The Hadley Centre SST data set version 1 (HadISST1) is employed for the SST analysis (Rayner et al., 2003). Subsurface ocean temperature data are taken from the Simple Ocean Data Assimilation (SODA) version 2.2.4 (Carton and Giese 2008). The precipitation field is extracted from the NOAA Climate Prediction Center (CPC) Merged Analysis of Precipitation (CMAP) dataset (Xie and Arkin 1997). Monthly atmospheric circulation data including winds, SLP, specific humidity, velocity potential, and air temperature employed in this study are procured from the National Centers for Environmental Prediction-National Center for Atmospheric Research (NCEP-NCAR) reanalysis version I (Kalnay et al., 1996). All datasets range from the years 1979–2019. We have also removed the long-term linear trend and seasonal cycle from all variables analyzed in this study.

To evaluate whether the effect of winter NAD on the following AWP can be reproduced in the current coupled model, the monthly outputs of 25 coupled atmosphere-ocean models participated in the Coupled Model Intercomparison Project, Phase 6 (CMIP6) from preindustrial control (pi-Control) experiments are used in this paper (Eyring et al., 2016). The variables analyze from the CMIP6 models included: SST, SLP, horizontal winds, precipitation, air temperature, and specific humidity. The SST variable output by the CMIP6 is converted to a horizontal resolution of 1° × 1°. In contrast, the other variables’ production is converted to a horizontal resolution of 2.5° × 2.5°, which concords with the resolution of the observed data and aids the calculation of the overall average. To circumvent the initial adjustment period caused by the model’s spin-up process, only the simulations run from the past century have been analyzed in this paper.

As per Ding et al. (2017), the NAD index (NADI) used here is defined as the difference of normalized (i.e., subtracting the mean and then dividing by the standard deviation) SLP anomalies between the southern (90°–60°W, 10°–30°N) and northern (75°–45°W, 55°–70°N) poles (southern minus northern pole) after linearly excluding the effect of the ENSO. The NAO index (NAOI) used in this study is defined as the difference in normalized SLP anomalies between the southern (80°W–30°E, 35°N) and northern (80°W–30°E, 65°N) poles over the North Atlantic sector (Li and Wang 2003). According to Zhou et al. (2019), the maximum seasonal variance of the NAD and NAO ranges from December to March (DJFM). Therefore, our research is based on the 4-month average DJFM NADI and NAOI. Figure 1B depicts the interannual time series of the DJFM-averaged NADI (blue line) and NAOI (red line) for 1979–2018, where their correlation was calculated at 0.49.

Following Wang and Enfield (2003), two types of AWP indices are employed in this study. One is the AWP area index (AWPAI), which is defined as the area within the 28.5°C isotherms at the AWP region’s sea surface. The other is the index of the depth enclosed by the 28.5°C isotherms in the AWP region (referred to as the AWP depth index, AWPDI).

The coupled model used in this study is the Flexible Global Ocean Atmosphere–Land System model grid-point version 2 (FGOALS-g2) developed at State Key Laboratory of Numerical Modeling for Atmospheric Sciences and Geophysical Fluid Dynamics (LASG), Institute of Atmospheric Physics (IAP), Chinese Academy of Sciences (Li et al., 2003). The FGOALS-g2 consist of the Grid-point Atmospheric Model of LASG/IAP version 2 (GAMIL2), the LASG/IAP Climate System Ocean Model version 2 (LICOM2), the Community Land Model version 3 (CLM3), and the improved version of Community Ice CodE version 4 by LASG (CICE4-LASG) connected with the coupler (CPL6).

To further interpret the NAD’s effect on the following AWP development, the control and assimilation experiments are conducted. The difference between the above two experiments is to assimilate the NAD-related winds (U, V) and air temperature (T) anomalies at 1000 hPa, 925 hPa, and 850 hPa in the western tropical North Atlantic and northeastern North America, which arises from regressions of the DJFM U, V and T anomalies onto the concurrent NADI (Figure 2). The control experiment is a historical run driven by external forcing observed from 1979 to 2018. Another experiment assimilated the NAD-related U, V, and T anomalies imposed on the climatological U, V, and T. The initialization is cycled at 40 years from 1979 to 2018 with a 1-month assimilation windows. The past 25 years’ output is applied to the composite analysis to minimize the possible internal impact.

Several statistical methods, including composite analysis, correlation analysis, and linear regression, are used in this study. The statistical significance is based on a two-tailed Student’s t-test, where the effective number ( N * ) of degrees of freedom is calculated as follows by Bretherton et al. (1999): N * ≈ N 1 − R x R y 1 + R x R y , (1) where N is the sample size, and R x 、 R y represents the lag-one autocorrelations of two-time series x and y , respectively.

To investigate the probable link between the NAD and AWP, the lead-lag correlation of the winter NADI (red line) with 3-month-averaged AWPAI and AWPDI for 1979–2018 is illustrated in Figure 3. The peak correlation between the NADI and AWPAI (R=−0.62; significant at the 99% confidence level) occurs during the months of May to July (MJJ), lagging the NAD’s peak during winter by about 5 months. For the AWP depth index, similar significant values (at the 99% confidence level) begin to appear during April–June (AMJ), peak (R=−0.60) during MJJ, and persist until the following fall. Moreover, we show the correlations of the AWPAI and AWPDI during MJJ with SLP anomalies in the previous winter (DJFM-averaged; Figure 4). It is observed that both the AWPAI and AWPDI are closely linked to the previous winter NAD-like dipole of SLP anomalies over the North Atlantic (Figure 1A).

We then perform a comprehensive analysis comparing the seasonal evolutions of the AWP area between positive and negative NAD events (Figure 5). It is defined as a positive (negative) NAD event when the winter NAD index is greater (less) than one positive (negative) standard deviation (Table 1). When the winter NAD is in its strong negative phase (blue line), the AWP develops in the Caribbean region during the months of April to June (AMJ) (Figure 5D). Thus, the eastern edge of the AWP extends eastward into the TNA region in the following MJJ and June–August (JJA) (Figures 5E, F), resulting in a twofold to threefold increase in area over the climatic state (Figures 5B, C). Contrarily, the positive phase of the winter NAD tends to inhibit the subsequent AWP development (see also Figures 5D–F; red line). The above results analyze changes in the area of the AWP, and a similar composite analysis regarding changes in the depth of the AWP is shown in Figure 6. A similar conclusion can be drawn that the strong positive (negative) phase of the winter NAD tends to inhibit (favor) the development of the AWP depth in subsequent months.

A similar lead-lag correlation of the winter NAOI (yellow line) with 3-month-averaged AWPAI and AWPDI is also investigated in Figure 3B. In contrast to the NADI, the NAOI shows weaker correlations with the AWPAI and AWPDI during MJJ. The previous winter NAO has little effect on the following area (Figures 5G–I) and depth (not shown) of the AWP. These results indicate that the NAD may play a more pivotal role in influencing the AWP due to its efficiency at forcing the TNA SST variability in contrast with the NAO. This suggests that the AWP variability may be a lagged response to NAD atmospheric forcing, rather than a lagged response to NAO atmospheric forcing. Although the above observational analysis suggests a close connection between the NAD and AWP, it is limited by the relatively small number of the NAD events. Further investigations employing more extended datasets from the CMIP6 models are necessary to support the results of the observed data.

To evaluate whether the effect of winter NAD on subsequent AWP can be reproduced in the current coupled model, we analyzed the monthly outputs of 25 coupled atmosphere-ocean models participated in CMIP6 (Table 2).

The models’ abilities in simulating the NAD are first assessed. Figure 9 illustrates the spatial patterns of the correlation coefficient between the DJFM-averaged NADI and concurrent SLP anomalies derived from observations and 25 CMIP6 models’ simulations. All the models can reproduce a distinct SLP dipole comprising of a low over northeastern North America and a high over the western tropical North Atlantic, which is the signature feature of the NAD. For the location of the NAD center, there is a northward displacement of the negative center of the NAD in most of the models when compared to the observations. Furthermore, most models can give an accurate representation of the NAD’s positive center, although several models (EC-Earth3, MIROC6, MPI-ESM-1-2-HAM, MPI-ESM1-2-HR, and MRI-ESM2-0) simulate the NAD’s positive centers extending too far to the east. Henceforth, a Taylor diagram (Taylor, 2001) is employed to quantitatively measure the difference between model simulations and observations and make inherent comparisons among different models (Figure 10). Two-thirds of the pattern correlation coefficients between the modeled and observed NAD pattern are evaluated at more than 0.70, indicating that the CMIP6 models generally reproduce well the NAD pattern. More than half of the ratios of the standard deviations between modeled and observed NAD patterns are valued to be less than 1.0, which indicates that most CMIP6 models underestimate the NAD variability. In general, HadGEM3-GC31-LL and FOGOALS-f3-L perform best at simulating the interannual NAD pattern.

The relationship between the NAD and AWP and the associated physical processes in model simulations are subsequently analyzed. Figure 11 shows simulated lead-lag correlations of the DJFM-averaged NADI with 3-month-averaged AWPAI. Most models (16 out of 25) reproduce a significant negative correlation occurring in spring or early summer. Also, as shown in Figure 12, more than half of the models simulate a negative correlation significant at the 90% confidence level between the previous winter NAD and the following MJJ AWP.

The aforementioned results show that a significant negative correlation between the boreal winter NAD and the following spring or early summer AWP variability can be reproduced in most CMIP6 models. Figure 13 shows correlation maps of the 3-month averaged SST and surface wind anomalies with the NADI derived from the multi-model ensemble means (MME) of 25 CMIP6 models. A tripole-like SST pattern in the North Atlantic is discerned during DJF. Negative SST anomalies in the NTA region reach maximum intensity during MAM. The NAD-related TNA cooling persists until MJJ, extending from the Caribbean Sea to the Gulf of Mexico. However, it is noted that the center of negative SST anomalies in the TNA region in models is found to be further east than that in observations, and southwesterly anomalies are much weaker afterMAM.

Numerous studies have yielded evidence of the importance of the AWP in modulating the climate system (Donders et al., 2011; Kim et al., 2020). For instance, the AWP may act as a source of moisture for precipitation in Central and North America (Donders et al., 2011). During the development and persistence of the large summer AWP, the North Atlantic brings record temperature rise to Florida (Drumond et al., 2011; Liu et al., 2015). Considering the close relationship between the NAD and the AWP, we speculate that the winter NAD may influence the North Atlantic region’s climate through its impact on the AWP in the following months.

To test this hypothesis, the correlations of DJFM-averaged NADI with precipitation, 700 hPa relative humidity, and surface air temperature anomalies in the following summer (the JJA season) have been calculated. As depicted in Figure 14A, the regions with significant negative precipitation anomalies associated with previous winter NAD cover the southern Gulf of Mexico, Caribbean Sea, and TNA during summer, consistent with recent studies (Zhou et al., 2019). The relative humidity field at 700 hPa exhibits zonal distribution characteristics (Figure 14B). Significant positive correlations are found in southwest North America, the Gulf of Mexico, and the southeastern Pacific Ocean. Meanwhile, significant negative correlations are found in the region between the north of the equator and south of 20°N.

By linearly removing the following MJJ AWPI, the regions of significant negative precipitation anomalies associated with previous winter NAD are significantly smaller and restricted to within the Gulf of Mexico (Figure 14C). Similar results are observed for the relative humidity field (Figure 14D). Based on the above analysis, the NAD-like atmospheric signal in winter may be stored in the following AWP, thus significantly affecting the TNA precipitation in the following summer. Therefore, the AWP may act as a bridge connecting the winter NAD to the subsequent summer precipitation in the TNA region.

Given that the winter NAD has a major influence on the following summer precipitation over the western TNA (85°W–50°W, 10°–25°N) through its impact on the following AWP, the winter NAD and the following MJJ AWP can be exploited to forecast summer precipitation over this region. Therefore, a simple empirical prediction model was established to predict the summer [JJA (+1)] precipitation (referred to as the precipitation index, PI) over western TNA by using the DJFM (0) NADI and the MJJ (+1) AWPAI. The model is based on a linear regression method, and the DJFM (0) NADI and MJJ (+1) AWPAI are chosen as variables. The model is as follows: P I t + 1 = α × N A D I t + β × A W P A I t , (2) where the parameters α and β are calculated by a least-squares fit to the DJFM (0) NADI and MJJ (+1) AWPAI, respectively. The time series of t + 1 , N A D I t , and A W P A I t have been standardized.

To evaluate the prediction performance of this model, the fourfold cross-validation is conducted. Figure 15A shows correlations between the observed and hindcast JJA (+1) precipitation anomalies obtained from the empirical prediction model. There are significant positive correlations in the western TNA region. The correlations between the observed values and fourfold cross-validated hindcasts are slightly lower than those shown in Figure 15A. However, they remain statistically significant in key areas of the study. The time series of observed and cross-validated hindcasted JJA (+1) PI are displayed in Figure 15C. The high correlation (R=0.72) between the observations and the cross-validated hindcasts indicates the good practicability of our model. Thus, it is practical to utilize this model to predict the summer precipitation over the western TNA region.