Introduction

The expected future slowdown of the Atlantic Meridional Overturning Circulation (AMOC) in response to anthropogenic climate change can have profound impacts on global and regional climate1,2,3. These impacts include variations of North Atlantic Ocean temperature and changes in its related predictability1,4, modulation of global air temperature2,5, regional changes of Northern/Southern hemisphere surface temperature6,7, changes of storm tracks and associated mid-latitude precipitation7, alteration of the monsoon systems including Sahelian rainfall8. Superimposed modes of North Atlantic sea surface temperature internal variability9 that are interconnected with AMOC10,11,12,13 may mediate some of these impacts4.

Recent modeling studies have shown that AMOC can react to accelerated warming in the tropical Indian Ocean (TIO) compared to the rest of the tropics on “fast” (monthly-to-decadal) or “slow” (multidecadal-to-centennial) timescales14,15,16. So far, studies primarily focus on the response of AMOC to persistent TIO warming, thus targeting potential changes under different levels of global warming and focusing on the Indian Ocean15. Variations of AMOC in response to internal (i.e., unforced) TIO surface temperature variations are unexplored as of yet. Here, we expand on existing research by using observational data sets as well as model simulation ensembles to further illustrate and document AMOC response to internal changes of tropical ocean temperature gradients generally and TIO more specifically.

Dynamic teleconnections originating from the Tropical Indian Ocean have been shown to influence Arctic and North Atlantic atmospheric patterns (i.e., the Northern Annular Mode & North Atlantic Oscillation, NAO)8,17,18 and atmospheric heat fluxes19,20 on monthly-to-decadal timescales. The NAO in turn affects both Arctic sea ice21 and AMOC22 variability. This mechanism is enhanced when relative warming of the TIO compared to the remaining tropics (relative TIO; rTIO) is considered15.

More recently, such atmospheric pathways have been shown to be complemented by oceanic pathways through the tropical Atlantic14,15. rTIO warming may drive tropical ocean temperature differences with the Atlantic and Pacific Ocean basins that locally enhances TIO precipitation and latent heat release in the atmosphere, inducing a global-scale Gill-type response that involves stationary Rossby and Kelvin waves. This strengthens the Walker circulation across the tropical Atlantic, resulting in increasing wind-induced evaporation and a northward shift of precipitation, ultimately increasing salinity throughout the tropical Atlantic basin. As a result, the tropical ocean temperature gradients related to rTIO warming induce positive density anomalies in the tropical Atlantic that are transported through ocean pathways to the North Atlantic and intensify AMOC some decades later (further details of the mechanisms e.g., in14,15).

The processes by which AMOC reacts to changes in anomalous TIO temperature have been highlighted in idealized coupled-model sensitivity experiments14,15 but not yet through observations or free model runs. Further, through their idealized setup, the previous studies targeted forced warming in the TIO and its implications for North Atlantic climate14,15,16. This is motivated by the evidence that the TIO has warmed by 0.15 °C dec−1 and also experienced enhanced warming of 0.05 °C dec−1 compared to the rest of the tropics since 196014. However, there was no detailed breakdown of the involved tropical ocean temperature gradients to the diagnosed rTIO-AMOC connection. In addition, unforced internal climate variations are important to understand global teleconnections and to forecast near-term climate variations on the decadal time scales and the associated impacts23,24 beyond the previously studied forced rTIO-driven AMOC intensification. Specifically, understanding and accounting for internal variability in the North Atlantic has been shown to be a potent tool in the attempt to predict climate over the adjacent continents on the seasonal-to-decadal time scales2,25,26. Since some modes of tropical Indian Ocean SST variability exhibit pronounced multi-annual variations27 which carry some decadal predictability28, the rTIO-AMOC link is a promising mechanism to better understand AMOC variations and predictability.

In this work, we examine internal rTIO-AMOC teleconnections diagnosed from observational data sets, following the methods of previous literature, to show how the two ocean basins are interconnected. We also take the opportunity to “zoom out”, analyzing in detail which ocean regions and tropical ocean basins contribute to the tropical temperature gradients that govern the rTIO-AMOC mechanisms from the literature. We then analyze model simulations to understand the robustness of the internal rTIO-AMOC connection. Finally, we draw conclusions on the importance of these internal mechanisms in the face of anthropogenic global warming, offering alternative interpretations of SST-based observed AMOC fingerprints in the context of internal variability.

Results

Observed tropical ocean SST teleconnections to AMOC

Direct observation of the influence of tropical ocean temperature gradients on AMOC is inhibited by a lack of long-term direct AMOC observations29,30. In this context, a North Atlantic SST index was proposed as a proxy for AMOC fluctuations10,31. This index is defined as the mean extended winter (November-May: NDJFMAM) difference between North Atlantic subpolar gyre and global mean SST31. It projects onto the annual mean AMOC31. Since SST has been observed and reconstructed for much longer than AMOC, we here invoke this observed SST-based AMOC index (henceforth SST AMOC index; SSTAMOC) to reconstruct AMOC variations and examine the tropical ocean—AMOC teleconnection mechanism in observational data sets. Furthermore, the objective of this study is to investigate the internal variability, separate from the anthropogenic forced signal. We separate the forced component from unforced internal climate variability using the residuals method that subtracts the multi-model ensemble mean of historical simulations from observations (see Methods)32 to focus on unforced effects. To this end, we estimate the forced signals from CMIP6 historical simulations (see Methods)32,33,34. Unless explicitly stated otherwise, the results we describe henceforth will be interpreted as unforced residuals, i.e., internal variability.

Continuing from previous literature14,15, we first define a rTIOcov index. We define this index as the TIO temperature minus the remaining tropical ocean temperature, weighted by its covariance with TIO temperature (TIO-TOcov, see Methods). Compared to the previous method, we thus now account for the co-variability between the indian ocean and the other tropical basins to focus on the signal related to the TIO and reduce contamination of our results from other decadal-to-multi-decadal climatic modes.

The full observed TIO, rTIOcov and SSTAMOC all display pronounced decadal-scale variability, as well as an accelerating trend over the later decades due to external forcing of the climate system (Fig. 1a–c). This includes both the unforced internal and the forced signal. A trough and peak are observed in 1890 and 1910 in rTIOcov, and in 1910 and 1930 in SSTAMOC. We find a temperature drop and subsequent recovery between 1920 and 1960 in rTIOcov and a substantial drop between 1940 and 1980 in SSTAMOC. The observed TIO and rTIOcov indices show a significant negative correlation with SSTAMOC when the latter leads by approximately 20 years (Fig. 1d, e; r = −0.92 and −0.60 respectively, solid black).

Fig. 1: Temperature time series for the Tropical Indian Ocean and the SSTAMOC index, as well as their cross-correlations in reanalysis data sets.
figure 1

Time series of the observed ERSST (black) and decomposed into the forced (red) and unforced internal (blue) components for the (a) tropical Indian Ocean (TIO) sea surface temperature (SST), (b) relative tropical Indian Ocean SST weighted by covariances (rTIOcov), and (c) AMOC SST-fingerprint index SSTAMOC. Thin lines represent the annual mean and the thick line the 21-year moving mean. d, e The lag-lead correlations of TIO and rTIOcov with the SSTAMOC for the observed (black), forced (red), and unforced (blue) filtered ERSST time series, respectively. The vertical blue line represents the time-lag of the peak correlation of the unforced signal (e). f The lag-lead correlation between the rTIOcov and SSTAMOC unforced indices for ERSST v5 (blue), HadiSST v1 (purple), and COBE v2 (orange). The vertical magenta line represents the mean of the maximum correlations between the three observational indices, the shading represents the 95% confidence interval of upper and lower bound uncertainty, and black dots highlight correlation significantly different from 0 at the 95% confidence level based on a Student’s t test and according for smoothing (see Methods). The years used for each observational data product is from 1871 to 2013.

Full size image

To improve our understanding of the underlying connections, we deconstruct the observed signal into a forced (red) and an unforced (blue) component. The forced TIO and rTIOcov signal are steadily warming, while forced SSTAMOC cools. The latter suggests AMOC weakening over time (red lines in Fig. 1a–c) related to anthropogenic climate change (as also argued in31). It is worth noting that, as estimated from the SSTAMOC proxy with forcing estimation from climate models, this study cannot make any statements about the emergence of AMOC weakening in observations. The unforced signal of TIO and SSTAMOC have a negative correlation maximum when the TIO lags by 17 years (Fig. 1d, solid blue, r = −0.80) and the modulations of rTIOcov and SSTAMOC have a positive correlation maximum when rTIOcov leads by 25 years (Fig. 1e, solid blue, r = 0.89). We do not find a significant correlation between the rTIOcov-SSTAMOC at short time lags, indicating that the mechanism that dominates the unforced signal in observations acts on multidecadal time scales.

The multidecadal relationship with rTIOcov leading AMOC is robust when considering sub-periods of the time series (Supplementary Fig. 1), indicating that it is a robust relationship that is not just dominated by the large drop in the SST time series around the 1960s. The time scale of some 25 years aligns with published estimates of a dynamic rTIOcov-AMOC pathway in climate models14,18,35. It therefore seems that the rTIOcov-AMOC connection is not only a forced response, as suggested by previous studies, but also a feature of unforced variability.

The decomposition of the relationship between rTIOcov and SSTAMOC shown here with ERSST v536 is robust in HadiSST v137 (r = 0.93, 27 years) and COBE v238 (r = 0.66, 27 years) (Fig. 1f and S1), with an average positive correlation maximum when the unforced rTIOcov leads by 26 ± 7 years (mean ± 99% confidence interval). Since the relationship of rTIOcov and SSTAMOC in response to global warming is opposite in sign, but the unforced correlation shows the described positive maximum (Fig. 1c) with some lags, processes involved in internal variability are likely different from those related to anthropogenic forcing, which warrants further exploration.

To better understand the tropical temperature gradients that precondition the influence of rTIOcov on AMOC, we set aside the proposed rTIOcov index and propose to further understand spatial patterns of the tropical temperature gradients using a lagged correlation analysis (Fig. 2). Observations show significant positive tropical SST correlation to SSTAMOC in the TIO and tropical eastern Pacific Ocean when tropical SST leads by more than 20 years. The tropical Atlantic Ocean (TAO) is significantly negatively correlated to SSTAMOC at the same time lag. This suggests that a tropical SST gradient involving the Indian and eastern Pacific (IEP) on the one hand, and the Atlantic on the other hand, may induce the slow mechanism. These patterns are robust across all three observational products and several time lags (not shown). We exclude the tropical western Pacific as it shows no robust pattern amongst the three observed products. From the identified IEP-TAO-TIO pattern, we can construct alternate indices of tropical SST gradients defined as the IEP minus the TAO weighted by their covariance (IEP-TAOcov). TIO-TAOcov and EP-TAOcov. These indices are examined to identify a potential dominance of one the tropical basins linking to SSTAMOC. Comparing the proposed tropical SST gradient metrics with the rTIOcov (Supplementary Fig. 2), the strongest correlations with SSTAMOC are still found with the rTIOcov approach (TIO-TOcov). Within the TGSST metrics, the TIO-TAOcov produces the highest correlations while the EP-TAOcov the lowest of the three within all three sets of observational products (Supplementary Fig. 3). Although the spatial pattern suggests an IEP-TAO relationship (Supplementary Fig. 3), this analysis suggests the importance of the TIO and TAO in determining the TGSST with potentially less importance of the EP. Yet, due to the overall highest and most consistent correlations of the rTIOcov with SSTAMOC among the tropical ocean temperature gradient indices in all three observations, we henceforth define the rTIOcov index as TIO-TOcov for comparison within coupled models (see Methods and Supplementary Fig. 2). To further understand its robustness and underlying processes, we now investigate the internal rTIOcov-AMOC relationship in climate model simulations in the absence of forcing.

Fig. 2: Correlations between the AMOC SST-fingerprint index SSTAMOC and SST fields, when SST leads SSTAMOC by 25 years, for the unforced signal in ERSST v5, HadiSST v1, and COBE v2.
figure 2

For each correlation, the time period 1871–2013 was used with a 21-year moving mean applied. A 25-year lag was selected based on the results in Fig. 1 and similar spatial patterns result when changing the lag time by ±5 years due to applying a 21-year moving mean to identify the low frequency signal. Hatching represents those spatial correlations that are not significant at the 95% confidence interval, taking into account the smoothing and autocorrelation.

Full size image

Unforced rTIO-AMOC connections in models

We examine pre-industrial control (piControl) simulations from 44 different models from the Coupled Model Intercomparison Project Phase 6 (CMIP6)39 archive. This kind of model simulation allows assessing climate variability due to internal processes only, as external forcings are kept constant to their 1850 estimated values. As in the analysis of observations, we analyze low-pass filtered model output. To account for smoothing-related end-point effects, we do not analyze the first or last 100 years of the simulations. The examined CMIP6 simulations produce a rTIOcov-SSTAMOC correlation maximum at 26 ± 6 years (99% confidence interval, Fig. 3), with a subset of 28 out of the 44 CMIP6 models within the observed estimate of 26 ± 7 years. In fact, the time lag in the CMIP6 model subset is on average 27 ± 2 years and shows a significant multi-model correlation maximum (r = 0.50) within the observed time lag (Fig. 3a; see Supplementary Table 1). The majority of models thus confirms the observed internal correlation. We now examine this subset of models more closely.

Fig. 3: The relationship between rTIOcov and SSTAMOC in the CMIP6 model archive.
figure 3

Correlations of (a) rTIOcov-SSTAMOC across different time lags for observations from Fig. 2 (thick blue, purple and yellow lines) and the different CMIP6 model pre-industrial control simulations. Those models that show a correlation maximum greater than the 0.30 threshold from bootstrapping and within 99% confidence interval of the observed rTIOcov-SSTAMOC lag of 26 years are shown in thin blue lines (CMIP6 subset), while those in thin gray are those outside the observed range. The thick red line represents the IPSL-CM6A-LR model shown in more detail in Fig. 4 and the thick black line is the CMIP6 subset model mean, both estimated using Fisher’s Z-transformation. b A scatter plot of the maximum rTIOcov-SSTAMOC correlation and time lag for the mean relationship in each CMIP6 model. Those of the CMIP6 subset are plotted in circles and those not in the subset are denoted with an “x”. The colorbar depicts mean AMOC strength at 45 °N within the model, which is only available for a few of the subset models and the 3 standard deviation thresholds are denoted with dashed magenta lines. The observed ERSST v5, HadiSST v1, and COBE v2 correlations are denoted with yellow stars within (b). A complete list of the corresponding models can be found in Supplementary Table 1. c, d as (a), but for rTIOcov-AMOC45N and AMOC45N-SSTAMOC, respectively. The thick black line indicates the mean correlation value across all models. Potential temporal offsets between correlation maxima are reflected in the black error bars, which represent the mean correlation value, mean maximum correlation time lag, and the respective 1–99% confidence intervals of the correlation peaks identified in each of the CMIP6 models using Fisher’s Z-transformation.

Full size image

Climate model simulations have the benefit of providing complete information about the (simulated) climate system, so that the actual AMOC can be diagnosed. We find that the aforementioned subset of models (28 of 44) also shows on average positive correlation between rTIOcov and AMOC at 45°N (AMOC45N) when rTIOcov leads by 17 ± 5 years (r = 0.43) (Fig. 3a–c). The observed ERSST value of r = 0.89 lies in the highest decile of the model correlation values (Fig. 3b). The AMOC45N and SSTAMOC relationship is very robust in this subset of CMIP6 piControl simulations: their average correlation is 0.60 when AMOC45N leads by 4 ± 3 years (Fig. 3d). At lower latitudes the AMOC lead becomes shorter (e.g., AMOC26N shows a maximum correlation to SSTAMOC at 2 ± 3 years, r = 0.80; not shown), illustrating the AMOC dynamics of subsurface southward propagation of AMOC anomalies11 and surface features represented by SSTAMOC. Overall, these findings highlight the robustness of the SSTAMOC fingerprint for estimating AMOC.

The low correlation values in models compared to observations can be partly attributed to the length of time series. The 44 piControl simulations have a length of between 300 and 2000 years, and therefore represent a multitude of possible climatic states or regimes, not all of which are necessarily representative of the relatively short observed period. In each model, we thus consider the 150-year long slices of the piControl simulations that show the maximum rTIOcov-SSTAMOC correlation within the observed time lag to mimic the length of the observational period (Fig. S3). 31 of 44 (~70%) models show time slices that lie within the observed time lag estimate of the rTIOcov-SSTAMOC relationship. In this analysis, the CMIP6 average maximum correlation values (black error bars in Fig. S3) of rTIOcov-SSTAMOC, rTIOcov-AMOC45N, and AMOC45N-SSTAMOC are 0.65 at 26 ± 2 years lag, 0.67 at 19 ± 6 years lag, and 0.81 at 7 ± 4 years lag, respectively. We find that most of the examined CMIP6 models are capable of reproducing the observed unforced rTIO-SSTAMOC relationship in at least one 150-year time slice. Yet, while the SSTAMOC-AMOC45N relationship is found to be robust as well, the rTIOcov-AMOC45N relationship is not coherent across all models, probably due to non-linearities and competing effects in the relationships.

We detail the specificities of the proposed rTIO-AMOC link in the most recent version of the IPSL climate model, the IPSL-CM6A-LR model (Fig. 4)40. In agreement with observations and the majority of examined CMIP6 models, the IPSL-CM6A piControl simulation shows a rTIOcov-SSTAMOC correlation peak when rTIOcov leads by 26 ± 4 years (Fig. 4a). This average peak is much lower than observed (0.36 vs. 0.89 in ERSST), yet significant. In 150-year long slices of the piControl simulation, the intermittent rTIO-SSTAMOC correlations reach values of up to 0.56 and a standard deviation of 0.12 between the slices when rTIOcov leads by 26 ± 4 years (Figs. 4, S4, S5). Moreover, values can approach correlations of 0.71 within IPSL-CM6A-LR, although occurring when lags are >33-years and exceeding the threshold based on the observed products, suggesting a potential relationship to modeled mean AMOC strength in determining the lag.

Fig. 4: A detailed view of the rTIOcov-AMOC relationship in the IPSL-CM6A-LR model.
figure 4

Lag-lead correlations between the (a) rTIOcov-SSTAMOC (red), (b) rTIOcov-AMOC45N (black), and (c) the AMOC-SSTAMOC (blue) indices in the IPSL-CM6A-LR piControl simulation. The thin lines represent individual 150-year segments, and the thick lines are the mean correlations of all segments, estimated from a Fisher’s Z-transformation. The crosses/error bars represent the standard deviation of the maximum correlation lag in all 150-year long segments of the 1200-year piControl run for correlations between the three different indices. The black error bars represent the approximate mean correlation and the respective 99% confidence interval of the maximum correlation lag in all 150-year long segments of the 1200-year piControl run for correlations between the three different indices.

Full size image

The 1200-yr long IPSL-CM6A-LR simulation shows a significant correlation (r = 0.59) between SSTAMOC and AMOC45N when AMOC45N leads by 6 years (Fig. 4b). Correlation between rTIOcov and AMOC45N is significant (r = 0.46) when rTIOcov leads by 19 years. These time lags align with those in observations and CMIP6 models. In 150 year time slices, the maximum correlation between rTIOcov and AMOC45N is 0.70 (standard deviation between time slices = 0.20) when rTIOcov leads by 19 ± 2 years. We find a similar spread between time slices in AMOC45N - SSTAMOC correlations (r = 0.70, standard deviation = 0.12) when AMOC45N leads by 6 ± 2 years (Fig. 4c). As a result, the internal relationship between rTIOcov temperature and AMOC45N in IPSL-CM6A-LR is weak and intermittent, but robust when rTIOcov leads by roughly 20 years.

The spatial signal of rTIO-AMOC connections

We now examine the physical pathways that connect the tropical ocean temperature gradient to SSTAMOC in the IPSL-CM6A-LR climate model. rTIOcov warming enhances geostationary Rossby wave generation in the TIO region through enhanced precipitation and latent heat release18,41, simultaneously driving a negative NAO-like pressure pattern in the North Atlantic alongside a southward shift in zonal wind stress over the subpolar Atlantic and an adjustment of meridional wind stress (Supplementary Fig. 5a–d). Ekman pumping in the subpolar North Atlantic responds to the atmospheric anomalies by increasing in magnitude, resulting in warmer, saltier, and denser waters (Supplementary Fig. 5e–h) within the eastern subpolar North Atlantic and deep convective regions. These climatic conditions persist for 10–15 years (Supplementary Fig. 6), demonstrating the low-frequency patterns in phase with the rTIO as also shown by35. Such prolonged NAO conditions impact AMOC, leading to changes in the entire water column (Supplementary Fig. 7) and southward propagation of an AMOC anomaly after 20 years (Supplementary Fig. 8)35. This AMOC anomaly feeds back onto SST, influencing the SSTAMOC index at a time lag of around 5 years (Supplementary Fig. 6f), which explains the overall observed and simulated time lag between rTIOcov and SSTAMOC, and explains differences in lags of the rTIOcov and AMOC45N ( ~ 19 years) and SSTAMOC ( ~ 26 years) in the IPSL model (Fig. 4) which we also demonstrated with CMIP6 (Figs. 3, S3).

Although some of our findings show characteristics of an alternative mechanism that describes a rTIO influence on the TAO42 and a propagation of that signal into the subpolar North Atlantic, we fail to clearly diagnose this mechanism in this work. This suggests that the direct connection to the North Atlantic via the NAO dominates the internal influence of rTIOcov on AMOC.

Discussion

A key assumption of this work is the use of the SST-based SSTAMOC index to estimate observed internal AMOC changes. This index has been designed to describe the full AMOC modulations and trends, but its suitability for addressing internal AMOC variations had to our knowledge not been assessed. Debates continue on the extent to which this index describes forced31 vs. internal atmospheric variability43 in the observed record. Using CMIP6 model simulations, we here show that there is overwhelming model agreement on the unforced internal correlation of SSTAMOC and AMOC45N when AMOC45N leads by around 6 years (Fig. 3d; as also suggested by44). The 150-year long time chunks in CMIP6 piControl simulations encompass correlation values of 0.60 ± 0.27 (0.81 ± 0.37 for the 44 best 150-year chunks), indicating that the use of the SSTAMOC index to assess internal fluctuations of AMOC45N is justified.

As discussed further up, various tropical ocean temperature gradients could have been chosen to be examined for their influence on AMOC in this study (cf. surface temperature patterns in Fig. 2). Comparing the observed influence of different options (Supplementary Fig. 2) we found that all selected indices show comparable results. Here, we chose to focus on rTIOcov due to its highest correlation to SSTAMOC, and because of its presence in published literature. The rTIOcov index is considered here to be representative of all examined tropical ocean temperature gradients, and our results are robust across these definitions of tropical ocean temperature gradients (Figs. 1, 2, S2, S3, and S5). Other indices potentially contribute to AMOC modulation through teleconnections, and further work is necessary to derive the exact mechanisms at work.

Our work assumes that the CMIP6 multi-model mean of historical simulations appropriately estimates the forced response of the observed climate system. Results presented here are robust to using the IPSL model single model ensemble mean instead of the multi-model mean for the forced response (Supplementary Fig. 9a, b), thereby giving credit to the assumption that an ensemble mean of historical simulations is a relevant estimation of the effect of external forcing on climatic variables. While models may not appropriately represent the response of the climate to forcing, which is a clear limitation to any study examining climate model ensemble averages, the CMIP6 historical simulations are the best available estimate of the forced response of the climate system at this time which, unlike statistical detrending, accounts for abrupt climate events such as volcanic eruptions or aerosol concentration changes. That being said, some evidence of powerful statistical detrending methods exist45, which go beyond what can be covered in this paper.

As we find the rTIOcov-AMOC relationship in about 60% of the CMIP6 historical members—a corresponding analysis using only the IPSL model historical ensemble, subtracting the IPSL ensemble mean to remove the forced response, shows comparative results (19/33 members reproduce the mechanism) –, this illustrates both that it is the dominant mechanism in these models and that it is inherently intermittent. Such intermittency is consistent with46. An analysis of the intermittency of the relationships of rTIOcov, SSTAMOC and AMOC45N (Figs. 4, S4) across the IPSL piControl simulation illustrates that the rTIOcov-AMOC45N relationship breaks down when the AMOC45N-SSTAMOC relationship decreases, while the teleconnection between rTIOcov and the SSTAMOC remains relatively stable. This implies changes in the North Atlantic climate, potentially related to NAO, that break down the AMOC45N-SSTAMOC relationship as probable causes of the intermittency in the rTIOcov-AMOC45N relationship46, while the interbasin atmospheric teleconnection remains intact.

Only a subset of CMIP6 models appear to consistently reproduce the observed internal rTIOcov-SSTAMOC relationship. Here, the non-stationarity of the relationship could be one possible reason, as the examined simulations cover different time lengths, and differences in the surface and subsurface features could differ in the different periods examined. Another potential reason is that some models might not capture the teleconnection mechanism due to coarse resolution, mean biases, the signal to noise problem in global climate models47, issues with mixing in the North Atlantic in some models48, or missing representation of the stratosphere which is crucial to simulating atmospheric waves49, along many other possible reasons. Alternative interpretations of this caveat are that the observed rTIOcov-SSTAMOC relationship is spurious in observations and erroneously reproduced by some climate models. Model differences that cause these discrepancies in the rTIOcov-AMOC45N mechanism will be important to study in future research.

This work could not clearly separate the influence of rTIOcov on AMOC45N from other factors influencing the AMOC45N internal variability. Given the relatively short observational record, the rTIOcov connection of AMOC might well be related to known inter-basin teleconnections50, part of a self-containing mechanism of AMOC variability51 or another teleconnection via the NAO18,35,52. However, this study implies a tropical SST gradient as a potentially important driver or at least pacemaker of AMOC changes, identifying additional TGSST indices that strongly rely on the variability of the tropical Indian and Atlantic Oceans. Disentangling the different drivers of AMOC variability is an exciting scientific question for the future.To conclude, we have analyzed several observational data sets as well as global climate model simulations from the CMIP6 archive to establish a pathway by which low-frequency internal, i.e., unforced, changes of tropical ocean temperature gradients may influence AMOC several decades later, leading to AMOC multi-decadal variability. This pathway includes shifts in geostationary waves that impact the NAO, changing surface winds with impacts on salinity and temperature distributions that change Ekman pumping, impacting AMOC. In models, internal rTIOcov warming strengthens AMOC ~ 20 years later. We also find that across model simulations a SST-based index that was proposed to indicate forced AMOC changes10,31 is reflective of internal AMOC changes ~6 years earlier, allowing a tracking of AMOC changes using relatively easily observable surface water characteristics. This study implies that rTIOcov temperature is an important pacemaker of AMOC variability as well as a potential early warning indicator of multi-decadal AMOC change.

Methods

Observations

As one estimate of observational uncertainty, we analyze the gridded observational data sets/reanalysis products ERSST v536, HadiSST v1.137, and COBE v238 for the period 1870–2014. All gridded products are regridded to a regular 1 × 1 degree grid prior to analysis, and anomalies to their respective long-term mean states are formed.

Models

We detail the mechanisms discussed in this paper in two different model setups. To find the response of the climate system to forcing, we analyze a 44-model ensemble-mean of historical simulations from CMIP639. The models are detailed in Supplementary Table 1. Prior to forming the multi-model mean, individual model ensemble means are calculated (i.e., we follow the one-model-one-vote approach53), regridded to a regular 1 × 1 degree grid, and then anomalies to their mean state formed to account for model mean bias. These model simulations are all characterized by the same forcing but utilizing different model setups and different initial conditions. The multi-model mean can therefore be regarded as a “best estimate” of the forced response of the system.

We also analyze the unforced physical pathways that connect Indian Ocean SST to the North Atlantic and AMOC in pre-industrial control simulations with the same GCMs (Supplementary Table 1). These model simulations are not subject to forcing and therefore represent the respective model’s interpretation of internal climate variability. Some CMIP6 GCMs show a pronounced centennial fluctuation in global SST, which may project onto local SST variability40. To account for this, we high-pass filter all model output that we analyze at a 100-year cut-off frequency.

Methods

In this paper, we analyze annual and boreal extended winter (November-May, NDJFMAM) mean SST, SSS, precipitation, evaporation minus precipitation and 500 hPa geopotential height and annual meridional overturning streamfunction, which are then—unless otherwise noted—low-pass filtered with a 21-year running mean. Two specific indices are considered in our analysis. First, we average annual mean SST in the tropical Indian Ocean (30° S–30° N, 40° W–100° W), and then subtract SST in the remaining Tropics (TO; 30° S–30° N) multiplied by the weighted covariances between the two to calculate the rTIOcov index, similar to14,15 but now accounting for the covariances of the tropical Indian SST modulations with the rest of the tropical ocean. This is illustrated in Eq. (1):

$${{\rm{rTIO}}}_{\mathrm{cov}}={\rm{TIO}}-\left({\rm{TO}}\times \frac{\mathrm{cov}({\rm{TIO}},{\rm{TO}})}{\mathrm{cov}({\rm{TO}},{\rm{TO}})}\right)$$
(1)

Other tropical basins similarly use 30° S–30° N as boundaries for the tropics and the EP is defined from 150 °E to 80 °E. As shown in Figs. 2 and S1, there are several tropical regions that could demonstrate a tropical gradient that relates to the SSTAMOC. We have chosen to define the rTIOcov as the gradient between the TIO and TO because of its use in previous literature, and because in observations we find the highest correlation to SSTAMOC when comparing with other indices for tropical temperature gradients (Supplementary Fig. 1). Several additional gradients that are based on the tropical Indian and Atlantic Oceans suggest that the potential key interactions rely on these two basins, with some influence from the tropical East Pacific. Second, in the face of a shortage of AMOC observations that go back in time further than 2005, we calculate the AMOC SST-fingerprint following the methods proposed in10,31 as a proxy. This SSTAMOC index is calculated by subtracting winter global mean SST (60° S–70° N) from subpolar North Atlantic winter SST, as defined in31, but the idea of a SST fingerprint of AMOC has been similarly been described previously11,54,55.

There are numerous approaches in the literature to remove the forced signal from models and observations32,33,34. We here analyze observed unforced climate variability, calculated following the “residuals”-approach32. This technique rescales the historical multi-model ensemble mean variability to the observed value in order to estimate the forced component using the ratio of observed and multi-model mean standard deviations32, and then subtracts on a grid-point basis the forced component from the full signal. The residual signal is treated as observed unforced internal variability in this study26. To compare correlations between the various temperature and AMOC indices in the models, the simulations are first high-pass filtered using a Butterworth filter of a 100-year cut-off frequency to remove a known 100-year centennial mode40, and then smoothed using a 21-year moving mean to remain consistent between the methods of the SSTAMOC and the other indices. Pearson correlation analysis is performed at different temporal lags to find statistical relationships between variables and tested for statistical significance using a bootstrapping of 1000 iterations and an alpha of 0.05. Given 11 years of diagnosed significant autocorrelation in the filtered time series (not shown), we account for autocorrelation by bootstrapping the data by blocks of 11 years. Given the sample size, significance threshold, and accounting for autocorrelation, the threshold of a significant correlation at a lag of 26 years (chosen from the observed lag) is 0.30. To help compare the lag-lead correlations of the indices, we transform them from a logarithmic to a linear scale using a Fisher’s Z transform to average correlations. However, the lag-lead correlation values vary in time and magnitude with index comparisons. To find the mean correlation of a lag-lead relationship, we identify the time of maximum peak or trough of the signal and use the Fisher’s Z transformation on the peaks/troughs to average the transformed correlation values centered around the maximum/minimum correlation.