Main

The tropical Pacific is home to intense convection, allowing for strong thermal and dynamical interactions between the upper ocean and the overlying atmosphere15. As warm SST anomalies propagate eastwards during extreme El Niño events (for example, 1982–83, 1997–98; see Extended Data Fig. 1a), nonlinear dynamical heating processes tend to intensify the anomalously warm SST9, while the western Pacific warm pool (water with temperature exceeding 28 °C) extends eastwards, moving the eastern edge of the warm pool beyond 160° W. This induces an eastward shift of equatorial rainfall and an extreme swing of the Southern Hemisphere’s largest rainband, the South Pacific Convergence Zone11, causing extreme hydroclimatic conditions that most severely affect vulnerable island countries in the Pacific11,16. Beyond the Pacific, almost every continent felt the impacts of the drastic shift in weather patterns during the 1982–83 extreme El Niño event, and in the USA alone crop losses were estimated to be around $10–12 billion4 (approximately $24–26 billion in 2013 dollars).

These profound impacts demand an improved understanding of ENSO propagation dynamics. Many studies have evaluated the relative importance of various ocean–atmosphere feedback processes17,18,19, yet the mechanism for the propagation asymmetry remains unresolved. Here we show that an asymmetry in the zonal flow along the equatorial Pacific upper ocean (hereafter referred to as the equatorial Pacific current) is the main cause.

Using various observational data assimilation systems (see Methods and Supplementary Table 1), we quantify the propagation characteristics of temperature anomalies (Ta) by compositing the equatorial warming and cooling rates (the time derivative of temperature, dT/dt) of the surface mixed layer for the strongest El Niño events on record (1982–83 and 1997–98; see Fig. 1a) and all La Niña events (Fig. 1b) since 1976. The composite covers the evolution over a two-year period, before and after the event peak (usually in January, denoted ‘Jan (1)’; see Fig. 1 legend). The contour of marks the peak of the temperature anomaly (dashed curve in Fig. 1a and b). A linear regression using the samples of zero-value rates is constructed (green line). The slope β describes the propagation characteristics: a positive slope indicating temperature anomalies peak earlier in the west, that is, eastward propagation; a negative slope indicating westward propagation; and the greater the amplitude, the slower the propagation. This analysis shows opposite zonal propagation of SST anomalies between these two types of events; eastwards during extreme El Niño events (β = 0.82) and westwards during La Niña events (β = −0.46). During moderate El Niño events the propagation is westwards (Extended Data Fig. 1b), similar to La Niña events.

Figure 1: Equatorial Pacific current and zonal propagation of ENSO SST anomalies.
figure 1

a, Warming and cooling rate (colour scale; in units of °C per month) of the equatorial Pacific surface mixed layer on interannual timescales (an average over 5° S–5° N) composited over an extreme El Niño (1982, 1997) lifecycle across all reanalysis products. Red (or blue) contours indicate the associated positive (or negative) SST anomalies. Statistically significant values at the 95% confidence level are shaded and contoured. The monthly composites are shown over the developing and decaying ENSO phases, denoted as year 0 and year 1, respectively, in which the events typically peak in January of year 1 (‘Jan(1)’). b, As in a but for post-1976 La Niña events. The mean of the linear-fit slope (green line) of the phase transition (dashed black line), β (in units of seconds per metre), across the products is 0.82 in a and −0.45 in b, both significant above the 95% confidence level (P < 0.05) based on a bootstrap method with 1,000 draws. c, Zonal current velocity averaged across the reanalysis products, over 5° S–5° N, 160° E–90° W, and over August to December. The dashed curve highlights interdecadal variation using a 13-year running mean. Grey shading denotes two standard deviations about each mean value, representing monthly spread and variations across reanalyses. Red and blue filled circles indicate occurrences of strong El Niño and La Niña events, respectively, with relative event intensity indicated by different marker sizes. Open circles indicate moderate ENSO events.

PowerPoint slide

Source data

Full size image

The direction of propagation has been understood as arising from three main competing positive feedback processes17,18. The zonal advective and Ekman pumping feedbacks associated with fluctuations in the trade winds involve advection of climatological SST along the Equator by anomalous zonal currents (ua) and upwelling (wa)20,21, respectively (that is, and ; where the overbar indicates climatological mean and superscript ‘a’ indicates anomaly. The thermocline feedback involves vertical advection () associated with eastward-propagating internal waves that influence SST in the eastern Pacific through the mean upwelling () of water at the thermocline (a narrow depth range of strong vertical temperature gradients below the mixed layer). These processes can establish propagation of SST anomalies in either direction: eastward if the thermocline feedback dominates, and westward otherwise17. Linear theories have highlighted a higher importance of the thermocline feedback in the decades since the mid-1970s8,22,23; however, this would predict an eastward propagation during La Niña events as well19, in contrast to observations10 (Fig. 1b).

La Niña anomalies can be viewed as an enhancement of the prevailing climate, with stronger westward-flowing surface currents. On the other hand, eastward current anomalies during El Niño associated with anomalously weak trade winds24 (Extended Data Fig. 2), oppose and even exceed in amplitude the background current, leading to a net eastward flow (Fig. 1c). This asymmetry in the total zonal current is apparent in all reanalyses (Extended Data Fig. 3, Supplementary Tables 2 and 3). Our heat budget analysis (see Methods) shows that advection of anomalous temperature by the total current, , which is strongly westward during La Niña events but eastward during extreme El Niño events (Fig. 1c; Extended Data Fig. 4), represents one salient asymmetric feature that should be considered.

We therefore examine how the mixed-layer heat balance changes when the total current-induced heat flux is removed from the heat-budget equation (see Methods). Evolution of the residual warming and cooling rates shows that without the effect of the total current, an eastward propagation would result for extreme El Niño and La Niña events (red dashed line in Figs 2a and b), both with a positive slope of β*, 0.55 and 0.69, respectively. This reconstructs a linear framework in which the thermocline feedback dominates, leading to an eastward propagation for all events after 1976 (ref. 19). Thus it is the westward total flow that plays a key part in determining the westward propagation during La Niña events.

Figure 2: Effect of total and anomalous currents on equatorial Pacific mixed layer heat balance during extreme El Niño events and all La Niña events.
figure 2

a, Composite evolution of the interannual-scale mixed layer warming and cooling rates (colour scale; in units of °C per month) during extreme El Niño events (1982, 1997), with advection of temperature anomalies by the total current (arrow) removed. The red dashed line and the green line indicate the altered slope β* and the original β, respectively. b, As in a but for post-1976 La Niña events. Only statistically significant values at the 95% confidence level are shaded in colour, contoured or marked by black arrows (grey arrows otherwise). β* in a and b are respectively 0.55 and 0.69, both significant above the 95% confidence level (P < 0.05). c, As in a but with the effect of current anomaly removed. d, As in c but for La Niña events. β* in c and d are respectively 0.05 and −0.14, both not statistically significant (P > 0.4).

PowerPoint slide

Source data

Full size image

The role of the total current can be further understood by decomposing it into one component that is due to the long-term mean current and another due to the ENSO-related current anomaly . We find that the westward long-term mean current favours a westward propagation during both types of events; without it the eastward propagation during extreme El Niño events is more prominent and the propagation during La Niña events reverses to eastward (red dashed line, Extended Data Fig. 1d and e). The current anomaly, on the other hand, has an opposite effect on the two types of events (Supplementary Table 4). Without the effect of the eastward current anomaly, the eastward tendency during extreme El Niño events is severely weakened (red dashed line in Fig. 2c). This eastward anomalous current, stronger than the westward mean current, leads to a flow reversal during extreme El Niño events (Fig. 1c), making the eastward propagation characteristic more prominent (Fig. 2a). During La Niña events, on the other hand, the anomalous current reinforces the effect of the westward mean flow for a more pronounced westward propagation (Fig. 2b and d and Extended Data Fig. 1e). During moderate El Niño events, the eastward current anomaly is far weaker than the mean current, and thus the total current remains westward (Fig. 1c, Extended Data Fig. 1c). Thus, for moderate ENSO events, there is no asymmetry, and SST anomalies for both El Niño and La Niña events propagate westwards (Supplementary Table 4).

The nonlinear effect is more prominent post-1976 (Supplementary Table 4) when El Niño events are stronger9,25 with large eastward current anomalies that are occasionally comparable to, or greater than, the mean background current (Fig. 1c). During such events, this effect reinforces eastward propagation induced by the thermocline feedback. During a La Niña event, the westward current, along with the zonal advective and Ekman pumping feedbacks, weakens the thermocline feedback effect, resulting in a net westward propagation (Fig. 3). This superposition of ENSO-related large current anomalies onto the long-term mean westward current invalidates the assumptions of linearity, making linear theories unable to explain the propagation asymmetry.

Figure 3: Schematic of competing effects on zonal propagation direction during ENSO events.
figure 3

a, Zonal currents in the equatorial Pacific (large grey arrow) have the effect of shifting the initial warm surface anomalies (dashed red patch) eastwards during extreme El Niño events, because the current anomaly ua (red arrow) is eastward and exceeds the strength of the westward background current (black arrow). This effect counters westward propagation as induced by the zonal advective and Ekman pumping feedbacks (blue arrow) and enhances eastward propagation induced by the thermocline feedback (pink arrow). b, During La Niña events, the zonal currents are prominently westward because the current anomaly always enhances the westward-flowing mean current. This weakens the thermocline feedback effect and enhances westward propagation as induced by the other two dynamical feedbacks.

PowerPoint slide

Full size image

Thus, interplay between the ENSO-related current anomaly and the climatological current determines the way the equatorial Pacific circulation influences the zonal propagation of SST anomalies. This means that a change in ENSO intensity or in the mean current can influence the extent to which the propagation asymmetry can be observed. The post-1976 prominence of the propagation asymmetry is partly because of the extremity of the 1982–83 and 1997–98 El Niño events. The mean current itself weakened through the 1980–2000 period (dashed curve in Fig. 1c), with a consistent weakening of the trade winds (Extended Data Fig. 2b). Although this mean current reduction could be interpreted as a rectification of a change in ENSO variability26,27, a weakened mean current will favour occurrences of an eastward propagation, even if El Niño intensity does not change. At present, there is no agreement among climate models on the magnitude of future ENSO events28,29. However, the consensus that emerges is a future with weaker equatorial mean westward currents12,13,14. Our study implies that this would increase the likelihood for occurrences of El Niño events with prominent eastward propagation characteristics.

To this end, we analysed 40 climate models that participated in the Coupled Model Intercomparison Project phases 3 and 5 (CMIP3 and CMIP5), subject to increasing atmospheric CO2 concentration (see Methods). With the large number of simulated ENSO events, the multi-model aggregate demonstrates robust statistics reaffirming the above conclusion that weaker mean currents and current reversals, which are projected to be more typical in the future, facilitate eastward propagation (see Methods and Extended Data Figs 5, 6, 7). Indeed, we find that a subset of models that are more realistic in terms of flow features and frequency of El Niño events with prominent eastward propagation (19 models; see Methods and Extended Data Figs 8, 9, 10 for model selection) simulate a 100% aggregate increase in the mean occurrence of such El Niño events (Fig. 4a), from about 2.7 events in 1907–1999 to about 5.4 events in 2006–2098.

Figure 4: Statistics of El Niño occurrence characterized by prominent eastward propagation in CMIP3 and CMIP5 models.
figure 4

a, Probability density function of 19 model ensemble means in the past (1907–1999; blue) and future (2006–2098; red) periods. Each probability distribution is generated from 5,000 bootstrap draws using the 19 model samples. The solid curves overlaid are the corresponding probability density functions of a fitted normal distribution, which are significantly different above the 95% confidence level (P = 0.01). Dashed red and blue vertical lines indicate the corresponding ensemble average. b, Multi-model histogram for the event occurrence as a function of long-term averaged equatorial Pacific current velocity in the past (blue) and future (red) periods, segregated into bins of 0.04 m s−1. c, As in b but for the Niño3 index, which is detrended and normalized by the standard deviation of the past period, segregated into bins of size 0.25. d, As in c but stratified according to the concurrence with (thick dark bars) and without (thin light bars) current reversals. The dotted vertical line in d marks 1.5 units of the normalized value.

PowerPoint slide

Source data

Full size image

The role of the current is further highlighted as the increase in eastward-propagating events is skewed towards weaker mean currents (Fig. 4b; Extended Data Fig. 10b), and this occurs for El Niño events of all magnitude (Fig. 4c). There is, in particular, a tendency for the increase to be larger in models that project stronger ENSO amplitude (Extended Data Fig. 10c), which is in turn associated with more occurrences of a current reversal (Extended Data Fig. 10d), a feature unique to the 1982–83 and 1997–98 El Niño events.

Stratifying the statistics (Fig. 4c) in terms of a current reversal or otherwise, 45% of the increase is found to be associated with current reversals, most of which are El Niño events stronger than the typical magnitude of past events (Fig. 4d). However, the inter-model consensus for ENSO amplitude projection is weak, despite a reduced mean current in all models (Extended Data Fig. 10b and c). This suggests that a weakened mean current is the determining factor for future increases in eastward propagating events of all magnitudes, including extreme El Niño events, either through the thermocline feedback effect or a current reversal, or both.

In summary, although different factors have been proposed to explain various nonlinear characteristics of ENSO9,25,30, none have been found to explain the cause for its propagation asymmetry, as observed in recent decades. Here we have provided observational and modelling evidence that the equatorial Pacific current is an important element of this asymmetry. The superposition of a current anomaly during ENSO onto the long-term mean westward flow enhances the westward currents during a La Niña event, but reverses the currents during extreme El Niño events. The role of the equatorial currents highlighted here casts a fresh perspective on the fundamentals of ENSO behaviour. Given the projected weakening of the background mean flow under global warming, our analysis not only resolves a perplexing scientific issue, but suggests that increased occurrences of ENSO propagation asymmetry will be a manifestation of global greenhouse warming, with important socioeconomic consequences.

Methods Summary

The propagation tendency of temperature anomalies during ENSO events is quantified as the slope of the zero-value contour of warming and cooling rates on ENSO timescales that tracks the peak of temperature anomalies as they evolve in time along the equatorial Pacific (Fig. 1a, b). A positive (or negative) slope in the time–longitude space indicates eastward- (or westward-) propagating temperature anomalies: the steeper the slope the slower the zonal propagation and thus the more observable the propagation characteristic. To investigate the factors that can cause temperature anomalies to propagate zonally, we conduct a heat budget analysis of the ocean surface mixed layer. All variables are derived from five ocean reanalysis systems that assimilate high-quality observational products going back to 1980 or earlier (Supplementary Table 1). Our surface heat balance explicitly expresses the zonal advection of temperature anomaly by the mean current and is considered to interact with the nonlinear component (that is, advection by anomalous current). Removing certain heat-flux components from the heat balance would alter the slope of the zero-value contour, and so by comparing the altered slope (β*) with the original (β) its influence on the propagation can be inferred. Because this study concerns asymmetry between El Niño and La Niña, a composite approach is adopted (see Methods for classification of ENSO events). The implication of our results for future climate is assessed through the analysis of 40 CMIP3 and CMIP5 climate models (see full Methods).

Online Methods

Heat budget analysis

We consider the heat balance of the surface mixed layer, which can be expressed as follows:

The variables T, u, v and w are potential temperature, and the zonal, meridional and vertical ocean current velocities respectively. Differential operators, x, y, z, and t, are along zonal, meridional, and vertical directions, and time, respectively. Superscript ‘a’ and overbar denote anomalous and long-term averaged quantities, respectively. All variables are averaged between 5° S–5° N, over the surface layer depth of 50 m. The rate of change of the mixed layer temperature (dT/dt) is calculated as monthly increments using a centred-difference approximation. Terms not explicitly expressed in equation (1), such as eddy effects and damping by net air–sea heat fluxes, are absorbed into ‘Res.’, such that the left- and right-hand sides of equation (1) are identical. Equation (1) is slightly different to that adopted in previous studies8,20,21, in that is expressed explicitly here and is considered to interact with the nonlinear advection term . This combination , where , is simply interpreted as the zonal advection of temperature anomalies by the total equatorial Pacific zonal current, which can be readily observed. The term tends to be overlooked because it is convolved into the continuity component via volume conservation—that is, —when the heat budget is expressed in flux form: .

Quantification of propagation characteristic

The contour of dT/dt = 0 marks the peak of Ta, so its positive (or negative) slope in time–longitude space indicates eastward- (or westward-) propagating Ta (Fig. 1a, b). The phase transition slope, β, is calculated by fitting a line via the least-squares method to the contour between 160° E–80° W and May(0)–May(1) to allow some room for temporal movement upon removal of the advection terms. The rationale for the longitudinal extent is described below under the heading ‘Mean currents and ENSO propagation structures across models’, and our results are not sensitive to this aspect of the calculation.

Removing an important advection component from the right-hand side of equation (1) would alter the spatial and temporal structure of dT/dt, thus affecting β. For example, a reversal from a negative slope (that is, westward propagation) to a positive slope (that is, eastward propagation) would suggest that the component removed is crucial in setting the westward propagation. In this way, the role of a certain advection term on the propagation tendency of Ta can be determined by comparing the altered slope β* to the original β. The 95% regression standard error for the slopes is considered in all analysis by setting any slopes to zero if they are not greater than their corresponding error estimates.

Data sets and data processing

The reanalysis products used are ECMWF ORA-S331, ECMWF ORA-S432, SODA-2.1633, SODA-2.2.434 and GODAS35 (Supplementary Table 1). Each reanalysis system assimilates the available observations (such as hydrographic profile data, moorings and satellites) into an ocean model forced by observed surface wind stress to calculate ocean currents. The reanalysis systems use different ocean models and data assimilation techniques. To focus on processes at ENSO timescales, a Butterworth low-pass filter36 is applied before analysis to remove signals with periods shorter than 18 months. Without filtering, the spatio-temporal structure of the warming and cooling rate dT/dt is noisy, given large high-frequency monthly fluctuations. On ENSO timescales, the rate of warming and cooling smoothly tracks the evolution of SST anomalies.

ENSO classification and statistical significance test

The classification of ENSO events is based on the Niño3 index derived from the National Oceanic and Atmospheric Administration (NOAA) extended reconstructed SST version-3b37, averaged over December–February, when ENSO events typically peak. ENSO events are defined if the Niño3 amplitude, within each of the pre-1976 (1959–1976) and post-1976 (1976–2011) periods, is greater than 0.5 units of standard deviation. We classify these as strong if Niño3 exceeds 1 unit of standard deviation, and as moderate or weak otherwise. This yields the following classification of events (the developing phase year is quoted): strong El Niño events in 1965, 1969, 1972, 1982, 1991, 1997 and 2009; strong La Niña events in 1970, 1973, 1975, 1988, 1998, 2007 and 2010; moderate El Niño events in 1963, 1976, 1987, 1994, 2002 and 2006; moderate La Niña events in 1964, 1967, 1984, 1995 and 2005. The 95% statistical significance for each composite is evaluated using a bootstrap approach38 in which samples of size N are randomly drawn repeatedly to obtain 1,000 mean values. N is the number of ENSO events within each respective period pooled together for all the reanalysis products. All significance levels are evaluated based on the two-sided P-value.

Analysis of climate models

The observational analysis results demonstrate that, in the backdrop of the effects by the three ENSO dynamical feedbacks, the equatorial Pacific current is an important element for the zonal phase propagation of ENSO SST anomalies (Fig. 3). The observation-based results are further corroborated through an analysis of 40 CMIP339 and CMIP540 climate models (see Extended Data Fig. 5 for the specific models). The 40 models, each of 186 years in record (inclusive of the past and future simulations), provide a large sample of ENSO events that is about 180 times larger than the observed sample of 25 events. Thus, the models, with their different mean climate states, provide a rigorous test bed for the effect of the current, which, along with the implications for the future, are discussed below.

The past and future climate simulations respectively correspond to the twentieth-century (1907–1999) and future projection scenarios (2006–2098) based on the Special Report on Emissions Scenarios (SRES) A1B for CMIP3 and representative concentration pathways (RCP) 4.5 for CMIP539,40. The time spans were necessarily chosen to include as many models as possible that cover the longest record without any missing data.

Mean currents and ENSO propagation structures across models

On the basis of the findings of early theoretical studies17,18 the prevalent direction of the basin-scale ENSO SST anomaly propagation along the Equator is an indicator for the dominant dynamical process over a given epoch: net eastwards for thermocline feedback and net westwards for zonal advective/Ekman feedback. This definition for the dominant ENSO dynamics has been adopted by previous studies41,42,43, which we refer to hereafter as the ‘ENSO propagation structure’ (rather than ‘ENSO mode’) to tie in with the topic of our study (zonal propagation).

The diagnosis for ENSO propagation structure in observations and models has been achieved previously through a lead-lag correlation between the Niño3 index and an east-minus-west SST index which is taken as the difference between the Niño4, representing SST variability in the Central Pacific, and the Niño1+2 for the far eastern Pacific41,44. The former is bounded in the west at 160° E and the latter in the east at 80° W, which is the exact longitudinal extent adopted in our study for calculating the phase transition slopes.

Here we diagnose the propagation structure in each past and future period in each model (Extended Data Fig. 5a) by the proportion of westward events (assigned as negative proportion) and eastward events (positive proportion) identified as El Niño and La Niña events with a statistically significant β. For each given period, the proportions of those four types of propagating events (that is, westward El Niño and La Niña, and eastward El Niño and La Niña; the red/blue bars and lines in Extended Data Fig. 5a) and non-propagating events (non-statistically significant slopes) add up to 1, and so the net propagation structure (grey circles for 1907–1999; black triangles for 2006–2098) can range from −1, if all of the events propagate westwards, to +1 if all propagate eastwards. For example, the past ENSO events in model number 3 consist of 10% westward El Niño events, 17% westward La Niña events, 28% eastward El Niño events, 19% eastward La Niña events, and 26% non-propagating El Niño and La Niña events. Summing the proportions of the propagating events and considering the directions (−0.1) + (−0.17) + 0.28 + 0.19 yields an eastward propagation structure with a relative scale of 0.2 as marked by the grey circle. Although our approach is different to the commonly used correlation-based methods3,4,5, in that we utilize β, the results using the two methods are largely consistent (figures not shown).

We find a significant positive inter-model correlation between ENSO propagation structure and mean equatorial currents (Extended Data Fig. 5b): models with weaker mean currents tend to generate a higher proportion of eastward-propagating ENSO events, and the tendency is statistically significant. This suggests that models with weak (or strong) mean currents tend to be more (or less) favourable for the thermocline feedback resulting in an eastward propagation structure (as explained in Fig. 3). Some of the models that simulate too many eastward-propagating La Niña events (Extended Data Fig. 5a; for example, models number 2, 3, 10, 17, 24 and 25), in contrast to observations (but consistent with linear theories), tend to have a weak mean current. Because the inter-model correlation between the propagation structures and mean zonal wind stresses is basically zero (Extended Data Fig. 5c), such an effect is evidence for the direct influence of the ocean currents (for example, related to specifications of the ocean model components), rather than, for instance, an effect of ENSO rectification onto the mean climate. These inter-model relationships also hold for the future simulations (see Extended Data Fig. 5 legend). Although this result has an important implication for ENSO modelling, this in itself is evidence that the ocean current does have an influence on ENSO zonal phase propagation, that is, a weaker mean current is more favourable for eastward propagation.

The model ensemble results in Extended Data Fig. 5b also imply that in a climate state with a weak background current, natural variability alone (within which the system supports naturally varying ENSO propagation structure) would more easily produce eastward-propagating events. With even weaker currents projected for the future, consistent with the weaker trade winds, the thermocline feedback effect for inducing eastward propagation is favoured further (Extended Data Fig. 6). Previous ENSO stability analysis for a number of the CMIP3 models45 demonstrated that the three main positive feedback processes are projected to increase, and would thus have competing effects on zonal phase propagation. The clear increase in the occurrences of eastward propagation events (Fig. 4a) can be more simply explained in terms of a weakened current as described in our study.

Effect of current reversals and models selection

One characteristic of the ENSO system is that the equatorial Pacific current anomaly is correlated with SST anomalies in the east (represented by the Niño3 index) in which the current leads Niño3 by about three months (Extended Data Fig. 4b). This highlights the tendency for an eastward (or westward) current anomaly in boreal autumn (September–December) to precede the peak of El Niño (or La Niña) in boreal winter (December–February). A particularly strong eastward anomalous current was observed during the 1982–83 and 1997–98 extreme El Niño events that leads to a re-intensified reversal in boreal autumn, a feature not seen in other events (Extended Data Fig. 4a). These extreme events are identified by their prominent eastward propagation with phase transition slope β that is stronger than in other events (Extended Data Fig. 4d). Here we demonstrate, using an aggregate of models, that current reversals have the effect of making eastward propagation characteristic more prominent. Because zonal propagation is the focus of our study, and given the dynamical links of the aforementioned features, we first select the models based on the following criteria: (1) The models must be able to simulate at least one prominent eastward-propagating El Niño event in either past or future simulation. Such an event is defined as one for which β is positive, greater than the linear-regression standard error, and above 0.5 standard deviations of all El Niño slopes (that is, following the definition for the observed slope prominence; Extended Data Fig. 4d). (2) The models must be able to simulate at least one current reversal during boreal autumn in either past or future simulation. (3) The models must produce a positive correlation between Niño3 and the current during any propagating El Niño events, a relationship also seen in observations (Extended Data Fig. 4c).

These criteria result in 24 models that simulate more realistic and distinctive current evolution between strong and moderate El Niño years (Extended Data Fig. 8) as expected from observations (Extended Data Fig. 4a), in contrast to that in the discarded models (Extended Data Fig. 9).

The effect of current reversal on zonal phase propagation is clearly exhibited by this aggregate of models, that is, it favours eastward propagation. This is due to the fact that the corresponding β tends to be more positive whenever the events coincide with a current reversal (Extended Data Fig. 7b). In the case where current reversals coincide with westward propagation, the westward slopes are found to be substantially weaker. Such an effect renders a positive correlation between the total current and β (Extended Data Fig. 7a), which is a characteristic also seen in observations (Extended Data Fig. 4d). This positive correlation further highlights the crucial role of the equatorial Pacific current on zonal phase propagation.

The effect of the total current on El Niño and La Niña propagation asymmetry is also reproduced (Extended Data Fig. 7c). The asymmetry becomes apparent with strong El Niño events, and more so when these co-occur with current reversals, similar to the observed counterpart (Fig. 2a, b; Supplementary Table 4).

An additional criterion is applied, resulting in a further exclusion of five models. Each of these excluded models already simulates 11 to 14 El Niño events with prominent eastward-propagation slope over the 93 years in the past simulation (Extended Data Fig. 10a). These are too frequent relative to the two events over the 53 years of the observational record, which translates to slightly less than four events for the 93 model years. The remaining 19 models simulate from 0 to 8 events (that is, double the expected observed frequency) in the past period. These 19 models also have climatological states that roam the regime of westward-propagation structure similar to the observed, as opposed to the five excluded models that tend to cluster about the eastward regime with already weak mean currents (Extended Data Fig. 5b; model numbers 8, 17, 24, 25 and 29). Given the extreme rarity, and to test whether a change in model climatological state can induce increased occurrence of such events, we retain the 19 models for future projections (Fig. 4).

A parallel between future projection and the late twentieth century

With the mean westward currents projected to weaken in the future (Extended Data Fig. 6), thus providing a more conducive condition for increased occurrences of current reversals (Extended Data Fig. 7d), it is expected that there will be more El Niño events exhibiting a prominent eastward-propagation characteristic in the future.

A 100% increase in the mean occurrence of such events is found (Fig. 4a), with 16 out of 19 models projecting an increase. Considering only models that simulate fewer than 8 events increases this to more than 116%, with model consensus consistently above 83%. As expected, retaining those that already simulate frequently occurring events (that is, saturated with eastward propagation) reduces the amount of increase to 76% when including models that already simulate up to 11 events, and 46% using all of the 24 models. Nonetheless, in all cases, the models as an aggregate simulate a notable increase in future occurrences of eastward-propagating El Niño events that is significant well above the 95% confidence level, with at least 75% of the models projecting an increase.

As revealed by the observational analysis, the emergence of an eastward propagation in the post-76 period is in part because the mean westward current is weaker, and in part because the eastward current anomalies associated with the extreme El Niño events are sufficiently large to reverse the current. In this regard, the variety of events and mean states provided by the different models point to a slightly different scenario for the future in which the importance of the projected current weakening is highlighted. This is evident as the model consensus is weak in the projection for stronger ENSO amplitude (11 out of 19 models; Extended Data Fig. 10c), but all of the models project a weaker mean current (Extended Data Fig. 10b). Despite this, there is still a tendency for stronger increase in the number of eastward-propagating events in models that also project a larger increase in ENSO amplitude (Extended Data Fig. 10c). This is through the contribution by current reversals (Extended Data Fig. 10d), which tend to occur with stronger El Niño events and have the effect of making the eastward propagation characteristic more prominent (Extended Data Fig. 7).

We note that although a weaker mean current facilitates current reversal, such that any modest eastward current anomaly can more easily exceed the background current, the increase in the number of current reversals in the future (Extended Data Fig. 7d) does not always translate to more occurrences in events having a prominent eastward-propagation characteristic (Extended Data Fig. 10d). This is expected, given the various kinds of event concurrences that the model aggregate provides (Extended Data Fig. 7). In fact, although all of the increase in eastward-propagating events is associated with weaker mean currents and El Niño events of all magnitudes (Fig. 4b and c), only 45% of this is associated with current reversal events, within which 85% are associated with large-magnitude El Niño events (Fig. 4d). Thus, given the weak model consensus in projecting an increase in ENSO amplitude, the most robust feature shared between the future projection and the change observed during the late twentieth century is the weaker westward mean current, which is projected by all of the models.