Increased ENSO sea surface temperature variability under four IPCC emission scenarios

Sea surface temperature (SST) variability of El Niño–Southern Oscillation (ENSO) underpins its global impact, and its future change is a long-standing science issue. In its sixth assessment, the IPCC reports no systematic change in ENSO SST variability under any emission scenarios considered. However, comparison between the 20th and 21st century shows a robust increase in century-long ENSO SST variability under four IPCC plausible emission scenarios. Sea surface temperature variability of the equatorial Pacific Ocean dictates the strength of El Niño–Southern Oscillation events. CMIP6 models under four IPCC emission scenarios show increased variability in the 21st century from the 20th century.

by CMIP6 models under various emission scenarios. However, sampling ENSO SST variability over relatively short 30-year running periods would be subject to a strong influence from decadal climate fluctuations such that stochastic noise and butterfly effect can obscure the impact of greenhouse forcing [14][15][16] . Even on centennial timescales, ENSO variability is influenced by internal climate fluctuations 17 . Here we show that, contrary to the AR6 assessment, ENSO SST variability is stronger in the 21st century than in the 20th century in four plausible emission scenarios considered in the IPCC AR6.
We construct monthly evolution of SST anomalies covering the period 1900-2099, under four shared socioeconomic pathways (SSP) 8 Table 1) (see 'CMIP6 models' in Methods). One experiment from each model is used as in the IPCC AR6. The SST anomalies are then quadratically detrended and averaged over traditional regions of ENSO SST indices. These include Niño3 (5° N-5° S, 150° W-90° W), Niño4 (5° N-5° S, 160° E-150° W) and Niño3.4, used to depict Eastern Pacific ENSO variability, Central Pacific ENSO variability, and their combined variability, respectively.
Under SSP585, a 'high emissions' scenario in the 21st century (2000-2099) sees greater Niño3.4 variability than in the 20th century  in a total of 38 out of 43 (88.4%) models (Fig.  1a). The multi-model ensemble mean increase (16.1%) is statistically significant according to a Bootstrap method 18 (see 'Bootstrap test' in Methods and Extended Data Fig. 1). A similar increase (15.7%) is obtained when inter-model differences in representing ENSO amplitude are removed (Extended Data Fig. 2). In terms of Niño3 and Niño4 SST variability, the level of inter-model consensus is also strong, both reaching 81.4%, consistent with recent studies 5-7 . Using more sophisticated indices leads to an even stronger consensus 7 .
Inter-model consensus is seen in all four SSP emission scenarios (Extended Data Figs. 3, 4 and 5), ranging between 78.6-88.4% for Niño3.4, 76.2-81.6% for Niño3, and 73.8-81.4% for Niño4 across emission scenarios. In the SSP126 scenario (Fig. 1b), which represents a strong mitigation path under the Paris Agreement to limit warming to 1.5-2.0 °C relative to the pre-industrial level, a total of 34 out of 39 models (87.2%) generate an increase in Niño3.4 SST variability. The multi-model ensemble mean increase (11.7%) is statistically significant. Thus, even if the Paris Agreement target is achieved, an increase in ENSO SST variability is projected, con-Increased ENSO sea surface temperature variability under four IPCC emission scenarios Wenju Cai 1,2 ✉ , Benjamin Ng 2 , Guojian Wang 1,2 , Agus Santoso 2,3 , Lixin Wu 1 ✉ and Kai Yang 4 sistent with previous results in terms of ENSO rainfall variability, which show a continuous increase after global mean temperature stabilizes 4 .
As mentioned above, influence from internal variability could explain the lack of inter-model consensus in the AR6 [14][15][16] . Using statistics over another longer period, for example, 50 years of the current climate of 1965-2015 and the future climate of 2050-2099, an inter-model consensus emerges, with a total of 34 out of 43 (79.1%) models under SSP585, and 27 out of 39 (69.2%) models under SSP126, generating an increase in Niño3.4 SST variability (Fig.  2a,b). Thus, as the period for calculating ENSO statistics lengthens, the influence of internal variability tapers off, whereas the effect of greenhouse warming is better diagnosed.
Studies based on CMIP5 models alone, or on a combination of CMIP5 and a smaller set of earlier available CMIP6 models 9,10 , have found that models better simulating ENSO dynamics tend to generate increased ENSO SST variability. Nonlinear Bjerknes feedback is particularly important, in which once warm SST anomalies establish deep atmospheric convection in the equatorial eastern Pacific, the response of equatorial zonal wind anomalies increases nonlinearly to further sea surface warming 19 , amplifying El Niño growth 10 . Models generally underestimate this feedback 10 . As a result, an inter-model consensus of increased ENSO SST variability emerges in a subset of models that simulate a reasonably realistic nonlinear positive feedback 10 .
Under greenhouse warming, an increased air-sea coupling arising from an enhanced upper equatorial ocean stratification underpins the increase in ENSO variability 10 . There is a tendency for models that simulate stronger nonlinear Bjerknes feedback to generate a greater future increase in ENSO variability, and vice versa, and this relationship is statistically significant above the 99% confidence level (Extended Data Fig. 6). As such, the skewness in three model outliers that project a decrease in ENSO variability in Figs. 1 and 2 is either small (UKESM1-0-LL) or negative (BCC-CSM2-MR, CESM2). Besides some improvements in the representation of the tropical Pacific climate 20 , more CMIP6 models simulate realistic ENSO nonlinear feedbacks 7 . The associated stronger air-sea coupling over the transient period of increasing greenhouse forcing  probably contributes to the strong inter-model consensus even when all available models are used. Putting our results in a broader context of recent studies, we note findings of decreased ENSO variability 21,22 in equilibrium state after an instantaneous doubling/quadrupling of CO 2 , which seemingly contradict our conclusion; thus, it is worth highlighting that our result of increased ENSO SST variability is based on the IPCC 21st century 'transient' warming scenarios. Further, ocean mesoscale eddy-induced vertical heat transport is part of ENSO dynamics 22 , but is not resolved in CMIP6 models; as such, whether or how incorporation of ocean mesoscale eddies alters our conclusion is not known. In addition, our conclusion is based solely on CMIP6 models, but a recent study based on 36 CMIP5 models and 20 CMIP6 models 23 finds a large spread in the projected ENSO change with no inter-model consensus; the discrepancy suggests a sensitivity of our result to model generations. We also note that our result is opposite to an observed decline in ENSO variability since 1999/2000 (ref. 24 ); however, any greenhouse warming-induced ENSO variability change in the two decades since 1999/2000 might be masked by decadal variability and not yet detectable. Finally, the impact of North Pacific SST variability on ENSO is more realistically simulated in CMIP6 than in CMIP5, and the impact is enhanced under warmer background SSTs, contributing to enhanced future ENSO SST variability 25 .
To conclude, in contrast to the IPCC AR6 assessment, there is in fact an inter-model consensus of increased ENSO SST variability in the 21st century from that of the 20th century under all four plausible SSP scenarios considered by the IPCC AR6 using CMIP6 models. The inter-model consensus is seen by comparing ENSO SST variability over the two 100-year periods of the 20th and 21st centuries. A long analysis period reduces the effect of internal variability, allowing a better detection of the greenhouse-forced effect.
The increased inter-model consensus in CMIP6 is facilitated by improvements in the representation of ENSO processes, for example, nonlinear positive feedbacks, thus underscoring the importance of sustained model improvement in a multi-model approach toward reliable future projections. The inter-model consensus, particularly in the scenario with the strong mitigation pathway, highlights the need to consider the possibility of increased ENSO SST variability in addition to increased ENSO-induced rainfall variability.

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/ s41558-022-01282-z.   We used one experiment from each model. Monthly climatology was constructed over the 1900-1999 period, and monthly SST anomalies referenced to the climatology were obtained. Time series of the 200-year SST anomalies averaged over regions of ENSO SST indices, Niño3 (5° N-5° S, 150° W-90° W), Niño4 (5° N-5° S, 160° E-150° W) and Niño3.4 (5° N-5° S, 170° W-120° W) was quadratically detrended over the full 200-year period. We compared the standard deviation of an ENSO SST index over two periods, either between the 2000-2099 and 1900-1999 (Fig. 1), or between the 50-year future climate of 2050-2099 and current climate of 1965-2014 (Fig. 2).
Bootstrap test. A bootstrap method 18 was used to examine whether the multi-model mean increase in variability of an ENSO SST index is statistically significant. For example, under SSP585, the 43 values of Niño3.4 SST variability in the 20th century were resampled randomly to form 10,000 realizations of 43-sampled sets. A histogram of the 10,000 realizations (blue bars), together with the 10,000 inter-realization standard deviation (grey shade), is shown in Extended Data Fig. 1. The same process was repeated for the 21st century. If the increased multi-model mean in Niño3.4 SST variability is greater than variability due to internal variability, that is, the sum of the 10,000 inter-realization standard deviation values for the 20th and 21st centuries, the increase is statistically significant above the 95% confidence level. Because ENSO SST variability varies vastly across models (Fig. 1), an alternative approach is to have an ENSO SST index in each model normalized by the standard deviation over the full 200-year period before conducting the multi-model average and the Bootstrap test. In this approach, the statistical significance holds (Extended Data Fig. 2).

Data availability
Data related to the paper can be downloaded from websites listed below: CMIP6 database, https://esgf-node.llnl.gov/projects/cmip6/. Detailed references and DOI URLs for each CMIP6 model are provided in Supplementary Information.

Code availability
Codes for calculating ENSO SST variability are available from the corresponding authors on request. Observed skewness based on 1900-2020 from a reanalysis 26 is shown as a black solid vertical line in each panel. Correlation coefficient and the P-value are shown in each panel, indicating statistically significant relationships above the 99% significance level. SST skewness in the eastern equatorial Pacific is an integral component of a parameter that also encapsulates skewness in the central Pacific to depict ENSO nonlinearity as a whole 7,10 . In terms of this nonlinearity parameter, a larger number of CMIP6 (79%) than CMIP5 (55%) models simulate greater than one third of the observed parameter.