Introduction

Tropical Pacific sea surface temperature (SST) changes modulate numerous aspects of global climate. The observed tropical Pacific SST trend since the mid-twentieth century reveals a relative warming in the west and cooling in the east, intensifying the zonal SST gradient1. The strengthening of the zonal SST gradient is tied to an enhanced Pacific Walker circulation, manifesting in the acceleration of the surface easterly trade winds through the so-called Bjerknes feedback2 which underpins the El Niño-Southern Oscillation (ENSO)3,4. ENSO is the most prominent year-to-year climate fluctuation in the tropical Pacific and exerts strong impacts on remote regions of the globe through atmospheric teleconnections5. ENSO also modulates the climate in vast regions across the globe, including over North America6 and South America5, which have experienced severe environmental changes in recent decades and exhibit significant precipitation trends as shown in this present study.

The Southwest United States suffered intense drought conditions since the start of the 21st century7,8. Winter-spring precipitation is the main water resource for the Southwest United States, and has been decreasing since 19809. Precipitation along the west coast of the United States is impacted by ENSO during winter, with La Niña years favouring below-average precipitation across southwestern North America10. This teleconnection occurs via poleward propagating Rossby waves11.

The Amazon rainforest plays a crucial role in the climate system. However, the Amazon has undergone severe deforestation which could influence the spatiotemporal pattern and amount of rainfall12,13. Rainfall in Amazonia is also sensitive to the tropical Pacific SST and it is supressed by warm ENSO events (El Niño). El Niño-like climate change has been shown to influence the annual mean rainfall over South America in state of the art climate models14. The Amazon is expected to face longer dry spells and a reduced rainfall in the dry season of the 21st century13,15. In contrast, the Amazon experienced a significant increase in wet season precipitation from the early 1990s to early 2010s, from a strengthening of the hydrological cycle16,17,18.

Several studies have found a discrepancy in long-term precipitation trends between observations and climate models19,20,21. Most models misrepresent the observed precipitation change over the historical period in the Amazon22 and North America23. Understanding the causes of this discrepancy is still lacking. In addition, it is well known that most state-of-the-art coupled climate models do not capture the pronounced observed strengthening of the trade winds and zonal SST gradient in the Pacific24,25. In contrast, they tend to show a weakening of the gradient resembling an El Niño-like pattern which continues into the future projections with increasing magnitude26, in stark contrast to the observed La Niña-like trend since the mid-twentieth century27. The source of this discrepancy has remained unclear28 but resolving it could lead to better regional climate simulations29.

In this study, we find that observations during 1979–2014 show statistically significant trends in precipitation over the Amazon and southwestern parts of North America. These trends are not captured by climate models during the same period. Since the Pacific trade wind trend is underestimated in the models, we ask whether this can explain the discrepancy in land precipitation. We test whether the simulated long-term precipitation trends can match their observed counterparts using the observed equatorial wind trends via the modelled interannual ENSO teleconnection. We show that the observed strengthening of the Pacific trade winds and modelled ENSO teleconnections can explain the majority of the discrepancy between modelled and observed precipitation trends. Reproducing the observed trends indicates that the discrepancy is linked to the inability of models to capture the observed tropical Pacific wind trends in the tropical Pacific Ocean.

Results

Discrepancies between observations and coupled models

The Eastern Pacific shows a cooling trend in observations (Fig. 1a), but the coupled models simulate a warming pattern in the Eastern Pacific (Fig. 1b). In Fig. 1a, a strengthening of the zonal SST gradient (i.e., La-Niña like pattern) leads to an increasing sea level pressure gradient. This means that the Walker circulation, coupled with SST gradient via the Bjerknes feedback, also strengthened30. Previous studies attributed this strengthening to internal variability, such as the Interdecadal Pacific Oscillation (IPO)31 or the Atlantic Multidecadal Oscillation (AMO)32; although some studies also point to a role for forced changes increasing the trade winds due to external factors, such as aerosol forcing33.

Fig. 1: Historical simulations do not reproduce the observed change in Pacific zonal SST gradient.
figure 1

Trends of the SST (shading, °C) and sea level pressure (SLP, contours, hPa) averaged over December to February (DJF) and the period 1979–2014 in the observations (a) and the CMIP6 multimodel mean historical simulations (b). Trends for SST and SLP are multiplied by 36 years to indicate the total linear change from 1979–2014. Stippling indicates SST trends are significant above the 90% level. The blue box represents the Niño4 region (5°N-5°S, 160°E-150°W).

Full size image

In contrast to the observations, however, coupled model simulations produce a weaker SST gradient pattern in the tropical Pacific and the SLP gradient is also decreased (Fig. 1b). The weakened SST gradient is consistent with an El-Niño like pattern and weaker trade winds.

We now compare the Niño4 850 hPa zonal wind trends from the observations with the multimodel mean from the Coupled Model Intercomparison Project Phase 6 (CMIP6, See Method). Observations show that Pacific trade winds have strengthened over 1979–2014 (Fig. 2a, b) but this trend is not captured by coupled climate models. The trend is similar but has less strengthening when the period is extended to 2022 in observations (-1.30\(\pm\)1.94 m·s-1, p = 0.28, Fig. 2a). The zonal wind trend over the Niño4 area in observations is -2.70\(\pm\)2.65 m·s-1 (p = 0.10, Fig. 2a, b), while the historical multimodel mean is 0.41\(\pm\)0.36 m·s-1 (p = 0.08, Fig. 2a, b). In other words, the coupled climate models simulate a weakened Walker circulation and an El-Niño like pattern. In contrast to this, the trend in the Atmospheric Model Intercomparison Project (AMIP) simulations, which use prescribed, observed SSTs, shows strengthened trade winds (-1.99\(\pm\)0.94 m·s-1, p < 0.01, Fig. 2a, b). Because of strong coupling between the atmosphere and ocean in the tropical Pacific, we hypothesise that the discrepancies between the Pacific trade wind trends in observation and coupled models is connected to the difference in SST patterns.

Fig. 2: Historical simulations fail to capture the Pacific trade wind intensification.
figure 2

Time series (a) and trends (b) of Niño4 850 hPa zonal wind (m·s-1) averaged over December to February (DJF) and the period of 1979–2014 for the observations (black), multimodel mean historical simulations (blue) and multimodel mean AMIP simulations (grey). Solid lines indicate the linear trend for 1979–2014 and dashed lines show the trend for 1979–2022 in (a). The trend for zonal wind is multiplied by 36 years to correspond to the total linear change from 1979–2014 in (b). The average of the ERA5 and JRA3Q observational analyses (denoted OBS) is represented in bold. The filled circle represents the average and the error bar indicates 90% confidence interval.

Full size image

ENSO adjustment of the precipitation trend

The distribution of precipitation trends over land is shown in Fig. 3. Observed regional rainfall trends show statistically significant drying over southwestern North America and wetting over the Amazon (Fig. 3a). However, the coupled models do not capture these observed rainfall trends over 1979–2014 (Fig. 3b).

Fig. 3: Adjusted precipitation trends better match observations.
figure 3

Trends of precipitation (%) averaged over December to February (DJF) and the period of 1979–2014 for the observations (a), the multimodel mean historical simulations (b), adjusted simulations (c) and multimodel mean AMIP simulations (d). The trend for precipitation is multiplied by 36 years to correspond to the total linear change from 1979–2014. The blue boxes represent the selected regions: southwestern North America (10°N–40°N, 115°W–95°W) and Amazon (10°S-10°N, 80°W-50°W). Stippling indicates the land precipitation trends are significant above the 90% level.

Full size image

We therefore carry out an adjustment of the coupled model simulations, to test whether the Pacific trade wind trends and associated ENSO teleconnections can explain the discrepancy in rainfall trend (see Methods). We used the modelled linear relationship in terms of interannual variability between the global precipitation and zonal wind over the Niño4 area and the observed zonal wind trend to create an adjustment term which was added to the modelled precipitation in the historical simulation.

We find that southwestern North America shows a clear and significant drying trend (trend: -56.04\(\pm\)34.87%, p = 0.01) in the observations (Fig. 4a). However, there is no significant trend in historical simulations (trend: -1.74\(\pm\)4.77%, p = 0.56, Fig. 4a). After the linear adjustment, the simulated precipitation trend is strengthened (trend: -19.13\(\pm\)4.79%, p < 0.01, Fig. 4a), and it is consistent with both the observed and simulated AMIP precipitation trends (trend: -17.33\(\pm\)6.26%, p < 0.01, Fig. 4a).

Fig. 4: Adjusted precipitation trends during 1979–2014.
figure 4

Trends of precipitation (%) averaged in December to February (DJF) over the period of 1979–2014 for the observations (black), the multimodel mean historical simulations (blue), the adjustments (red) and the multimodel mean AMIP simulations (grey) over southwestern North America (a) and Amazon (b), respectively. The trend for precipitation is multiplied by 36 years to correspond to the total linear change from 1979–2014. The average of the four products is represented in bold (denoted OBS). The filled circle represents the average and the error bar indicates 90% confidence interval.

Full size image

The Amazon shows increasing rainfall over 1979–2014 in the observations (trend: 11.46\(\pm\)11.06%, p = 0.10, Fig. 4b) but the historical model simulations again fail to reproduce any significant trend (trend: 3.74\(\pm\)5.03%, p = 0.24, Fig. 4b). The precipitation in the Amazon also has a well-established relationship with ENSO, with La-Niña events enhancing the convection in the Amazon area13,34. The simulated historical precipitation trend over Amazon after the ENSO adjustment is 16.56 \(\pm\)5.03%, p < 0.01 (Fig. 4b) which becomes again similar to the AMIP simulations (trend: 14.38\(\pm\)3.18%, p < 0.01, Fig. 4b). It is also now significant and consistent with observed changes. In summary, after the linear adjustment, the historical CMIP6 simulations capture the observed rainfall trends, especially in southwestern North America and the Amazon (Fig. 3c) and are comparable with the AMIP simulations (Figs. 3d and 4).

The strengthening of the equatorial zonal wind during La Niña strengthens the Pacific Walker Circulation, which means there is an anomalous ascending branch from the western Pacific, descending on the Central/Eastern Pacific. There is also ascent over northern South America (Supplementary Fig. S1). This favours increasing rainfall over Amazon area18. The changes in upper tropospheric tropical divergence also result in a poleward and eastward stationary Rossby wave which usually results in a poleward shift of the mid-latitude jet stream (Supplementary Fig. S2), affecting momentum transport by transient eddies and anomalous eddy-driven descent that suppresses precipitation in southern North America35.

Discussion

For the multidecadal period of 1979–2014, eastern Pacific SST experienced a cooling trend, leading to an increased zonal SST gradient and stronger trade winds24,36. In this study, we find that coupled ocean-atmosphere models fail to reproduce the most prominent and statistically significant observed precipitation trends during boreal winter in southwestern North and northern South Americas. This occurs in conjunction with much weaker Pacific trade wind changes in models than observations24,25.

A linear adjustment based on modelled ENSO teleconnections and observed tropical Pacific wind trend leads to model reproduction of the rainfall trends over the Americas during 1979–201411,37. We note however, that even after the adjustment, the modelled trends tend to remain weaker than observed. This may be related to weak extratropical teleconnections in climate models and the so called “signal-to-noise paradox”38,39 in which models typically underestimate climate signals, possibly due to their weaker response to boundary and external forcings. The underestimated teleconnections in climate models may explain the weaker ensemble-mean trends, even after correcting for the ENSO trend40.

There is an interplay with local internal variability in the discrepancy between models and observations. Outside the tropical regions of the Americas, there are some regions in the models where rainfall trends become more consistent with the observations; for instance, in northern North America (Fig. 3a, c). There are also regions where the multi-model mean trends become statistically significant, but inconsistent with observations, e.g., over high-latitude eastern North America and Asia (Fig. 3c), potentially indicating the additional role of internal variability in observations such as from the North Atlantic Oscillation41. The presence of the La Niña-like trends in the Pacific is associated with a positive Southern Annular Mode (SAM). The strengthening of the southern hemisphere mid-latitude jet and poleward movement of the storm track corresponded to an increasing pressure gradient between mid-latitudes and the Antarctic42 (Supplementary Fig. S3) and enhanced precipitation in southeastern Australia43 (Fig. 3a, c).

If we extend the end of the time period to 2022 (medium radiative forcing scenario used for historical models from 2015–2022), the observed trade wind trend still has the same sign but becomes less significant (-1.30\(\pm\)1.94 m·s-1, p = 0.28, Fig. 2a, Supplementary Fig. S4). The observed trends in the SST zonal gradient are also weaker over this extended period21. This is likely due to the 2015/2016 extreme El-Niño event44; when 2015 and 2016 are removed from the time series, the trade wind shows a more strengthening trend (-1.91\(\pm\)1.98 m·s-1, p = 0.12). Nevertheless, the linear adjustment also works for this extended period (Supplementary Figs. S5, S6).

Finally, we note that our adjustment method may also be useful in constructing ‘storylines’ of future climate change. As we do not know the cause of the trade wind strengthening, there remain two plausible storylines: one is the persistent strong trade wind and the other is weak trade wind in the model projections which will drive different rainfall changes. The Pacific trade winds are coupled with the Pacific zonal SST gradient which not only plays an important role in tropical Pacific, but also climate over remote regions. This study sheds light on the effect of the tropical Pacific trends on North America and the Amazon climates which are expected to change in the future. It indicates that simulating rainfall trends over the Americas and projecting their future evolution requires successful simulation of long-term tropical Pacific ocean-atmosphere trends, as well as ENSO and its teleconnections. Climate change and multidecadal variability in the trade winds and their effects on regional rainfall are therefore crucial for future trends and water availability over North and South America.

Methods

Data

The monthly 850hPa zonal wind, sea level pressure and precipitation are taken from ERA545 (0.25° × 0.25°) and JRA3Q46 (1.25° × 1.25°) reanalyses over the period of 1979–2022. We averaged them over December-January-February (DJF). Both reanalyses are interpolated onto a 2.5° × 2.5° latitude-longitude grid and the average of the two is used for the analysis. The resolution of the dataset does not affect the results.

The observed precipitation datasets are from the Global Precipitation Climatology Project (GPCPv2.347) and the Climate Prediction Center Merged Analysis of Precipitation (CMAP48). Both are based on the gauge stations and satellites on a 2.5° global grid.

The observed sea surface temperature datasets (HadISST) are from the Hadley Centre49.

We compare the observed changes to the models from CMIP6. We pick all 14 models (Table S1) containing both CMIP historical simulations and AMIP simulations and DAMIP simulations. The historical simulations are fully coupled runs, forced with observed radiative forcings over the period 1850–2014. The AMIP simulations are the atmosphere only runs forced with observed radiative forcings, sea surface temperatures, and sea ice concentrations over the period 1979–2014. One realization from each model is chosen for both types of simulations. All data are interpolated to a common 2.5° × 2.5° latitude-longitude grid before performing any analyses.

Definition of precipitation

In this study, we define the relative percentage change of precipitation (%) as it compared to 1981–2010 climatology over the period 1979–2014 December-January-February (DJF) mean over the land.

Adjustment of precipitation trends based on ENSO teleconnection

The linear adjustment of precipitation trends based on the observed strengthening trade winds over Niño4 area (5°N-5°S, 160°E-150°W) and modelled ENSO teleconnections is defined as:

$$\frac{\partial {Pr }_{{ENSO}{{\_}}{adj}}}{\partial t}={r}_{Pr ,U}\cdot \frac{\partial {U}_{{obs}}}{\partial t}$$
(1)
$${r}_{Pr ,U}=\,\frac{\partial {Pr }_{{hist}}}{\partial {U}_{{hist}}}$$
(2)

where \(\frac{\partial {\Pr }_{{\rm{ENSO\_adj}}}}{\partial {\rm{t}}}\) is the adjustment precipitation trend based on modelled ENSO teleconnection. \({{\rm{r}}}_{\Pr ,{\rm{U}}}\) is the regression coefficient of the precipitation regressed onto the Niño4 850 hPa zonal wind. \({\Pr }_{{\rm{hist}}}\) and \({{\rm{U}}}_{{\rm{hist}}}\) are the interannual variabilities of precipitation and 850 hPa zonal wind over Niño4 area for every CMIP6 model in historical runs, respectively. The interannual variability is calculated after removing trends computed using least-squares regression. \(\frac{\partial {{\rm{U}}}_{{\rm{obs}}}}{\partial {\rm{t}}}\) is the Niño4 850 hPa zonal wind trend for the observed. The trend is calculated using the least-squares regression.

Then we add this adjustment term to the original CMIP6 historical multi-model mean trend. The adjusted linear trend of the precipitation (\(\frac{\partial {\Pr }_{{\rm{hist\_adj}}}}{\partial {\rm{t}}}\)) is given by:

$$\frac{\partial {\Pr}_{{\rm{{hist}\_adj}}}}{\partial {\rm{t}}}=\frac{\partial {\Pr}_{{\rm{hist}}}}{\partial {\rm{t}}}+\frac{\partial {\Pr}_{{{{\rm{{ENSO}}}_{\_}{{\rm{adj}}}}}}}{\partial {\rm{t}}}$$
(3)

where \(\frac{\partial {\Pr }_{{\rm{hist}}}}{\partial {\rm{t}}}\) is the multi-model mean trend of precipitation for the models with historical runs. The trend is using the least-squares regression coefficient.

Significance test

In this study, the trends are significantly different from zero when they exceed 1.65 standard errors (90% confidence level). The standard error (\({\rm{S}}{{\rm{E}}}_{{\rm{obs}}}\)) for the observations is defined as:

$$S{E}_{{obs}}=\,\sqrt{\frac{1}{n-2}\cdot \frac{{\sum \left(y-\hat{y}\right)}^{2}}{{\sum \left(x-\bar{x}\right)}^{2}}}$$
(4)

where y is the real values, \(\hat{y}\) is the estimated values, x is the year, \(\bar{x}\) is the mean year, n is the size of the observations.

The standard error (\({{\rm{SE}}}_{{\rm{models}}}\)) for the multimodel is defined as:

$${{\rm{SE}}}_{{\rm{models}}}=\,\frac{{\rm{\sigma }}}{\sqrt{{\rm{m}}}}$$
(5)

where \(\sigma\) is the standard deviation of trends among all the models and m is the total number of the models.