An excessive cold tongue error in the equatorial Pacific has prevailed in several generations of climate models. However, the causes of this problem remain a mystery, partly owing to uncertainty and/or a lack of observational datasets. Based on the multimodel ensemble from phase 5 of the Coupled Model Intercomparison Project (CMIP5), this study introduces a novel intermodel approach to identify the bias source by going beyond comparison with observational datasets. Intermodel statistics show that the excessive cold tongue bias could be traced back to a too strong oceanic dynamic cooling linked to a too shallow thermocline along the equatorial Pacific. A heat budget analysis suggests that the excessive oceanic dynamic cooling is balanced by the surface latent heat flux (LHF) adjustment. This is consistent with a variety of oceanic and atmospheric observations but at odds with the popular objectively analyzed air–sea heat fluxes (OAFlux) products. Further analyses suggest an alarming overestimation of OAFlux net surface heat flux (Qnet) into the tropical Pacific, mainly ascribed to observational uncertainly in air specific humidity. Implications for intermodel statistics in assessing model processes, validating observational data, and regulating future climate projections are discussed.
Along the equatorial Pacific, easterly winds force oceanic upwelling and an eastward shoaling of the thermocline, which help bring the subsurface cold water close to the surface. The strong oceanic dynamic cooling could be balanced by large net surface heat flux (Qnet). As a result, the equatorial Pacific is observed to display a minimum sea surface temperature (SST) extending from the coast of South America to the central Pacific, commonly referred to as the cold tongue (Wyrtki 1981; Xie 2013). It is well known that large interannual SST anomalies of the equatorial Pacific cold tongue associated with El Niño–Southern Oscillation (ENSO) have considerable effects on the regional and global climate, such as the Indian summer monsoon (ISM; Webster and Yang 1992; Kumar et al. 1999; Wu et al. 2012), East Asian climate (e.g., Wang et al. 2000; Zhou and Wu 2010; Yang et al. 2014), north Asian climate (e.g., Ropelewski and Halpert 1986; Larkin and Harrison 2005a; Zou et al. 2014), and global sea level pressure (SLP), surface temperature, and precipitation (e.g., Ropelewski and Halpert 1987; Kiladis and Diaz 1989; Klein et al. 1999; Trenberth and Caron 2000; Alexander et al. 2002; Larkin and Harrison 2005b; Banholzer and Donner 2014).
Coupled general circulation models (CGCMs) do not perform very well in the tropical Pacific, producing an excessive and overly narrow equatorial Pacific cold tongue that extends too far westward (e.g., Mechoso et al. 1995; Yu and Mechoso 1999; Latif et al. 2001; Lin 2007; de Szoeke and Xie 2008; Zheng et al. 2012; Li and Xie 2014; see also Fig. 1). This excessive cold tongue error has prevailed in several generations of CGCMs, limiting their skill in simulating and predicting ENSO as well as its global teleconnections (Latif et al. 2001; Guilyardi 2006; AchutaRao and Sperber 2006; Wittenberg et al. 2006; Ham and Kug 2012; Kim et al. 2014).
The causes of the cold tongue bias in CGCMs are complex. There have been several possible hypotheses to address this problem. For example, this cold tongue bias could be reduced by requiring a more realistic simulation of stratus cloud amount off the Peruvian coast (Yu and Mechoso 1999). Improving the convection schemes (e.g., Zhang and Song 2010) and/or changing wind stress formulations (Luo et al. 2005) in atmospheric models can also help reduce the cold tongue bias. In addition, the excessive equatorial Pacific cold tongue in CGCMs may be also partly due to a poor representation of ocean–atmosphere feedbacks (Lin 2007) and large coupling intervals between atmospheric and oceanic models (Misra et al. 2008).
In particular, Zheng et al. (2012) carried out a recent heat budget analysis for 15 coupled simulations from phase 3 of the Coupled Model Intercomparison Project (CMIP3) by comparisons of surface heat fluxes from the objectively analyzed air–sea heat fluxes (OAFlux) project (Yu and Weller 2007) and ocean temperatures and currents from the Simple Ocean Data Assimilation (SODA) reanalysis (Carton and Giese 2008). Surprisingly, while most models produce weaker Qnet into the equatorial Pacific compared to OAFlux, the ocean heat advection in models also contributes more cooling to the upper ocean than SODA does. These two findings physically contradict each other because excessive (insufficient) warming from Qnet in CGCMs would be balanced by more (less) cooling from upper-ocean heat transport for long-term means and vice versa. In other words, large uncertainties in OAFlux and/or SODA datasets make it difficult to determine what physical processes result in an excessive equatorial Pacific cold tongue in CGCMs.
This study introduces a novel intermodel approach to revisit the equatorial Pacific cold tongue bias. We show that the cold tongue bias in 20 CGCMs from phase 5 of CMIP (CMIP5) could be traced back to errors in upper-ocean heat transport; models with a stronger oceanic dynamic cooling tend to feature cooler SST and less precipitation along the equatorial Pacific. Our analysis also reveals an alarming uncertainty in the widely used OAFlux dataset, with an overestimation of Qnet being up to about 30 W m−2 in magnitude over the tropical Pacific.
The rest of the paper is organized as follows. Section 2 describes models and datasets used in this study. Section 3 investigates the source of excessive equatorial Pacific cold tongue in CMIP5 models as well as uncertainty in observational datasets by using an intermodel analysis. Section 4 is a summary with discussion.
2. Models and datasets
We examine the 50-yr (1950–99) climate of the historical simulations from 20 CMIP5 CGCMs (Taylor et al. 2012). Though more than 30 CMIP5 modeling groups (centers) had submitted their simulations at the time of this analysis, for simplicity we adopt data from only 20 models, all developed at major modeling groups (centers). Table 1 shows the model names, modeling groups (centers), and labels. The description of individual models can be obtained online (at http://www-pcmdi.llnl.gov; Taylor et al. 2012). Here, we use only one ensemble member (r1i1p1) run for each model. Monthly mean model outputs are used, including SST, downward and upward shortwave (SW) and longwave (LW) radiation, latent heat flux (LHF), and sensible heat flux (SHF) at the sea surface; 2-m specific humidity (qa); 10-m, 925-hPa winds; precipitation; and ocean temperature. All model outputs are interpolated to a common 2° × 2° grid unless otherwise specified.
For comparisons, we examine both the observed and reanalyzed (assimilated) datasets (for simplicity referred to as observations), including the SODA ocean temperature (Carton and Giese 2008) for 1950–99, Hadley Centre Sea Ice and SST dataset (HadISST; Rayner et al. 2003) for 1950–99, precipitation from the Global Precipitation Climatology Project (GPCP; Adler et al. 2003) for 1979–2008, and 925-hPa winds from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis (Kalnay et al. 1996) for 1950–99 and the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40; Uppala et al. 2005) for 1958–2001. In addition, surface heat fluxes and related variables such as 10-m wind speed (WS), qa, SST, and sea surface specific humidity (qs) from the OAFlux (Yu and Weller 2007), CORE version 2.0 (v2.0) (Large and Yeager 2009), ERA-40 (Uppala et al. 2005), and National Oceanography Centre Southampton Flux Dataset version 2.0 (NOCS v2.0; Berry and Kent 2009) datasets are also used.
a. Intermodel diversity
Figure 2 compares the annually averaged SST along the equatorial Pacific between HadISST observations (black line) and 20 CMIP5 CGCMs (colored lines). Almost all CMIP5 CGCMs still suffer from a cool SST error and excessive westward extension of equatorial Pacific cold tongue relative to observations. In particular, the intermodel diversity is considerable with a zonal mean SST difference of up to 3°C, being comparable in magnitude to SST peak anomalies for the strongest ENSO event in nature.
We examine the intermodel variability of annual mean SST climatology in the equatorial Pacific (2°S–2°N, 120°E–80°W) by performing an intermodel empirical orthogonal function (EOF) analysis for observations and 20 CMIP5 CGCMs. Figure 3a shows the regressed patterns of SST, precipitation, and 925-hPa winds onto the first intermodel principal component (PC1). Here, the SST and precipitation for observations and each CGCM are normalized by their tropical Pacific (20°S–20°N, 120°E–80°W) means. The first intermodel EOF mode (EOF1), explaining 77.6% of total intermodel SST variability, captures the cool SST and insufficient precipitation in the Pacific cold tongue, flanked by the excessive precipitation and warm SST on both sides. On the equator, easterly wind anomalies in the western basin help develop the excessive equatorial cold tongue by enhancing (shoaling) the ocean upwelling (thermocline) in the east. This relationship among wind, precipitation, and SST/thermocline biases in the equatorial Pacific is suggestive of Bjerknes feedback.
The PC1 is highly correlated with the SST (Fig. 3b) and precipitation (Fig. 3c) in the equatorial Pacific cold tongue (2°S–2°N, 150°E–110°W), with the strong intermodel correlations of −0.86 and −0.92, respectively. Namely, models with higher (lower) PC1 tend to have a stronger (weaker) equatorial Pacific cold tongue. Here, we choose four models of the highest PC1 values as the cool cold tongue (cCT) models (M6, M8, M17, and M18), and four models of the lowest PC1 values as the warm cold tongue (wCT) models (M1, M5, M7, and M10).
b. Cause of the cold tongue bias
For long-term means, the upper-ocean heat budget can be written as (Li and Xie 2012)
where QSW, QLW, QLHF, and QSHF denote net SW, net LW, LHF, and SHF at the sea surface (positive downward), respectively; Do is the ocean heat transport effect due to three-dimensional advection and mixing (including entrainment at the base of the mixed layer); and QSW + QLW + QLHF + QSHF = Qnet. As some oceanic variables (such as the vertical ocean current velocity) are unavailable from CMIP5 model outputs, it is hard to directly estimate the upper-ocean heat processes. Here we use Qnet to diagnose the Do term in Eq. (1).
The intermodel diversity from surface heat fluxes and ocean heat transport could be responsible for that of SST. For example, too strong an oceanic upwelling could bring the cold water from the thermocline close to the surface and thus cool SST. On the other hand, the enhanced trade winds might also tend to generate more LHF and help cool SST, indicative of wind–evaporation–SST (WES) feedback (Xie and Philander 1994). Figure 4a shows the zonal mean differences in the tropical Pacific of various terms in Eq. (1) between cCT and wCT models. cCT models have cooler SST on the equator than wCT models, and the main contributor to cool SST at the equator is the term Do, which is balanced by LHF. Intermodel correlations support this hypothesis. Indeed, models with a stronger Do cooling tend to produce cooler SST and less precipitation in the equatorial Pacific cold tongue (Fig. 5). The differences of SW, LW, and SHF between cCT and wCT models in the equatorial Pacific cold tongue are small (Fig. 4a). This suggests that the increased Do cooling generates an excessive equatorial Pacific cold tongue in cCT models, and cooler SST induces less LHF away from the ocean to balance Do anomalies.
Comparisons with the SODA ocean temperature and NCEP–NCAR wind at 925 hPa also support the hypothesis that an excessive equatorial Pacific cold tongue in CGCMs is due to a too strong Do cooling. On the equator, models have too shallow a thermocline (represented by the 20°C isotherm) in the eastern basin relative to SODA observations, which helps bring cold subsurface water to cool SST, consistent with their equatorial easterly biases in the western half of the basin relative to NCEP–NCAR and ERA-40 wind observations (Fig. 6; Bellenger et al. 2014). Indeed, models with a stronger eastward shoaling of thermocline tend to have cooler SST and less precipitation in the equatorial Pacific cold tongue (Figs. 7a,b); in addition, models with a stronger easterly wind bias in the western half of the basin would also have a stronger equatorial Pacific cold tongue (Figs. 7c,d). Similar intermodel relationships can be found between the biases of the SST/precipitation in the cold tongue and zonally averaged zonal wind along the equator but with a slightly weaker correlation (not shown). Clearly, the SODA ocean temperature and NCEP–NCAR wind observations are in good agreement with the above intermodel statistics (see black stars in Fig. 7).
c. Uncertainty in observational datasets
Figure 4b compares the meridional profiles of Qnet in the tropical Pacific between the multimodel ensemble mean (MME) simulation and various surface heat flux datasets. Compared to three surface heat flux datasets (CORE v2.0, ERA-40, and NOCS v2.0), CGCMs have an excessive warming from Qnet to balance a too-strong Do cooling on the equator and thus maintain an overly cool SST in the equatorial cold tongue. This is consistent with the intermodel statistics as explained above, as well as the SODA ocean temperature and NCEP–NCAR wind observations. But these also imply a large overestimation of the OAFlux Qnet in the equatorial Pacific (Fig. 4b), indicating that Zheng et al.’s (2012) contradictory findings are likely due to uncertainty in OAFlux Qnet (see the introduction).
Figure 8a compares the Qnet, net SW, and LHF in the equatorial Pacific cold tongue between the MME simulation and various surface heat flux datasets. Models have more Qnet into the ocean than CORE v2.0, ERA-40, and NOCS v2.0 observations, which is mainly due to less LHF away from the ocean, consistent with the intermodel statistics in Fig. 4a. Surface LHF is calculated based on a bulk formula (Yu and Weller 2007; Li et al. 2011a,b):
where L is the latent heat of evaporation, is the surface air density, is the transfer coefficient, WS is the surface wind speed, RH is the surface relative humidity, T and are the SST and the difference from surface air temperature, is the saturation humidity following the Clausius–Clapeyron (CC) equation, , , and is the gas constant for water vapor. Surface LHF variation can be decomposed into an SST effect term , and the atmospheric term given by that represents atmospheric forcing due mostly to changes in WS, RH, and . Here the overbar and prime denote the mean and perturbation, respectively. In particular, the atmospheric forcing due to changes in WS is obtained by linearizing Eq. (2) into (Xie et al. 2010). In general, surface LHF variation can be represented as
where Res is a residual. Figure 8b shows the MME biases in LHF and the corresponding SST and WS effect terms in the equatorial Pacific cold tongue compared to the CORE v2.0, ERA-40, and NOCS v2.0 observations. The insufficient surface LHF release in CGCMs is mainly due to a too cool SST effect in the equatorial Pacific cold tongue.
However, OAFlux observations have a significant overestimation of Qnet relative to CMIP5 CGCMs and other observations, due to both a large overestimation of net SW into the ocean and an underestimation of LHF away from the ocean (Fig. 8a). Both the OAFlux and CORE v2.0 use the same International Satellite Cloud Climatology Project (ISCCP) surface radiation source (Zhang et al. 2004; Yu and Weller 2007; Large and Yeager 2009). While CORE v2.0 adopts a large tropical reduction for ISCCP solar radiation to agree with in situ measurements and other products (Large and Yeager 2009), the OAFlux utilizes only the raw radiation, resulting in a large overestimation of net SW into the ocean.
To further explore the source of uncertainty in observational LHF, Fig. 9 shows the SST, difference between qs and qa (i.e., Dq), and WS in the equatorial Pacific cold tongue for various observations. While OAFlux observations have similar magnitudes in both SST and WS compared to other observations, they have too low Dq, leading to a large underestimation of LHF. Since qs is only a function of SST following the CC equation, such a low Dq in OAFlux is caused by an overestimation of qa.
4. Summary and discussion
An excessive equatorial cold tongue remains the most prominent error for the tropical Pacific climate simulation in CMIP5 CGCMs. The causes of this bias have always been a mystery, partly due to large uncertainties in observational datasets. The use of intermodel statistics is a novel aspect of this study, enabling us to identify the source of model biases and uncertainty in observational datasets. The present analysis suggests that the equatorial cold tongue bias with an intermodel SST spread of up to 3°C in CMIP5 CGCMs could be traced back to errors in upper-ocean heat transport, with stronger Do cooling models biasing cooler SST and less precipitation in the cold tongue. Too shallow a thermocline in CGCMs helps induce a too strong Do cooling in the equatorial Pacific cold tongue, leading to cooler SST and less precipitation via Bjerknes feedback. The cool SST could result in a too weak surface LHF damping and then allow more Qnet into the ocean to balance the Do anomalies. Such intermodel results are consistent with a variety of observational products (such as the SODA ocean temperature, NCEP–NCAR winds, and CORE v2.0, ERA-40, and NOCS v2.0 surface heat fluxes) that have independent data sources, and dramatically identify a large overestimation of Qnet in OAFlux.
Such an alarming uncertainty in OAFlux Qnet is found not only on the equator but also in the whole tropical Pacific with about 30 W m−2 in magnitude (Fig. 2b). This overestimation in OAFlux Qnet cannot be tolerated in many flux applications. For example, while observed ocean temperature changes suggest only a small positive imbalance in the global mean Qnet with about 0.2 W m−2 in magnitude (Levitus et al. 2005; Large and Yeager 2009), the climatological global mean Qnet in OAFlux is up to about 30 W m−2 (not shown). This study finds that the overestimated OAFlux Qnet largely originates in an underestimation of LHF release, mainly due to observational uncertainty in qa. OAFlux qa is synthesized from the satellite-based Special Sensor Microwave Imager (SSM/I) column water vapor retrievals (Chou et al. 2003) and NCEP-1, NCEP-2, and ERA-40 outputs. Although satellite remote sensing produces good WS and SST observations, it is still difficult to retrieve air humidity at a few meters above the sea surface by using the simple precipitable water information. On the other hand, NCEP analyses are indeed found to be too moist above the tropical Pacific based on the comparison with in situ measurements (Large and Yeager 2009).
In recent years, the OAFlux project is devoting great effort to constructing a higher-resolution product of qa by objectively merging two recent satellite-derived analyses [i.e., the multi-instrument microwave regression (MIMR; Jackson et al. 2009) and the Goddard Satellite-based Surface Turbulent Fluxes, version 3 (GSSTF3; Shie et al. 2012)] with the existing OAFlux analysis and atmospheric reanalyses (Jin et al. 2015). However, we notice that the qa values of both MIMR and GSSTF3 are even higher averaged over the global ice-free ocean than those of the OAFlux analysis (see Fig. 4a of Jin et al. 2015), particularly in the tropical oceans (see Figs. 5a and 7a of Jin et al. 2015). This implies that the error of an LHF underestimation due to the overestimated qa would not be reduced in the ongoing OAFlux product (so far the new OAFlux data are not obtained online as the work is still going on). Thus, further research into improving and validating qa observations will be essential to improving surface heat flux products. The benefits of improved surface heat flux products are enormous for numerous oceanic and atmospheric studies, including a more reasonable evaluation of physical processes in CGCMs.
Uncertainty and/or a lack of observational datasets usually make it very difficult to assess model physical processes. Considering the large intermodel diversity in the CMIP5 multimodel ensemble, our intermodel statistics provide new insight for model evaluation to go beyond comparison with observations and assess physical processes via intermodel differences/correlations. In addition, the example for identifying uncertainty in OAFlux Qnet provides us some confidence that this new intermodel approach could be also used for validating observational data and even calibrating future climate projection in the multimodel ensemble. For example, CGCMs tend to produce an amplitude of Indian Ocean dipole (IOD; Saji et al. 1999; Webster et al. 1999) mode greater than observations (Cai and Cowan 2013; Liu et al. 2014; Li et al. 2015a) and a slightly weaker amplitude of Indian Ocean basin (IOB; Klein et al. 1999; Lau and Nath 2000) mode relative to observations (Li et al. 2015b). The recent studies (Cai et al. 2011; Li et al. 2015a) show that models with a larger IOD amplitude bias would have a stronger IOD-like climate projection under global warming in the CMIP3 or CMIP5 multimodel ensemble. This strong intermodel statistical correlation between the present IOD simulation bias and future IOD-like climate projection might offer a good chance to calibrate regional climate projection in the future by correcting the present climate simulation to observations.
This work was supported by the National Basic Research Program of China (Grant 2012CB955603), the Natural Science Foundation of China (Grant 41406026), the Guangdong Natural Science Foundation for Distinguished Young Scientists, the Pearl River S&T Nova Program of Guangzhou (201506010094), the Strategic Priority Research Program of the CAS (XDA11010103 and XDA11010203), the CAS/SAFEA International Partnership Program for Creative Research Teams, the Technology Foundation for Selected Overseas Chinese Scholars (Ministry of Human Resources and Social Security of the People’s Republic of China), the Open Project Program of the Key Laboratory of Meteorological Disaster of the Ministry of Education (Nanjing University of Information Science and Technology; Grant KLME1402), and the Open Project Program of the State Key Laboratory of Tropical Oceanography (Grant LTOZZ1202). We also wish to thank the climate modeling groups (Table 1) for producing and making available their model outputs, the WCRP’s Working Group on Coupled Modelling (WGCM) for organizing the CMIP5 analysis activity, the Program for Climate Model Diagnosis and Intercomparison (PCMDI) for collecting and archiving the CMIP5 multimodel data, and the U.S. Department of Energy Office of Science for supporting these datasets in partnership with the Global Organization for Earth System Science Portals.