Table. 1. The detailed design of experiments. For WACCM4, the experiments are based on F_1955-2005_WACCM_CN compset. The experiments of CAM5 are based on F_AMIP_CAM5 compset.
Methods
In our study, the region (30°S-30°N, 30°E-180°) was used to represent IPWP. For observational data and reanalysis data, anomalies are calculated by subtracting the monthly mean. Linear trends were calculated by linear regression analysis in the least square method. Linear regression analysis was also used to remove the signal of IPWP. In addition, Ensemble Empirical Mode Decomposition (EEMD) was carried out to further study the nonlinear trends of SWV and tropical tropopause temperature. It is an improvement of EMD, which is an adaptive time-space analysis method that suits the non-stationary and non-linear time series (Salisbury and Wimbush 2002; Wu and Shen 2016). It can extract the Intrinsic Mode Functions (IMF) at different frequencies and obtain a residual series that can represent the non-linear long-term trend. To depict the relationship between tropical SWV entry and SST of IPWP, the correlation coefficients were calculated. Because the SWV entry and tropical tropopause temperature has large interannual variabilities, they have been eliminated to some extent by the five-year running average on raw timeseries before calculating the correlation coefficient. The statistical significance is calculated by Mann-Kendall non-parametric test for trend. The statistical significance of correlation is using the Student’s t-test and the degrees of freedom are estimated as follows (Bretherton et al. 1999):
\(N_{\text{eff}}\)≈N\(\frac{1-r_{1}r_{2}}{1+r_{1}r_{2}}\) (1)

Results

Observed reduction in SWV entry for the period 1984-2020
Fig. 1a shows the time series of SWV entry anomalies in the tropics from 1984 to 2020. A few facts are first reviewed to make sure the datasets used can capture recent SWV changes. The reportedly two severe drops which were registered around 2000 (Randel et al. 2006; Dhomse et al. 2008; Wang et al. 2017) and 2011(Urban et al. 2014) respectively are also shown here. The interannual cycle is also well captured in the two datasets with an approximate 2-year cycle mainly controlled by QBO (Fueglistaler et al. 2005; Diallo et al. 2018; Diallo et al. 2022). Now, we assess the long-term trend in the lower SWV for the period 1984-2020. Both SWOOSH and ERA5 datasets show a significantly decreasing trend in SWV entry anomalies from 1984 to 2020. The linear decreasing rate is 0.106 ± 0.021 ppmv per decade in SWOOSH and 0.037 ± 0.022 ppmv per decade in ERA5. The decades-long decreasing trend is statistically significant at 99% confidence level. This decline is also manifested in their mean values. For SWOOSH, the SWV amount was about 3.62 ppmv in the first five years of 1984-2020 and it decreased to 3.43 ppmv in the last five years, with a drop of ~0.2 ppmv. A similar decadal drop with a smaller amplitude can be also seen in ERA5. But this drop amplitude is smaller than that estimated by linear tread. The reason for this disagreement will be explained later.