical structure) accompanies the net decrease rather than the upward shift of Cm, and hence the positive ΔSWcld (cf. Fig. 8). The contribution of ΔCl is of secondary importance over the convective regime, but can be significant when is sufficient at levels where a positive ΔQc appears (η = 0.8–0.9; Fig. 8). In summary, a combination of these effects due to ΔCl and ΔCm results in the nonlinear dependence of ΔSWcld on the mean clouds, with the maximum value of ΔSWcld occurring somewhere in the middle. An intersection of the extrapolated curve with the mean cloud obtained from ISCCP suggests that a near-neutral or weakly negative ΔSWcld is plausible over the convective regime. However, the reasons responsible for the diversity in MPE-A are specific to our models, and the validity of the curve estimated in Fig. 11a for the MME should be severely tested.
Likewise, the sensitivity of ΔSWcld to the mean low cloud over the subsidence regime is nonlinear (Fig. 11b). Again, ΔSWcld is positive with less (<20%) in MIROC3 PPE-S, whereas MIROC5 PPE-A shows a nearly neutral ΔSWcld with = 30%, having a small spread. All the MPE models exhibit positive ΔSWcld, with the minimum values occurring in models with a moderate amount of . Over the subsidence regime, is roughly proportional to the mean PBL depth in MPE (not shown, but the PBL height varies from 950 to 1200 m among the eight models). The inversion strength is larger with deeper PBL, so that this proportionality will be reasonable. However, negative ΔQc is not uniform in the PBL but is larger near the surface and the free troposphere. Thus, models generating clouds at too-low (η > 0.9) or too-high (η < 0.8) levels are likely to show a greater decrease in Cl. A comparison of with the ISCCP climatology suggests that a model showing a weak positive feedback is plausible, which corresponds to being generated at around η = 0.85. For further constraints, validation of the PBL height and the dominant processes that control the height will be desirable (Medeiros et al. 2005).
The nonmonotonic relationship between ΔSWcld and the mean cloud amounts suggests the former depending on the latter in MPE-A, which otherwise was not identified in each PPE. However, the possible processes responsible for the above relationship will heavily depend on how they are represented in parameterization schemes, so simply carrying our result into MMEs should be done with caution. This is discussed further in the next section.
6. Concluding discussion
In this study, we constructed a model ensemble in which each of eight atmospheric models is structurally different. The differences were systematically formed by replacing one or more parameterization schemes for the atmospheric processes (i.e., cumulus convection, cloud, and turbulence) between two versions of MIROC. This ensemble, called MIROC5 MPE-A in the present paper, enabled us to connect two base models showing opposite cloud–shortwave feedback ΔSWcld to global warming, and was used to explore the cause of diversity in ΔSWcld in the IPCC-class GCMs.
Climate change simulations carried out using the MIROC5 MPE-A and results from MPE-A and three other PPEs based on MIROC, lead to the following conclusions:
MPE-A spans the gap in terms of ECS between two PPEs on which MPE-A was based. This indicates the validity of MPE for systematically exploring the structural differences between different climate models.
The difference in the tropical ΔSWcld, which is the major driver for the wide range of ECS in the MPE-A as well as in the many IPCC-class models, is controlled by the mutual iteration of at least two processes but not by any single process. In particular, the coupling of the cloud scheme with either turbulence or the convection scheme was found to be critical, while the details of how these processes are coupled when modifying the cloud behavior were not fully clear.
There appears to be a dependence of ΔSWcld in the ensemble on the differences in the mean vertical structure of clouds in the control simulation. The latter often generated opposite signs of the cloud response to a similar change in the thermodynamic field (section 4). For each of the circulation regimes, ΔSWcld and cloud changes in MPE had a nonlinear, but systematic, relationship with the mean cloud amount, which may be constrained using satellite estimates. The analysis suggests a weak positive feedback over the subsidence regime and a near-neutral or weak negative ΔSWcld over the convective regime (section 5).
It will be essential to evaluate the processes controlling the cloud response to global warming to understand the diversity in ΔSWcld. Possible key parameters are the change in the PBL height and the PBL wetness. If the PBL gets thicker because of destabilization with increasing SST, this will work to increase low clouds (Xu et al. 2010). The PBL can also be thinner, and hence Cl decreases, if the cloud-top entrainment is weakened for any reason (Lauer et al. 2010). Such changes in the PBL depth appear to occur in fine-resolution models that resolve a subtle change in the inversion height. The PBL height in MPE changes by about ±10 m, which is much less than the change found in Lauer et al. (2010). Another factor, the PBL wetness change, is determined by a balance between the increased moisture sources from surface evaporation and the altered advection of dry air from the free troposphere (Wyant et al. 2009; Brient and Bony 2012). In the MIROC5 MPE-A, the PBL over the subsidence regime became drier when the ocean warmed, which supports the positive low-cloud feedback. However, we further determined that the magnitude of ΔSWcld varies depending on the mean Cl and the PBL height because the change in the saturation excess is not uniform within the PBL.
Because the strong subsidence regime occupies a small area in the tropics (cf. Fig. 4), some recent studies have emphasized the importance of the change in shallow cumulus cloud that occurs for weak ω500 (Medeiros et al. 2008; Gettelman et al. 2012). While many studies examined the feedback associated with shallow cumulus or stratocumulus over the subsidence regime, the cloud composite shown in Fig. 8 clearly indicates that the shallow convective clouds are smoothly extended to the convective regime, where the cumulus congestus and cumulonimbus sometimes overlap. It is therefore reasonable that ΔSWcld over the convective regime changed the most when the cumulus convection scheme in STD was replaced (i.e., CLD+CNV). Despite the possibility that this sensitivity to the convection scheme is different in a model implementing the shallow convection scheme, one of our conclusions—that the change in the middle cloud (or the cumulus congestus) in the transition region is also a crucial factor for different ΔSWcld among the models—is likely to hold in other models, given an accumulated result in previous studies (Medeiros et al. 2008; Klocke et al. 2011). While all the CMIP3 models underrepresent cumulus congestus, and thus its importance for cloud–shortwave feedback, we might expect the updated models in CMIP5 to show a higher sensitivity of ΔSWcld to cumulus congestus in transition circulation regimes.
We have demonstrated that our MPE has advantages over a PPE in understanding sources of different cloud feedbacks in GCMs. There are, however, limitations of the present work. One is the insufficient number of possible hybrid configurations. While the diversity in ECS in our PPEs was covered by the MPE-A, we should extend the present approach by reverting other parts of the GCM for a thorough investigation. Furthermore, the MPE has a similar weakness to PPEs in a qualitative sense. Namely, the diversity in the ensemble crucially depends on the two base models that are to be connected by the MPE. For example, we tried to constrain ΔSWcld by using the mean cloud amount, which may have a different curve between ΔSWcld and either or when the base models are different. In this regard, the constrained ΔSWcld for the convective and subsidence regimes may still be tentative. A similar difficulty was encountered in Klocke et al. (2011) when applying the relationship found in a PPE to an MME. Ideally, an MPE based on multiple GCMs could eventually fill the gaps among structurally different GCMs in the CMIP3 MME. In reality, such modeling is not possible unless a coordinated collaboration between the modeling centers is established.
The authors are grateful to three anonymous reviewers for their constructive comments. This work was supported by the Innovative Program of Climate Change Projection for the 21st Century (“Kakushin” program), Grant-in-Aid 23310014 and 23340137 from MEXT, Japan, and by the Mitsui & Co., Ltd. Environment Fund C-042. The computation was carried out on the Earth Simulator, NEC SX at NIES, and HITACHI HA8000 at the University of Tokyo.
Renamed the Atmosphere and Ocean Research Institute as of April 2011.