ENSO spatial and temporal properties in hybrid CGCMs

Fig. 2.4. Same as fig.1 from Belmadani et al. (2010), but for the hybrid models. From left to right and top to bottom: SODA 1.4.2, CSIRO-MK3.0, INGV-ECHAM4, INM-CM3.0, IPSL-CM4, UKMO-HadCM3, UKMO- HadGEM1.

All the CGCMs tend to have a wider zonal extent and a narrower meridional extent (except maybe UKMO-HadGEM1) than SODA. However, like SODA they all exhibit a meridionally asymmetric pattern for SST anomalies associated to ENSO, with stronger variability in the Southern tropical Pacific than in the Northern tropical Pacific, highlighting the influence of large-scale ENSO anomalies on nearshore variability off Peru. In terms of amplitude of SST variability, most models fall within the observed range. Computing the spatial correlation coefficient between patterns of first SVD mode SST anomalies from the CGCMs and from SODA allows quantifying the differences in spatial distribution of ENSO-related anomalies (table 2.2): whereas INGV-ECHAM4, IPSL-CM4, UKMO-HadCM3 and UKMO-HadGEM1 are well correlated to SODA, CSIRO-MK3.0 and especially INM-CM3.0 exhibit weaker values, which has to be related to the misplaced variability peaks in these models. The latter CGCM tends to produce a central Pacific El Niño (Larkin and Harrisson, 2005; Ashok et al., 2007, Kao and Yu, 2009), which is known to have very little influence on the South American coast (Kug et al., 2009), thereby limiting the interest of such model for the study of Peru- Chile climate change.

Model Explained covariance (%)

Explained SST variance (%)

Explained τx

variance (%)

PC time series correlation (%)

SST amplitude (adim.)

ENSO period (years)

Spatial correlation*

with SODA (%)

CSIRO-MK3.0 77.3 28.9 7.4 85 0.84 2.0 72.4

INGV-ECHAM4 89.6 23.7 9.3 85 0.93 4.0 88.6

INM-CM3.0 69.3 18.6 11.4 88 1.02 3.3 49.3

IPSL-CM4 93.0 34.8 10.2 94 0.82 2.5 88.2

UKMO-HadCM3 89.6 22.2 18.8 92 0.79 2.9 82.2

UKMO-HadGEM1 68.3 14.6 7.4 82 0.99 4.0 87.0

SODA-1.4.2 86.3 29.3 19.4 91 0.90 3.7 100.0

Table 2.2. Spatial and temporal characteristics of the first SVD mode between anomalies of wind stress and SST in the tropical Pacific (fig. 2.4) for the CGCMs and SODA. * The domain used for the spatial correlation is (11°S-11°N, 135°E-80°W)

In the “real world” characterized here by SODA, more than 85% of the covariance between anomalies of surface wind stress and SST can be explained by ENSO (table 2.2). In the model world, this rate varies between about 70% to more than 90%. In particular, CSIRO-MK3.0, INM-CM3.0 and UKMO-HadGEM1 exhibit weaker explained covariance than SODA. In terms of SST variance, about 30% is explained by ENSO in the case of SODA, whereas most CGCMs show weaker values, especially INM-CM3.0 and the two UK Met Office models (HadCM3 and HadGEM1). In terms of wind variability, a significantly weaker part is explained by ENSO in the climate models compared to SODA, with the exception of UKMO- HadCM3.

The period of the ENSO cycle is heterogeneous among the hybrid models, ranging from biennial (CSIRO-MK3.0) to quadriannual (INGV-ECHAM4, UKMO-HadGEM1), versus 3.7 years for the reanalysis. This reflects the fact that even if hybrid CGCMs have a closer ENSO

time scale on average to that of SODA than CGCMs with a dominant zonal advective feedback or a dominant thermocline feedback (Belmadani et al., 2010), other processes such as the spatial distribution of the wind patch for instance are also likely to affect the ENSO time scale (Guilyardi et al., 2004; Zelle et al., 2005; Capotondi et al., 2006; Merryfield, 2006):

a peak of wind variability displaced to the west (resp. east) tends to favour shorter (resp.

longer) time scales (An and Wang, 2000), due to a shorter (resp. longer) time taken by upwelling equatorial Rossby waves to reach the western boundary where they are reflected as Kelvin waves and eventually act as a negative feedback to the growth of El Niño, according to the delayed oscillator paradigm (Suarez and Schopf, 1988). In addition, the meridional gradient of the wind stress can also have an effect on the ENSO time scale, as the phase velocity of the excited Rossby waves decreases poleward (Kirtman, 1997). In this respect, INGV-ECHAM4, INM-CM3.0 and UKMO-HadGEM1 exhibit the most realistic time scales among the ensemble.

Fig. 2.5. Autocorrelation of (left) Niño3 SST and (right) Niño4 SST indices over an 18-month lead–lag period for the CGCMs and for SODA. The horizontal dotted line indicates a value of e^-1=0.368.

Another measure of the ENSO time scale is provided by the autocorrelation functions of ENSO indices such as Niño3 SST and Niño4 SST anomalies (fig. 2.5): the width at decorrelation values (e^-1) provides an estimate of El Niño or La Niña event duration. A first order approximation for the ENSO period can thus be obtained by doubling such value.

INGV-ECHAM4 exhibits a significantly longer cycle (~34 months i.e. 2.8 years for Niño3 and ~42 months i.e. 3.5 years for Niño4) than the other models and SODA, which is qualitatively consistent with the estimate derived from the SVD analysis (table 2.2). On the other hand, CSIRO-MK3.0, INM-CM3.0 and IPSL-CM4 tend to have the shortest periods among the hybrid models: ~1.9 years for Niño3 and between 2.0 and 2.3 years for Niño4,

which is also consistent with the SVD analysis, except maybe for INM-CM3.0. The UK Met Office models exhibit the best agreement with SODA: ~2.2-2.3 years for Niño3, 2.7 years for Niño4. In the SVD analysis, the HadGEM1 model exhibits a quadriannual ENSO, which is also the case for SODA (3.7 years); HadCM3 exhibits a shorter time scale (2.9 years), though with a secondary peak beyond 4 years. Overall, these results are qualitatively consistent with the SVD analysis for most models. As the autocorrelation function only provides an estimate of event duration and not of cycle length, quantitative differences with the ENSO period estimates derived from the SVD analysis should not be over-interpreted. Note that the same diagnosis was used by Joseph and Nigam (2006) for a subset of the CMIP3 models, including UKMO-HadCM3 for which results were found to be consistent with the ones presented here.

Following Joseph and Nigam (2006), the decorrelation threshold in the opposite phase (-e^-1) is used to assess the ability of the models to shift from El Niño to La Niña and back (fig. 2.5).

Only CSIRO-MK3.0 and IPSL-CM4 are able to cross this threshold, opposite to the other models and SODA, which means that these two models tend to have a too regular cycle. In the real world, a La Niña event does not always follow an El Niño event (and vice-versa): in some cases neutral conditions can prevail for several years, such in 1978-1982 or in 1959- 1962; in other cases several consecutive El Niño or La Niña events can occur, such as in 1999-2001 or in 2002-2005 (fig. 1.19). None of the hybrid models stays in the same phase over the considered 36-month time interval unlike the MIROC3.2-HIRES model for instance (see Joseph and Nigam, 2006 – their figure 2), confirming the oscillatory behaviour of ENSO in these models. Note that whereas CSIRO-MK3.0 and IPSL-CM4 reach their minima in the considered interval, it is not the case for the other models, which reflects the fact that the latter tend to have ENSO periods exceeding 36 months. This might provide an explanation for the quantitative differences with estimates from the SVD analysis.

Model Niño3 SSTA:

RMS (°C)

Niño3 SSTA:

skewness (°C)

Correlation (%):

Niño3 SSTA, Niño4 TXA

Niño4 SSTA:

RMS (°C)

Niño4 SSTA:

skewness (°C)

Correlation (%):

Niño4 SSTA, Niño4 TXA CSIRO-MK3.0 0.91 -0.17 61.5 0.82 +0.20 57.8 INGV-ECHAM4 0.79 -0.02 65.9 0.60 +0.05 73.3 INM-CM3.0 0.90 +0.23 54.1 1.02 +0.50 55.6

IPSL-CM4 1.00 +0.11 70.5 0.70 -0.30 67.8

UKMO-HadCM3 0.88 +0.32 67.5 0.81 +0.16 74.5 UKMO-HadGEM1 0.69 +0.03 49.2 0.56 +0.09 57.6 SODA-1.4.2 0.96 +0.91 71.5 0.64 -0.32 78.2 Table 2.3. Statistics of Niño3 and Niño4 SST indices for the CGCMs and for SODA.

Niño3 and Niño4 indices are frequently used in the literature to assess simple ENSO statistics, such as ENSO amplitude (RMS of the time series), skewness (defined as in Belmadani et al., 2010) or Gill-type linearity (assessed with the correlation coefficient between western Pacific

wind anomalies and eastern Pacific SST anomalies). Table 2.3 presents these statistics computed for our model ensemble.

Most CGCMs feature an amplitude of eastern Pacific ENSO in relatively good agreement with that of SODA, except INGV-ECHAM4 and UKMO-HadGEM1 which underestimate ENSO amplitude by ~18% and ~28%, respectively. This is qualitatively consistent with the results of van Oldenborgh et al. (2005), and more importantly, quantitatively consistent with Guilyardi (2006) who used the same diagnostic as the one presented here. A notable exception is UKMO-HadCM3 for which we find an amplitude of 0.88°C versus 0.77°C for Guilyardi (2006). Note however that Guilyardi (2006) indicates an error bar of +/-0.09°C, i.e. an amplitude ranging from 0.68°C to 0.86°C. He also uses a longer time span than the one considered here: 341 years versus 148 years, respectively. Hence, the differences between the two studies for this model might be related to decadal and multi-decadal ENSO modulation - which are significant for this model (see below) -, an issue which has been illustrated by Wittenberg (2009) with a 2000-year simulation performed with the GFDL-CM2.1 model. On the other hand, most CGCMs tend to overestimate western Pacific ENSO amplitude, except INGV-ECHAM4, IPSL-CM4 and UKMO-HadGEM1 which exhibit a rather good agreement with SODA. Such behaviour may be related to the overestimated westward extension of the ENSO SST anomaly pattern for most CMIP3 models (fig.2.4), itself related to the overestimated westward extension of the cold tongue due to stronger zonal winds than observed (see Belmadani et al., 2010). Note that the former statement is less true for UKMO- HadGEM1 and to a smaller extent for INGV-ECHAM4 (fig.2.4), which tends to confirm this hypothesis.

It is known that SST anomalies are positively skewed in the eastern tropical Pacific and negatively skewed in the western tropical Pacific, a see-saw pattern which is captured by the SODA reanalysis (table 2.3). Consistently with results from Belmadani et al. (2010), IPSL- CM4 is the only hybrid model with a negative skewness in Niño4 (not only is the sign correct but also the magnitude). Belmadani et al. (2010) proposed to attribute such behaviour to the rather low nonlinear zonal advection for all the hybrid models except IPSL-CM4 which presents a value close to that of SODA (see their figure 6). INGV-ECHAM4 and UKMO- HadGEM1 have the smallest positive values, which is likely related to the smaller SST variability. ENSO skewness presents marked biases also in the eastern Pacific: all the models underestimate it by at least a factor three (for UKMO-HadCM3) and two of them even show negative values (CSIRO-MK3.0 and INGV-ECHAM4).

In terms of linearity, most models have Niño3SSTA/Niño4TXA correlation values of the order of SODA (71.5%), except INM-CM3.0 and UKMO-HadGEM1 which have a lower degree of linearity (respectively 54.1% and 49.2%). This is consistent with the results of Guilyardi (2006) (his figure 8) who find larger scatter between Niño3SSTA and Niño4TXA for these two models than for the other hybrid models. In the western Pacific, three models have significantly lower correlation values between SST anomalies and local wind anomalies:

CSIRO-MK3.0, INM-CM3.0 and UKMO-HadGEM1 (table 3).

Fig. 2.6. Seasonal phase-locking of ENSO: standard deviation of (left) Niño3 SST and (right) Niño4 SST indices for each calendar month, for the CGCMs and for SODA.

Another important feature of ENSO is its seasonal phase locking. Indeed, as noted in the introductory chapter, El Niño tends to peak in the central to eastern tropical Pacific during the boreal winter season (e.g. Rasmusson and Carpenter, 1982). El Niño Modoki also tends to be synchronized with the seasonal cycle, with a peak season in the central to western tropical Pacific ranging from boreal summer to winter (e.g. Ashok et al., 2007). Following Joseph and Nigam (2006), we use the standard deviation of ENSO indices (Niño3 for the conventional ENSO and Niño4 for ENSO Modoki) for each calendar month to assess phase locking in the CGCMs (fig. 2.6). Models exhibit widespread behaviour in terms of ENSO peak season in both the eastern and the western Pacific. In the eastern Pacific, IPSL-CM4, UKMO- HadGEM1 and UKMO-HadCM3 have their ENSO peaking in boreal winter (consistently with Joseph and Nigam (2006) for the latter model), just like SODA. On the other hand, CSIRO-MK3.0 and INM-CM3.0 peak in boreal spring and summer, respectively. INGV- ECHAM4 does not exhibit any clear seasonal preference. In the central/western Pacific, CSIRO-MK3.0, UKMO-HadCM3 and UKMO-HadGEM1 have their peak season in boreal fall, like for SODA. IPSL-CM4 tends to peak in winter, whereas INM-CM3.0 peaks in

summer, though with a marked semi-annual cycle and a secondary maximum in fall. Again, INGV-ECHAM4 is not seasonally-discriminating.

As explained in the introductory chapter and in the first section of the present chapter (Belmadani et al., 2010), ENSO is subject to low-frequency modulation, which can impact significantly its structure and dynamics. Here we use wavelet analysis together with the N3VAR and N4VAR (equivalent to N3VAR for the Niño4 region) indices to document such decadal modulation in the CGCMs (figures 2.7 and 2.8). As expected, SODA exhibits significant low-frequency modulation of ENSO amplitude, with clear shifts in the mid- to late 70s and in the early 90s in both Niño3 and Niño4 regions. A change from bi-annual to quadri- annual cycles in the mid-70s is also evident in the corresponding wavelet spectra. Most CGCMs tend to underestimate decadal modulation of ENSO amplitude in the eastern Pacific, which is particularly obvious for UKMO-HadGEM1, INGV-ECHAM4 and to a smaller extent for INM-CM3.0. UKMO-HadCM3 and CSIRO-MK3.0 exhibit the best agreement with SODA, whereas IPSL-CM4 tends to slightly overestimate decadal modulation. Changes in the frequency of ENSO can be detected for most models except UKMO-HadGEM1 which tends to maintain a quadri-annual ENSO during the considered 150-year study period. Note that these results are not always consistent with those of Lin (2007), who finds for instance significant interdecadal variability of ENSO amplitude and frequency for UKMO-HadGEM1 and little variability of the amplitude for IPSL-CM4. However, his study is based on a different climate scenario, in which GHG forcing slowly increases at the observed rate (20C3M), which may provide an explanation for such differences. In addition, his findings are based on a qualitative assessment of Niño3 SST wavelet spectra, rather than on a quantitative assessment such as the one performed here on ENSO amplitude with N3VAR.

Opposite to the eastern Pacific, most CGCMs tend to strongly overestimate decadal modulation in the western Pacific, except INGV-ECHAM4 and UKMO-HadGEM1 which exhibit a rather good agreement with SODA in terms of N4VAR mean value and standard deviation. Such behaviour may be related to the overestimated SST variability in this region.

Opposite to eastern Pacific ENSO, changes in the frequency of central/western Pacific ENSO are detected for all the hybrid models.

Fig. 2.7: Low-frequency modulation of ENSO for the CGCMs and for SODA. For each model: a) Wavelet power spectrum of Niño3 SST anomalies (only values above the 90% significance level are plotted); b) Corresponding 2-7 year scale-average time series (N3VAR). Units are (°C)².

Fig. 2.8. Same as fig. 2.7, except for Niño4.

Overall, INGV-ECHAM4, IPSL-CM4 and UKMO-HadCM3 exhibit the ‘best’ agreement with SODA in terms of spatial and temporal properties of ENSO, whereas INM-CM3.0 is the most unrealistic among the hybrid models. CSIRO-MK3.0 does a good job with respect to ENSO amplitude and decadal modulation, whereas UKMO-HadGEM1 performs well in terms of ENSO pattern, period, phase shift and phase locking. We decided here to select the latter model rather than the former, because the afore-mentioned three “best” models do not perform very well in terms of ENSO period (except INGV-ECHAM4), and because at least one of them does not perform very well in terms of ENSO phase shift and phase locking. In contrast, all these three models have good skills in terms of ENSO amplitude and decadal modulation of ENSO period, and only one of them (INGV-ECHAM4) features a significantly underestimated decadal modulation of ENSO amplitude. In addition, as mentioned at the end of the introductory chapter, climate change downscaling experiments are performed over 10- year periods, which are too short for the assessment of decadal modulation at the regional scale. For the rest of the analysis, we thus decided not to consider the CSIRO-MK3.0 and INM-CM3.0 models, and focus the analysis of the behaviour of the IPCC models in the eastern South Pacific on the four ‘best’ models in terms of large scale tropical variability:

INGV-ECHAM4, IPSL-CM4, UKMO-HadCM3 and UKMO-HadGEM1.

At this point, it is important to note that although we have tried to take into account as many aspects of ENSO as possible for the skill assessment of CMIP3 CGCMs, this analysis does not pretend to be exhaustive by any means. ENSO is a complex multiform phenomenon resulting from interactions between the ocean and the atmosphere on a range of time scales, from intraseasonal to multi-decadal. Even though our understanding of ENSO has remarkably increased since the early works of Bjerknes in the 1960s, there are still numerous issues regarding its structure, dynamics and teleconnections that are yet to be understood. Here we tried to combine a dynamical approach, allowing to diagnose biases of the models in the representation of complex coupled feedbacks, with a more conventional statistical approach, based on commonly used ENSO metrics, in order to identify the most relevant CGCMs for the study of the impact of global warming on the Humboldt Current System, at least in terms of large-scale equatorial forcing. Other metrics and/or diagnostics might provide a different group of candidates for such a regional downscaling project. However, we have shown in the first section of the present chapter that our results are consistent to a large extent with those of Guilyardi (2006), who used different diagnostics to identify the different feedbacks (or ENSO modes) in the CMIP3 CGCMs. In addition, we have checked that the results presented in the present section are consistent to a large extent with other studies using similar simple ENSO

statistics to assess model performance. This mixed dynamical/statistical approach and the consistency of our results with those from previous studies gives us confidence in the choice we have made. Last, downscaling experiments are meant to be performed with forcings from several CGCMs, in order to be able to provide error estimates for projected changes at the regional scale. It is an a posteriori way to check the relevance of our choice of CGCMs.

It is also important to mention that we have chosen to assess ENSO behaviour in the models before analyzing their characteristics at the boundaries of the regional domain – instead of the opposite, or instead of performing both analyses for all the models – mainly for a practical reason. Indeed, most CGCM fields that were required for the analysis of ENSO variability presented here are either surface fields (SST, zonal wind stress) or vertical equatorial sections (temperature, salinity), which represent a relatively reasonable – though significant – amount of data over the CMIP3 multi-model ensemble. Conversely, the diagnostics performed at the regional scale require several bidimensional fields at the three open boundaries of the regional domain (zonal and meridional currents, temperature, salinity) as well as surface atmospheric fields such as zonal and meridional wind stress (see next section). Such a higher number of bidimensional fields requires much more disk space and processing, hence the relevance of limiting the number of models to be considered in the regional analysis, which is presented in the next section.