Editing IPCC:AR6/WGI/Chapter-8 (section)

== 8.5 What Are the Limits for Projecting Water Cycle Changes? ==

<div id="h1-6-siblings" class="h1-siblings"></div>

Understanding the limits to projecting water cycle changes are fundamental for refining climate and hydrological models needed to develop successful climate change adaptation strategies. Regional water cycle projections depend on a range of model-dependent responses ( [[#8.5.1|Section 8.5.1]] ) and are also strongly influenced by internal variability, especially in the near term ( [[#8.5.2|Section 8.5.2]] ; [[#Hawkins--2012|Hawkins and Sutton, 2012]] ; [[#Rowell--2012|Rowell, 2012]] ; [[#Orlowsky--2013|Orlowsky and Seneviratne, 2013]] ; [[#Kent--2015|Kent et al., 2015]] ; [[#Fatichi--2016|Fatichi et al., 2016]] ; [[#Greve--2018|Greve et al., 2018]] ; [[#Chegwidden--2019|Chegwidden et al., 2019]] ). CMIP6 models show that different model responses to the same forcing scenario remain the main source of uncertainty for projected changes in regional precipitation (Figure 8.23; [[#Lehner--2020|Lehner et al., 2020]] ). [[#8.5.3|Section 8.5.3]] assesses the potential for non-linear responses when shifting from low- to high-global warming levels ( [[#8.4.2.4|Section 8.4.2.4]] ; [[#James--2017|James et al., 2017]] ). While regional uncertainties related to downscaling methods ( [[IPCC:Wg1:Chapter:Chapter-10#10.3.3|Section 10.3.3]] ) and impact models (WGII Chapter 4) are not covered here, the added value of regional climate models is briefly discussed ( [[#8.5.1.2.2|Section 8.5.1.2.2]] ) with a focus on water cycle changes.

<div id="_idContainer066" class="Basic-Text-Frame"></div>

[[File:0b8f2c03f5b75fc8ee3c79acde78d727 IPCC_AR6_WGI_Figure_8_23.png]]

'''Figure 8.23 |''' '''Geographical and zonal mean distribution of the percentage of variance explained by the three sources of uncertainty in CMIP6 projections of 20-year mean precipitation changes in''' '''2021–2040''' '''(top),''' '''2041–2060''' '''(middle) and''' '''2081–2100''' '''(bottom) relative to the''' '''1995–2014''' '''base period:''' '''Internal climate variability (left), model response uncertainty (middle) and scenario uncertainty (right, considering four plausible concentration scenarios:''' '''SSP1-2.6''' ''',''' '''SSP2-4.5''' ''',''' '''SSP3-7.0''' '''and''' '''SSP5-8.5''' ''').''' Percentage numbers give the area-weighted global average value for each map. Right panels show the zonal mean fractions over both land and sea (solid lines) and over land only (dashed line). The figure was adapted from Figure 4a in [[#Lehner--2020|Lehner et al. (2020)]] , https://creativecommons.org/licenses/by/4.0/ . The relative contributions of internal variability, models and emissions scenarios to the total uncertainty depend on both region and time horizon. The scenario uncertainty is relatively low in near and mid-term time horizons while it increases in the long term mostly over the high latitudes. The model response uncertainty is the most influential factor across all time horizons. Internal variability also plays a key role in the near term, especially in the subtropics. Further details on data sources and processing are available in the chapter data table (Table 8.SM.1).

<div id="8.5.1" class="h2-container"></div>

<span id="model-uncertainties-of-relevance-for-the-water-cycle"></span>
=== 8.5.1 Model Uncertainties of Relevance for the Water Cycle ===

<div id="h2-17-siblings" class="h2-siblings"></div>

Model response uncertainty is typically estimated as the inter-model spread (range) projected by a set of climate models for a given emissions scenario. It is best estimated at the end of a high-emissions scenario when internal variability has a limited contribution to total uncertainty (Figure 8.23). Even for aggregated quantities, like decadal-mean precipitation averaged over relatively large domains, model response uncertainty is substantial and can exceed scenario uncertainty ( [[#Hawkins--2011|Hawkins and Sutton, 2011]] ; [[#Lehner--2020|Lehner et al., 2020]] , 1.5.4, 4.4.1.3). This can also be true for other water cycle variables such as soil moisture, runoff and streamflow at the regional scale, either derived directly from global climate models (GCMs) or produced by ‘offline’ using global hydrological models (GHMs) driven by the same GCMs (Orlowsky and Seneviratne, 2013; [[#Giuntoli--2015|Giuntoli et al., 2015]] , 2018; [[#Chegwidden--2019|Chegwidden et al., 2019]] ). Although some of the model response uncertainty is related to climatological biases ( [[#Grose--2017|Grose et al., 2017]] ; G. [[#Li--2017|]] [[#Li--2017|]] [[#Li--2017|]] [[#Li--2017|Li et al., 2017]] ; [[#Lehner--2019|Lehner et al., 2019]] ; [[#Samanta--2019|Samanta et al., 2019]] ), model biases are not the only way to assess the reliability of climate projections (compare with Box 4.1). Therefore, our focus here is on the representation of key processes that are not completely resolved in current-generation GCMs ( [[#8.5.1.1|Section 8.5.1.1]] ) and on the model improvements associated with increased horizontal resolution ( [[#8.5.1.2|Section 8.5.1.2]] ).

<div id="8.5.1.1" class="h3-container"></div>

<span id="fitness-for-purpose-and-poorly-constrained-key-processes"></span>
==== 8.5.1.1 Fitness-for-purpose and Poorly Constrained Key Processes ====

<div id="h3-43-siblings" class="h3-siblings"></div>

The AR5 ( [[IPCC:Wg1:Chapter:Chapter-7|Chapter 7]] recognized that the simulation of clouds and precipitation remains challenging for state-of-the-art GCMs. Model development and evaluation have continued since AR5, with a particular emphasis on the representation of new model components, like interactive vegetation, aerosols and biogeochemical cycles. For example, the comparison of simulated tropical precipitation across three successive generations of CMIP models (including CMIP6) indicates overall little improvement for the summer monsoons, the double-ITCZ bias, the diurnal cycle and the frequency of precipitation ( [[#Fiedler--2020|Fiedler et al., 2020]] ). Some of these issues are related to inherent model limitations in three specific areas: atmospheric convection, cloud – aerosol interactions and land surface processes (ocean and cryosphere-related processes are addressed in Chapter 9). These limitations do not weaken the overall progress made in the large-scale simulation of present-day climate (FAQ 3.3 and [[IPCC:Wg1:Chapter:Chapter-3#3.3.2.3|Section 3.3.2.3]] ), even though the improvement of CMIP6 compared with CMIP5 models is limited (Figure 3.12) and is generally less systematic or obvious at the regional scale (e.g., Gusain et al. , 2020; Monerie et al. , 2020; Oudar et al. , 2020a) . Instead, they call for a careful interpretation of hydrological projections with the full range of plausible outcomes, rather than only considering the most likely scenarios ( [[#Sutton--2018|Sutton, 2018]] , 2019).

<div id="8.5.1.1.1" class="h4-container"></div>

<span id="atmospheric-convection"></span>
===== 8.5.1.1.1 Atmospheric convection =====

<div id="h4-29-siblings" class="h4-siblings"></div>

Moist convection is fundamental to the water cycle through its vertical transport of momentum, heat, and moisture across the atmosphere. It is particularly active in the tropics where it contributes to more than half of annual precipitation and to the development of severe weather events. Given limitations in computing resources, the current-generation GCMs cannot yet represent small-scale cloud processes and consequently shallow and deep convection is determined by sub-grid-scale parametrizations. While such parametrizations can be evaluated against field observations (e.g., [[#Abdel-Lathif--2018|Abdel-Lathif et al., 2018]] ), it remains challenging to estimate convective entrainment that is valid for both shallow and deep convection (G.J. [[#Zhang--2016|]] [[#Zhang--2016|]] [[#Zhang--2016|]] [[#Zhang--2016|Zhang et al., 2016]] ). Comparisons between regional projections with explicit compared with parametrized convection also highlight the limitations of parametrized convection for assessing climate change ( [[#Kendon--2019|Kendon et al., 2019]] ; [[#Jackson--2020|Jackson et al., 2020]] ).

Atmospheric convection is particularly important for a realistic simulation of tropical precipitation intensities ( [[#Pendergrass--2014a|Pendergrass and Hartmann, 2014a]] ; [[#Kendon--2019|Kendon et al., 2019]] ). Many CMIP5 models produce rainfall at water vapour amounts lower than in observations ( [[#Takahashi--2018|Takahashi, 2018]] ), as well as too light and too frequent precipitation events ( [[#Sun--2015|Sun et al., 2015]] ; [[#Trenberth--2017|Trenberth et al., 2017]] ). Such biases can be explained by a lack of convective inhibition ( [[#Rochetin--2014a|Rochetin et al., 2014a]] , b) and by too much convective and too little non-convective precipitation ( [[#Chen--2019|Chen and Dai, 2019]] ). Tropical convection controls the amount of precipitable water simulated over the equatorial Indian Ocean, which has been identified as a key metric for differentiating model skill in simulating South Asian monsoon precipitation ( [[#Hagos--2019|Hagos et al., 2019]] ). Many models have difficulty in adequately simulating the diurnal cycle of precipitation over land ( [[#Couvreux--2015|Couvreux et al., 2015]] ), the rainfall intensity distribution associated with the West African monsoon ( [[#Roehrig--2013|Roehrig et al., 2013]] ), and the intensity of tropical cyclones (Sections 10.3.3.4 and 11.7.1.3), phenomena for which atmospheric convection also plays a key role.

Since AR5, there have been improvements in the representation of convective clouds and related precipitation in GCMs. For instance, the drizzle issue (too light and too frequent rainfall events) has led to modifications in the deep convection triggering scheme (Rochetin et al. , 2014b; Han et al. , 2017; Xie et al. , 2018; Wu et al. , 2019) . Although high-resolution studies have highlighted these limitations, most GCMs still rely on a convective available potential energy (CAPE) closure which has been adapted to various cloud regimes ( [[#Bechtold--2014|Bechtold et al., 2014]] ; [[#Han--2017|Han et al., 2017]] ; [[#Walters--2019|Walters et al., 2019]] ) or evaluated against convection-permitting models (CPMs; J. [[#Chen--2020|]] [[#Chen--2020|Chen et al., 2020]] a). To increase the sensitivity of convection to tropospheric humidity, several models now include a representation of deep convective entrainment dependent on relative humidity (Bechtold et al. , 2008; Han et al. , 2017; M. Zhao et al. , 2018; Walters et al. , 2019) . Other efforts have focused on the improvement of shallow convection and low-level cloudiness due to their major contribution to uncertainty in climate sensitivity ( [[IPCC:Wg1:Chapter:Chapter-7#7.4.2.4|Section 7.4.2.4]] ). A cloud-regime-based study however highlights an apparent disconnection between cloud and precipitation processes in GCMs ( [[#Tan--2018|Tan et al., 2018]] ), suggesting that a good representation of clouds does not lead to systematic improvement in simulated precipitation. A global simulation in which the parametrized convection is switched off shows a strong influence of parametrized convection on daily precipitation extremes(P. [[#Maher--2018|]] [[#Maher--2018|Maher et al., 2018]] ). Regional simulations at a 25km resolution suggest that an explicit deep convection can be beneficial even at such a relatively coarse resolution ( [[#Vergara-Temprado--2020|Vergara-Temprado et al., 2020]] ). Perturbed physics ensembles (PPE, [[IPCC:Wg1:Chapter:Chapter-1#1.4.4|Section 1.4.4]] ) make it possible to identify parameters in the convection scheme that are most important in determining future precipitation changes ( [[#Bernstein--2016|Bernstein and Neelin, 2016]] ).

Since AR5, spatial aggregation of tropical convection has also received growing attention in both observational ( [[#Holloway--2017|Holloway et al., 2017]] ) and modelling studies ( [[#Muller--2015|Muller and Bony, 2015]] ; [[#Wing--2017|Wing et al., 2017]] ; [[#Tan--2018|Tan et al., 2018]] ). The '''changing degree of convective organization was highlighted as a key mechanism for dynamic changes in extreme precipitation ( [[#Pendergrass--2020a|Pendergrass, 2020a]] ).''' Yet, convective parametrizations do not represent all aspects of mesoscale convective systems ( [[#Hourdin--2013|Hourdin et al., 2013]] ; [[#Park--2019|Park et al., 2019]] ). This is related to the complexity of mechanisms involved from synoptic to mesoscale dynamics, which are only partially resolved by models. Cloud-resolving models (CRMs, [[#8.5.1.2.2|Section 8.5.1.2.2]] ) represent a useful benchmark for improving the parametrization of mesoscale convective systems. Machine learning can also be used to parametrize moist convection after training the model with a conventional or a super parametrization scheme ( [[#Gentine--2018|Gentine et al., 2018]] ; [[#O’Gorman--2018|O’Gorman and Dwyer, 2018]] ), but has not yet been used in the CMIP framework.

While some global modelling centres have reported progress in their parametrization of convection and in their simulation of seasonal, daily and sub-daily precipitation (e.g., [[#Danabasoglu--2020|Danabasoglu et al., 2020]] ; [[#Roehrig--2020|Roehrig et al., 2020]] ), CMIP6 models as a whole only show limited improvements in their simulation of the tropical precipitation climatology compared to CMIP5 (Figure 3.10; [[#Fiedler--2020|Fiedler et al., 2020]] ). For instance, the double-ITCZ syndrome is still prominent ( [[#Tian--2020|Tian and Dong, 2020]] ) despite being reduced in some models (e.g., [[#Qin--2018|Qin and Lin, 2018]] ) . This systematic bias was shown to arise from atmospheric processes including cloud feedbacks ( [[#Tian--2015|Tian, 2015]] ; [[#Dixit--2018|Dixit et al., 2018]] ; [[#Talib--2018|Talib et al., 2018]] ) and the SST threshold at which deep convection occurs in the tropics ( [[#Oueslati--2015|Oueslati and Bellon, 2015]] ; [[#Xiang--2017|Xiang et al., 2017]] ; [[#Adam--2018|Adam et al., 2018]] ). Such biases can also arise from a too weak sensitivity of seasonal tropical precipitation to local SSTs compared with observations ( [[#Good--2021|Good et al., 2021]] ). These biases are large enough to alter forced precipitation changes, and consequently limit our confidence in projected precipitation changes ( [[#Samanta--2019|Samanta et al., 2019]] ; [[#Aadhar--2020|Aadhar and Mishra, 2020]] ). Observational constraints can be used to narrow model response uncertainties ( [[#DeAngelis--2015|DeAngelis et al., 2015]] ; G. [[#Li--2017|]] [[#Li--2017|]] [[#Li--2017|]] [[#Li--2017|Li et al., 2017]] ; [[#Ham--2018|Ham et al., 2018]] ; [[#Watanabe--2018|Watanabe et al., 2018]] ), although there is still no consensus that model selection or weighting is a reliable alternative to the ‘one-model-one-vote’ approach used in [[#8.4|Section 8.4]] (Box 4.1). The detrimental influence of model errors can also be mitigated by focusing on phenomena or events (Polson and Hegerl, 2017; [[#Weller--2017|Weller et al., 2017]] ), implementing bias adjustment techniques ( [[IPCC:Wg1:Chapter:Chapter-10#10.2.3.2|Section 10.2.3.2]] ), or adopting a non-probabilistic storyline approach (Zappa and Shepherd, 2017).

In summary, since AR5 empirical convective parametrization schemes and associated precipitation biases have improved in some but not all global climate models. There is still ''low confidence'' in their ability to accurately simulate the spatio-temporal features of present-day precipitation, especially in the tropics where a double-ITCZ bias is still apparent in many models. While such biases limit the reliability of precipitation projections in some cases, there is currently only ''medium confidence'' that model selection or weighting is a better alternative to the one-model-one-vote approach (Box 4.1). Improved water cycle projections can be achieved by focusing on phenomena or weather events, such as a thermodynamic intensification of convective events ( ''high confidence'' , [[#8.2.2.1|Section 8.2.2.1]] ), however accurate quantitative estimates are currently hampered by complex, model-dependent dynamical responses ( [[#8.2.2.2|Section 8.2.2.2]] ).

<div id="8.5.1.1.2" class="h4-container"></div>

<span id="aerosol-microphysical-effects-on-clouds-and-precipitation"></span>
===== 8.5.1.1.2 Aerosol microphysical effects on clouds and precipitation =====

<div id="h4-30-siblings" class="h4-siblings"></div>

In AR5 Chapter 7, there was ''low confidence'' in the representation of cloud–aerosol interactions in climate models. Despite progresses in this field since AR5, cloud–aerosol interactions remain a major obstacle to understanding climate and severe weather ( [[#Varble--2018|Varble, 2018]] ). High aerosol concentrations have been observed to suppress rain in water clouds ( [[#Campos%20Braga--2017|Campos Braga et al., 2017]] ; [[#Fan--2020|Fan et al., 2020]] ). However, such aerosol effects are muted in GCMs, which tend to produce precipitation from shallow clouds too frequently at the expense of rain intensity ( [[#Suzuki--2015|Suzuki et al., 2015]] ; [[#Jing--2017|Jing et al., 2017]] ). This arises from incomplete knowledge of how clouds adjust to aerosol primary effects such as cloud condensation nuclei (CCN). The adjustment occurs mainly as a dynamic response to the impacts of CCN on cloud droplet size and number concentrations on precipitation-forming processes ( [[#Rosenfeld--2008|Rosenfeld et al., 2008]] ; [[#Goren--2014|Goren and Rosenfeld, 2014]] ; [[#Koren--2014|Koren et al., 2014]] ; [[#Camponogara--2018|Camponogara et al., 2018]] ). Uncertainties are large for deep clouds, as their processes are much more complex and include also the impacts of aerosols on ice-precipitation processes. Aerosols can substantially invigorate ( [[#Rosenfeld--2008|Rosenfeld et al., 2008]] ; [[#Koren--2014|Koren et al., 2014]] ; [[#Fan--2018|Fan et al., 2018]] ) and electrify ( [[#Thornton--2017|Thornton et al., 2017]] ; Q. [[#Wang--2018|]] [[#Wang--2018|]] [[#Wang--2018|]] [[#Wang--2018|]] [[#Wang--2018|]] [[#Wang--2018|Wang et al., 2018]] ) deep tropical convective clouds. High-resolution atmospheric simulations suggest that high aerosol concentrations can increase environmental humidity by producing clouds that mix more condensed water into the surrounding air, which in turn favours large-scale ascent and strong convective events ( [[#Abbott--2021|Abbott and Cronin, 2021]] ). Further assessment of uncertainties in aerosol – cloud interactions for shallow water clouds is provided in [[IPCC:Wg1:Chapter:Chapter-7#7.3.3.2|Section 7.3.3.2]] .

A major challenge in representing convective clouds and related precipitation events in GCMs is a lack of sophisticated cloud microphysics in convective parametrization schemes (e.g., [[#Fan--2016|Fan et al., 2016]] ). Most of these schemes only include simple microphysical treatments, such as direct partition between cloud condensation and precipitation, and do not include advanced treatment of conversion among different types of hydrometeors. As such these schemes are unable to simulate microphysical cloud and precipitation responses to aerosol-related perturbations in cloud droplet concentration and ice crystals (see Box 8.1), or perturbations in thermodynamical states from global warming. Efforts have been made to include more advanced cloud microphysical treatment in cumulus parametrizations ( [[#Song--2011|Song and Zhang, 2011]] ; [[#Grell--2014|Grell and Freitas, 2014]] ; [[#Berg--2015|Berg et al., 2015]] ) or to use explicit cloud microphysics schemes in climate models with a ‘super parametrization’ ( [[#Wang--2015|Wang et al., 2015]] ), which have been shown to improve the performance in simulating cloud properties and precipitation. However, few of these improvements have been incorporated into CMIP6 climate models so the projected precipitation response to anthropogenic perturbation may still be hindered by the inadequate microphysical treatment in cumulus parametrization ( [[#Smith--2020|Smith et al., 2020]] ).

In summary, there is still ''low confidence'' in the simulated influence of the aerosol microphysical effects on future precipitation changes.

<div id="8.5.1.1.3" class="h4-container"></div>

<span id="land-surface-processes"></span>
===== 8.5.1.1.3 Land surface processes =====

<div id="h4-31-siblings" class="h4-siblings"></div>

Land surface processes determine the partitioning of net surface radiation into sensible, latent and ground heat fluxes, the partitioning of precipitation into evapotranspiration and runoff, and the net terrestrial carbon flux at the Earth’s surface. They are relevant for simulating the terrestrial water cycle responses to climate change, as well as the response to land use change (FAQ 8.1). Even basic land surface properties such as albedo ( [[#Terray--2018|Terray et al., 2018]] ) or the ratio of transpiration to total evaporation ( [[#Chang--2018|Chang et al., 2018]] ) still need to be improved in state-of-the-art coupled GCMs. Runoff sensitivities are also not well constrained in these models, which display a large spread for the present-day climate, influencing simulated changes under global warming ( [[#Lehner--2019|Lehner et al., 2019]] ). Earth System Models (ESMs) incorporate some combined biophysical and biogeochemical processes to a limited extent, and many relevant processes about how plants and soils interactively respond to climate changes are yet to be considered (e.g., Y. [[#Liu--2020|]] [[#Liu--2020|]] [[#Liu--2020|Liu et al., 2020]] ). Consequently, land surface processes and their atmospheric coupling contribute to the range in water cycle projections ( [[#Jia--2019|Jia et al., 2019]] ).

Since AR5, development of new and existing processes in land surface models (LSMs) have been evaluated. These include soil freezing and permafrost (Vergnes et al. , 2014; Chadburn et al. , 2015; K. Yang et al. , 2018; Gao et al. , 2019) , soil and snow hydrology ( [[#Brunke--2016|Brunke et al., 2016]] ; [[#Decharme--2016|Decharme et al., 2016]] ), glaciers ( [[#Shannon--2019|Shannon et al., 2019]] ), surface waters and rivers ( [[#Decharme--2012|Decharme et al., 2012]] ), as well as vegetation ( [[#Bartlett--2015|Bartlett and Verseghy, 2015]] ; [[#Betts--2015|Betts et al., 2015]] ; [[#Knauer--2015|Knauer et al., 2015]] ; [[#Tang--2015|Tang et al., 2015]] ) and the representation of hydraulic gradients throughout the soil–plant–atmosphere continuum (Bonan et al., 2014). Such land surface model developments have led to significant improvements in global offline hydrological simulations driven by observed atmospheric forcings (e.g., [[#Li--2017|]] [[#Li--2017|]] [[#Li--2017|]] [[#Li--2017|C. Li et al., 2017]] ; [[#Decharme--2019|Decharme et al., 2019]] ).

Progress in the representation of land surface heterogeneity has been made, in the form of improved mapping of root zone storage capacity (Wang-Erlandsson et al., 2016), improved vegetation stand, disturbance and fire dynamics (F. [[#Li--2013|]] [[#Li--2013|Li et al., 2013]] ; [[#Fisher--2018|Fisher et al., 2018]] ; [[#Haverd--2018|Haverd et al., 2018]] ; [[#Yue--2018|Yue et al., 2018]] ; [[#Zou--2019|Zou et al., 2019]] ), better representation of urban surfaces (Box 10.3), and the explicit representation of inland water bodies ( [[#Gu--2015|Gu et al., 2015]] ; [[#Verseghy--2017|Verseghy and MacKay, 2017]] ). The representation of realistic snow and vegetation cover significantly affects the simulation of the land surface energy and water budgets at multiple time scales (Loranty et al., 2014; [[#Bartlett--2015|Bartlett and Verseghy, 2015]] ; [[#Thackeray--2015|Thackeray et al., 2015]] ; [[#Qiu--2016|Qiu et al., 2016]] ; [[#Thackeray--2016|Thackeray and Fletcher, 2016]] ; L. [[#Wang--2016|]] [[#Wang--2016|]] [[#Wang--2016|Wang et al., 2016]] ; [[#Alessandri--2017|Alessandri et al., 2017]] ). Groundwater remains inadequately represented in many models, which limits our current understanding of the two-way interactions between groundwater and the rest of the hydrologic cycle (R.G. Taylor et al. , 2013a; Leng et al. , 2014; Vergnes et al. , 2014; Pokhrel et al. , 2015; [[#Maxwell--2016|Maxwell and Condon, 2016]] ; [[#Collins--2017|Collins, 2017]] ; Scanlon et al. , 2018; Condon et al. , 2020) . Land management exerts an increasing influence on the water cycle ( [[#Abbott--2019|Abbott et al., 2019]] ) whose representation in the current-generation climate models is generally incomplete ( [[IPCC:Wg1:Chapter:Chapter-10#10.3.3.7.2|Section 10.3.3.7.2]] ).

Aside from land surface models (LSMs), global hydrological models (GHMs) have been further developed for off-line simulations of the hydrological impacts of both climate change and water management (Jiménez Cisneros et al. , 2014; Schewe et al. , 2014; Döll et al. , 2016, 2018; Pokhrel et al. , 2016, 2017; Veldkamp et al. , 2018) . GHMs can equal or outweigh the contribution of GCMs to uncertainties in hydrological projections at the regional scale (Giuntoli et al., 2015). Historical GHM simulations are currently not sufficient to improve regional water cycle projections, due to modelling uncertainties in both the driving GCMs and land surface hydrology (Pechlivanidis et al. , 2017; Samaniego et al. , 2017; Hattermann et al. , 2018; Krysanova et al. , 2018) . Biophysical vegetation processes are still not accounted for in many GHMs, which may lead to inadequate projections of terrestrial runoff and water resources. However, hydrological models that do simulate these effects often disagree ( [[#Prudhomme--2014|Prudhomme et al., 2014]] ), so do not necessarily provide the added value of a more sophisticated representation of vegetation processes and land surface conditions ( [[#Döll--2016|Döll et al., 2016]] ).

Since AR5, there has been increasing recognition of the need to better understand the role of land–atmosphere coupling and related feedbacks (Joetzjer et al. , 2014; Berg et al. , 2016; Catalano et al. , 2016; [[#Berg--2018a|Berg and Sheffield, 2018a]] ; Santanello et al. , 2018) . This has led to the development of dedicated field campaigns (Song et al., 2016; [[#Phillips--2017|Phillips et al., 2017]] ; [[#Dirmeyer--2018|Dirmeyer et al., 2018]] ), remotely sensed observations (Ferguson and Wood, 2011; [[#Roundy--2017|Roundy and Santanello, 2017]] ), and tailored diagnostics (Tawfik et al., 2015a, b; [[#Miralles--2016|Miralles et al., 2016]] , 2019; [[#Dirmeyer--2017|Dirmeyer and Halder, 2017]] ). Dynamic vegetation models have been introduced in global ESMs but they need further evaluation (Medlyn et al., 2015; [[#Prentice--2015|Prentice et al., 2015]] ; [[#Cantú--2018|Cantú et al., 2018]] ; [[#Franks--2018|Franks et al., 2018]] ) to provide valuable information on potential vegetation feedbacks. Plant migration and mortality, increased disturbances from wild fires, insects and extreme events, interactive nitrogen cycle, or the impact of increased levels of tropospheric ozone are often ignored or poorly represented in the current-generation of ESMs (Bonan and Doney, 2018; [[#Fisher--2018|Fisher et al., 2018]] ).

The physiological response of plants to increasing atmospheric CO <sub>2</sub> is generally accounted for, but only using empirical models of stomatal conductance that are characterized by a single critical parameter of intrinsic water-use efficiency (Franks et al., 2017, 2018). This reflects a lack of structural diversity and caution about the consensus of the photosynthesis response to increasing CO <sub>2</sub> ( [[#Knauer--2015|Knauer et al., 2015]] ; [[#Huang--2016|Huang et al., 2016]] ), which has implications for the ability of the current-generation models to account for uncertainty in future evapotranspiration changes. Most CMIP5 models underestimate the ratio of plant transpiration to total terrestrial evapotranspiration, which may suggest that they also underestimate the impact of plant physiology on the water cycle (Lian et al., 2018). Plant hydraulics are not explicitly considered in many land surface models, which may lead to an underestimation of the influence of the increasing atmospheric moisture stress on plant transpiration under climate change (Massmann et al. , 2019; Grossiord et al. , 2020; Y. Liu et al. , 2020) . Most ESMs underestimate the water use efficiency measured at many sites and, consequently overestimate the ratio of evapotranspiration to precipitation (J. [[#Li--2018|]] [[#Li--2018|]] [[#Li--2018|Li et al., 2018]] ).

In summary, since AR5 substantial advances have been made in the representation of land surface processes in current-generation Earth System Models (ESMs). Offline hydrological models allow the application of bias-adjusted atmospheric forcings, but there is ''low confidence'' of an improved response compared to coupled climate models, given their inherent limitations (Box 10.2). While improvements in the representation of complex land surface feedbacks relevant to the water cycle are needed, there is currently ''low confidence'' that they will systematically improve the reliability of water cycle projections.

<div id="8.5.1.2" class="h3-container"></div>

<span id="added-value-of-increased-horizontal-model-resolution"></span>
==== 8.5.1.2 Added Value of Increased Horizontal Model Resolution ====

<div id="h3-44-siblings" class="h3-siblings"></div>

Coarse spatial resolution of climate models has often been considered a key limitation in global climate projections ( [[#Di%20Luca--2015|Di Luca et al., 2015]] ; [[#Roberts--2018|Roberts et al., 2018]] ). Proposed and tested solutions include a uniform or regional increase in the resolution of GCMs, or the use of regional climate models (RCMs). The increase in computing resources has also led to the development of convection-permitting models ( [[#Prein--2015|Prein et al., 2015]] ), which have been integrated over larger domains, but are still unsuitable for CMIP simulations. Statistical downscaling tools are also widely used to generate fine-scale regional climate information necessary for climate impacts and adaptation studies. A comprehensive assessment of the added value of increased spatial resolution and of the benefits and shortcomings of statistical downscaling tools are addressed in [[IPCC:Wg1:Chapter:Chapter-10|Chapter 10]] ( [[IPCC:Wg1:Chapter:Chapter-10#10.3.3|Section 10.3.3]] ).

<div id="8.5.1.2.1" class="h4-container"></div>

<span id="high-resolution-global-climate-models"></span>
===== 8.5.1.2.1 High-resolution global climate models =====

<div id="h4-32-siblings" class="h4-siblings"></div>

Since AR5, horizontal resolution has increased in most global climate models, which has led to several improvements in the simulation of the water cycle (see also [[IPCC:Wg1:Chapter:Chapter-10#10.3.1.1|Section 10.3.1.1]] ), not only in areas with steep or complex orography, but also over the tropical oceans and within the North Pacific and North Atlantic storm tracks (Piazza et al. , 2016; Roberts et al. , 2018; Bui et al. , 2019; [[#Chen--2019|Chen and Dai, 2019]] ; Vannière et al. , 2019) . Yet, the added value of higher resolution global climate models is not systematic (Johnson et al. , 2016; Ogata et al. , 2017; D. Huang et al. , 2018; Mahajan et al. , 2018; Vannière et al. , 2019) and needs careful assessment ( [[#Haarsma--2016|Haarsma et al., 2016]] ; [[#Caldwell--2019|Caldwell et al., 2019]] ). Several AGCM studies suggest that increased spatial resolution leads to better simulation of the atmospheric moisture transport from ocean to land, the geographical distribution of annual mean precipitation ( [[#Demory--2014|Demory et al., 2014]] ), and the frequency distribution of daily precipitation intensities (L. [[#Zhang--2016|]] [[#Zhang--2016|]] [[#Zhang--2016|]] [[#Zhang--2016|Zhang et al., 2016]] ; [[#Chen--2019|Chen and Dai, 2019]] ) including extremes in many( [[#Jacob--2014|Jacob et al., 2014]] ; [[#Westra--2014|Westra et al., 2014]] ), but not all cases ( [[#Bador--2020|Bador et al., 2020]] ).

Part of the improvement in simulated precipitation accuracy is related to improved simulation of the frequency and/or mean intensity of tropical ( [[#Roberts--2015|Roberts et al., 2015]] ; [[#Walsh--2015|Walsh et al., 2015]] ) and extratropical ( [[#Hawcroft--2016|Hawcroft et al., 2016]] ) cyclones. Idealized regional experiments also show that the North Atlantic storm track response to global warming can be amplified in higher resolution models ( [[#Willison--2015|Willison et al., 2015]] ). Increased atmospheric horizontal resolution can be also important for simulating Northern Hemisphere (NH) blockings ( [[#Davini--2017|Davini et al., 2017]] ; [[#Schiemann--2017|Schiemann et al., 2017]] ) and synoptic features of the East Asian summer monsoon ( [[#Yao--2017|Yao et al., 2017]] ; [[#Kusunoki--2018|Kusunoki, 2018]] ). Variable resolution based on grid stretching may be a valuable alternative for simulating regional phenomena like monsoons (Sabin et al. , 2013; Krishnan et al. , 2016) or tropical cyclones ( [[#Harris--2016|Harris et al., 2016]] ; [[#Chauvin--2017|Chauvin et al., 2017]] ), while avoiding inconsistencies in the forcings or physics that can be found in RCMs driven by GCMs ( [[#Boé--2020|Boé et al., 2020]] ; [[#Tapiador--2020|Tapiador et al., 2020]] ).

Increasing horizontal model resolution in CMIP5 and CMIP6 models leads to a systematic increase in global mean precipitation, enhanced moisture advection to land in close connection with increased orographic precipitation, and a partial reduction of the long-standing double ITCZ bias ( [[#Demory--2014|Demory et al., 2014]] ; [[#Caldwell--2019|Caldwell et al., 2019]] ; [[#Vannière--2019|Vannière et al., 2019]] ). Recent studies based on HighResMIP simulations ( [[#Haarsma--2016|Haarsma et al., 2016]] ) confirm the added value of increased horizontal resolution (at least 50 km in the atmosphere and 25 km in the ocean) for the simulation of tropical ( [[#Roberts--2020|Roberts et al., 2020]] ) and extratropical cyclones ( [[#Priestley--2020b|Priestley et al., 2020b]] ). CMIP6 model biases in annual mean precipitation are only slightly reduced at higher resolution (Figure 3.10).

High resolution representation of the land surface is also important for simulating many features of the terrestrial water cycle, such as orographic precipitation, snow, runoff and streamflow in complex topography areas ( [[#Zhao--2015|Zhao and Li, 2015]] ). However, the added value may be easier to assess in offline rather than online land surface simulations ( [[#Döll--2016|Döll et al., 2016]] ) given the possible use of bias-corrected atmospheric forcings. Offline high-resolution GHMs are routinely used to monitor water resources or to assess the hydrological impacts of bias-adjusted global climate projections ( [[#Davie--2013|Davie et al., 2013]] ; S. [[#Huang--2017|Huang et al., 2017]] , 2018). Yet, the development and calibration of ‘hyper-resolution’ hydrological models, with gridcells of typically 100 m to 1 km, raises a number of issues given the lack of comprehensive surface or subsurface information ( [[#Bierkens--2015|Bierkens et al., 2015]] ) and the lack of coupling with the atmosphere (Berg and Sheffield, 2018a).

In summary, there is ''high confidence'' that increasing horizontal resolution in GCMs can reduce a number of systematic model errors of relevance for the water cycle, including synoptic circulation and the statistics of daily precipitation. High-resolution GCMs and GHMs provide improved representation of land surfaces, including topography, vegetation and land use change, which are required to accurately simulate changes in the terrestrial water cycle. However, there is ''low confidence'' that the higher horizontal resolution simulations currently available provide more accurate projections of the large-scale features of the water cycle.

<div id="8.5.1.2.2" class="h4-container"></div>

<span id="regional-climate-models-and-convective-permitting-models"></span>
===== 8.5.1.2.2 Regional climate models and convective-permitting models =====

<div id="h4-33-siblings" class="h4-siblings"></div>

Regional Climate Models (RCMs) are used to dynamically downscale global model simulations for a particular region (usually at a spatial resolution of the order of 10 to 50 km; see [[IPCC:Wg1:Chapter:Chapter-10#10.3.3|Section 10.3.3]] ). The AR5 reported that RCMs are useful for regions with variable topography and for small-scale phenomena. However, they inherit biases from their driving GCMs and thus may lack physical consistency with them. Since AR5, the application of RCMs has largely increased due to international model intercomparison projects such as CLARIS-LPB ( [[#Sánchez--2015|Sánchez et al., 2015]] ). Many studies have focused on present-day climatological precipitation, showing with ''high confidence'' improvements in its monthly to seasonal accumulation and spatial distribution ( [[#Dosio--2015|Dosio et al., 2015]] ; [[#Giorgi--2016|Giorgi et al., 2016]] ; [[#Bozkurt--2019|Bozkurt et al., 2019]] ; [[#Falco--2019|Falco et al., 2019]] ; [[#Di%20Virgilio--2020|Di Virgilio et al., 2020]] ), although the modelling of precipitation remains the ‘Achilles heel’ of both GCMs and RCMs and should be considered cautiously when informing regional climate change adaptation strategies ( [[#Tapiador--2019b|Tapiador et al., 2019b]] ).

Regional Convective Permitting Models (CPMs), typically run at a resolution less than 10 km, have been implemented over increasingly large domains. Compared to models with parametrized convection ( [[#8.5.1.1.1|Section 8.5.1.1.1]] ), they generally show improved simulation of key features of the water cycle such as orographic precipitation, sea breeze dynamics, the diurnal cycle in precipitation, soil-moisture–precipitation feedbacks, daily precipitation persistence, sub-daily to daily precipitation intensities and related extremes ( [[#8.2.3.2|Section 8.2.3.2]] ; Birch et al. , 2015; Prein et al. , 2015; Kendon et al. , 2017; Leutwyler et al. , 2017; Willetts et al. , 2017; [[#Hohenegger--2018|Hohenegger and Stevens, 2018]] ; Berthou et al. , 2019b; [[#Takahashi--2019|Takahashi and Polcher, 2019]] ; Fumière et al. , 2020; Scaff et al. , 2020; Caillaud et al. , 2021) . A growing number of studies have also assessed the potential added value of using CPMs for regional climate projections (Ban et al. , 2015; Giorgi et al. , 2016; Fosser et al. , 2017; Kendon et al. , 2017, 2019; C. Liu et al. , 2017; Lenderink et al. , 2019; Rasmussen et al. , 2020; see also [[IPCC:Wg1:Chapter:Atlas|Atlas]] 5.6.3) . Although projected changes in rainfall occurrence in CPMs are broadly and qualitatively consistent with the results of GCMs and RCMs ( [[#Kendon--2017|Kendon et al., 2017]] ), there is a tendency towards stronger changes in both wet and dry extremes (Berthou et al. , 2019a; Kendon et al. , 2019; Lenderink et al. , 2019; Finney et al. , 2020a) . While both GCMs and RCMs project an overall decrease in summer precipitation over the Alps, RCMs simulate an increase over the high Alpine elevations that is not present in the global simulations ( [[#Giorgi--2016|Giorgi et al., 2016]] ).

Recent studies based on both GCMs and CPMs indicate that both CAPE and convective inhibition will increase in a warmer climate ( [[#8.2.3.2|Section 8.2.3.2]] ; J. [[#Chen--2020|]] [[#Chen--2020|Chen et al., 2020]] a), consistent with a shift from moderate to less frequent but stronger convective events ( [[#Rasmussen--2020|Rasmussen et al., 2020]] ). If underestimated by models with parametrized convection, such a mechanism could explain the underestimation of both projected increase in precipitation extremes ( [[#Borodina--2017|Borodina et al., 2017]] ; [[#Yin--2018|Yin et al., 2018]] ) and land surface drying ( [[#Douville--2017|Douville and Plazzotta, 2017]] ) in the extratropics. CMIP5 models with a larger increase in extreme precipitation also exhibit larger declines or smaller increases in light to moderate events ( [[#Thackeray--2018|Thackeray et al., 2018]] ).

In summary, there is ''high confidence'' that dynamical downscaling using limited area models adds value in simulating precipitation and related water cycle processes at the regional scale, especially in complex orography areas ( [[IPCC:Wg1:Chapter:Chapter-10#10.3.3.5.1|Section 10.3.3.5.1]] ). There is ''high confidence'' that the explicit simulation of atmospheric convection can improve the representation of weather phenomena, including the life cycle of convective storms and related precipitation extremes. Even with an improved simulation of small-scale processes, there is only ''medium confidence'' that there will be an improvement in RCM-based water cycle projections as they rely on GCM boundary conditions.

<div id="8.5.2" class="h2-container"></div>

<span id="role-of-internal-variability-and-volcanic-forcing"></span>
=== 8.5.2 Role of Internal Variability and Volcanic Forcing ===

<div id="h2-18-siblings" class="h2-siblings"></div>

Beyond modelling uncertainties, internal variability and unpredictable natural forcings may also lower the degree of confidence in projected water cycle changes, especially in the near term (2021-2040) and regional-scale projections ( [[#Hawkins--2011|Hawkins and Sutton, 2011]] ; Kent et al. , 2015; Thompson et al. , 2015; Fatichi et al. , 2016; [[#McKinnon--2018|McKinnon and Deser, 2018]] ; [[#Chen--2019|Chen and Brissette, 2019]] ; Lehner et al. , 2020) . Although there is ''low confidence'' that the main modes of climate variability (Annex IV) are altered in a warmer climate (Sections 4.4.3 and 4.5.3), increasing contrast between wet and dry weather regimes ( [[#8.2.2.1|Section 8.2.2.1]] ) will amplify their influence on water cycle variability ( [[#8.4.2.9|Section 8.4.2.9]] ) and therefore contribute to uncertainties in near-term precipitation changes (Figure 8.23). The role of internal variability as source of uncertainties in regional climate projections is assessed in [[IPCC:Wg1:Chapter:Chapter-10#10.3.4.3|Section 10.3.4.3]] . Here we assess the role of internal variability in influencing water cycle projections using paleoclimate reconstructions, pre-industrial model simulations, and large single model ensembles ( [[#8.5.2.1|Section 8.5.2.1]] ). Implications for the predictability of near-term water cycle changes are specifically assessed, as they show significant but model-dependent regional hydrological fingerprints over land [[#8.5.2.2|Section 8.5.2.2]] ). The role of volcanic eruptions is also briefly assessed in terms of consequences and uncertainties in water cycle projections ( [[#8.5.2.3|Section 8.5.2.3]] ).

<div id="8.5.2.1" class="h3-container"></div>

<span id="quantification-of-water-cycle-internal-variability"></span>
==== 8.5.2.1 Quantification of Water Cycle Internal Variability ====

<div id="h3-45-siblings" class="h3-siblings"></div>

Estimating internal variability is an important challenge in the assessment of human-induced changes in the water cycle since its magnitude and range of variability can exceed the anthropogenic signal, at least at the regional scale and for near-term projections or low-emissions scenarios (Sections 4.4.1.4 and 8.4.2.9; Deser et al. , 2012; [[#Shepherd--2014|Shepherd, 2014]] ; Xie et al. , 2015; Sarojini et al. , 2016; [[#Dai--2019|Dai and Bloecker, 2019]] ; Lehner et al. , 2020) . Underestimating internal variability in models may result in the overestimation of anthropogenic climate change because the ‘noise’ in the signal-to-noise ratio is underestimated ( [[#Knutson--2018|Knutson and Zeng, 2018]] ). There is ''medium confidence'' that this underestimation affects global water cycle projections, for instance, in terms of drought persistence and severity in the south-western USA, eastern Australia, southern Africa, the Mediterranean, the southern Amazonian basin and China ( [[#Ault--2014|Ault et al., 2014]] ; [[#Cook--2018|Cook et al., 2018]] ; [[#Gu--2018|Gu et al., 2018]] ). In CMIP6 models, the uncertainty in future projections of 20-year mean precipitation changes attributable to internal variability ranges from 41% in the near term (2021–2040) to 5% in the long term (2081–2100) (Figure 8.23). For decadal-mean precipitation changes, the relative contribution of internal variability is even larger when using large ensembles ( [[#Lehner--2020|Lehner et al., 2020]] ).

Over the 20th century, CMIP5 models show a realistic magnitude of decadal precipitation variability, if not a slight overestimation in some regions ( [[#Knutson--2018|Knutson and Zeng, 2018]] ). However, the relatively short and human-influenced instrumental record limits our ability to quantify the magnitude of internal variability in the water cycle, particularly over long time scales (decadal and beyond). Global extended reanalyses ( [[IPCC:Wg1:Chapter:Chapter-1#1.5.2|Section 1.5.2]] ) have been used to derive long-term variability in the regional water cycle components ( [[#Caillouet--2017|Caillouet et al., 2017]] ), merged with historical meteorological and hydrological local observations ( [[#Bonnet--2017|Bonnet et al., 2017]] ; [[#Devers--2020|Devers et al., 2020]] ). Specific assessment of these types of methodology and related uncertainties is provided in [[IPCC:Wg1:Chapter:Chapter-10|Chapter 10]] (Sections 10.2 and 10.3). Paleoclimate archives (tree rings, corals, ice core, speleothems, lake and ocean sediments) provide extended reconstructions of key water cycle metrics and large-scale circulation features. Some studies have suggested that CMIP5 models underestimate internal variability at decadal and longer time scales, and therefore may be missing important processes in the climate system (Ault et al. , 2012, 2013; Bunde et al. , 2013; Franke et al. , 2013; Cheung et al. , 2017; Hope et al. , 2017; [[#Kravtsov--2017|Kravtsov, 2017]] ; Cassou et al. , 2018) . However, recent assessments using paleoclimate records have found that CMIP5 models are able to reproduce decadal-to-centennial variability, including the severity, persistence and spatial extent of megadroughts (Coats et al., 2015; [[#Stevenson--2015|Stevenson et al., 2015]] ; [[#PAGES%20Hydro2K%20Consortium--2017|PAGES Hydro2K Consortium, 2017]] ), once signal reddening (autocorrelation) in proxy archives is accounted for (Deeet al., 2017; [[#PAGES%20Hydro2K%20Consortium--2017|PAGES Hydro2K Consortium, 2017]] ). Implementation of proxy system models, that is, functions that transform model variables into proxy units, has reduced model–proxy disagreement, although some differences in the magnitude of internal variability remain, particularly at centennial time scales (Deeet al., 2017; [[#Parsons--2017|Parsons et al., 2017]] ). It is unclear whether remaining discrepancies represent limitations of the climate models, or limitations of the proxy system models. Therefore, there is ''medium to high confidence'' (i.e., depending on the region) that climate models do not underestimate water cycle internal variability.

The mechanisms driving internal variability in the water cycle in climate model simulations varies. While models indicate that cool SSTs in the eastern tropical Pacific (La Niña or the cool phase of the PDO) are associated with drought in south-western North America, they also show that atmospheric internal variability may be a more prominent driver (Coats et al., 2015, 2016; [[#Stevenson--2015|Stevenson et al., 2015]] ; [[#Parsons--2018|Parsons et al., 2018]] ). Simulations of the last millennium from CMIP5–PMIP3 reproduce the observed negative correlation between eastern Australian rainfall and the central equatorial Pacific SSTs with varying skill, and also display periods when the ENSO teleconnection weakens substantially for several decades (Brownet al., 2016a). Differences in simulated internal variability have been found to be responsible for the inter-model spread in predicted shifts in subtropical dry zones for a given shift in the Hadley cell (Seviour et al., 2018). CMIP5 models show that both internal variability and anthropogenic forcings are responsible for the drying over the South Atlantic Convergence Zone region, though with large uncertainties ( [[#Zilli--2021|Zilli and Carvalho, 2021]] ). Moreover, the detection of the anthropogenic forcing on the South Atlantic Convergence Zone is strongly dependent on the characterization of model internal variability ( [[#Talento--2012|Talento and Barreiro, 2012]] ).

Beyond the tropics, North Pacific decadal variability (Annex IV.2.6, 2.4.5, 3.7.6) exerts a strong modulation of extratropical ENSO teleconnections, but also influences low-frequency variability of the Walker circulation, which is underestimated by most CMIP5 models (England et al., 2014). Atlantic Multi-decadal Variability (Annex IV.2.7, 2.4.6, 3.7.7) teleconnections show a high model spread among CMIP5 models, both in terms of persistence and spatial coherence (Qasmiet al., 2017), which has potential consequences for the water cycle variability simulated over Europe. For example, internal variability will continue to play an important role in the variability of river flows over France in coming decades ( ''medium confidence'' ) (Giuntoli et al., 2013; [[#Boé--2014|Boé and Habets, 2014]] ; [[#Bonnet--2017|Bonnet et al., 2017]] ).

Ensembles of atmosphere-only simulations driven by observed or reconstructed SST are useful for evaluating the ability of models to capture the circulation and/or precipitation variability observed over the historical period ( [[#Zhou--2016|Zhou et al., 2016]] ; [[#Deng--2018|Deng et al., 2018]] ; [[#Douville--2019|Douville et al., 2019]] ). However, limitations of such AGCM-based attribution methods, that is, related to the lack of air–sea interactions in the response, may lead to erroneous attribution conclusions in some regions for local circulation and mean and extreme precipitation ( [[#Dong--2017|Dong et al., 2017]] ). Other methods to measure the portion of precipitation variability include the partitioning into dynamical as opposed to thermodynamical components ( [[#Saffioti--2016|Saffioti et al., 2016]] ; [[#Fereday--2018|Fereday et al., 2018]] ; [[#Lehner--2018|Lehner et al., 2018]] ), the analysis of variance ( [[#Dong--2018a|Dong et al., 2018a]] ) and direct characterization of stochastic weather-noise ( [[#Short%20Gianotti--2014|Short Gianotti et al., 2014]] ).

Single-model initial condition large ensembles (SMILEs) are a powerful tool for estimating the magnitude of internal variability in historical and future climates ( [[IPCC:Wg1:Chapter:Chapter-1#1.4.4|Section 1.4.4]] ). Using SMILEs, it has been shown, for example, that internal NAO variability imparts substantial uncertainty to future changes in European precipitation (Figure 8.24; [[#Deser--2017|Deser et al., 2017]] ). For the South Asian summer monsoon, internal variability can overshadow the forced monsoon rainfall trend, thereby increasing near-term projection uncertainties (X. [[#Huang--2020|Huang et al., 2020]] a). Specific regional applications of the use of large ensembles are further assessed in Sections 10.3.4.3 and 10.3.4.4.

<div id="_idContainer068" class="Basic-Text-Frame"></div>

[[File:752cba23648b8c84e80e7f016363d6c2 IPCC_AR6_WGI_Figure_8_24.png]]

'''Figure 8.24 |''' '''Impact of the North Atlantic Oscillation (NAO) on''' '''2016–2045''' '''climate trends.''' '''(a)''' Regressions of winter sea level pressure (SLP) and precipitation trends upon the normalized leading principal component (PC) of winter SLP trends in the CESM1 Large Ensemble, multiplied by two to correspond to a two standard deviation anomaly of the PC (as internal climate variability component); '''(b)''' CESM1 ensemble-mean winter SLP and precipitation trends (as forced climate variability component); '''(c)''' b – a (forced minus internal climate variability component); '''(d)''' b + a (forced plus internal climate variability component). Precipitation in colour shading (mm day <sup>–1</sup> per 30 years) and SLP in contours (interval = 1 hPa per 30 years with negative values dashed). Figure adapted from [[#Deser--2017|Deser et al. (2017)]] , https://creativecommons.org/licenses/by/4.0/ ; further details on data sources and processing are available in the chapter data table (Table 8.SM.1).

Since AR5, SMILEs have helped quantify the time of emergence of climate change signals (see Sections 1.4.2.2 and 10.4.3). Results from SMILEs indicate that by 2000–2009 (compared to 1950–1999), simulated anthropogenic shifts in mean annual precipitation already emerged over 36–41% of the globe including high latitudes ( [[#Frankcombe--2018|Frankcombe et al., 2018]] ; [[#Kumar--2018|Kumar and Ganguly, 2018]] ), the eastern subtropical oceans, and the tropics (Zhang andDelworth, 2018). By 2050 (2100), more than 60% (85%) of the globe is projected to show detectable anthropogenic shifts in mean annual precipitation (Zhang andDelworth, 2018). Other SMILE results for the 1950–2100 period ( [[#Kay--2015|Kay et al., 2015]] ; [[#Sigmond--2016|Sigmond and Fyfe, 2016]] ) indicate that internal variability can obscure the detection of the anthropogenic hydroclimatic signal until the middle to late 21st century in many parts of the world for both mean and extreme precipitation ( [[#Martel--2018|Martel et al., 2018]] ; [[#Dai--2019|Dai and Bloecker, 2019]] ). A common finding is that changes in the characteristics of wet extreme events will emerge earlier than changes in average conditions ( [[#Gaetani--2020|Gaetani et al., 2020]] ; [[#Hawkins--2020|Hawkins et al., 2020]] ; [[#Kusunoki--2020|Kusunoki et al., 2020]] ). An assessment of the methods used to estimate time of emergence is presented in [[IPCC:Wg1:Chapter:Chapter-10|Chapter 10]] ( [[IPCC:Wg1:Chapter:Chapter-10#10.3.4.3|Section 10.3.4.3]] ). For specific regional examples of climate change attribution and emergence of anthropogenic signal, see [[IPCC:Wg1:Chapter:Chapter-10#10.4.2|Section 10.4.2]] .

In summary, there is ''medium confidence'' that climate models reproduce the general magnitude and character of internal variability that influences water cycle variables. There is ''high confidence'' that internal variability will continue to be a major source of uncertainty, at least for near-term water cycle projections at the regional scale. There is ''low confidence'' in the region-dependent time of emergence of water cycle changes (see also [[IPCC:Wg1:Chapter:Chapter-10#10.4.3|Section 10.4.3]] ), but there is ''medium confidence'' that changes in wet extreme events will emerge earlier than changes in average conditions.

<div id="8.5.2.2" class="h3-container"></div>

<span id="implications-for-near-term-water-cycle-projections"></span>
==== 8.5.2.2 Implications for Near-Term Water Cycle Projections ====

<div id="h3-46-siblings" class="h3-siblings"></div>

Adapting water resource management in the face of climate change will greatly benefit from improved prediction of land surface hydrology at the decadal time scale. Climate predictions ( [[IPCC:Wg1:Chapter:Chapter-1#1.4.4|Section 1.4.4]] ) differ from climate projections by constraining the initial state of the slow components of the climate system (i.e., the ocean, the cryosphere and the terrestrial hydrology) as well as volcanic aerosols and ozone depleting substances with observations. Anthropogenic and natural radiative forcing and low-frequency modes of variability (e.g., AMV and PDV, Annex IV.2.7 and IV.2.6) suggest the possible predictability of climate in the first decade or so of the 21st century, in addition to the projected response to the anthropogenic forcing (Sections 4.2.3 and 4.4.1.3).

In AR5, decadal prediction of precipitation over some land areas showed improved skill due to specified radiative forcing, with almost no added value from ocean initialization. Since AR5, more studies have been devoted to understanding the potential or effective water cycle predictability related to ocean multi-decadal variability. Decadal hindcast experiments based on large ensembles highlight increasing skill scores in annual mean precipitation three to seven years ahead, at least over the Sahel and Europe ( [[#Yeager--2018|Yeager et al., 2018]] ). There is relatively high predictability of the AMV impacts over the Mediterranean basin, Central Asia and the Americas (from the USA to northern South America) during boreal summer, but in boreal winter the signal-to-noise ratio shows only weak predictability over land ( [[#Yamamoto--2016|Yamamoto and Palter, 2016]] ; [[#Ruprich-Robert--2017|Ruprich-Robert et al., 2017]] ). The link between South Asian summer monsoon changes and the AMOC and the decadal variability in the Pacific Ocean open the possibility of increased predictability for the near future (Kushnir et al. , 2017; X. Huang et al. , 2020b; Sandeep et al., 2020) .

The additional skill associated with the initialization of the cryosphere and the land surface has received limited attention. However, there is observational evidence that oceanic decadal variations can propagate into the atmosphere and, consequently accumulate into terrestrial land surface reservoirs(e.g., [[#Bonnet--2017|Bonnet et al., 2017]] ) and vegetation (e.g., [[#Zeng--1999|Zeng et al., 1999]] ). This land surface memory, like in soil moisture ( [[#Alessandri--2008|Alessandri and Navarra, 2008]] ; [[#Catalano--2016|Catalano et al., 2016]] ) or snow( [[#Loranty--2014|Loranty et al., 2014]] ), may also contribute to the decadal predictability of the terrestrial component of the water cycle, but remains difficult to assess given the limitations of observational records. Vegetation initialization seems to generate as much noise as signal and does not necessarily translate into improved skill in early decadal predictions based on ESMs ( [[#Weiss--2014|Weiss et al., 2014]] ).

Decadal hydrological predictability in an idealized setting has also been investigated through offline land surface hindcast experiments, driven by observed atmospheric forcing and/or initial conditions, suggesting the potential for skilful predictions for terrestrial water storage, deep soil moisture, and groundwater ( [[#Yuan--2018|Yuan and Zhu, 2018]] ). Yet, a real-world assessment is hampered by the lack of observations and is only feasible when multi-decadal records of satellite estimates of terrestrial water storage, snow mass or soil moisture are available.

In summary, there is ''high confidence'' that the water cycle changes that have already emerged from internal variability will become more pronounced in near-term (2021–2040) projections. However, there is ''low confidence'' in decadal predictions of precipitation changes, particularly over most land areas, because internal variability remains difficult to predict and can offset or amplify the forced water cycle response.

<div id="8.5.2.3" class="h3-container"></div>

<span id="volcanic-forcing"></span>
==== 8.5.2.3 Volcanic Forcing ====

<div id="h3-47-siblings" class="h3-siblings"></div>

Volcanic eruptions can affect climate projections in the near term (2021–2040; [[IPCC:Wg1:Chapter:Chapter-4#4.4.4|Section 4.4.4]] and Cross-Chapter Box 4.1). In this chapter, they are of interest because they can trigger a transient departure from the water cycle response to anthropogenic radiative forcing. Major volcanic eruptions temporarily reduce total global and wet tropical region precipitation ( ''high confidence'' ) ( [[#Iles--2014|Iles and Hegerl, 2014]] ), can weaken or shift the ITCZ ( [[#Iles--2014|Iles and Hegerl, 2014]] ; [[#Colose--2016|Colose et al., 2016]] ; [[#Liu--2016|Liu et al., 2016]] ), and reduce summer monsoon rainfall ( ''medium confidence'' ) (Pausata et al. , 2015b; [[#Zambri--2016|Zambri and Robock, 2016]] ; Zambri et al. , 2017; Zuo et al. , 2019; M. Singh et al. , 2020) . Monsoon precipitation in one hemisphere can be enhanced by the remote volcanic forcing occurring in the other hemisphere ( ''medium confidence'' ) (Pausata et al. , 2015a; Liu et al. , 2016; Zuo et al. , 2019) . Over the Sahel, the sign of hydrological changes depend on the hemisphere where the volcanic eruptions occur (J.M. [[#Haywood--2013|]] [[#Haywood--2013|Haywood et al., 2013]] ). Out of phase changes in the Sahel and the Amazonian basin are expected from the effect of volcanic aerosols on tropical Atlantic SST and the ITCZ ( [[#Hua--2019|Hua et al., 2019]] ). Over the last millennium, uncertainties remain in the symmetry/asymmetry of the monsoon response because it is difficult to estimate the exact latitude and season of past volcanic eruptions further back in time ( [[#Colose--2016|Colose et al., 2016]] ; [[#Fasullo--2019|Fasullo et al., 2019]] ).

Data for six major eruptions over the last century along with CMIP5 historical experiments indicate that volcanic eruptions cause a detectable decrease in streamflow in northern South America, Central Africa, high-latitude Asia and in wet tropical–subtropical regions, and a detectable increase in south-western North America and southern South America ( [[#Iles--2015|Iles and Hegerl, 2015]] ). Attempts to include volcanic forcing in future projections show enhanced precipitation variability on annual to decadal time scales with small reductions in Asian monsoon rainfall ( [[#Bethke--2017|Bethke et al., 2017]] ). The occurrence of volcanic eruptions in the coming century, either as single large events or clustered smaller ones, can alter the water cycle (see also Cross-Chapter Box 4.1), and regional drought events may be enhanced by co-occurring volcanic ( [[#Liu--2016|Liu et al., 2016]] ; [[#Gao--2017|Gao and Gao, 2017]] ; [[#Zambri--2017|Zambri et al., 2017]] ) and GHG (e.g., [[#Cook--2018|Cook et al., 2018]] ) forcing ( ''low confidence'' ). Volcanic eruptions may also lead to widespread precipitation anomalies up to several years following an eruption through their potential influence on the El Niño Southern Oscillation ( ''low confidence'' ) ( [[#Stevenson--2016|Stevenson et al., 2016]] ; [[#Dee--2020|Dee et al., 2020]] ; [[#McGregor--2020|McGregor et al., 2020]] ).

In summary, large volcanic eruptions reduce global mean precipitation, as well as precipitation in tropical wet regions ( ''high confidence'' ). There is ''low confidence'' in specific regional and seasonal responses, primarily due to the limitations of the observational record.

<div id="8.5.3" class="h2-container"></div>

<span id="non-linearities-across-global-warming-levels"></span>
=== 8.5.3 Non-linearities Across Global Warming Levels ===

<div id="h2-19-siblings" class="h2-siblings"></div>

The AR5 concluded that annual and seasonal mean precipitation changes can be estimated by linear pattern-scaling techniques ( [[#Santer--1990|Santer and Wigley, 1990]] ; [[#Arnell--2016|Arnell and Gosling, 2016]] ; [[#Greve--2018|Greve et al., 2018]] ), which represent regional changes in precipitation as a linear function of global mean temperature change. However, there are a number of caveats when pattern-scaling is applied to low-emissions scenarios or to scenarios where localized forcing (e.g., anthropogenic aerosols) are significant and vary in time ( [[#Collins--2013|Collins et al., 2013]] ). Here the focus is in on non-linear water cycle responses to increasing global warming levels, as estimated for instance from the difference between the first 2°C of global warming, and the next 2°C of warming (Figure 8.25), and their possible underlying mechanisms.

<div id="_idContainer070" class="Basic-Text-Frame"></div>

[[File:9324edb0441675e3d1c9de729b4087f7 IPCC_AR6_WGI_Figure_8_25.png]]

'''Figure 8.25 |''' '''Effect of first versus second 2°C of global warming relative to the 1850–1900 base period on seasonal mean precipitation (mm day''' ''–1'' ''').''' CMIP6 multi-model ensemble mean December–January–February (left panels) and June–July–August (right panels) precipitation difference for '''(a, b)''' SSP5-8.5 at +2°C '''(c, d)''' SSP5-8.5 at +4°C minus SSP5-8.5 at +2°C (second 2°C warming); '''(e, f)''' second minus first 2°C fast warming (c–a and d–b). Only models reaching the +4°C warming levels in SSP5-8.5 are considered. Differences are computed based on 21-year time windows centred on the first year reaching or exceeding the selected global warming level using a 21-year running mean global surface atmospheric temperature criterion. Uncertainty is represented using the simple approach. No overlay indicates regions with high model agreement, where ≥80% of models agree on sign of change. Diagonal lines indicate regions with low model agreement, where &lt;80% of models agree on sign of change. For more information on the simple approach, please refer to the Cross-Chapter Box Atlas.1. Further details on data sources and processing are available in the chapter data table (Table 8.SM.1).

<div id="8.5.3.1" class="h3-container"></div>

<span id="non-linearities-in-large-scale-atmospheric-circulation-and-precipitation"></span>
==== 8.5.3.1 Non-linearities in Large-scale Atmospheric Circulation and Precipitation ====

<div id="h3-48-siblings" class="h3-siblings"></div>

Since AR5, there is further evidence that the pattern-scaling technique has limitations ( [[#Lopez--2014|Lopez et al., 2014]] ; [[#Wartenburger--2017|Wartenburger et al., 2017]] ; [[#Tachiiri--2019|Tachiiri et al., 2019]] ), and that alternative approaches, such as multiple regressions using the land–sea warming contrast as an additional predictor, offer added value ( [[#Joshi--2013|Joshi et al., 2013]] ). The simplest traditional pattern-scaling approach approximates future changes by the product of a time-evolving global surface temperature change and a pattern that varies spatially but is constant across time, scenarios, and models. This technique was shown to be more robust across scenarios rather than across models, with better results for temperature compared with precipitation ( [[#Tebaldi--2014|Tebaldi and Arblaster, 2014]] ; see also [[IPCC:Wg1:Chapter:Chapter-4#4.2.4|Section 4.2.4]] ). One approach which avoids scaling is to consider a period in a different scenario with the same global surface temperature change ( [[#Herger--2015|Herger et al., 2015]] ). It is attractive as it provides patterns of any temporal resolution that are consistent across variables. Nonetheless, this technique is still only based on global surface temperature and is not necessarily suitable for precipitation changes projected in stabilized versus transient scenarios (at the same global warming level) given the fast-atmospheric adjustment to GHG radiative forcing (Sections 8.2.1 and 8.4.1.1).

Even in a theoretical climate system governed by linear processes, pattern-scaling assumptions can fail because the different forcing time response of different parts of the Earth system cause evolving spatial warming patterns ( [[#Good--2016a|Good et al., 2016a]] ). This occurs primarily because different feedbacks occur at different time scales ( [[#Armour--2013|Armour et al., 2013]] ; [[#Andrews--2015|Andrews et al., 2015]] ), which in turn implies that the atmospheric circulation and water cycle is dependent both on the level of warming and the rate of change ( [[#Ceppi--2018|Ceppi et al., 2018]] ). The usual distinction between the fast adjustment to increased GHG concentrations and the slower response to SST warming ( [[#8.2.2.2|Section 8.2.2.2]] ) may, however, not be sufficient to explain the time evolution of the hydroclimatic response at the regional scale, especially in subtropical land areas where this response critically depends on shifts in atmospheric circulation associated with distinct ‘fast’ (typically five to ten years, that is however much slower than the atmospheric adjustment assessed in [[#8.2.1|Section 8.2.1]] ) and slow SST warming patterns ( [[#Zappa--2020|Zappa et al., 2020]] ). The changing balance between the water cycle response to anthropogenic GHG and aerosol forcings is another source of non-linearity across time and global warming levels (Ishizaki et al. , 2013; Rowell et al. , 2015; Y. Liu et al. , 2019b; Wilcox et al., 2020) .

Non-linearities in the climate response are thought to arise from multiple factors. These include state-dependent ice-albedo feedback and its potential influence on Northern Hemisphere (NH) storm tracks (Peings and Magnusdottir, 2014; [[#Semenov--2015|Semenov and Latif, 2015]] ; see also Cross-Chapter Box 10.1 and [[#8.6.1|Section 8.6.1.2]] ); a state-dependent sensitivity of tropical precipitation to increased SST ( [[#Schewe--2017|Schewe and Levermann, 2017]] ; [[#He--2018|He et al., 2018]] ); a complex response of the Atlantic meridional overturning circulation (AMOC; Sections 9.2.4.1 and 8.6.1.1) and its model- and magnitude dependent teleconnections with regional temperature and precipitation (Kageyama et al., 2013; [[#Jackson--2015|Jackson et al., 2015]] ; [[#Qasmi--2017|Qasmi et al., 2017]] , 2020); and other atmospheric and terrestrial ( [[#8.5.3.2|Section 8.5.3.2]] ) processes such as cloud and land surface feedbacks (Ceppi and Gregory, 2017; [[#King--2019|King, 2019]] ). The response of convective precipitation may exhibit non-linearities because it is itself modulated by both dynamics and atmospheric water content, each responding independently to warming (Chadwick and Good, 2013; [[#Neupane--2013|Neupane and Cook, 2013]] ).

Based on a simple model, it was also suggested that the Indian summer monsoon may exhibit a moisture-advection feedback which allows multiple stable states as boundary conditions change (Zickfeld et al., 2005). However, limitations of this theory and comprehensive GCMs suggest a near-linear monsoon response to a broad range of radiative forcings (Boos and Storelvmo, 2016). Non-linear precipitation responses to global warming have been reported in the Indo-Pacific, where a linear increase in SSTs can trigger non-linear changes in precipitation and a shift in the ITCZ depending on the relative amplitudes of uniform and structured SST anomalies ( C.T.Y. [[#Chung--2014|]] [[#Chung--2014|Chung et al., 2014]] ; [[#Toda--2018|Toda and Watanabe, 2018]] ).

Compared to atmospheric circulation and seasonal mean precipitation, extreme precipitation has been found to scale more accurately with local and global mean temperature (Chou et al., 2012; [[#Pendergrass--2015|Pendergrass et al., 2015]] ). The projected increase in the magnitude of extreme precipitation is generally proportional to the global warming level, with an increase of around 7% per 1°C warming ( [[IPCC:Wg1:Chapter:Chapter-11#11.4.5|Section 11.4.5]] ) although this rate shows seasonal and geographical variations and is slightly less for five-day than for one-day precipitation maxima. Projected changes in extreme precipitation are the result of both thermodynamical and more model-dependent and potentially less linear dynamical contributions ( [[#Pfahl--2017|Pfahl et al., 2017]] ). Projected changes in precipitation extremes are also potentially sensitive to a non-linear response of spatial convective organization (Pendergrass et al., 2016), and can exhibit a quadratic rather than linear response to global warming (Pendergrass et al., 2019).

Within CMIP6, the linearity to CO <sub>2</sub> forcing can be assessed through the comparison of the model response to abrupt doubling compared with abrupt quadrupling of atmospheric CO <sub>2</sub> ( [[#Webb--2017|Webb et al., 2017]] ). Preliminary analyses based on CMIP5 models showed that annual precipitation changes following a doubling step change in CO <sub>2</sub> from pre-industrial levels are not necessarily consistent with the response to the step from doubling to quadrupling despite a similar change in radiative forcings ( [[#Good--2016a|Good et al., 2016a]] ; [[#Ceppi--2017|Ceppi and Shepherd, 2017]] ). Beyond the visual comparison of the climate response at various global warming levels (e.g., Figure 4.35), the linearity across global warming levels can be assessed by using the highest emissions scenario and comparing seasonal mean relative precipitation changes at +2°C versus +4°C above pre-industrial (1850–1900) temperatures (Figure 8.25). The results support the previous finding ( [[#Good--2016b|Good et al., 2016b]] ) that a second 2°C warming does not necessarily lead to the same precipitation anomaly pattern as the first 2°C, especially in the tropics where regional differences can be large but not necessarily consistent among different models. They are also consistent with a recent analysis of CMIP5 models showing that the projected drying in the Mediterranean and in Chile is substantially faster than the increase in GSAT, and therefore does not scale linearly with global warming ( [[#Zappa--2020|Zappa et al., 2020]] ).

In summary, there is ''high confidence'' that continued global warming will further amplify GHG-induced changes in large-scale atmospheric circulation and precipitation. Nonetheless, there are cases where regional water cycle changes are not linearly related to global warming due to the interaction of multiple forcings, feedbacks and time scales ( ''medium confidence'' , see also Sections 4.2.4, 7.4.3 and 8.2.1). Aridity in subtropical regions is highly sensitive to fast shifts in large-scale atmospheric circulation so are particularly susceptible to such non-linearities.

<div id="8.5.3.2" class="h3-container"></div>

<span id="non-linearities-in-land-surface-processes-and-feedbacks"></span>
==== 8.5.3.2 Non-linearities in Land Surface Processes and Feedbacks ====

<div id="h3-49-siblings" class="h3-siblings"></div>

Land surface responses and feedbacks represent a potential source of non-linearity for the water cycle response, at least at regional and local scales. The forced response of soil moisture and freshwater resources not only depends on precipitation, but also on evaporation ( [[#Laîné--2014|Laîné et al., 2014]] ), snowmelt ( [[#Thackeray--2016|Thackeray et al., 2016]] ), and runoff(X. [[#Zhang--2018|]] [[#Zhang--2018|]] [[#Zhang--2018|]] [[#Zhang--2018|]] [[#Zhang--2018|Zhang et al., 2018]] ) which are intrinsically non-linear processes depending on soil moisture or temperature thresholds. Bare ground evaporation is, for instance, usually estimated as a non-linear function of surface soil moisture ( [[#Jefferson--2015|Jefferson and Maxwell, 2015]] ). Plant transpiration requires more complex formulations with non-linear dependencies on multiple environmental factors including root-zone soil moisture and atmospheric CO <sub>2</sub> concentration ( [[#Franks--2017|Franks et al., 2017]] ). Globally, land surface evaporation is both energy and soil-moisture limited, but one of these limitations can become dominant depending on regions and seasons. Non-linearities may be particularly strong in transitional regimes where and when soil moisture limitation plays a major role ( [[#Berg--2018b|Berg and Sheffield, 2018b]] ).

Snowmelt is a nonl-inear process and projected changes in snowfall are also a non-linear combination of changes in total precipitation and in the fraction of solid precipitation. In cold regions, snowfall may first increase because of the increased water capacity of a warmer atmosphere and then decrease because snow falls as rain in an even warmer atmosphere. Such non-linearities can contribute to elevation, latitudinal and seasonal contrasts in the observed and projected retreat of the Northern Hemisphere (NH) snow cover ( [[#Shi--2015|Shi and Wang, 2015]] ; [[#Thackeray--2016|Thackeray et al., 2016]] ). Mountain glaciers also represent source of non-linear runoff responses since the annual runoff can first increase due to additional melting and then decrease as the glaciers shrink ( [[#Kraaijenbrink--2017|Kraaijenbrink et al., 2017]] ; [[#Shannon--2019|Shannon et al., 2019]] ). [[IPCC:Wg1:Chapter:Chapter-9#9.5.1.3|Section 9.5.1.3]] concludes with ''high confidence'' that the average annual runoff from glaciers will generally reach a peak at the latest by the end of the 21st century, and decline thereafter. This peak may have already occurred for small catchments with little ice cover, but tends to occur later in basins with large glaciers. Permafrost thawing is another mechanism which can trigger a non-linear hydrological response in the high latitudes of the NH( [[#Walvoord--2016|Walvoord and Kurylyk, 2016]] ), whose magnitude and potential abruptness is assessed in [[IPCC:Wg1:Chapter:Chapter-5#5.4.3.3|Section 5.4.3.3]] .

Land surface runoff and groundwater recharge are highly non-linear process, depending for instance on rainfall intensity, soil infiltration capacity, vertical profile of soil moisture and water table depth. A non-linear relationship between rainfall and groundwater recharge was observed in the tropics where intense seasonal rainfalls associated with internal climate variability contribute disproportionately to recharge (R.G. Taylor et al. , 2013a; Cuthbert et al. , 2019a) . Groundwater fluxes in arid regions are generally less responsive to climate variability than in humid regions, which can temporarily buffer climate change impacts on water resources or lead to a long, initially hidden, hydrological responses to global warming ( [[#Cuthbert--2019a|Cuthbert et al., 2019a]] ). Hydrological model simulations driven by individual and combined forcing show that decreased precipitation can cause larger deficits in soil moisture, streamflow and water table depth than other forcings, but also that these factors are not linearly cumulative when applied in combination ( [[#Hein--2019|Hein et al., 2019]] ). Surface runoff was found to scale only approximately with global warming ( [[#Tanaka--2017|Tanaka et al., 2017]] ). Significant non-linearities were found in the projected annual mean runoff response to global warming in CMIP5 projections, which could not be entirely explained by precipitation changes(X. [[#Zhang--2018|]] [[#Zhang--2018|]] [[#Zhang--2018|]] [[#Zhang--2018|]] [[#Zhang--2018|Zhang et al., 2018]] ). Similar non-linear behaviours are found in CMIP6 models over the Amazon, Yangtze, Niger, Euphrates and Mississippi river basins (Figure 8.26), highlighting the need to reassess the assumption of linearity when estimating regional water cycle changes.

<div id="_idContainer072" class="Basic-Text-Frame"></div>

[[File:06341e23b271a4184c52705376bb9afe IPCC_AR6_WGI_Figure_8_26.png]]

'''Figure 8.26 |''' '''Rate of change in basin-scale annual mean runoff with increasing global warming levels.''' Relative changes (%) in basin-averaged annual mean runoff estimated as multi-model ensemble median from a variable subset of CMIP6 models for each SSP over nine major river basins: '''(a)''' Mississippi, '''(b)''' Danube, '''(c)''' Lena, '''(d)''' Amazon, '''(e)''' Euphrates, '''(f)''' Yangtze, '''(g)''' Niger, '''(h)''' Indus, and '''(i)''' Murray. The basin averages have been estimated after a first-order conservative remapping of the model outputs on the 0.5° by 0.5° river network of [[#Decharme--2019|Decharme et al. (2019)]] . The shaded area indicates the 5–95% confidence interval of the ensemble values across all SSPs. Note that the y-axis range differs across basins and is particularly large for Niger and Murray (panels g and i). The number of models considered is specified for each scenario in the legend located inside panel b. Further details on data sources and processing are available in the chapter data table (Table 8.SM.1).

Beyond changes in land surface water fluxes, non-linearities in the response of soil moisture and freshwater reservoirs have not been well documented in global climate projections but deserve further attention given the complex interactions between the water, energy and carbon cycles ( [[#Berg--2018a|Berg and Sheffield, 2018a]] ), the growing direct human influence on rivers and groundwater ( [[#Abbott--2019|Abbott et al., 2019]] ), and a possible offset between the linear components of changes in precipitation and evapotranspiration. Significant non-linearities were found in water scarcity projections, as seen by the stronger sensitivity to the first 2°C increase in global warming ( [[#Gosling--2016|Gosling and Arnell, 2016]] ).

In summary, there is both numerical and process-based evidence that terrestrial water cycle changes can be non-linear at the regional scale ( ''high confidence'' ). Non-linear regional responses of runoff, groundwater recharge and water scarcity have been documented based on both CMIP5 and CMIP6 models, and highlight the limitations of simple pattern-scaling techniques ( ''medium confidence'' ). Water resources fed by melting glaciers are particularly exposed to such non-linearities ( ''high co'' ''nfidence'' ).

<div id="8.6" class="h1-container"></div>

<span id="what-is-the-potential-for-abrupt-change"></span>