Understanding how the El Niño–Southern Oscillation (ENSO) will change under increasing greenhouse gas emissions is critical due to ENSO's widespread impacts, which include changes to floods, droughts, and fishery production e.g.,. While previous work demonstrates model agreement on future intensification of ENSO's atmospheric impacts e.g., precipitation;, debate remains as to whether ENSO-driven sea surface temperature (SST) variability will increase or decrease in the future as disparities between projections from different climate models persist e.g.,. In this study we present ENSO SST projections for the first time from 14 individual single model initial-condition large ensembles (SMILEs) with the time-evolving forced response cleanly separated from internal variability.

Previous work finds a diverse range of projections of ENSO variability in multi-model ensembles (CMIP; coupled model intercomparison projects). In CMIP3 a range of ENSO variability projections were found, leading to a review stating that it was not yet possible to determine whether ENSO would change in the future . In CMIP5 the models again demonstrate a variety of responses with the multi-model mean change indistinguishable from zero . Other studies have further investigated ENSO projections by excluding models that have strong ENSO biases in the historical period. When selecting for models that correctly represent ENSO skewness, find an increase in eastern Pacific ENSO SST variability. This result is supported by the recent release of CMIP6 wherein demonstrate that 34 of 43 models show an increase in ENSO SST variability in SSP585 (27 of 39 in SSP126), in agreement with another early CMIP6 assessment .

There is also some evidence for an increase in ENSO variability in the observational record, with paleoclimate data suggesting that ENSO variability strengthened post-1950 . Coral records indicate that ENSO variability has been 25 % larger over the last 5 decades compared to the pre-industrial era . This agrees with tree ring records, which show a strengthening of ENSO variance in the late 20th century . A study using linear inverse modeling based on instrumental data agrees that ENSO SST variance has increased since 1976 . The authors, however, note that such differences could be due to multi-decadal variability or the impact of climate change, and the cause of this observed change remains unclear. A recent study also noted the strong contribution of internal variability in the observed record . conclude that 65 % of the observed increase in extreme El Niño events is attributable to the internal variability (largely multi-decadal variability from the Atlantic Multidecadal Oscillation), with the rest of the increase attributed to a changing climate.

While consensus between multi-model ensembles and paleoclimate records suggests that ENSO variability may increase with increasing greenhouse gas emissions, other studies that look at the equilibration to strong greenhouse gas forcing complicate this result. A recent study using a very-high-resolution climate model (0.1∘ ocean; 0.25∘ atmosphere) finds that ENSO SST variability decreases substantially under strong (4 × CO2) forcing . This study further argues that CMIP-class models are too low in horizontal resolution to correctly capture some important aspects of ENSO ocean dynamics, which could lead to an artificial increase in SST variance in these models. On longer timescales, ENSO amplitude is also found to decrease in CMIP-class standard-resolution (1 to 2∘) models . Using output from the Long Run Model Intercomparison Project, find diverging transient ENSO variability projections but a consistent decrease in SST variability across models when the response has stabilized. Indeed, suggest that ENSO variability in CMIP5 models is not linear in time, implying an important role for analyzing multiple time periods. Collectively, these studies demonstrate the importance of further research to reconcile these discrepancies.

Differences between the projections of ENSO variability in individual models have been linked to the pattern of mean-state warming across the tropical Pacific e.g.,. For instance, in CMIP5 and CMIP6 models there is a weak negative correlation between the magnitude of the tropical Pacific SST gradient and ENSO SST variability . In this case, an increase in ENSO SST variability is linked to a weakening of the tropical Pacific SST gradient. This relationship varies between models and is found to be more robust in models that have more realistic subsurface nonlinear dynamical heating and ENSO asymmetry when compared to observations . Periods of weaker tropical Pacific SST gradient are also linked to higher ENSO variability, with periods with a stronger gradient linked to periods of smaller ENSO SST variability . Models that show reduced ENSO variability in response to anthropogenic forcing also show a strengthening of the tropical Pacific SST gradient . This relationship between ENSO amplitude and the mean-state tropical Pacific SST gradient is identified in Paleoclimate Modelling Intercomparison Project (PMIP) models, although it is not constant through time . However, paleoclimatic proxies suggest that there is no relationship between mean zonal SST gradient and ENSO variability , demonstrating that this link is either not found in our observations or, as in the PMIP models, is not constant in time. We emphasize that the east–west Pacific SST gradient is only one of many mean-state metrics that are potentially important to influence (and vice versa be influenced by) ENSO variability. In fact, different aspects of the climate mean state affect ENSO feedbacks in multiple ways, leading to a potentially high sensitivity of the linear ENSO growth rate (and hence SST variability) to these factors .

How the tropical Pacific SST gradient itself will change in the future is also a point of contention. Most CMIP-class models agree on a projected weakening of the SST gradient El Niño-like warming;. This, however, must be reconciled with the recent observed increase in the SST gradient La Niña-like warming;. Some studies suggest that the inconsistencies between models and observations can be explained by internal variability and observational uncertainty , while others argue that the observations are truly outside the model range . One hypothesis is that under transient forcing the ocean thermostat mechanism, whereby upwelling waters delay warming in the eastern equatorial Pacific, overwhelms the long-term weakening of the Walker circulation and SST gradient , leading to temporary La Niña-like warming. Another study concludes that a more accurate representation of ENSO nonlinearity leads to La Niña-like warming in at least one realistic climate model . As such, while most models project an El Niño-like warming, this is not the only possible future response. Alternatively, realistic ENSO nonlinearity does not necessarily lead to a La Niña-like warming pattern. While realistic representations of ENSO nonlinearities are a necessary condition for the simulation of rectified mean-state changes via nonlinear dynamical heating, diverging projected changes in future ENSO SST variability among models with largely realistic ENSO nonlinear dynamic heating can lead to either ENSO-induced damped or amplified eastern tropical Pacific mean-state warming . Additionally, it has been hypothesized that SST mean-state biases in the Atlantic Ocean are responsible for low-biased interbasin connectivity (and hence also low-biased SST variability) between the Atlantic and Pacific . This suite of studies highlights the need for further research and continued observations to reduce uncertainties in how the tropical Pacific SST gradient will change in the future.

A primary reason for uncertainty in future changes in ENSO variability and the mean-state tropical Pacific SST gradient is that ENSO itself experiences large multi-decadal variability , meaning that long averaging periods or large ensembles are needed to identify robust forced changes e.g.,. Projections using multi-model ensembles can be difficult to interpret because the spread due to internal variability is comparable to the spread due to inter-model differences . With this in mind, large ensembles are needed to quantify the time-varying ENSO response . Without large ensembles, previous studies were forced to rely on long time averages to robustly evaluate changes in ENSO variability (e.g., 2000–2099 compared to 1900–1999).

In this study, we present ENSO projections from 14 SMILEs, for which the time-dependent response of ENSO to external forcing can be isolated from internal variability through ensemble averaging. The unprecedented number of SMILEs used allows a detailed examination of the model dependence of future ENSO projections. In particular, this paper aims to

isolate the forced response of ENSO in individual models,

The 14 SMILEs used in this study include both CMIP5- and CMIP6-class models and use one of three external forcing scenarios (Table ). Models used in this study have between 10 and 100 ensemble members for the historical and future scenarios. For those with single forcing we allow as few as three members due to the limited availability of models with any more than a single ensemble member. Previous work has investigated how many ensemble members are needed to investigate ENSO characteristics and find that this depends on the metric used, the length of time averaging, the level of acceptable error, and the model itself . find, for example, that for ENSO amplitude 10–47 members are sufficient depending on which model is used to make the estimate, while this range widens to 1–45 members for ENSO seasonality. determine that to look at ENSO variance over 30-year periods 10 members are sufficient for a 20 % level of acceptable error, while 30–40 are needed to reduce this to 10 %, with 12–20 sufficient for a 15 % error . Fewer members are needed to estimate mean-state ENSO metrics than its variability . In this study we accept models with a minimum of 10 ensemble members to include as many models as possible; however, we note that there are larger errors associated with estimates of ENSO variability for the smaller ensemble sizes. We additionally show the time series of ENSO indices in Fig.  as is acts to demonstrate the inherent variability due to different ensemble sizes.

While the purpose of this study is to provide a detailed overview of ENSO projections in all available SMILEs, it is useful to provide some observational context. A detailed comparison of ENSO characteristics in both CMIP models and SMILEs can be found in the following studies for a large range of ENSO metrics, and we do not repeat this here, although we note that in general CMIP6 models outperform CMIP5 in 8 out of 24 metrics and are only degraded in 1 and that ENSO-related SST biases and its teleconnections are also improved . We do, however, compare model-simulated and observed El Niño and La Niña composites (Figs. S1 and S2 in the Supplement). We find diverse errors in ENSO patterns, amplitudes, and transitions. All models overestimate the westward extension of ENSO SST anomalies and either overestimate or underestimate the ENSO amplitude, and some models overestimate the duration of ENSO events (e.g., MIROC6 and MIROC-ES2L).

When considering ENSO SST variability we find that only some models capture the observed variability within the ensemble spread (Figs.  and S3), but we note that this is not a reliable predictor for future change. For example, both MPI-GE and EC-EARTH3 compare favorably with the observed magnitude of ENSO SST variability but have different future responses (Fig. ). MPI-GE shows no change in ENSO SST variability, while EC-EARTH3 has strongly increasing ENSO SST variability. We additionally note that there is large multi-decadal variability in the amplitude of ENSO variability in individual ensemble members. This indicates that we need a SMILE to robustly detect a forced change, but also that an actual forced change in the real world might be masked by this large multi-decadal variability.

In this section we evaluate the response of ENSO variability in each individual SMILE in one of three future scenarios dependent on availability (RCP8.5, SSP370, SSP585). We note that there may be differences between SSP370 and RCP85/SSP585, particularly regarding early 21st century sulfate aerosols, which have potential consequences for the evolution of ENSO .

The warming rate over the equatorial Pacific varies seasonally and as a function of longitude (Fig. ). While some models have larger warming trends in the eastern Pacific (El Niño-like warming) for all months (MIROC6, CESM1-LE, EC-EARTH, CESM2-LE), others demonstrate a more complicated seasonal cycle of warming rate, with the location of the largest warming dependent on which month is considered. In the MEM, the eastern Pacific warms more in all seasons; however, from December to June the strong warming trend extends into the central Pacific, which warms with almost equal magnitude to the east. The MEM warming rate is larger over the eastern equatorial Pacific during late boreal spring (the climatological warm season of the equatorial Pacific) compared to boreal autumn (the cold season), but the season of strongest warming is model-dependent. For example, GFDL-CM3 has the largest eastern equatorial Pacific warming in boreal summer, while CESM1-LE has the largest warming in boreal fall and winter.

The time evolution of tropical Pacific SST gradient changes is investigated by considering the SST difference between the eastern and western equatorial Pacific (Fig. ). A decrease in this gradient indicates stronger warming in the eastern Pacific, i.e., El Niño-like warming (red in Fig. ), while a strengthening of the gradient corresponds to La Niña-like warming (blue in Fig. ). We find El Niño-like warming in most models, which is evident in the MEM. This is, however, not the case for CSIRO-Mk36 and is seasonally dependent for GFDL-CM3, GFDL-ESM2M, IPSL-CM5A-LR, and MPI-GE. With the exception of CSIRO-Mk36, models with a stronger climatological gradient tend to have stronger El Niño-like warming (Fig. S3b), although the correlation between the gradient climatology and the gradient change is not significant. The season during which projected changes are strongest is again model-dependent, but SST gradient changes are often concentrated in the season during which the climatological SST gradient is strongest in the specific model considered. The following section relates the projected changes in ENSO variability to the changes in the tropical Pacific SST gradient to investigate whether these changes are linked.

The projected change in ENSO SST variability is linearly related to the projected change in the tropical Pacific SST gradient across ensemble members from all models in the period 2021–2050 (Fig. ). This relationship breaks down in the later period (2070–2099) for models that have large SST gradient changes (e.g., CESM2-LE, CSIRO-Mk3.6, CanESM2). In general, an ensemble member that experiences El Niño-like warming is more likely to see an increase in ENSO variability, while a La Niña-like warming is associated with the opposite response. While this does not occur for all ensemble members, the larger the response in either the tropical Pacific SST gradient or ENSO SST variability the more likely that a concurrent change in the other variable will be seen.

We further investigate the relationship between changes in ENSO SST variability and the change in the tropical Pacific SST gradient by plotting the evolving relationship between the two variables over time (Fig. ). Many models have a hook-shaped ensemble mean (forced) response that corresponds to the following time-dependent response. First, there is an increase in ENSO SST variability concurrent with uniform warming across the Pacific (i.e., no change in tropical Pacific SST gradient) or weak El Niño-like warming. Then there is El Niño-like warming concurrent with a plateau in ENSO SST variability. In some models, this is followed by a decrease in ENSO SST variability as the mean state continues to warm in an El Niño-like fashion. This behavior is seen in seven models (CESM1-LE, GFDL-SPEAR-MED, MIROC6, MIROC-ES2L, CanESM5, CESM2-LE, IPSL-CM6A-LR), with an additional four that demonstrate a portion of the hook-shaped response (EC-EARTH3, ACCESS-ESM1-6, CanESM2, MPI-GE). The three outliers are GFDL-ESM2M, which exhibits an increase then a decrease in both ENSO SST variability and the mean-state tropical Pacific SST gradient, GFDL-CM3, which shows an increase then a decrease in ENSO SST variability without much change in the mean-state tropical Pacific SST gradient, and CSIRO-Mk36, which has strong La Niña-like warming and an increase in ENSO SST variability.

The time evolution of ENSO SST variability and the tropical Pacific SST gradient (Fig. ) provide context for the results in Fig. . The internal variability relationship that is the primary contribution to Fig. a clearly shows the role of rectification in the mean state e.g.,, but for the forced changes this is only one of several mechanisms at play, so the forced changes can depart from this linear relationship (Figs. b and ). The observational changes between 1951–1980 and 1990–2019 also show a notable departure from this linear relationship, with a strong La Niña-like mean-state trend together with a weak increase in Niño3.4 variance (Fig. ). The multi-decadal La Niña-like mean-state change in observations is near the edge of what could result from the internal variability found in models (z score = -1.88), suggesting that this mean state is likely partially forced see. Our results highlight the value of using SMILEs to analyze the time-dependent response of the tropical Pacific in climate models. We additionally note that the individual ensemble member spread is much larger for ENSO SST variability projections than the tropical Pacific SST gradient (Fig. ). This means that more ensemble members are needed to retrieve the forced change in ENSO SST variability under warming as well as higher-order moments of other Earth system variables;, but somewhat fewer are needed to isolate the forced change in the tropical Pacific SST gradient e.g.,.

Using two SMILEs (CanESM5 and MIROC6) that have single-forcing simulations available, we separate the contribution of aerosols, greenhouse gases, and natural forcing to historical changes in both ENSO SST variability and the tropical Pacific SST gradient (Fig. 10). Both models have forced increases in ENSO variability over this time period (1985–2014 compared to 1951–1980), although this is not evident in all ensemble members. This increase is consistent with observations and is driven by anthropogenic forcing (greenhouse gases and aerosols), not natural forcing. The larger ensemble (MIROC6) has strong variability across ensemble members in the natural forcing ensemble, with large increases or decreases observed in a single realization (or observations) due to this variability alone. This indicates that ENSO SST variability changes have not yet emerged from internal variability, consistent with previous work e.g.,.

Over the historical period, the tropical Pacific SST gradient weakens in MIROC6 and does not change in CanESM5 in response to all forcings. This is a result of the combination of greenhouse gas and aerosol forcings, which generally oppose each other (Fig. 10). Specifically, greenhouse gases act to weaken the gradient in both models, while aerosols strengthen the gradient in CanESM2 and have a minimal effect in MIROC6. The observed change in the tropical Pacific SST gradient sits right at the edge of the ensemble spread in the all-forcing historical scenario. This suggests that the observed gradient could be due to internal variability; however, it would be an extreme event. Differences between these models and observations could also be due to an incorrect model response to external forcing.

While the MEM change in ENSO SST variability in this study compares well with previous literature, the SMILEs add information by allowing for the isolation of the time-dependent ENSO response in individual models. In most models, although not all, ENSO amplitude is projected to increase, in agreement with and . However, while previous work used the differences between long time periods (i.e., 2000–2099 compared to 1900–1999) to isolate the projected change, we leveraged SMILEs to separate the time-dependent response of ENSO amplitude to external forcing from internal variability. Individual SMILEs have a range of projected evolutions of ENSO amplitude that are not linear in time. Additionally, in CESM2-LE, GFDL-CM3, and GFDL-ESM2M ENSO amplitude first increases, then decreases. In these models, the response captured using long time averaging is dependent on the specific period used. Here, SMILEs add value as an important new tool to fully capture the time-dependent ENSO variability projections.

The spatial pattern of projected changes in ENSO variability also differs between models. In the MEM, we find a general shift toward more variability in the central Pacific. However, the individual model patterns that make up the MEM are diverse, and more work is needed to understand the model differences that we find. The MEM increase in the central Pacific variability is also clear in the MEM composite of El Niño event evolution. Additionally, La Niña events appear to strengthen in the MEM more than El Niño events, highlighting nonlinear behavior. In general, for both pattern and event evolution changes, the individual model responses are similar between the two time periods considered (2021–2050 and 2070–2099), albeit smaller in magnitude for the earlier period. However, this is not true for all models, with CESM2-LE a clear outlier showing opposite changes in the earlier and later periods. Model differences could be related to different climatological biases, different patterns of transient mean-state warming, and different ENSO feedbacks and dynamics e.g.,, with additional work needed to diagnose why individual models behave the way they do.

We find that the seasonal synchronization of ENSO SST variability increases in most models investigated in this study. This manifests in an increase in ENSO SST variability in boreal fall and winter and a decrease in spring and early summer. These changes have implications for understanding ENSO's intrinsic dynamics e.g.,, changes to remote ENSO impacts (e.g., potentially amplifying the ENSO impact in one season for a given region and reducing it in a different season), and potential changes to ENSO predictability e.g.,. Thus, changes in variability should be assessed seasonally and not by using annual averages. We note that determining the dynamical cause for the increased ENSO seasonal synchronization in most of the models will require a detailed ENSO feedback analysis e.g.,, including assessing potential future changes in the southward wind shift mechanism e.g.,.

The tropical Pacific SST gradient response also agrees with previous work showing an El Niño-like warming in most models and the MEM e.g.,. However, again the individual model responses that go into the MEM are diverse, with the warming varying both seasonally and longitudinally. In general, the season of strongest SST gradient weakens the most; however, this is again not consistent across all models. In most models, weakening of the SST gradient intensifies in the second half of the 20th century, consistent with a weakening of the Walker circulation gradually overwhelming the ocean thermostat mechanism .

Similar to the previous hypotheses, we find a link between the projected tropical Pacific SST gradient and the change in ENSO SST variability . However, the SMILEs provide a much clearer picture of the time evolution of anomalies. We find a large subset of models that have an increase in ENSO SST variability that precedes the mean-state change, with ENSO SST variability plateauing, and in some models decreasing as the mean state warms in an El Niño-like fashion. This result begins to conceptually link together previous work that found an increase in ENSO SST variability over the coming century e.g., and a decrease in the longer-term equilibrated state e.g.,. The individual model responses may also put into context a recent high-resolution study that finds a decrease in ENSO SST variability and argues that CMIP-class models cannot capture all aspects of ENSO ocean dynamics correctly, leading to a potentially incorrect projected increase in ENSO SST variability. Based on our study, their model could be capturing a longer-term decrease in ENSO SST variability due to the strong forcing used (4 × CO2). Alternatively, we do find CMIP-class SMILEs that show a decrease in ENSO SST variability, highlighting the need to use multiple models for robustness. Based on these results, high-resolution modeling studies that use multiple models are needed to reconcile projections and determine the robustness of the higher-resolution result compared to standard-resolution CMIP-class models.

In the two models that have single-forcing SMILE experiments, greenhouse gas and aerosol forcings contribute to a historical increase in ENSO SST variability. While one might expect aerosol forcing to have the opposite effect of greenhouse gas increases, aerosols are hemispherically asymmetric unlike greenhouse gas forcing, leading to shifts in the Intertropical Convergence Zone that can also influence ENSO . In contrast, greenhouse gas and aerosol forcings have competing influences on historical trends in the tropical Pacific SST gradient. The observed change in the tropical Pacific SST gradient sits right at the edge of all ensemble members of the all-forcing historical scenario consistent with. This suggests that the modeled SST gradient response to greenhouse gas forcing could be too strong or incorrect, the modeled SST gradient response to aerosol forcing could be too weak, the observed change could be an extreme event caused by internal variability, the modeled internal variability could be too small, or some combination of the above. Further investigation is needed as more single-forcing ensembles are released to determine the relative roles of these possibilities.

In this study we highlight the value of SMILEs for investigating ENSO projections. SMILEs allow the isolation of the time-dependent response of ENSO SST variability to increasing greenhouse gases. By using SMILEs we can better understand projected changes in ENSO variability and the tropical Pacific SST gradient and isolate inter-model differences in their response to anthropogenic forcing.

Our results show the following.

ENSO SST projections a.

Projections of ENSO amplitude are not linear in time.

These results present an extensive picture of future changes in ENSO in 14 SMILEs. The use of SMILEs means that the diverse responses across models can be truly attributed to model differences, rather than including contributions from internal variability. These diverse responses demonstrate a need for further investigation into the processes causing model differences in ENSO projections and provide a baseline for future research. Additionally, our results highlight time-dependent behavior including nonlinear changes in ENSO SST variability and changes in the tropical Pacific SST gradient that intensify near the end of the 21st century. This has important implications for ENSO's teleconnections and impacts, as a nonlinear change in ENSO SST variability likely has nonlinear time-dependent changes in its impacts as well. Our results may also help to reconcile previous work that suggests a transient increase in ENSO SST variability, but a long-term equilibrated decrease, by isolating the time-dependent behavior of ENSO SST variability projections. While many models agree on the trajectory of projected changes, not all models behave the same way. This highlights the need for further research on the mechanisms of inter-model differences in ENSO projections. There is a rich diversity of future ENSO changes projected by climate models, and more work is needed to understand which aspects of these projections are robust.