Introduction

The Atlantic Niño has attracted considerable attention from the climate community1,2,3,4, as it comprises the dominant source of interannual climate variability in the tropical Atlantic region5,6 and serves as a potential precursor of El Niño-Southern Oscillation (ENSO)7,8,9,10,11,12.

The Atlantic Niño could be driven by multiple mechanisms13,14,15. For example, Bjerknes feedback is considered as the fundamental mechanism of Atlantic Niño5,16. Thermodynamic processes are also suggested to be non-negligible for the Atlantic Niño17,18. Perturbations across the off-equatorial areas of the Atlantic Ocean19,20,21,22 and the tropical Indian Ocean23 can trigger the Atlantic Niño as well. In particular, the South Atlantic Ocean Dipole (SAOD), which is a southwest–northeast oriented dipole of sea surface temperature anomalies (SSTA) in the South Atlantic, is found to be closely related to the appearance of Atlantic Niño events24,25.

The Atlantic Niño predictability is relatively less studied. The prediction of Atlantic Niño remains challenging and unsatisfactory26,27,28,29,30. For example, the skillful prediction is restricted to only approximately 4-month lead31 using the North American Multi-Model Ensemble (NMME)32,33. Model predictions only capture the evolution of some extremely strong Atlantic Niño events at a 1-month lead34. In addition, it was shown that models with a stronger connection between the boreal autumn Indian Ocean Dipole and the following winter Atlantic Niño appear to have a higher Atlantic Niño prediction skill35. All these studies examined the prediction skill over a fixed length of time period and did not investigate the variation of prediction skill over a long period of time.

Recently, it has been revealed that the Atlantic Niño variability has significantly declined (~30%) during 2000–2017, in comparison with that during 1982–1999, which could be caused by a weaker Bjerknes feedback and stronger latent heat flux damping36. Additionally, the Atlantic Niño variability was projected to be suppressed under the global warming impacts37,38. Although 2019 and 2021 experienced the resurgence of strong events39,40, it is still interesting and important to examine whether there would be a change in the Atlantic Niño prediction skill given the weakened Atlantic Niño variability, and what could be driving it.

Based on above considerations, we examine the changes in Atlantic Niño prediction skill by evaluating the NMME predictions over two time periods (1982–2000 and 2000–2018). More importantly, potential sources responsible for the changed prediction skill are explored.

Results and discussion

Decline in Atlantic Niño variability and prediction skill

The Atlantic Niño is represented using the ATL3 index41. Its variability and prediction skill are shown in Fig. 1. The 15-year sliding ATL3 STDs calculated separately from HadISST42 and ERSST.v543 datasets both show a decrease in ATL3 magnitude since 2000, which is consistent with the previous study36. Figure 1b displays the prediction skills of the multiple-model ensemble mean (MME) over the two subperiods. The MME anomaly correlation coefficients (ACCs) at all lead times are significantly lower over the second subperiod, supported by a bootstrap test with resampling times of 10000. The MME ACC decreases ~0.1 at the 0-month lead and ~0.3 at the 5-month lead. At the 2-month lead, the MME decreases most significantly by 38% from 0.74 to 0.46. Although the decrease in prediction skill is consistent with the decrease in variability, the former may not necessarily be driven by the latter. Scatter plots of ATL3 prediction skills against ATL3 variabilities (Supplementary Fig. 1) suggest that the Atlantic Niño prediction skill decline has no clear linear relationship with the weakened variability.

Fig. 1: The Atlantic Niño variability and prediction skill.
figure 1

a 15-year sliding standard deviations (STD) of ATL3 index calculated separately with the HadISST and ERSSTv5 datasets. b ACCs of the multiple-model ensemble mean (MME) over the two subperiods. c Persistence prediction skills over the two subperiods. The x-axis of (a) indicates the last year of each 15-year sliding window. Error bars in (b) and (c) indicate the 95% confidence interval determined by a bootstrap test with resampling times of 10000.

Figure 1c shows that the persistence prediction skill declines over the second subperiod as well, indicating a decrease in predictability. However, the persistence prediction skill decreases significantly only at the short leads and does not even change at the 5-month lead. Sliding ACCs of MME and persistence at different lead times, and their difference (Supplementary Fig. 2) confirm that the MME prediction skill decreases more significantly over the second subperiod and that there is an inconsistency between the changes in the MME prediction skill and persistence. As a result, the MME and persistence prediction skills become comparable at the end of the examined time period (Supplementary Fig. 2c), meaning the failure of dynamical predictions. All these results imply that the significant decline in Atlantic Niño prediction skill in the NMME may relate not only to predictability but also to deficiencies in dynamical predictions, which have great impact on prediction skills44,45,46.

Changing relationship between the global ocean and Atlantic Niño

To investigate the difference between the two subperiods, we focus on the 2-month lead prediction that shows the most pronounced decline (Fig. 1). For convenience, verification data are referred to as OBS unless otherwise noted. The observed and predicted ATL3 index are regressed on the global SSTA over each subperiod, separately for OBS (Fig. 2a, b) and MME (Fig. 2e, f). Pre-2000, the Atlantic Niño in OBS not only shows a strong correlation with the equatorial Atlantic SSTA, but also shows a significant correlation with global SSTAs over regions such as the subtropical Atlantic Ocean, western Pacific, and eastern Indian Ocean (Fig. 2a). This regression pattern exhibits an inter-Pacific–Atlantic SST gradient and zonal wind anomalies over the equatorial Atlantic, favoring the Bjerknes feedback and development of Atlantic Niño7,15,47. Post-2000, the Atlantic Niño is less associated with SSTA in the tropical Atlantic as well as other ocean basins (Fig. 2b). However, the SSTA dipole structure in the South Atlantic seems to be strengthened and the Atlantic Niño is more closely related to the anomalous cyclone in the South Atlantic Ocean (Fig. 2b), which has been revealed to have a robust link to the Atlantic Niño21,25,48. The correlations between the observed ATL3 and SWP (southwest pole of the SAOD) indices are −0.09 and −0.38 over the first and second subperiod, respectively, indicating the enhanced ATL3-SWP relationship over the second subperiod. It has been found that perturbations of the subtropical anticyclone centered at approximately 30 °S to be the predominant mechanism linking the southern extra-tropics and equatorial Atlantic variability, and leading to the SAOD. As such, the SAOD may play a role in the prediction skill decline24,25.

Fig. 2: Changing relationship between the Atlantic Niño index and various climate fields over the two subperiods.
figure 2

a, b regressions of ATL3 index onto the global SSTA (shading) and 850 hPa wind anomalies (vectors) in observations; c, d regressions of ATL3 index onto the global sea level pressure (shading) and 200 hPa wind anomalies (vectors) in observations (OBS); e, f regressions of ATL3 index onto the global SSTA at the 2-month lead in the multi-model ensemble mean (MME) predictions. a, c, e show results pre-2000; b, d, f show results post-2000. Only regressions with significance level of p < 0.05 are plotted. Black boxes in (e, f) indicate the northeastern pole (NEP) and southwestern pole (SWP) of the South Atlantic Ocean Dipole (SAOD). Wind anomalies are missing in (e, f) since wind data are not available from the NMME.

Previous studies have noticed the shift of the Atlantic–Pacific connection and explained it via the modulation of low-frequency variability modes49,50,51. For example, there is a shift from a period of strong Atlantic–Pacific connection to a period of weak Atlantic–Pacific during the 1970s7. In the strong connection period, the Walker circulation is strengthened, favoring the development of coupled processes, and leading to a good correlation between ENSO and the Atlantic Niño, and vice versa in the weak connection period7. Moreover, the Atlantic Niño resembles the canonical equatorial mode in the strong connection period, while it behaves as a dipolar structure during the weak connection period52, notably similar to the SAOD24,25. Figure 2c, d further displays the regression maps of ATL3 against the global sea level pressure and 200 hPa wind speed anomalies, supporting the weakening Atlantic–Pacific connection since 2000, and indicating a diminished Walker circulation7,52. All these facts indicate that a climatic regime shift occurred around 2000 when the connection of Atlantic Niño and global SSTAs switches from strong to weak. The changing relationship between the global ocean and Atlantic Niño may be responsible for the decreased Atlantic Niño predictability and persistence.

The MME prediction at the 2-month lead well reproduces the climatic regime shift. As can be seen in Fig. 2e, f, the connection of Atlantic Niño and global SSTAs switches from strong to weak in the MME as well. However, in contrast to OBS (Fig. 2b), the MME does not show an apparent dipole structure over the South Atlantic post-2000 (Fig. 2f). This could be a model deficiency in the NMME and may explain the inconsistency between prediction skill declines in the MME and persistence, which will be discussed later.

Intrinsic reason responsible for the decline in Atlantic Niño prediction skill

One might intuitively think that the changed ENSO-ATL3 relationship would be responsible for the skill decline. However, no significant correlation is found between the ATL3 prediction skills and ENSO-ATL3 correlations (Supplementary Fig. 3). As previously discussed, the SAOD may also play a role in the prediction skill decline. Therefore, we contrast the Atlantic Niño prediction skills against ENSO-SAOD correlations in Fig. 3. It turns out that the declined prediction skills correlate significantly with the ENSO-SAOD correlations, with less negative ENSO-SAOD relationships leading to lower Atlantic Niño prediction skills. The correlation between Atlantic Niño prediction skills and ENSO-SAOD correlations is −0.64 at the 0-month lead and −0.84 at the 2-month and 3-month lead times, indicating a more important impact of the ENSO-SAOD relationship at these lead times.

Fig. 3: Relationship between the Atlantic Niño prediction skill and ENSO-SAOD correlation.
figure 3

Scatter plots of ATL3 prediction skills against ENSO-SAOD correlations over the two subperiods for predictions at a 0, b 1, c 2, d 3, e 4, f 5 lead months. Blue and red markers indicate prediction skills pre- and post-2000, respectively. Blue and red dashed lines indicate the observed ENSO-ATL3 correlations pre- and post-2000, respectively.

To demonstrate how the South Atlantic SSTA is associated with ENSO over the two subperiods, the observed SSTA is regressed onto the observed Niño3.4 index (Fig. 4). Pre-2000, the Pacific Niño leads to the enhanced easterlies over the western equatorial Atlantic, which can deepen thermocline and trigger the cooling over the eastern equatorial Atlantic53. The cooling over the east can form an inter-basin SSTA gradient with the warming over the tropical Pacific, favoring the coupling feedback and the development of Atlantic Niña events47. In addition, the Pacific Niño enhances the background southeast trade wind over the NEP (northeast pole of the SAOD) region and the prevailing northwesterly wind over the SWP region (Fig. 4a). The former enhances the cooling over the NEP region via a couple of processes such as subsidence, evaporation, and equatorward advection, while the latter leads to a warming over the SWP region by sending more warm and moist air from the tropics to the subtropics, which suppresses the evaporation25. As the Pacific Niño decays, the cooling over the ATL3 and NEP regions weakens (Fig. 4c, e), but the SSTAs and wind anomalies over the NEP and SWP regions can sustain for more than 6 months, reflecting the stronger negative ENSO-SAOD relationship pre-2000. Post-2000, the Pacific Niño leads to the enhanced easterlies over the western equatorial Atlantic as well (Fig. 4b). However, the easterlies over the western equatorial Atlantic do not lead to an obvious cooling over the east, which could be related to the northerly wind anomalies over the NEP regions, as the cross-equatorial winds can modulate upwelling patterns and cause anomalous SST54. The northerly wind anomalies over the NEP region are present for more than 6 months as the Pacific Niño decays (Fig. 4f), which can weaken the background southeast trade wind and curb the cooling over the region, explaining the weak ENSO-SAOD relationship post-2000. Note that there appears to be stronger SST anomalies in the south Atlantic post-2000, although the ENSO influence on the SAOD decreases, which is beyond the scope of this study but deserves further attention.

Fig. 4: Regressions of the Niño3.4 index against the South Atlantic SSTA over the two subperiods in observations.
figure 4

a, b simultaneous regressions of Niño3.4 index against the SSTA and 850hPa wind anomalies. c, d same as (a, b), but with the Niño3.4 index leading by 3 months; e, f same as (a, b), but with the Niño3.4 index leading by 6 months; Only regressions with significance level of p < 0.05 are plotted. Black boxes indicate the northeastern pole and southwestern pole of the South Atlantic Ocean Dipole. Red box indicates the ATL3 region.

Previous studies have shown that both the remote influence and local oceanic conditions are important for the development of Atlantic Niño13,53. The SSTA gradient between the equatorial Pacific and Atlantic favors the coupling feedbacks and thus the persistence of Atlantic Niño47, and may explain the higher Atlantic Niño prediction skill pre-2000. In short, the more negative ENSO-SAOD relationship could be the intrinsic reason to have the higher Atlantic Niño prediction skill pre-2000, and the less negative ENSO-SAOD relationship could lead to the lower Atlantic Niño prediction skill post-2000.

Model deficiency responsible for the decline in Atlantic Niño prediction skill

Although the ENSO-SAOD correlation can explain a large part of the Atlantic Niño prediction skill, model deficiency can give rise to the declined Atlantic Niño prediction skill as well. Figure 2f has indicated that the models may not well capture the enhanced ATL3-SWP relationship. Supplementary Fig. 4 further shows regressions of ATL3 time series onto the South Atlantic SSTA for OBS and the MME predictions at different lead months. Pre-2000, the MME can well capture the observed relationship between the Atlantic Niño and the SWP SSTA, and it changes little as lead time increases. Post-2000, however, MME can basically capture the observed relationship only at 0-month lead, but fails to capture the relationship as lead time increases. It indicates that the dynamical predictions of MME cannot well reflect the relationship between the ATL3 and SWP at the long leads post-2000.

To determine the extent to which the SWP contributes to the prediction skill of Atlantic Niño, Supplementary Fig. 5 shows results of statistical predictions based on linear regressions as described by Eq. (1). It shows that when the ATL3 index is the only predictor, the prediction skill is clearly higher pre-2000 than that post-2000. When the SWP index is used together with the ATL3 index as the dual predictors, the prediction skills over the two subperiods become comparable, suggesting the complementary role of SWP to the prediction skill decline post-2000. The prediction skill over the second subperiod even exceeds that over the first subperiod at 4- and 5-lead month. Additionally, the Indian Dipole Mode Index55, and North Tropical Atlantic Index9 can contribute to the prediction skill over the first subperiod, consistent with previous studies22,23,35, but no apparent increase in prediction skill is obtained post-2000. These results reveal that the SWP SSTA contributes to a substantial proportion of the Atlantic Niño predictability post-2000. Therefore, it is very likely that the misrepresentation of ATL3-SWP correlation may be responsible for the much more declined MME prediction skill.

As we try to show a relationship between the change of Atlantic Niño prediction skill and the change of ATL3-SWP correlation, it is found that models cannot well capture the observed ATL3-SWP correlation post-2000 (Supplementary Fig. 6). The negative ATL3-SWP correlation is weakened in most models, while we know that it is strengthened in OBS (Supplementary Figs. 4 and 6, and Fig. 2). This may explain why the Atlantic Niño prediction skill declines more significant than the persistence. Specifically, OBS shows that the weakened ENSO-SAOD relationship corresponds to the enhanced ATL3-SWP correlation post-2000. The former leads to part of the decline in prediction skill, but the latter can add additional prediction skill to the Atlantic Niño post-2000. As the NMME models fail to faithfully capture the ATL3-SWP correlation post-2000 (Supplementary Figs. 4 and 6), it makes sense that the NMME models show the significant prediction skill decline over the time period. That being said, model experiments should be conducted to verify this idea.

Conclusions

This study aimed to answer how the Atlantic Niño prediction skill would vary as a result of the weakening Atlantic Niño since 2000. By assessing the NMME hindcasts, this study revealed a significant decline in the Atlantic Niño prediction skill post-2000 that cannot be simply explained by changes in persistence. It was found that the significant decline in Atlantic Niño prediction skill in the NMME relates not only to predictability, but also to deficiencies in dynamical predictions. Further analysis indicated that the changing ENSO-SAOD relationship is the intrinsic reason responsible for the decreased Atlantic Niño persistence, and that the misrepresentation of ATL3-SWP correlation is responsible for the much more declined MME prediction skill. This is because the weakened ENSO-SAOD relationship corresponds to the enhanced ATL3-SWP correlation post-2000. The former leads to part of the decline in the prediction skill, but the latter can add additional prediction skill to the ATL3 SSTA post-2000. That the NMME models fail to faithfully capture the ATL3-SWP correlation post-2000 leads to the significant prediction skill decline.

Previous studies have reported what may lead to the climatic regime shift around 2000 and how the climatic regime shift could lead to changes in ENSO teleconnections. For example, the Atlantic multidecadal oscillation could modulate the Atlantic Niño variability50, and the relationship between ENSO and the SSTA over the NEP region may be associated with the Pacific decadal oscillation51. However, the primary driver of the climate regime shift is beyond the scope of this study as most necessary data are unavailable from the NMME models. Given the various model constructions and initialization schemes included in the NMME predictions, results of this study could suggest a common deficiency in current climate models and prediction systems. By resolving this, it is expected that the dramatic decline in Atlantic Niño prediction skill would be mitigated.

Methods

Model data

The NMME is an ensemble of predictions provided by a number of state-of-the-art climate models from various U.S. and Canadian modeling centers. The NMME is broadly evaluated, conveniently accessible, and widely considered among the most sophisticated predictions available32,33. All models covering the period from 1982 to 2018 were adopted in this study except NCAR CCSM3 and NCEP CFSv2, as both of them show a false increase in Atlantic Niño variability and a low prediction skill. The ensemble mean of each model was used to form a seven-member multi-model ensemble. Supplementary Table 1 provides a brief description of these models

Observational data

Observational data used in this study were the sea surface temperature data from HadISST42 and ERSST.v5 datasets43. The wind and sea level pressure data were from NCEP/DOE reanalysis products56. All data were interpolated onto a uniform grid of 1° × 1°.

Defined indices

The Atlantic Niño has the primary variability peak in JJA, but it can also have a secondary peak in the boreal winter, which is called Atlantic Niño II57. Therefore, the current study examines the prediction skill of SST anomalies in the equatorial Atlantic across all calendar months. The Niño3.4 and ATL3 indices are defined as the area-averaged SSTAs over the Niño3.4 (5 °S–5 °N, 120°–170 °W) and Atlantic 3 (3 °S–3 °N and 20 °W–0°)41 regions, respectively. The SAOD is defined as the difference between the area-averaged SSTA over its NEP (0°–15 °S, 10 °E–20 °W) and SWP (25°–40 °S, 10°–40 °W) according to previous studies24,25. To show changes of Atlantic Niño prediction skill, the studying time period was separated into two subperiods following the previous study36, i.e., 1982–2000 and 2000–2018. The climatology and linear trend were removed for each subperiod separately. The prediction skill was evaluated via the ACC. The persistence ACC was calculated as the autocorrelation of the observed ATL3 index as a function of time lag. The 0-month lead time refers to as one-month integration of prediction from the initial condition. The DMI is defined as the difference in SSTA between the tropical western Indian Ocean (50 °E–70 °E, 10 °S–10 °N) and the tropical south-eastern Indian Ocean (90 °E–110 °E, 10 °S–0)55. The NTA index is defined as the area-averaged SSTAs over the north tropical Atlantic region (80°W–20 °E, 0 °N–15 °N)9.

The statistical model

To explore the impacts of different factors on prediction skill of the Atlantic Niño, a statistical prediction based on linear regression and leave-one-out cross-validation was established following the previous study9. The predictors included climate indices such as the ATL3, Niño3.4, Indian Dipole Mode Index55, and North Tropical Atlantic Index9 at different leading months, whereas the predictand was the ATL3 index. The regression equation was as follows,

$$\,y=a+{\sum }_{i=0}^{n}{b}_{{li}}{x}_{{li}}, \,\,{l}=0,\,\ldots ,\,5$$
(1)

where \(y\) indicated the ATL3 index, \(x\) the index of predictors, \(l\) the leading months, \(i\) the \(i\)-th predictor, \(a\) the intercept, and \(b\) the regression coefficients. All data over the examined time period, excluding the observed value at the target month and the initial value at the start month, were used to establish the regression models. Statistical prediction was conducted across the two subperiods using different combinations of predictors.