Content uploaded by Nicolas Jourdain
Author content
All content in this area was uploaded by Nicolas Jourdain on Dec 08, 2015
Content may be subject to copyright.
Further insights on the influence of the Indian Ocean Dipole on the1
following year’s ENSO from observations and CMIP5 models2
Nicolas C. Jourdain1,2,3∗, Matthieu Lengaigne4,5, J´
erome Vialard4
3
Takeshi Izumo4and Alexander Sen Gupta3,6
4
1CNRS, LGGE, F-38402 Grenoble, France5
2Univ. Grenoble Alpes, LGGE, F-38402 Grenoble, France6
3ARC Centre of Excellence for Climate System Science, UNSW, Sydney, Australia7
4LOCEAN-IPSL, Sorbonne Univ. (UPMC, Univ Paris 06)-CNRS-IRD-MNHN, Paris, France8
5Indo-French Cell for Water Sciences, IISc-NIO-IITM-IRD Joint International Laboratory, NIO,
Goa, India
9
10
6Climate Change Research Centre, UNSW, Sydney, Australia11
∗Corresponding author address: Nicolas C. Jourdain,Laboratoire de Glaciologie et G´
eophysique de
l’Environnement, 54 rue Moli`
ere, Domaine Universitaire, BP96, 38402 St Martin d’H`
eres Cedex,
France
12
13
14
E-mail: njourdain@lgge.obs.ujf-grenoble.fr15
Generated using v4.3.2 of the AMS L
A
T
E
X template 1
LaTeX File (.tex) Click here to download LaTeX File (.tex, .sty, .cls, .bst, .bib) jourdain_JCLI_2015_resub.tex
ABSTRACT
2
Recent observational studies have suggested that negative / positive Indian
Ocean Dipole events (nIOD and pIOD respectively) favor a transition towards
El Ni˜
no / La Ni˜
na events one year later. These statistical inferences are how-
ever limited by the length and uncertainties in the observational records. In
this paper, we compare observational datasets with 21 155-year CMIP5 his-
torical simulations to assess IOD and El Ni˜
no Southern Oscillation (ENSO)
properties along with their synchronous and delayed relationships. In the ob-
servations and most CMIP5 models, we show that: El Ni ˜
nos tend to be fol-
lowed by La Ni˜
nas, but not the opposite; pIODs co-occur more frequently
with El Ni˜
nos than nIODs with La Ni˜
nas; nIODs tend to be followed by El
Ni˜
nos one year later less frequently than pIODs by La Ni˜
nas; including an
IOD index in a linear prediction based on the Pacific warm water volume im-
proves ENSO peak hindcasts at 14 months lead. The IOD-ENSO delayed
relationship partly results from a combination of ENSO intrinsic properties
(e.g. the tendency for El Ni˜
nos to be followed by La Ni˜
nas) and from the
synchronous IOD-ENSO relationship. Our results however reveal that this is
not sufficient to explain the high prevalence of pIOD-Ni˜
na transitions in the
observations and 75% of the CMIP5 models, and of nIOD-Ni˜
no transitions in
60% of CMIP5 models. This suggests that the tendency of IOD to lead ENSO
by one year should be explained by a physical mechanism that however re-
mains elusive in the CMIP5 models. The ability of many CMIP5 models to
reproduce the delayed influence of the IOD on ENSO is nonetheless a strong
incentive to explore extended-range dynamical forecasts of ENSO.
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
3
1. Introduction39
The El Ni˜
no Southern Oscillation (ENSO) is the dominant mode of climate variability at inter-40
annual timescales (e.g. McPhaden et al. 2006). El Ni˜
no events are characterized by warm Sea41
Surface Temperature Anomalies (SSTA) in the central and eastern tropical Pacific, associated with42
enhanced deep atmospheric convection and westerly wind anomalies in the western and central43
Pacific. ENSO emerges from internal dynamics of the ocean-atmosphere coupled system in the44
tropical Pacific, with climate impacts at an almost global scale through atmospheric teleconnec-45
tions (e.g. Trenberth et al. 1998). These global climate impacts are a strong incentive for improved46
long-term forecasts of ENSO. ENSO is however associated with a complex set of ocean and at-47
mosphere processes. Improving the skill of ENSO forecasts has thus remained a major challenge48
for the scientific community over the last 30 years.49
The basic physics of the evolution of El Ni˜
no/La Ni˜
na events are now reasonably well under-50
stood (see the reviews by Wang and Picaut 2004; Collins et al. 2010; Clarke 2014). The so-called51
Bjerknes feedback (Bjerknes 1969), a positive air-sea feedback loop in the Pacific, provides the52
necessary instability for an El Ni˜
no to develop, but favorable conditions are needed for the Bjerk-53
nes feedback to induce an El Ni˜
no event. In particular, the importance of heat stored in the upper54
tropical Pacific Ocean is now widely recognized within the framework of the recharge/discharge55
oscillator (Jin 1997; Wyrtki 1985; Meinen and McPhaden 2000). Prior to an El Ni ˜
no, there is a56
build-up in the Warm Water Volume (WWV, defined as the volume of sea water above the 20◦C57
isotherm in the Equatorial Pacific, Meinen and McPhaden 2000). This anomalously high WWV58
favors the growth of SSTAs during the ensuing El Ni ˜
no, sustained by the Bjerknes feedback. The59
ENSO-WWV relationship appears to change on decadal or longer timescales (Lee and McPhaden60
2010; Tang and Deng 2010; McPhaden 2012), as is the case with other ENSO characteristics such61
4
as its preferred frequency or spatial pattern (An and Wang 2000; Leloup et al. 2007; Wittenberg62
2009). Izumo et al. (2014) nonetheless found that the WWV lead on ENSO was statistically sig-63
nificant over the period 1872-2008.64
During an El Ni˜
no, the eastward shift of the Walker cell induces anomalous subsidence, in-65
creased downward surface shortwave fluxes, and reduced near-surface wind over the Indian Ocean.66
This favors a tropical Indian Ocean Basin-wide (IOB) warming during the peak of El Ni˜
no events67
(e.g. Klein et al. 1999; Ohba and Ueda 2005, 2009b) that is maintained throughout boreal summer68
by local air-sea interactions (e.g. Xie et al. 2009). The tropical Indian Ocean is also home to the69
Indian Ocean Dipole (IOD, Reverdin et al. 1986; Saji et al. 1999; Murtugudde et al. 2000; Web-70
ster et al. 1999). A positive IOD (pIOD) is characterized by strong negative SSTA near the coast71
of Sumatra and weaker positive SSTA in the western Indian Ocean. Similar feedbacks to those72
occurring for ENSO in the Pacific explain the IOD growth during boreal summer until its peak in73
boreal fall (Saji et al. 1999; Cai and Cowan 2013), even though fundamental differences of air-sea74
interactions have been reported between the Pacific and Indian basins (Li et al. 2003). There is75
now strong evidence that the IOD is an intrinsic mode of variability of the Indian Ocean (e.g. An-76
namalai et al. 2003; Fischer et al. 2005; Behera et al. 2006; Luo et al. 2008, 2010). El Ni ˜
no events77
do however have a tendency to trigger pIOD events, due to the easterly wind anomalies that they78
favor over the eastern equatorial Indian Ocean (Annamalai et al. 2003). As a result, pIODs often79
co-occur with El Ni˜
no events and negative IODs (nIODs) co-occur with La Ni˜
na events (Anna-80
malai et al. 2003; Li et al. 2003; Ashok et al. 2003; Bracco et al. 2005; Fischer et al. 2005; Behera81
et al. 2006). This relationship between co-developing IOD and ENSO events (hereafter referred82
to as the ”synchronous IOD-ENSO relationship”) is highlighted by the ∼0.5 to 0.7 correlation83
between the ENSO and IOD indices in boreal fall (correlation at zero-lag in Fig. 1). A decadal84
modulation of the IOD has been noted by Tozuka et al. (2007) and Ummenhofer et al. (2009).85
5
The strength of the synchronous IOD-ENSO relationship also undergoes interdecadal variability,86
as shown by Yuan and Li (2008), Santoso et al. (2012), and Izumo et al. (2014) who all noted a87
weaker IOD-ENSO relationship over the period 1950-1970 than over the recent period.88
Distinct asymmetries exist between the positive and negative phases of both ENSO and IOD, in89
terms of their intensity, duration or phase transition (e.g. Ohba and Ueda 2009a; Okumura et al.90
2011; Ohba and Watanabe 2012; Choi et al. 2013). The intensity of both ENSO and IOD is asym-91
metric in that the amplitude of El Ni˜
no (pIOD) events tends to be larger than the amplitude of92
La Ni˜
na (nIOD) events (Burgers and Stephenson 1999; Jin et al. 2003; Wu et al. 2010; Cai et al.93
2012). There is also an asymmetry in the duration of ENSO events: most El Ni˜
no events terminate94
rapidly after their peak, while La Ni˜
na events tend to persist into the following year (Ohba and95
Ueda 2009a; Ohba et al. 2010; Okumura et al. 2011). Finally, ENSO exhibits a transition asym-96
metry, whereby there is a higher probability for El Ni ˜
no to be followed by La Ni˜
na the following97
year than vice versa (Ohba et al. 2010; Choi et al. 2013).98
Previous studies have proposed that variability in the Indian Ocean sector could influence ENSO99
during the following year (e.g. Meehl 1987; Clarke and van Gorder 2003; Kug and Kang 2006).100
For instance, it has been suggested that the IOB could foster a rapid transition from El Ni˜
no to101
La Ni˜
na (Kug and Kang 2006; Ohba and Ueda 2007, 2009a; Kug and Ham 2012; Santoso et al.102
2012; Ohba and Watanabe 2012) by favoring the development of easterly wind anomalies over103
the western Pacific in winter (Annamalai et al. 2005; Dayan et al. 2015). Similarly, Izumo et al.104
(2010) have suggested that IOD events could be precursors of ENSO. Based on satellite data, they105
showed that positive (negative) IODs tended to be followed by La Ni˜
na (El Ni˜
no) events approx-106
imately 14 months later (Fig. 1). They demonstrated significant skill in predicting ENSO events107
14 months in advance by jointly using IOD and WWV as predictors in a simple linear regression108
model. Hereafter, this influence of the IOD on following year’s ENSO state is referred to as the109
6
”delayed IOD-ENSO relationship”. Izumo et al. (2010) suggested the following mechanism to110
explain that delayed relationship. The warm anomaly in the Eastern Indian Ocean associated with111
a negative IOD induces an easterly anomaly in the Equatorial Pacific. This negative anomaly and112
the associated wind signal over the Pacific abruptly disappear in November-December at the ter-113
mination of the IOD event. The equatorial wave response to this wind anomaly and its sudden114
relaxation drives an eastward current anomaly in the western-central Equatorial Pacific. The latter115
favors the onset of a delayed El Ni˜
no by pushing the warm pool eastward. This IOD influence116
interacts with the intrinsic Pacific WWV preconditioning. For example, a nIOD can interact con-117
structively with positive WWV conditions to favor the onset of an El Ni˜
no. More recently, Izumo118
et al. (2015) suggested that the IOD may reinforce the IOB tendency to favour ENSO phase tran-119
sition. When IOD and IOB events co-occur, the abrupt IOD eastern pole demise at the end of120
fall indeed favours a faster development of IOB-induced wind anomalies in the western Pacific in121
winter-spring. As predicted by linear wave theory, this faster wind change enhances the central122
Pacific SST response by ∼25-50% relative to the sole IOB influence.123
The analysis of Izumo et al. (2010) was based on only ∼30 years of data. A later study (Izumo124
et al. 2014) explored the robustness of the delayed IOD-ENSO relationship over the 1872-2008125
period using IOD and ENSO indices based on an optimal combination of sparse historical obser-126
vations, and an efficient proxy of WWV interannual variations based on the temporal integral of127
Pacific zonal wind stress. They found that a linear hindcast model based on both DMI (Dipole128
Mode Index, see section 2) and WWV proxies in SON can explain ∼50% of the variance of the129
ENSO peak 14 months in advance, with significant contributions from both IOD and WWV over130
the historical period.131
Uncertainties however remain about the mechanism proposed by Izumo et al. (2010) as the132
relationship between the IOD and ENSO-independent wind anomalies in the western Pacific is133
7
not robust across all wind products (Dayan et al. 2014). In addition, the significance tests used134
by Izumo et al. (2010) did not account for the biennial tendency of the IOD and the tendency of135
the IOD to co-occur with ENSO. Taken together these tendencies could partly explain why IODs136
precede ENSOs without the need for a physical mechanism, i.e., ENSO often induces a positive137
IOD and transition to La Ni˜
na the following year, hence resulting in a lead relation between a138
positive IOD and La Ni˜
na that is not necessarily associated with a causality. In addition, very little139
work has been done on the transition asymmetry of the delayed IOD-ENSO relationship despite140
the existence of asymmetries between the positive and negative phases of both ENSO and IOD.141
While oceanic and atmospheric observational data have a good accuracy and spatial sampling142
since the beginning of the satellite era, they do not resolve the main patterns of the IOD or ENSO143
very well before the 1980s (e.g. Izumo et al. 2014). On the other hand, the fifth Coupled Model144
Intercomparison Project (CMIP5) offers a large database in which IOD-ENSO relationships can145
be tested based on long temporal series. Many CMIP5 models produce reasonable ENSO ampli-146
tude and spatial SST patterns (Kim and Yu 2012; Bellenger et al. 2014; Taschetto et al. 2014).147
The CMIP5 models also capture the IOD variability, despite a tendency to overestimate its ampli-148
tude (Cai and Cowan 2013). Models from the previous generation (CMIP3) had similar skills to149
CMIP5 models in terms of ENSO and IOD characteristics (Bellenger et al. 2014; Cai and Cowan150
2013); most of them did not capture the transition/duration asymmetry of ENSO (Ohba et al.151
2010). Despite the IOD and ENSO biases in individual models, the CMIP5 ensemble provides an152
opportunity to sample the diversity of ENSO and IOD behaviors.153
In this paper, we use 23 155-year CMIP5 simulations to assess the robustness of the IOD-ENSO154
relationships, the asymmetry of their temporal sequences, as well as to question the physical or155
statistical nature of their origin. After a brief presentation of the datasets and methods (section 2),156
we show that many characteristics of ENSO, IOD and of their synchronous relationship are cor-157
8
rectly represented in a majority of the 23 CMIP5 models analyzed in this study (section 3). In158
section 4, we show that the CMIP5 delayed IOD-ENSO relationship is also in good agreement159
with observations, and that the use of an IOD index significantly enhances hindcast skill scores of160
the ENSO peak 14 months later. We also demonstrate that the delayed IOD-ENSO relationship161
does not seem to be just a statistical artifact, and therefore needs to be explained, at least partly, by162
a physical mechanism. Finally, we highlight the importance of ENSO asymmetries for the delayed163
IOD-ENSO relationship.164
2. Datasets and methods165
We use four monthly SST gridded products based on ship and buoy observations:166
•HadSST2 covers the period from 1850 onward at 5◦resolution and there is no interpolation167
to fill grid cells where there are no observations (Rayner et al. 2006).168
•HadISST covers the period from 1870 onward at 1◦resolution and optimal interpolation is169
used to produce a complete spatial coverage (Rayner et al. 2003).170
•The Centennial in-situ Observation-Based SST Estimates (COBE) covers the period from171
1850 onward at 1◦resolution and optimal interpolation is used for gridding purposes (Ishii172
et al. 2005).173
•Version 3b of the Extended Reconstructed Sea Surface Temperature (ERSST) covers the174
period from 1854 onward at a 2◦resolution with a gridding technique based on Empiri-175
cal Orthogonal Function projections (see Smith et al. 2008 about ERSSTv3, but note that176
ERSSTv3b does not include satellite data).177
There are large uncertainties in these four datasets before the 1880s due to lack of observations178
(Yasunaka and Hanawa 2011), so only the period 1890-2012 is considered in this paper.179
9
To estimate WWV over the observed period (see definition in the Introduction), we first use180
the 1980-2012 monthly WWV time series provided by the TAO Project Office, NOAA/PMEL,181
USA (Meinen and McPhaden 2000). In addition, we calculate monthly WWV from the Simple182
Ocean Data Assimilation reanalysis SODA-2.2.4 that covers the period 1871-2008 (Carton and183
Giese 2008). The correlation between September-November (hereafter SON) time series of SODA184
WWV and NOAA-BMRC WWV is 0.97 over the overlapping period (1980-2008), demonstrating185
that the SODA reanalysis is adequate for estimating WWV variability, at least over the recent186
period. The correlation between SODA SON WWV and the SON WWV-proxy derived in Izumo187
et al. (2014) is 0.39 over 1890-2001 and 0.67 over 1980-2001, which is far from perfect knowing188
that both are constrained by the same wind stress product (20CR, Compo et al. 2011). The SODA189
SON WWV compares much better with an ”independent” WWV (correlation of 0.69 over 1890-190
2000, and 0.90 over 1980-2000) calculated from a regional ocean reanalysis covering the tropical191
Pacific and based on a different ocean model, wind forcing, and assimilation scheme than SODA192
(Tang and Deng 2010). All the aforementioned correlations have a p-value lower than 0.01, as193
estimated using effective numbers of degrees of freedom based on auto-correlation calculations.194
Finally, we analyze 23 historical CMIP5 simulations in which greenhouse gas, aerosol and ozone195
concentrations are prescribed to follow observations spanning ∼1850 to 2005 (Taylor et al. 2011).196
The various models used in this paper are listed in Tab. 1. More CMIP5 models are available,197
but the three-dimensional data needed to calculate WWV (which is needed for our study) were198
not available for these models at the time of writing. Some of the models we consider only dif-199
fer by their grid resolution or the parameterization of certain physical processes: for instance,200
IPSL-CM5A-LR and IPSL-CM5A-MR are based on the same code but are run at two different201
resolutions. Some models have a common oceanic component (e.g., ACCESS-1.0 and GFDL-202
CM3) or a common atmospheric component (e.g., GFDL-ESM2M and GFDL-ESM2G; CCSM4203
10
and NorESM1-M; ACCESS-1.0 and HadGEM2-ES), even though the exact version of these com-204
mon components often differs. In this paper, we only consider the first ensemble member of each205
model.206
ENSO is characterized through the NINO34 index (SST anomalies averaged between 5◦S -5◦N207
and 170◦W -120◦W ) and IOD is examined through the DMI index defined by Saji et al. (1999).208
The IOB is characterized through the Tropical Indian Ocean index (TIO, defined as SST anomalies209
averaged between 30◦S -25◦N and 20◦E -120◦E ). Considering both CMIP3 and CMIP5 models,210
Cai and Cowan (2013) have found a cross-model correlation of 0.89 between DMI amplitude and211
the IOD amplitude calculated using the principal component of Indian Ocean SST anomalies. This212
suggests that the use of these simple DMI and NINO34 indices is able to capture with reasonable213
fidelity the covariation of ENSO and IOD despite structural biases in the models. In order to214
isolate interannual variability and exclude interdecadal and lower frequency variability, all indices215
are filtered by removing the Hann-filtered time series using averaging weights of 1.0 in the center216
of the 11-year running window, 0.75 at ±2 years, and 0.25 at ±4 years. Most of the results217
presented in this paper were compared against a simple 7-year window running average (i.e. over218
±3 years) and results were very similar to using the Hann filter.219
Classical techniques like composites or Pearson correlations can be misleading because the in-220
dices are asymmetric (e.g. ENSO exhibits a positive skewness). The meaning of the standard221
deviation thus varies across the models, as well as the proportion of positive and negative phases,222
which makes model intercomparisons complex. Here, we use a different approach: positive phases223
(pIOD and El Ni˜
no) are defined as part of the upper quartile of the index distribution, while neg-224
ative phases (nIOD and La Ni˜
na) are defined as part of the lower quartile. These thresholds cor-225
respond approximately to 60-80% of the standard deviation. This framework allows a fair com-226
parison of phase transitions between models and observations because there are 25% of positive227
11
and negative phases for every dataset. The positive and negative phases are defined for the peak228
months, that is, NDJ for NINO34 and SON for DMI. We use a Monte-Carlo method to estimate if229
a given phase transition (e.g. El Ni ˜
no to La Ni˜
na) is significantly different from a random process.230
To that end, we generate 105synthetic time series by randomly resampling the actual time series231
(with replacement). If, say, less than 10% of the synthetic time series present more transitions232
from an El Ni˜
no to a La Ni˜
na than the actual time series, we consider that the Ni˜
no-Ni˜
na transition233
is significantly different from a random transition at the 90% significance level. For example, 25%234
of El Ni˜
no events would be followed by a La Ni˜
na if the transition were purely random and the235
time series were infinitely long. But because the CMIP5 time series are only 155-year long, the236
Monte-Carlo method indicates that the transition can be considered as non-random at the 90%237
significance level if more than ∼34% or less than ∼16% of El Ni˜
no events are followed by a La238
Ni˜
na.239
3. IOD and ENSO in the CMIP models240
To assess the mechanism proposed by Izumo et al. (2010) in the CMIP5 models, it is important241
to ascertain that these models accurately reproduce IOD and ENSO seasonal phase locking and242
spatial patterns (section 3a). The recharge mechanism (i.e. lead/lag relation between WWV and243
ENSO) and biennial tendency of ENSO are then assessed in section 3b. Finally, the synchronous244
relation of the IOD and ENSO, and its phase asymmetry, is evaluated in section 3c.245
a. Seasonal phase locking and spatial patterns246
The monthly standard deviation of NINO34 and DMI are displayed in Fig. 2 for the 23 CMIP5247
models and observations. As noted in previous studies (Taschetto et al. 2014; Bellenger et al.248
2014), the inter-model spread in NINO34 peak amplitude (in NDJ) is very large, the later being249
12
underestimated by 20-50% in 13 models, and overestimated by 10-50% in 9 models (Fig. 2a).250
In order to emphasize ENSO seasonality rather than its amplitude, Fig. 2c shows the normalized251
amplitude of the NINO34 seasonal cycle. In most models the NINO34 variance correctly peaks252
in austral summer or early spring, except for the IPSL-CM5A-LR and IPSL-CM5A-MR (Fig. 2c).253
The duration of the peak however lasts too long in a number of models, extending into February-254
March (inmcm4, MRI-CGCM3, and the two MPI models) or into September-October (NorESM1-255
M, GISS-E2-H and CNRM-CM5), as already noted by Jourdain et al. (2013) and Taschetto et al.256
(2014). Consequently, the 23-model-mean underestimates NINO34 amplitude in NDJ and over-257
estimates it in AMJ.258
Consistent with Cai and Cowan (2013), we find that 21 out of 23 CMIP5 models overestimate the259
DMI peak amplitude (in SON, Fig. 2b), the two exceptions being GISS-E2-H and MRI-CGCM3.260
The majority of models nonetheless peak in SON. In the following, we retain only 21 CMIP5261
models, excluding IPSL-CM5A-LR and IPSL-CM5A-MR from our analysis because, as noted262
above, they have a highly unrealistic ENSO seasonality (Fig. 2c), and secondary DMI peaks in263
austral spring and winter (Fig. 2d).264
As found in Taschetto et al. (2014) and Jourdain et al. (2013), the 21 CMIP5 models generally265
realistically simulate largest NINO34 SST anomalies in the eastern and central Pacific, but the SST266
anomaly extends too far west and has a too narrow meridional extent (not shown). Consistently267
with SST anomalies, ENSO wind anomalies are underestimated in the central Pacific and overes-268
timated in the far western part of the basin in most models (not shown). The IOB pattern appears269
in SST regressions on NINO34 (not shown). As found by Cai and Cowan (2013), the two poles of270
the IOD display a realistic location in the 21 CMIP5 models, but the IOD signal is stronger than271
observed (not shown).272
13
b. ENSO sequences273
We first assess the relationship between equatorial Pacific warm water build-up and ENSO across274
CMIP5 models. The selected CMIP5 models all simulate a warm water recharge-discharge asso-275
ciated with ENSO: the correlation between MAM WWV and the following NDJ NINO34 ranges276
between 0.45 and 0.75 for the 21 models versus 0.70 for NOAA-BMRC over 1980-2010 and 0.60277
for SODA over 1890-2008. All CMIP5 models and observational datasets also exhibit a statisti-278
cally significant negative correlation between the peak of ENSO and the WWV 2-3 seasons after279
(Fig. 3a), indicative of a WWV discharge (or recharge for La Ni ˜
na events) starting at the ENSO280
peak. For a few models, this WWV discharge however occurs too early, in particular for GISS-281
E2-H and ACCESS-1.0. A few other models (e.g. MIROC5) also simulate an unrealistic warm282
water build-up more than a year before the ENSO peak (Fig. 3a). Overall, the multi-model mean283
however agrees well with the observations.284
We now further investigate the biennial tendency of ENSO. The CMIP5 multi-model ENSO285
autocorrelation exhibits a remarkable agreement with observations, with a tendency of ENSO to286
flip its phase every year (Fig. 3b). Despite this overall agreement, a few individual models fail287
to reproduce this biennial behavior (MIROC models, MPI-ESM models and HadGEM2-ES). In288
addition, some CMIP5 models display an auto-correlation after the peak that remains significant289
a few months longer than in the observations, an indication that ENSO events tend to have a late290
termination in these models, as previously noted by Leloup et al. (2008) for the CMIP3 mod-291
els. This linear technique based on auto-correlation does not allow an assessment of the transition292
asymmetries between El Ni˜
no and La Ni˜
na. To that end, we calculate the percentage of El Ni˜
no293
events followed by a La Ni ˜
na, and vice-versa. Observations and CMIP5 multi-model ensemble294
(with confidence exceeding the p=0.10 level) show that the percentage of El Ni˜
nos followed by295
14
La Ni˜
nas is significantly larger than what would be expected from a random transition (Fig. 4a).296
On the other hand, the proportion of La Ni˜
nas followed by El Ni˜
nos is not significantly distin-297
guishable from a random sequence in observations and the CMIP5 multi-model mean (Fig. 4a).298
This asymmetry in ENSO transition is also captured by most CMIP5 models: 19 out of 21 models299
(FGOALS-s2 and IPSL-CM5B-LR excepted) show more transitions from El Ni˜
no to La Ni˜
na than300
from La Ni˜
na to El Ni˜
no (Fig. 4). The CMIP5 multi-model mean (with confidence exceeding the301
p=0.10 level) and a majority of the models nonetheless tend to overestimate the occurrence of El302
Ni˜
no to La Ni˜
na transitions. The relatively good representation of the asymmetry in ENSO phase303
transitions in CMIP5 may indicate some improvement from CMIP3 for which Ohba et al. (2010)304
reported significant biases.305
c. Synchronous IOD-ENSO relationship306
In this subsection, we investigate the ability of CMIP5 models to simulate the tendency for pIOD307
(nIOD) to co-occur with El Ni˜
no (La Ni˜
na). As shown in Fig. 5, 19 out of 21 models produce308
a significant synchronous correlation between DMI and NINO34 (exceptions are MIROC5 and309
INMCM4). For a majority of models, this correlation ranges within ±0.15 of the observed one.310
For the observations, this positive correlation starts being significant around May during the early311
development of the IOD event. By contrast, the positive correlation starts being significant earlier312
in most CMIP5 models, and half of them simulate an unrealistically high correlation between SON313
DMI and the previous year’s NDJ NINO34. For a few models (e.g. MIROC and MPI models), this314
feature is likely related to the overestimated persistence of ENSO events until the following SON315
(Fig. 3b). The inter-model variations in the strength of the synchronous IOD-ENSO relationship316
is significantly related to the inter-model variations in ENSO amplitude (correlation of 0.56, not317
shown), as well as to the inter-model variations in IOD amplitude (correlation of 0.46). There is318
15
also a significant cross-model correlation between the amplitude of ENSO and the amplitude of319
IOD (correlation of 0.62, not shown) likely due to the similar mechanisms involved in both modes320
of variability(Liu et al. 2014; McPhaden and Nagura 2014) or to the fact that IOD can be triggered321
by ENSO.322
We now assess asymmetries in the synchronous IOD-ENSO relationship. If the DMI and323
NINO34 were randomly distributed, 25% of pIODs (defined as upper quartile DMI events) would324
occur synchronously with an El Ni˜
no event, and 25% with a La Ni˜
na event. Instead, 57 to 68% of325
observed pIODs are associated with an El Ni˜
no event in the observations (Fig. 4). Two third of the326
models and the multi-model mean (with confidence exceeding the p=0.10 level) underestimate the327
occurrence of pIOD to El Ni˜
no transitions. The later is nonetheless significantly larger than what328
would be expected from a random transition (as is the nIOD to La Ni˜
na transition). The observed329
proportion of nIODs associated with a La Ni˜
na event is 38 to 54%, i.e. weaker than the opposite330
transition, but still significantly higher than what a purely random processes would produce. The331
multi-model ensemble reproduces this asymmetry along with a large majority of the 21 models332
(Fig. 4b): only three models exhibit an opposite asymmetry (most pronounced in the GISS-E2-H).333
4. IOD as an early precursor of ENSO334
We now assess the ability of CMIP5 models to simulate the tendency for IOD events to be335
followed by an ENSO event in the following year (Izumo et al. 2010, 2014). We then discuss336
whether the delayed IOD-ENSO relationship can be explained by a statistical artifact arising from337
the intrinsic properties of ENSO sequences and the synchronous IOD-ENSO relationship. Finally,338
we investigate the possibility of a physical mechanism.339
16
a. Statistical relationship and transition asymmetry340
All CMIP5 models except MIROC5 simulate a significant negative correlation between DMI341
in SON and NINO34 14 months later, as in the observations (Fig. 5) and as previously described342
by Izumo et al. (2010). Correlation coefficients are slightly weaker than observed in two thirds343
of CMIP5 models, but reach -0.50 in six models (versus -0.40 in the observations). Contrary to344
the synchronous correlation, the inter-model variations in CMIP5 delayed correlations are neither345
related to ENSO (r=-0.15) nor to IOD (r=0.04) amplitude, and there is no relationship between the346
synchronous and delayed correlations across the models (r=0.00). There is a transition asymmetry347
in this delayed IOD-ENSO relationship: the delayed pIOD-Ni ˜
na transitions are more frequent than348
the delayed nIOD-Ni˜
no transitions in the observations (43-46% vs 23-38%, Fig. 4c). The multi-349
model mean and two thirds of the individual models are able to capture this asymmetry (44±1% vs350
38±1% for the multi-model mean, Fig. 4c). The delayed pIOD-Ni˜
na transitions are significantly351
more frequent (at the 90% level) than what would be expected from a random process in all the352
observational products, in the multi-model mean and in 19 out of 21 individual models. The353
delayed nIOD-Ni˜
no transition is above this limit for one observational product only (HadiSST),354
for 18 models, and for the multi-model mean.355
In order to assess the potential influence of IOD events on ENSO one year later, we compare356
14-month lead linear statistical ENSO hindcasts based on WWV only to hindcasts based on WWV357
and DMI. Our approach is similar to Izumo et al. (2010) except that we use a more challenging358
cross-validation method, leaving 50% of the dataset to ”train” the prediction scheme and keeping359
50% for its evaluation (versus ”leave one out” in Izumo et al. 2010). The training/evaluation360
partition is randomly resampled 105times to achieve statistical robustness. For each resampled361
set, we calculate the hindcast correlation skill score for one (WWV) or two (WWV and DMI)362
17
predictors at various lags, and the results are expressed in terms of enhanced explained NINO34363
variance (Fig. 6a). The explained NINO34 variance in SODA is increased by ∼40% when DMI364
is included as a predictor over the period 1980-2008. This is is a strong increase, even though this365
is slightly lower than the ∼60% found by Izumo et al. (2010, their Fig. 1), mostly because we366
use a more challenging cross-validation method (not shown). The increase in explained variance367
is much weaker over the period 1890-2008, reaching only ∼20%, which could be due to inter-368
decadal variability or to the fact that SODA is poorly constrained by observations over the late369
19thand early 20th centuries. The CMIP5 multi-model mean correlation skills are similar to the370
SODA correlation skills over 1890-2008 when both WWV and DMI are used (not shown). The371
inclusion of DMI as predictor of ENSO increases the explained NINO34 variance by 12% in the372
multi-model mean, and by up to 25% in HadCM3 and GFDL-CM3 (Fig. 6a). The inclusion of373
DMI as a predictor of the following year’s ENSO significantly increases the NINO34 explained374
variance at the 90% level in 16 out of 21 CMIP5 models (see legend of Fig. 6). The enhanced375
predictability resulting from the inclusion of TIO (IOB index) is comparatively lower (Fig. 6b),376
which is likely due to the strong correlation between IOB and ENSO. This will be addressed in377
the discussion section.378
We now analyze whether there is an asymmetry in the enhanced ENSO predictability resulting379
from the inclusion of DMI as a predictor. We consider that a hindcast is successful when the pre-380
dicted phase corresponds to the actual phase. In the SODA reanalysis over the period 1890-2008381
and in the CMIP5 ensemble, 50% of the linear hindcasts from WWV and DMI are successful382
for both El Ni˜
no and La Ni˜
na (Fig. 7a). This percentage is much larger than the 25% chance383
of success that would arise from a random process. To further assess the uncertainty on these384
proportions for the observations, we have accounted for an observational uncertainty of ±0.2 K385
on both DMI and NINO34 (http://stateoftheocean.osmc.noaa.gov/sur/ind/dmi.php386
18
and http://stateoftheocean.osmc.noaa.gov/sur/pac/nino34.php) by re-calculating 105
387
times the percentage of successful predictions but adding random perturbations in the range388
±0.2 K to both DMI and NINO34. This calculation gives the 90% confidence intervals for the two389
proportions: [42 −58]% of successful El Ni˜
no predictions and [42 −54]% of successful La Ni˜
na390
predictions. Hence, an asymmetry leading to a difference of 16% or less between the proportion391
of successful El Ni˜
no and the proportion of successful La Ni˜
na predictions would be undetectable392
over the observed period because of observational uncertainties and short observational records.393
The CMIP5 ensemble exhibits a very small asymmetry in success rate, with 51.7% successful La394
Ni˜
na hindcasts vs 48.6% successful El Ni˜
no hindcasts (Fig. 7b). Only four individual models have395
a higher success rate for El Ni˜
no hindcasts than for La Ni˜
na hindcasts (HadGEM2-ES, inmcm4,396
MPI-ESM-LR and MRI-CGCM3; not shown). Among the 17 other models with a higher suc-397
cess rate for La Ni˜
na hindcasts, the difference in success rate between Ni˜
no and Ni˜
na hindcasts398
is smaller than 4% in five models, and smaller than 10% in nine of the 12 other models. Over-399
all, prediction of La Ni˜
na phases appears slightly more successful than the prediction of El Ni˜
no400
phases in the CMIP5 models, but the difference is so small that accounting for this asymmetry is401
probably useless in an operational context.402
b. First approach to test the IOD influence on the following year’s ENSO: probabilities403
We now question whether the delayed IOD-ENSO relationship derived from our analysis could404
purely be a statistical artifact, i.e. unrelated to a physical process linking the IOD and ENSO one405
year later. For instance, an El Ni˜
no tends to synchronously induce a pIOD in the Indian Ocean, but406
also tends to be followed by a La Ni˜
na: this results in a tendency for pIOD events to precede a La407
Ni˜
na event without the need of an actual influence of the IOD on following year’s ENSO. Our null408
hypothesis is hence that delayed relationships between IOD and ENSO can entirely arise from the409
19
intrinsic properties of ENSO temporal sequences and the synchronous ENSO-IOD relationship,410
without the need to invoke a physical mechanism for the delayed IOD-ENSO relationship. If the411
null hypothesis were true, the proportion Pof pIODs followed by a La Ni˜
na 14 months later and412
the proportion of nIODs followed by an El Ni˜
no 14 months later would be given by:413
414
PpIODy0→ninay1=PpIODy0→ninay0.Pninay0→ninay1
+PpIODy0→neuty0.Pneuty0→ninay1
+PpIODy0→ninoy0.Pninoy0→ninay1
(1)
and:415
416
PnIODy0→ninoy1=PnIODy0→ninay0.Pninay0→ninoy1
+PnIODy0→neuty0.Pneuty0→ninoy1
+PnIODy0→ninoy0.Pninoy0→ninoy1
(2)
where y0 and y1 denote year 0 and year 1 and, for example, P(pIODy0→ninay1)is the prob-417
ability for a pIOD to be followed by a delayed La Ni˜
na (i.e. peaking 14 months later), whereas418
P(pIODy0→ninay0)is the probability for a pIOD to be followed by a synchronous La Ni˜
na (i.e.419
peaking 2-3 months later).420
Let us first examine the case of pIOD transitions to a La Ni˜
na in the following year (Eq. 1). Our421
null hypothesis is tested in Fig. 8a showing the scatter plot of the actual probability for a delayed422
pIOD-Ni˜
na transition vs the probability calculated from the right hand side in Eq. 1. The actual423
probability is greater than the one derived from Eq. 1 for all the observational products and 20 out424
of 21 CMIP5 models. This is a clear indication that the tendency for pIODs to co-occur with an425
20
El Ni˜
no and for El Ni˜
no events to be followed by a La Ni ˜
na cannot alone explain the tendency426
for pIOD to precede La Ni˜
na 14 months later. This means that we can reject the null hypothesis427
for the delayed pIOD-Ni˜
na transition in both the observations and the CMIP5 models. Let us now428
examine the opposite transition, i.e. from a nIOD to an El Ni ˜
no in the following year (Eq. 2). The429
actual probability is greater than the one derived from Eq. 2 for 18 out of 21 CMIP5 models, but430
there is no agreement across the observational products (Fig. 8a). This means that we can reject431
the null hypothesis for the delayed nIOD-Ni˜
no transition in most CMIP5 models but not in the432
observations.433
Another interesting feature in Fig. 8a is that the actual probabilities of delayed pIOD-La Ni˜
na434
transitions are highly correlated to the ones expected from Eq. 1-2 across the CMIP5 models435
(r=0.85). This indicates that even if the null hypothesis is rejected, a significant part of the de-436
layed pIOD-Ni˜
na relationship actually results from intrinsic properties of ENSO sequences and437
the synchronous IOD-ENSO relationship. Further investigations of the contribution of the vari-438
ous terms in Eq.1 indicate that this high correlation is tightly related to the high probability of El439
Ni˜
nos preceding La Ni˜
nas combined with the high probability of pIOD co-occurring with El Ni˜
no440
(third line of Eq. 1, not shown). The delayed nIOD-Ni˜
no relationship exhibits a different behav-441
ior in the sense that the actual probability is not highly correlated to the RHS in Eq. 2 (Fig. 8a).442
This is mostly because nIOD tends to co-occur with La Ni˜
na, but only a small proportion of La443
Ni˜
na events are followed by El Ni ˜
no, so most delayed nIOD-Ni˜
no transitions occur when nIOD444
co-occurs with a neutral ENSO event (not shown).445
c. Second approach to test the IOD influence on the following year’s ENSO: synthetic time series446
To test the robustness of the conclusions derived from the previous analysis an alternative447
method is examined using synthetic time series instead of phase probabilities. Our null hypothesis448
21
is again that the delayed relationships between the IOD and ENSO can arise from the intrinsic449
properties of ENSO sequences and a synchronous linear ENSO-IOD relationship, without the450
need for a physical mechanism in the delayed IOD-ENSO relationship. Here we also assume451
the synchronous IOD-ENSO relationship is linear. To test our null hypothesis, we extract the452
residuals ε(y)from the following linear fit to NINO34 time series from individual CMIP5 models453
or observations:454
455
DMI(y) = ˆ
kNINO34(y) + ε(y)(3)
where DMI(y)is the SON DMI time series, NINO34(y)the NDJ NINO34 time series (of the456
same year), and ˆ
ka constant parameter obtained from a least-mean-square calculation for each457
observed or CMIP5 dataset. Then we assume that residuals εresult from random processes, and458
we randomly re-sample εto build 104synthetic DMI time series:459
460
DMIsynth(y) = ˆ
kNINO34(y) + ε(r)(4)
where ris a random index.461
To investigate whether Eq. 3 can account for the statistics of the delayed IOD-ENSO relationship462
in the individual CMIP5 models and observations, we first calculate the proportions of delayed463
pIOD-Ni˜
na and nIOD-Ni˜
no in every synthetic time series. Then, we calculate the probability464
(across the 104time series) of obtaining a smaller proportion of delayed IOD-ENSO transitions465
than the actual one. This calculation gives the statistical significance for a rejection of the null466
hypothesis. For example, if 90% of the synthetic proportions are smaller than the actual one, we467
22
consider that the null hypothesis for this transition can be rejected at the 90% confidence level.468
This second method provides a better quantification of the results than the first method, but relies469
on an assumption of linearity (i.e. no asymmetry in synchronous IOD-ENSO transitions). Note470
that very similar results to the ones presented below are obtained if the model of Eq. 3 also accounts471
for intrinsic biennial tendency of the DMI index (i.e. a term ˆ
k2DMI(y−1)is added in the right472
hand side of Eq. 3-4, not shown). We also obtained very similar results when accounting for first-473
order non-linearity by fitting DMI on ˆ
k1NINO34(y)+ ˆ
k2NINO34(y)2instead of ˆ
kNINO34(y)(not474
shown).475
For the observational products, we find that the null hypothesis is rejected at the 99% confidence476
level for the delayed pIOD-Ni˜
na transition, which means that the observed transition occurs more477
often than in 99% of the synthetic time series (Fig. 8b). By contrast, the null hypothesis cannot478
be rejected for the observed delayed nIOD-Ni˜
no transition: the statistical confidence level for the479
rejection is less than 75%, i.e. the observed transition occurs less often than in more than 25% of480
the synthetic time (Fig. 8b). These observation-based results are consistent with those obtained481
in the previous subsection and suggest that there might be a physical mechanism amplifying the482
effects of the internal sequential properties of ENSO and the synchronous IOD-ENSO relationship483
on the delayed pIOD-Ni˜
na transitions.484
We also find that the null hypothesis is rejected at the 90% confidence level in 15 CMIP5 mod-485
els for the delayed pIOD-Ni˜
na transition, and in 12 models for the delayed nIOD-Ni˜
no transition486
(Fig. 8b). These results suggest that in more than half of the CMIP5 models there might be a phys-487
ical mechanism increasing the proportions of pIOD-Ni˜
na and of nIOD-Ni˜
no delayed transitions,488
but that its influence may be stronger or more common for the former transition. The results for489
CMIP5 are also consistent with the results from the previous subsection: both the orange and blue490
”clouds” in Fig. 8a are above the 1:1 line, and the models for which the null hypothesis cannot be491
23
rejected even at low confidence levels in the second approach are the same models that are close492
to the 1:1 line (see italisized characters in Fig. 8a). Overall, this second method also suggests that493
the intrinsic properties of ENSO sequences and the synchronous IOD-ENSO relationship alone494
are not sufficient to explain the high prevalence of delayed IOD-ENSO transitions. Hence this495
strongly suggests that a physical mechanism may influence the transition from IOD to ENSO in496
the following year.497
d. Physical mechanism498
We now investigate whether the physical mechanism proposed by Izumo et al. (2010) to explain499
the influence of the IOD on the following year’s ENSO is evident in the CMIP5 models. As this500
mechanism was originally based on the analysis of a relatively short time period (1981-2009),501
we first reproduce their analysis (their Fig. 2) using SODA and 20CR over the entire 1890-2008502
period and over 1980-2008 (Fig. 9c,d). In this analysis, the delayed response to the IOD is ob-503
tained by lagged multiple linear regression of a selection of variables onto DMI and NINO34 of504
year 0, and the signals induced by synchronous ENSO is removed by only considering the regres-505
sion coefficient related to DMI (see figure caption). By this procedure, we also removed most of506
the signature of the IOB that co-occurs with the peak of ENSO at year 0/1 because it is highly507
correlated to NINO34 (see Discussion section). The SST anomalies associated with the develop-508
ing phase of a negative IOD event (June-October year 0) increase easterly winds in the Western509
and Central Equatorial Pacific. Composites of 500hPa vertical velocity indicate that this wind510
anomaly is consistent with an amplification of the Walker circulation that would be caused by511
the eastern pole of the IOD (not shown). In the mechanism proposed by Izumo et al. (2010), the512
easterly wind anomalies force a downwelling Rossby wave that contributes to the build-up of heat513
in the western Pacific (see T300 in Fig. 9c,d), providing an efficient pre-conditioning for an El514
24
Ni˜
no onset (recharged state). They also force an upwelling Kelvin wave that reflects at the eastern515
boundary as an upwelling Rossby wave, favoring eastward currents in the central Pacific a couple516
of months later (delayed advective-reflective feedback; Picaut et al. 1997) and hence positive SST517
anomalies that can favor the onset of an El Ni˜
no. As shown by Izumo et al. (2015), changes in518
western Pacific zonal wind stress from boreal summer-fall to winter-spring matter as much as the519
actual stress anomaly in winter-spring for the development of spring SST anomalies in the central520
Pacific. The occurrence of an IOD always enhances the summer-fall to winter-spring change in521
western Pacific zonal wind stress, as explained below for the case of a nIOD (either ”pure” or522
associated with a basin-wide cooling). In the case of a pure nIOD event (i.e. no IOB as in Fig. 9),523
the eastern IOD pole disappearance in late November of year 0 induces a relaxation of the easterly524
anomaly in the western Pacific. This favors positive SST anomaly at the beginning of year 1 in525
the central Pacific (Fig. 9c,d). In the case of a nIOD co-occuring with the development of an IOB526
cooling, the western pole (a signature associated with the IOB) and eastern poles of the IOD tend527
to have compensating effects on winds in the western Pacific (Annamalai et al. 2005). In such a528
configuration, the disappearance of the IOD eastern pole in late November of year 0 enables the529
development of a westerly wind anomaly in the western Pacific, which also favors positive SST530
anomaly in the central Pacific at the beginning of year 1 (Izumo et al. 2015). The negative T300531
anomaly that appears in the Eastern Pacific during the easterly wind anomaly, with signs of east-532
ward propagation and followed by positive heat content anomalies in late winter and spring (at the533
beginning of year 1), is qualitatively consistent with the equatorial wave response described by534
Izumo et al. (2010, 2015).535
To assess the aforementioned mechanism in the CMIP5 models, we separately consider the536
CMIP5 models that clearly show a delayed IOD-ENSO relationship and the models that do not,537
i.e. the upper and lower terciles (seven models in each tercile) in terms of enhanced predictability538
25
resulting from the inclusion of DMI as a predictor of the following year’s ENSO (according to539
Fig. 6). Figure 9 shows that the models displaying the strongest IOD influence on the following540
year’s ENSO also display signals that are consistent with the observations in terms of upper ocean541
heat content (see T300 in Fig. 9a,b). By contrast, anomalies are weaker in CMIP5 models with542
a weak IOD influence on the following year’s ENSO (Fig. 9b). The heat content in these mod-543
els clearly displays a warm-water build-up in the Western Pacific during the IOD event, as well544
as a deepening of the thermocline in the central Pacific shortly after the end of the IOD-induced545
easterly anomalies in the central Pacific. This easterly wind anomaly in the western Pacific how-546
ever begins much earlier in the CMIP5 model mean than in SODA: the CMIP5 multi-model mean547
easterly anomaly is already well developed during the build up of the IOD event and decays at548
its peak, whereas the SODA easterly anomaly grows synchronously with the IOD event. These549
decaying easterlies during the development of the IOD event coincide with significant negative550
SST anomalies in the Pacific. Moreover, the western Pacific wind anomaly in DJF of year 0/1551
also displays a different behavior in CMIP5 models and observations. The shift from easterly to552
westerly anomalies occurs later in the CMIP5 models (around January-February; Fig. 9a,b) than in553
the observations (around October-November; Fig. 9c). In summary, despite realistic relationship554
between the IOD and the following year’s ENSO, the physical mechanism of this interaction re-555
mains elusive in the CMIP5 models, and further work (e.g. investigating the mechanisms for each556
CMIP model separately) will be needed to identify other processes that may affect the IOD-ENSO557
interactions.558
To investigate asymmetries in the mechanism described in Fig. 9, we have built composites559
anomalies by subtracting composites of each ENSO phase (La Ni˜
na, neutral, El Ni˜
no) from com-560
posites of pIOD co-occurring with these ENSO phases (pIOD-Ni˜
na, pIOD-neutral and pIOD-Ni˜
no561
respectively), and similarly for nIOD co-occurring with these ENSO phases. The results (not562
26
shown) corroborate the linear regression analysis above and do not suggest a strong asymmetry563
between pIOD and nIOD related mechanisms (the signatures of ENSO are only ∼15% stronger564
on the years following a pIOD than on the years following a nIOD).565
5. Discussion566
This paper brings further evidence that observed pIODs (nIODs) tend to be followed by La Ni˜
na567
(El Ni˜
no) events 14 months later. These delayed sequences are also present in a majority of the568
21 CMIP5 models examined in this paper. Two statistical tests developed in this paper show that a569
physical mechanism is needed to explain the delayed pIOD-Ni˜
na transition in the observations and570
in most CMIP5 models, although the ENSO effect on the synchronous IOD and tendency of La571
Ni˜
nas to follow El Ni˜
nos clearly favors this transition. This physical mechanism however remains572
elusive in the CMIP5 models.573
a. Discussion on the physical mechanism in the observations574
The 1890-2008 IOD-induced wind anomaly over the Pacific starts 3-4 months earlier and is 2-3575
times weaker than in Izumo et al. (2010). This may be related to the use of 20CR in our analysis576
(instead of NCEP2 in Izumo et al. 2010) because we note similar differences over 1980-2008 as577
over 1890-2008 in Fig. 9c,d. The warm water build-up in SODA and the associated SST anomalies578
over 1890-2008 are approximately half the magnitude of those over 1980-2008 (which are similar579
to those found by Izumo et al. over a similar period). It is difficult to know whether this smaller580
amplitude over the entire period arises from possible biases in the first half of the SODA and 20CR581
reanalyses or from interdecadal variability in the delayed IOD-ENSO relationship (Izumo et al.582
2014 have suggested that the IOD influence was weaker during the mid-20thcentury). Nonetheless,583
27
our analysis suggests that the mechanism proposed by Izumo et al. (2010) may operate over the584
entire 1890-2008 period, but probably with significant interdecadal variations.585
Despite some consistency with Izumo et al. (2010) mechanism, the exact timing of the develop-586
ment of westerly anomalies over the Pacific (Fig. 9c,d at the end of year 0) is puzzling. Indeed,587
the transition from easterly to westerly anomaly occurs in October-November, slightly before the588
end of the Indian Ocean warm anomaly. These westerly anomalies become well developed over589
the western Pacific approximately at the same time as Pacific positive SST anomalies around the590
dateline, i.e. in December-January. As shown in Fig. 9c,d, these westerly anomalies in winter591
are not related to any significant SST signal in the Indian Ocean (such as IOB). Nonetheless, it592
is possible that the development of westerly anomalies occurs through a physical mechanism that593
is more complex than previously described. Another source of uncertainty is of course the linear594
nature of the statistical analyses such as that of figure 11, which may not properly capture, e.g.,595
asymmetries between positive and negative events or thresholds in the convection response to SST596
anomalies.597
b. Discussion on the difference between observations and CMIP5598
It is important to question whether the physical mechanism operating in CMIP5 models and599
observations are the same. The systematic overestimation of the IOD amplitude in CMIP5 models600
(Fig. 2) may indeed imply an overly strong atmospheric response and therefore an overestima-601
tion of the asynchronous IOD-ENSO relationship in CMIP5. The overly strong IOD in CMIP5602
might for example explain why a physical mechanism is needed to explain the delayed nIOD-603
Ni˜
no transition in a majority of CMIP5 models but not in some observational products (section604
4c). However, the absence of a significant correlation between the DMI amplitude and the 14-605
month lagged DMI-NINO34 correlation across the CMIP5 models (r=0.04, not shown) suggests606
28
that this is not the case. This non-significant correlation could be related to the fact that the western607
Pacific wind sensitivity to the IOD amplitude varies across models, e.g. due to different convec-608
tion schemes and SST mean states (that affects the threshold for convection). In addition, even if609
the wind response to the IOD amplitude were the same in every models, Bellenger et al. (2014)610
have shown that there is a large model dispersion in terms of Bjerkness feedback and therefore611
in ENSO response to wind anomalies. Similarly, the probabilities of randomly obtaining each612
delayed transition (Fig. 8) are not significantly rank-correlated to the DMI amplitude (r=0.25 and613
r=-0.05 for pIOD-Ni˜
na and nIOD-Ni˜
no respectively). This suggests that the overestimation of the614
IOD amplitude is not the main reason for the prevalence of a delayed IOD-ENSO relationship in615
the CMIP5 models.616
The second major concern about the mechanism in CMIP5 models is that there are differences in617
the details of the timing of IOD-induced wind anomalies compared to observations. In section 4,618
we have reported that the easterly wind anomaly in the western Pacific begins much earlier in the619
CMIP5 models than in SODA. Going back to Fig. 5, we can see that approximately half of the620
CMIP5 models display a moderate correlation between DMI and the previous year’s NINO34,621
which is not found in observations. As suggested in section 3, this could be partially related to622
ENSO events lasting too long in some models. Using a similar methodology as in Fig. 11, we623
removed the linear influence of the previous years ENSO (either in NDy−1Jy0or in MJJy0). The624
pattern is almost unchanged compared to Fig. 9, although weakened for the top CMIP5 models625
(not shown). This negative SST anomaly does not appear to emerge from a previous ENSO, but626
peaks in June. We have have found that this SST anomaly co-occurs with decreased convection627
/ enhanced subsidence in the western-central equatorial Pacific and increased convection over the628
maritime continent (hence an amplification of the Walker circulation considering the cold tongue629
bias of CMIP5 models) (not shown) but its origin remains unclear and understanding its effects on630
29
the winds would require specific modeling work. Another difference between the observations and631
the CMIP5 models is that the shift from easterly to westerly wind anomaly in the Western Pacific632
occurs later in the CMIP5 models than in the observations. This bias is possibly partly related to633
the IOD eastern pole SST anomaly lasting longer in CMIP5 models (until ∼December-January)634
than in observations (until ∼November).635
Further analyses will be necessary to understand whether these differences lead to fundamen-636
tally distinct mechanisms. A possible strategy would be to select the CMIP5 models showing the637
best agreement with observations in terms of Indo-Pacific relationship, then to perform ensemble638
experiments from neutral, discharged, and recharged Pacific WWV, with either a climatological639
SST, pure IOD, or pure IOB pattern specified in the Indian Ocean.640
Finally, it has been shown that even though a physical mechanism is required to explain the641
delayed IOD-ENSO relationship in the CMIP5 models, a part of this relationship is favored by the642
intrinsic properties of ENSO temporal sequences and the synchronous ENSO-IOD relationship643
(section 4b,c). From this point of view, the overestimated occurrence of El Ni˜
no to La Ni˜
na644
transitions in the CMIP5 models (Fig. 4a) tends to favor the occurrence of delayed pIOD to La645
Ni˜
na transitions. However, the underestimated occurrence of synchronous pIOD to El Ni˜
no in the646
CMIP5 models (Fig. 4b) has the opposite effect on the occurrence of delayed pIOD to La Ni˜
na647
transitions. These different statistical properties between CMIP5 models and observations further648
complicate the comparison between the physical mechanism in the observations and in the CMIP5649
models.650
c. Limitations of this study651
The present study has only considered the role of IOD on ENSO transitions. Nonetheless, we652
would like to point out that the present study does not rule out the influence of IOB on the Pacific653
30
evolution. Indeed, this influence has been shown to be important for ENSO transitions (Kug and654
Kang 2006; Ohba and Ueda 2007, 2009a; Kug and Ham 2012; Santoso et al. 2012; Ohba and655
Watanabe 2012; Dayan et al. 2015). However, it is very difficult to isolate the role of ENSO from656
the role of IOB on the following year’s ENSO using simple statistical techniques. Indeed, the657
co-variance between synchronous NDJ NINO34 and JF TIO is 60% in the observations and 45%658
in the CMIP5 multi-model mean, while the covariance between SON DMI and the following JF659
TIO is only ∼25% in both the observations and the CMIP5 multi-model mean (not shown), and660
the covariance between SON DMI and NINO34 of the same year is also ∼25%. This supports the661
idea that the IOB can to a large extent be considered as an intrinsic part of ENSO while the IOD is662
more independent from both ENSO and the IOB. This is the assumption that has been made along663
the previous subsections. In other words, what we have previously called ”intrinsic properties of664
ENSO temporal sequence” (in particular in the two statistical approaches used in section 4) is not665
purely intrinsic (in the sense of its usual description limited to the Pacific) but may partly result666
from interactions with the IOB.667
Due to the high covariance of synchronous NINO34 and TIO, it is very difficult to distinguish the668
roles of ENSO and IOB in our methodologies. We have attempted to remove the influence of IOB669
in the composites of Fig. 9 (i.e. calculating the multi-linear regression Y=ag
DMI +b
^
NINO34 +670
cg
T IO, then plotting a). However, the multi-linear regression appears to be very similar to the one671
plotted in Fig. 9 (not shown). Again, we do not claim that IOB has no influence on the following672
year’s ENSO, but as NINO34 and TIO are highly correlated, removing the NINO34 signal already673
removes a large part of the TIO signal. To properly compare the roles of IOD and IOB on the674
following year’s ENSO, a sensible method would probably be to run partially coupled experiments675
that are specifically designed to answer this question. Such dedicated coupled experiments, in676
the CMIP models that reproduce the ENSO-IOD relations well, would also probably allow to677
31
investigate the mechanisms of the Indian Ocean influence of ENSO beyond what could be achieved678
through the linear statistical analyses of the present paper. However, this is beyond the scope of679
this paper.680
Beyond the role of IOB, a further limitation of our study is that we have only analyzed the phys-681
ical mechanisms in terms of an ”atmospheric bridge” as in Izumo et al. (2010). Some authors682
have suggested that a mechanism through the Indonesian Throughflow (”oceanic bridge”) could683
better explain the delayed IOD-ENSO relationship than the atmospheric bridge (Yuan et al. 2011,684
2013) while other studies suggest this oceanic bridge has only a marginal influence (e.g. Clarke685
1991; Schwarzkopf and B¨
oning 2011; Izumo et al. 2015; Kajtar et al. 2015). While we have not686
attempted to examine this oceanic mechanism, we have shown that the atmospheric bridge is effec-687
tive and statistically robust in the CMIP5 models and, to a less robust extent, in the observations.688
d. Implications for ENSO predictions689
Most studies dedicated to dynamical ENSO forecasts based on ocean/atmosphere models have,690
so far, not attempted to extend their forecasts beyond the spring barrier, i.e. 9 months ahead of691
ENSO peak (e.g. Barnston and Tippett 2013; Ham et al. 2014), with a few exceptions (e.g. Luo692
et al. 2008). Our study however suggests that most CMIP5 models do reproduce a contribution693
of Indian Ocean variability to ENSO predictability at leads of up to ∼14 months prior to the694
ENSO peak. We have shown that this additional predictability is not purely a statistical artifact695
arising from, e.g., the tendency of El Ni˜
nos to induce a synchronous pIOD in the Indian Ocean696
and to be followed by a La Ni˜
na. Even if the physical mechanism that explains this increased697
ENSO predictability requires further scrutiny, this ability of many CMIP5 models to reproduce698
the 14-months lead influence of the IOD on the Pacific Ocean is a strong incentive to explore699
extended-range dynamical forecasts of ENSO.700
32
Finally, we would like to emphasize that the IOD has more potential than the IOB in terms of701
ENSO forecast. An ENSO prediction based on TIO (the IOB index, see section 2) can only be702
issued in January-February because TIO is less robust before that season. Hence this would be a703
prediction ∼11 months ahead, while DMI allows predictions ∼14 months ahead. Notwithstanding704
this 3-month forecast delay, we have calculated the increase of ENSO explained variance when705
including JF TIO as a predictor of ENSO in addition to the JF Warm Water Volume (WWV). The706
predictability of ENSO from WWV is much higher for WWV in JF than for WWV in SON (not707
shown). As a result, including JF TIO as a second predictor does not enhance the predictability708
of ENSO as much as SON DMI. Indeed, Including JF TIO as a predictor increases the explained709
NINO34 variance in SODA by less than 9% while including SON DMI increases that variance710
by 20% (Fig. 6). The explained NINO34 variance in the CMIP5 multi-model mean is also much711
weaker for JF TIO as a predictor than for SON DMI (5% vs 12% in Fig. 6), and the increased712
explained NINO34 variance is only significant at the 90% level in 5 models for JF TIO versus 16713
models for SON DMI (see legend of Fig. 6). Thus it appears that the SON IOD has more potential714
than the JF IOB in terms of ENSO forecasts (given that WWV is also used as a predictor).715
6. Conclusion716
In this paper, we have first used extended observational time series and output from 21 CMIP5717
historical model simulations to describe the properties of ENSO sequences and the co-occurrence718
between synchronous IOD and ENSO events. In the observations as in the CMIP5 models, ∼40%719
of El Ni˜
no events are followed by La Ni ˜
na, a significantly larger proportion than that expected from720
a random sequence (25%), while the proportion of La Ni˜
na events followed by El Ni ˜
no (∼25%)721
is not distinguishable from a random distribution at the 90% significance level. Most models722
reproduce the observed tendency of IOD events to co-occur with synchronous ENSO events. There723
33
is also a clear transition asymmetry in this synchronous relationship, whereby the proportion of724
pIODs co-occurring with an El Ni˜
no (∼50-60%) is significantly higher than the proportion of725
nIODs co-occurring with a La Ni˜
na (∼40%) in both the observations and CMIP5 models.726
Then, this paper has further addressed the possible influence of IODs on the following year’s727
ENSO. Nearly all CMIP5 models produce a robust relationship between IOD and the following728
year’s (i.e. 14 months later) ENSO, with a multi-model mean lag-correlation of -0.40 as in the729
observations. Consequently, the predictability of ENSO from a linear combination of DMI and730
the Pacific warm water volume in the previous year is enhanced compared to the predictability731
of ENSO from the Pacific warm water volume only. The enhanced predictability corresponds to732
13% in explained ENSO variance for the CMIP5 multi-model mean and to 22% for the 120-year733
SODA ocean reanalysis. We have also found that the proportion of pIODs followed by a La Ni˜
na734
14 months later (∼45%) is greater than the proportion of nIODs followed by an El Ni˜
no 14 months735
later (∼30-40%) both in the CMIP5 models and in the observations, but the consequences of this736
asymmetry in terms of operational forecasts are very limited.737
Two statistical tests were developed to examine if the delayed IOD-ENSO relationship can arise738
solely from intrinsic properties of the ENSO temporal sequences (i.e. tendency of El Ni˜
nos to739
be followed by La Ni˜
nas) and the synchronous IOD-ENSO relationship, i.e. without the need to740
invoke a physical mechanism whereby an IOD event causes a response in the equatorial Pacific741
that interferes with the ENSO dynamics. The probability of pIODs leading La Ni ˜
na events by742
14 months is significantly larger than what would be expected from ENSO sequences and syn-743
chronous IOD-ENSO relationship in all the observational products and in 15 out of 21 CMIP5744
models. The probability of nIOD events leading El Ni˜
nos by 14 months is also larger that that de-745
rived from the properties of ENSO temporal sequences and of synchronous IOD-ENSO transitions746
in 12 out of 21 CMIP5 models and in one observational product.747
34
The main conclusion of this paper is therefore that a physical mechanism is needed to explain748
the delayed pIOD-Ni˜
na transition in the observations and in most CMIP5 models, even though the749
tendency of El Ni˜
nos to induce synchronous pIODs and to be followed by La Ni˜
nas clearly favors750
this transition. A mechanism is also needed to explain the delayed nIOD-Ni ˜
no relationship in 12751
out of 21 CMIP5 models. But no such mechanism is needed to explain the delayed nIOD-Ni˜
no re-752
lationship in the nine other models, and there is no consensus for this transition in the observations.753
The analyses in this paper are broadly consistent with Izumo et al. (2010, 2015) mechanisms, al-754
though several inconsistencies do not really allow to conclude unambiguously about a common755
mechanism across observational datasets and CMIP models.756
Acknowledgments. This study was conducted in the context of the ARC project DP110100601757
and the Agence Nationale de la Recherche (ANR) project METRO (2010-BLAN-616-01). This758
work was supported by the NCI National Facility at the ANU via the provision of computing759
resources to the ARC Centre of Excellence for Climate System Science. We acknowledge the760
World Climate Research Program’s Working Group on Coupled Modelling, which is responsible761
for CMIP, and we thank the climate modeling groups (Tab. 1) for producing and making their762
model output available. Support for the Twentieth Century Reanalysis Project dataset is provided763
by the U.S. Department of Energy, Office of Science Innovative and Novel Computational Impact764
on Theory and Experiment (DOE INCITE) program, and Office of Biological and Environmental765
Research (BER), and by the National Oceanic and Atmospheric Administration Climate Program766
Office. We also acknowledge the following centers for making their data available: the TAO767
Project Office, NOAA/PMEL for their WWV estimates, the University of Maryland for SODA,768
the UK Met Office for HadSST2 and HadISST, NOAA for ERSST, the Japan Meteorological769
Agency for COBE SST.770
35
References771
An, S.-I., and B. Wang, 2000: Interdecadal Change of the Structure of the ENSO Mode and Its772
Impact on the ENSO Frequency. J. Climate,13 (12), 2044–2055.773
Annamalai, H., R. Murtugudde, J. Potemra, S. P. Xie, P. Liu, and B. Wang, 2003: Coupled dy-774
namics over the Indian Ocean: Spring initiation of the zonal mode. Deep Sea Research Part II:775
Topical Studies in Oceanography,50 (12), 2305–2330.776
Annamalai, H., S. P. Xie, J. P. McCreary, and R. Murtugudde, 2005: Impact of Indian Ocean Sea777
Surface Temperature on Developing El Ni˜
no. J. Climate,18 (2), 302–319.778
Ashok, K., Z. Guan, and T. Yamagata, 2003: A look at the relationship between the ENSO and the779
Indian Ocean Dipole. Journal of the Meteorological Society of Japan,81 (1), 41–56.780
Barnston, A. G., and M. K. Tippett, 2013: Predictions of Nino3.4 SST in CFSv1 and CFSv2: a781
diagnostic comparison. Clim. Dynam.,41 (5-6), 1615–1633.782
Behera, S. K., J. J. Luo, S. Masson, S. A. Rao, H. Sakuma, and T. Yamagata, 2006: A CGCM783
study on the interaction between IOD and ENSO. J. Climate,19 (9), 1688–1705.784
Bellenger, H., E. Guilyardi, J. Leloup, M. Lengaigne, and J. Vialard, 2014: ENSO representation785
in climate models: from CMIP3 to CMIP5. Clim. Dynam.,42 (7-8), 1999–2018.786
Bentsen, M., and Coauthors, 2012: The Norwegian Earth System Model, NorESM1-M-Part 1:787
Description and basic evaluation. Geoscientific Model Development Discussions,5, 2843–2931.788
Bjerknes, J., 1969: Atmospheric teleconnections from the equatorial pacific. Mon. Wea. Rev.,789
97 (3), 163–172.790
36
BOM, 2010: Operational implementation of the ACCESS numerical weather prediction system.791
NMOC Operations Bulletin No.83. Tech. rep., Bureau of Meteorology (BOM). URL http://792
www.bom.gov.au/australia/charts/bulletins/apob83.pdf.793
Bracco, A., F. Kucharski, F. Molteni, W. Hazeleger, and C. Severijns, 2005: Internal and forced794
modes of variability in the Indian Ocean. Geophys. Res. Lett.,32 (12).795
Burgers, G., 1999: The El Ni ˜
no stochastic oscillator. Clim. Dynam.,15 (7), 521–531.796
Burgers, G., and D. B. Stephenson, 1999: The ”normality” of El Ni ˜
no. Geophys. Res. Lett.,26 (8),797
1027–1030.798
Cai, W., and T. Cowan, 2013: Why is the amplitude of the Indian Ocean Dipole overly large in799
CMIP3 and CMIP5 climate models? Geophys. Res. Lett.800
Cai, W., P. van Rensch, T. Cowan, and H. H. Hendon, 2012: An Asymmetry in the IOD and ENSO801
Teleconnection Pathway and Its Impact on Australian Climate. J. Climate,25 (18).802
Carton, J. A., and B. S. Giese, 2008: A reanalysis of ocean climate using Simple Ocean Data803
Assimilation (SODA). Mon. Wea. Rev.,136 (8), 2999–3017.804
Choi, K.-Y., G. A. Vecchi, and A. T. Wittenberg, 2013: ENSO Transition, Duration, and Amplitude805
Asymmetries: Role of the Nonlinear Wind Stress Coupling in a Conceptual Model. J. Climate,806
26 (23).807
Chylek, P., J. Li, M. K. Dubey, M. Wang, and G. Lesins, 2011: Observed and model simulated 20th808
century Arctic temperature variability: Canadian earth system model CanESM2. Atmospheric809
Chemistry and Physics Discussions,11 (8), 22 893–22 907.810
Clarke, A. J., 1991: On the reflection and transmission of low-frequency energy at the irregular811
western Pacific Ocean boundary. J. Geophys. Res.,96 (S01), 3289–3305.812
37
Clarke, A. J., 2014: El Ni ˜
no Physics and El Ni˜
no Predictability. Annual review of marine science,813
6, 79–99.814
Clarke, A. J., and S. van Gorder, 2003: Improving El Ni ˜
no prediction using a space-time integra-815
tion of Indo-Pacific winds and equatorial Pacific upper ocean heat content. Geophys. Res. Lett.,816
30 (7).817
Collins, M., S. Tett, and C. Cooper, 2001: The internal climate variability of HadCM3, a version818
of the Hadley Centre coupled model without flux adjustments. Clim. Dynam.,17 (1), 61–81.819
Collins, M., and Coauthors, 2010: The impact of global warming on the tropical Pacific Ocean820
and El Ni˜
no. Nature Geoscience,3 (6), 391–397.821
Collins, W., and Coauthors, 2011: Development and evaluation of an Earth-system model–822
HadGEM2. Geosci. Model Dev. Discuss,4, 997–1062.823
Compo, G. P., and Coauthors, 2011: The twentieth century reanalysis project. Q. J. Roy. Meteorol.824
Soc.,137 (654), 1–28.825
Dayan, H., T. Izumo, J. Vialard, and M. Lengaigne, 2014: Does sea surface temperature outside826
the tropical Pacific contribute to enhanced ENSO predictability? Clim. Dynam.,43, 1311–1325.827
Dayan, H., T. Izumo, J. Vialard, M. Lengaigne, and S. Masson, 2015: Do regions outside the trop-828
ical Pacific influence ENSO through atmospheric teleconnections ? Clim. Dynam., published829
online.830
Donner, L. J., B. L. Wyman, R. S. Hemler, L. W. Horowitz, Y. Ming, M. Zhao, J. C. Golaz,831
and Coauthors, 2011: The Dynamical Core, Physical Parameterizations, and Basic Simulation832
Characteristics of the Atmospheric Component AM3 of the GFDL Global Coupled Model CM3.833
J. Climate,24 (13), 3484–3519.834
38
Dufresne, J.-L., M.-A. Foujols, S. Denvil, A. Caubel, O. Marti, O. Aumont, Y. Balkanski, and835
Coauthors, 2012: Climate change projections using the IPSL-CM5 Earth System Model: from836
CMIP3 to CMIP5. submitted to Clim. Dynam.837
Fischer, A. S., P. Terray, E. Guilyardi, S. Gualdi, and P. Delecluse, 2005: Two independent triggers838
for the Indian Ocean dipole/zonal mode in a coupled GCM. J. Climate,18 (17), 3428–3449.839
Gent, P. R., G. Danabasoglu, L. J. Donner, M. M. Holland, E. C. Hunke, S. R. Jayne, D. M.840
Lawrence, and Coauthors, 2011: The Community Climate System Model Version 4. J. Climate,841
24, 4973–4991.842
Ham, Y.-G., S. Schubert, Y. Vikhliaev, and M. J. Suarez, 2014: An assessment of the ENSO843
forecast skill of GEOS-5 system. Clim. Dynam.,43 (9-10), 2415–2430.844
Ishii, M., A. Shouji, S. Sugimoto, and T. Matsumoto, 2005: Objective analyses of sea-surface845
temperature and marine meteorological variables for the 20th century using ICOADS and the846
Kobe Collection. International journal of climatology,25 (7), 865–879.847
Izumo, T., M. Lengaigne, J. Vialard, J.-J. Luo, T. Yamagata, and G. Madec, 2014: Influence of848
Indian Ocean Dipole and Pacific recharge on following year’s El Ni˜
no: interdecadal robustness.849
Clim. Dynam.,42 (1-2), 291–310.850
Izumo, T., J. Vialard, H. Dayan, M. Lengaigne, and I. Suresh, 2015: A simple estimation of equa-851
torial Pacific response from windstress to untangle Indian Ocean Dipole and Basin influences852
on El Ni˜
no. Clim. Dynam. in press.853
Izumo, T., and Coauthors, 2010: Influence of the state of the Indian Ocean Dipole on the following854
year’s El Nino. Nature Geoscience,3 (3), 168–172.855
39
Jin, F.-F., 1997: An equatorial ocean recharge paradigm for ENSO. Part I: Conceptual model. J.856
Atmos. Sci.,54 (7), 811–829.857
Jin, F.-F., S.-I. An, A. Timmermann, and J. Zhao, 2003: Strong El Nino events and nonlinear858
dynamical heating. Geophys. Res. Lett.,30 (3), 20–1.859
Jourdain, N. C., A. Sen Gupta, A. S. Taschetto, C. C. Ummenhofer, A. F. Moise, and K. Ashok,860
2013: The Indo-Australian monsoon and its relationship to ENSO and IOD in reanalysis data861
and the CMIP3/CMIP5 simulations. Clim. Dynam.,41 (11-12), 3073–3102.862
Kajtar, J. B., A. Santoso, M. H. England, and W. Cai, 2015: Indo-Pacific Climate Interactions in863
the Absence of an Indonesian Throughflow. J. Climate,28, 5017–5029.864
Kim, S. T., and J.-Y. Yu, 2012: The two types of ENSO in CMIP5 models. Geophys. Res. Lett.,865
39, L11704,39, L11 704.866
Klein, S. A., B. J. Soden, and N.-C. Lau, 1999: Remote sea surface temperature variations during867
ENSO: Evidence for a tropical atmospheric bridge. J. Climate,12 (4), 917–932.868
Kug, J.-S., and Y.-G. Ham, 2012: Indian Ocean Feedback to the ENSO Transition in a Multimodel869
Ensemble. J. Climate,25 (20), 6942–6957.870
Kug, J.-S., and I.-S. Kang, 2006: Interactive feedback between ENSO and the Indian Ocean. J.871
Climate,19 (9), 1784–1801.872
Lee, T., and M. J. McPhaden, 2010: Increasing intensity of El Ni ˜
no in the central-equatorial873
Pacific. Geophys. Res. Lett.,37 (14), L14 603.874
Leloup, J. A., Z. Lachkar, J.-P. Boulanger, and S. Thiria, 2007: Detecting decadal changes in875
ENSO using neural networks. Clim. Dynam.,28 (2-3), 147–162.876
40
Li, T., B. Wang, C. P. Chang, and Y. Zhang, 2003: A Theory for the Indian Ocean Dipole-Zonal877
Mode. J. Atmos. Sci.,60 (17), 2119–2135.878
Liu, L., S.-P. Xie, X.-T. Zheng, T. Li, Y. Du, G. Huang, and W.-D. Yu, 2014: Indian Ocean879
variability in the CMIP5 multi-model ensemble: the zonal dipole mode. Clim. Dynam.,43 (5-880
6), 1715–1730.881
Luo, J.-J., S. Masson, S. K. Behera, and T. Yamagata, 2008: Extended ENSO predictions using a882
fully coupled ocean-atmosphere model. J. Climate,21 (1), 84–93.883
Luo, J. J., R. Zhang, S. K. Behera, Y. Masumoto, F. F. Jin, R. Lukas, and T. Yamagata, 2010:884
Interaction between El Nino and extreme Indian ocean dipole. J. Climate,23 (3), 726–742.885
Martin, G. M., N. Bellouin, W. J. Collins, I. D. Culverwell, P. R. Halloran, S. C. Hardiman, T. J.886
Hinton, and Coauthors, 2011: The HadGEM2 family of Met Office Unified Model Climate887
configurations. Geosci. Model Dev. Discuss.,4, 765–841.888
McPhaden, M. J., 2012: A 21st century shift in the relationship between ENSO SST and warm889
water volume anomalies. Geophys. Res. Lett.,39 (9).890
McPhaden, M. J., and M. Nagura, 2014: Indian Ocean dipole interpreted in terms of recharge891
oscillator theory. Clim. Dynam.,42, 1569–1586.892
McPhaden, M. J., S. E. Zebiak, and M. H. Glantz, 2006: ENSO as an integrating concept in earth893
science. Science,314 (5806), 1740–1745.894
Meehl, G. A., 1987: The annual cycle and interannual variability in the tropical Pacific and Indian895
Ocean regions. Mon. Wea. Rev.,115 (1), 27–50.896
41
Meinen, C. S., and M. J. McPhaden, 2000: Observations of warm water volume changes in the897
equatorial Pacific and their relationship to El Ni˜
no and La Ni˜
na. Journal of Climate,13 (20),898
3551–3559.899
Murtugudde, R., J. P. McCreary, and A. J. Busalacchi, 2000: Oceanic processes associated with900
anomalous events in the Indian Ocean with relevance to 1997-1998. J. Geophys. Res.,105 (C2),901
3295–3306.902
Ohba, M., D. Nohara, and H. Ueda, 2010: Simulation of asymmetric ENSO transition in WCRP903
CMIP3 multimodel experiments. J. Climate,23 (22), 6051–6067.904
Ohba, M., and H. Ueda, 2005: Basin-wide warming in the equatorial Indian Ocean associated with905
El Ni˜
no. Scientific Online Letters on the Atmosphere,1 (0), 89–92.906
Ohba, M., and H. Ueda, 2007: An impact of SST anomalies in the Indian Ocean in acceleration907
of the El Ni˜
no to La Ni˜
na transition. Journal of the Meteorological Society of Japan,85 (3),908
335–348.909
Ohba, M., and H. Ueda, 2009a: Role of nonlinear atmospheric response to SST on the asymmetric910
transition process of ENSO. J. Climate,22 (1), 177–192.911
Ohba, M., and H. Ueda, 2009b: Seasonally different response of the Indian Ocean to the remote912
forcing of El Ni˜
no: Linking the dynamics and thermodynamics. Scientific Online Letters on the913
Atmosphere,5, 176–179.914
Ohba, M., and M. Watanabe, 2012: Role of the Indo-Pacific interbasin coupling in predicting915
asymmetric ENSO transition and duration. J. Climate,25 (9), 3321–3335.916
Okumura, Y. M., M. Ohba, C. Deser, and H. Ueda, 2011: A Proposed Mechanism for the Asym-917
metric Duration of El Ni˜
no and La Ni˜
na. J. Climate.918
42
Picaut, J., F. Masia, and Y. Du Penhoat, 1997: An advective-reflective conceptual model for the919
oscillatory nature of the ENSO. Science,277 (5326), 663–666.920
Raddatz, T., and Coauthors, 2007: Will the tropical land biosphere dominate the climate–carbon921
cycle feedback during the twenty-first century? Clim. Dynam.,29 (6), 565–574.922
Rayner, N. A., P. Brohan, D. E. Parker, C. K. Folland, J. J. Kennedy, M. Vanicek, T. J. Ansell, and923
S. F. B. Tett, 2006: Improved analyses of changes and uncertainties in sea surface temperature924
measured in situ since the mid-nineteenth century: The HadSST2 dataset. J. Climate,19 (3),925
446–469.926
Rayner, N. A., D. E. Parker, E. B. Horton, C. K. Folland, L. V. Alexander, D. P. Rowell, E. C. Kent,927
and A. Kaplan, 2003: Global analyses of sea surface temperature, sea ice, and night marine air928
temperature since the late nineteenth century. J. Geophys. Res,108 (D14), 4407.929
Reverdin, G., D. L. Cadet, and D. G., 1986: Interannual displacements of convection and surface930
circulation over the equatorial Indian Ocean. Quarterly Journal of the Royal Meteorological931
Society,112 (471), 43–67.932
Saji, N. H., B. N. Goswami, P. N. Vinayachandran, and T. Yamagata, 1999: A dipole mode in the933
tropical Indian Ocean. Nature,401 (6751), 360–363.934
Santoso, A., M. H. England, and W. Cai, 2012: Impact of Indo-Pacific feedback interactions on935
ENSO dynamics diagnosed using ensemble climate simulations. J. Climate,25 (21), 7743–936
7763.937
Schmidt, G. A., and Coauthors, 2006: Present-day atmospheric simulations using GISS ModelE:938
Comparison to in situ, satellite, and reanalysis data. J. Climate,19 (2), 153–192.939
43
Schwarzkopf, F. U., and C. W. B¨
oning, 2011: Contribution of Pacific wind stress to multi-decadal940
variations in upper-ocean heat content and sea level in the tropical south Indian Ocean. Geophys.941
Res. Lett.,38 (12).942
Smith, T. M., R. W. Reynolds, T. C. Peterson, and J. Lawrimore, 2008: Improvements to NOAA’s943
historical merged land-ocean surface temperature analysis (1880-2006). Journal of Climate,944
21 (10), 2283–2296.945
Tang, Y., and Z. Deng, 2010: Tropical Pacific upper ocean heat content variations and ENSO946
predictability during the period from 1881–2000. Adv. Geosci,18, 87–108.947
Taschetto, A. S., A. Sen Gupta, N. C. Jourdain, A. Santoso, C. C. Ummenhofer, and M. H. Eng-948
land, 2014: Cold tongue and warm pool ENSO events in CMIP5: mean state and future projec-949
tions. J. Climate,27 (8), 2861–2885.950
Taylor, K. E., R. J. Stouffer, and G. A. Meehl, 2011: An overview of CMIP5 and the experiment951
design. Bull. Am. Meteorol. Soc.952
Tozuka, T., J.-J. Luo, S. Masson, and T. Yamagata, 2007: Decadal modulations of the Indian953
Ocean dipole in the SINTEX-F1 coupled GCM. J. Climate,20 (13), 2881–2894.954
Trenberth, K. E., G. W. Branstator, D. Karoly, A. Kumar, N.-C. Lau, and C. Ropelewski, 1998:955
Progress during TOGA in understanding and modeling global teleconnections associated with956
tropical sea surface temperatures. J. Geophys. Res.,103 (C7), 14 291–14 324.957
Ummenhofer, C. C., M. H. England, P. C. McIntosh, G. A. Meyers, M. J. Pook, J. S. Risbey,958
A. Sen Gupta, and A. S. Taschetto, 2009: What causes southeast Australia’s worst droughts?959
Geophys. Res. Lett.,36 (4), L04 706.960
44
Voldoire, A., and Coauthors, 2012: The CNRM-CM5. 1 global climate model: Description and961
basic evaluation. Clim. Dynam. published online.962
Volodin, E. M., N. A. Dianskii, and A. V. Gusev, 2010: Simulating present-day climate with963
the INMCM4. 0 coupled model of the atmospheric and oceanic general circulations. Izvestiya964
Atmospheric and Oceanic Physics,46 (4), 414–431.965
Wang, C., and J. Picaut, 2004: Understanding ENSO physicsA review. Earth’s Climate, 21–48.966
Watanabe, M., and Coauthors, 2010: Improved Climate Simulation by MIROC5: Mean States,967
Variability, and Climate Sensitivity. J. Climate,23 (23), 6312–6335.968
Watanabe, S., and Coauthors, 2011: MIROC-ESM: model description and basic results of CMIP5-969
20c3m experiments. Geosci. Model Dev. Discuss.,4, 1063–1128.970
Webster, P. J., A. M. Moore, J. P. Loschnigg, and R. R. Leben, 1999: Coupled ocean–atmosphere971
dynamics in the Indian Ocean during 1997–98. Nature,401 (6751), 356–360.972
Wittenberg, A. T., 2009: Are historical records sufficient to constrain ENSO simulations. Geophys.973
Res. Lett,36, L12 702.974
Wu, B., T. Li, and T. Zhou, 2010: Asymmetry of Atmospheric Circulation Anomalies over the975
Western North Pacific between El Ni˜
no and La Ni˜
na. J. Climate,23 (18), 4807–4822.976
Wyrtki, K., 1985: Water displacements in the Pacific and the genesis of El Ni˜
no cycles. J. Geophys.977
Res.,90 (C4), 7129–7132.978
Xie, S.-P., K. Hu, J. Hafner, H. Tokinaga, Y. Du, G. Huang, and T. Sampe, 2009: Indian Ocean ca-979
pacitor effect on Indo-western Pacific climate during the summer following El Ni ˜
no. J. Climate,980
22 (3), 730–747.981
45
Yasunaka, S., and K. Hanawa, 2011: Intercomparison of historical sea surface temperature982
datasets. International Journal of Climatology,31 (7), 1056–1073.983
Yongqiang, Y., Z. Xuehong, and G. Yufu, 2004: Global coupled ocean-atmosphere general circu-984
lation models in LASG/IAP. Advances in Atmospheric sciences,21 (3), 444–455.985
Yuan, D., H. Zhou, and X. Zhao, 2013: Interannual Climate Variability over the Tropical Pacific986
Ocean Induced by the Indian Ocean Dipole through the Indonesian Throughflow. J. Climate,987
26 (9).988
Yuan, D., and Coauthors, 2011: Forcing of the Indian Ocean Dipole on the interannual variations989
of the tropical Pacific Ocean: roles of the Indonesian Throughflow. J. Climate,24 (14), 3593–990
3608.991
Yuan, Y., and C. Li, 2008: Decadal variability of the IOD-ENSO relationship. Chinese Science992
Bulletin,53 (11), 1745–1752.993
Yukimoto, S., and Coauthors, 2001: The new Meteorological Research Institute coupled994
GCM(MRI-CGCM 2)- Model climate and variability. Papers in Meteorology and Geophysics,995
51 (2), 47–88.996
997
46
998
47
LIST OF TABLES999
Table 1. CMIP5 model names, providing institutes, and related reference. . . . . . . 481000
48
Model Institute References
ACCESS-1.0 CSIRO-BOM, Australia BOM 2010
CanESM2 CCCMA, Canada Chylek et al. 2011
CCSM4 NCAR, CO, USA Gent et al. 2011
CNRM-CM5 CNRM-CERFACS, France Voldoire et al. 2012
FGOALS-g2 LASG-CESS, China Yongqiang et al. 2004
FGOALS-s2 LASG-IAP, China Yongqiang et al. 2004
GFDL-CM3 NOAA, GFDL, USA Donner et al. 2011
GFDL-ESM2G NOAA, GFDL, USA Donner et al. 2011
GFDL-ESM2M NOAA, GFDL, USA Donner et al. 2011
GISS-E2-H NASA/GISS, NY, USA Schmidt et al. 2006
HadCM3 MOHC, UK Collins et al. 2001
HadGEM2-CC MOHC, UK Martin et al. 2011
HadGEM2-ES MOHC, UK Collins et al. 2011
INMCM4 INM, Russia Volodin et al. 2010
IPSL-CM5A-LR IPSL, France Dufresne et al. 2012
IPSL-CM5B-LR IPSL, France Dufresne et al. 2012
IPSL-CM5A-MR IPSL, France Dufresne et al. 2012
MIROC5 AORI-NIES-JAMSTEC, Japan Watanabe et al. 2010
MIROC-ESM-CHEM AORI-NIES-JAMSTEC, Japan Watanabe et al. 2011
MPI-ESM-LR MPI-N, Germany Raddatz et al. 2007
MPI-ESM-MR MPI-N, Germany Raddatz et al. 2007
MRI-CGCM3 MRI, Japan Yukimoto et al. 2001
NorESM1-M NCC, Norway Bentsen et al. 2012
TABL E 1. CMIP5 model names, providing institutes, and related reference.
49
LIST OF FIGURES1001
Fig. 1. Correlation between DMI in SON (show by the gray bar) and lagged monthly NINO34 in1002
four observational datasets over 1890-2012. The 90% significance level is indicated by the1003
horizontal dashed lines (t-test with 100 degrees of freedom, which corresponds to 123 years1004
of data with a significant auto-correlation over 15 months as noted in Burgers 1999). . . . 511005
Fig. 2. Standard deviation of NINO34 (a) and DMI (b) for 23 CMIP5 models (color lines), the1006
CMIP5 multi-model mean (solid black), and the observational multi-dataset mean (dashed1007
black), in K. (c) and (d) are similar to (a) and (b), except that the curve is normalized1008
(anomaly with respect to the mean, divided by the standard deviation). . . . . . . . 521009
Fig. 3. (a) Correlation between NINO34 in NDJ (gray bar) and lagged monthly WWV for 211010
CMIP5 models (colors), the CMIP5 multi-model mean (solid black), the SODA reanaly-1011
sis (long-dashed black), and the NOAA-BMRC observational dataset (short-dashed black).1012
(b) Auto-correlation between NINO34 in NDJ (gray bar) and lagged monthly NINO34 for1013
21 CMIP5 models (colors), the CMIP5 multi-model mean (solid black), and the multi-1014
observation mean (dashed black). The 90% level of statistical significance for 100 degrees1015
of freedom is indicated by the dashed line (see caption of Fig. 1). . . . . . . . . . 531016
Fig. 4. (a) Proportion (%) of Ni ˜
nos followed by La Ni˜
na (yellow) and of La Ni˜
nas followed by El1017
Ni˜
no (green). (b) Proportion of pIODs immediately followed by El Ni˜
no (red) and of nIODs1018
immediately followed by La Ni˜
na (blue). (c) Proportion of pIODs followed by a La Ni˜
na1019
14 months later (orange) and of nIODs followed by El Ni˜
no 14 months later (blue). The1020
bars falling between the two dashed lines cannot be considered as different from the random1021
proportion of 25% (90% confidence interval based on Monte-Carlo statistical method, see1022
text). The right-end column represents the CMIP5 multi-model mean, and the error bars1023
indicates the 90% confidence interval on the mean (t-value, all CMIP5 distributions being1024
normally distributed at the 99% confidence level according to a Shapiro-Wilk test). . . . . 541025
Fig. 5. Correlation between DMI in SON (shown by the gray bar) and lagged monthly NINO34 in1026
21 CMIP5 models. The 90% significance level is indicated by the horizontal dashed lines1027
(t-test for 100 degrees of freedom, see caption of Fig. 1). . . . . . . . . . . . 551028
Fig. 6. Increase of explained NINO34 variance when a second predictor of ENSO is used in ad-1029
dition to WWV. (a) Increased variance due to DMI (i.e. first predictor is SON WWV and1030
second predictor is SON DMI). (b) Increased variance due to TIO (i.e. first predictor is1031
JF WWV and second predictor is JF TIO). The plot shows the mean value over 105cross-1032
validations leaving 50% of the samples (randomly chosen) to train the prediction model and1033
keeping 50% for its evaluation. The significance of the increase in variance obtained from1034
this method is indicated within brackets (bold when 90% significant). The date on which1035
the forecast is issued is indicated by the gray bar, and the peak of ENSO in NDJ of year 1 is1036
indicated by the light red bar. . . . . . . . . . . . . . . . . . . 561037
Fig. 7. (a) NINO34 values in SODA (y-axis) vs NINO34 hindcasts based on a linear function of1038
WWV and DMI 14 months before (x-axis). (b) Same for the 21 CMIP5 simulations together1039
instead of SODA. The solid lines indicate the upper and lower quartiles defining El Ni˜
no and1040
La Ni˜
na events. The shaded areas indicate successful hindcasts of La Ni ˜
na (blue), neutral1041
(gray), and El Ni˜
no events. The ratio of successful hindcasts is indicated for each shaded1042
area. The mean values of the points located in each shaded area are indicated by a cross,1043
and the size of each segment of the cross corresponds to the 90% confidence interval on the1044
mean(t-test).......................571045
50
Fig. 8. (a) Scatter plot of the actual probability of the delayed pIOD-Ni ˜
na and nIOD-Ni˜
na transi-1046
tions (Y-axis) vs the probability calculated from Eq. 1-2 (X-axis). The cross-model corre-1047
lation coefficient ris indicated for both transitions. The characters are displayed as bold1048
when the null hypothesis is also rejected at the 90% confidence level when using the sec-1049
ond method (see panel b), and in italic otherwise. (b) Probability to have a synthetic DMI1050
timeseries (see Eq. 4) with a proportion of delayed pIOD-Ni˜
na (orange) and nIOD-Ni˜
no1051
(blue) transitions lower than the actual proportion in the observed and modelled timeseries.1052
This probability can be considered as the statistical confidence level for rejecting the null1053
hypothesis related to our second method (see text). The 10% and 90% levels are indicated1054
withdashedlines......................581055
Fig. 9. Longitude-time sections showing the influence of SON DMI on Equatorial Indo-Pacific1056
SSTs (averaged over 10◦S -0◦in the Indian Ocean and 5◦S -5◦N in the Pacific), Equatorial1057
temperature over 0-300 m (T300, averaged between 3◦S -3◦N ), and zonal wind (averaged1058
between 3◦S -3◦N ). The linear influence of synchronous NDJ NINO34 is removed as fol-1059
lows. For any variable Y(SST, T300 or zonal wind), we calculate the coefficients of the1060
following multi-linear regression: Y=a
]
DMI +b
^
NINO34, with]denoting non-dimensional1061
variables (divided by their standard deviation), and −ais plotted here to illustrate the case of1062
a negative IOD (even though positive IODs are also accounted for in this plot). Gray shad-1063
ing indicate no data for T300. The upper row shows the CMIP5 multi-model mean anomaly1064
based on the best tercile of models in terms of enhanced predictability of ENSO from DMI1065
(according to Fig. 6c). The second row is similar but for the lowest CMIP5 tercile. Multi-1066
model mean composites consist of the average of regression coefficients from individual1067
models. The two lower rows are based on SODA and 20CR over 1890-2008 (third row) and1068
1980-2008 (fourth row). The black contours indicate the 90% confidence level (multi-model1069
meanforCMIP5)......................591070
51
SO ND J F MAM J J A SO ND J F MAM J J A S OND J F MAM J J A SO ND
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
HadSST2
HadISST
ERSST
COBE
zero lag
NINO34 leads DMI leads
FIG . 1. Correlation between DMI in SON (show by the gray bar) and lagged monthly NINO34 in four
observational datasets over 1890-2012. The 90% significance level is indicated by the horizontal dashed lines
(t-test with 100 degrees of freedom, which corresponds to 123 years of data with a significant auto-correlation
over 15 months as noted in Burgers 1999).
1071
1072
1073
1074
52
J F M A M J J A S O N D
0.2
0.4
0.6
0.8
1
1.2
1.4
J F M A M J J A S O N D
0
0.2
0.4
0.6
0.8
1
1.2
J F M A M J J A S O N D
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
J F M A M J J A S O N D
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
ACCESS1−0
CanESM2
CCSM4
CNRM−CM5
FGOALS−g2
FGOALS−s2
GFDL−CM3
GFDL−ESM2G
GFDL−ESM2M
GISS−E2−H
HadCM3
HadGEM2−CC
HadGEM2−ES
inmcm4
IPSL−CM5B−LR
MIROC5
MIROC−ESM−CHEM
MPI−ESM−LR
MPI−ESM−MR
MRI−CGCM3
NorESM1−M
IPSL−CM5A−LR
IPSL−CM5A−MR
CMIP5 mean
Observations
(a) NINO34 std (b) DMI std
(c) normalized NINO34 std (d) normalized DMI std
FIG . 2. Standard deviation of NINO34 (a) and DMI (b) for 23 CMIP5 models (color lines), the CMIP5 multi-
model mean (solid black), and the observational multi-dataset mean (dashed black), in K. (c) and (d) are similar
to (a) and (b), except that the curve is normalized (anomaly with respect to the mean, divided by the standard
deviation).
1075
1076
1077
1078
53
J FMAM J J ASOND J FMAM J J A SOND J FMAM J J ASOND J FMAM J J ASOND
−1
−0. 8
−0. 6
−0. 4
−0. 2
0
0. 2
0. 4
0. 6
0. 8
1NDJ NINO34 (zero−lag)
ACCESS1−0
CanESM2
CCSM4
CNRM−CM5
FGOALS−g2
FGOALS−s2
GFDL−CM3
GFDL−ESM2G
GFDL−ESM2M
GISS−E2−H
HadCM3
HadGEM2−CC
HadGEM2−ES
inmcm4
IPSL−CM5B−LR
MIROC5
MIROC−ESM−CHEM
MPI−ESM−LR
MPI−ESM−MR
MRI−CGCM3
NorESM1−M
CMIP5 mean
NOAA-BMRC 1980−2012
SODA 1890−2008
NINO34 leads
WWV leads
J FMAMJ J ASOND J FMAMJ J ASOND J FMAM J J ASOND J FMAMJ J ASOND
−0. 6
−0. 4
−0. 2
0
0. 2
0. 4
0. 6
0. 8
1NDJ NINO34 (zero−lag)
ACCESS1−0
CanESM2
CCSM4
CNRM−CM5
FGOALS−g2
FGOALS−s2
GFDL−CM3
GFDL−ESM2G
GFDL−ESM2M
GISS−E2−H
HadCM3
HadGEM2−CC
HadGEM2−ES
inmcm4
IPSL−CM5B−LR
MIROC5
MIROC−ESM−CHEM
MPI−ESM−LR
MPI−ESM−MR
MRI−CGCM3
NorESM1−M
CMIP5 mean
OBS mean (HadSST2,
HadISST,COBE,ERSST)
(a)
(b)
zero lag
FIG . 3. (a) Correlation between NINO34 in NDJ (gray bar) and lagged monthly WWV for 21 CMIP5 models
(colors), the CMIP5 multi-model mean (solid black), the SODA reanalysis (long-dashed black), and the NOAA-
BMRC observational dataset (short-dashed black). (b) Auto-correlation between NINO34 in NDJ (gray bar) and
lagged monthly NINO34 for 21 CMIP5 models (colors), the CMIP5 multi-model mean (solid black), and the
multi-observation mean (dashed black). The 90% level of statistical significance for 100 degrees of freedom is
indicated by the dashed line (see caption of Fig. 1).
1079
1080
1081
1082
1083
1084
54
HadISST
ERSST
COBE
ACCESS1−0
CanESM2
CCSM4
CNRM−CM5
FGOALS−g2
FGOALS−s2
GFDL−CM3
GFDL−ESM2G
GFDL−ESM2M
GISS−E2−H
HadCM3
HadGEM2−CC
HadGEM2−ES
inmcm4
IPSL−CM5B−LR
MIROC5
MIROC−ESM−CH
MPI−ESM−LR
MPI−ESM−MR
MRI−CGCM3
NorESM1−M
pIOD and El Niño nIOD and La Niña (b) Synchronous DMI-ENSO
(c) Delayed DMI-ENSO pIOD to La Niña nIOD to El Niño
80
60
40
20
0
80
60
40
20
0
60
40
20
0
(a) year-to-year ENSO-ENSO Niño to Niña Niña to Niño
CMIP5 mean
FIG . 4. (a) Proportion (%) of Ni˜
nos followed by La Ni˜
na (yellow) and of La Ni˜
nas followed by El Ni˜
no
(green). (b) Proportion of pIODs immediately followed by El Ni ˜
no (red) and of nIODs immediately followed
by La Ni˜
na (blue). (c) Proportion of pIODs followed by a La Ni˜
na 14 months later (orange) and of nIODs
followed by El Ni˜
no 14 months later (blue). The bars falling between the two dashed lines cannot be considered
as different from the random proportion of 25% (90% confidence interval based on Monte-Carlo statistical
method, see text). The right-end column represents the CMIP5 multi-model mean, and the error bars indicates
the 90% confidence interval on the mean (t-value, all CMIP5 distributions being normally distributed at the 99%
confidence level according to a Shapiro-Wilk test).
1085
1086
1087
1088
1089
1090
1091
1092
55
SON D J FMA MJ J A SON D J FMA MJ J A SON D J FMAM J J A SON D
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
zero lag
NINO34 leads
DMI leads ACCESS1−0
CanESM2
CCSM4
CNRM−CM5
FGOALS−g2
FGOALS−s2
GFDL−CM3
GFDL−ESM2G
GFDL−ESM2M
GISS−E2−H
HadCM3
HadGEM2−CC
HadGEM2−ES
inmcm4
IPSL−CM5B−LR
MIROC5
MIROC−ESM−CHEM
MPI−ESM−LR
MPI−ESM−MR
MRI−CGCM3
NorESM1−M
CMIP5 mean
Observations
FIG . 5. Correlation between DMI in SON (shown by the gray bar) and lagged monthly NINO34 in 21 CMIP5
models. The 90% significance level is indicated by the horizontal dashed lines (t-test for 100 degrees of freedom,
see caption of Fig. 1).
1093
1094
1095
56
year 1 ENSO peak
IOB peak
year 1 ENSO peak
O ND J F M AM J J A S O N D J F MA M J J A S
0
0.1
0.2
0.3
0.4
0.5
0.6
0
0.1
0.2
0.3
0.4
0.5
0.6
J F M A M J J A S O N D J F M A M
year 0 IOD peak
[a] [b]
ACCESS1-0
CanESM2
CCSM4
CNRM-CM5
FGOALS-g2
FGOALS-s2
GFDL-CM3
GFDL-ESM2G
GFDL-ESM2M
GISS-E2-H
HadCM3
HadGEM2-CC
HadGEM2-ES
inmcm4
IPSL-CM5B-LR
MIROC5
MIROC-ESM-CHEM
MPI-ESM-LR
MPI-ESM-MR
MRI-CGCM3
NorESM1-M
CMIP5 mean
SODA 1890-2008
SODA 1980-2008
(a: 0.94 / b: 0.74)
(a: 1.00 / b: 0.83)
(a: 0.96 / b: 0.77)
(a: 1.00 / b: 0.92)
(a: 1.00 / b: 0.97)
(a: 0.83 / b: 0.66)
(a: 1.00 / b: 0.74)
(a: 0.94 / b: 0.44)
(a: 0.95 / b: 0.81)
(a: 0.64 / b: 0.53)
(a: 1.00 / b: 0.99)
(a: 0.95 / b: 0.87)
(a: 0.81 / b: 0.73)
(a: 0.90 / b: 0.90)
(a: 0.97 / b: 0.98)
(a: 0.37 / b: 0.82)
(a: 0.75 / b: 0.52)
(a: 0.92 / b: 0.35)
(a: 0.93 / b: 0.80)
(a: 0.92 / b: 0.57)
(a: 0.98 / b: 0.89)
(a: 0.98 / b: 0.91)
(a: 0.91 / b: 0.51)
FIG . 6. Increase of explained NINO34 variance when a second predictor of ENSO is used in addition to
WWV. (a) Increased variance due to DMI (i.e. first predictor is SON WWV and second predictor is SON DMI).
(b) Increased variance due to TIO (i.e. first predictor is JF WWV and second predictor is JF TIO). The plot
shows the mean value over 105cross-validations leaving 50% of the samples (randomly chosen) to train the
prediction model and keeping 50% for its evaluation. The significance of the increase in variance obtained from
this method is indicated within brackets (bold when 90% significant). The date on which the forecast is issued
is indicated by the gray bar, and the peak of ENSO in NDJ of year 1 is indicated by the light red bar.
1096
1097
1098
1099
1100
1101
1102
57
NINO34 linear hindcast from IOD and WWV
Observed NINO34
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
(a) SODA
NINO34 linear hindcast from IOD and WWV
Simulated NINO34
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
(b) CMIP5
51.7%
50%
48.6%
50%
57%
58.7%
FIG . 7. (a) NINO34 values in SODA (y-axis) vs NINO34 hindcasts based on a linear function of WWV and
DMI 14 months before (x-axis). (b) Same for the 21 CMIP5 simulations together instead of SODA. The solid
lines indicate the upper and lower quartiles defining El Ni˜
no and La Ni˜
na events. The shaded areas indicate
successful hindcasts of La Ni˜
na (blue), neutral (gray), and El Ni˜
no events. The ratio of successful hindcasts is
indicated for each shaded area. The mean values of the points located in each shaded area are indicated by a
cross, and the size of each segment of the cross corresponds to the 90% confidence interval on the mean (t-test).
1103
1104
1105
1106
1107
1108
58
10 20 30 40 50 60
#
#
$
$
&
&
A
A
B
B
C
C
D
D
E
E
F
F
G
G
H
H
I
I
J
J
K K
L
L
M
M
N
N
O
O
P
P
Q
Q
R
R
S
S
T
T
U
U
Actual probability (%)
Probability from sync. IOD−ENSO and delayed ENSO−ENSO (%)
# HadISST
$ ERSST
& COBE
A ACCESS1-0
B CanESM2
C CCSM4
D CNRM-CM5
E FGOALS-g2
F FGOALS-s2
G GFDL-CM3
H GFDL-ESM2G
I GFDL-ESM2M
J GISS-E2-H
K HadCM3
L HadGEM2-CC
M HadGEM2-ES
N inmcm4
O IPSL-CM5B-LR
P MIROC5
Q MIROC-ESM-CHEM
R MPI-ESM-LR
S MPI-ESM-MR
T MRI-CGM3
U NorESM1-M
pIOD to La Niña
r = 0.84
nIOD to El Niño
r = 0.36
HadISST
ERSST
COBE
ACCESS1−0
CanESM2
CCSM4
CNRM−CM5
FGOALS−g2
FGOALS−s2
GFDL−CM3
GFDL−ESM2G
GFDL−ESM2M
GISS−E2−H
HadCM3
HadGEM2−CC
HadGEM2−ES
inmcm4
IPSL−CM5B−LR
MIROC5
MIROC−ESM−CH
MPI−ESM−LR
MPI−ESM−MR
MRI−CGCM3
NorESM1−M
100
90
80
70
60
50
40
30
20
10
0
pIOD to La Niña
nIOD to El Niño
(a)
(b)
60
50
40
30
20
10
Probability (%)
FIG . 8. (a) Scatter plot of the actual probability of the delayed pIOD-Ni˜
na and nIOD-Ni˜
na transitions (Y-
axis) vs the probability calculated from Eq. 1-2 (X-axis). The cross-model correlation coefficient ris indicated
for both transitions. The characters are displayed as bold when the null hypothesis is also rejected at the 90%
confidence level when using the second method (see panel b), and in italic otherwise. (b) Probability to have a
synthetic DMI timeseries (see Eq. 4) with a proportion of delayed pIOD-Ni˜
na (orange) and nIOD-Ni˜
no (blue)
transitions lower than the actual proportion in the observed and modelled timeseries. This probability can be
considered as the statistical confidence level for rejecting the null hypothesis related to our second method (see
text). The 10% and 90% levels are indicated with dashed lines.
1109
1110
1111
1112
1113
1114
1115
1116
59
CMIP5 top tercile
40E 120E 160W 80W
year0 year1
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec CMIP5 top tercile
40E 120E 160W 80W
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec CMIP5 top tercile
40E 120E 160W 80W
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec
CMIP5 lowest tercile
40E 120E 160W 80W
year0 year1
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec CMIP5 lowest tercile
40E 120E 160W 80W
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec CMIP5 lowest tercile
40E 120E 160W 80W
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec
SODA 1890-2008
40E 120E 160W 80W
year0 year1
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec SODA 1890-2008
40E 120E 160W 80W
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec 20CR 1890-2008
40E 120E 160W 80W
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec
SODA 1980-2008
Longitude
40E 120E 160W 80W
year0 year1
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec SODA 1980-2008
Longitude
40E 120E 160W 80W
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec 20CR 1980-2008
Longitude
40E 120E 160W 80W
Feb
Apr
Jun
Aug
Oct
Dec
Feb
Apr
Jun
Aug
Oct
Dec
SST (K) T300 (K) zonal wind (m/s)
SST, T300 (K) wind (m/s)
0 0.2 0.4 0.6 >0.8-0.2-0.4-0.6<-0.8 0 0.3 0.6 0.9 >1.2-0.3-0.6-0.9<-1.2
[a]
[b]
[c]
[d]
FIG . 9. Longitude-time sections showing the influence of SON DMI on Equatorial Indo-Pacific SSTs (av-
eraged over 10◦S -0◦in the Indian Ocean and 5◦S -5◦N in the Pacific), Equatorial temperature over 0-300 m
(T300, averaged between 3◦S -3◦N ), and zonal wind (averaged between 3◦S -3◦N ). The linear influence of
synchronous NDJ NINO34 is removed as follows. For any variable Y(SST, T300 or zonal wind), we calculate
the coefficients of the following multi-linear regression: Y=ag
DMI +b
^
NINO34, withedenoting non-dimensional
variables (divided by their standard deviation), and −ais plotted here to illustrate the case of a negative IOD
(even though positive IODs are also accounted for in this plot). Gray shading indicate no data for T300. The up-
per row shows the CMIP5 multi-model mean anomaly based on the best tercile of models in terms of enhanced
predictability of ENSO from DMI (according to Fig. 6c). The second row is similar but for the lowest CMIP5
tercile. Multi-model mean composites consist of the average of regression coefficients from individual models.
The two lower rows are based on SODA and 20CR over 1890-2008 (third row) and 1980-2008 (fourth row). The
black contours indicate the 90% confidence level (multi-model mean for CMIP5).
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
60
