1. Introduction

The intertropical convergence zone (ITCZ) is a zonally elongated narrow band of enhanced low-level wind convergence, cloudiness, and rainfall, marking the upward branch of the Hadley circulation cell. A fascinating feature displayed by the ITCZ over the Pacific and Atlantic Oceans is the off-equatorial preference location, in the Northern Hemispheric 4°–12° latitude belt. The existence of one single ITCZ straddling the Northern Hemisphere has been puzzling theoreticians for quite a long time, trying to understand the causes of such an asymmetric response to the essentially symmetric solar radiation forcing (Charney 1971; Holton et al. 1971; Waliser and Somerville 1994; Philander et al. 1996). Idealized experiments performed with aquaplanet model settings forced by highly symmetric SST distributions show no unequivocal responses, with either two off-equatorial ITCZs (Hayashi and Sumi 1986; Swinbank et al. 1988) or one single ITCZ centered on the equator, coincident with the maximum SST location (Lau et al. 1988). Hess et al. (1993), using similar aquaplanet model configurations, identify a strong dependency of ITCZ location on the adopted parameterization for convection and the strength of the SST meridional gradient. The vast majority of coupled general circulation models (CGCMs) show the occurrence of an overly strong ITCZ in the southeastern Pacific region, in a broad region off Peru near 10°S (Mechoso et al. 1995). While the appearance of a Southern Hemispheric ITCZ in March–April is an observed feature of the tropical Pacific climate, its overestimation represents a well-known syndrome affecting state-of-the-art climate models, which is generally referred to as double ITCZ (DI) (Mechoso et al. 1995). This bias affects climate modeling ability to correctly reproduce some of the most prominent climatological features of the tropical Pacific. In particular, the representation of the mean state in the Pacific sector displays an anomalous symmetric structure about the equator, contrasting with the asymmetry characterizing the observed annual mean patterns of rainfall, sea surface temperature, and wind—possibly reflecting the interhemispheric differences for the oceans and continents distribution (Philander et al. 1996; Ma et al. 1996; Yu and Mechoso 1999). DIs in CGCMs are generally associated with an anomalously extended cold tongue on the equator (a quite distinctive feature with respect to aquaplanet configurations), and they typically manifest themselves with a wide spectrum of behaviors (Mechoso et al. 1995; Lin 2007). de Szoeke and Xie (2008) classify the error associated with ITCZ representation in Fourth Assessment Report (AR4) models according to the mean seasonal evolution of precipitation, identifying two distinct error typologies: a persistent double-ITCZ error (rain persisting too long in the Southern Hemisphere) and an alternating ITCZ error (precipitation maxima crossing the equator with the season). Both of them lead to a spuriously high annual mean precipitation in the southeastern tropical Pacific.

Since the early assessment of Mechoso et al. (1995), the overall performance of climate models has been gradually improving through the years (Meehl et al. 2005). However, the erroneous representation of the tropical climate remains a severe limitation for the current generation of CGCMs, recently employed to perform climate projections within the Intergovernmental Panel on Climate Change (IPCC) AR4 (Lin 2007; de Szoeke and Xie 2008), ultimately impacting the predictability and simulation of tropical variability modes (El Niño–Southern Oscillation and Madden–Julian oscillation) on seasonal and interannual time scales. The availability of AR4 experiments archived by the Program for Climate Model Diagnosis and Intercomparison (PCMDI) allows for the cross comparison of an ensemble of CGCMs, combining different parameterizations of unresolved physics as well as spatial resolutions and numerical schemes. Recently, the DI issue (and, more broadly, the models systematic errors in the tropical eastern Pacific) in AR4 simulations of the twentieth-century climate has been examined under different perspectives. Lin (2007) approaches the DI bias in relation to the representation of the main ocean–atmosphere feedbacks, whereas de Szoeke and Xie (2008) focus on the role played by the meridional wind biases in relationship with the interaction with the complex Central American orography.

The purpose of this study is to analyze the relationships between the bias on precipitation in the southeastern tropical Pacific and the systematic errors affecting the underlying large-scale atmospheric vertical circulation regimes in the IPCC AR4 CGCMs. Large-scale vertical motions in the atmosphere are responsible for heat and moisture transport, and thus play a crucial role on determining atmospheric stability, cloudiness, and precipitation.

Specifically, the following issues are addressed. What is the partition of the spurious DI rainfall signal between shallow and deep convective components? Unraveling the deep from the shallow convection precipitation is a fundamental step to further disclosing the nature of DI in climate models. Another crucial question concerning the DI bias is whether the detected rainfall anomaly in the southeastern Pacific is caused by an overly frequent (either deep or shallow) convective activity or by anomalously strong precipitation associated with individual convective events. To this end, a useful approach is provided by a composite methodology first proposed by Bony et al. (2004) for cloud feedback studies, generally referred to as regime sorting, which will be applied here to study the model-dependent relationships between precipitation and vertical circulation regimes in the region affected by the DI systematic error. Exploring a geophysical quantity in the space defined by another variable as an alternative to the standard analysis in the time–space domain allows a better identification of the physical mechanisms relating the two fields under examination. An additional advantage derived from the use of such methodology is the identification of thresholds in the physical space defined by the two selected variables. Specifically, this approach is here extended to the SST–vertical circulation physical space so as to identify critical SSTs setting the transition to deep convection for each member of the AR4 ensemble. The interplay between errors on the SST–deep convection coupling and the biases on SST will also be investigated. The present analysis will particularly focus on the southeastern tropical Pacific region, where the DI systematic error manifests itself.

We address these questions for both coupled and, where available, the corresponding uncoupled [i.e., Atmospheric Model Intercomparison Project (AMIP) type] AR4 simulations of the twentieth-century climate. The cross comparison between coupled and AMIP simulations will provide some insight on the role played by ocean–atmosphere coupling, as compared to atmospheric internal dynamics, in modifying the relationship between the DI spurious precipitation signal and the underlying vertical circulation regimes.

The paper is structured as follows. In section 2, the model and observational data used in this analysis are described. The space–time structure of the systematic errors affecting precipitation and vertical circulation in the tropical eastern Pacific is described in section 3. Results from the regime-sorting analysis applied to the AR4 coupled simulations of the twentieth century are shown in section 4. In section 5 the same analysis is extended to a smaller set of AMIP simulations. The role of biases affecting the representation of SST and the critical SST leading to convection on the amplitude and structure of DI is investigated in section 6. Summary and conclusions are given in section 7.

2. AR4 models and validation data

The analyses shown in the present work are based on monthly outputs from a subset of 20 AR4 CGCMs (except for the mean seasonal cycle of precipitation, where a larger 23-member ensemble is used instead). Also, a smaller set of 13 twin simulations conducted with the atmospheric-only component of the corresponding coupled models, under prescribed SSTs (AMIP type), is analyzed. The models employed in this study are listed in Table 1.

This study focuses on the IPCC Climate of the Twentieth Century (20C3M) simulation, for the 1960–2000 period. Model results are compared with both observational datasets and reanalyses (for simplicity, in the paper we will refer to both types of data as “observations”). In particular, the observed global Climate Prediction Center (CPC) Merged Analysis of Precipitation (CMAP) dataset (Xie and Arkin 1997) is used for precipitation, while for ω fields the 40-yr ECMWF Re-Analysis (ERA-40) (Uppala et al. 2005) has been used. Finally, the global Hadley Centre Global Sea Ice and Sea Surface Temperature (HadISST) analyses (Rayner et al. 2003) were used for sea surface temperatures.

3. Mean seasonal cycle

In Fig. 1 the mean seasonal cycle of precipitation over the eastern Pacific (averaged between 150° and 100°W) is plotted against latitude for observations (Xie and Arkin 1997) and AR4 models. This diagnostic essentially updates the evaluation made by Mechoso et al. (1995; see also de Szoeke and Xie 2008), and portrays the current status of CGCMs, as far as the DI bias is concerned. Assessing the occurrence of DI in models can be quite subjective, in that observations do show a hint of DI manifesting itself as a weakening of the Northern Hemisphere ITCZ and the concomitant appearance of a southern ITCZ, from about February to April (Zhang 2001). Following Mechoso et al. (1995), in Fig. 1 we highlight precipitation in excess of 6 mm day⁻¹, which is set as an arbitrary threshold so that models overcoming this critical value south of the equator are considered to be affected by the DI bias (a more objective metric of precipitation in the southeastern Pacific is defined below). An immediate outcome is that, except for flux-adjusted models and one nonadjusted model, all of the analyzed CGCMs display a spurious precipitation signal south of the equator (around 10°S) mostly affecting the boreal late winter–early spring period. To quantify the DI bias for each coupled model in a more objective way, we introduce a southern ITCZ (SI) index, which is simply defined as the annual mean precipitation over the 20°S–0°, 100°–150°W window. The selected spatial domain is chosen so as to account for the large intramodel spread of the bias structure. In particular, the latitude extent of the box is sufficiently wide to account for differing DI meridional locations in the various models. While this index reflects the integral behavior of a model over a region of the southeastern Pacific, from Fig. 1 it is legitimate to assume that the resulting index values are mainly affected by the presence of the DI. In Fig. 2 the SI index for each member of the extended (23 members) AR4 ensemble and observations is shown. It is evident that among the models displaying a smaller discrepancy with observations are those models that make use of flux adjustments on both heat and water [CCCma Coupled General Circulation Model, version 3.1 T47 resolution (CGCM3.1-T47), CGCM3.1 T63 resolution (CGCM3.1-T63), MRI CGCM2.3.2a, and MIUB ECHO-G Model (MIUBECHOG)]. On the opposite side there are mainly models showing a persistent DI through most of the annual cycle [(CNRM-CM3), NCAR Parallel Climate Model version 1 (PCM1), INM Coupled Model, version 3.0 (INM-CM3.0), GISS Model E-H (GISS-EH)]. This index provides a tool to rank the AR4 CGCMs based on model performance in the DI region, and it will be used in the final discussion. (See Table 1.)

The nature of the spurious precipitation signal in the eastern Pacific is now analyzed by looking at the mean seasonal cycle of the large-scale vertical circulation. We use the Lagrangian pressure tendency ω at 500 hPa (hPa day⁻¹, hereafter ω₅₀₀) as a proxy of the large-scale vertical circulation. Positive values of ω₅₀₀ identify regions of large-scale subsidence, while negative values of ω₅₀₀ indicate regions characterized by convective regimes.

The mean seasonal cycle of ω₅₀₀ in the 20°S–0°, 100°–150°W range is shown in Fig. 3. The persistently positive values shown by observations (from ERA-40) indicate that this region of the eastern Pacific is characterized by a predominant subsidence regime (see comments in the next section). The bulk of AR4 models, on the other hand, reveal a seasonal inversion of the large-scale circulation regime, with rising (ω₅₀₀ < 0) from January to May–June and sinking (ω₅₀₀ > 0) during the rest of the year. Some of the models reveal an almost persistent convective regime (NCAR PCM1 and INM-CM3.0). A few notable exceptions are represented by the third climate configuration of the Met Office Unified Model (HadCM3), Hadley Centre Global Environmental Model version 1 (HadGEM1), and MRI CGCM2.3.2a models, displaying constantly positive ω₅₀₀ values, consistent with the observations (although the latter is a flux-corrected model). The relationship between mean precipitation and ω₅₀₀ in the DI region during January–June (JFMAMJ) and July–December (JASOND) is illustrated in Fig. 4. Models prone to display a pronounced DI bias, associated with anomalously high rainfall, are generally characterized by a consistently large ascent signal (i.e., negative ω₅₀₀ values). The largest model–observation discrepancies as well as intermodel scatter are found in JFMAMJ, while intermodel correlations between precipitation and ω₅₀₀ are 0.79 and 0.85 for JFMAMJ and JASOND, respectively.

4. Regime-sorting analysis

In the previous section it has been shown that the occurrence of a DI in CGCMs is generally associated with midtroposheric large-scale rising motion (ω₅₀₀ < 0). However, spatial averages do not allow a clear inspection of the dynamical causes underlying the process under examination. In particular, it is unclear what is the relative role played by deep versus shallow convection in driving the anomalous precipitation associated with the DI. To better clarify this point we apply to each AR4 model a compositing methodology (illustrated in Bony et al. 2004), where precipitation events are regime sorted based on the large-scale vertical circulation regime (as represented by the ω₅₀₀ dynamical proxy). This procedure is applied to monthly outputs of ω₅₀₀, split into bins of 10 hPa day⁻¹ width [see Hourdin et al. (2006) for further details] in the 20°S–0°, 100°–150°W region. Before applying the regime-sorted analysis, the probability density function (PDF) of the ω₅₀₀ for models and observations is computed (Fig. 5). The PDF provides the normalized frequency of occurrence for a given regime, and it must be considered as a relative weight for the regime-sorted precipitation. Observations show a marked peak around 20 hPa day⁻¹, with a sharp decline for larger ω₅₀₀ values, and a smoother tail of negative values. This distribution essentially reflects the dominance of subsidence regimes in the tropics, which is in turn determined by the clear-sky radiative cooling characterizing this particular region (Chéruy and Chevallier 2000; Bony et al. 2004). All of the AR4 models largely agree with the observed PDF. However, the frequency of occurrence of moderate-to-intense convective events (ω₅₀₀ < −20 hPa day⁻¹) is generally overestimated in the models, while the opposite tendency is exhibited in the −10 < ω₅₀₀ < 10 hPa day⁻¹ range and for subsidence rates larger than 40 hPa day⁻¹. Regimes around the PDF peak (20–30 hPa day⁻¹), on the other hand, occur with a typically higher frequency compared to the observations.

The distribution of precipitation in the 20°S–0°, 100°–150°W region, regime sorted as a function of ω₅₀₀, is shown in Fig. 6. The dynamical link between large-scale circulation and precipitation manifests itself with the largest rainfall events occurring in concomitance with deep convective regimes, contrasted by the relatively weaker precipitation signals associated with moderate and shallow convection. Moreover, under midtropospheric subsidence regimes precipitation appears to be weakly dependent on the strength of sinking motion.

The comparison with observations reveals two important aspects regarding the general behavior of AR4 models. First, none of the model precipitation curves falls below the observed distribution regardless of the specific dynamical regime, indicating a systematic model rainfall overestimate for a given vertical circulation regime. Second, model–observation discrepancies are generally low for shallow convection regimes, but gradually increase with −ω₅₀₀ under deep convection conditions.

An aspect worth examining is the relative contribution to the precipitation bias, derived from the ω₅₀₀ PDF and the regime-sorted precipitation, as measured by Δω/ω and ΔPr/Pr respectively, with Δω(ΔPr) the pointwise difference between modeled and observed ω₅₀₀ PDF (regime-sorted precipitation). In the negative ω₅₀₀ axis, where the largest precipitation events occur, Δω/ω reaches peaks of as much as 6, while ΔPr/Pr remains confined below 1. In other words, it is the spuriously large frequency of occurrence of deep convection regimes that sets the intensity of the bias, rather than the amount of precipitation that falls for a given vertical velocity.

To obtain a more quantitative estimate of precipitation for different vertical circulation regimes, the regime-sorted values of precipitation need to be weighted by the frequency of occurrence of each ω₅₀₀ regime interval. After combining the regime-sorted precipitation with the corresponding ω₅₀₀ PDF, we obtain the distributions shown in Fig. 7. Observations show that the largest contribution to precipitation in the area under examination clearly derives from shallow convective processes, the maximum signal occurring for ω₅₀₀ values around 10–20 hPa day⁻¹, while moderate and intense deep convective events play a relatively minor role. The ensemble of AR4 models, on the other hand, displays a much wider range of behavior. Based on the specific shape of the regime-weighted distributions, models can be gathered into three distinct clusters (shown in Fig. 8). A first group, identified as SUB, collects models displaying the ability of capturing the dominance of precipitation under subsidence regimes with a maximum around 20 hPa day⁻¹, consistent with the observed pattern. A second group, identified as INT, gathers models that exhibit a maximum contribution to precipitation in the deep convection regimes of moderate intensity, with a broad intramodel peak in the (−30 to −10) hPa day⁻¹ range. Finally, a third cluster, labeled as HYBRID, can be identified, which collects models displaying two relative maxima, for both deep and shallow convection regimes, thus mixing together the features of the SUB and INT groups.

We now evaluate the error associated with the regime-sorted precipitation for each single AR4 ensemble member as the model–observation rms error (RMSE) over the (−100 to +80) hPa day⁻¹ ω₅₀₀ range and compare the resulting estimates for each cluster (Fig. 9). From the comparison, it turns out that models referring to the SUB group, except for one single outlier (GISS-ER; see comments below), show a RMSE that is on average lower than the average error as estimated for INT and HYBRID clusters. This indicates that models that qualitatively capture the observed rainfall pattern in the regime-sorted space (SUB cluster) largely minimize the associated error on precipitation. On the other hand, the presence of spurious precipitation under deep convection regimes, particularly for intermediate strengths of convective motions, contributes to a systematically larger model error.

While the adopted approach proves to be generally skillful in segregating models that capture the dominant subsidence regime of the southeastern Pacific (low error on precipitation) from those showing a spuriously high occurrence of deep convection (large error on precipitation), there is one notable exception, represented by the GISS-ER. This model, despite qualitatively capturing the correct regime-sorted rainfall distribution, features an overly strong precipitation signal under subsidence conditions, which leads to a consistently high RMSE (Fig. 9). To further clarify this anomalous behavior the regime-sorted analysis on precipitation was extended by including the lower-tropospheric 700- and 850-hPa compositing levels (not shown). In particular, the weighted Pr(ω₈₅₀) distribution displays a primary maximum for negative ω bins and a secondary maximum around 20 hPa day⁻¹. While the primary maximum confirms that most of the detected precipitation occurs under shallow convection conditions (consistent with observations), a nonnegligible contribution is associated with the secondary maximum, indicating intense convection under dynamical subsidence conditions in the low troposphere, which is clearly a model bias. A detailed explanation of the behavior exhibited by this single model is beyond the scope of the present work. However, it is worth mentioning that the GISS-ER is an outlier within the IPCC AR4 ensemble, as documented elsewhere. In particular, Lin (2007) reveals that this model features a permanent El Niño–like mean SST pattern, with almost no east–west SST gradient in the tropical Pacific, and exceedingly high precipitations over the eastern Pacific.

5. AMIP simulations

The regime analysis performed on the coupled models highlighted the role of convective events of moderate intensity on the setup of the DI. Specifically, the coupled systems appear to reside in a region of the parameter space characterized by convective regimes for a longer time compared to what is known from observations. To single out the effects of the ocean–atmosphere coupling on the DI bias from the contribution deriving from the atmospheric component of the coupled model, we analyze the DI structure in the available AMIP simulations of the twentieth century stored at the PCMDI. Compared to the full set of coupled model experiments, the AMIP experiments constitute a smaller 13-member ensemble. Each AMIP simulation has been performed using observed SSTs as a lower boundary condition for the atmospheric model in a stand-alone configuration. The mean seasonal cycle of precipitation in the eastern tropical Pacific for the AMIP ensemble is shown in Fig. 10. As expected, the SST-forced experiments show a reduced intermodel spread compared to the CGCM ensemble. All of the examined simulations display a reasonable agreement with the observations [except for the Institute of Atmospheric Physics (IAP) model, revealing an overall weak precipitation signal—a feature shared by the corresponding coupled simulation], with no pronounced seasonal excursion of the ITCZ in the Southern Hemisphere. The regime analysis previously applied to the coupled ensemble is now extended to the AMIP simulations. The PDF of the ω₅₀₀ for models and observations is shown in Fig. 11 (left panel). To facilitate a direct comparison with the coupled systems the PDFs of the corresponding CGCM simulations are also shown (Fig. 11, right panel). AMIP runs considerably overestimate the frequency of occurrence of subsidence regimes around the 20 hPa day⁻¹ peak with respect to both observations and coupled models, leading to a generally higher kurtosis of the ω₅₀₀ distributions. The most striking difference between AMIP and coupled simulations lies in the larger frequency of occurrence of deep convection regimes featured by the CGCM experiments, basically reflecting the previously emphasized discrepancies between coupled models and observations. Regime-sorted precipitation in AMIP and in the corresponding coupled runs are overall consistent in both distribution and magnitude (Fig. 12). The reduced occurrence of deep convection events displayed by AMIP experiments when compared to coupled simulations leads to a consistently reduced intermodel spread of ω₅₀₀-weighted regime-sorted precipitation (Fig. 13). The clusters previously identified for the coupled ensemble collapse into one single behavioral group (essentially reproducing the SUB cluster features) when the AMIP set of experiments is considered.

6. SST–large-scale circulation relationship

The intercomparison between coupled and AMIP simulations revealed that the SST constraint plays a crucial role in controlling the frequency of occurrence of convective regimes and, as a consequence, on the strength of the spurious precipitation signal in the eastern Pacific. To establish, in a more quantitative way, the relation between the onset of deep convection and the thermal conditions of the surface ocean, the regime-sorting approach is here extended to the ω₅₀₀–SST physical space so as to obtain ω₅₀₀ distributions sorted by surface thermal regimes. This analysis allows one to clearly identify SST thresholds leading to the onset of deep convection events for each model.

Using the same procedure outlined in section 4, the SST domain is split into bins of 0.5° width. Then, the average ω₅₀₀ is computed for each SST bin over the previously defined longitude–latitude box. The model ω₅₀₀ distributions sorted by thermal regimes display a typical elbowlike pattern that is qualitatively consistent with that observed (Fig. 14; see also Bony et al. 1997; Lau et al. 1997). The midtroposphere vertical velocity is typically positive (indicating subsidence conditions) and approximately constant for surface temperatures lower than a threshold value, beyond which the system enters into a different dynamical regime characterized by deep convection. A (model dependent) critical SST leading to the regime transition is here simply identified as the surface temperature corresponding to the zero crossing of the regime-sorted ω₅₀₀. Alternative options are clearly possible, the most obvious one being the SST corresponding to the elbow in the thermally sorted ω₅₀₀ curve. However, the latter can be potentially affected by large uncertainties, owing to the smooth transition displayed by some of the model realizations. Moreover, the zero-crossing criterion applied to the adopted dynamical proxy objectively separates subsidence from ascending conditions.

The vast majority of AR4 models display a regime transition for temperatures lower than the observed 28°C threshold, with a relatively large spread within the 26°–29.5°C range. The intramodel scatter displayed by the zero-crossing SSTs clearly reflects the differing sensitivities of deep convection on ocean surface thermal conditions displayed by the various coupled models. To establish whether the thermal conditions of the surface ocean are more likely to lie beyond or below the corresponding model SST threshold, we need to associate the corresponding frequency of occurrence to each thermal regime. The PDF for each selected SST bin (shown in Fig. 15) reveals a wide spectrum of model SST distributions, symmetrically spread around the observed distribution. Combining together the SST corresponding to the most likely thermal state (i.e., the SST corresponding to the PDFs peak in Fig. 15, hereafter MLT) with the SST threshold previously identified (hereafter THR), it is possible to estimate the likelihood for a given model to undergo a deep convection event in the examined region. In other words, models whose most likely thermal state is warmer (colder) than THR are more (less) likely to be in the deep convection region of the phase space. In Fig. 16, the difference between THR and MLT is shown for models and observations. Negative (positive) values in Fig. 16 correspond to models whose SST is warmer (colder), most of the time, than the deep convection threshold and are thus expected to feature a more (less) pronounced bias on precipitation. Almost all of the models pertaining to the previously identified INT and HYBRID clusters lie below the zero line. Three of them [IAP, L’Institut Pierre-Simon Laplace Coupled Model, version 4 (IPSL CM4), and NCAR Community Climate System Model, version 3 (CCSM3)] display a positive value for this index, although they are very close to the zero limit. On the other hand, all of the models showing a THR − MLT difference ≥1° fall within the SUB model group.

To further corroborate the hypothesis of a strict relationship between the DI systematic error and the THR–MLT index, models are displayed in the two-dimensional parametric space defined by the SI index, discussed in section 3, and the THR–MLT index (Fig. 17). Here, the split between the SUB (low DI bias, positive THR − MLT) and HYBRID and INT (strong DI bias, negative or marginally positive THR − MLT) model populations appears to be more evident. The grossly linear relationship emerging between these two indices (correlated at the 0.84 level) suggests causality between the amplitude of the systematic error on precipitation and the combined bias on the critical SST leading to deep convection and on the surface thermal state in the southeastern Pacific. Consistent with this interpretation, models displaying a DI persisting through most of the year (Fig. 1) are those whose surface temperatures are prone to be systematically warmer than the SST threshold leading to the onset of deep convection. On the other hand, a large THR − MLT difference acts as a deterrent for the start of deep convection. The Met Office’s HadCM3 provides a particularly insightful example as it features an overly large THR (29.5°C) combined with a MLT consistent with the observed one, leading to a fairly reduced DI error. Different reasons (i.e., a THR close to observations and a cold-biased MLT) induce a similar performance in the Met Office HadGEM1.

7. Summary and conclusions

In this study the double-ITCZ systematic error affecting the climate of the tropical eastern Pacific in the current generation of coupled models is examined in relation with the representation of atmospheric vertical circulation regimes, using a regime-sorting approach. The analysis, applied to both coupled and uncoupled (atmosphere only) IPCC AR4 simulations of the twentieth century, reveals that the excess of precipitation detected in the southeastern Pacific (the DI region) in CGCMs is mainly due to the overly frequent onset of deep convection, whereas the error on rainfall magnitude associated with individual convective events is overall consistent with errors already present in the corresponding AMIP-type simulations. Through the present analysis we also identified three distinct model behavioral groups within the AR4 ensemble, thus associating to each model a DI rainfall fingerprint: two unimodal distributions, SUB and INT featuring maximum precipitation under subsidence and deep convection regimes respectively, and one bimodal distribution, HYBRID, displaying both components. A simple metric for precipitation, based on the model–observation rms error, but defined in the vertical circulation regime space, reveals that models that correctly capture (at least, qualitatively) the observed regime-sorted rainfall pattern in the eastern Pacific (SUB cluster) do also minimize the rms error. Thus, the most intense DI occurrences are associated with the spurious deep convective precipitation bulge in the ω₅₀₀ space displayed by models in the INT and HYBRID clusters. The relative homogeneity displayed by CGCMs in the DI precipitation signature sharply contrasts with the richness of deep convection schemes (and corresponding closure/triggers) adopted by individual climate models in the AR4 ensemble (see Table 1 in Lin 2007 for a synoptic view). Each single identified cluster displays a wide variety of deep convection parameterizations (not shown); thus, there is no obvious relationship between a given model group and a specific deep convection scheme. Clearly, the AR4 experimental set is not optimal in that different realizations of the twentieth-century climate are produced with model configurations differing by several aspects (parameterizations of unresolved processes, spatial resolution, dynamical cores, etc.) so that intramodel differences can be hardly ascribed to one single element, but may rather result from the additional effect of changing many different model features (Schneider 2002). However, the apparent lack of sensitivity to convective schemes shown by models within each single cluster seems to indicate that there is a more fundamental factor overcoming the differences between the adopted deep convection schemes, thus forcing very different systems to behave in a similar way. As pointed out in previous studies, SST is a primary candidate to explain the DI location and strength in both observations (Zhang 2001) and models (Mechoso et al. 1995; Yu and Mechoso 1999). Relative maxima in SST control the regions where the largest upward vertical motions occur (Schneider 1977) that are, in turn, responsible for the vertical advection of moisture—an important prerequisite for the onset of deep convection. The comparison between coupled and AMIP experiments further confirms that it is the drift of surface thermal conditions from a realistic pattern that determines the intramodel spread in the manifestation of DIs, as GCMs behavioral clusters collapse into one single group under prescribed SSTs. This suggests that the existence of homogeneous CGCM classes can be traced back to the different ways that the coupled models represent the SST–deep convection relationship.

The composite analysis of ω₅₀₀ in the space defined by surface thermal states shows that the critical SST, setting the transition to deep convective unstable conditions in the AR4 models population, can vary within a wide range of values, likely reflecting the aforementioned variety of deep convection schemes featured by climate models. Similarly, the SST biases in the tropical Pacific do also exhibit a large intramodel scatter. However, a model displaying anomalously warm surface temperatures over the eastern Pacific does not necessarily favor the onset of deep convection (thus producing overly strong precipitations) unless the corresponding convective SST threshold lies, on average, below the surface temperature in that region. Thus, the distance between the critical convective SST and the average SSTs over the DI region largely control the model clustering process. This result is consistent with the finding that model SSTs in the tropical eastern Pacific are symmetrically distributed around the observations (within a belt about 2°C wide), while precipitation is systematically higher than the observational estimates (Figs. 6 and 15, see also Fig. 2 in de Szoeke and Xie 2008). The symmetric model SST distribution indicates that cold-biased models can, in principle, result in anomalously large rainfall and spurious southern ITCZ if the convective SST threshold is consistently lower than the average SST. Specularly, a warm-biased model may display unfavorable conditions to the setup of anomalous deep convection if the SST threshold for deep convection is sufficiently larger than the mean SSTs in the eastern Pacific (see the HadCM3 case).

Summarizing, an important outcome of this study is that, by splitting the total DI bias on precipitation into 1) an error in the frequency of occurrence of deep convection and 2) an error in the magnitude of precipitation for an individual convective event, it is possible to state that the first is caused by ocean–atmosphere interactions, whereas the second can be attributed to the atmospheric GCM component only, with the former playing a major role on the total amplitude of the DI bias.

A dominant paradigm among the theories trying to explain the DI in climate models invokes the well-known deficit of low-level stratocumulus clouds in the southeastern Pacific (Philander et al. 1996; Yu and Mechoso 1999). The lack of stratocumulus clouds and the implied enhancement of solar radiation income, in turn, determine a warm bias at the ocean surface, which ultimately triggers deep convection and precipitation in a region where the observed dominant regime is subsidence with consistently low rainfall. However, a closer look at the zonally averaged mean meridional SST profile in the region under examination reveals that not all of the models display a warm bias therein, but there is a rather symmetric scatter of warm and cold SST biases around the observations (see Fig. 2 in de Szoeke and Xie 2008). Precipitation, on the other hand, is mostly skewed toward positive anomalies with almost all of the models overestimating precipitation in the tropical southeastern Pacific. Assuming a thermally driven nature for the DI bias, with the present analysis we suggest a possible explanation for the aborementioned apparent inconsistency between SST and precipitation biases, with the missing link identified in the model-dependent critical SST setting the transition to a deep convective regime. This parameter partly decouples the SST bias from the bias on precipitation as models with a cold bias may still be featuring overly strong rainfall, provided that their convective SST threshold is consistently low.

While the local impact of SSTs is, in our view, a major driver of the DI systematic error, the analysis of the AMIP simulations (section 5) suggests that the constraint of SSTs over regions that are far from the eastern Pacific may exert a similarly important control on the rainfall bias over the examined region. The nonlocal factor, which most likely influences the tropical rainfall pattern in an atmospheric stand-alone simulation with prescribed observed SSTs, is related to the presence of a correct SST gradient across the equator. Numerical experiments performed using AGCMs with simplified water-covered earth configurations forced by idealized zonally symmetric SST profiles (Aqua-Planet Experiment Project; information online at http://www-pcmdi.llnl.gov/projects/amip/ape) suggest that transitions from a single- to a double-ITCZ equilibrium may arise (at least in some models) after gradually reducing the SST meridional gradient (from peaked to flat) around the equator. However, the large intermodel spread in AGCM response to idealized SST gradients cast large uncertainties as to the precise mechanisms governing the relation between the nonlocal SST forcing and the latitudinal ITCZ location.

Double-ITCZs have been found to be largely controlled by the combined effect of SST–deep convection coupling and the amplitude of the SST bias in the eastern Pacific. The SST threshold setting the transition to deep convection, even if not directly disposable, may be partly controlled through an appropriate modeling of the triggers characterizing a given deep convection scheme. The criteria used to determine the initiation of convection considerably vary from one scheme to another, including cloud-base buoyancy (Gregory and Rowntree 1990), moisture convergence (Tiedtke 1989), and convective available potential energy (Zhang and McFarlane 1995), to mention a few. Each of these quantities is ultimately controlled by the SST via processes occurring in the boundary layer. Revisiting the deep convection schemes used by AR4 models in view of the above considerations may represent a possible pathway to alleviate the DI syndrome in CGCMs. This will require additional efforts, including numerical experiments to be performed by individual modeling groups, specifically designed to address errors in the representation of SST–deep convection coupling and their impact on the double-ITCZ bias.

The construction of a set of standard metrics aimed to evaluate climate model performance is a most urgent need since the multimodel intercomparison framework is becoming a standard procedure in climate science. A wide consensus is emerging that application-dependent metrics are more valuable and physically justifiable compared to single indices of overall model performance, as the latter are typically based on a somewhat arbitrary set of nonhomogeneous metrics (Gleckler et al. 2007). The onset of a split ITCZ is a bias affecting the vast majority of state-of-the-art CGCMs in a region crucial to the development of El Niño. Thus, building a specific metric to rank models with respect to this particular systematic error is a relevant step toward the definition of a set of process-oriented metrics. The regime-sorting methodology has been found to be a particularly insightful instrument in the diagnosis of double ITCZs in CGCMs. Metrics based on this approach may represent a useful complement to existing diagnostics in the assessment of a model’s ability to reproduce the climate of the tropics, within the framework of upcoming multimodel intercomparison efforts.