Wind stress measurements from the Quick Scatterometer (QuikSCAT) satellite and two atmospheric reanalysis products are used to evaluate the annual mean and seasonal cycle of wind stress simulated by phases 3 and 5 of the Coupled Model Intercomparison Project (CMIP3 and CMIP5). The ensemble CMIP3 and CMIP5 wind stresses are very similar to each other. Generally speaking, there is no significant improvement of CMIP5 over CMIP3. The CMIP ensemble–average zonal wind stress has eastward biases at midlatitude westerly wind regions (30°–50°N and 30°–50°S, with CMIP being too strong by as much as 55%), westward biases in subtropical–tropical easterly wind regions (15°–25°N and 15°–25°S), and westward biases at high-latitude regions (poleward of 55°S and 55°N). These biases correspond to too strong anticyclonic (cyclonic) wind stress curl over the subtropical (subpolar) ocean gyres, which would strengthen these gyres and influence oceanic meridional heat transport. In the equatorial zone, significant biases of CMIP wind exist in individual basins. In the equatorial Atlantic and Indian Oceans, CMIP ensemble zonal wind stresses are too weak and result in too small of an east–west gradient of sea level. In the equatorial Pacific Ocean, CMIP zonal wind stresses are too weak in the central and too strong in the western Pacific. These biases have important implications for the simulation of various modes of climate variability originating in the tropics. The CMIP as a whole overestimate the magnitude of seasonal variability by almost 50% when averaged over the entire global ocean. The biased wind stress climatologies in CMIP not only have implications for the simulated ocean circulation and climate variability but other air–sea fluxes as well.
The reliability of future climate projections using climate models depends heavily on the fidelity of the climate models. The latter can be assessed by evaluating the ability of the climate models to simulate the present climate using available observations (e.g., Pierce et al. 2006; Gleckler et al. 2008; Waliser et al. 2009; Su et al. 2013). The World Climate Research Programme's (WCRP) phase 3 of the Coupled Model Intercomparison Project (CMIP3) (Meehl et al. 2007) has made available a suite of twentieth-century simulations by various coupled general circulation models (CGCMs) that have facilitated many evaluation efforts (e.g., Guilyardi 2006; Capotondi et al. 2006, 2012; Su et al. 2006; Yu and Kim 2010a; Jamison and Kravtsov 2010; Kwok 2011; Li et al. 2012a). Recently, phase 5 of the Coupled Model Intercomparison Project (CMIP5) has released many “historical” simulations that encompass the twentieth century. The comparison of CMIP3 and CMIP5 in the context of observations is an important initiative of the WCRP that helps identify potential improvements and remaining issues in climate models that contribute to CMIP and the Intergovernmental Panel on Climate Change (IPCC) assessments (e.g., Gleckler et al. 2011). Examples of such studies include evaluating the fidelity of CMIP GCMs to represent water vapor and clouds (Jiang et al. 2012), clouds and radiation (Li et al. 2012a,b; Li et al. 2013), and observed phenomena such as the structure of El Niño–Southern Oscillation (ENSO) (e.g., Kim and Yu 2012).
A key parameter of climate models that has not been evaluated systematically across a suite of CMIP using global observations is ocean surface wind stress. Hereafter, we simply refer to this parameter as wind stress for the sake of convenience. Wind stress is an important variable in the coupled climate system as it reflects the momentum flux between the ocean and atmosphere, which is a major forcing of ocean circulation. Direct observations of wind stress are extremely sparse. Satellite scatterometers such as the Quick Scatterometer (QuikSCAT) provided a decade (from mid-1999 to late 2009) of global measurements of wind stress with unprecedented spatial and temporal sampling. In the present study, we use wind stress measurements from QuikSCAT and auxiliary atmospheric reanalysis products to evaluate CMIP3 and CMIP5 and discuss the implications of the biases in CMIP wind stress to the simulation of ocean circulation and natural climate variability.
The evaluation of the CMIP wind stress can also help understand the positive aspects and limitations of other wind-dependent ocean–atmosphere fluxes (e.g., latent heat flux) simulated by CMIP. It is also relevant to the assessment of the simulated state of the ocean by CMIP, such as the structure of sea surface temperatures (SST), pycnocline and sea level, horizontal and meridional circulation, and meridional transport of heat and freshwater, which all have significant dependence on wind stress. This study focuses on the evaluation of the wind stress climatology from CMIP, including the annual mean and seasonal cycle. The climatological state of CMIP has strong implications to the simulated natural variability and probably to climate change projections based on these models. Therefore, the present study is a necessary first step toward further evaluation for the variability on other time scales (e.g., intraseasonal, interannual, and decadal variability) and for climate change projection.
2. CMIP and reference datasets
We analyzed the wind stress from the twentieth-century simulations of 18 models in CMIP3 and from the historical simulations of 11 models in CMIP5 that are available at the beginning of our analysis (the model acronyms and expansions used in CMIP3 and CMIP5 are provided in Table 1). The CMIP3 twentieth-century simulations end in 2000. Those for the CMIP5 historical run extend beyond 2000. We choose to analyze a common 3-decade period for CMIP3 and CMIP5 from 1970 to 1999. This period corresponds to the era with more robust observation methods and sampling of ocean surface wind. Consequently, atmospheric reanalysis products that use these modern-era observations are more reliable. CMIP3 and CMIP5 simulations for the period of 1970–99 are used to produce the respective monthly climatology of wind stress and mapped on a common 2° × 2° grid.
The observational reference dataset used in this study is based on satellite scatterometer measurements derived from the QuikSCAT mission of the National Aeronautics and Space Administration (NASA). Launched in June 1999, QuikSCAT provided over a decade of wind stress measurements (until November 2009). QuikSCAT observations have revolutionized our capability to estimate the dynamical forcing of the ocean from basin to mesoscale and to study the related air–sea interaction processes (e.g., McPhaden and Zhang 2002; Chelton et al. 2004; Lee and McPhaden 2008; also cf. review articles by Liu 2002; Chelton and Xie 2010; Lee et al. 2010; Bourassa et al. 2010, and references therein). The Scatterometer Climatology of Ocean Winds (SCOW) based on QuikSCAT measurements (Risien and Chelton 2008) is used in this study to evaluate the climatology of CMIP. This dataset is available online (from http://cioss.coas.oregonstate.edu/scow/). Satellite scatterometer observations of wind stress are also available for the 1990s from European Remote Sensing Satellite-1 (ERS-1) and ERS-2. We used the QuikSCAT data because the SeaWinds instrument on QuikSCAT is a more advanced scatterometer and QuikSCAT has a much better sampling (covering 90% of the global ocean every day). Moreover, a carefully produced QuikSCAT wind stress climatology (i.e., SCOW) is available and has been widely used for various scientific investigations and evaluations of climate models and atmospheric reanalysis products (e.g., Slingo et al. 2009; Kanzow et al. 2010; Roquet et al. 2011; Xue et al. 2011; Johnson et al. 2012).
In addition to QuikSCAT measurements, we also use two atmospheric reanalysis products as additional reference. These are the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) Global Reanalysis 1 (NCEP-1; Kalnay et al. 1996) and the European Centre for Medium-Range Weather Forecast (ECMWF) Interim Re-Analysis (ERA-Interim; Dee et al. 2011). The atmospheric reanalysis products assimilate different types of atmospheric observations into the underlying atmospheric models. So the wind stress estimates from these products can be used as auxiliary reference products for evaluating CMIP. However, we are mindful about the fact that the wind fields based on atmospheric data assimilation products are subject to the limitations of the underlying atmospheric models and the data assimilation methods and have themselves various biases relative to satellite observations (e.g., Mestas-Nuñez et al. 1994; Chelton and Frielich 2005).
For the sake of convenience, we refer to the two reanalysis products as NCEP-1 and ERA-Interim. The NCEP-1 product encompasses the period from 1948 to the present while the ERA-Interim is from 1978 to the present. For NCEP-1, the climatology was computed for the period of 1970–99, the same period over which the CMIP climatologies were computed. For ERA-Interim, the wind stress climatology was computed for the period of 1978–2011. There are about half a dozen other atmospheric reanalysis products available from NCEP, ECMWF, the Japan Meteorological Agency (JMA), and the NASA Global Modeling and Assimilation Office (GMAO) that we have not analyzed. It is not our objective to perform a systematic evaluation of atmospheric reanalysis products using QuikSCAT data. We included NCEP-1 and ERA-Interim just as auxiliary reference products primarily because of the limited period of the QuikSCAT observations. Moreover, our results show that many of the biases of CMIP wind stress relative to QuikSCAT are similar to the biases relative to the two reanalysis products. The climatology from atmospheric reanalysis products were interpolated and the QuikSCAT climatology (originally on a 0.5° × 0.5° grid) was bin averaged to the same 2° × 2° grid as CMIP to facilitate the comparison.
Even though the QuikSCAT climatology is based on observations primarily in the 2000s and the CMIP simulations are for the late twentieth century, the comparison is still relevant because the CMIP internal variability does not match that observed in real time. The only real-time information in the CMIP simulations is the radiative and aerosol forcing. These forcings are not expected to cause a major difference in climatology on decadal time scales. The utility of the atmospheric reanalysis products also helps to address the potential issue of the dependence of climatology on natural decadal variability. We found that the differences of the climatology for the atmospheric reanalysis products between the entire periods of the products and the QuikSCAT period are much smaller than the difference in climatology between the CMIP and the reanalysis products. This justifies the utility of the QuikSCAT climatology to evaluate CMIP.
The annual-mean zonal wind stress from QuikSCAT, NCEP-1, ERA-Interim, the ensemble average of the 18 CMIP3 models, and the ensemble average of the 11 CMIP5 models are shown in Figs. 1a,b,d,f,h. The differences between the reanalysis or CMIP ensemble averages and QuikSCAT are presented in Figs. 1c,e,g,i. Note that the maximum color scale of the zonal wind stress difference (0.1 N m−2) is smaller than that for the zonal wind stress itself (0.25 N m−2). The differences between the reanalysis products and QuikSCAT are generally smaller than those between CMIP ensemble averages and QuikSCAT. This is understandable because the reanalysis products assimilate atmospheric observations. The differences between the atmospheric reanalysis products and QuikSCAT are primarily not because of the difference in periods over which the respective climatology was computed. We have analyzed the NCEP-1 and ERA-Interim products for the QuikSCAT period and found the dominant differences from QuikSCAT to be similar.
There are systematic biases of CMIP zonal wind at different latitude ranges. The most conspicuous differences are at midlatitude westerly wind regions, with CMIP eastward wind stress being too strong. This is shown by the positive differences in Figs. 1g,i at midlatitudes. At subtropical–tropical transition latitudes where the predominant wind is easterly, the CMIP westward wind stresses are too strong (negative differences from QuikSCAT). At high-latitude regions, CMIP zonal wind stresses tend to have a westward bias (negative differences from QuikSCAT). The mid- and high-latitude biases in CMIP wind stress climatology are related to the equatorward bias in the positions of midlatitude jets simulated by CMIP3 and CMIP5. The equatorward position bias of the austral midlatitude jet in CMIP3 twentieth-century simulations has been noted previously (e.g., Kidston and Gerber 2010). The eastward bias at midlatitudes and westward bias at subtropical–tropical transition latitudes and at high latitudes correspond to too strong an anticyclonic wind stress curl over the subtropical ocean gyres and too strong a cyclonic wind stress curl over the subpolar ocean gyres. These have the effect of increasing the strength of both the subtropical and subpolar gyres.
Horizontal gyre circulations contribute to meridional heat transport, especially in high-latitude oceans where the meridional overturning circulation carries little heat transport because the vertical temperature gradient is small at these latitudes (Lee et al. 2010). Therefore, inaccurate simulation of the strength of horizontal gyres because of biased wind stress would affect these meridional transports. Over the Southern Ocean, the biased wind stress and wind stress curl may influence the structure of the Antarctic Circumpolar Current (ACC), the meridional overturning circulation in the region, and surface heat flux and water mass formation. The Southern Ocean is believed to be an important sink for atmospheric carbon dioxide (CO2). The biased CMIP wind over this region thus has significant implications to CO2 flux in the region as well.
Figure 2 is the counterpart of Fig. 1 but for meridional wind stress. Note that the maximum color scale of the meridional wind stress difference (0.05 N m−2) is smaller than that for the zonal wind stress itself (0.01 N m−2). In the North Pacific and Atlantic Oceans, CMIP has substantially stronger midlatitude northward meridional wind stress (positive difference from QuikSCAT) and somewhat stronger southward meridional wind stress in subtropical–tropical latitudes (negative differences from QuikSCAT). In the Southern Ocean, the meridional wind stress in CMIP is generally more negative than QuikSCAT (more so in CMIP3 than CMIP5). In the eastern equatorial Pacific Ocean, the negative CMIP–QuikSCAT difference (blue color in Figs. 2g,i) indicates that the northward component of the southeasterly trade wind is too weak in CMIP. Périgaud et al. (1997) found that meridional wind in the equatorial Pacific influenced the magnitude of ENSO variability in a simple coupled model. Whether the biased meridional wind in the eastern equatorial Pacific Ocean in CMIP would affect the simulated behavior of ENSO warrants a further investigation.
The biases of CMIP ensemble averages relative to QuikSCAT discussed earlier are representative of many CMIP models. To illustrate this, the differences between the annual-mean zonal wind stress of individual CMIP3 models are shown in Fig. 3b–s. The QuikSCAT data are shown in Fig. 3a as a reference. Figure 4 is a similar representation to Fig. 3 but showing the differences of individual CMIP5 models from QuikSCAT. In Figs. 3 and 4, the maximum color scale for the differences between individual CMIP models and QuikSCAT are the same as that for QuikSCAT wind stress itself. This is a different presentation of color scale from Fig. 1 because the differences between individual CMIP models and QuikSCAT shown in Figs. 3 and 4 are generally larger than that between the CMIP ensemble and QuikSCAT shown in Fig. 1. The main characteristics of the model–data differences in CMIP ensemble averages discussed earlier (in reference to Figs. 1g,i) can be identified in many CMIP models (especially in terms of the positive model–data difference in midlatitude regions).
Since the model–data differences in zonal wind stress shown above tend to be more or less zonally oriented, we present a comparison of the zonally averaged zonal wind stress in Fig. 5. The QuikSCAT, NCEP-1, and ERA-Interim products are shown as black solid, black dashed, and black dashed–dotted curves. The thick red curves denote CMIP ensemble averages. The thin color curves correspond to the individual CMIP models. The color curves in the upper and lower panels represent CMIP3 and CMIP5, respectively. One could readily identify the major biases of CMIP zonal wind stress discussed earlier: the eastward biases at midlatitude westerly wind regions (30°–50°N and 30°–50°S) and westward biases at subtropical–tropical easterly wind regions (15°–25°N and 15°–25°S) and at high-latitude regions (poleward of 55°S and 55°N). To provide some quantitative description of the maximum biases of annual-mean zonal wind stress at different latitudes, the differences between the mean of ensemble averages for CMIP3 or CMIP5 and QuikSCAT are shown in Fig. 5c. The largest positive bias at midlatitude westerly wind regime occurs near 39°N and 39°S, with CMIP being stronger than QuikSCAT by approximately 55% (averaged between 41°N and 41°S). The largest negative bias for the subtropical–tropical easterly wind regime occurs near 19°N and 19°S, with CMIP being too strong by 26%. The largest negative bias in the high-latitude Southern Ocean occurs at 61°S, with CMIP being too weak by 27%. In the high-latitude northern oceans, the largest difference occurs at 63°N with CMIP being 668% too strong (because QuikSCAT is very small). At these latitudes, CMIP ensemble zonal wind stress generally has an opposite sign from QuikSCAT.
Near the equator, there is little model–data difference in the zonally averaged zonal wind stress when integrating over all ocean basins. This apparent consistency is actually misleading. As we show in the following, there are substantial biases in the individual oceans that tend to cancel out when integrated over all the basins. Figure 6 presents the equatorial zonal wind stress averaged between 2°S and 2°N as a function of longitude for individual CMIP models (thin color curves), CMIP3 or CMIP5 ensemble average (thick red curves), and QuikSCAT and the two reanalysis products (thick black curves). From the figure, the CMIP3 and CMIP5 ensemble averages look surprisingly similar (in part because of the vertical scale of the figure). Figure 7a consolidates Fig. 6 by plotting only QuikSCAT and the CMIP3 and CMIP5 ensemble averages to better illustrate the similarity and difference between CMIP3 and CMIP5. The CMIP zonal wind stresses over the equatorial Indian and Atlantic Oceans are both weaker than QuikSCAT (and the reanalysis products). In the equatorial Pacific, CMIP ensemble–average zonal wind stresses are fairly close to QuikSCAT in the east. However, they are weaker than QuikSCAT in the central Pacific and stronger than QuikSCAT in the western Pacific. When integrated zonally, there is partial compensation of the model–data differences between the central and western equatorial Pacific and between the equatorial Atlantic and Indian Oceans. That explains why there is such little model–data difference in global zonal–average zonal wind stress near the equator (Fig. 5).
The biased CMIP zonal wind stresses in the equatorial oceans are expected to affect the zonal structure (e.g., east–west slope) of the equatorial sea level or pycnocline in individual basins. In the equatorial Indian Ocean, the positive zonal wind stress causes sea level (pycnocline) to tilt upward (downward) toward the east. The opposite is true for the Atlantic Ocean. The weak equatorial zonal wind stresses in CMIP3 and CMIP5 for these two oceans imply that the zonal tilts of sea level or pycnocline in these basins would be too weak. This is demonstrated in Fig. 7b, which compares the longitudinal variation of sea level (SSH) from CMIP3 and CMIP5 ensemble averages with the observation-based estimate from Maximenko et al. (2009). The (zonal) sea level anomalies for individual basins were referenced to the equatorial average (2°S–2°N) for each basin both for the observation and for the CMIP models. This removes the bias of equatorial zonal–average sea level from individual models in individual basins and isolates the zonal variation. For the Indian Ocean Basin and Atlantic Ocean basin where the CMIP equatorial zonal wind stresses are too weak, the CMIP sea level slopes are also consistently too weak. For the Pacific, the too strong (weak) zonal wind stress in the western (central) part of the basin in CMIP ensemble averages also affects the zonal gradient of sea level (and thus pycnocline). They include most of the CMIP5 models analyzed in the present study. Despite the fact that the CMIP5 models are not identical, the biased ensemble-average sea level can be explained by the biased zonal wind stress in the equatorial zone.
Equatorial zonal wind stress and sea level (pycnocline) structure are important to various climate modes originating in the tropics as well as their interaction with multidecadal variability and climate change. For instance, the too weak zonal wind stress and too flat pycnocline in the Indian and Atlantic Oceans would affect the simulated behavior of the Indian Ocean zonal/dipole mode and Atlantic Niño. For instance, some CGCMs have difficulties capturing Atlantic Niño because of a too weak wind stress and too flat thermocline (Deser et al. 2006; Richter et al. 2012), which are believed to be related to the biased zonal wind and sea level (pycnocline) slope in the equatorial Atlantic. Cai and Cowen (2013) found that CMIP3 and CMIP5 models overestimated the amplitude of Indian Ocean dipole because the pycnocline depth in the eastern equatorial Indian Ocean is too shallow owing to the weak equatorial zonal wind. The biases in the equatorial Pacific also have implications to the behavior of interannual variability. In particular, the biased zonal wind stress structure in the central and western equatorial Pacific may affect the behavior of the anomalous warming events that occur in the central equatorial Pacific Ocean also known as the central Pacific El Niño (Kao and Yu 2009; Yu and Kim 2010b; Lee and McPhaden 2010), warm-pool El Niño (Kug et al. 2009), date line El Niño (Larkin and Harrison 2005), or El Niño Modoki (Ashok et al. 2007). The NCEP-1 equatorial zonal wind stress is generally weaker than that of QuikSCAT and ERA-Interim in much of the Indian, Pacific, and Atlantic Oceans. The weaker NCEP-1 equatorial easterly wind was also reported by Wittenberg (2004) in comparison with the in situ–based Florida State University (FSU) wind product.
To further illustrate the overall agreement of the spatial structure of the annual-mean wind stress among CMIP models, QuikSCAT data, and the reanalysis products, we present the Taylor diagram (Taylor 2001) of the annual-mean wind stress for CMIP3 (Fig. 8) and CMIP5 (Fig. 9). In these diagrams, QuikSCAT is used as the reference dataset (indicated by “OBS” in Figs. 8, 9). The statistics presented in the diagram, the normalized spatial standard deviation (along the radius) for each product and the spatial correlation with the QuikSCAT data (along the arc), are average statistics for zonal and meridional wind stress. For both figures, the four panels show the Taylor diagrams for the global ocean, 30°S–30°N, north of 30°N, and south of 30°S. For both CMIP3 and CMIP5, the models are more consistent among one another at low latitudes (30°S–30°N) than at higher latitudes, as indicated by the tighter clustering of the dots in Figs. 8b and 9b than in Figs. 8c,d and 9c,d. The consistency of the CMIP models with QuikSCAT data and the reanalysis products are also generally better at the 30°S–30°N latitudes than at higher latitudes. This is evident from the generally smaller distance from the model dots to QuikSCAT and to the reanalysis products denoted by OBS, N, and E. Also evident from the Taylor diagrams is that the ensemble averages of CMIP models C in Figs. 8 and 9 are generally closer to the observation than most of the individual CMIP models.
We next investigate the representation of seasonal anomalies by CMIP models (with reference to their respective annual mean). Comparison of the seasonal anomalies of CMIP3 and CMIP5 models with QuikSCAT data suggest that many models have a tendency to overestimate the overall magnitude of the seasonal anomalies. An example is shown in Fig. 10 for October anomalies of zonal wind stress where many of the CMIP3 models show more red and blue colors (larger anomalies) than QuikSCAT. The overestimation of the magnitude of seasonal anomalies in CMIP is more evident during the spring and fall (e.g., April or October) than during winter and summer (e.g., January and July), regardless of the hemisphere (not shown). To further illustrate the model–data comparison of the seasonal temporal changes, we present the Taylor diagrams for the temporal variability of seasonal anomalies, averaged over the global ocean and averaged between the statistics of zonal and meridional wind stress (Fig. 11). The magnitude of the seasonal anomalies of the CMIP3 and CMIP5 models (temporal standard deviation) as a whole is almost 1.5 times larger than that of the QuikSCAT data or the reanalysis products.
The diagnostics of the CMIP3 and CMIP5 wind stress presented above indicate that CMIP3 and CMIP5 are overall fairly similar to each other as a whole, especially in terms of the ensemble-average wind stress (e.g., Figs. 1, 2). The biases relative to QuikSCAT are similar in many CMIP3 and CMIP5 models (e.g., Figs. 3–7, 11). Using the Taylor diagrams as a metric, CMIP5 appears to be marginally better than CMIP3 in the representation of the structure of annual-mean wind stress (Fig. 8a versus Fig. 9a). This also appears to be the case for the representation of the seasonal cycle (Fig. 11a versus Fig. 11b), where the standard deviation of the CMIP5 models as a whole is close to 1.5 while that of CMIP3 is larger than 1.5; the correlation of the CMIP5 models with QuikSCAT cluster around 0.7, while those for CMIP3 are mostly less than 0.7. The standard deviation and correlation for the ensemble average of CMIP5 are also slightly better than those for CMIP3 (cf. the distance of the C dots to the reference datasets OBS, N, and E). Another feature revealed by Fig. 11 is that the ensemble averages of CMIP models are distinctively closer to observed seasonal wind stress anomalies than any individual CMIP model. This is reflected by the smaller distances of the C dots to the reference datasets (OBS, N, and E) than the distances between individual model dots and the reference datasets. There are less CMIP5 models (11) analyzed than CMIP3 models (18). The statistics based on a future comparison using the same families of models from CMIP3 and CMIP5 would be more rigorous.
Wind forcing acts on the ocean currents. The wind stress observed by QuikSCAT reflects the momentum transfer between the atmosphere and the moving ocean surface (i.e., “relative wind,” as opposed to the “absolute wind” reference to the stationary continent). However, the wind stress from CMIP and reanalysis products do not take into account the ocean current. In regions with a strong ocean current such as the ACC, this effect could be significant (e.g., Duhaut and Straub 2006). The fact that the reanalysis products are more similar to QuikSCAT than CMIP ensemble averages over much of the global oceans (e.g., Fig. 5) suggests that the effect of not accounting for the moving ocean currents may be minor in comparison to the fundamental biases in the CMIP models.
4. Concluding remarks
We have used QuikSCAT satellite scatterometer measurements and NCEP-1 and ERA-Interim products to evaluate the annual mean and seasonal cycle of wind stress simulated by 18 CMIP3 and 11 CMIP5 models. The ensemble CMIP3 and CMIP5 wind stresses are found to be fairly similar to each other and thus similar in terms of their differences from QuikSCAT and the reanalysis products. Generally speaking, there is a lack of significant improvement of CMIP5 over CMIP3. There are systematic biases of CMIP zonal wind stress at different latitude ranges. The CMIP annual-mean zonal wind stress has eastward biases at the midlatitude (30°–50°N and 30°–50°S) regions where the predominant winds are westerly, westward biases in subtropical–tropical transition latitudes (15°–25°N and 15°–25°S) where the predominant winds are easterly, and westward biases at high-latitude regions (poleward of 55°S and 55°N). At midlatitude westerly (subtropical–tropical easterly) regions, CMIP zonal wind stress is stronger than QuikSCAT by approximately 55% (26%). At the high-latitude Southern Ocean, CMIP zonal wind stress is weaker than QuikSCAT by about 27%. These biases correspond to too strong anticyclonic wind stress curl over the subtropical ocean gyres and too strong cyclonic wind stress curl over the subpolar ocean gyres, which would make the subtropical and subpolar gyres too intense and influence oceanic meridional heat transport. The biased wind stress over the high-latitude Southern Ocean would affect various aspects of the Southern Ocean circulation.
When integrated zonally across all oceans, the equatorial zonal wind stresses are relatively close to QuikSCAT. However, significant biases exist in individual basins. In the equatorial Atlantic and Indian Oceans, CMIP ensemble zonal wind stresses are both too weak. Consistent with these wind biases, the zonal slope of sea level from CMIP3 and CMIP5 ensemble averages are both too small in these basins, implying the same for zonal slope of the pycnocline. In the equatorial Pacific, CMIP zonal wind stresses are fairly close to QuikSCAT in the east but are too weak in the central Pacific and too strong in the western Pacific. These biases in equatorial zonal wind stress are expected to affect the simulation of climate modes originating in the tropics such as the Indian Ocean zonal/dipole modes, Atlantic Niño, and ENSO (especially the so-called central Pacific El Niño).
The CMIP models tend to overestimate the magnitude of seasonal variability, especially during spring and autumn. The standard deviation of seasonal variability averaged over the global ocean from CMIP models as a whole is almost 50% larger than that observed by QuikSCAT and inferred from the reanalysis products.
The biases in CMIP wind stress climatology are not sensitive to the decades over which the climatology was defined for CMIP models and for the reference datasets (see appendix). The cause of the biases in CMIP wind stress climatology needs to be examined. Our findings also prompt investigations of the consequences of the biased CMIP wind stress on the simulated ocean circulation, meridional transports of properties (e.g., heat), the representation of interannual and decadal variability that could be affected by the climatological state, and other air–sea fluxes (e.g., latent heat and CO2). The potential effects of the biased climatological state on climate projection also need to be assessed.
We acknowledge the GCM modeling groups, the Program for Climate Model Diagnosis and Intercomparison (PCMDI), and the WCRP's Working Group on Coupled Modeling for their roles in making available the WCRP CMIP3 and CMIP5 multimodel datasets. Support of these data sets is provided by the Office of Science, U.S. Department of Energy. This research was carried out in part at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.
Uncertainties due to Decadal and Diurnal Variability
A possible concern of this study might be how decadal variability affects our conclusions given the dissimilar periods for climatologies (i.e., 1970–99 for NCEP-1 and CMIP; 1978–2011 for ERA-Interim; and 1999–2009 for QuikSCAT). We performed further analysis of climatologies defined over different periods and found that our conclusions were not affected.
We first examined the sensitivity of CMIP5 climatology to the period over which the climatology is defined. The climatology for 1970–99 as presented in the paper is extremely similar to that computed over a 1.5-century (1850–2005) period (Fig. A1, red and blue curves). Figure A1 is taken from Figs. 5b and 6b without the individual CMIP models (the thin color curves) and with the addition of the 1.5-century CMIP5 climatology (the blue dashed curve). The differences of CMIP5 from the reference datasets are much larger than the differences in CMIP5 climatology defined over different periods. Therefore, the 30-yr climatology used in this study is representative of the climatology over a much longer period (i.e., not subject to any significant effect of decadal variability).
Figure A2 presents a comparison of CMIP5, NCEP-1, and ERA-Interim climatology for the common period of 1978–2005. This figure is very similar to Fig. A1, where the climatology was defined over different periods. The findings remain the same: for example, CMIP5 zonal wind stress has westward biases at midlatitudes and eastward biases at high latitudes and at subtropical–tropical latitudes, and CMIP5 zonal wind stresses are too weak in the equatorial Indian and Atlantic Oceans and too strong in the western equatorial Pacific Ocean. From Figs. A1 and A2, it is seen that these biases in CMIP wind stress relative to all three reference datasets are similar.
Another concern of this study might be the sampling error in the QuikSCAT data. QuikSCAT, sampling approximately 90% of the World Ocean on a daily basis, is insufficient to capture diurnal variability of the wind field. The magnitude of the bias of time mean QuikSCAT wind stress caused by the limited sampling (e.g., Fig. 2 in Lee et al. 2008; Guan et al. 2013) is several times smaller than the differences between CMIP and QuikSCAT shown in this study.