Changes in mean and extreme precipitation scale universally with global 1 mean temperature across and within climate models

Projections of precipitation from global climate models are crucial for risk assessment and adaptation strategies under different emission scenarios, yet model uncertainty limits their application. Here, we assess inter-model differences by separating the response of precipitation to anthropogenic forcing within 21 individual, bias-adjusted CMIP6 models using a pattern filtering technique. The forced response of mean precipitation, the number of wet days and the intensity and frequency of daily extremes are identified using low-frequency component analysis. Inter-model agreement in the sign of local change is moderate across land areas, with better agreement for extreme metrics (90% of models agree on 51, 41, 61, 61% of land area, for each metric respectively). Differences in the average magnitude of local changes are also large but can be explained well by the magnitude of global surface warming, despite model differences in the sign of local change (R2 of 0.81, 0.79, 0.69, 0.79). Moreover, we show that these temperature-precipitation scaling relationships can be identified robustly within individual climate models from inter-temporal changes in the detected forced response (median R2 of 0.82, 0.82, 0.76, 0.87). Inter-model spread in these relationships is considerable (coefficient of variation of 22, 33, 26, 17%), thus diagnosing a source of the uncertainty in the magnitude of projected precipitation change. These results suggest that despite uncertainty in the sign of regional change, the magnitude of future precipitation changes is well constrained by temperature-scaling relationships both across and within models. They may offer a new avenue to constrain the magnitude of future projections. 9


27
The hydrological cycle is likely to account for a considerable portion of the impacts of future 28 climate change. Key aspects of social well-being, such as agricultural productivity (Liang et al. 29 (2017)), flood damages (Davenport et

47
(2021)) dominated by a thermodynamic contribution with small inter-model discrepancy (Pfahl 48 et al. (2017)). Dynamical changes from atmospheric circulation cause only regional differences in 49 the magnitude of these increases, but contribute the majority of the uncertainty between models 50 (Pfahl et al. (2017)). 51 For seasonal and annual averages, thermodynamic processes are expected to lead to a "rich-52 get-richer" effect in which historical differences in regional precipitation are intensified (Seager  2018)). In general, changes to seasonal and annual averages are 60 projected to be heterogeneous with large inter-model uncertainty, often even in the sign of regional 61 change (Chadwick et al. (2016)). 62 With the aim of better constraining precipitation projections, we here provide an assessment of 63 future changes across 21 bias-adjusted (Lange (2019(Lange ( , 2021) members of the CMIP-6 ensemble.

64
To assess characteristics of the distribution of precipitation with relevance to societal outcomes of each precipitation characteristic. We find that despite large differences in the spatial pattern and 70 sign of regional change, the average magnitude of local changes scales strongly with global mean 71 2-m temperature (GMT) change, both across and within models. This suggests that even when 72 dynamic processes dominate, resulting in regionally heterogeneous changes with large inter-model 73 uncertainty, the intensity of these changes can be related back to the underlying thermodynamic 74 driver. These clear relations may help inform probabilistic assessments of the magnitude of 75 regional precipitation change (Chadwick et al. (2016)), valuable while the dynamical atmospheric 76 response and the resulting signs of regional change remain uncertain (Shepherd (2014)). Moreover, 77 the identification of precipitation-temperature scaling relationships for individual climate models

101
(2021)). One approach to overcome this issue is to use large ensembles of a single climate model 102 in which internal variability can be characterised and removed by initialising ensemble members 103 from different initial conditions (Kay et al. (2015)). However, to consider the full range of structural 104 model differences which can bias the forced response, a variety of climate models must be assessed. 105 We do so using the multi-model ensemble CMIP6, and instead apply a pattern filtering technique  Here we provide a conceptual summary of LFCA and of its application to identifying the climatic

135
a. The precipitation response to anthropogenic-forcing in individual climate models 136 We identify the response of mean daily precipitation, the number of wet days and the intensity  where less than 80% of models agree on the sign of change. Disagreement between models is often 189 concentrated at the boundary between regional increases and decreases, likely where uncertainty   On average, the rates at which precipitation scales with temperature change within individual 306 models is consistent with that identified between models. However, there is considerable inter-307 model heterogeneity in these rates, the distribution of which is shown in Fig. 12. EC-Earth appears 308 to be a consistent outlier, for mean precipitation and the intensity and frequency of daily extremes 309 in particular. Excluding this model, we calculate coefficients of variation of 17-33% for these 310 distributions, demonstrating that large inter-model uncertainty exists in the modelled rate at which 311 precipitation changes scale with temperature. These inter-model differences in intra-model scaling 312 rates are significant at the 10% level given our methodological uncertainty (estimated using 1000 313 bootstrapped replacements of inter-temporal changes) for between 47 and 64% of unique model 314 pairs, see Fig. D2. 315 We find weak evidence that these differences in the scaling rate may depend on the equilibrium 321 climate sensitivity (Fig. D3). There is, however, clear evidence for co-varying scaling rates across 322 models between the mean precipitation and the number of wet days, and between the frequency 323 and intensity of daily extremes, suggesting that common physical drivers underlie the model biases 324 in these indices (Fig. D4). Weaker evidence for co-varying rates between mean precipitation and 325 the daily extremes is also noted (Fig. D4).                  but without taking absolute values of regional precipitation change. The mean intra-model scaling and the inter-model scaling (identified in Fig. 7) are shown as dashed horizontal lines in black and red respectively. Using these uncertainty distributions of the scaling rate of each model, we calculate that inter-model differences in the scaling rate are significantly non-zero at the 10% level for 47, 60, 58 and 64% of unique model pairs for mean precipitation, the number of wet days, Rx1 and R>99p respectively.