The SSTs analyzed in this study are taken from the HadISST v1.1. data set (Rayner et al., 2003) covering the period 1948–2017 and released by the Met Office Hadley Center. 10-m wind data were taken from the National Centers for Environmental Prediction/National Center for Atmospheric Science (NCEP/NCAR) reanalysis data for the same period (Kalnay et al., 1996).

In order to overcome limitations associated to the relatively short data set used here for addressing decadal variability in the SPMM/ENSO relationship, we use long-term simulations of the NCAR Community Earth System Model (CESM) (Kay et al., 2015), a Coupled General Circulation Model skillful in representing ENSO diversity (Karamperidou et al., 2015; Dewitte and Takahashi, 2017; Cai et al., 2018; Carréric et al., 2020) and the NPMM characteristics (Liguori and Di Lorenzo, 2019). The 42 members of the historical runs (1920–2005) of the CESM Large Ensemble Project (CESM-LE) are used here consisting in a total of 3,600 years. This allows to estimate the spread between the members and thus uncertainties due to internal climate variability in some diagnostics. Additionally we use 62 models from the CMIP archive (Phases 5 and 6) (see Table 1).

We estimate the lagged correlation between ENSO and the SPMM indices (SST and winds) over the full period (1948–2017). The results are presented in Figure 2. First we note that the correlation at zero lag between the SPMM index and ENSO indices is zero by construction. The Figure 2A indicates that the correlation between the SPMM index and the ENSO indices reach statistically significant values for both positive and negative lags. However, the relationship is stronger for positive lags. In particular, the lag-correlation between the SPMM index and the E index peaks at 0.5 at lag 10 months for SST while it is much weaker (not significant) for the C index, indicating that the SPMM relates more to EP El Niño events than to CP El Niño. We checked that the results are only weakly impacted when the fields are not filtered prior to performing the SVD analysis to derive the SPMM index (not shown). The peak negative correlation (r = −0.3) is reached for a lag of ~17 months (SPMM leads ENSO), meaning that a positive phase of the SPMM tends to yield a reduction of the E index almost 1 year and a half later. In order to evaluate the stationarity of the relationship between SPMM and El Niño types, we extend the previous analysis by considering the lead-lagged relationship of the El Niño indices (NINO34, E and C) and the SPMM index over 21-years running segments (Figures 2B–D). Results indicate that the SPMM/E relationship for positive lags is rather stationary (E leads SPMM by ~10 months over most of the record) although its strength is enhanced (i.e., larger values of maximum correlation) after 1975 probably due to the enhanced variability of the E index (e.g., stronger ENSO) after the 1977 climate shift (An and Wang, 2000). For negative lags (SPMM leads ENSO), the inverse relationship between the SPMM and E is mostly observed over the period 1950–1995. On the other hand, the SPMM/NINO34 relationship (Figure 2B) exhibits a change from before and after 1980, which is similar to that of the SPMM/C relationship (Figure 2D). There is in particular a strong negative relationship between the SPMM-induced SST and the C index (SPMM ahead C) during the period 1984–2000, indicating that SPMM tends to precede La Niña events. This can be interpreted as resulting from some SPMM events culminating after El Niño events that are followed by La Niña events, like for the 1987/88 and 1997/98 El Niño events.

Due to limitations associated with the short observational record, we used the CESM-LE simulations to access a larger number of events and evaluate the influence of natural variability on the SPMM/ENSO relationship. The CESM model has also been shown to be skillful in accounting for ENSO diversity and asymmetry (Karamperidou et al., 2017; Cai et al., 2018; Carréric et al., 2020) and used for investigating aspects of the Meridional Modes (Liguori and Di Lorenzo, 2018, 2019). The SPMM variability is also rather realistic (Figure 3A), which motivates the comparison with observation (Figure 1). Figure 3B shows the ensemble mean lead-lagged correlation between the SPMM index and the E and C indices. It indicates that, like in the observations, that the SPMM/ENSO relationship is strongest for positive lags, that is when ENSO leads. The maximum correlation is reached for lag ~10 months and is larger for the E index than for the C index, like in the observations. On the other hand, the SPMM leads the C index with an inverse relationship (i.e., negative correlation), meaning that in this model a positive phase of the SPMM tends to precede La Nina conditions by ~7 months, which is distinct from what is observed (Figure 2A). The dispersion of the correlation amongst the ensemble is relatively weak (see shading in Figure 2A), suggesting a rather stable SPMM/ENSO relationship in this model. As a further evaluation of the role of low-frequency variability onto the statistics, 21-year running lagged-correlation analysis are performed on the 42 members of the historical runs (1920–2005) of the model yielding a total of 33,264 realizations of lead-lag correlation values between the SPMM index and the ENSO indices (E and C). The density function as a function of lag is then estimated, providing the distribution of the correlation values associated to the expected SPMM/ENSO relationship (Figures 3C, D). The “shape” of the distribution provides a metric of the decadal variability of the SPMM/ENSO relationship. Where it is wide and/or deviate from Gaussianity, this indicates that decadal modulation is likely and that the observational record may have limitations, assuming a “perfect” model. Note that the dashed blue line in Figures 3B, C corresponds to the ensemble mean of the lagged correlation between the SPMM and ENSO indices (i.e., curves of Figure 3C). The deviation of the value of correlation corresponding to the peak of the distribution (~50% percentile in case the distribution is close to Gaussianity) from this mean value would also indicate that decadal modulation in the SPMM/ENSO relationship is notable; This is because this would mean that a large number of correlation values for 21-years chunks does not match with the mean value of the correlation between SPMM and ENSO indices over the whole record. As an indication of the structure of the distribution, we indicated on the panels where the distribution passes the “dip test” (Hartigan, 1985) at the 95% level. This test is based on the hypothesis that the given data have more than one mode in their distribution. The test generates an unimodal distribution function that has the smallest difference values from the empirical distribution function (EDF). The maximum of these differences is referred to as “dip statistic”. The larger the value of the “dip statistic”, the more probable the empirical data have multiple modes. Where the distribution is either likely multimodal (at lags where red diamonds are indicated in Figures 3C, D) or skewed, this indicates a deviation from Gaussianity and therefore the influence of decadal variability on the statistics of the SPMM/ENSO relationship. Overall, results indicate some spread in the lagged correlation that can be visualized by the distance between the 10 and 90% percentiles (curves in thick black lines). Despite this spread, the SPMM index has almost no delayed influence on the E mode (i.e., SPMM leading E) because the correlation values for the 10 and 90% percentiles remains low. However, the distribution for negative lags of the relationship between the SPMM and the C mode (Figure 6C) suggests that there can be periods when the correlation reaches −0.5 (10% percentiles) from lag ~6 months and beyond, suggesting that the SPMM could be a precursor of either CP El Niño or La Niña events over some particular decades. That is a positive (negative) phase of the SPMM would precede a CP La Niña (CP El Niño) event. For positive lags (i.e., ENSO leads SPMM) and the E index, the model is in good agreement with the observations (i.e., the E index preceding the SPMM by ~10 months; Figure 3B). This is consistent with the distribution having a more leptokurtic shape than for negative lags, with a high value for frequency at correlation values near 0.5 (peak correlation at lag 10 months is 0.47 in the observations). This means that this high correlation coefficient is a robust feature with regards to the sensitivity to the period over which the statistics is done. Note that the correlation at lag 10 months reaches 0.70 for the 90% percentiles, indicating that for some decades, the SPMM is tightly linked to EP El Niño events. The 10-month lag indicates that the SPMM enters its peak phase when the coincident EP El Niño event enters either in its decay phase or has reversed into a La Niña event as it is often the case for strong EP El Niño events in that model (Okumura and Deser, 2010; DiNezio et al., 2017). This can explain why the SPMM is ahead the C index with an inverse relationship (i.e., negative correlation) because the C index captures La Niña events. Overall, this indicates that the SPMM follows EP El Niño events in both observations and the CESM-LE simulations. This is further illustrated by the scatter plots of the maximum value of the SPMM index between June and February vs. the DJF E index (Figure 4). The latter reveals in particular that strong SPMM events (i.e., events for which the amplitude of the SPMM index is larger than 1.75 time its standard deviation) are associated with El Niño events with a DJF E index larger than 1, which approximately corresponds to strong El Niño events based on the NINO34 index (i.e., events for which DJF NINO34 larger than 1.75 time its standard deviation). This has taken place 2 times over the observational record (2015/16, 1982/83); that is 66% of the time. Note, while the 1997/98 El Niño does not enter this category it is however associated with a moderate SPMM event. In the CESM-LE, this situation concerns 60% of the SPMM events (Table 2).

Considering the strong relationship between the E index and SPMM, we hypothesize that strong SPMM events are triggered/favored by coastal warming along the coast of central Chile and that the latter is induced by the equatorial oceanic teleconnection during EP El Niño event. As a first step, in order to diagnose the oceanic teleconnection along the coast, we estimate the lead-lagged correlation between the SPMM index and SST and sea level height (SSH) anomalies along the west coast of South America in the CESM-LE simulations (Figure 5). It shows that maximum correlation is found for positive lags (SPMM behind SST/SSH) and that the latter reduces with increasing latitude, implying a poleward propagation of SST/SSH anomalies of equatorial origin in relation with the SPMM. The correlation peaks in the latitude range 15°-30°S (15°-35°S) for SST (SSH), which corresponds to the location where the SPMM pattern seems to originate from the coast and has a strong loading (see Figure 3A). In order to infer the remote forcing mechanisms of these propagating SST/SSH anomalies, we estimate the composite evolution of SST and decomposed SSH anomalies (cf. method section) along the coast during strong SPMM events (Figure 6). Since upwelling favorable winds along the coast are modulated by the ENSO atmospheric teleconnection through fluctuation of the SPO, we also consider the composite evolution of 10-meter winds (brown arrows in the bottom panels) along with that of the SPO index (brown curves on the top panels). Figure 6 first indicates that strong SPMM events are associated with El Niño conditions as evidenced by the composite evolution of the equatorial Kelvin wave in the eastern Pacific that peak around DJF (year Y0).

Another manifestation of the occurrence of El Niño is the increase magnitude of the SPO index in Austral winter (Y0) (You and Furtado, 2017). The latter is associated with reduced upwelling favorable winds along the coast of Chile south of ~30°S (brown arrows). The coastal conditions feature a double-dip warming associated with downwelling waves originating from the equatorial region, the first one in Austral winter (Y0) and the other one at the peak of El Niño in Austral summer. Note that these waves are either Rossby waves that radiate from the coast or Kelvin waves trapped along the coast or a combination of the two kinds. In particular in Austral winter, El Niño start developing in conjunction with the forcing of intraseasonal Kelvin waves that are trapped along the coast while as El Niño culminates an interannual Kelvin wave is forced and radiate along the western coast as Rossby waves (Clarke and Shi, 1991), which results in a mixed-type of waves along the coast. While the sucessive SST anomalies off Peru are well defined with comparable propagating characteristics, they tend to dissipate from 15°S with altered southward propagation. In particular, the SST anomalies peaking in Austral summer propagates much slower from 15°S until 20°S, which is associated with the third baroclinic mode wave. In particular, the maximum amplitude of the third baroclinic mode contribution to sea level anomalies along the coast coincides with the persistent warming between 15°S and 30°S until Austral winter (Y1). This is the season when the El Niño event has transitioned to neutral to cold conditions [see reversal of SST anomalies from July (Y1) within 0° and 10°S] and the SPO activity is enhanced as evidenced by the large dispersion of the SPO index (brown shading). The coastal warming off Central Chile along with episodic positive SPO events is favorable for the triggering of a SPMM event. Values of the SPO [JJA Y(1)] index above its standard deviation take place 30% of the time during strong SPMM events associated with strong El Niño events in CESM, while 25% of these same SPMM events are associated with a negative SPO index in JJA (Y1) (i.e., below minus the standard deviation). This means that a large portion of the strong SPMM events following strong El Niño events (70%) can take place independently of a favorable state of the SPO, supporting the key role of the oceanic teleconnection along the coast of Peru/Chile in favoring the SPMM growth.

In this section we evaluate how models participating in the CMIP behave in terms of the SPMM/ENSO relationship evidenced here. We find that only a few models (10 out of 62) simulate to some extent this observed SPMM/ENSO relationship (Figure 7). Interestingly, this is independent of whether or not they can simulate realistically ENSO diversity (see Table 1), although the few models with skill (i.e., max correlation between SPMM and E index superior or equal to 0.35) do simulate a realistic ENSO non-linearity/diversity (α¯= −0.39 ± 0.11, see Table 1). The large number of models with low skill can be interpreted as resulting from biases in both mean state and equatorial variability. In particular the warm bias in the eastern Pacific (Cai et al., 2020) extends poleward along the coast (Figure 8), which may decouple surface processes with wave dynamics through deepening the thermocline and/or favoring faster wave propagation. This in turn may weaken the equatorial oceanic teleconnection associated with the SPMM. Also models with low skill have on average a weaker ENSO non-linearity than observations (α¯= −0.17 ± 0.14, see Table 1), which is associated to a tendency to simulate weaker EP El Niño events and therefore a weaker oceanic teleconnection along the west coast of South America. They have also a larger mean SST bias on average (Figure 8), which may affect both the oceanic teleconnection and regional air-sea interactions propitious for the development of the SPMM.