Introduction

One of the salient characteristics of the tropical Pacific is a large sea surface temperature (SST) difference between warm waters in the western Pacific warm pool and cold waters in the central-to-eastern tropical Pacific1,2. This zonal SST gradient in the mean state is further enhanced during La Niña events, whereas it weakens substantially during El Niño events. Considering that the tropics can significantly affect extratropical weather and climate via atmospheric teleconnections2,3,4,5,6 and that the characteristics of El Niño-Southern Oscillation (ENSO), which is the major predictability source1,2, are closely connected to the tropical Pacific mean state1,2,7, the tropical Pacific SST response to sustained anthropogenic emissions is a component of the climate response that is of particular importance5,8,9,10,11,12,13,14,15. Under sustained anthropogenic warming, a majority of earlier and current generations of climate models tend to produce a weakened zonal SST gradient across the tropical Pacific8,16,17,18,19,20, which is linked to a weakening of the tropical atmospheric circulation resulting from increased dry static stability21 and/or differential atmospheric moisture and precipitation responses22. In contrast, the observational record over the satellite era indicates a distinct strengthening characterized by enhanced warming in the northwest/southwest Pacific but modest cooling in the tropical central-to-eastern Pacific6,10,11,12, which is accompanied by intensified Hadley and Pacific Walker circulations12,23,24,25. Given the close connection between SST and deep convection in the tropics, this model-observation discrepancy has led to a questioning of the credibility of model-projected future climate change12,13,17,18,20.

The enhanced zonal SST gradient that has developed over the satellite era may have resulted from greenhouse gas forcing via ocean memory, with this manifested through either the upwelling of cold subsurface waters along the equatorial Pacific (i.e., an ocean dynamical thermostat mechanism)17,26, a weakening of the amplitude of non-linear ENSO fluctuations27, a strengthened atmospheric circulation12, a greater warming over the Atlantic and/or Indian Ocean23, or an enhanced Southern Ocean cooling10,11,28. Such processes may not be accurately represented in climate models, with this resulting in model-observation discrepancies17,18,20,29. Hence, it is possible that there will be La Niña-like SST gradient changes in the future8,17. However, it cannot be ruled out that the enhanced zonal SST gradient has arisen, at least in part, from a transient delay in eastern Pacific warming19, non-greenhouse gas forcing such as Antarctic stratospheric ozone depletion14,30 and anthropogenic aerosols19,30,31,32, or internal variability30,33,34. These aspects imply that multi-decadal timescales such as the time interval that spans the satellite era might not be long enough to confidently determine the forced response of the tropical Pacific SST to greenhouse gas forcing. This issue may be resolved by computing centennial trends from reconstructed SST data sets. Potential uncertainties in these data sets, stemming from sparse in-situ measurements in both time and space, especially prior to the 1950s, as well as changes in measurement practice and instrumentation over time, however, hinder unambiguous determination35,36,37,38. While some data sets were found to exhibit a reduced zonal SST gradient consistent with a weakening of the Pacific Walker circulation inferred from independent observations of sea level pressure (SLP)39 and surface winds40, others exhibit a distinct cooling in the central-to-eastern equatorial Pacific.

In order to take into account the limitations presented by the finite duration of the time of extensive data availability and potential model deficiencies in assessing the zonal SST gradient change resulting from greenhouse gas forcing, we utilized SST reconstructions from paleoclimate proxy records over the past four million years. The utility of the SST reconstruction stems from the large atmospheric carbon dioxide (CO2) range over the late Cenozoic period, with the magnitude of CO2 changes comparable to that of projected future changes41,42,43, and they can thereby provide a unique opportunity for evaluating changes in the zonal SST gradient and thereby help to constrain model performance41,42. In particular, the mid-Pliocene is the most relevant distant past period because atmospheric CO2 concentrations of ~400 ppm were much higher than under pre-industrial condition41,44. Paleoclimate proxy data over the historical period, rather than over the distant past, might be employed to determine the forced response of the tropical Pacific SST to sustained global warming. However, unambiguous attribution to greenhouse gas forcing may not be feasible due to influences of human-induced non-greenhouse gas forcing agents14,19,30,31,32 and low-frequency internal climate variability30,33,34 over the post-war period.

In this study, we first examine whether the time duration of either the satellite or post-war era, over which high-quality SST observations exist, is long enough to unambiguously determine the forced response of tropical Pacific SST to greenhouse gas forcing. Then, we analyze paleoclimate SST records during the mid-Pliocene relative to pre-industrial and assess whether model-simulated changes in the zonal SST gradient under sustained global warming (also under sustained global cooling in the opposite sense) are in line with the change inferred from paleoclimate proxy records. Finally, we explore the temporal evolution of the zonal SST gradient and corresponding global-mean surface temperature over the past four million years in conjunction with a variety of model simulations to determine the sign of the forced response of the zonal SST gradient to large greenhouse gas forcing.

Results

Limitation of instrumental record

We begin by examining observed SST trends over the tropical Pacific. The ERSST5 and HadISST datasets (Methods) indicate a slight cooling in the central-to-eastern tropical Pacific over 1979–2023, which is in stark contrast to pronounced warming in the northwest and southwest Pacific and the western Pacific warm pool (Fig. 1a,b). If this La Niña-like SST trend pattern is a forced signal resulting from anthropogenic global warming, a similar but more pronounced spatial pattern may be expected to have emerged over an extended historical period, such as spanning back to the mid-20th century. However, the spatial pattern over 1950–2023 is noticeably different from that over 1979–2023 and moreover there are disagreements between the reconstructed datasets (Fig. 1c,d). Unlike ERSST5, the HadISST dataset does exhibit an enhanced zonal SST gradient across the equatorial Pacific, and this could reflect an ocean dynamical thermostat mechanism17,26. However, the related muted warming or modest cooling in the central-to-eastern equatorial Pacific and tropical southeastern Pacific is not statistically significant. Furthermore, corresponding trends in SLP estimated from reanalysis datasets (Methods) indicate a reduced zonal gradient over the tropical Pacific over the post-war period (Supplementary Fig. 1c,d), implying a greenhouse gas forcing-induced weakening of the Pacific Walker circulation39,40, which is at odds with a distinct strengthening over 1979–2023 (Supplementary Fig. 1a,b). It is also noted that surface zonal winds over the central equatorial Pacific exhibit inconsistent trends between the period 1979–2023 and the post-war period (Supplementary Fig. 1, vectors). A previous study38 also showed that instrumental SST records underestimate the 20th century warming trend in the eastern Pacific cold tongue. This raises the question as to whether the La Niña-like SST trend pattern over the satellite era can be attributed solely to greenhouse gas forcing, independent of low-frequency internal variability and non-greenhouse gas forcing14,30,31,32,33,34.

Fig. 1: Uncertainty in the attribution of observed SST trends over the tropical Pacific.
figure 1

a Annual-mean SST trends (units: K decade−1) over the period 1979–2023 computed from ERSST5 with stippling indicating statistical significance of the computed trends at the 95% confidence level. b Same as in a but for HadISST. c Same as in a but over the period 1950–2023. d Same as in b but over the period 1950–2023.

Full size image

A potential role for low-frequency internal variability and non-greenhouse gas forcing in the La Niña-like SST trend pattern over the satellite era leads us to explore whether the forced response of the zonal SST gradient to anthropogenic forcing has already emerged in large ensemble model simulations (Methods). We identify the timescales over which the forced signal emerges from the background internal variability by comparing the magnitude of the ensemble-mean anomaly relative to the 1950–1969 climatology with two standard deviations across the ensemble members of a given model, with the zonal SST gradient defined, following ref. 34, as the SST difference between the western equatorial Pacific (5°S-5°N, 110°E-180°) and the eastern equatorial Pacific (5°S-5°N, 180°-80°W). For the case of CanESM5 under historical forcing (up to year 2014) and subsequently the Shared Socioeconomic Pathway (SSP) 3-7.0 (2015–2100) scenario of the Coupled Model Intercomparison Project phase 6 (CMIP6), the ensemble-mean anomalies exhibit a gradual weakening of the zonal SST gradient over time (Fig. 2a), with the ensemble-mean SST change averaged over 2081–2100 clearly indicating more pronounced warming over the eastern equatorial Pacific than over the western equatorial Pacific (Fig. 2b). However, the magnitude of the ensemble-mean change is mostly smaller than the inter-ensemble spread (Fig. 2a), implying the dominance of internal variability over the forced response over the satellite or post-war periods. Although the CESM2 Large Ensemble (CESM2 LE) also projects a distinct reduction in the zonal SST gradient at the end of the 21st century (Fig. 2e), the temporal evolution of ensemble-mean change noticeably differs from CanESM5. The ensemble-mean zonal SST gradient increases from the 1970s to the early 2010s (Fig. 2d), which was attributed to non-greenhouse gas forcing, especially, Antarctic stratospheric ozone depletion and changes in anthropogenic aerosols30, although a marked reduction of the zonal SST gradient eventually follows.

Fig. 2: Inter-model discrepancies in the forced response of the zonal SST gradient across the equatorial Pacific to greenhouse gas forcing.
figure 2

a Changes in the SST difference (units: °K) between the western Pacific (5°S-5°N, 110°E-180°) and the eastern Pacific (5°S-5°N, 180°-80°W) relative to the climatology over 1950–1969 in CanESM5 under the CMIP6 historical protocol (1950–2014) and the SSP3-7.0 (2015–2100) scenario with the line and shading denoting the ensemble mean and two standard deviations across 50 ensemble members. b Ensemble-mean SST anomalies averaged over 2081–2100 relative to the climatology over 1950–1969, in which the tropical mean (30°S-30°N) is removed to highlight spatial inhomogeneity. c SST anomalies averaged over the years 131–150 of the CMIP6 abrupt-4×CO2 experiment relative to the corresponding pre-industrial control simulation for a single member of CanESM5, in which the tropical mean is removed. d–f Same as in a–c but for CESM2 and related large ensemble simulations (50 ensemble members with smoothed biomass burning emissions). g–i Same as in a–c but for GFDL-ESM2M under the CMIP5 historical protocol (1950–2004) and the RCP8.5 (2005–2100) scenario for 30 ensemble members and the abrupt-4×CO2 (131–150) scenario for a single member.

Full size image

Given well-known model biases (e.g., excessively cold cold-tongue and double inter-tropical convergence zone biases), the zonal SST gradient may strengthen in response to greenhouse gas forcing in reality through mechanisms such as the ocean dynamical thermostat mechanism, inter-basin teleconnection associated with greater warming over the Atlantic and/or Indian Ocean, and ENSO non-linear rectification15. Indeed, unlike other climate models participating in CMIP5, GFDL-ESM2M was found to simulate a La Niña-like mean state change under global warming as a result of ENSO non-linear rectification27. Considering that simulation results from a single ensemble member were used in ref. 27, we analyze further the GFDL-ESM2M Large Ensemble (GFDL-ESM2M LE) under CMIP5 historical forcing (1950–2004) followed by the Representative Concentration Pathway (RCP) 8.5 (2005–2100) scenario to determine the forced response of the zonal SST gradient more robustly. Fig. 2g,h indicate a slight increase in the zonal SST gradient at the end of the 21st century relative to the 1950–1969 climatology along with markedly different temporal evolution in comparison to CanESM5 and the CESM2 LE. However, the temporal evolution of the ensemble-mean change does not appear to align with the pathway of greenhouse gas emissions in the RCP8.5 scenario, with the magnitude of the ensemble-mean change being markedly smaller than the inter-ensemble spread even at the end of the 21st century (Fig. 2g). As this indistinct enhancement in the GFDL-ESM2M LE might result from opposing impacts of non-greenhouse gas forcing, the response of the zonal SST gradient to greenhouse gas forcing is additionally explored by analyzing the spatial pattern of SST anomalies relative to pre-industrial conditions in an idealized experiment for which atmospheric CO2 concentrations were abruptly quadrupled from pre-industrial conditions and then were held fixed for at least 150 years (i.e., abrupt-4×CO2 scenario). Unlike in Fig. 2h, the SST anomalies averaged over years 131–151 exhibit a slightly greater warming over the central-to-eastern equatorial Pacific than over the western Pacific (Fig. 2i), which is qualitatively consistent with CanESM5 (Fig. 2c) and CESM2 (Fig. 2f). However, the difference between Fig. 2h and Fig. 2i appears to be indistinct. Given the possibility that the changes shown in Fig. 2i might be still far from the full response to the imposed abrupt quadrupling of CO2, the temporal evolution of simulated zonal SST gradient change is examined over the full simulation period (300 years for GFDL-ESM2M and 999 years for CESM2). As shown in Supplementary Fig. 2a, the change in the zonal SST gradient is distinctly weaker in GFDL-ESM2M in comparison to CESM2. Nevertheless, GFDL-ESM2M also does not indicate a strengthening of the zonal SST gradient in a near-equilibrium climate, which is at odds with Fig. 2g. Furthermore, the SST anomalies averaged over the period 281–300, shown in Supplementary Fig. 2c, more clearly exhibit an enhanced warming in the eastern equatorial Pacific relative to the western equatorial Pacific, which is qualitatively similar to that for CESM2 over 980–999 (Supplementary Fig. 2b) in comparison to Fig. 2i. This suggests that even a centennial timescale might not be long enough to confidently determine the forced response of the zonal SST gradient to greenhouse gas forcing.

Based on the role of the Pacific Walker circulation in shaping the SST distribution in the tropical Pacific, we also examine for the temporal evolution and spatial pattern of model-simulated SLP changes whether the time duration of the instrumental record is long enough to confidently identify the forced response of the Pacific Walker circulation to greenhouse gas forcing (Supplementary Fig. 3). In CanESM5 and the CESM2 LE, the zonal SLP difference between the eastern equatorial Pacific (5°S-5°N, 160°W-80°W) and the western equatorial Pacific (5°S-5°N, 80°E-160°E)39 is projected to distinctly weaken at the end of the 21st century, despite the discrepancy in the trajectory of ensemble-mean change between the two models along with large inter-ensemble spread. In contrast, the GFDL-ESM2M LE simulates a slight strengthening at the end of the 21st century; however, the spatial pattern at the end of the 21st century is inconsistent with that for the abrupt-4×CO2 experiment, particularly in the western Pacific warm pool, which exhibits an indistinct change or slight weakening. These results therefore indicate that due to the influence of internal variability and non-greenhouse gas forcing, the time duration of the instrumental record covering the satellite or post-war periods might not be sufficiently long to robustly determine the response to greenhouse gas forcing.

Tropical Pacific zonal SST gradient in distant past climates

Having shown the potential limitation of high-quality SST records over the satellite or post-war era, we now turn our attention to paleoclimate proxy records during the mid-Pliocene warm period as this distant past period with high atmospheric CO2 concentrations (~400 ppm) may shed light on future change resulting from large greenhouse gas forcing44,45. The PRISM3 SST reconstruction46 (Methods) indicates that annual-mean SSTs over the western Pacific warm pool during the mid-Pliocene epoch (~3.3–3.0 million years ago) were similar in magnitude to modern values, with the SST difference (Pliocene minus modern) of 0.2 °K at ODP (Ocean Drilling Program) 769, 0.0 °K at DSDP (Deep Sea Drilling Project) 586 and 0.0 °K at ODP 806 (circles in Fig. 3a). In comparison, the sites located in the eastern tropical Pacific exhibit strongly positive SST anomalies of 2.5 °K at ODP 852, 3.0 °K at ODP 847, 1.9 °K at ODP 677 and 5.4 °K at ODP 1237 except for DSDP 573 (−0.6 °K) (circles in Fig. 3a). The near-zero SST differences in the western Pacific warm pool, but large positive anomalies in the eastern tropical Pacific, are seemingly at odds with what is seen for the present period (Fig. 1a,b).

Fig. 3: Zonal contrast in the annual-mean SST anomalies in the tropical Pacific.
figure 3

a Mean SST differences (units: °K) for CESM2 between the CMIP6 midPliocene-eoi400 and pre-industrial control simulations. Superimposed circles, stars and triangles represent, respectively, the SST anomalies (mid-Pliocene minus modern) from the PRISM3 dataset, the SST anomalies (~3.2 million years ago minus ~0.1 million years ago) from TEX86 (TetraEther indeX of 86 carbons)-based reconstructions47, and the SST anomalies (~3.2 million years ago minus pre-industrial) inferred from the alkenone paleotemperature proxy44. b Reduced space reconstruction of SST differences, derived from alkenone proxy data, between the mid-Pliocene and pre-industrial conditions (ref. 44). c Mean SST differences for CESM2-FV2 between the CMIP6 lgm and pre-industrial control simulations. Superimposed triangles represent the SST anomalies inferred from geochemical proxy data (ref. 49). d Similar to c but for the average of three paleoclimate data assimilation products, i.e., ref. 49, ref. 51 and ref. 52. e Mean SST differences for CESM2 between the CMIP6 abrupt-4×CO2 (averaged over the years 131–150) and pre-industrial control simulations. f Mean SST differences for CESM2 between the CMIP6 abrupt-0.5×CO2 (averaged over the years 131–150) and pre-industrial control simulations.

Full size image

According to the TEX86-based analyses of ref. 47 (Methods), the western Pacific warm pool was warmer during the late Miocene-to-Pliocene (~12.0–2.6 million years ago) than in the present day, whereas the tropical Pacific zonal SST gradient was not enhanced during these past high CO2, warm climate states47,48. More specifically, the TEX86-based SST reconstruction (stars in Fig. 3a) indicates that ODP site 806, located in the western Pacific warm pool, had a warmer SST (~1.4 °K) 3.17 million years ago relative to 0.1 million years ago. A larger SST difference of ~2.9 °K between 3.16 million years ago and 0.14 million years ago was, however, estimated at ODP site 850, located in the eastern equatorial Pacific, suggesting a reduced zonal SST gradient during the mid-Pliocene period.

Given the potential cold bias of TEX86 paleotemperature data, especially over the eastern equatorial Pacific45, as well as the uncertainties inherent in Mg/Ca paleotemperature data due to changes in seawater Mg/Ca ratio over time45,47, ref. 44 reconstructed SST records based solely on alkenone (\({U}_{37}^{{K}^{\prime}}\)) paleotemperature data using a Bayesian calibration method to account for the upper limit issue of alkenone-based estimation (Methods). Their reconstruction also indicates a reduced zonal SST gradient during the mid-Pliocene period relative to pre-industrial conditions (triangles in Fig. 3a): although ODP site 806 exhibited a warmer SST (~1.1 °K) during the mid-Pliocene warm period around 3.2 million years ago relative to the pre-industrial state, the sites located in the eastern Pacific were estimated to have warmer SST anomalies than the western Pacific site (i.e., ODP 806) except for IODP (Integrated Ocean Drilling Project) site U1338 (ODP 846: ~3.6 °K, ODP 847: ~2.2 °K, ODP 850: ~1.7 °K, ODP 1237: ~1.9 °K, ODP 1239: ~2.4 °K, OPD 1241: ~1.4 °K, IODP U1338: ~0.7 °K). Therefore, despite substantial discrepancies in the magnitude between the studies, paleoclimate proxy records during the mid-Pliocene warm period collectively suggest that the zonal SST gradient is predisposed to weaken during warm climate states with high CO2 levels relative to colder climates with lower CO2.

Next, on the premise that there is a causal relationship between CO2 forcing and changes in the tropical Pacific zonal SST gradient, we qualitatively evaluate the fidelity of model simulations using mid-Pliocene paleoclimate proxy records as a benchmark. Consistent with the PRISM3, TEX86-based and alkenone-based SST reconstructions, the CESM2 model simulates positive SST anomalies over the tropical Pacific under the CMIP6 Paleoclimate Modelling Intercomparison Project phase 4 (PMIP4) midPliocene-eoi400 scenario (Methods) relative to pre-industrial conditions, with more pronounced warming in the central-to-eastern tropical Pacific than in the western Pacific warm pool (shading in Fig. 3a). Although the spatial pattern and magnitude of simulated tropical SSTs exhibit substantial inter-model spread, other models contributing to CMIP6-PMIP4 simulate a reduced zonal SST gradient (or at least enhanced warming in the eastern equatorial Pacific which is at odds with the La Niña-like SST change pattern over the satellite era) during the mid-Pliocene warm period relative to the pre-industrial state (Supplementary Fig. 4).

Despite an overall reduction in the zonal SST gradient in both paleoclimate proxy records and model simulations, paleoclimate proxy records exhibit indistinct warm anomalies over part of the eastern equatorial Pacific in comparison to model simulations (Fig. 3a). Considering that the accuracy of paleotemperature reconstructions can be affected by a variety of factors, we also analyzed a reconstructed Pliocene Pacific SST field derived from alkenone proxy data44. As reported in ref. 44, Fig. 3b more clearly indicates a reduced zonal SST gradient during the mid-Pliocene warm period.

The paleoclimate proxy records during the mid-Pliocene warm period conversely imply an enhanced zonal SST gradient during colder climates with lower CO2 levels. Considering that the Last Glacial Maximum (LGM) was characterized by a much colder climate with lower CO2 levels (~200 ppm) relative to pre-industrial conditions, the PMIP community has selected the LGM as a particularly relevant past period for understanding near-future climate change41. Hence, we also analyzed CMIP6-PMIP4 lgm simulations along with proxy records to assess the robustness of the conclusion drawn from Fig. 3a,b and Supplementary Fig. 4.

As CMIP6-PMIP4 lgm scenario simulations were not performed with CESM2, we used output from CESM2-FV2 and CESM2-WACCM-FV2, with both of these representing similar climate states to CESM2. For both CESM2-FV2 (Fig. 3c) and CESM2-WACCM-FV2 (Supplementary Fig. 5a), the SST differences between the lgm and pre-industrial experiments exhibit negative values in the tropical Pacific with more enhanced cooling in the central-to-eastern Pacific than in the western Pacific warm pool, which qualitatively aligns with inferences from the paleoclimate proxy records during the mid-Pliocene warm period. The contrast in the zonal SST gradient change between the mid-Pliocene and LGM periods is qualitatively consistent with the difference between two idealized CO2-only experiments of CMIP6, one with atmospheric CO2 concentrations instantaneously quadrupled from pre-industrial levels (i.e., abrupt-4×CO2; Fig. 3e) and the other where concentrations were instantaneously halved (i.e., abrupt-0.5×CO2; Fig. 3f). These qualitative similarities appear to indicate that the trajectory of large atmospheric CO2 changes has played a major role in governing the change in the zonal SST gradient across the equatorial Pacific in the distant past, although climate change during the LGM is of course also linked to other factors including orbital parameters, ice sheet, and land-sea configuration41. Note that the enhanced cooling in the central-to-eastern equatorial Pacific is also represented by INM-CM4-8 and MIROC-ES2L (Supplementary Fig. 5b,c). In comparison, the spatial distribution of SST anomalies over the tropical Pacific for MPI-ESM1-2-LR differs from other models; however, an enhanced cooling in the eastern equatorial Pacific is also evident (Supplementary Fig. 5d). Moreover, as in other climate models providing simulation output for both abrupt-4×CO2 and lgm or midPliocen-eoi400, MPI-ESM1-2-LR also exhibits a reduced zonal SST gradient in response to a quadrupling of atmospheric CO2 concentrations (Supplementary Fig. 6).

In comparison, SST anomalies in the LGM period, inferred from geochemical proxy records49, exhibit an indistinct change in the zonal SST gradient, although the zonal SST gradient does not appear to have increased (Fig. 3c, triangles). While there has been a lack of consensus on the spatial distribution of SST anomalies over the tropical Pacific during the LGM due to conflicting proxy reconstructions49,50, such large uncertainties have been substantially reduced by the paleoclimate data assimilation technique. Indeed, the SST anomalies during the LGM averaged over three paleoclimate data assimilation products49,51,52 indicate enhanced cooling in the eastern equatorial Pacific relative to the western equatorial Pacific (Fig. 3d), which is supported by an independent study50 indicating a deepened mixed layer induced by a stronger Pacific Walker circulation during the LGM in comparison to pre-industrial or current conditions.

Despite a qualitatively similar increase in the zonal SST gradient between Fig. 3c and Fig. 3d, we note that there is a noticeable discrepancy in the spatial pattern: while the paleoclimate data assimilation products suggest the strongest LGM cooling in the eastern equatorial Pacific, CESM2-FV exhibits strong cooling in both the central and eastern equatorial Pacific. Likewise, although the zonal SST gradient is enhanced in both Fig. 3c,d (i.e., LGM) and Fig. 3f (abrupt-0.5×CO2 experiment), the cooling is weaker at the dateline in comparison to 150°W along the equator in the former but stronger in the latter. These discrepancies might have resulted from model deficiency or suggest that the zonal SST gradient change during the LGM period might not be consistent with that for the halved CO2-only experiment.

Changes in the zonal SST gradient and global-mean surface temperature over the late Cenozoic

We also consider whether the reduced zonal SST gradient during the mid-Pliocene warm period may have existed independently of warmer global temperatures by analyzing changes in the zonal SST gradient and global-mean surface temperature over the past four million years. A variety of paleoclimate proxy records suggest an overall cooling of global-mean surface temperature over the past four million years (Supplementary Fig. 7a). Despite discrepancies in the magnitude of estimated cooling between studies43, this cooling over time has been linked, in large part, to decreased atmospheric CO2 concentrations (Supplementary Fig. 7b)43.

According to the alkenone-based SST reconstructions of ref. 44, the SST difference between the western Pacific warm pool and the eastern equatorial Pacific has undergone an overall increase during the late Cenozoic period, albeit with substantial temporal variations (Supplementary Fig. 7c, purple line). Also presented in Supplementary Fig. 7c is a time series of the zonal SST difference estimated by ref. 43, in which Mg/Ca ratio and faunal proxies were used together with alkenone-based proxies to reconstruct global and regional SSTs. Due to differences in proxies, core sites and calibration methods between ref. 43 and ref. 44, the time series exhibit substantial discrepancies. However, the zonal SST difference based on the reconstruction of ref. 43 also indicates an overall increase over the same period, which is further supported by ref. 45 and ref. 47. Hence, the zonal SST difference appears to have strengthened in response to global cooling. To show this aspect more quantitatively, a scatter plot is created using the reconstruction of ref. 43 over the past four million years. Figure 4a indicates that although there is a large spread, the zonal SST difference tends to have weakened during warm climate states in comparison to much colder climates in the distant past. More specifically, the reconstruction of ref. 43 suggests that the SST difference between the western Pacific warm pool and the equatorial eastern Pacific has decreased by ~0.3 ºK with respect to a 1-K increase in global-mean surface temperature (Fig. 4a). The zonal SST differences estimated from the reconstruction of ref. 44 points to a similar conclusion with a nearly identical regression slope, albeit with some non-linearity (Fig. 4b).

Fig. 4: Changes in the zonal SST gradient across the equatorial Pacific in response to global warming.
figure 4

a A scatter plot highlighting changes in the zonal SST gradient relative to pre-industrial versus corresponding global-mean surface temperature change over the past four million years based on the reconstruction of ref. 43. b Same as in a but with change in the zonal SST gradient based on the reconstruction of ref. 44. c Model-simulated response of the SST difference between the western Pacific (5°S-5°N, 110°E-180°) and the eastern Pacific (5°S-5°N, 180°-80°W) to global warming. Open squares in red (blue) denote the SST anomalies averaged over the years 131–150 of abrupt-4×CO2 (abrupt-0.5×CO2) relative to the corresponding pre-industrial control simulation for a given model. Filled triangles in red (blue) represent the SST differences between the midPliocene-eoi400 (lgm) and pre-industrial control simulations. Open circles in red (blue) denote the SST anomalies averaged over the years 131–150 of abrupt-solp4p (abrupt-solm4p) relative to the corresponding pre-industrial control simulation for a given model. Each number in the figure denotes a distinct climate model. d Same as in c but for the SLP differences between the eastern Pacific (5°S-5°N, 160°W-80°W) and the western Pacific (5°S-5°N, 80°E-160°E).

Full size image

Next, we show that climate models are able to broadly reproduce the relationship inferred from paleoclimate proxy records. In Fig. 4c, model-simulated zonal SST gradient is related to corresponding global-mean surface air temperature for midPliocene-eoi400 (filled triangles in red), lgm (filled triangles in blue), abrupt-4×CO2 (open squares in red), abrupt-0.5×CO2 (open squares in blue), abrupt-solp4p (similar to abrupt-4×CO2, except that the solar constant was instantaneously increased by 4%, open circles in red) and abrupt-solm4p (similar to abrupt-4×CO2, except that the solar constant was instantaneously reduced by 4%, open circles in blue). Although the contrast between midPliocene-eoi400 and lgm is less distinct in comparison to both abrupt-4×CO2/abrupt-0.5×CO2 and abrupt-solp4p/abrupt-solm4p, model simulations are roughly consistent with paleoclimate proxy records in that the tropical Pacific zonal SST gradient tends to weaken under sustained global warming and vice versa, although the regression slope is just around −0.1 K K-1. The smaller magnitude compared to the paleoclimate proxy records implies that climate models might underestimate the response of the zonal SST gradient to global warming. Fig. 4d indicates that the weakening of the zonal SST gradient under sustained global warming is in general accompanied by a slowdown of the Pacific Walker circulation. Although the contrast in the SLP gradient change between midPliocene-eoi400 and lgm is again less distinct, climate models, for which simulation output for abrupt-4×CO2 is available, are found to exhibit a distinctly weakened zonal SST gradient in response to CO2 quadrupling (Supplementary Fig. 6).

Finally, based on the qualitative agreement of CESM2 with mid-Pliocene proxy records (Fig. 3a,b), simulated changes in SLP and 850-hPa winds under midPliocene-eoi400 relative to pre-industrial experiment are compared with ensemble-mean changes for year 2100. Although the simulated changes are not identical, both distant past and future warm periods are characterized by an overall weakening of the Pacific Walker circulation (Supplementary Fig. 8a,b), in line with previous studies39,40,44,53. It is also noted that both distant past and future simulations exhibit an increased freshwater forcing over the central-to-eastern equatorial Pacific in warm climates (Supplementary Fig. 8c,d), which can be expected to impede the upwelling of cold subsurface waters.

Summary and Discussion

Considering the far-reaching influences of the tropical Pacific SSTs, the marked model-observation discrepancy in the changes of the zonal SST gradient across the tropical Pacific over the satellite era could potentially undermine our confidence in model-projected human-induced future climate change. However, we found large inter-model discrepancies in the temporal evolution of the forced response of the zonal SST gradient in large-ensemble model simulations along with substantial inter-ensemble spread. In addition, it is found that GFDL-ESM2M also eventually exhibits a slight weakening of the zonal SST gradient in response to CO2 quadrupling, thereby differing from the slight strengthening at the end of the 21st century for the same model, implying that timescales longer than a century might be needed to confidently identify the forced response to greenhouse gas forcing. The finding of previous studies that the observed La Niña-like SST trend pattern is linked in part to non-greenhouse gas forcing14,19,30,31,32 further complicates this challenge. Due to this potential limitation of the effective duration over which high-quality historical SST observations exist, we analyzed the response of the tropical Pacific SST to large past climate changes using paleoclimate proxy records over the past four million years. Paleoclimate proxy records indicate that the tropical Pacific zonal SST gradient weakened during high CO2, warm periods, such as the mid-Pliocene, relative to colder periods with lower CO2 levels. We found that idealized climate model simulations are able to broadly capture this contrasting feature. Synthesizing all of these results therefore leads us to conclude that the tropical Pacific zonal SST gradient is likely to decrease under sustained large global warming.

Previous studies suggested that both the ocean dynamical thermostat linked to ocean thermal stratification26 and the tropical atmospheric circulation, especially the Pacific Walker circulation22, govern the response of tropical Pacific SST to external forcing16,35. In other words, the zonal SST gradient can be enhanced in response to greenhouse gas forcing if the ocean dynamical thermostat dominates over a weakened Pacific Walker circulation, and vice versa. The results inferred from paleoclimate proxy data seem to suggest that the ocean dynamical thermostat could play a dominant role at first but its influence might diminish under sustained global warming with continued emissions of greenhouse gases16,19,35. In addition to increased dry static stability 21 and/or differential atmospheric moisture and precipitation responses22, enhanced warming over the extra-tropics relative to the tropics might lead to a slowdown of the Pacific Walker circulation under sustained global warming, i.e., weakened mid-latitude westerlies due to reduced meridional temperature gradient may induce changes in subtropical anticyclonic circulations and related Pacific trade winds.

Although paleoclimate proxy records during the mid-Pliocene warm period consistently imply a weakening of the zonal SST gradient with global warming, substantial uncertainties remain in terms of the spatial pattern and magnitude due to uneven proxy data sampling in both space and time, proxy uncertainties and methodological limitations42,43,45,47. Thus, as a complement to developments with research efforts in modeling, concurrent continued expansion of the spatio-temporal resolution of paleo-proxy records needed for determining more spatially and temporally coherent climate response to external forcing, as well as the accuracy of these records, will constitute an essential contribution to advancing our understanding of the response of tropical Pacific SST to sustained anthropogenic forcing.

Methods

Observed SST data and paleoclimate SST proxy records

Tropical Pacific SST trends are analyzed using NOAA’s Extended Reconstructed Sea Surface Temperature version 5 (ERSST5)54, which covers the period from 1854 to present, along with the Hadley Centre Sea Ice and Sea Surface Temperature (HadISST) SST dataset55, which covers the period from 1870 to present. For estimating SLP trends, the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5) dataset56 and the Japanese 55-year Reanalysis (JRA-55)57 are used in this study. ERA5 and JRA-55 use HadISST2 and COBE (Centennial in situ Observation-Based Estimates) SST, respectively, for boundary conditions. Hence, there might be a mismatch between SST trends shown in Fig. 1 and SLP trends shown in Supplementary Fig. 1. However, both ERA5 and COBE-SST258 indicate that the La Niña-like SST trend pattern over 1979–2023 is not evident over 1950–2023 (Supplementary Fig. 9).

To explore the SST difference between the western Pacific warm pool and the eastern tropical Pacific during the mid-Pliocene epoch (~3.3–3.0 million years ago), we used the Pliocene Research Interpretation and Synoptic Mapping (PRISM, version 3) annual mean SST data products available through the Deep Sea Drilling Project (DSDP) and Ocean Drilling Program (ODP) sites. These products were estimated from foraminifera, radiolarian, diatom, ostracod, the Mg/Ca ratio of planktic foraminifera, and the unsaturation index of alkenones46,59,60. The PRISM3 sites located in the western Pacific warm pool are ODP 769 (8.78°N, 121.29°E), DSDP 586 (0.50°S, 158.50°E), and ODP 806 (0.31°N, 159.36°E) and the sites located in the eastern tropical Pacific include DSDP 573 (0.49°N, 133.30°W), ODP 852 (5.28°N, 110.07°W), ODP 847 (0.18°N, 95.32°W), ODP 677 (1.20°N, 83.74°W), and ODP 1237 (16.00°S, 76.37°W). With PRISM346,59,60, the modern SST fields used to remove differences inherent in proxy calibrations are from ref. 61 for SST proxies based on the Mg/Ca ratio of planktic foraminifera and the unsaturation index of alkenones, and from ref. 62 for faunal-based SST proxies.

The accuracy of paleotemperature reconstructions can be affected by diagenetic overprinting, changes in seawater chemistry, and calibration limitations42,43,47,49. Given the possibility that proxy-based SSTs in the western Pacific warm pool might be underestimated for warm climate states in the distant past due to both changes in seawater Mg/Ca ratio and the upper limit of alkenone-based temperatures (~28.5 °C or ~301.65 °K), ref. 47 reconstructed SST for the past 12 million years using both TEX86 (TetraEther indeX of 86 carbons) with the temperature calibration of ref. 63 and alkenone temperature proxies at ODP sites 769, 806, and 1143 (9.36°N, 113.29°E) in the western Pacific warm pool and ODP site 850 (1.30°N, 110.52°W) in the eastern tropical Pacific. This proxy dataset constructed by ref. 47 is also used in our analysis.

Due to potential cold bias of TEX86-based paleotemperature data, especially over the eastern equatorial Pacific45, and uncertainties in Mg/Ca paleotemperature data45,47, ref. 44 reconstructed SST records using only alkenone proxies with a Bayesian calibration method to account for the upper limit issue of alkenone-based SST estimation. The sites located in the western Pacific warm pool are ODP 806 and ODP 1143, and the sites located in the eastern tropical Pacific include ODP 846 (3.09°S, 90.82°W), ODP 847, ODP 850, ODP 1237, ODP 1239 (0.67°S, 82.08°W), ODP 1241 (5.84°N, 86.44°W), and Integrated Ocean Drilling Project (IODP) site U1338 (2.51°N, 117.97°W). We used median values at each site for estimating SST anomalies during the mid-Pliocene relative to pre-industrial shown in Fig. 3a and ensemble-mean values averaged over the western Pacific warm pool and the eastern equatorial Pacific for the time series of zonal SST difference shown in Supplementary Fig. 7c. In addition, a reconstructed Pliocene Pacific SST field derived from alkenone proxy data44 is used in this study.

For the case of LGM SSTs, we used geochemical proxy records from ref. 49. Given conflicting proxy reconstructions for the LGM period49,64,65, we also used paleoclimate data assimilation products49,51,52, in which large uncertainties stemming from incomplete and uneven proxy data sampling in both space and time were substantially reduced. In particular, ref. 49 exclusively used geochemical SST proxies (i.e., the unsaturation index of alkenones, TEX86, δ18O and the Mg/Ca ratio) to avoid no-analog problems and lack of comparable Bayesian models.

Ref. 43 produced a global SST stack covering the last 4.5 million years using 127 published SST records based on alkenone, Mg/Ca ratio, and faunal proxies with taking into account long-term changes in seawater Mg/Ca ratio and used the global SST stack to reconstruct global-mean surface temperature change over the same period by applying a scaling formula derived from climate model simulations. These paleoclimate temperature data are used to estimate the temporal evolution of global-mean surface temperature as well as the zonal SST difference between sites located in the western Pacific warm pool and sites located in the eastern tropical Pacific (Fig. 4 and Supplementary Fig. 7). It is noted that the temporal evolution of global-mean surface temperature estimated from ref. 43 is qualitatively consistent with that from ref. 66 and ref. 67 (Supplementary Fig. 7a).

Model simulation output

The temporal evolution of model-simulated tropical Pacific zonal SST and SLP gradients is evaluated over the period 1950–2100 using the Community Earth System Model version 2 Large Ensemble (CESM2 LE) simulations5 forced with CMIP668 historical forcing over the period 1850–2014 and Shared Socioeconomic Pathway (SSP) forcing scenario SSP3-7.0 over the period 2015–2100. The CESM2 LE consists of two 50-member subsets, but we only used the second subset forced by a smoothed version of the biomass burning dataset over the period 1990–2020. Detailed information on the CESM2 LE initialization procedure can be found in ref. 5. Given large inter-model spread, CanESM5 is also used to examine whether the time duration of the instrumental record is sufficiently long for confidently determining the forced response. We used simulation output for 50 ensemble members forced by the CMIP6 historical protocol and the SSP3-7.0 scenario.

The GFDL-ESM2M model is exceptional in that a La Niña-like mean state change is simulated in response to greenhouse gas forcing27 standing in contrast to other models that produce an El Niño-like mean state change. However, given that simulation output from a single ensemble member was analyzed in ref. 27, it is uncertain whether the reported La Niña-like mean state change is a forced response due to the potential role of internal variability30,33,34,69,70. For this reason, along with the CESM2 LE and CanESM5, we analyzed the GFDL-ESM2M Large Ensemble71, which consists of 30 ensemble members run under CMIP5 historical forcing (1850–2004) and subsequently the Representative Concentration Pathway (RCP) 8.5 scenario.

Working under the assumption that a qualitative agreement with paleoclimate proxy record lends some credence to the robustness of model-projected future climate change under sustained anthropogenic forcing, model simulation output for the CMIP6-PMIP4 midPliocene-eoi400 was evaluated against the paleoclimate proxy data described in the preceding subsection. The midPliocene-eoi400 experiment was designed to examine the response of the climate system to stable long-term CO2 forcing (~400 ppm), similar to the present day, during the mid-Pliocene warm period (~3.2 million years ago) when mean surface temperature was warmer than that for the pre-industrial period43,44,46. In the midPliocene-eoi400 experiment, astronomical parameters, solar activity, and tropospheric aerosols were prescribed as in the pre-industrial control experiment, while the experimental design included modifications to ice sheets, topography, ocean bathymetry, and vegetation. On the premise that there is a causal relationship between CO2 forcing (or global warming) and changes in the zonal SST gradient, we also analyzed the CMIP6-PMIP4 lgm experiment which focuses on the LGM period (~21 thousand years ago) when CO2 concentrations were ~200 ppm with boundary conditions being substantially different from the present day. We note that the magnitude of the changes in forcing and global mean surface temperature between the LGM and pre-industrial periods is comparable to that projected from the present to the end of the 21st century72. More detailed information on the CMIP6-PMIP4 midPliocene-eoi400 and lgm experiments can be found in ref. 73 and in ref. 72, respectively.

Although the CESM2 LE simulations do not encompass these periods in the relatively distant past, CESM2 was a contributing model to the midPliocene-eoi400 group of simulations under CMIP6-PMIP4. Additionally, CESM2-FV2 and CESM2-WACCM-FV2, which are closely related to CESM274, were also contributions to the CMIP6-PMIP4 lgm exercise. Hence, these model simulations along with corresponding pre-industrial control simulations are used in our analysis. Considering that there are potentially large model structural uncertainties across CMIP6 models, we also included additional CMIP6 models participating in the CMIP6-PMIP4: EC-Earth3-LR, GISS-E2-1-G, HadGEM3-GC31-LL, IPSL-CM6A-LR, and NorESM1-F for the midPliocene-eoi400 experiment and INM-CM4-8, MIROC-ES2L, and MPI-ESM1-2-LR for the lgm experiment.

To examine whether climate models are able to reproduce the relationship inferred from paleoclimate proxy records between global-mean surface temperature and the zonal SST gradient, we analyzed a range of idealized experiments along with the midPlocene-eoi400 and lgm, i.e., abrupt-4×CO2 (coupled model simulation in which atmospheric CO2 concentrations were abruptly quadrupled from pre-industrial levels and then were held fixed for at least 150 years), abrupt-0.5×CO2 (similar to abrupt-4×CO2, except that CO2 concentrations were abruptly halved), abrupt-solp4p (similar to abrupt-4×CO2, except that the solar constant was instantaneously increased by 4%), and abrupt-solm4p (similar to abrupt-solp4p, except that the solar constant was reduced by 4%). For these idealized experiments forced by abrupt change, we analyzed SST and SLP anomalies averaged over years 131–150 relative to pre-industrial values. Supplementary Table 1 provides information on the model simulations analyzed in this study.

Statistical information

We used the standard least squares linear regression approach for computing trends. Statistical significance of the computed trends is determined using a two-sided Student’s t-test at the 95% confidence level with reduced degrees of freedom to account for autocorrelation in a given time series.