The primary Indian Ocean mode is the IOD, discovered following a strong positive event in 1997 . The magnitude of IOD events is commonly measured using the Dipole Mode Index (DMI; see Sect. for definition). The majority of IOD events are triggered by the El Niño–Southern Oscillation ENSO;. During an El Niño, the atmospheric convection in the Walker circulation shifts eastward, forcing anomalous subsidence over the Maritime Continent and easterly wind anomalies along the equatorial IO , triggering a positive IOD event. Alternatively, some IOD events are initiated without any Pacific influence . Once SST anomalies form, regardless of their origin, several positive feedbacks sustain and enhance the IOD . As the zonal SST gradient initially decreases, the seasonal easterly winds strengthen, shoaling the thermocline and resulting in cold upwelling, which reinforces the initial cold anomaly . Anomalous Ekman pumping warms the western Indian Ocean . IOD events are characteristically phased-locked to begin developing around June, peak in boreal fall, and rapidly decay following the reversal of the trade winds in November . The IOD has a periodicity of 3–5 years with positive events being more intense than negative events . This skewness is mainly attributed to asymmetry in the thermocline–SST feedback, which overpowers increased damping from the SST–cloud–radiation feedback .

Projected future changes in the Indian Ocean's mean state will have significant impacts on the IOD by modulating the efficiency of the feedbacks that control its evolution . Shoaling of the equatorial thermocline near Indonesia will strengthen the thermocline–SST feedback, encouraging stronger IOD events . On the other hand, increased atmospheric static stability will weaken the atmospheric components of the Bjerknes feedback . The changing feedbacks will modify the characteristics of the IOD, reducing skewness and shifting events towards boreal summer . Although the frequency of positive IOD events has been rising, the consensus on the future frequency and strength of IOD is weak . The overall frequency of IOD events is not projected to increase , yet proxies suggest past positive IOD-like mean states have experienced more frequent IODs . Some studies find no increase in IOD magnitude under future warming scenarios, whilst others have . The projected frequency of extreme positive IOD events increases substantially as the zonal SST gradient and winds weaken . However, an opposite response may occur for moderate positive events as tropospheric warming limits Ekman pumping .

Investigations of past interannual variations in the Indian Ocean have facilitated insights into the nature of coupling between altered mean states and climate variability in this basin. Proxy reconstructions reveal elevated IOD variability during periods when the zonal SST gradient was reduced relative to present-day conditions, accompanied by enhanced mean cooling in the eastern Indian Ocean (see for a review). Such periods of intensified IOD activity alongside eastern cooling have been uncovered during the last millennium , the mid-Holocene period , and the Last Glacial Maximum (LGM) . Various lines of proxy evidence reinforce the finding that the Indian Ocean is capable of harbouring larger interannual climate variations than observed today . Coeval records from the Pacific point to mean-state-dependent feedbacks maintaining pan-tropical IOD and ENSO interactions over different palaeoclimate periods. For example, strong IOD activity recorded over the last millennium appears to be tightly linked to strengthened ENSO variability as well . During the mid-Holocene and the LGM however, when global climate and forcing were drastically altered relative to the last millennium, reconstructions point to stronger interannual variability and weakened ENSO-related variability in the Pacific . Though the ENSO–IOD relationship may depend on the palaeoclimate mean state and boundary conditions therein, proxy records support a tight link between the zonal SST gradient and IOD intensity , with robust evidence for variations larger than those observed in the instrumental period.

There have been several previous studies looking at the Indian Ocean Dipole using palaeoclimate simulations by a single model. and use the FOAM model and find a slight weakening in IOD amplitude at the mid-Holocene and little change at the LGM. Meanwhile find a stronger IOD in the mid-Holocene in MIROC5. showed that changes in IOD amplitude do not scale global mean temperature changes across a multi-model ensemble including palaeoclimate simulations. Combined, though, the literature does suggest IOD changes with changes in the seasonal cycle; something supported by the emerging, independent research of , who look at the mechanisms behind this response.

The Indian Ocean Basin Mode (IOBM) is a basin-wide anomaly that is the leading mode of interannual Indian Ocean SST variability. IOBM events often follow IOD events, developing in boreal winter and peaking in the following spring . This seasonality implies a connection with the ENSO mature phase in boreal winter , although modelling suggests IOBM can occur without ENSO . Nonetheless, ENSO is the dominant forcing mechanism for the IOBM as an El Niño induces anomalous Walker circulation over the equatorial Indo-Pacific . This modulates surface winds, reducing evaporation and increasing downward short-wave radiation to warm the Indian Ocean . Diffusion of tropospheric temperature anomalies from the eastern Pacific contributes significantly . The IOBM–ENSO relationship displays interdecadal variation with strong ENSO elevating the relationship as the thermocline shoals in the south-west Indian Ocean . Most Coupled Model Intercomparison Project (CMIP) 5 models capture the key features of the IOBM whilst overestimating its amplitude , due to winter rainfall and thermocline biases . Initial investigations of the CMIP6 ensemble indicate underestimated SST variability results from inaccurate latent heat fluxes and wind-driven ocean processes .

IOBM future projections must consider the response and uncertainty of ENSO simulation . Under warming, the tropical Pacific moves towards an El Niño-like mean state and a consensus is developing that ENSO-induced precipitation variability will intensify, canonical El Niño will strengthen, and both El Niño Modoki and extreme ENSO events will increase in frequency . These changes will exacerbate the future global impacts of ENSO and should drive a stronger IOBM . Indeed, many modelling studies suggest an enhanced IOBM and capacitor effect under global warming . This has been found alongside ENSO activity that is reduced or unchanged . The increased IOBM response results from strengthened air–sea interactions and a greater tropospheric temperature response . A recent study projects a weaker IOBM response in early summer suggesting a decreased capacitor effect, yet the IOBM feedback on ENSO transition is strengthened . The IOBM is also projected to have an increasing influence over the East Asian summer climate .

Few studies have focused on reconstructing past IOBM variations. Complexities arise in assessing past IOBM behaviour due to the potentially confounding influence of the IOD, where warming in the east of the basin associated with negative IOD events could be conflated with anomalous IOBM warming . Moreover, the lack of available interannually resolved records from the western Indian Ocean presents an additional challenge in isolating past IOBM intensity.

Climate models, with their basis in geophysical fluid dynamics, are our best source of information about the future changes in climate. The vehicle that coordinates and collates simulations of future climate is the Coupled Model Intercomparison Project (CMIP) . The aspect of CMIP that looks at simulations of past is the Paleoclimate Model Intercomparison Project (PMIP). Here, we combine simulations from both PMIP phase 4 PMIP4; and phase 3 (PMIP3) . This acts to increase the ensemble size and the two phases have shown to be statistically indistinguishable .

Models must have completed one or more palaeoclimate simulations to be included here and to have provided monthly surface temperature and precipitation for at least 30 years for both this simulation and the preindustrial control (see Sect. ). The resulting 34 models are listed in Table , with a total of 34 426 simulated years analysed. Further information about the CMIP6 models is available in on the PMIP4 website (https://pmip4.lsce.ipsl.fr/doku.php/database:participants, last access: 7 July 2022), whilst details of the CMIP5 models can be found in Table 9.A.1 of .

This research uses simulations run under five different experiments defined under either the CMIP or PMIP protocols. All models have performed a preindustrial control (henceforth piControl) which approximates to constant 1850 CE boundary conditions . This simulation acts as a baseline from which changes are computed under all the other experiments. The piControl simulations vary in length but are a minimum of 100 years long (Table ). The “abrupt4xCO2” experiment is another of the required deck of CMIP simulations and is an idealized forcing experiment where the carbon dioxide concentrations are instantaneously quadrupled and then the model is left to respond to them. These simulations are useful in estimating a model's climate sensitivity , and the resulting values are shown in Table . Here we use the average of years 101–150 after the forcing change to provide an idea of climate changes in response to increasing greenhouse gases. This experiment and a segment of it are selected over other options as they are (a) unchanged between CMIP5 and CMIP6, (b) available for a large number of models, and (c) closer to equilibrium than the alternative “1pctCO2” experiment. Efforts are made to account for any transience in this segment through linear detrending (see Sect. ).

The “midHolocene” experiment is the most simulated of all PMIP experiments, having both been part of PMIP since the beginning and not requiring changes in the land–sea mask. The midHolocene experiment aims to replicate the conditions of 6000 years ago. The primary change in the simulation is alteration in the orbital configuration, although a small reduction in greenhouses gases is incorporated in the most recent simulations . We consider simulations from both PMIP3 and PMIP4 together, as their forcings and boundary conditions are statistically indistinguishable.

The “lig127k” experiment represents some of the warming conditions seen during the last interglacial, focused at 127 000 years ago . The experiment is included in PMIP4 for the first time and has been completed by 14 modelling groups . It is predominantly an orbitally forced experiment, with accompanying changes in greenhouse gases: the approximate +9 m of sea level rise during the last interglacial are not incorporated into the boundary conditions. The magnitude of the orbital forcing is substantially larger than seen during the midHolocene, with the anomaly in NH summer insolation being twice as strong and occurring for longer .

The world was much colder at the time of the Last Glacial Maximum (LGM), by a similar order of magnitude to the warming seen in projections e.g.. As such, it has long been seen as an important test for climate models and has featured in PMIP since its inception. The “lgm” experimental protocol requires the imposition of large ice sheets, associated changes in the land–sea mask, and reductions in greenhouse gases. This can be a challenging experiment to deploy for a climate model, and therefore a smaller number of modelling groups have completed it than either of the interglacial experiments . Only a handful of groups have posted output on the Earth System Grid Federation (ESGF) (Table ) at the time of writing, so we additionally incorporate PMIP3 simulations to increase the ensemble size. Although the PMIP4 models show some differences in circulation and encompass a greater ensemble spread in temperatures, we note that conclude that they “are not fundamentally different from the PMIP3-CMIP5 results”.

This research looks at coupled phenomena, and so self-consistency between all observed variables is an important factor. Therefore, we adopt a reanalysis product instead of multiple observational datasets. For this work, the 20th-Century Reanalysis is used, although we do not expect our main conclusions in the model evaluation to be sensitive to this choice. Version 3 of the 20th-Century Reanalysis extends back to 1836 and uses HadISST to provide its bottom boundary conditions . Initial analysis using the shorter second version did expose a sensitivity in the IOD rainfall teleconnection over India, presumably relating to the ENSO–IOD relationship in the earlier period. Evaluation of the models is performed through comparison of the 20th-Century Reanalysis to the piControl simulations. It would be preferable to use the historical simulations instead of piControl. However, not all models have a historical simulation available, and previous work suggests the difference in biases are not substantial .

The analysis undertaken in this research follows the workflow described by . This involves the creation of a curated replica of the relevant simulation output available on the ESGF. Then calendar adjustments are made to re-aggregate the monthly output from a present-day calendar to those representing 30∘ of the orbit for the past climate experiments using the PaleoCalAdjust software . A modified version of the Climate Variability Diagnostics Package CVDP; is run on an individual simulation to calculate multiple pertinent time series and spatial fields. Please see for further details and instructions on performing such analysis yourself.

All of the modes of climate variability explored in this research are defined through the use of area-averaged sea surface temperature anomalies (or more strictly “skin temperature”, which is used as an alternative in CVDP). All anomalies are computed with respect to a climatology calculated over the full simulation length and additionally have a linear trend removed in case the simulations are not equilibrated. The same process is undertaken in the abrupt4xCO2 experiment, except only years 101–150 are considered. ENSO is monitored using the Niño3.4 region (5∘ S–5∘ N, 120–170∘ W; ). The Dipole Mode Index is used to define the IOD . It is the difference between the area average of 10∘ S–10∘ N, 50–70∘ E and 0–10∘ S, 90–110∘ E. The Indian Ocean Basin Mode is defined using variations in the Tropical Indian Ocean time series (averaged over 15∘ S–15∘ N, 40–110∘ E; ). These modes adopt the definitions used by the IPCC , who demonstrate that they are equivalent to definitions based on empirical orthogonal functions.

The El Niño-like mode of variability proposed by is best defined using wind fields . However, those fields are not as universally available in the ESGF archive, especially for PMIP experiments. Instead, we analyse the area-averaged SST anomalies of the Eastern Equatorial Indian Ocean (EEIO) over the region 2.5∘ S–2.5∘ N, 70–95∘ E , the impact of which will be discussed in Sect. .

We present spatial patterns associated with the Indian Ocean Dipole, which are computed via regression. Detrended monthly anomalies of surface temperature and precipitation are computed at every grid point across the globe. These anomalies are then regressed against the Dipole Mode Index (defined as described above). The resulting patterns are therefore expressed as the local change in temperature or precipitation seen under a positive IOD event when the Dipole Mode Index is 1 ∘C the level of warming reached during the extreme event in 2019;. Both the IOD spatial patterns and the mean climate changes presented in Sect.  show the multi-model ensemble average. Each model's changes are computed on their own native grid and then interpolated onto a regular 1∘ grid before averaging across the ensemble. Stippling in figures is used to indicate where the ensemble is “not consistent” in its direction of change; this is computed as being less than two-thirds of the ensemble array on the sign of the change.

It is necessary to assess the appropriateness of the PMIP ensemble against observations prior to looking at the response in experiments. This is performed through the inspection of the patterns of surface climate variables over the region of interest (Fig. ). Here we use a single atmospheric reanalysis dataset 20th-Century Reanalysis;Sect.  driven by the prescribed SSTs from the HadISST dataset . The annual pattern of SST consists of a pool of warm water extending from Indonesia in the east to nearly all the way across the basin (Fig. a). This extent is relatively well captured by the ensemble mean (Fig. b), although it overestimates the amount of cooler water in the Arabian Sea and near Madagascar. The annual mean precipitation pattern reflects the seasonal march of the Intertropical convergence zone (ITCZ), with an intense oceanic band at 10∘ S and a more diffuse terrestrial pattern in the Northern Hemisphere (NH) especially focused over orography (Fig. c). The ensemble mean simulated pattern shares most of these features but is much smoother (Fig. d).

The Indian Ocean Dipole is characterized by a region of cooling off Java and Sumatra counterbalanced by a warming in the western half of the basin Fig. e. This pattern in SST is fundamentally replicated by the ensemble (Fig. f). The models show more cooling around the Indonesian Archipelago, as well as a less extensive patch of warm SSTs in the western half of the basin, especially south of the Equator. The rainfall pattern of the IOD is strongest over the interior of the IO, with the direction controlled by the underlying SST pattern (Fig. g). This aspect of the rainfall pattern is well simulated by the ensemble (Fig. h). Both models and observations capture a reduced rainfall over India.

The IOD is positively correlated with the Indian summer monsoon rainfall , although the strength of that teleconnection seems to be changing . find that CMIP5 models (roughly half the current ensemble) have too strong a summer rainfall teleconnection. show that the quality of the IOD simulation improved in CMIP6. However, this literature-based description of the IOD's teleconnection to rainfall over India is of a different sign to the response shown in Fig. g and h. This discrepancy results from the analysis approach used here combined with model mean state biases. demonstrate that the sign of the Indian summer monsoon IOD teleconnection is dependent on the analysis technique used (for compositing versus correlation). The IOD teleconnection shown here is computed by regressing detrended monthly anomalies (Sect. ) for all seasons; this already acts to convolve IOD impacts on the summer and winter monsoons. More importantly, the regression approach we adopt can be overridden by an ENSO response which is correlated to the DMI time series . Additionally, the models underestimate the annual rainfall over central India on average (Fig. c, d) because they do not capture the full strength of the summer rainfall, which further acts to reduce the summer contribution to the teleconnection. A separate investigation of the role of the Indian Ocean Dipole on the Indian monsoon rainfall using an alternate analysis approach would be required to fully isolate and understand the present and past response. It is sufficient to note here that the models do replicate the large-scale pattern emerging from the observations, but there remains doubt about the IOD teleconnections over the Indian subcontinent. Therefore, this analysis does not draw conclusions about the impact of past interannual variability changes on the Indian monsoon.

The tropical climate in the experiments analysed here primarily responds to one of two forcings: changes in the orbital configuration or in the concentrations of greenhouse gases. These different forcings result in two distinct climate responses, as demonstrated in the simulations. First we present and discuss the midHolocene and lig127k experiments together and then move onto a description of the lgm and abrupt4xCO2 experiments. We note this latter comparison is not perfect because the high-latitude ice sheets do impact the tropics through bathymetry changes and atmospheric teleconnections .

Precession leads to changes in the day of the year at which the Earth is closest to the Sun (the perihelion). At present, this occurs during NH winter, yet this was nearer the autumn equinox during the mid-Holocene and occurred near the summer solstice at 127 ka . This results in a different distribution of the incoming solar radiation throughout the year. The eccentricity of the orbit was also larger during the last interglacial, exacerbating these changes in the lig127k experiment . Intuitively, more sunlight in summer and less in winter will lead to a stronger seasonal temperature range in the Northern Hemisphere, especially outside of the tropics. This is visible in Fig. b, c for the midHolocene simulations in Asia north of the Himalayas and even more so in lig127k Fig. e, f;. Over most of the region, the seasonal temperature changes “cancel” themselves out, leading to minimal changes in the annual mean (Fig. a, d). This is especially true over the ocean, where its high effective heat capacity acts to damp seasonal variations.

There is an obvious exception to the simple narrative of precession-modulated temperature response outlined above, with summer cooling over land stretching from the Indian subcontinent over Arabia and into Ethiopia (Fig. c, f). This can be thought of as being related to a poleward shift in the summer ITCZ driven by the stronger interhemispheric temperature gradients , mainly through dynamic processes rather than thermodynamic ones . The impact of this process is greatest over West Africa and is underestimated in model simulations , and both experiments fall under separate instances of an African humid period . Summer rainfall increases over East Africa are clearly visible (Fig. c, f), which extend into the Arabian peninsula. This increased rainfall comes with an increase in convection and cloud cover, leading to a reduction in incoming solar radiation at the surface as well as latent heat flux changes associated with the precipitation. The situation is different over India, however. There is no increase in total rainfall in the Indian summer monsoon Fig. , but rather a redistribution of it, so that more falls in the foothills of the Himalayas instead of on the central Indo-Gangetic Plain. A similar style of dipole response occurs over the Western Ghats. The band of anomalous summer drying extends east from central India over South-East Asia to the Philippines . Meanwhile the Indonesian Archipelago is wetter in JJA and drier in DJF, as is northern Australia.

Over the Indian Ocean itself, there is an orbitally driven increase in DJF rainfall that is strongest near the African coast (Fig. b, e). This accompanies, and could be driving, a slight cooling in DJF in the region (Fig. b, e). In JJA, the orbital forcing results in a west–east dipole of rainfall changes (Fig. c, f), which is reminiscent of the IOD teleconnection (Fig. h). This is accompanied by some low-amplitude, annual mean SST changes that project onto the IOD (Fig. h), although these are only visible in the lig127k ensemble (Fig. d). The weaker orbital forcing in the midHolocene experiment means that the changes are swamped by the impact of the reduced greenhouse gases in the PMIP4 experiment .

The Last Glacial Maximum was a substantially colder period, with global mean cooling of around 5–7 ∘C . The lower greenhouse gas concentrations played a dominant role in determining the tropical climate changes, although impacts on climate variability from the Laurentide ice sheet were felt across the globe . The tropical Indian Ocean itself cools by 2–3 ∘C in the PMIP ensemble (Fig. g), roughly in line with estimates from palaeodata assimilation . Conversely a quadrupling of CO2 leads to an increase in temperatures by 4–5 ∘C (Fig. j). Both responses show little seasonal variation and greater amplitudes of temperature change over the land (Fig. ). The two patterns mirror each other to first order, except for a stronger response over the Indonesian Archipelago associated with the exposure of the Sunda Shelf .

The colder climate of the lgm experiment leads to drier conditions across the region Fig. g;, whilst future warming leads to increased rainfall across most of the basin (Fig. j). There is greater seasonal variation in the rainfall changes (Fig. k, l) than the temperature response (Fig. k, l), reflecting the stronger seasonal cycle in rainfall. There is an interesting non-linearity in the JJA rainfall changes between the LGM and idealized warming to the south-east of Sumatra and Java (cf. Figs. g and j) in that both high- and low-CO2 experiments demonstrate a drying in the region. This non-linearity is centred on the eastern region of the Dipole Mode Index (Sect. ) and therefore could be expected to influence the response of the Indian Ocean Dipole. This arises from the exposure of the Sunda Shelf , which has already been shown to have an influence on the ensuing variability . Even though the underlying drivers of climatic forcing for the lgm and abrupt4xco2 experiments differ, they trigger a similar mean zonal response in the tropical Indian Ocean that has been suggested to modulate interannual variability across the basin .

Having established that the experiments lead to consistent changes in the mean Indian Ocean climate, we next explore whether they also alter its variability. We start by looking at the IOD, whose pattern is relatively well captured by the models (Sect. ). The amplitude of the IOD can be characterized by the standard deviation of the monthly Dipole Mode Index (Sect. ). The standard deviation of the (detrended) Dipole Mode Index in the HadISST dataset, used by the 20th-Century Reanalysis, is 0.46 ∘C. There is a spread in the amplitude of the IOD in the piControl simulations that ranges from 0.24 ∘C in INM-CM4-8 to 0.75 ∘C in CSIRO-Mk3-6-0, with an ensemble median that is close to the observed value (Fig. a). The precise composition of models that have undertaken each experiment varies, leading to subtly different control ensembles. Those models that have performed an LGM experiment are generally ambiguous about its impact on the IOD amplitude: 7 of the 13 models show a slight decrease, with the other 6 showing an increase. MIROC-ESM shows the largest change (an increase of 63 %), although its successor model has a very small reduction (-0.01 ∘C). The median value also remains the same (Fig. a), leading us to conclude to that the last glacial conditions did not impact the magnitude of the Indian Ocean Dipole. A similar situation occurs for the lig127k experiment: equivocal simulated changes leading to minimal changes in the IOD amplitude distribution. Despite having a weaker orbital forcing, the midHolocene sees a reduction in the standard deviation of the DMI of roughly 10 %. The individual models within the ensemble still show some disagreement about the signal, with only 20 out of 29 models simulating a reduction of some magnitude (CSIRO-Mk3-6-0 is an outlier, whose already high IOD amplitude increases by a further 35 %). There is a robust decrease in IOD amplitude in the abrupt4xCO2 experiments (Fig. ; taken between years 101-150 with a linear trend removed), with 18 of the 24 simulations showing a reduction of some magnitude. This result was noted in PMIP3 by and fits with , who anticipate a reduction in amplitude should the mean state changes resemble a positive IOD, conditions that are found in Fig. l, although this will be explored in more in Sect. . The abrupt4xCO2 experiments are still not equilibrated, and there may be some impacts of slow feedbacks that have not yet fully emerged .

Our approach to define the IOD pattern through linear regression (Sect. ) allows spatial changes to be considered separately from changes in amplitude. The orbital forcing experiments both see an expansion of the IOD cold pole eastwards along the Equator (Fig. a, c), which is stronger in the lig127k than the midHolocene. This spills out over the Equator into the Northern Hemisphere, especially into the Bay of Bengal. In the lig127k experiment, this acts to make Bay of Bengal SSTs cool during a positive IOD event in every model. This is a region where many models already have trouble replicating the warming signal seen in observations (Fig. e, f), but even the five models that successfully capture this in the piControl simulations switch to cold SSTs in the lig127k (not shown). This is accompanied by a constriction of the western SST pole of the IOD pattern back towards the East African coast (Fig. c). This results in the boundary between the dipolar phases of the IOD occurring to the west of Sri Lanka, rather than to its east, in the majority of models. The shifts in the midHolocene experiments are not as strong but are of a similar nature (Fig. a, c).

The expansion of the SST cold pole along the Equator during the orbital simulations (Fig. a, c) is accompanied by a reduction in rainfall that extends further westward during a positive IOD (Fig. b, d). There is also a slight weakening of the strong drying response in the south-east Indian Ocean (east of Java) and an intensification of the wet conditions off the East African coast, which do not extend into the continent. Both the lig127k and midHolocene ensembles suggest that the IOD had a wetter response over central India (Fig. b, d), although this could indicate a greater contribution from the IOD rather than ENSO throughout the simulation (see Sect. ).

There is little change in the SST pattern of the IOD in the LGM experiment (Fig. c), with the only substantial changes occurring within the Maritime Continent. This is likely associated with the sub-aerial exposure of the Sunda Shelf , which is handled differently by the different models and with varied responses in the atmospheric circulation . The rainfall teleconnection pattern associated with the IOD is severely damped in the LGM experiment. This likely occurs as a consequence of the Clausius–Clapeyron relationship; the regression is performed on absolute rather than relative rainfall, and there is a substantial reduction in climatological precipitation across the tropics (Fig. g).

The change in the IOD pattern of SST in the abrupt4xCO2 experiment (Fig. d) sees an equatorward shift in the cold pole, combined with a slight expansion westwards along the Equator. The ensemble does not show a consistent response of changes in the warm pole, although there is some consistency in the Arabian Sea with most models showing an expansion of the warm pole there. Interestingly, in the abrupt4xCO2 experiment anomalous rainfall variations associated with positive IOD events (which generate drying over the east – see Fig. h) reduce across the Indian Ocean (Fig. h).

The Indian Ocean Basin Mode (IOBM) involves synchronous fluctuations across the whole basin. Its amplitude can be measured by looking at the standard deviation of the tropical Indian Ocean (Fig. b). The PMIP4 ensemble shows large variations in the quality of their simulation in the piControl, ranging from roughly one-third to 1.5 times the amplitude seen in the 20th-Century Reanalysis 0.31 ∘C. In the LGM experiment, there is no change in the median IOBM amplitude, despite more than two-thirds of the ensemble suggesting an increase and COSMOS-ASO simulating a dramatic one. The midHolocene ensemble shows a slight reduction in IOBM amplitude, seen in 22 of the 28 members with a mean reduction of -3.7 %. A similar behaviour is seen in the lig127k ensemble, although the amplitude reduction is slightly stronger, with an ensemble mean reduction of -6.4 % (Fig. b). There is a slight increase in IOBM amplitude under the abrupt4xCO2 experiment as suggested by, e.g., , although the ensemble as a whole is equivocal, with only 12 of the 25 members agreeing in sign with this mean.

In this work, we have looked at both the response of the mean state of the Indian Ocean and its variability under a variety of experiments. We found changes in both of these properties and have discussed some spatial relationships between their ensemble mean responses. A natural extension of this is to ask if there exist underlying relationships that act across all the different experiments. The changes in the Indian Ocean Basin Mode (Fig. b) were weak in every experiment, yet there were substantial basin-wide SST changes in the LGM and abrupt4xCO2 experiment (Fig. ). It should therefore come as little surprise that PMIP4–CMIP6 demonstrates no significant relationship between the changes in the mean tropical Indian Ocean Index and changes in its standard deviation (Fig. a).

The literature suggests that changes in the mean state of the Indian Ocean can map onto the Dipole Mode Index . This is further suggested to result in changes in the IOD via changes in the equatorial thermocline . The PMIP4–CMIP6 ensemble shows little consistency among changes in the mean gradient between the two poles of the Dipole Mode Index and the variability of the index (Fig. b). Even separating out just individual experiments shows a muted relationship. Given the large ensemble size and that the majority of the experiments have reached quasi-equilibrium, there is little scope for the initial conditions of internal variability to influence the IOD changes under this experiment, highlighted as an important source of uncertainty in future projections by . It should be noted that the method adopted here not only combines positive and negative modes of the IOD but also moderate and strong positive events, which have been shown to respond differently under future warming . However, orbital changes result in different seasonal temperature changes in the west and east Indian Ocean, most visibly in the lig127k experiment (Fig. e, f). This means that the annual cycle of the SST difference measured by the Dipole Mode Index varies under orbital forcing and also happens to occur in the other experiments. One way to explore this seasonality is to look the amplitude of the annual cycle (i.e. the difference between the maximum and minimum monthly SST gradients). Changes in the amplitude of the seasonal cycle are inversely proportional to the changes in the IOD strength (Fig. c). Further independent research supports the importance of the changes in seasonal cycle in the midHolocene ensemble and looks at the physical mechanisms behind their influence. Taken together these findings suggest there is potential to constrain the future projections of the IOD with (palaeo-) observations of the seasonal change in SST gradients.

There is a strong link between variability in the Pacific Ocean, predominantly the El Niño–Southern Oscillation (ENSO), and that in the Indian Ocean. This impacts both the IOD and the IOBM . have already explored changes in ENSO across the PMIP4–CMIP6 ensemble. They saw consistent reductions in ENSO amplitude in both the lig127k and midHolocene experiments but little consistency across the other two experiments presented here. Similar to the results shown in Fig. a and b, there is no universal relationship between mean state changes in the equatorial Pacific and the amplitude of ENSO . The amplitude of ENSO in the various experiments, as measured by the standard deviation of the smoothed monthly Niño3.4 SST anomalies, is shown in Fig. a. The damping of ENSO in the orbital simulations is most visible in the lig127k experiment. There are hints of an increase in the abrupt4xCO2 experiment; however, the ensemble is equivocal, with 11 out of 25 models showing a reduction in ENSO amplitude. Despite this lack of consistency in the simulated direction of change, the ensemble demonstrates a strong positive relationship between the changes in the amplitude of ENSO and those of the IOBM (Fig. b), which is statistically significant. This conforms nicely with our pre-existing knowledge of the role of ENSO as a driver for IOBM events .

The relationship between changes in ENSO and the IOD are more nuanced. ENSO is known to play a role in a proportion of IOD events , although by no means all of them . Therefore if ENSO is damped one might reasonably expect that the IOD also becomes weaker. Alternatively one could envisage that ENSO's relative importance for the IOD becomes less, as has been posited before for palaeoclimate periods . We explore both possibilities through scatter plots between and the ENSO amplitude (Fig. c) and the IOD amplitude (Fig. d). Statistically significant lines of best fit could be drawn in both cases for the ensemble as a whole, yet it is unclear if this a wise choice. There is a strong contribution from the abrupt4xco2 experiments between changes in the amplitude of ENSO and both the IOD amplitude and its correlation with ENSO (shown as solid red lines in Fig. c, d). However, neither line passes through the origin, which would be required as these are scatter plots of anomalies from the piControl and thus call for invoking model biases or other uncertainties. Equally, the relationships are not supported across the pair of orbital experiments, both of which do show a linear regression of the same sign (yellow–green lines in Fig. c, d) but neither of which are statistically significant.

propose the existence of a third mode of interannual variability in the Indian Ocean, unobserved in the modern climate but which may have been active during the Last Glacial Maximum. Sediment cores in the eastern equatorial basin display evidence of interannual variability that cannot be explained by either the IOD or the IOBM alone. Instead, demonstrate that an equatorial mode, characteristically similar to ENSO, drove SST variability in much the same way as is currently observed in the equatorial Pacific. The disappearance and possible future re-emergence of this mode are related to changes in the mean state tropical climate . Few PMIP models are capable of capturing the changes in regional hydroclimate inferred from LGM proxies . However, those that can, also exhibit a weakening of the Indian Walker circulation induced by the exposure of the Sunda and Sahul shelves. Anomalous divergence and easterly wind anomalies over the exposed shelves then increase coastal upwelling off Indonesia, shoaling the thermocline and steepening the equatorial SST gradient . These conditions favour a mode of variability which, unlike the IOD, is characterized by a distinctively equatorial pattern with interannual warm and cool events of equivalent magnitude . Similar to a Pacific El Niño, its warm phase is preceded by a Kelvin wave in the thermocline that propagates from west to east. The projected response of the IO to greenhouse warming bears similarities to the simulated LGM, including a shoaling of the thermocline and a steeper equatorial SST gradient . It has therefore been proposed that this Indian Ocean “Niño-like mode” could re-emerge in certain emission scenarios .

Here, we define the Niño-like mode index as the standard deviation of SST anomalies in the eastern equatorial Indian Ocean (EEIO; Sect. ). demonstrate that SST anomalies are at best an approximation for the Niño-like mode, since the equatorial shift in the IOD in some climates can also contribute to SST anomalies in this region. Ideally, the mode would be identified by its atmospheric precursor: wind anomalies in the western basin. However, these fields are not universally available. The EEIO region overlaps in its bottom corner with that used to identify the southern-eastern pole of the IOD and is clearly contained in the basin-wide anomalies used to track the IOBM. Nonetheless, we note that the IOD influence over equatorial SSTs is small relative to the Niño-like mode in simulations by , and there is little relationship between the amplitude changes in IOD and EEIO (not shown). The amplitude of the Niño-like mode in each experiment is shown in Fig. a. The piControl ensemble exhibits a range of standard deviations between 0.18 and 0.53 ∘C, yet the ensemble median (0.37 ∘C) is rather close to that of the reanalysis (0.38 ∘C).

The most robust response is observed in the lig127k ensemble, in which 13 of 14 models agree on an increase in amplitude during the last interglacial. Only MIROC-ES2L simulates a decrease in variability (of 46 %) with the remaining models averaging a 17 % increase. In contrast, the other experiments show little to no agreement between models, with the mean amplitude going largely unchanged from the piControl. Even the lgm experiment, which is the only past climatic period previously analysed for the existence of this mode, has little agreement: the amplitude increases in six models and decreases in six models (the substantial increase from COSMOS-ASO does shift the ensemble mean in Fig. a though). Across PMIP, a roughly 1:1 relationship emerges between amplitude changes in variability within the EEIO region and across the whole basin (as measured by IOBM strength), which is statistically significant. This behaviour would be expected in the absence of the Niño-like mode. However, the relationship does not exist for the lig127k ensemble, where the consistent increase in EEIO standard deviation (Fig. a) occurs despite a decrease in variability across the whole basin (Fig. b). It is also interesting that the dominant period of interannual variability of in the EEIO region reduces under orbital forcing (not shown; see the “Code and data availability” section). These two observations provide further evidence that a Niño-like mode can emerge under certain climate states.

and both use CESM1.2 in their model analysis. However, its CMIP6 successor, CESM2, lacks an lgm experiment at the time of writing. Its predecessor, CCSM4, is included here and exhibits an 8 % decline in variability. However, explicitly note that CCSM4 fails to simulate the eastern drying and western wetting reconstructed from proxies, as do most CMIP5 models. They argue that this is due to its inadequate atmospheric response to shelf exposure, a key mechanism in the onset of the Niño-like mode. Although this is the first analysis of the Niño-like mode in CMIP6, we acknowledge that this systematic bias may persist into our results. An updated review of PMIP4 with respect to glacial hydroclimate proxies, similar to , would be required to clarify this point. Nonetheless, the robust increase in variability in the lig127k experiment presents an opportunity for the Niño-like mode to be studied in past climates other than the LGM.

How the Niño-like mode interacts with other modes of variability is relatively unexplored. The proximity of the Niño-like mode's equatorial arm to the IOD region off of Sumatra and Java and the observation that both modes peak in boreal autumn lead to the reasonable assumption that the two might influence each another through either changes in mean state or variability. link the future re-emergence of the Niño mode with a shoaling of the thermocline and enhanced upwelling in the eastern basin, conditions which resemble a positive IOD. Furthermore, there is a realistic possibility that ENSO could influence the Niño-like mode as it does the IOBM (Fig. b) and, to a lesser extent, the IOD (Fig. c). Figure b shows a statistically significant relationship between the mean equatorial SST gradient and the amplitude of the Niño-like mode in the lig127k experiment, which also passes through the origin. However, this relationship is not shared across the rest of the ensemble. Potentially this arises from the strong and consistent changes in the seasonality seen in the lig127k experiment, as this was shown to impact the IOD (Fig. c).

We also investigate influences on the relationship between the Niño-like mode and the IOD through an analysis of their correlation. Although demonstrated the independence of the Niño-like mode through its unique precursors, neither they nor quantified its synchronicity with the IOD. Here, we find that 88 of 114 simulations across all experiments exhibit a negative correlation between the Dipole Mode Index and the EEIO index, with a mean of correlation coefficient of -0.31. The strong positive IOD event in 1997–1998 was associated with ENSO and produced SST anomalies in the EEIO . However, there is no connection between changes in the correlation of the IOD and Niño-like modes and changes in either ENSO (Fig. c) or the IOD amplitude (Fig. d). We also find that the correlation between the DMI and EEIO index becomes stronger (i.e. more negative) in all models during the lig127k experiment. Distinguishing between a “switching on” of the Niño-like mode in the lig127k experiment and the IOD pattern bleeding into equatorial regions (Fig. c, similar to the strong positive IOD in 1997–1998) is challenging. Further detailed investigation into the role of changes in seasonality and phase locking of both modes and their atmospheric precursor would be particularly useful.