Indonesian Throughflow Variability and Linkage to ENSO and IOD in an Ensemble of CMIP5 Models
AGUSSANTOSO,a,b,cMATTHEWH. ENGLAND,a,bJULESB. KAJTAR,d,eANDWENJUCAIc,f,g
aARC Centre of Excellence for Climate Extremes, University of New South Wales, Sydney, Australia
bClimate Change Research Centre, University of New South Wales, Sydney, Australia
cCentre for Southern Hemisphere Oceans Research (CSHOR), CSIRO Oceans and Atmosphere, Hobart, Tasmania, Australia
dInstitute for Marine and Antarctic Studies, University of Tasmania, Hobart, Tasmania, Australia
eARC Centre of Excellence for Climate Extremes, University of Tasmania, Hobart, Tasmania, Australia
fKey Laboratory of Physical Oceanography/Institute for Advanced Ocean Studies, Ocean University of China, Qingdao, China
gQingdao National Laboratory for Marine Science and Technology, Qingdao, China
(Manuscript received 24 June 2021, infinal form 15 December 2021)
ABSTRACT: Understanding variability of the Indonesian Throughflow (ITF) and its links to El Nino˜ –Southern Oscilla- tion (ENSO) and the Indian Ocean dipole (IOD), and how they are represented across climate models constitutes an impor- tant step toward improved future climate projections. These issues are examined using 20 models from phase 5 of the Coupled Model Intercomparison Project (CMIP5) and the SODA-2.2.4 ocean reanalysis. It is found that the CMIP5 models overall simulate aspects of ITF variability, such as spectral and vertical structure, that are consistent with the reanalysis, although intermodel differences are substantial. The ITF variability is shown to exhibit two dominant principal vertical structures: a surface-intensified transport anomaly (ITFM1) and an anomalous transport characterized by opposingflows in the surface and subsurface (ITFM2). In the CMIP5 models and reanalysis, ITFM2is linked to both ENSO and the IOD via anomalous Indo-Pacific Walker circulation. The driver of ITFM1however differs between the reanalysis and the CMIP5 models. In the reanalysis ITFM1is a delayed response to ENSO, whereas in the CMIP5 models it is linked to the IOD associ- ated with the overly strong IOD amplitude bias. Further, the CMIP5 ITF variability tends to be weaker than in the reanaly- sis, due to a tendency for the CMIP5 models to simulate a delayed IOD in response to ENSO. The importance in considering the vertical structure of ITF variability in understanding ENSO and IOD impact is further underscored by the close link between greenhouse-forced changes in ENSO variability and projected changes in subsurface ITF variability.
KEYWORDS: Indian Ocean; Pacific Ocean; El Nino; ENSO; La Nina; Ocean circulation; Ocean dynamics; Climate change; Climate variability; Climate models; Ocean models; Reanalysis data; Interannual variability; Oceanic variability;
Tropical variability; Maritime Continent
1. Introduction
The Indo-Pacific region is a major component of the global climate system, which hosts some of Earth’s most dominant sources of climate variability, El Nino˜ –Southern Oscillation (ENSO) and the Indian Ocean dipole (IOD).
The world’s warmest and most expansive pool of ocean waters straddles the Maritime Continent, regulating the global climate. Within the warm pool, vigorous atmospheric convection forms the ascending branch of the Walker circu- lation, fed by the easterly Pacific trade winds, which pile up water toward the Maritime Continent, thereby creating a pressure difference that drives voluminous oceanic flow from the Pacific to the Indian Ocean through the Indonesian Archipelago (Wyrtki 1987). This cross-basin oceanic flow,
termed the Indonesian Throughflow (ITF), forms an inte- gral part of the global ocean circulations, maintaining the state of Earth’s climate and its variability (e.g., Hirst and Godfrey 1993; Gordon and Fine 1996; Murtugudde et al.
1998;Vranes et al. 2002;Jochum et al. 2009;Santoso et al.
2011;Sprintall et al. 2014;Kajtar et al. 2015;Hu et al. 2015;
Sprintall et al. 2020).
Changes to ITF transport influence heat and freshwater balance in the Indo-Pacific region that is important for climate across various time scales (Vranes et al. 2002; Feng et al.
2013;Ummenhofer et al. 2017;Jin and Wright 2020). The ITF is also involved in the recharge and discharge of equatorial Pacific warm water during ENSO events (e.g.,McGregor et al.
2014), which can trigger IOD occurrences (e.g.,Yang et al.
2015). Thus, not only do ENSO and the IOD affect the ITF (e.g.,Sprintall and Revelard 2014;Hu and Sprintall 2016), but changes in the ITF may in turn impact ENSO and the IOD (Lee et al. 2002;Yuan et al. 2013;Kajtar et al. 2015). Under- standing how the ITF is linked to ENSO and the IOD is therefore important for discerning the impacts of these inter- actions, particularly as the climate system is expected to con- tinue to change under global warming with projected increase in the frequency of extreme ENSO and IOD events (Cai et al.
2020,2021). While our knowledge of Indo-Pacific linkages has improved (Wang 2019; Cai et al. 2019), the extent to which Denotes content that is immediately available upon publica-
tion as open access.
Supplemental information related to this paper is available at the Journals Online website: https://doi.org/10.1175/JCLI-D-21- 0485.s1.
Corresponding author: Agus Santoso, [email protected]
ITF variability and its links to ENSO and the IOD are repre- sented across a wide range of climate models that are used to make future projections is still not clear. Here we investigate this issue using 20 climate models that participated in phase 5 of the Coupled Model Intercomparison Project (CMIP5).
As inferred from observational and model-based studies, the warm (El Ni˜no) and cold (La Ni˜na) phases of ENSO gen- erally correspond with anomalously weak and strong ITF transport, respectively (Clarke and Liu 1994; Meyers 1996;
England and Huang 2005;Santoso et al. 2011;Gordon et al.
1999; Sprintall and Revelard 2014; van Sebille et al. 2014;
Hu and Sprintall 2016). The Walker circulation weakens during an El Nino, as the Paci˜ fic easterly trade winds relax, associated with anomalously warm central-eastern equatorial Pacific sea surface. This leads to lower-than-normal sea level in the western Pacific, consequently reducing the interbasin pressure gradient that drives the ITF. The converse occurs during a La Ni˜na. Given the prominence of ENSO on ocean– atmospheric circulation changes in the Indo-Pacific region, the impact of the IOD can be difficult to discern. The positive phase of the IOD, marked by anomalously cold and warm sea surface in the eastern and western tropical Indian Ocean, respectively, often co-occurs with developing El Ni˜no (e.g., Cai et al. 2011). Conversely, negative IODs tend to occur with La Nina events. There is, however, increasing evidence for a˜ significant IOD impact on the ITF (e.g.,Sprintall et al. 2009;
Liu et al. 2015; Sprintall and Revelard 2014; Pujiana et al.
2019). For instance,Pujiana et al. (2019) found up to 40%
reduction in ITF transport during the weak 2016 La Nina that˜ was attributed to the concurrent negative IOD, in which downwelling Kelvin waves associated with anomalous west- erly winds off Sumatra and Java led to higher-than-normal sea level at the ITF outflow region, thereby reducing the interbasin pressure gradient.
Our understanding of ITF variability is hampered by sparse observations over discontinuous periods of direct current mea- surement within the Indonesian seas since only the 1990s, pri- marily in the Makassar Strait (e.g.,Gordon et al. 2008;Susanto et al. 2012)}a key ITF channel that accounts for about 80% of the total ITF transport (Gordon et al. 2010,2019). As such, many investigations of ITF variability rely on estimates of ITF transport derived using observed temperature profiles across northwestern Australia and Java (Meyers 1996), on remotely sensed altimeter data (Sprintall and Revelard 2014), and on the use of numerical models, such as ocean reanalysis and state-of-the-art climate models (e.g.,Murtugudde et al. 1998;
England and Huang 2005;Potemra and Schneider 2007;van Sebille et al. 2014). The link between ITF transport and ENSO reported across these studies varies, with the strength of the correlation coefficients ranging from about 0.3 to 0.6. These dif- ferences should largely stem from the contrasting methodolo- gies, data products, and models used, but there are physical processes modulating ENSO impacts that may be captured to varying extents across studies. Using an intermediate complex- ity model, Murtugudde et al. (1998) found thatfixing winds over the Indian Ocean to climatology led to a dramatic increase in the ENSO–ITF correlation (from 0.31 to 0.65), illus- trating that Indian Ocean processes counteract the impact of
ENSO (see alsoMeyers 1996;Masumoto 2002).Potemra and Schneider (2007)showed using an ocean reanalysis and two coupled models that the ITF exhibits opposing transport anom- alies between the upper 100 m and the deeper layer (see also Potemra et al. 2003). An enhanced upper-layer transport was found to be associated with anomalous easterly winds south of Java. A reduced transport at the subsurface below 100 m was linked to anomalous westerlies in the equatorial Pacific and anomalous easterlies in the equatorial Indian Ocean, a typical condition during an El Ni˜no event. The latter also occurs dur- ing a positive IOD. Thus, the impact on the total depth-inte- grated ITF transport would not be as apparent than if the anomalous transports were of uniform polarity across depths.
The partitioning of ITF variability into upper and lower layers has also been shown in observation-based studies (Molcard et al. 1996, 2001; Sprintall et al. 2009; Susanto et al. 2012;
Sprintall and Revelard 2014;Gordon et al. 2019). For instance, during the weak 2006/07 El Nino, which coincided with a˜ strong positive IOD event, an enhanced and reduced south- wardflow was observed above and below 100 m, respectively, in both the Makassar Strait (Susanto et al. 2012) and the ITF outflow passages (Sprintall et al. 2009).
It is not yet clear to what extent ITF variability and its link to ENSO and IOD vary across climate models that are used for future climate projections. Climate models simulate a diverse range in the representation of ENSO and IOD, exhib- iting notable biases (e.g.,Saji et al. 2006;Cai et al. 2011;Liu et al. 2013; Taschetto et al. 2014; Jourdain et al. 2016;
McKenna et al. 2020). In particular, the overly strong IOD amplitude bias has persisted throughout generations of cli- mate models, due to an overly active Bjerknes feedback in the southeastern tropical Indian Ocean (Cai and Cowan 2013). In this study, we investigate how ITF variability and its interplay with ENSO and IOD are represented across models and how this may be affected by model biases. Our analysis shows that ITF variability is model dependent, strongly influ- enced by the simulated ENSO, the IOD, and their linkage.
The simulated ITF variability is shown to exhibit vertical structures which are distinctly linked to ENSO and the IOD.
We demonstrate that by examining these vertical structures, the impact of model biases can be better understood, and this is also important for understanding the response of ITF to cli- mate change.
The rest of the paper is organized as follows.Section 2pre- sents the implemented models and the analysis approach. The results are presented insection 3, covering ITF seasonal cycle, characteristics of the variability, relationships with ENSO and IOD, and intermodel correlations to reveal the factors that influence the simulated ITF variability.Section 4 concludes the paper with a summary and discussion on the ramification of our results on future ITF variability.
2. Data and methods
Our analysis focuses on the historical simulations of the 20 CMIP5 models listed inFig. 1. We analyze a common 93-yr historical period (1907–99) across the models to obtain the longest span without any missing data in our model archives.
As a qualitative comparison and to facilitate further under- standing of the CMIP5 models result, we also examine the SODA-2.2.4 ocean reanalysis (Giese and Ray 2011) from 1970 to 2008, while keeping in mind that a reanalysis itself would differ from direct observations. SODA-2.2.4 is based on the Parallel Ocean Program (POP) ocean model assimilating vari- ous observations, with an average resolution of 0.25830.48340 vertical levels, and the 5-day averaged output is mapped onto monthly averaged uniform 0.5830.58340 level grid (Carton and Giese 2008). For ease of calculation across several different models, which have varying geographical configuration of the Maritime Continent depending on model resolutions (seeFig. S1 andTable S1in the onlinesupplemental material), we define the ITF as the depth-integrated transport across ocean por- tions along a specified transect line between Sumatra and Australia (green line in Fig. 1a) that separates the Pacific
from the Indian Ocean. The bathymetry along this transect for all the models is provided inFig. S2. The transport inte- grated across this transect and over the full depth (i.e., ITF total transport) is equivalent to the difference in depth- integrated mass streamfunction between the easternmost and westernmost boundary of the ITF gateway, thus strictly defining the ITF based on mass conservation of the global ocean circulation (Santoso et al. 2011). The sign convention here is positive for an enhanced ITF and negative for a weak- ened ITF. The CMIP5 models simulate a multimodel mean of 15.3 Sv (1 Sv≡106m3s21) in ITF total transport, compa- rable to the SODA-2.2.4 reanalysis of 16.9 Sv.
To further investigate the characteristics and drivers of ITF variability, an empirical orthogonal function (EOF) analysis is applied on the integrated transport within each vertical layer to extract the vertical modes of ITF variability in each model (a) Transect location
Longitude
Latitude
ITFtotal 16.9 Sv (SODA) 15.3 Sv (CMIP5)
50 100 150
30S 20S 10S 0 10N 20N
1 2 3 4 5 6 7 8 9 10 11 12
−10
−5 0 5 10
(b) ITF
total annual anomaly
Calendar month
(Sv)
bcc−csm1−1 CanESM2 CCSM4 CNRM−CM5 FGOALS−g2 FGOALS−s2 GFDL−CM3 GFDL−ESM2G GFDL−ESM2M HadGEM2−CC HadGEM2−ES IPSL−CM5A−LR IPSL−CM5A−MR IPSL−CM5B−LR MIROC−ESM−CHEM MPI−ESM−LR MPI−ESM−MR MRI−CGCM3 NorESM1−M NorESM1−ME
(c) ITF annual anomaly (SODA)
Depth (m)
Calendar month
1 2 3 4 5 6 7 8 9 10 11 12
−1000
−800
−600
−400
−200
(d) ITF annual anomaly (CMIP5)
Calendar month
(x 10 −2 Sv m −1)
1 2 3 4 5 6 7 8 9 101112
−5
−2.5 0 2.5 5
FIG. 1. Annual cycle of depth-integrated ITF transport (ITFtotal) calculated across the transect in (a) presented as the anomaly in (b) over 1970–2008 for the SODA-2.2.4 reanalysis (thick black line) and for the historical period of 1907–99 for the CMIP5 models (ensemble mean in thick red line). Sign convention is positive for enhanced ITF trans- port. Colored thin curves correspond to colored model names. This color coding is used in the rest of all applicablefig- ures. (c),(d) The vertical structure of ITF transport annual cycle per unit depth for the reanalysis and CMIP5 multimo- del mean, respectively. The mean values of the ITF depth-integrated transport across the green transect line for the reanalysis and CMIP multimodel mean are indicated in (a). The transect line stretches from 08, 1038E to 178S, 1308E where the ITF calculation for all the models is done over ocean grid-points (see supplementalFig. S2for model bathymetry along this transect).
and reanalysis. To facilitate intermodel comparisons, ocean transport for all the models isfirst linearly interpolated onto the vertical levels of the SODA reanalysis. Given the uneven distribution of the data due to irregular vertical grid thickness, each anomaly matrix isfirst weighted by the thickness of the vertical grids prior to EOF computation. EOF analysis has been previously used to study observed ITF variability (Molcard et al. 1996,2001;Susanto et al. 2012), although the period of analysis was too short to sufficiently resolve interan- nual variations.
To reveal the drivers of ITF variability, we perform correla- tion analysis between ITF transport time series with climate variables such as sea surface temperature (SST), wind stress, and upper ocean heat content (averaged temperature above 300 m) within the SODA reanalysis and each of the CMIP5 models. For ENSO and IOD variability, we utilize the Nino-˜ 3.4 (SST anomalies averaged over 58S–58N, 1708–1208W) and Dipole Mode Index (DMI; difference in anomalous SSTs between 108S–108N, 508–708E and 108S–08, 908–1108E; Saji et al. 1999), respectively. ENSO events are defined as when the amplitude of Nino-3.4 anomalies averaged from December of˜ ENSO development year to the subsequent February (DJF) exceeds 0.5 standard deviation. IOD events are defined as when the amplitude of DMI anomalies averaged from September to November (SON) exceeds 0.5 standard deviation. The SODA Nino-3.4 and DMI time series (detrended with seasonal cycle˜ removed) are highly correlated to those of the NOAA Extended Reconstructed SST (ERSST) version 3b, with correlation coefficients of 0.98 and 0.85, respectively. The ENSO magnitude is comparable between the two reanalyses (Fig. S3a), but the SODA IOD amplitude is about 20% larger, although it is still at the low end of the CMIP5 models of which the multimodel mean magnitude is∼60% larger than in ERSST (Fig. S3b). Overly large IOD amplitude is a persistent bias over generations of climate models, with many of the CMIP5 models also exhibiting overly regular ENSO and IOD variability (McKenna et al. 2020) as indicated by sharp spectral peaks (Figs. S3c,d).
Unless stated otherwise, all time series have been detrended using a second-order polynomialfit with monthly climatologi- cal means (entire record as baseline) removed. Analyses focus- ing on interannual variability further incorporate a removal of 11-yr running mean and an application of a Butterworth low- pass filter to remove signals with periodicities shorter than 18 months. Statistical significance for the correlation of time series is evaluated based on consideration of the effective degrees of freedom arising from autoregressive properties of a geophysical time series (Davis 1977).
3. Results
a. General characteristics of ITF variability
The CMIP5 multimodel mean and the SODA reanalysis are consistent in exhibiting an apparent annual cycle of the ITF, reaching a maximum in austral winter and a minimum in austral summer (Fig. 1b), consistent with previous studies based on models and observations (e.g., Masumoto and
Yamagata 1996;Lee et al. 2010;Shinoda et al. 2012;Liu et al.
2015; Gordon et al. 2019). The CMIP5 ensemble exhibits notable intermodel spread (Fig. 1b): thepeak-to-troughampli- tude of seasonal cycle in the ITF total transport (Fig. 1b) ranges from 3.6 to 18 Sv, with a mean of∼9 Sv, similar to the SODA reanalysis. This intermodel range for the 20 CMIP5 models is notably larger than the 4.9–8.7-Sv range for the 14 ocean reanalysis products analyzed byLee et al. (2010), which is expected since CMIP models, unlike ocean reanalysis, are not constrained to assimilated observational data.
The ITF annual cycle is confined in the upper 100 m above the mean thermocline due to monsoonal forcing. Semiannual anomalies occur below, consistent with semiannual wind varia- tions in the equatorial Indian Ocean that generate Kelvin waves toward the Maritime Continent (e.g., Sprintall et al. 2000;
Iskandar et al. 2005;Lee et al. 2010;Susanto et al. 2012;Gordon et al. 2019). Upward phase propagation reveals a signature of downward penetration of Kelvin wave energy (e.g.,Sprintall et al. 2009;Drushka et al. 2010). Both the reanalysis and the CMIP5 multimodel mean exhibit such vertical structures in a strikingly close correspondence to each other (Figs. 1c,d).
Beyond the seasonal cycle, the ITF displays apparent inter- annual variability as highlighted inFig. 2a, which shows the ITF total transport time series in the SODA reanalysis with the seasonal cycle removed. Also visible inFig. 2is an increas- ing trend since the 1990s, linked to the strengthening of the tropical Pacific Walker circulation (e.g.,Feng et al. 2011) and the associated salinity effect (Hu and Sprintall 2017). The transport range is∼15 Sv, which is of similar order to that seen in observations at different survey sections (e.g.,Liu et al.
2015;Gordon et al. 2019). With the long-term trend removed, the SODA reanalysis exhibits a standard deviation of 2.7 Sv in the ITF total transport, while the CMIP5 models range from 0.8 to 3.1 Sv with a multimodel mean of 2.2 Sv, compara- ble to the reanalysis. Power spectral analysis reveals that the total ITF transport variability in the SODA reanalysis exhib- its peak variability with periods of 4–8 years per cycle (Fig. 2b), coinciding more with the ENSO time scale (3–5 years per cycle) than the IOD (2–4 years per cycle). This tendency is somewhat similar in the CMIP5 multimodel (Fig. 2b, thin curves). The power spectrum at each depth level for the reanalysis and CMIP5 multimodel mean (Figs. 2c,d) shows the majority of the variability being contained in the upper 100 m across a diverse range of time scales, with inter- annual variability becoming more dominant with depth. Cal- culating the power spectrum with transport variability at each depth normalized by its respective standard deviation reveals the dominant time scales (Figs. 2c,d, contours). Specifically, the interannual variability is most dominant at 100–200 m, with intraseasonal variability being prominent within the sur- face layer (0–50 m). The vertical structure of ITF variability is relevant for understanding the impact of ENSO and IOD as discussed next.
b. Link with ENSO and IOD
We first note that the year-to-year variability of the ITF total transport in the SODA reanalysis exhibits temporal
behavior that is broadly consistent with observations across the IX1 transect that provide a geostrophic transport estimate based on expendable bathythermograph (XBT) temperature records (Meyers 1996;Liu et al. 2015;Feng et al. 2018). Rela- tive to existing surveys at other choke points (Molcard et al.
1996;Sprintall et al. 2009;Gordon et al. 2019), the IX1, which stretches from Fremantle, Western Australia, to the Sunda Strait, Indonesia, would be the closest to our transect (Fig. 1a;
seeFig. S4a) and has the longest record dating back to 1983.
Feng et al. (2018)presented the observed time series of the IX1 geostrophic transport anomalies referenced to 700 m (their Fig. 2). Although an exact match with our time series (Fig. 2a) is not expected, there is notable agreement between the two, with transport anomalies appearing to coincide with ENSO phases [cf. our Fig. S4b with Fig. 2 of Feng et al.
(2018)]: anomalously high transports occurred in 1988/89, 1995/96, 1999–2001, early 2006, and 2007/08, coincident with La Ni˜na conditions, whereas anomalously low transports in 1986–88, 1991/92, 1997/98, and 2002–05 were concurrent with El Nino conditions.˜
The tendency for a weaker and stronger ITF to occur with El Ni˜no and La Ni˜na, respectively, is marked by a statistically significant negative correlation inFigs. 3aand3c. The CMIP5 models overall capture the expected link between ENSO and ITF, albeit with large intermodel differences, namely a weaker ITF corresponding with an El Ni˜no and a stronger
ITF corresponding with a La Ni˜na. The CMIP5 multimodel averaged lag correlation between the Ni˜no-3.4 index and ITF total transport (ITFtotal;Fig. 3a) reveals a maximum correla- tion coefficient (r) of about20.30, comparable to the SODA reanalysis (r5 20.38; statistically significant above the 95%
confidence level), and the correlations double when consider- ing only variability on interannual time scales (Fig. 3c). The CMIP5 intermodel spread in the correlations is notably large, with a range of about 0.4.
The relationship between ITFtotal with the DMI, on the other hand, is not statistically significant for either the CMIP5 multimodel mean or the reanalysis (Figs. 3b,d), even though IOD events are associated with large atmospheric and oceanic anomalies that affect sea level at the ITF outflow passages (Sprintall and Revelard 2014). The apparent lack of a link between ITFtotaltransport and IOD can be expected to be a result, at least in part, of a counteracting effect of ENSO.
During a positive IOD event, which peaks in austral spring, anomalous surface-layer cooling occurs off Java and Sumatra, and thus a lower sea level at the ITF outflow region, which tends to increase the throughflow (e.g.,Sprintall and Revelard 2014); as such, the IOD should be positively correlated to the ITF, but it is not the case inFigs. 3band3d. However, a posi- tive IOD often coincides with an El Nino, through anomalous˜ atmospheric subsidence over the Maritime Continent as the Indo-Pacific Walker circulation weakens. This ENSO–IOD Year
(Sv)
(a) ITF total
70 74 78 82 86 90 94 98 02 06 5
10 15 20 25
Period (yr/cycle)
PSD (standardized unit)
(b) Power spectra
ITFtotal
Nino3.4 DMI
8 4 2 1 0.5
0 10 20 30
Period (yr/cycle)
Depth (m)
1612 6 2
4 4
4 6
2
6 8
(c) ITF power spectrum (reanalysis)
(x10−4 Sv2 cpy−1)
8 4 2 1 0.5
−600
−500
−400
−300
−200
−100 0
0 1 2 3 4
>5
8 4 12
2 2
8 4
4
2
2
(d) ITF power sepctrum (CMIP5)
Period (yr/cycle)
8 4 2 1 0.5
−600
−500
−400
−300
−200
−100 0
FIG. 2. (a) Time series of ITFtotalwith monthly climatological mean removed (black) in SODA-2.2.4. The detrended time series is shown in blue, highlighting the early-twenty-first-century enhancement of ITF transport shown in black.
(b) Power spectrum of standardized ITFtotal(black), Nino-3.4 (red), and DMI (blue) time series in the SODA reanaly-˜ sis (thick curves) and CMIP5 ensemble mean (light colored thin curves). (c) Power spectrum of ITF time series at each depth level in the SODA reanalysis (red shading), with the power spectrum based on standardized time series shown in purple contours. The transport time series isfirst detrended with monthly means removed. (d) As in (c), but for the power spectrum averaged across the 20 CMIP5 models.
co-occurrence is reflected in a positive correlation between the Nino-3.4 and DMI of˜ r50.49 using the SODA monthly data (r50.39 for ERSST), significant above the 99% level, when the DMI leads Nino-3.4 by 2 months (r˜ ∼0.7 if correlat- ing September–November averaged DMI and December– February averaged Nino-3.4 for both reanalyses). Such co-˜ occurring tendency means that the positive IOD-enhanced ITF tends to be counteracted by El Nino˜ –induced ITF weakening via an anomalously low western Pacific sea level associated
with weaker Pacific trade winds; the converse applies for negative IOD and La Nina co-occurrences. The signi˜ ficant negative ENSO-ITFtotalcorrelation and the absence of a sig- nificant positive IOD–ITFtotal correlation (Fig. 3), which would otherwise indicate a dominating IOD effect, suggests the overall dominance of the ENSO forcing on the ITF total transport.
The counteracting interplay between ENSO and IOD on the ITF involves opposing transport anomalies between the
Correlation coef.
(a) ITF
total vs Nino3.4
−12 0 12
−1
−0.75
−0.5
−0.25 0 0.25 0.5
(b) ITF
total vs DMI
−12 0 12
−1
−0.75
−0.5
−0.25 0 0.25 0.5
Correlation coef.
(c) ITF
total vs Nino3.4 (interannual)
Lag (month)
−12 0 12
−1
−0.75
−0.5
−0.25 0 0.25 0.5
(d) ITF
total vs DMI (interannual)
Lag (month)
−12 0 12
−1
−0.75
−0.5
−0.25 0 0.25 0.5
FIG. 3. (a) Correlations between ITFtotaland Nino-3.4 as a function of lag time in months for˜ the SODA reanalysis (thick black line) and CMIP5 models (thin colored curves; ensemble mean in thick red line). (b) As in (a), but for DMI. Positive (negative) time lags indicate Nino-3.4 and˜ DMI leading (lagging) ITF time series. (c),(d) As in (a) and (b), respectively, but based onfil- tered time series to isolate interannual variability. Correlation coefficient cut-off values for the statistical significance at the 95% confidence level are indicated by red and black dashed horizon- tal lines, respectively, for the reanalysis and CMIP5 models.
surface and subsurface layers, as revealed by compositing transport anomalies according to ENSO and IOD phases (Fig. 4). For both the CMIP5 ensemble (Fig. 4a) and reanal- ysis (Fig. 4e), anomalously weak transport occurs in the sub- surface during an El Nino, peaking at around 100˜ –200-m depth where interannual variability dominates (Figs. 2c,d).
A signature of upward propagation of anomaly is apparent, which is analogous to that occurring on subannual time scales (Figs. 1c,d), transmitting the subsurface anomaly pro- gressively toward the surface over about 6 months (see also Sprintall et al. 2009). Upward propagation is also evident in Makassar Strait current observation from 2004 to 2017 (Gordon et al. 2019) which shows transport over 300–760 m leading that in the upper 0–300-m layer. This upward
propagation feature contributes to a tendency for a longer lagged response of the ITF to ENSO toward the surface as seen in the reanalysis (supplemental Figs. S5 and S6), a response that varies across the CMIP5 models, which will be discussed insection 3c. Upward propagation has been sug- gested to be indicative of downward penetration of Kelvin wave energy (McCreary 1984) associated with zonal wind forcing, which has strong intraseasonal component in the equatorial Indian Ocean. However, the sequence of interan- nual climate forcing could also contribute as illustrated below.
Overlying the anomalously weak subsurface transport is an anomalously enhanced surface-layer transport. A similar ver- tical dipole pattern is seen in the positive IOD composite (Figs. 4b,f), with the surface intensification, which is particularly 28.9
(a) El Nino
CMIP5
Depth (m)
2 6 10 14 18 22
−500
−400
−300
−200
−100 0
28.35
(b) positive IOD
CMIP5
(x10−2 Sv m−1) 2 6 10 14 18 22
−1 −0.5 0 0.5 1
31.5 (c) La Nina
CMIP5 2 6 10 14 18 22
−500
−400
−300
−200
−100 0
31.3
(d) negative IOD
CMIP5 2 6 10 14 18 22
12 (e) El Nino
SODA
Depth (m)
Month 2 6 10 14 18 22
−500
−400
−300
−200
−100 0
10
(f) positive IOD
SODA Month 2 6 10 14 18 22
13
(g) La Nina
SODA Month 2 6 10 14 18 22
−500
−400
−300
−200
−100 0
14
(h) negative IOD
SODA Month 2 6 10 14 18 22
FIG. 4. Composites of interannual transport anomalies at each depth level (i.e., in units of Sv m21) according to (a) El Nino events,˜ (b) positive IOD events, (c) La Nina events, and (d) negative IOD events, averaged across the 20 CMIP5 models. (e)˜ –(g) As in (a)–(d), but for the SODA reanalysis. Shading indicates composites that are significant above the 90% confidence level. The composites are shown from January of the ENSO and IOD development year (months 1–12) to December of the following year (months 13–24). ENSO and IOD respectively peak around December to February (months 12–14) and September to November (months 9–11). Value in the bottom left of each panel indicates the respective number of events, of which the proportion is overall comparable between the CMIP5 and reanal- ysis relative to the different record lengths (93 years for CMIP5, 39 years for reanalysis).
prominent in the CMIP5 models, being notably more pro- nounced compared to the El Nino composite. These tendencies,˜ but in the opposite sense, are overall also seen in the La Nina˜ and negative IOD composites (Figs. 4c,d,g,h). Such structure of opposing anomalous transports between the surface and subsur- face has been identified in limited observations, particularly during concurrent ENSO and IOD phases (Sprintall et al. 2009;
Susanto et al. 2012; Sprintall and Revelard 2014). This has been suggested to be associated with large-scale atmospheric divergence/convergence over the Maritime Continent during ENSO–IOD concurrence (Potemra and Schneider 2007).
During a positive IOD (and a strong El Ni˜no, which tends to induce a positive IOD condition), easterly wind anomalies prevail during boreal summer to autumn in the eastern tropi- cal Indian Ocean, which leads to upwelling off Java and Sumatra in turn locally enhancing the easterly anomalies (e.g.,Schott et al. 2009). Through the associated lowered sea level height and enhanced southward Ekman transport, an increased ITF transport occurs in the upper layer. Downwel- ling Kelvin waves tend to occur below, though intermittently (e.g.,Horii et al. 2008), and can contribute to transport reduc- tion at depth (e.g.,Potemra and Schneider 2007). In the west- ern Pacific, upper-ocean heat content decrease linked to El Nino˜ –induced westerly winds in the equatorial Pacific, which peaks following El Nino maturity (e.g.,˜ McPhaden et al.
2020), leads to subsurface transport reduction. Reduced trans- port can be further prolonged at the surface through an increase in heat content off Sumatra and Java associated with the swing toward the opposite phase of the IOD (Feng and Meyers 2003). This, coupled with the tendency for the effect of ENSO and IOD to cancel each other out during ENSO developing phase, may explain the lagged response of the depth-integrated ITF to ENSO seen in observations as noted byLiu et al. (2015)andFeng et al. (2018). In several models (Figs. 3a,c) though, the ITFtotal instead leads Nino-3.4 by˜ 3–6 months. This is associated with a more robust deep trans- port response in these models that appears to peak during ENSO developing phase (around August;Fig. 4), and a pro- longed surface anomaly associated with the delayed IOD bias (discussed insection 3c). The latter would have a persistent counteracting effect on subsurface anomaly extending into ENSO decay phase.
When ENSO co-occurs with IOD events (i.e., El Nino co-˜ occurring with positive IOD, or La Ni˜na co-occurring with negative IOD), the transport anomaly amplitude tends to be comparable between the surface and subsurface (Fig. S5), as opposed to the stronger surface anomaly during the IOD (Fig. 4). Taken together, these results suggest that the IOD and ENSO have a stronger impact at the surface and subsur- face, respectively. The weaker subsurface impact of the IOD might be attributed to the strong presence of intraseasonal Kelvin waves in the equatorial Indian Ocean (Horii et al.
2008; Iskandar et al. 2005; Drushka et al. 2010), which can lead to destructive interference between upwelling and downwelling anomalies. Support for an IOD impact on sur- face transport is underscored by the pervasiveness of the sur- face anomalies in the CMIP5 IOD composite (Figs. 4b,d) which, relative to the reanalysis (Figs. 4f,h), potentially
indicates an impact of the overly strong IOD amplitude bias (see below). In addition, the CMIP5 composite for positive IOD events without El Ni˜no, which is a rare occurrence in observations, exhibits even stronger surface transport intensi- fication, with no appreciable anomalies at depth (figure not shown).
Some asymmetry between positive and negative climate phases is noticeable, especially in the reanalysis, although this still points to the distinct impact between ENSO and IOD dis- cussed above. For instance, the negative surface anomalies during La Nina are not statistically signi˜ ficant (Fig. 4g), unlike in the El Ni˜no case. The positive subsurface anomalies in the negative IOD composite are notably weaker than in the La Ni˜na composite (cf. Figs. 4g,h), while in the positive IOD counterpart they are comparable to the El Nino composite˜ (cf.Figs. 4e,f). This asymmetry partly stems from fewer co- occurrences of negative IOD events with La Ni˜na events than co-occurrences of positive IOD events with El Ni˜no events, over the reanalysis period used here in which all positive IOD events co-occurred with an El Nino (Fig. 4f; see also˜ Fig. S5c).
Nonlinearity in the general properties of ENSO and IOD is also expected to play a role, although this may be less appar- ent in climate models, which overall tend to underestimate the observed nonlinearity (e.g.,McKenna et al. 2020). Investi- gation into such asymmetry is beyond the scope of this paper, and here we focus on the distinct impact of ENSO and IOD in general.
To further characterize the vertical structures of ITF vari- ability and examine the associated mechanisms, we perform an empirical orthogonal function (EOF) decomposition on transport anomalies above 1200 m capturing most of the transport variability, which intensifies toward the surface (Fig. 2). The patterns of anomalous SST, zonal wind stress, and upper ocean heat content associated with these EOFs are then assessed. The analysis was applied to the CMIP5 and reanalysis data with bandpassfiltering to isolate processes on interannual time scales (Fig. 5). For comparison, an analysis based on raw data with the long-term trend and seasonal cycles removed was also conducted (Fig. S7). The first (EOF1) and second (EOF2) modes together explain more than 80% of the total variability (Fig. 5a; see alsoFig. S7a) and are thus the focus of our discussion, focusing on variabil- ity over interannual time scales. These two principal struc- tures are analogous to those extracted using EOFs on 1-yr- long current observations in two key ITF outflow passages, the Timor Passage (March 1992–April 1993; Molcard et al.
1996) and Ombai Strait (December 1995–November 1996;
Molcard et al. 2001).
The CMIP5 multimodel mean and the SODA reanalysis (Figs. 5b,c; see alsoFigs. S7b,c) show that EOF1 exhibits sur- face-intensified anomalous transport (hereafter referred to as ITFM1), while EOF2 exhibits opposite anomalies between the surface and subsurface (ITFM2). Correlating the respective principal component (PC) time series against Nino-3.4 and˜ DMI reveals that the link between ITF transport with both ENSO and IOD is captured by ITFM2. The reanalysis and CMIP5 multimodel mean are consistent in exhibiting high correlation coefficients that are significant above the 95%
confidence level (Figs. 5f,g; see alsoFigs. S7f,g). This dem- onstrates that the baroclinic structure, characterized by anomalously strong surface transport accompanied by weak subsurface transport, is associated with El Ni˜no and positive IOD events (and the converse for La Nina and negative˜ IOD events). In the case of ITFM1though, which appears to be equivalent barotropic in structure, there is apparent dis- agreement between the reanalysis and the CMIP5 models (Figs. 5d,e; see alsoFigs. S7d,e). In the reanalysis, ITFM1and ENSO are linked, with El Nino leading to weaker-than-nor-˜ mal surface-intensified transport about 6 months later. In the CMIP5 multimodel mean, there is no such clear relationship, given the large intermodel differences. Conversely for the IOD, the CMIP5 models consistently simulate anomalously
strong surface-intensified transport associated with a positive IOD (Fig. 5e), but no such relationship is found in the reanal- ysis. This contrast between CMIP5 and reanalysis is likely due to a combination of spurious IOD influence in CMIP5 models associated with the overly strong IOD bias, and a lack of ENSO-independent IOD events in observations.
However, agreement between the CMIP5 and reanalysis is seen in the unfiltered data (Fig. S7e), thus indicating that ITFM1is also associated with intraseasonal processes in the Indian Ocean, and that the disagreement is on interannual time scales.
Lag correlations between PC time series and grid point zonal winds and upper-ocean heat content on interannual time scales (Fig. 6) reveal consistency between the reanalysis
1 2 3 4 5 6
0 20 40 60 80
Mode
(%)
(a) Variance explained
−1 0 1 2
−600
−500
−400
−300
−200
−100 0
(b) EOF1 (ITFM1)
Depth (m)
(x10−2 Sv m−1)
−1 −0.5 0 0.5 1
−600
−500
−400
−300
−200
−100 0
(c) EOF2 (ITFM2)
(x10−2 Sv m−1)
−18−12−6 0 6 12 18
−1
−0.5 0 0.5 1
Correlation coef.
Lag (month) (d) Nino3.4 vs PC1
−18−12−6 0 6 12 18 Lag (month) (e) DMI vs PC1
−18−12−6 0 6 12 18 Lag (month) (f) Nino3.4 vs PC2
−18−12−6 0 6 12 18 Lag (month) (g) DMI vs PC2
FIG. 5. First two leading EOF modes of the vertical profile of interannual ITF transport variability and link with ENSO and IOD.
(a) Percentage of variance explained by each mode of variability. Black circles indicate those for the SODA reanalysis, and colored crosses for the CMIP5 models. The (b)first (ITFM1) and (c) second (ITFM2) modes are shown as the regression of transport per unit depth onto the corresponding principal component (PC) time series. (d)–(g) Lag correlation between the corresponding PC time series with the cli- mate indices. Positive lags indicate climate indices leading ITFM1and ITFM2. Thick black and thick red lines indicate quantities for the reanalysis and CMIP5 multimodel mean, respectively. Correlation coefficient cut-off values corresponding to statistical significance at the 95% confidence level are indicated by black and red dashed horizontal lines respectively for the reanalysis and CMIP5 multimodel mean.
and CMIP5 multimodel mean for the forcing of ITFM2, which is underpinned by large-scale anomalous divergence/
convergence over the Indo-Pacific warm pool (Figs. 6b,d).
The association implies that interannual transport variabil- ity marked by anomalously strong surface transport and anomalously weak subsurface transport, as captured by ITFM2, is linked to easterly wind anomalies across the equa- torial Indian Ocean and westerly anomalies across the equa- torial Pacific, consistent with the results of Potemra and Schneider (2007)based on a reanalysis and two models. The atmospheric divergence reduces upper-ocean heat content, and thus a lowered sea level and shoaled thermocline, across the western Pacific to eastern Indian Ocean, and increases heat content in the eastern Pacific and western Indian Ocean. Such conditions are characteristic of a co- occurrence between El Nino and a positive IOD (the˜
converse for La Nina and negative IOD), as indicated in the˜ correlation pattern for SSTs (Figs. S8b,d).
For the ITFM1 interannual forcing, the reanalysis and CMIP5 multimodel mean are not consistent with each other (Figs. 6a,c). In the reanalysis, stronger and weaker surface intensified transports are respectively associated with anoma- lous easterly and westerly winds in the equatorial Pacific that occur several months prior. This suggests that the reanalysis ITFM1is a lagged response to ENSO, as indicated by a La Ni˜na–like pattern at 6-month lead time inFig. S8c. In con- trast, the CMIP5 ITFM1is associated with anomalous winds and upper-ocean heat content changes within the tropical Indian Ocean, thus suggesting that ITFM1(i.e., surface-inten- sified transport variability) in CMIP5 models tends to be asso- ciated with processes internal to the Indian Ocean, evidently the IOD (Fig. 5e; see alsoFig. S8a).
6 month
−0.2 (a) ITF
M1 τx, HC
CMIP5
20S 0 20N
−0.4
−0.2 −0.2
0.6 0.4 0.2 (b) ITF
M2 τx, HC
CMIP5
4 month
−0.4
−0.2
20S 0 20N
−0.4
−0.2
−0.4
−0.2 0.6
0.2 0.4
2 month
−0.4
−0.2
20S 0 20N
−0.4
−0.2
−0.4 0.4 −0.2 0.2
0 month
−0.4
−0.2
20S 0 20N
−0.2
−0.4 0.4
−2 month
−0.4
−0.2
50E 150E 110W
20S 0 20N
−0.2
−0.2 0.2
50E 150E 110W
−0.5
−0.3 0.3 (c) ITF
M1 τx, HC
SODA
−0.3
−0.3 0.5 0.3
0.3 (d) ITF
M2 τx, HC
SODA
−0.3−0.5
0.3 −0.5
−0.3
0.7 0.3 0.3
−0.5
−0.3
−0.7
−0.5
−0.3 −0.5
0.7 0.3
−0.3
−0.3 0.3
−0.7
−0.3
−0.3 0.7 0.3
0.3
−0.5
−0.3
50E 150E 110W
−0.3 −0.7
−0.5
−0.3 0.5
0.3
50E 150E 110W
−1 −0.5 0 0.5 1
FIG. 6. Correlation patterns of gridpoint zonal wind stress (tx; contours) and upper-ocean heat content (HC; color shading) against (a),(c) ITFM1and (b),(d) ITFM2in (a),(b) the CMIP5 multimodel mean and (c),(d) reanalysis at different lag times. Black and purple con- tours mark positive and negative correlations betweentxand ITF anomalies, respectively. Positive lags indicate ITFM1and ITFM2lagging txand heat content anomalies. Only correlation coefficients that are statistically significant at the 95% confidence level are shown. Analysis is based on interannual time series.
Thus, the difference between the CMIP5 models and the SODA reanalysis is largely manifested in ITFM1 transport variability and the associated forcing. This disagreement applies exclusively for interannual time scale processes, as the unfiltered data show a closer agreement between the CMIP5 models and reanalysis, in particular between ITFM1and IOD (Fig. S7e). In the unfiltered data, in association with ITFM1, both CMIP5 and reanalysis indicate localized zonal wind anomalies just south of the Maritime Continent between Java and northern Australia (Figs. S9a,c), with weak anomalous cooling in the eastern equatorial Indian Ocean signifying a positive IOD (Fig. S10). Associated with anomalous easterlies in this region, upper-ocean heat content is reduced south of Java, contributing to an increased ITF. The local wind is typi- cally rich in high-frequency variability. Indeed, unlike ITFM2, which is dominated by interannual variability, ITFM1contains significant intraseasonal variability (Fig. S11), and so does the DMI (Fig. 2b). Thus, the most apparent discrepancy between the reanalysis and CMIP5 is attributed to the interannual vari- ability of surface-intensified transport as depicted by ITFM1. c. Intermodel relationships
In this section, we investigate the CMIP5 intermodel differ- ences by examining intermodel correlations across various variables. First, there is a strong link between the magnitude of interannual variability in the ITF total transport (ITFtotal) across the models with the simulated ENSO amplitudes: mod- els simulating stronger ENSO producing stronger interannual variability in ITFtotal(r50.70;Fig. 7a). The intermodel corre- lation is not statistically significant however when using the unfiltered data (Table 1), underscoring the role of higher-fre- quency variability. In both unfiltered andfiltered (i.e., interan- nual) cases, the intermodel correlation is significant for the DMI amplitude (Table 1), given substantial intraseasonal component in the DMI (Fig. 2b). The ITFtotal variability is also found to be positively correlated to the magnitude of zonal wind variability south of the Maritime Continent (txIO) (Table 1). To assess for a possible influence of model outliers on the results, we recalculate the correlations by excluding two models, GFDL-ESM2M and MIROC-ESM-CHEM, which respectively simulate the strongest and weakest ampli- tude of both ENSO and IOD. The choice of these two excluded models (hereafter referred to as EXM) in terms of exhibiting the extreme ends of ENSO and IOD amplitude constitutes a more stringent test, given the focus is on the impact of ENSO and IOD itself.
As suggested by the analysis insection 3b, ENSO exerts a stronger imprint on ITF transport variability in the subsurface than in the surface layer, which in contrast is more strongly linked to the IOD. This is reflected in the intermodel relation- ships: stronger ENSO and IOD amplitudes across models tend to correspond with stronger interannual variability in the subsurface (100–300 m; ITF100–300) and surface layer (0–100 m; ITF0–100), respectively (Figs. 7c–f). The correlations are statistically significant above the 95% confidence level.
Without EXM, the strengths of the linkage decrease, but the relative importance between the ENSO and IOD influence
on surface and subsurface transports remains robust. Note that in the unfiltered data the relationship with IOD ampli- tude is stronger for ITF100–300than ITF0–100(Table 1). This however stems from the tendency for models that simulate stronger ENSO amplitude to also produce stronger IOD vari- ability (Table 1). The association is stronger in the unfiltered data, indicating that such tendency may not necessarily be solely due to the coupling between ENSO and IOD but also to other factors such as the level of stochasticity in the models that can influence the variability of both ENSO and the IOD.
ENSO and the IOD are linked through the atmospheric Walker circulation, and this interaction in turn affects atmo- spheric circulation particularly in the vicinity of the Maritime Continent over which anomalous atmospheric convergence and divergence due to ENSO and IOD occur. As such, there is a significant intermodel correlation between the magnitude of local wind forcing (txIO) and the magnitude of both ENSO and IOD variability (Table 1). The strength oftxIOvariability is in turn strongly linked to the amplitude of ITF variability across the models, particularly in the surface layer (ITF0–100), which is more directly impacted by surface winds. Local wind however has its own internal variability independent of ENSO and IOD; for example, the correlation between the magnitude oftxIOvariability and that of ENSO and IOD is not statistically significant without EXM (Table 1). IOD variability also con- tains ENSO-independent components, and the extent varies across models. This is reflected by the large intermodel range in the correlation between DJF-average Ni˜no-3.4 and SON- average DMI, ranging from near 0 (MIROC-ESM-CHEM) to about 0.8 (CNRM-CM5), with the upper range being closer to the reanalysis correlation of 0.7 (e.g.,Fig. 8e;yaxis). Further, while there is a statistically significant relationship between ENSO amplitude and ENSO–IOD correlation across models (r50.65;pvalue,0.01), implying that ENSO is to a certain extent a driver of the IOD, this relationship is sensitive to model sampling as the correlation drops tor50.38 (pvalue5 0.12) without EXM. Nonetheless, given the impact of interplay between ENSO and IOD on ITF variability (section 3b), the varying degree of ENSO–IOD linkage across models can influence the simulated ITF variability, as further illustrated below.
In the real system, the IOD tends to lead ENSO by about a season (Fig. 8a), reflecting the peak of IOD in boreal autumn preceding ENSO maturity in winter. However, in 15 out of the 20 CMIP5 models, it tends to be the converse, with El Ni˜no and La Ni˜na respectively leading positive and negative IOD by an average of 3 months (Fig. 8a). This bias seems to be related to the simulated ENSO teleconnection, rather than simply the representation of the IOD itself, as the delayed bias is much more apparent in the multimodel composite of monthly DMI according to El Ni˜no and La Ni˜na phases (Fig. 8b) than the composite based on IOD events (Fig. 8c).
Previous studies have pointed out that positive IOD events in models can occur following El Nino events or persist longer˜ than observed (Cai et al. 2005;Saji et al. 2006;Cai et al. 2011;
Jourdain et al. 2016). In the extreme case,Cai et al. (2005) found in their model that positive IOD events often occur three seasons following El Ni˜no events. They suggested that
this was a consequence of crudely representing Java–Timor topography as a single zonal landmass that would lead to spu- rious intrusion of El Nino˜ –induced upwelling Rossby waves from the Pacific into the southern coast of Java. In the real ocean, these waves propagate off the coast of northwestern Australia. There is an indication that model resolution indeed plays a role: models with the coarsest horizontal resolution
(CanESM2, the three IPSL models, MIROC-ESM-CHEM, and MPI-ESM-LR; see Table S1) exhibit strong delayed IOD bias (3–6 months). The longer persistence of the simu- lated IOD can also be attributed to persistent El Ni˜no (e.g.,Jourdain et al. 2016). We find that the delayed IOD occurrence relative to ENSO (i.e., a positive ENSO–IOD time lag) contributes to the overall longer persistence of the
(a) ENSO vs ITF
total
variability
ITF total std dev (Sv)
r=0.70 (0.00)
0.5 0.75 1 1.25 1.5 0
0.5 1 1.5 2
(c) ENSO vs ITF
0−100
variability
ITF 0−100 std dev (Sv)
r=0.41 (0.07)
0.5 0.75 1 1.25 1.5 0.25
0.5 0.75 1 1.25 1.5
(e) ENSO vs ITF
100−300
variability
ITF 100−300 std dev (Sv)
r=0.75 (0.00)
Nino3.4 std dev (
°C)
0.5 0.75 1 1.25 1.5 0.25
0.5 0.75 1 1.25 1.5
(b) IOD vs ITF
total
variability
r=0.58 (0.01)
0.1 0.2 0.3 0.4 0.5 0.6 0
0.5 1 1.5 2
(d) IOD vs ITF
0−100
variability
r=0.52 (0.02)
0.1 0.2 0.3 0.4 0.5 0.6 0.25
0.5 0.75 1 1.25 1.5
(f) IOD vs ITF
100−300
variability
r=0.41 (0.08)
DMI std dev (
°C)
0.1 0.2 0.3 0.4 0.5 0.6 0.25
0.5 0.75 1 1.25 1.5
FIG. 7. Intermodel relationships between ENSO and IOD amplitude and ITF variability: (a),(b) total transport, (c),(d) transport between surface to 100 m (ITF0–100), and (e),(f) transport between 100 and 300 m (ITF100–300). ENSO amplitude is measured as standard deviation of Nino-3.4 index. IOD amplitude measured as standard deviation of the˜ DMI. Analysis is based on data detrended with seasonal mean removed andfiltered to focus on interannual variability.
Multimodel mean is indicated with a red square and reanalysis with a black square. Correlation coefficients significant above the 95% confidence level are indicated in boldface;pvalues are shown in parentheses.
simulated IOD than the reanalysis (measured as the time when the DMI autocorrelation crosses zero). The intermodel correla- tion (r50.66;r50.71 without EXM) is significant above the 99% confidence level (Fig. 8d). This positive ENSO–IOD time lag also explains why the ENSO–IOD synchronous correlation across the models appears to be weaker overall compared to reanalysis (Fig. 8e), as underscored by a high intermodel
correlation ofr5 20.74 (r5 20.61 without EXM) significant above 99% confidence level. On the other hand, it also reflects a stronger tendency for positive IOD events to occur in the year following El Ni˜no events in the models (Fig. 8f).
The prolonged/delayed IOD in the models is expected to affect how ENSO influences the ITF, especially in the surface layer where the impact of IOD is most prominent. Because TABLE1. Intermodel correlation of variability amplitude between pairs of indices that represent ENSO (Ni˜no-3.4), IOD (DMI), zonal wind stress south of the Maritime Continent (txIO; 158–58S, 1058–1408E), and ITF transports integrated across entire depth (ITFtotal), over top 100 m (ITF0–100), and at 100–300 m (ITF100–300). The amplitude is taken as standard deviation of each index.
Shown are correlations based on filtered and unfiltered data. Values in brackets are correlations excluding GFDL-ESM2M and MIROC-ESM-CHEM models that respectively exhibit the strongest and weakest ENSO and IOD amplitudes. Correlation coefficients outside and inside brackets that are significant above the 95% and 90% confidence levels, respectively, are shown in boldface.
Filtered Unfiltered
Amplitude Ni˜no-3.4 DMI txIO Ni˜no-3.4 DMI txIO
ITFtotal 0.70[0.60] 0.58[0.44] 0.70[0.58] 0.24 [0.15] 0.47[0.45] 0.71[0.72]
ITF0–100 0.41 [0.10] 0.52[0.36] 0.78[0.73] 0.15 [20.02] 0.47[0.43] 0.86[0.88]
ITF100–300 0.75[0.58] 0.40 [0.11] 0.68[0.54] 0.58[0.36] 0.57[0.40] 0.70[0.65]
Ni˜no-3.4 0.65[0.43] 0.67[0.38] 0.74[0.58] 0.40 [0.06]
DMI 0.59[0.37] 0.59[0.46]
(a) Nino3.4 vs DMI (interannual)
Lag (month)
Correlation coef.
−18 −12 −6 0 6 12 18
−0.5
−0.25 0 0.25 0.5 0.75
1 (b) DMI composite
(° C)
Month
El Nino
La Nina 2 4 6 8 10 12 14 16 18 20 22 24
−0.3
−0.2
−0.1 0 0.1 0.2 0.3
(c) DMI composite
(° C)
Month pIOD
nIOD 2 4 6 8 10 12 14 16 18 20 22 24
−0.6
−0.4
−0.2 0 0.2 0.4 0.6
IOD persistence (month)
Lag Nino3.4 DMI (month) (d)
r=0.66 (0.00)
−4 −2 0 2 4 6 8
7 9 11 13
Lag Nino3.4 DMI (month) Cor Nino3.4 DJF DMI SON
(e)
r=−0.74 (0.00)
−4 −2 0 2 4 6 8
0 0.2 0.4 0.6 0.8
Lag Nino3.4 DMI (month)
Cor Nino3.4 DJF DMI SON+1 (f)
r=0.56 (0.01)
−4 −2 0 2 4 6 8
−0.4
−0.2 0 0.2 0.4 0.6 0.8
FIG. 8. (a) Lag correlation between ENSO and IOD. Positive lag indicates ENSO leading IOD, and vice versa for negative lag. Red curve indicates CMIP5 multimodel mean, solid black for SODA, and dashed black for ERSST. (b) Composite of DMI according to El Nino and La Ni˜ na phases indicated in red and blue, respectively. (c) Composite of DMI according to positive IOD and negative IOD˜ phases indicated in red and blue, respectively. In (b) and (c), solid (dashed) curves indicate reanalysis (CMIP5 multimodel mean) with filled circle, empty circle, and triangle markers indicating statistical significance above the 90% confidence level for SODA, ERSST, and CMIP5 multimodel mean, respectively. (d)–(f) Intermodel relationship between ENSO–IOD time lag and (d) IOD persistence, (e) correla- tion coefficients of DJF average Nino-3.4 vs SON average DMI, and (f) correlation coef˜ ficients of DJF average Nino-3.4 vs following year˜ SON average DMI. ENSO-IOD time lag is defined as the time (in months) at which the correlation between Ni˜no-3.4 and DMI [in (a)]
reaches a positive maximum. IOD persistence is defined as the time at which the autocorrelation of the DMI crosses zero. Red square indi- cates CMIP5 multimodel mean, black square for SODA, and empty circle for ERSST. All analyses are based on interannual time series.
El Ni˜no and La Nina correspondingly lead to a weaker and˜ stronger surface transport in the following year (Fig. 4), these anomalies would tend to be counteracted by the pro- longed/delayed effect of positive and negative IODs, respec- tively (see section 3b). Indeed, the longer the IOD lags ENSO, the weaker the ENSO influence is on the ensuing ITF0–100(Fig. 9a), underscored by an intermodel correlation of 0.61 (r50.76 without EXM) significant above the 99%
confidence level. Specifically, this means that the longer a positive IOD lags an El Nino, the weaker the in˜ fluence of El Ni˜no is on the ensuing reduction in upper-layer ITF trans- port. As described insection 3b, anomalous surface-intensi- fied ITF transport represents the leading mode of ITF variability (ITFM1;Figs. 5a,b). In the reanalysis, a reduced surface transport associated with ITFM1is more of a lagged response to an El Nino with less apparent link to the IOD˜ (Fig. 5d). In the CMIP5 models, on the other hand, ITFM1
is instead prominently linked to the IOD (Fig. 5e) with positive and negative IOD respectively corresponding to enhancement and reduction in ITFM1-associated transport.
Thus, the models’ tendency for a delayed positive IOD counteracts the upper-layer ITF reduction that typically fol- lows an El Nino. The prevalence of the IOD in driving˜ ITFM1in the CMIP5 models is linked to the magnitude of txIOvariability, which is notably stronger than the reanalysis (Fig. 9b;r50.52; 0.44 without EXM). Models with stronger txIOvariability tend to exhibit a stronger link between DMI and ITFM1. In the reanalysis, there is less apparent link between ITFM1and the IOD, and consistently thetxIOvari- ability is much weaker than in the CMIP5 models. In addi- tion, there is a tendency for models with more prevalent ITFM1relative to ITFM2to exhibit a weaker link between ENSO and ITFtotalvariability (Fig. 9c;r50.65; 0.68 without EXM). These compensating effects potentially explain why the ITF transport variability in the CMIP5 models tends to be weaker than in the reanalysis, despite the IOD amplitude being much stronger than the reanalysis (Fig. 7).
4. Summary and discussions
Up until now, there has been no systematic multimodel study on ITF variability linked to Indo-Pacific climate vari- ability; namely ENSO and the IOD. This is examined here using 20 CMIP5 models and the SODA-2.2.4 reanalysis. We found that overall the CMIP5 models capture several proper- ties of ITF transport that are qualitatively consistent with the SODA reanalysis, although the CMIP5 intermodel differ- ences are substantial. For instance, the CMIP5 models simu- late an ITF total transport (ITFtotal) that reaches a maximum in austral winter and a minimum in austral summer, with a peak-to-troughamplitude of∼9 Sv in the multimodel mean seasonal cycle, consistent with the reanalysis, but with a large intermodel range of∼4–18 Sv. In terms of variability of the ITF total transport, with long-term trends removed, the stan- dard deviation of the CMIP5 models ranges from 0.8 to 3.1
Lag Nino3.4 DMI (month) Max cor Nino3.4 vs IT F
0−100r=0.61 (0.00) (a)
−4 −2 0 2 4 6 8
−0.8
−0.6
−0.4
−0.2 0 0.2
τ
xIOstd dev (x10
−2N m
−2)
Max cor DMI vs PC1
r=0.52 (0.02) (b)
0.5 0.7 0.9
0.2 0.4 0.6 0.8
Ratio var EOF1 vs EOF2 Max cor Nino3.4 vs IT F
total(c)
r=0.65 (0.00)
1 2 3 4 5 6 7
−1
−0.8
−0.6
−0.4
−0.2 0
Fig. 9. Intermodel relationships of (a) ENSO–IOD time lag (seeFig. 8) and maximum negative correlation between Nino-3.4˜ and ITF0–100 over which Nino-3.4 leads ITF˜ 0–100 (supplemental Fig. S6c), (b) standard deviation of zonal wind variability and maxi- mum correlation between DMI and ITFM1(refer toFig. 5e), and (c) the ratio of variance explained by ITFM1and ITFM2(Fig. 5a) and maximum negative correlation between Ni˜no-3.4 and ITFtotal (Fig. 3c). Red square indicates CMIP5 multimodel mean; black square indicates SODA.
Sv, with a multimodel mean of 2.2 Sv that is slightly weaker than that of the reanalysis at 2.7 Sv. The discrepancy is larger when focusing on variability at interannual time scales where the CMIP5 multimodel mean is ∼1 Sv and the reanalysis 1.5 Sv. This indicates the different roles of the simulated ENSO and IOD in the CMIP5 models on the ITF variability.
The ITF total transport is found to weaken during El Nino˜ and strengthen during La Nina marked by CMIP5 multimodel˜ mean in maximum correlation between the Nino-3.4 index˜ and ITFtotalof∼20.3, statistically significant above 95% confi- dence level, with a large intermodel range of ∼0.4. The relationship with the IOD is in contrast not statistically signifi- cant. Considering that ENSO and IOD are the dominant driv- ers of Indo-Pacific climate variability, these relationships appear disproportionately weak. This seemingly weak rela- tionship is due to the vertical structure of ITF variability in which surface and subsurface transports tend to exhibit opposing anomalies, a feature that was previously revealed by Potemra and Schneider (2007)based on two climate models and a reanalysis, and identified in studies based on limited observations (e.g.,Sprintall et al. 2009;Susanto et al. 2012;
Sprintall and Revelard 2014).
Here we further show that ITF variability can be decom- posed into two primary vertical structures, with one exhibiting a surface intensified anomaly (ITFM1), and the other exhibit- ing anomalous opposingflows between the surface and subsur- face in the upper 300 m (ITFM2). Separating the variability into these two structures reveals a discrepancy between the CMIP5 multimodel ensemble and the reanalysis. While the
CMIP5 models and reanalysis are consistent in terms of ITFM2, which is shown to be a response to both ENSO and IOD, there is a strong disagreement associated with ITFM1. In the reanalysis, ITFM1is a lagged response to ENSO, such that an El Ni˜no leads to a surface-intensified reduction in transport about 6 months later. On the other hand, ITFM1in the CMIP5 models is linked to the IOD, in which a positive IOD corre- sponds to a surface-intensified transport enhancement. The prevalence of an IOD impact on surface transport in the CMIP5 models is related to the overly strong IOD amplitude, which impacts surface transport anomalies through local wind variability. However, despite the strong IOD amplitude, the CMIP5 ITF variability tends to be weaker than that in the reanalysis. This damped ITF variability is partly attributed to the propensity for the CMIP5 models to simulate a delayed or prolonged IOD in response to ENSO, of which the effects tend to counteract each other. As noted in section 3c, the delayed IOD bias could be related to model resolution, with models with the coarsest horizontal resolution exhibiting the strongest bias.
Our results underscore that the ITF exhibits vertical struc- ture that responds differently to ENSO and the IOD, and thus diagnosing the variations of the ITF under different climate states requires a consideration of processes in the different layers. To highlight this point, we present some results on ITF changes in response to greenhouse forcing, which clearly exhibit distinct responses in the surface and subsurface layer transports. We utilize the CMIP5 simulations under represen- tative concentration pathways (RCP) 4.5 scenario (Taylor et al.
(x10−2 Sv)
Depth (m)
(a) ITF mean state difference
−1 0.5 0 0.5 1
−1500
−1000
−500 0
r=−0.05 (0.84) (b) Variability dif. (ENSO vs ITF
0−100)
Δ std ITF 0−100 (x10−1 Sv)
−0.2 0 0.2 0.4
−4
−3
−2
−1 0 1 2
Δ std ITF 100−300 (x10−1 Sv)
Δ std Nino3.4 (°C)
(d) Variability dif. (ENSO vs ITF100−300)
r=0.93 (0.00)
−0.2 0 0.2 0.4
−4
−2 0 2 4
r=−0.14 (0.54) (c) Variability dif. (IOD vs ITF
0−100)
−0.2 −0.1 0 0.1 0.2
−4
−3
−2
−1 0 1 2
Δ std DMI (°C) r=0.68 (0.00) (e) Variability dif. (IOD vs ITF100−300)
−0.2 −0.1 0 0.1 0.2
−4
−2 0 2 4
FIG. 10. (a) Differences between future (2006–98; RCP4.5) and historical (1907–99) periods in ITF transport at depth levels. Multimodel difference is denoted by thick dashed line; thick red line indicates statistically significant difference at 95% significance level (evaluated using bootstrap mean test with 1000 draws). (b)–(d) Intermodel relationships between the future change in ENSO and IOD amplitude and that of ITF variability in the surface (ITF0–100) and subsurface (ITF100–300) layer.
2012), comparing thefuture(2006–98) andhistorical(1907–99) periods. First, in terms of mean state there is a lack of a robust change in the surface layer transport (Fig. 10a), consistent with the fact that the local wind (txIO) changes are diverse across the models (figure not shown). On the other hand, a robust ITF transport slowdown is found at depths below 100 to 1200 m (Fig. 10a), which is consistent with the projected weakening in the global ocean circulations (Sen Gupta et al. 2016). In terms of interannual variability, the response is diverse across the CMIP5 models (Figs. 10b–e). Nonetheless, there is a strong link between the change in ITF variability and the change in ENSO amplitude, and that link is found in the subsurface layer with a high intermodel correlation of 0.93 (Fig. 10d). The asso- ciation with the IOD amplitude change is also significant, although not as strong as that of ENSO (r50.68;Fig. 10e), likely stemming from the intermodel link between ENSO and IOD amplitude change (r50.65;figure not shown). On the other hand, the connections between changes in the surface transport variability and those of ENSO and IOD amplitude are not statistically significant (Figs. 10b,c), which is an expected consequence of the counteracting effect between ENSO and IOD as discussed above, as well as various other factors that may influence the surface transport.
Previous studies have shown that there tends to be no inter- model consensus in the projected change in ENSO and