1. Introduction

Water vapor provides most of the greenhouse effect of the earth’s atmosphere. Clouds are a major contributor to the planetary albedo (Kiehl and Trenberth 1997). A small change in these radiative effects of water vapor and clouds can either offset or greatly amplify the perturbation to the earth’s radiation balance from anthropogenic effects (Houghton et al. 2001). Therefore, it is imperative for climate models on which our economical polices are increasingly relying to narrow the uncertainties in their simulations of the feedbacks from water vapor and clouds. Toward that objective, we have to critically evaluate how well the existing leading climate models simulate the feedbacks from water vapor and clouds.

Two methods have been employed to shed insight onto the question how well climate models simulate the feedbacks from water vapor and clouds. The first one is to check the differences in the feedbacks of water vapor and clouds in global warming among different models. A pioneering study using this method was carried out by Cess et al. (1990, 1996). Their analysis revealed that the cloud feedbacks differ greatly among models, while the globally averaged feedback from water vapor in the models follows that of a constant relative humidity model. A later study by Soden and Held (2006) reached the same conclusion for the Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report (AR4) models (Meehl et al. 2007). These results underscore the uncertainties in the cloud feedbacks in the climate models, but do not address the question which model has the right cloud feedbacks. Another limitation of these results is that consistency in the simulation of water vapor feedback does not rule out the possibility that all the models have a biased water vapor feedback.

The second method is to compare the response of water vapor and clouds to SST changes over the time scales for which observational data are available. A frequently used natural signal in the SST is El Niño warming (Sun et al. 2003; Sun and Held 1996; Soden 1997; Held and Soden 2000; Sun et al. 2006). By comparing the observed changes in the water vapor and clouds with those from the model with the observed SST boundary conditions [Atmospheric Model Intercomparison Project (AMIP) simulations], these studies suggest that the sign of the water vapor feedback in the GCMs is probably correct on the time scale of ENSO, even when averaged over the entire tropics. The study of Sun et al. (2006) shows, however, that models tend to overestimate the positive feedback of water vapor over the immediate region of the El Niño warming. The study of Zhang and Sun (2008) further shows that at least for the National Center for Atmospheric Research (NCAR) models, the overestimate of the positive feedback of water vapor during El Niño warming is due to an excessive response of upper-tropospheric water vapor to the surface warming. A more serious concern raised by the study of Sun et al. (2006) is the finding of a common bias in the simulation of the cloud albedo feedback in the leading climate models: with the exception of the Geophysical Fluid Dynamics Laboratory (GFDL) model, all of the models they analyzed in that study underestimate the response of cloud albedo to the surface warming. Nonetheless, the finding that at least the GFDL model may have a cloud albedo feedback as strong as the observed strikes an optimistic tone.

The study of Sun et al. (2006) used the AMIP simulations. Using AMIP simulations of these models to estimate feedbacks has an inherent limitation: the feedbacks are the feedbacks operating in the immediate neighborhood of the observed climatology. Because the mean climate is free to drift to the state that is in turn determined by the feedbacks, the feedbacks may change in the process of integration of the coupled model. In other words, if the coupled system—the models thereof, in particular—is not strictly a linear feedback system, and if the SST of the equilibrium state of the coupled run differs significantly from the observed, the feedbacks estimated about the equilibrium state of the coupled runs could be significantly different from those estimated from the corresponding AMIP runs. The coupled models do have a significantly different climatological SST from that observed—they all have an excessive cold tongue in the equatorial central Pacific (Sun et al. 2006). The purpose of this paper is to further assess the feedbacks from water vapor and clouds over the tropical Pacific region by directly using the outputs from fully coupled runs, and to highlight the impact of the excessive cold tongue in the coupled models on the feedbacks of water vapor and clouds over that region. Lin (2007) recently examined the cloud albedo feedback over the same region using the shortwave forcing of clouds at the surface, but he did not examine the water vapor feedback. Also, in his calculation of the feedback from the shortwave forcing of clouds at the surface, he did not separate the seasonal signal from the interannual signal. It is also not clear whether the temporal variations are separated from the regional variations in his calculation. This mixing-together variability of different origins makes interpretation of the results more challenging. Another important study that is closely related to the present effort is that of Bony and Dufresne (2005). Using monthly mean midtropospheric vertical velocity, they first decomposed the tropical circulations into a series of dynamical regimes and then examined the sensitivity of the cloud forcing in these regimes to a change in local SST. They highlighted the importance of boundary layer clouds by noting that the response of these clouds to a change in local SST in the model simulations of the present climate disagree most with observations. The present effort is focused on the response of the clouds and water vapor over the central and eastern equatorial Pacific to El Niño warming. We will show that even over the region where the peak El Niño warming takes place, models have systematic biases in the simulated feedbacks from water vapor and clouds.

A more difficult question to address is then whether the biases in the water vapor and cloud feedbacks shown up in the ENSO variability in the models will be carried over to the feedbacks of water vapor and clouds in the simulated global warming. To shed some light on this question, we take note that there are significant variations among different models in the simulated feedbacks of water vapor and clouds on both time scales—ENSO and global warming. If the model that has the weakest cloud albedo feedback during its simulated ENSO variability is also the model that has the weakest cloud albedo feedback in its simulated global warming, and more generally the variations in this feedback simulated by different models are strongly positively correlated on these two time scales—ENSO variability and global warming—we may have a piece of evidence in hand to support a positive connection between biases in the cloud albedo feedback on these two different time scales. Conversely, if no such strong positive correlations are found, we may conclude that the feedbacks in ENSO are not a harbinger of the feedbacks in global warming. Thus, we will examine whether variations among the models in the feedbacks during ENSO and during global warming have significant correlations. The issue of whether the feedbacks on these two different time scales are linked has been addressed to some degree and from a different angle by Bony and Dufresne (2005) and Zhu et al. (2007). Bony and Dufresne (2005) compared the interannual feedbacks with climate feedbacks. They noted that the large intermodel differences in the feedback in the simulated global change in their weakly subsiding case are not presented in the feedback in the interannual variability. Zhu et al. (2007) showed that changes in the low cloud amount in response to the 1997/98 El Niño are different from the changes in the low cloud amount in response to simulated global warming. The former is more linked to the changes in the position of the convective activities, while the latter is more linked to a change in the vertical stability.

The paper is organized as follows. The methodology is briefly described in section 2. In section 3, we first report the estimates of the feedbacks of water vapor and clouds in ENSO variability using the coupled runs of those models analyzed in Sun et al. (2006). These coupled runs are control runs. We then do the same feedback analysis using the twentieth-century simulations by an expanded group of models from the IPCC AR4 archive (Meehl et al. 2007). We will see that the bias in the cloud albedo feedback and the bias in the water vapor feedback identified in Sun et al. (2006) exist in all of these coupled simulations—control runs of the original group as well as these twentieth-century simulations runs of the expanded group of models. In section 4, we will attempt to address the question of whether the variations among models in the simulated cloud and water vapor feedbacks in ENSO variability are correlated with variations in the feedbacks of water vapor and clouds in global warming. Summary and conclusions are provided in section 5.

2. Methodology

In estimating the feedbacks of water vapor and clouds in the ENSO cycle, we will follow Sun et al. (2006): we will use the response of tropical convection to ENSO forcing to obtain the feedbacks of water vapor and clouds associated with tropical convection. We will first analyze the group of coupled models whose AMIP runs we analyzed in the study of Sun et al. 2006. In that study, we examined nine AGCMs. Seven of the nine AGCMs have a corresponding fully coupled GCM whose control runs are available for our analysis. These models are the NCAR Community Climate System Model, version 1 (CCSM1; Boville and Gent 1998), NCAR CCSM, version 2 (CCSM2; Kiehl and Gent 2004), NCAR CCSM, version 3 (CCSM3) at, respectively, T42 and T85 resolution (Collins et al. 2006; additional information online at http://www.ccsm.ucar.edu/experiments/ccsm3.0/), the third climate configuration of the Met Office Unified Model (HadCM3; Collins et al. 2001), the French Institut Pierre Simone Laplace (IPSL) Coupled Model version 4 (CM4) model (Marti et al. 2005), and the GFDL Climate Model version 2.0 (CM2.0; Delworth et al. 2006). Unless explicitly stated, calculations for this group of models use a 50-yr-long segment of the control runs of the models.

We then extend the analysis to a larger set of coupled models using the twentieth-century simulations submitted to the third phase of the Coupled Model Intercomparison Project (CMIP3; Meehl et al. 2007) by various groups (available online at http://www-pcmdi.llnl.gov/ipcc/model_documentation/ipcc_model_documentation.php). We limit our analysis to those models that do not use flux adjustment. In the AR4 archive, there are 17 models that do not use flux adjustment, but only 13 of them have all of the variables that we need to compute the feedbacks from water vapor and clouds. After excluding one more model from these 13 models because the interannual variability in the cold tongue SST in this model is too weak to be used to calculate the feedbacks, we end up with an expanded group of 12 models. These models are the French Centre National de Recherches Météorologiques Coupled Global Climate Model, version 3 (CNRM-CM3) and IPSL-CM4; the U.S. GFDL CM2.0, GFDL CM, version 2.1 (CM2.1), Goddard Institute for Space Studies Model E-H (GISS-EH), and NCAR CCSM3; the Chinese Institute of Atmospheric Physics (IAP) Flexible Global Ocean–Atmosphere–Land System Model (FGOALS); the Japanese Model for Interdisciplinary Research on Climate 3.2, medium-resolution version [MIROC3.2(medres)] and MIROC 3.2, high-resolution version [MIROC3.2(hires)]; the German ECHAM5/Max Planck Institute (MPI); and the U.K. HadCM3 and Hadley Centre Global Environmental Model version 1 (HadGEM1). Details for each of these models, including the spatial resolutions for the atmosphere and ocean components of these models, can be viewed on the above-mentioned Web site.

3. Feedbacks in the ENSO cycle

Applying the same linear regression technique of Sun et al. (2006) to the simulations of tropical interannual variations by those coupled models, we obtain Table 1. The numbers in the parenthesis are the results from the corresponding AMIP runs that are available for us. The observational results listed in the table are obtained using the dataset of Zhang et al. (2004). This dataset is based on International Satellite Cloud Climatology Project (ISCCP) data (Rossow and Schiffer 1999) and covers a much longer period than the 4-yr-long Earth Radiation Budget Experiment (ERBE; Barkstrom et al. 1989) period used in Sun et al. (2006). The estimate of the feedbacks from the greenhouse effect of water vapor clouds from the ISCCP data are quite close to those from ERBE. The estimate of the negative feedback from the shortwave forcing from this extended period covered by ISCCP data is significantly larger than that estimated from ERBE—about −14 W m⁻² K⁻¹ from ISCCP versus −11 W m⁻² K⁻¹ from ERBE. Over the period that these two datasets overlap (i.e., the ERBE period), the feedback from the shortwave forcing estimated from ISCCP data is −12.5 W m⁻² K⁻¹. Thus, only part of the difference, ∼1.5 W m⁻² K⁻¹, is due to the length of the data used. Note that the shortwave forcing from ISCCP is derived from a radiation model constrained by ISCCP observations. Exact agreement between ISCCP data and the ERBE period is not guaranteed. It is probably safe to assume that the feedback from the shortwave forcing of clouds is between −11 and −14 W m⁻² K⁻¹.

The problems uncovered in the previous analysis also show up in this extended analysis. First, models tend to underestimate the strength of the negative feedback from cloud albedo. This is true whether one uses the estimate from ERBE or from ISCCP. The model that now has the strongest negative cloud albedo feedback is IPSL-CM4, but the feedback is only 50%–70% of the estimate from the observations. Compared to the corresponding value estimated from their AMIP runs, substantial weakening in the simulated strength of the cloud albedo feedback occurs in all of the four models that were identified as better models in the previous analysis using their AMIP runs (the NCAR CCSM3 at T85 resolution, HadCM3, IPSL-CM4, and GFDL CM2.0). This reduction in the strength of the cloud albedo feedback is particularly notable for GFDL CM2.0—the value changed from −15.43 W m⁻² K⁻¹, the value estimated from the AMIP run, to −6.14 W m⁻² K⁻¹, the value estimated from the coupled run.

Figures 1 and 2 provide, respectively, a basinwide view of the response of the shortwave forcing of clouds (C_s) to El Niño warming in the coupled models and in their corresponding AMIP runs. Using the data from Zhang et al. (2004) results in a pattern for the observations that is very similar to that from the ERBE data—a negative response of C_s in the central and eastern Pacific is flanked by a positive response in the western Pacific, the Southern Pacific convergence Zone, and the north subtropical Pacific. The positive response in the north subtropical Pacific is consistent with the finding by Bony and Dufresne (2005). Except in the region immediately adjacent to the east coast and slightly off equator, observations (Figs. 1a and 2a) do not show a positive feedback from C_s in the far eastern equatorial Pacific (5°S–5°N). A positive feedback from C_s in the far equatorial eastern Pacific was noted in Lin (2007), who used the same dataset as the present study, but had the seasonal cycle included in his calculation of the feedback from C_s. The absence of this feature in our calculation suggests that the positive feedback from C_s in the far equatorial eastern Pacific found in the study of Lin (2007) is from the seasonal cycle. Interestingly, many models do have a positive feedback on the ENSO time scale in the far equatorial eastern Pacific.

The spatial pattern of the response of C_s in the coupled models resembles that obtained from their corresponding AMIP runs, but the maximum response of C_s is located more westward in the coupled models by about 20°. With the exception of NCAR CCSM2 and CCSM3 at T42, the maximum response of C_s in the coupled simulations is also weaker than that in the corresponding AMIP runs. In the AMIP runs, the maximum response of C_s in NCAR CCSM3 at T85, HadCM3, IPSL-CM4, and GFDL CM2.0 has a value that exceeds −30 W m⁻² K⁻¹. In their corresponding coupled runs, however, the maximum response of C_s in these four better models is significantly reduced. The reduction in the maximum response of C_s in NCAR CCSM3 at T85, IPSL-CM4, and GFDL CM2.0 is about 10 W m⁻² K⁻¹. The reduction in the maximum response of C_s in HadCM3 is even higher (∼15 W m⁻² K⁻¹).

There are exceptions to this general weakening in the response of C_s, seen in NCAR CCSM2 and NCAR CCSM3 at T42. NCAR CCSM2 has hardly any negative response in its AMIP run, but now develops a substantially negative response in the region immediately west to the date line (150°–170°E). The improvements in NCAR CCSM3 at T42 are even more substantial. The response of C_s in this model is comparable to that in GFDL CM2.0. Such a “self-correction” of the cloud albedo feedback clearly indicates the importance of nonlinearity in the coupling between the atmosphere and ocean. While it is encouraging to see that ocean–atmosphere coupling allows a self-correction to take place, it is disappointing to see that this self-correction is limited to the two models whose cloud albedo feedback assessed from their AMIP runs (i.e., about the observed SST because AMIP runs use the observed SST as the boundary forcing) has the largest error. As indicated by the results from their AMIP runs, the negative feedback from cloud albedo barely exists in these two models in the immediate neighborhood of the observed SST (see the numbers in the parenthesis). All other models that have been judged from their respective AMIP runs to have a significantly negative feedback of cloud albedo at the observed SST are found to have an even weaker cloud albedo feedback at their respective equilibrium SST.

The general weakening of the response of C_s in the coupled models (relative to the values estimated from the AMIP runs) is linked to the weakened precipitation response. Figures 3 and 4 show, respectively, the precipitation response to El Niño warming in the coupled models and in the corresponding AMIP runs. Contrasting Figs. 3 and 4, one finds that the reduction in the maximum response of the precipitation in NCAR CCSM3 at T85, HadCM3, IPSL-CM4, and GFDL CM2.0 exceeds 30%. The location of the maximum response of precipitation in these coupled models also shifts westward relative to that in the AMIP runs. The general reduction in the precipitation and the westward shift of the response suggest that the excessive cold tongue in these models plays a role in further weakening the response of C_s to SST changes in that region. Lin (2007) also noted that coupling shifts the deep convection westward and results in a reduction in the precipitation over the equatorial Pacific. [Interestingly, the two models (NCAR CCSM2 and CCSM3 at T42) that have an improved cloud albedo feedback are the only two models that do not have a significant weakening in their maximum response in the precipitation.]

To further examine the suggestion that the weakened precipitation response over the cold tongue region is a cause of the weakened response in the shortwave forcing of clouds, we plotted the scatter diagram of precipitation and SST over the cold tongue region (Fig. 5a). The figure shows that about their respective climatology (i.e., near the zero point as plotted here), the precipitation increases with SST increases at a faster rate in the observations than that in the models. Only when the positive SST anomalies are very large does the rate of increase of precipitation with respect to SST in some models become comparable to that in the observations. The corresponding figure for the surface level solar radiation shows that the relationship between the surface solar radiation and SST mirrors the relationship between precipitation and SST (Fig. 5b). About its respective climatology, the surface solar radiation decreases at a faster rate in the observations than in the models.

The new estimates also confirm another common bias existing in the climate models: the overestimate of the positive feedback of water vapor. Comparing the estimates from the coupled runs with those from the corresponding AMIP runs reveals that estimating from their AMIP runs of the climate models can result in a stronger positive feedback of water vapor. For example, the new estimate of the water vapor feedback in the HadCM3 is now significantly closer to the observed value than that from its AMIP runs. Still, all of the models have a stronger water vapor feedback than that indicated in observations. The overestimate ranges from about 25% in NCAR CCSM2 to about 45% in NCAR CCSM3 and IPSL-CM4. Discrepancies of this magnitude cannot be accounted for by errors in the observations. In any case, the existence of a significant spread in this discrepancy among these models suggests that any agreement in the globally averaged values in this regard must be linked to error cancellations among different regions. Figure 6 shows the spatial pattern of the response of G_a. Clearly, the models do not just differ from the observations over the immediate region of surface warming resulting from El Niño, they also differ in the surrounding regions of the far western Pacific and the subtropical regions. Figure 7 shows the pattern of the greenhouse effect of water vapor (G_a) from the corresponding AMIP runs. Comparing the spatial pattern of G_a in these two figures indicates a westward shift of convection in the coupled models relative to their AMIP runs.

With the exception of the NCAR CCSM2 and CCSM3, the estimates using the coupled simulations also yield a lower value for the positive feedback from the greenhouse effect of clouds than from the corresponding AMIP runs. The decrease in the strength of the positive feedback from the greenhouse effect of clouds is consistent with the decrease in the strength of the negative feedback from the cloud albedo. The exceptional behavior in the two NCAR models in this regard is also consistent with the exceptional behavior in these two models in simulating the cloud albedo feedback (recall that the feedback from cloud albedo in these two models is estimated to be more negative than that estimated from AMIP runs). Figures 8 and 9 show, respectively, the spatial pattern of the response of the greenhouse effect of clouds (C_l) in the coupled and AMIP simulations of the models. The contribution to the increase in the value of as shown in Table 1 in CCSM2 and CCSM3 mainly comes from the western edge of the equatorial cold tongue region defined here (150°–250°E). Judging from Fig. 8, and the corresponding pattern of the precipitation response (Fig. 3), it is likely that the large positive initial bias in the cloud albedo feedback—the bias shown up in the corresponding AMIP runs—in these two models enhances the deep convection in the region of about 150°–170°E (where SST is already warm) to a degree that the weakening effect from the excessive cold tongue is offset.

Although the estimates from the coupled simulations yield a lower value of the feedback from C_l for the majority of the models, the lack of cancellation between the feedback from C_l and the feedback from C_s in the models, as first revealed in the estimates from the AMIP runs, continues to exist in the estimates from the coupled simulations because of the larger bias in the feedback from C_s in the new estimates. Column 5 of Table 1 lists the net clouds feedback (C_l + C_s). In the observations, the net cloud feedback is slightly negative. All of the models, judging from the estimates from the coupled simulations at least, have a positive net cloud feedback. The relationship between C_s and C_l in the models is biased. In contrast, not all of the models overestimate the feedback from the total greenhouse effect of water vapor and clouds (G_a + C_l) (column 3, Table 1). Two of the models are actually found to underestimate the combined feedback from G_a and C_l (NCAR CCSM2 and HadCM3). Clearly, the relationship between G_a and C_l is not the same in all of the models.

As in Sun et al. (2006), we also assess the corresponding feedback from the atmospheric transport. We deduce this feedback from the net surface heating from the energy balance equation of the atmosphere. We derive the net surface heat flux by combing the latent and sensible heat fluxes from the National Centers for Environmental Prediction (NCEP) with the radiative fluxes from ISCCP. All of the models underestimate the negative feedback from the net surface heat flux. All of the models in the table also underestimate the negative feedback from the atmospheric transport. Seen in the net atmospheric feedback and the feedback from the net surface heating , the discrepancy with the observations in the new estimates increases in the models that were identified as better models in this regard in the previous analysis (GFDL CM2, IPSL-CM4, HadAM3, and NCAR CAM3 with T85 resolution), but decreases in the models that were identified as the worse models in this regard (NCAR CAM1, NCAR CAM2, and NCAR CAM3). Still, no models in the new estimates have a regulatory effect that is comparable to the observations. The best model identified in the previous analysis, the GFDL CM2, however, remains the one that has the strongest the regulatory effect, though the new estimate suggests that this regulatory effect from deep convection in this model is still too weak compared to observations (about −10 W m⁻² K⁻¹ in the model versus about −22 W m⁻² K⁻¹).

We have also extended the analysis to more models by using the twentieth-century simulations submitted to the CMIP3 archive by various modeling centers (Meehl et al. 2007). We limit our analysis here only to those models that do not use flux adjustment and have all of the variables we need in the archive. We then eliminate one model that does not have a significant ENSO signal. This comes down to 12 models for our extended analysis. The results are summarized in Table 2. The numbers in the parentheses are the corresponding feedbacks from the AMIP runs of these 12 models. (Note that not all of these models have an AMIP run or have a complete set of variables from the AMIP runs.) Judging from the estimates from the coupled simulations, all of the models listed in Table 2 have a weaker cloud albedo feedback than what is seen in observations. Again, we see that the estimates using coupled simulations of ENSO result in a weaker cloud albedo feedback than that estimated from the corresponding AMIP simulations. Note that judging from the estimates from the AMIP simulations of this extended list of models alone, three models actually have a stronger cloud albedo feedback than what is seen in the observations. Still, most of the models already have a weaker cloud albedo feedback than observations in their AMIP simulations. All of the listed models in Table 2, with the exception of HadGEM1, also overestimate the feedback from water vapor. Thus, the two common biases identified in our earlier analysis are in fact prevalent among the coupled models. Regression maps like those in Figs. 3 and 4 for these models again show a weakened and more westward shift of the precipitation response in these coupled models than in their corresponding AMIP runs (not shown here to avoid redundancy).

4. Can feedbacks in ENSO be harbingers of the feedbacks in global warming?

The above analysis establishes that there are common biases in the simulated feedbacks in the ENSO cycle: models tend to underestimate the negative feedback from cloud albedo and overestimate the positive feedback from water vapor. The question is then whether these systematic biases will be carried over to the global warming simulations. To address this question, we examine whether variations in the feedbacks during global warming among different models are correlated with the variations in the same feedbacks during the ENSO cycle among these models. In other words, we examine whether the models that have a stronger feedback in ENSO than their peer models also tend to have a stronger feedback in global warming than their peer models. We have estimated the feedbacks in the global warming using the A1B scenario runs by computing the linear trends of the SST, the cloud shortwave forcing, and the greenhouse effect of water vapor over 2000–2100. We then obtain the feedback of cloud albedo as the ratio between the trend of the shortwave cloud forcing and the trend of the SST, and the feedback of water vapor as the ratio between the trend of the greenhouse effect of water vapor over the trend of SST. Figure 10a is a scatter diagram showing the variations in the cloud albedo feedback in global warming among models (the vertical axis) over the cold tongue region against the variations in the same feedback in the ENSO cycle over the same region (the horizontal axis). The figure shows that there are no correlations between the variations on these two different time scales. The CNRM-CM3 has the strongest negative feedback of cloud albedo in the ENSO cycle, but has almost the weakest negative feedback of cloud albedo in global warming. For the feedback from the greenhouse effect of water vapor, there is some correlation between the variations among model-simulated feedback in global warming and the variations among the model simulated feedback in the ENSO cycle (Fig. 10b), but the correlation is not strong (about 0.57). Although the models that have a relatively weaker feedback of water vapor generally line up [e.g., HadGEM1, MIRO3.2(medres), and MIRO3.2(hires)], the models that have a relatively stronger water vapor feedback do not. The model that has the strongest water vapor feedback during ENSO is the IPSL-CM4, but this model does not have the strongest feedback in the global warming (in fact, measured by the strength of the feedback during global warming, it only ranks as the eighth strongest).

We have also computed a global mean cloud albedo feedback—the ratio between the trend in the shortwave cloud forcing averaged over the global oceans and the trend in the global mean SST. Variations in this feedback among the models are not correlated with the variations in the feedback during the ENSO cycle over the cold tongue region (Fig. 11a). This is also true for the relationship between the global mean water vapor feedback during global warming and the regional water vapor feedback over the cold tongue in the ENSO cycle (Fig. 11b).

These results suggest that, at least in the models, the feedbacks of water vapor and clouds estimated from the ENSO cycle cannot be used as harbingers for the feedbacks of water vapor and clouds during global warming.

5. Conclusions

The extended calculation using coupled runs confirms the earlier inference from the AMIP runs that underestimating the negative feedback from cloud albedo and overestimating the positive feedback from the greenhouse effect of water vapor over the tropical Pacific during ENSO is a prevalent problem of climate models. The estimates from the coupled simulations of both the cloud albedo feedback and the water vapor feedback differ from the estimates from the corresponding AMIP simulations. The changes in the cloud albedo feedback are particularly significant. The previous analysis of Sun et al. (2006) has suggested that the GFDL CM2 may have a cloud albedo feedback that is as strong as that of the observations. The new estimate with the coupled runs puts this suggestion in doubt because the new estimate is significantly weaker than the previous estimate. All of the models we have examined in this analysis are found to have a weaker negative feedback from the net surface heating than that from observations, indicating that deep convection over the equatorial Pacific in the models has a weaker regulatory effect over the SST in that region. The differences between the values estimated from the coupled runs and the values estimated from the corresponding AMIP runs appear to be linked to the excessive cold tongue in the climatology in the coupled models.

The two common biases, shown in the ENSO cycle, however, do not appear to be carried over the global warming simulations. The variations in the cloud albedo feedback among different models are not correlated with the variations in the same feedback in the global warming simulations among different models. The variations in the water vapor feedback among different models during ENSO over the cold tongue are positively correlated with the variations in the water vapor feedback during global warming, but the correlation is weak. There is no correlation between the feedbacks over the cold tongue region during ENSO and the globally averaged feedbacks during global warming. Therefore, the overestimate of the water vapor feedback and the underestimate of the cloud albedo feedback during the ENSO cycle in the models do not necessarily imply that the sensitivity of the mean tropical climate to anthropogenic forcing is overestimated by the models. As noted by Zhu et al. (2007), in two leading GCMs the changes in the cloud amount in response to ENSO and to global warming may involve different mechanisms. On the other hand, we are not suggesting that the prevalence of these two biases in the models during ENSO should not be of concern for the accuracy of global warming simulated by the models. This is because the lack of correlation in the models between the feedbacks on these two time scales could also be due to error cancellations in the models. Conversely, the lack of the spread among models in the simulated interannual feedbacks does not guarantee that the intermodel differences in the corresponding climate feedbacks are not large (Bony and Dufresne 2005). In any case, the present results highlight the continuing difficulty that models have in simulating the water vapor and cloud feedbacks accurately in the deep tropics on the time scale over which we have observations to compare with model simulations. The results should also be of value to the diagnosis of the causes of the biases in the ENSO amplitude in the models.