Tightening of tropical ascent and high clouds key to precipitation change in a warmer climate

The change of global-mean precipitation under global warming and interannual variability is predominantly controlled by the change of atmospheric longwave radiative cooling. Here we show that tightening of the ascending branch of the Hadley Circulation coupled with a decrease in tropical high cloud fraction is key in modulating precipitation response to surface warming. The magnitude of high cloud shrinkage is a primary contributor to the intermodel spread in the changes of tropical-mean outgoing longwave radiation (OLR) and global-mean precipitation per unit surface warming (dP/dTs) for both interannual variability and global warming. Compared to observations, most Coupled Model Inter-comparison Project Phase 5 models underestimate the rates of interannual tropical-mean dOLR/dTs and global-mean dP/dTs, consistent with the muted tropical high cloud shrinkage. We find that the five models that agree with the observation-based interannual dP/dTs all predict dP/dTs under global warming higher than the ensemble mean dP/dTs from the ∼20 models analysed in this study.


GISS
. Relationships between the inter-model spreads in tropical high cloud fraction sensitivity and longwave radiative sensitivities. Colored circles mark the individual models' interannual (blue) and centennial (red) sensitivities to surface temperature for tropical-mean high cloud fraction (CF), longwave radiative cooling (LWC), all-sky OLR, clearsky OLR (OLR clr ), and longwave cloud radiative effect (CRE lw ). Multi-model-means are shown in black "+". The left y-axis is for high cloud fraction sensitivity and the right y-axis is for longwave radiative sensitivities. The across-model correlations between dCF/dT s and the radiative sensitivities are shown above the x-axis for interannual (blue) and centennial (red) rates.

Supplementary Discussion
The dominance of longwave radiative control on precipitation sensitivity. The dominance of longwave radiation in controlling the global-mean precipitation change is remarkable. We find that the inter-model spread in global-mean dP/dT s is highly correlated with the model differences in global-mean dLWC/dT s (R = 0.94 for interannual and 0.84 for centennial rates, Supplementary  Figure 10). This indicates that high cloud fraction sensitivity is not a dominant factor in driving the model discrepancy in ECS.

Contribution of upper tropospheric water vapor to the longwave radiative feedback biases.
We have analyzed the simulated upper tropospheric water vapor sensitivity to surface warming in the models and find all models capture the increase of upper tropospheric water vapor path (UTWVP) with surface temperature, at the rates between 9.2% K −1 and 15.6% K −1 , approximately consistent with the Clausius-Clapeyron relation. The multi-model-mean is 11.9% K −1 , higher than that derived from the combined AIRS and MLS water vapor observations, 9.6% ± 1.5% K −1 (Supplementary Figure 12), although within the AIRS and MLS data uncertainty of ~25%. All models (except CNRM_cm5) produce greater upper tropospheric moistening with surface warming than the observations. The CNRM_cm5 model has the largest decrease of high cloud fraction with surface warming (Figure 2a) and also the weakest upper tropospheric moistening (Supplementary Figure 12b). The moist biases in the rest of the models, consistent with relatively weak decreases of high cloud fraction (the correlation between the spreads in dUTWVP/dT s and dCF/dT s for the 13 models is 0.57), would contribute to the low biases in the magnitudes of dOLR/dT s . However, based on the calculations using the radiative kernels Error! Reference source not found.,Error! Reference source not found.
, we find that the ensemble-mean moist bias of 2% K −1 in the upper troposphere would only contribute to a small fraction of the low bias in dOLR/dT s , on the order of 0.05 W/m 2 K −1 .
In addition, the inter-model spread in dUTWVP/dT s has rather weak correlations with the spreads in dOLR/dT s , dOLR clr /dT s and dCRE lw /dT s (R = −0.31, −0.23, −0.26, respectively) on the interannual time scale. Therefore, we conclude that the misrepresentation of dCF/dT s is a dominant source for the model spread in dOLR/dT s across the CMIP5 models and the moist bias in the upper troposphere associated with the muted high cloud shrinkage contributes only slightly to the low bias in the magnitude of dOLR/dT s . Determination of the observation-based interannual precipitation sensitivity. Two sets of observations are used to determine the best estimate of the interannual precipitation sensitivity.
First, we obtain an OLR-constrained L v dP/dT s based on the approximately linear relationship between the model simulated interannual tropical-mean dOLR/dT s and global-mean L v dP/dT s and the CERES observed tropical-mean dOLR/dT s , i.e., the OLR-constrained L v dP/dT s = A⋅ (dOLR/dT s ) CERES + B + ε, where A and B are the slope and intercept for the least squares regression across the models, respectively, and ε is the linear fitting residual. The statistical distributions of the slope and intercept for the regression between the modeled dOLR/dT s and L v dP/dT s are determined by 10000 bootstrap iterations with replacement. The resulting mean slope and intercept are 0.33 and 0.56, respectively. With the CERES dOLR/dT s at 3.8±0.4 W m −2 K −1 , the mean value of the OLR-constrained L v dP/dT s = 0.33 × 3.8 + 0.56 = 1.8 W m −2 K −1 with the standard deviation of 0.1 W m −2 K −1 . Assuming that L v dP/dT s contains random variations not captured by the linear relation with dOLR/dT s , the statistics of the fitting residual ε is characterized by all the models' fitting residuals, which yield a standard deviation of 0.39. Thus, the OLR-constrained L v dP/dT s has a mean of 1.8 W m −2 K −1 and a standard deviation of 0.4 = 0.39 2 + 0.1 2 W m −2 K −1 . Hence, the value of the OLR-constrained L v dP/dT s at the 95% confidence level (within two times of the standard deviation) is 1.8 ± 0.8 W m −2 K −1 .
Second, we compute the interannual precipitation sensitivity directly from the least squares regression of the GPCP precipitation onto the HadCRUT4 surface temperature for the period of 1995 to 2005. The 5-month running averaging is applied onto the de-seasonalized anomalies.
This gives the GPCP L v dP/dT s at 2.7±0.9 W m −2 K −1 .
Third, we choose the overlapped range of the two observational measures of the interannual L v dP/dT s , 1.8-2.6 W m −2 K −1 as the best estimate of the observation-based short-term precipitation sensitivity at the 95% confidence level (Figure 4).
Significance tests for the correlations in the study. For correlation coefficients involving 21 models, the 2-sided student-t test requires R ≥ 0.433 for the 95% significance level and R ≥ 0.558 for the 99% significance level. Hence, all the correlations relevant to our conclusions are statistically significant at the 95% level and in many cases at 99% significance level.