### Observational data

To validate the simulated precipitation variability, daily precipitation observations from IMERG [0.1° × 0.1°; (*23*)] and GPCP [1° × 1°; (*24*)] were used. We compared the present-day mean precipitation and precipitation variability over their common period of 2001–2019 (fig. S2). We note that precipitation intensity statistics, such as variability, can be affected by horizontal resolution. For fair comparison, the higher-resolution IMERG data were regridded to the lower model resolution (N216) via area conservative remapping. The GPCP data were analyzed on its original grids. The comparison shows that the climatological mean and variability of precipitation simulated by the PPE ensemble are well within the range of observational uncertainty between IMERG and GPCP (partly related to their different horizontal resolutions; fig. S2).

### Model simulations

We used the 14-member PPE simulations conducted with HadGEM3-GC3.05 developed by the Met Office Hadley Centre (*13*, *14*). HadGEM3-GC3.05 is a state-of-the-art coupled global climate model with a high resolution of N216 (60 km in the mid-latitudes) in horizontal and 85 levels in vertical. The ensemble members are generated by simultaneously perturbing 47 parameters controlling key processes in the atmosphere, land, and aerosol model components within expert-specified limits. The simulation includes historical simulations and RCP 8.5 projections from 1900 to 2099.

The different model variants in the PPE sample a range of uncertainty in regional precipitation responses that is comparable to CMIP5 despite sampling a relatively narrow range of global temperature changes (*14*). The advantage of using a single model framework is that the members all share a common structural bias so that emergent relationships across the PPE tend to be clearer than those in a multimodel ensemble, in which structural bias is an extra source of uncertainty. Different parameterizations across the PPE variants can lead to different precipitation and sea surface temperature biases and sea surface warming patterns, which are relevant for precipitation projections. Detailed physical understanding is not allowed in CMIP5/6 ensembles because of the unavailability of daily circulation data.

To investigate the response to global warming, we compared precipitation variability between the first and last 60-year periods of the simulation, i.e., the baseline 1900–1959 and the projection in 2040–2099, to maximize the signal-to-noise ratio in changes. The use of 60-year-long periods ensures sufficient sampling and, hence, reliable estimates of variability. The changes are scaled with global mean near-surface air temperature changes to eliminate the uncertainty of different warming rates across model variants. As precipitation variability increases quasi-linearly with global warming within the level projected up to the end of the 21st century, the scale provides a reasonable estimate of the response to warming (fig. S10).

### Calculation of precipitation variability

We considered precipitation variations at typical time scales, including the synoptic variation at 2 to 10 days, the monthly variation at 25 to 35 days, the intraseasonal variation at 30 to 80 days, and the interannual variation at 1 to 8 years. The climatological annual cycle and linear trend were first removed from the time series before applying a filter according to the typical frequency for specific time scales. The variability was then estimated as the standard deviation of the filtered time series (*16*, *17*). The total precipitation variability was estimated as the standard deviation of the unfiltered time series.

### Assessing “wet-get-more variable” at the regional scale

To assess the applicability of the “wet-get-wetter” and “wet-get-more variable” paradigms at the regional scale, we compared regional changes in mean precipitation and the variability (Fig. 5A). We ranked all the grid points globally from the driest to the wettest according to their precipitation climatology. To reduce spatial noise, we applied spatial smoothing before the sort by regridding to a coarser resolution (from ~60 to ~360 km; regridding to different resolutions is tested, which yields similar results). The changes in the variability and mean state of precipitation at each grid point were then sorted on the basis of their ranks in climatology. We interpret the changes in a probabilistic perspective by aggregating over the neighboring grids (301 grid points used here; slightly different aggregations are tested, which yield similar results) to derive a probabilistic distribution of changes. Then, we show the median change with the 10th to 90th percentile ranges (curve and shading in Fig. 5A). We performed the analysis for each model variant separately and then show the ensemble median results in Fig. 5A. Hence, the change patterns are examined in each of the model worlds, taking into account the potential differences in the distributions of wet and dry regions among different model variants.

For the wet-get-wetter paradigm, in addition to the precipitation-based ranks (Fig. 5A), we also examined the paradigm on the basis of ranks by precipitation minus evaporation (*P* − *E*; figure not shown). At the regional scale, the wet-get-wetter paradigm does not hold for either *P* or *P* − *E*, largely because of spatial shifts in convection and convergence in the tropics under future warming (*25*, *40*).

### Moisture budget diagnostics

We applied the moisture budget diagnostics to investigate the physical processes driving the changes in precipitation variability. It has been widely used to diagnose changes in mean and extreme precipitation (*6*, *27*). In a climate state, precipitation (*P*) is balanced by evaporation (*E*) and vertical (− < ω∂_{p}*q*>) and horizontal (− < **V** · ∇ *q*>) moisture advection that are related to low-level convergence and horizontal winds, respectively

(2)where *q* is specific humidity, ω is vertical velocity, **V** is horizontal wind vector, δ_{0} is the residual, and

denotes vertical integral throughout the troposphere. Such a balance also holds at a specific time scale

$$encoding>{P}_{f}-{E}_{f}=-\mathrm{\omega}{\mathrm{\partial}}_{p}q{}_{f}-\mathbf{V}\xb7\nabla q{}_{f}+{\mathrm{\delta}}_{1}$$(3)where the subscript *f* denotes variation at a specific time scale derived from the filter.

Next, we applied simplifications to the full moisture budget to determine the moisture process that dominates the variation of precipitation and, based on which, to further understand the mechanisms for the projected changes.

First, among the moisture budget terms, the vertical moisture advection dominates precipitation variation at all time scales

$$encoding>{P}_{f}\approx -\mathrm{\omega}{\mathrm{\partial}}_{p}q{}_{f}$$(4)

It largely captures the phase and magnitude of precipitation variation, as supported by the high temporal correlation (or explained variance) and low root mean square deviation (RMSD) with precipitation (fig. S6). Therefore, the vertical moisture advection reasonably reproduces the climatological precipitation variability (fig. S5)

$$encoding>{P}_{f}\approx {[-\frac{{\mathrm{\omega}}_{m}({q}_{\mathrm{l}}-{q}_{\mathrm{u}})}{g}]}_{f}$$(5)where ω* _{m}* represents the vertical motion at mid-troposphere, which is closely related to precipitation. As atmospheric moisture is concentrated in the lower troposphere, the right-hand side of Eq. 5 is dominated by its lower-level component. Hence, the variation in precipitation can be further approximated as that in the vertical advection of lower-level moisture

(6)

Here, ω* _{m}* is represented by mid-tropospheric vertical velocity at 500 hPa and

*q*

_{l}is represented by specific humidity at 925 hPa.

In terms of change, the vertical moisture advection reasonably reproduces changes in precipitation variability under global warming (comparing Figs. 4A and 6A)

$$encoding>\begin{array}{c}\u2206\mathrm{\sigma}[{P}_{f}]\approx \u2206\mathrm{\sigma}\left[{(-\frac{{\mathrm{\omega}}_{m}{q}_{\mathrm{l}}}{g})}_{f}\right]=\\ \mathrm{\sigma}\left[{(-\frac{{\mathrm{\omega}}_{m1}{q}_{\mathrm{l}1}}{g})}_{f}\right]-\mathrm{\sigma}\left[{(-\frac{{\mathrm{\omega}}_{m0}{q}_{\mathrm{l}0}}{g})}_{f}\right]\end{array}$$(7)where σ denotes variability (estimated by standard deviation) and ∆ denotes the change between the baseline (1900–1959) and future (2040–2099), which are indicated by the subscripts 0 and 1, respectively.

We further separated the contributions from thermodynamics, dynamics, and nonlinear processes using idealized models. To estimate the thermodynamic (TH) contribution, which is related to changes in atmospheric moisture only, we change the specific humidity (*q*_{l}) to the future value and keep the circulation (ω_{m}) as in the baseline. The estimated changes in variability with this idealized model relative to the base period are regarded as the role of thermodynamics

(8)

Likewise, for the dynamic (DY) contribution, which is due to changes in circulation only, we change the vertical velocity (ω_{m}) to the future value and keep the humidity (*q*_{l}) as in the baseline. Hence, the DY contribution is estimated as the changes in variability with this configuration relative to the baseline

(9)

The nonlinear (NL) effect involves interactions between changes in humidity and circulation. It is estimated as the residual between the full changes in vertical moisture advection and the TH and DY contributions estimated from Eqs. 8 and 9

$$encoding>\mathit{NL}\approx \u2206\mathrm{\sigma}\left[{(-\frac{{\mathrm{\omega}}_{\mathrm{m}}{q}_{\mathrm{l}}}{g})}_{f}\right]-\mathit{TH}-\mathit{DY}$$(10)

To provide a theoretical understanding of the thermodynamic and dynamic effects, we investigated the moisture budget in a further simplified framework. As the variability in vertical motion is far larger than that in humidity, the variation in vertical moisture advection is largely governed by that in vertical motion. Hence, by neglecting the variation in humidity and its interaction with circulation, Eq. 6 can be simplified as follows

$$encoding>{P}_{f}\approx {(-{\mathrm{\omega}}_{\mathrm{m}}{q}_{\mathrm{l}}/g)}_{f}\approx -\frac{{({\mathrm{\omega}}_{\mathrm{m}})}_{f}\overline{{q}_{\mathrm{l}}}}{g}$$(11)where

$encoding>\overline{{q}_{\mathrm{l}}}$ is the climatological mean low-level humidity in a climate state. Thus, the variability in precipitation is proportional to that in vertical motion (σ[ − (ω_{m})* _{f}*])

(12)

Under this framework, the thermodynamic effect can be estimated as

$encoding>\frac{\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}0})}_{f}]}{g}(\overline{{q}_{\mathrm{l}1}}-\overline{{q}_{\mathrm{l}0}})$, which is determined by changes in atmospheric moisture and climatological circulation variability. The dynamic effect can be estimated as

$encoding>\frac{\overline{{q}_{\mathrm{l}0}}}{g}\{\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}1})}_{f}]-\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}0})}_{f}]\}$, which is determined by changes in circulation variability and climatological moisture availability.

The contributions of each term can be expressed as the percentage with respect to climatological precipitation variability, which then measures the direct contributions to the percentage precipitation variability change. The advantage is that the contributions of moisture and circulation changes are clearly separated between thermodynamics and dynamics

$$encoding>\mathit{TH}(t)\approx \frac{\{\frac{\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}0})}_{f}]}{g}(\overline{{q}_{\mathrm{l}1}}-\overline{{q}_{\mathrm{l}0}})\}}{\left\{\frac{\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}0})}_{f}]\overline{{q}_{\mathrm{l}0}}}{g}\right\}}=\mathrm{\delta}\overline{{q}_{\mathrm{l}}}$$(13)

$$encoding>\mathit{DY}(t)\approx \frac{\{\frac{\overline{{q}_{\mathrm{l}0}}}{g}\{\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}1})}_{f}]-\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}0})}_{f}]\}\}}{\left\{\frac{\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}0})}_{f}]\overline{{q}_{\mathrm{l}0}}}{g}\right\}}=\mathrm{\delta}\mathrm{\sigma}[-{({\mathrm{\omega}}_{\mathrm{m}})}_{f}]$$(14)where *TH*(*t*) and *DY*(*t*) indicate the theoretical estimations of their contributions, respectively, and δ denotes a percentage change. To first order, the *TH* effect acts to enhance precipitation variability by the rate of background moistening (Eq. 13). The dynamic effect is associated with changes in the variability of vertical motion (Eq. 14).