Editing IPCC:AR6/WGI/Chapter-10 (section)

=== 10.3.4 Managing Uncertainties in Regional Climate Projections ===

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

Regional climate projections are affected by three main sources of uncertainty (Sections 10.2.2, 1.4.3 and 4.2.5): unknown future external forcings, imperfect knowledge and implementation of the response of the climate system to external forcings, and internal variability ( [[#Lehner--2020|Lehner et al., 2020]] ). In a regional downscaling context, uncertainties arise in every step of the modelling chain. Here the propagation of uncertainties ( [[#10.3.4.1|Section 10.3.4.1]] ), the management of uncertainties ( [[#10.3.4.2|Section 10.3.4.2]] ), the role of the internal variability for regional projections ( [[#10.3.4.3|Section 10.3.4.3]] ), and the design and use of ensembles to account for uncertainties ( [[#10.3.4.4|Section 10.3.4.4]] ) will be assessed. Observational uncertainty, in particular for the calibration of statistical downscaling methods ( [[#10.2.3.1|Section 10.2.3.1]] ), also contributes to projection uncertainty.

<div id="10.3.4.1" class="h3-container"></div>

<span id="propagation-of-uncertainties"></span>
==== 10.3.4.1 Propagation of Uncertainties ====

<div id="h3-34-siblings" class="h3-siblings"></div>

Modelling chains for generating regional climate information range from the definition of forcing scenarios to the global modelling, and potentially to dynamical or statistical downscaling and bias adjustment ( [[#10.3.1|Section 10.3.1]] ). The propagation and potential accumulation of uncertainties along the chain has been termed the cascade of uncertainty ( [[#Wilby--2010|Wilby and Dessai, 2010]] ). Even within one model, like a global model, uncertainty propagates across scales. From a process point of view, these uncertainties are related to forcings and global climate sensitivity, and errors in the representation of the large-scale circulation ( [[#10.3.3.3|Section 10.3.3.3]] ; [[#McNeall--2016|McNeall et al., 2016]] ) and regional processes ( [[#10.3.3.4|Section 10.3.3.4]] ), feedbacks ( [[#10.3.3.5|Section 10.3.3.5]] ) and drivers ( [[#10.3.3.6|Section 10.3.3.6]] ). From a modelling point of view, these uncertainties are related to the choice of dynamical and statistical models ( [[#10.3.1|Section 10.3.1]] ) and experimental design ( [[#10.3.2|Section 10.3.2]] ). The overall uncertainty can be statistically decomposed into the individual sources ( [[#Evin--2019|Evin et al., 2019]] ; [[#Christensen--2020|Christensen and Kjellström, 2020]] ), although there might be non-linear dependencies between them.

Uncertainty propagation often increases the spread in regional climate projections when comparing global model and downscaled results, which has been used as an argument against top-down approaches to climate information ( [[#Prudhomme--2010|Prudhomme et al., 2010]] ). Increased spread in the modelling chain may also arise from a more comprehensive representation of previously unknown or underrepresented uncertainties ( [[#Maraun--2018b|Maraun and Widmann, 2018b]] ). The increased spread in this case goes together with a better representation of processes and thus an increased model fitness-for-purpose ( [[#10.3.3.9|Section 10.3.3.9]] ).

<div id="10.3.4.2" class="h3-container"></div>

<span id="representing-and-reducing-uncertainties"></span>
==== 10.3.4.2 Representing and Reducing Uncertainties ====

<div id="h3-35-siblings" class="h3-siblings"></div>

Climate response uncertainties (Chapter 1) can be represented by multi-model ensembles, although the sampled uncertainty typically underestimates the full range of uncertainty ( [[#Collins--2013b|Collins et al., 2013b]] ; [[#Shepherd--2018|Shepherd et al., 2018]] ; [[#Almazroui--2021|Almazroui et al., 2021]] ). Traditionally, climate response uncertainty has been characterized by the ensemble spread around the multi-model mean change. The change has then further been qualified in terms of the agreement across models and compared to estimates of internal climate variability ( [[#Collins--2013b|Collins et al., 2013b]] ). Since AR5, several limitations of this approach have been identified ( [[#Madsen--2017|Madsen et al., 2017]] ) such as the failure to address physically plausible, but low-likelihood, high-impact scenarios (Chapters 1, 4, 8 and 9; [[#Sutton--2018|Sutton, 2018]] ) or that qualitatively different or even opposite changes may be equally plausible at the regional scale ( [[#Shepherd--2014|Shepherd, 2014]] ). In a multi-model mean these different responses would be lumped together, strongly dampened, and qualified as non-robust, whereas in fact high impacts might occur. Further, the multi-model mean itself is often implausible because it is a statistical construct ( [[#Zappa--2017|Zappa and Shepherd, 2017]] ). Overall, there is ''high confidence'' that some regional future climate changes are not well-characterized by multi-model mean and spread.

Since AR5, physical climate storyline approaches (see also Chapter 1, [[#10.5.3|Section 10.5.3]] , Box 10.2, and Atlas.2.5.2) have been developed to better characterize and communicate uncertainties in regional climate projections ( [[#Shepherd--2019|Shepherd, 2019]] ). A special class of such storylines attempts to attribute regional uncertainties to uncertainties in remote drivers. For instance, the Dutch Meteorological Service has presented climate projections for the Netherlands for different plausible changes of the mid-latitude atmospheric circulation and different levels of European warming ( [[#van%20den%20Hurk--2014|van den Hurk et al., 2014]] ). [[#Manzini--2014|Manzini et al. (2014)]] have quantified the impact of uncertainties in tropical upper troposphere warming, polar amplification, and stratospheric wind change on Northern Hemisphere winter climate change. Based on these results, [[#Zappa--2017|Zappa and Shepherd (2017)]] separated the multi-model ensemble into physically consistent sub-groups or storylines of qualitatively different projections in relevant remote drivers of the atmospheric circulation. In a similar vein, ( [[#Ose--2020|Ose et al., 2020]] ) trace uncertainties in projections of the East Asian summer monsoon and [[#Mindlin--2020|Mindlin et al. (2020)]] conditioned the response of Southern Hemisphere mid-latitude circulation and precipitation to greenhouse gas forcing on large-scale climate indicators ( [[IPCC:Wg1:Chapter:Chapter-8#8.4.2.9.2|Section 8.4.2.9.2]] ).

These physical climate storylines help to physically explain contradicting regional projections and thus make the conveyed information a better representation of the true uncertainty ( [[#Hewitson--2014a|Hewitson et al., 2014a]] ). Additionally, the attribution of regional uncertainties to drivers may in principle help reduce uncertainty in the case where some storylines can be ruled out because the projected changes in the driving processes appear to be physically implausible ( [[#Zappa--2017|Zappa and Shepherd, 2017]] ). There is thus ''high confidence'' that storylines attributing uncertainties in regional projections to uncertainties in changes of remote drivers aid the interpretation of uncertainties in climate projections.

Another approach that has continued to develop for characterising and reducing projection uncertainties is the use of emergent constraints (Chapters 1, 4, 5 and 7; [[#Hall--2019|Hall et al., 2019]] ). The idea is to link the spread in climate model projections via regression to the spread in present climate model biases for relevant driving processes. Models with lower biases are assigned higher weight in the projections, which in turn reduces the spread of the projections in a physical way and may additionally reduce projection uncertainty. For instance, [[#Simpson--2016|Simpson et al. (2016)]] have reduced the spread in projections of North American winter hydroclimate by linking this spread to model biases in the representation of relevant stationary wave patterns. Other examples of using emergent constraints in a regional context are Brown et al. (2016), G. [[#Li--2017|]] [[#Li--2017|Li et al. (2017)]] , [[#Giannini--2019|Giannini and Kaplan (2019)]] , [[#Ose--2019|Ose (2019)]] and [[#Zhou--2019|Zhou et al. (2019)]] .

<div id="10.3.4.3" class="h3-container"></div>

<span id="role-of-internal-variability"></span>
==== 10.3.4.3 Role of Internal Variability ====

<div id="h3-36-siblings" class="h3-siblings"></div>

A regional climate projection based on a single simulation from a single global model or driving a single RCM alone will inevitably be affected by not considering the internal variability (Figure 10.10). This is mainly due to the dominant influence of the chaotic atmospheric circulation on regional climate variability, in particular at mid- to high latitudes. Internal variability is an irreducible source of uncertainty for mid- to long-term projections with an amplitude that typically decreases with increasing spatial scale and lead time (Sections 1.4.3 and 4.2.1). However, regional-scale studies show that both large- and local-scale internal variability together can still represent a substantial fraction of the total uncertainty related to hydrological cycle variables, even at the end of the 21st century ( [[#Lafaysse--2014|Lafaysse et al., 2014]] ; [[#Vidal--2016|Vidal et al., 2016]] ; [[#Aalbers--2018|Aalbers et al., 2018]] ; [[#Gu--2018|Gu et al., 2018]] ).

<div id="_idContainer036" class="Basic-Text-Frame"></div>

[[File:b5a447f469f04d352b1f3ff6157251f9 IPCC_AR6_WGI_Figure_10_10.png]]

'''Figure 10.10''' '''|''' '''Observed and projected changes in austral summer (December to February) mean precipitation in Global Precipitation Climatoloy Centre (GPCC), Climatic Research Unit Time Series (CRU TS) and 100 members of the Max Planck Institute for Meteorology Earth System Model (MPI-ESM. (a)''' 55-year trends (2015–2070) from the ensemble members with the lowest (left) and highest (right) trend (% per decade, baseline 1995–2014). '''(b)''' Time series (%, baseline 1995–2014) for different spatial scales (from top to bottom: global averages; South-Eastern South America; grid boxes close to São Paulo and Buenos Aires) with a five-point weighted running mean applied (a variant on the binomial filter with weights [1-3-4-3-1]). The brown (green) lines correspond to the ensemble member with weakest (strongest) 55-year trend and the grey lines to all remaining ensemble members. Box-and-whisker plots show the distribution of 55-year linear trends across all ensemble members, and follow the methodology used in Figure 10.6. Trends are estimated using ordinary least squares. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Analysis of multi-model archives such as CMIP or CORDEX simulation results cannot easily disentangle model uncertainty and uncertainty related to internal variability. Since AR5, the development of single-model (global model and/or RCM) initial-condition large ensembles (SMILEs) has emerged as a promising way to robustly assess the regional-scale forced response to external forcings and the respective contribution of internal variability and model uncertainty to future regional climate changes ( [[IPCC:Wg1:Chapter:Chapter-4#4.2.5|Section 4.2.5]] ; [[#Deser--2014|Deser et al., 2014]] , 2020; [[#Kay--2015|Kay et al., 2015]] ; [[#Sigmond--2016|Sigmond and Fyfe, 2016]] ; [[#Aalbers--2018|Aalbers et al., 2018]] ; [[#Bengtsson--2019|Bengtsson and Hodges, 2019]] ; [[#Dai--2019|Dai and Bloecker, 2019]] ; [[#Leduc--2019|Leduc et al., 2019]] ; [[#Maher--2019|Maher et al., 2019]] ; [[#von%20Trentini--2019|von Trentini et al., 2019]] ; [[#Lehner--2020|Lehner et al., 2020]] ). The recent development of a multi-model archive of SMILE simulations facilitates the quantification and comparison of the influence of internal variability on global model-based regional climate projections between different models ( [[#Deser--2020|Deser et al., 2020]] ; [[#Lehner--2020|Lehner et al., 2020]] ). Another related development is the more frequent use of observation-based statistical models to assess the influence of internal variability on regional-scale global and regional model projections ( [[#Thompson--2015|Thompson et al., 2015]] ; [[#Salazar--2016|Salazar et al., 2016]] ). However, these methods often implicitly assume that regional-scale internal variability does not change under anthropogenic forcing, which is a strong assumption that does not seem to hold at regional and local scales ( [[#LaJoie--2016|LaJoie and DelSole, 2016]] ; [[#Pendergrass--2017|Pendergrass et al., 2017]] ; W. [[#Cai--2018|]] [[#Cai--2018|Cai et al., 2018]] ; [[#Dai--2019|Dai and Bloecker, 2019]] ; [[#Mankin--2020|Mankin et al., 2020]] ; [[#Milinski--2020|Milinski et al., 2020]] ).

The appropriate ensemble size for a robust use of SMILEs depends on the model and physical variable being investigated, the spatial and time aggregation being performed, the magnitude of the acceptable error and the type of questions one seeks to answer ( [[#Deser--2012|Deser et al., 2012]] , 2017b; [[#Kang--2013|Kang et al., 2013]] ; [[#Wettstein--2014|Wettstein and Deser, 2014]] ; [[#Dai--2019|Dai and Bloecker, 2019]] ; [[#Maher--2019|Maher et al., 2019]] ). It is noteworthy that the recent development of ensembles with a very large ensemble size (greater than 100) have led to new insights and methodologies to robustly assess the required ensemble size for questions such as the estimation of the forced response to external forcing or a forced change in modes of internal variability, such as ENSO, and its associated teleconnections ( [[#Herein--2017|Herein et al., 2017]] ; [[#Maher--2018|Maher et al., 2018]] ; [[#Haszpra--2020|Haszpra et al., 2020]] ; [[#Milinski--2020|Milinski et al., 2020]] ).

The use of SMILEs assumes that they have a realistic representation of internal variability and its evolution under anthropogenic climate change ( [[#Eade--2014|Eade et al., 2014]] ; [[#McKinnon--2017|McKinnon et al., 2017]] ; [[#McKinnon--2018|McKinnon and Deser, 2018]] ; [[#Chen--2019|Chen and Brissette, 2019]] ). Assessing the realism of simulated internal variability for past and current climates remains an active research field with a number of issues such as the shortness and uncertainties of the observed record, in particular in data-scarce regions ( [[#10.2.2.3|Section 10.2.2.3]] ), the signal-to-noise paradox ( [[IPCC:Wg1:Chapter:Chapter-4#4.4.3.1|Section 4.4.3.1]] ; [[#Scaife--2018|Scaife and Smith, 2018]] ), uncertainty in past observed external forcing estimates (Chapters 2, 6 and 7) and the limitations of assumptions underlying the statistical methods used to derive observational large ensembles ( [[#McKinnon--2017|McKinnon et al., 2017]] ; [[#McKinnon--2018|McKinnon and Deser, 2018]] ; [[#Castruccio--2019|Castruccio et al., 2019]] ). Calibration methods inspired by weather and seasonal forecasts can be used to improve the reliability of regional-scale climate projections from large ensembles ( [[#Brunner--2019|Brunner et al., 2019]] ; [[#O’Reilly--2020|O’Reilly et al., 2020]] ). Interestingly, reliability is improved when the calibration is performed separately for the dynamical and residual components of the ensemble resulting from dynamical adjustment ( [[#10.4.1|Section 10.4.1]] ; [[#O’Reilly--2020|O’Reilly et al., 2020]] ).

Importantly, accurately partitioning uncertainty in regional climate projections can provide an incentive for immediate action, accepting a large range of possible outcomes due to internal variability, while confounding model uncertainty with internal variability may be understood as a lack of knowledge and lead to delayed action in adaptation decision-making ( [[#10.5.3|Section 10.5.3]] ; [[#Maraun--2013b|Maraun, 2013b]] ; [[#Mankin--2020|Mankin et al., 2020]] ).

There is ''high confidence'' that the availability of SMILEs allows a robust assessment of the relative contributions of model uncertainty and internal variability in regional-scale projection uncertainty. There is ''high confidence'' that the use of SMILEs with appropriate ensemble size leads to an improved estimate of regional-scale forced response to an external forcing as well as of the full spectrum of possible changes associated with internal variability. There is ''high confidence'' that these improved estimates are beneficial for characterizing the full distribution of outcomes that is a key ingredient of climate information for robust decision-making and risk-analysis frameworks.

<div id="10.3.4.4" class="h3-container"></div>

<span id="designing-and-using-ensembles-for-regional-climate-change-assessments-to-take-uncertainty-into-account"></span>
==== 10.3.4.4 Designing and Using Ensembles for Regional Climate Change Assessments to Take Uncertainty Into Account ====

<div id="h3-37-siblings" class="h3-siblings"></div>

Ensembles of climate simulations play an important role in quantifying uncertainties in the simulation output (Sections 10.3.4.2 and 10.3.4.3). In addition to providing information on internal variability, ensembles of simulations can estimate scenario uncertainty and model (structural) uncertainty. Chapter 4, especially Box 4.1, discusses issues involved with evaluating ensembles of global model simulations and their uncertainties. In a downscaling context, further considerations are necessary, such as the selection of global model–RCM combinations when performing dynamical downscaling. This is a relevant issue when resources are limited. The structural uncertainty of both the global model and the downscaling method can be important (e.g., Mearnset al., 2012; [[#Dosio--2017|Dosio, 2017]] ), as well as further potential uncertainty created by inconsistencies between the global model and the downscaling method (e.g., [[#Dosio--2019|Dosio et al., 2019]] ), which could include, for example, differences in topography or the way to model precipitation processes ( [[#Mearns--2013|Mearns et al., 2013]] ).

An important consideration is which set of global models should be used for global model–RCM combinations. If adequate resources exist, then large numbers of global model–RCM combinations are possible ( [[#Déqué--2012|Déqué et al., 2012]] ; [[#Coppola--2021|Coppola et al., 2021]] ; [[#Vautard--2021|Vautard et al., 2021]] ). However, coordinated downscaling programmes can be limited by the human and computational resources available, for producing ensembles of downscaled output, which limits the number of feasible global model–RCM combinations. With this limitation in mind, a small set of GCMs may be chosen that span the range of equilibrium climate sensitivity in available global models (e.g., [[#Mearns--2012|Mearns et al., 2012]] , 2013; [[#Inatsu--2015|Inatsu et al., 2015]] ), though this range may be inconsistent with the likely range (Chapter 4), or some other relevant measure of sensitivity, such as the projected range of tropical SSTs ( [[#Suzuki-Parker--2018|Suzuki-Parker et al., 2018]] ). A further choice is to emphasize models that do not have the same origins or that do not use similar parametrizations and thus might be viewed as independent, a criterion that could be applied to both global models (Chapter 4) and RCMs ( [[#Evans--2014|Evans et al., 2014]] ). Global models and RCMs could also be discarded that unrealistically represent processes controlling the regional climate of interest ( [[#McSweeney--2015|McSweeney et al., 2015]] ; [[#Maraun--2017|Maraun et al., 2017]] ; [[#Bukovsky--2019|Bukovsky et al., 2019]] ; [[#Eyring--2019|Eyring et al., 2019]] ). Box 4.1 offers a more detailed discussion of the issues surrounding these approaches. Finally, global models may be selected to represent different physically self-consistent changes in regional climate ( [[#Zappa--2017|Zappa and Shepherd, 2017]] ). Statistical methods can provide estimates of outcomes from missing global model–RCM combinations in a large matrix ( [[#Déqué--2012|Déqué et al., 2012]] ; [[#Heinrich--2014|Heinrich et al., 2014]] ; [[#Evin--2019|Evin et al., 2019]] ).

However, even using a relatively small set of global models can still involve substantial computation that strains available resources, both for performing the simulations and for using all simulations in the ensemble for further impacts assessment. The NARCCAP programme ( [[#Mearns--2012|Mearns et al., 2012]] ) used only a subset of its possible global model–RCM combinations that balanced comprehensiveness of sampling the matrix with economy of computation demand, while still allowing discrimination, via ANOVA methods, of global model and RCM influences on regional climate change ( [[#Mearns--2013|Mearns et al., 2013]] ). An advantage of the sparse, but balanced matrix for those using the downscaling output for further studies, is that they have a smaller, yet comprehensive set of global model–RCM combinations to work with. Alternatively, data-clustering methods can clump together downscaling simulations featuring similar climate-change characteristics, so that only one representative simulation from each cluster may be needed for further impacts analysis, again systematically reducing the necessary number of simulations to work with (Mendlik andGobiet, 2016; [[#Wilcke--2016|Wilcke and Bärring, 2016]] ).

Independently of the resources, participation of multiple models in a simulation programme such as CORDEX for RCMs or CMIP for global models creates ensembles of opportunity, which are ensembles populated by models that participants chose to use without there necessarily being an overarching guiding principle for an optimum choice. As discussed in Chapter 4, these ensembles are likely suboptimal for assessing sources of uncertainty. An important contributor to the suboptimal character of such an ensemble is that the models are not independent. Some may also have larger biases than others. Yet often, the output from models in these ensembles has received equal weight when viewed collectively, as was the case in much of the AR5 assessment (e.g., [[#Collins--2013b|Collins et al., 2013b]] ; [[#Knutti--2013|Knutti et al., 2013]] ; [[#Flato--2014|Flato et al., 2014]] ; [[#Kirtman--2014|Kirtman et al., 2014]] ). A number of emerging methodologies aim at optimizing the ensembles available by weighting the simulation results according to a number of criteria relevant at the regional scale that aim at obtaining more realistic estimates of the uncertainty ( [[#Sanderson--2015|Sanderson et al., 2015]] ; [[#Brunner--2020|Brunner et al., 2020]] ).

There is ''high confidence'' that ensembles for regional climate projections should be selected such that models unrealistically simulating processes relevant for a given application are discarded, but at the same time, the chosen ensemble spans an appropriate range of projection uncertainties.

<div id="cross-chapter-box-10.2" class="h2-container box-container"></div>

'''Cross-Chapter Box 10.2 | Relevance and Limitations of Bias Adjustment'''

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

'''Coordinators:''' Alessandro Dosio (Italy), Douglas Maraun (Austria/Germany)

'''Contributors:''' Ana Casanueva (Spain), José Manuel Gutiérrez (Spain), Stefan Lange (Germany), Jana Sillmann (Norway/Germany)

Bias adjustment is an approach to post-process climate model output and has become widely used in climate hazard and impact studies ( [[#Gangopadhyay--2011|Gangopadhyay et al., 2011]] ; [[#Hagemann--2013|Hagemann et al., 2013]] ; [[#Warszawski--2014|Warszawski et al., 2014]] ) and national assessment reports ( [[#Cayan--2013|Cayan et al., 2013]] ; [[#Georgakakos--2014|Georgakakos et al., 2014]] ). Despite its wide use, bias adjustment was not assessed in AR5 ( [[#Flato--2014|Flato et al., 2014]] ). Several problems have been identified that may arise from an uncritical use of bias adjustment, and that may result in misleading impact assessments. The rationale of this Cross-Chapter Box is to provide an overview of the use of bias adjustment in this Report, and to assess key limitations of the approach.

Bias-adjusted climate model output is used extensively throughout this Report. Several results from Chapter 8, and many of the climatic impact-drivers in [[IPCC:Wg1:Chapter:Chapter-12|Chapter 12]] ( [[IPCC:Wg1:Chapter:Chapter-12#12.2|Section 12.2]] ) are based on bias adjustment. The ( [[IPCC:Wg1:Chapter:Atlas|Atlas]] presents many results both as raw and bias-adjusted data (Atlas.1.4.5). The application of bias adjustment in the WGI report was informed by the assessment in Chapter 10 and this Cross-Chapter Box. Finally, bias adjustment is crucial for many studies assessed in the WGII report. An overview of bias adjustment can be found in [[#10.3.1.3|Section 10.3.1.3]] , a general performance assessment of individual method classes in [[#10.3.3.7|Section 10.3.3.7]] . The fitness of bias adjustment for climate change applications is assessed in [[#10.3.3.9|Section 10.3.3.9]] .

'''Relevance of bias adjustment'''

An argument made for the use of bias adjustment is the fact that impact models are commonly very sensitive, often non-linearly, to the input climatic variables and their biases, in particular when threshold-based climate indices are required ( [[#Dosio--2016|Dosio, 2016]] ). There are, however, cases where bias adjustment may not be necessary or useful, such as: when only qualitative statements are required; when only changes in mean climate are considered (instead of absolute values); when percentile-based indices are used.

'''Modification of the climate change signal'''

Bias adjustment methods like quantile mapping can modify simulated climate trends, with impacts on changes to climate indices, in particular, extremes ( [[#Haerter--2011|Haerter et al., 2011]] ; [[#Dosio--2012|Dosio et al., 2012]] ; [[#Ahmed--2013|Ahmed et al., 2013]] ; [[#Hempel--2013|Hempel et al., 2013]] ; [[#Maurer--2014|Maurer and Pierce, 2014]] ; [[#Cannon--2015|Cannon et al., 2015]] ; [[#Dosio--2016|Dosio, 2016]] ; [[#Casanueva--2020|Casanueva et al., 2020]] ). Some argue that these trend modifications are implicit corrections of state-dependent biases ( [[#Boberg--2012|Boberg and Christensen, 2012]] ; [[#Gobiet--2015|Gobiet et al., 2015]] ). However, others argue that the modification is generally invalid because the modification is linked to the representation of day-to-day rather than long-term variability ( [[#Pierce--2015|Pierce et al., 2015]] ; [[#Maraun--2017|Maraun et al., 2017]] ); a given temperature value does not necessarily belong to the same weather state in present and future climate ( [[#Maraun--2017|Maraun et al., 2017]] ); the modification affects the models climate sensitivity ( [[#Hempel--2013|Hempel et al., 2013]] ); and is affected by random internal climate variability ( [[#Switanek--2017|Switanek et al., 2017]] ). Thus, trend preserving quantile mapping methods have been developed ( [[#10.3.1.3.2|Section 10.3.1.3.2]] ), although some authors found no clear advantage of these methods ( [[#Maurer--2014|Maurer and Pierce, 2014]] ). Further research is required to fully understand the validity of trend modifications by quantile-mapping.

'''Bias adjustment in the presence of large-scale circulation errors'''

The large-scale circulation has a strong impact on regional climate, thus circulation errors will cause regional climate biases ( [[#10.3.3.3|Section 10.3.3.3]] ). As bias adjustment in general does not account for circulation errors, it is therefore important to understand the impact of these errors on the outcome of the bias adjustment ( [[#Addor--2016|Addor et al., 2016]] ; [[#Photiadou--2016|Photiadou et al., 2016]] ; [[#Maraun--2017|Maraun et al., 2017]] ). If the frequency of precipitation-relevant weather types is biased, a standard bias adjustment (not accounting for this frequency bias) would remove the overall climatological bias, but the precipitation falling in a given weather type could still be substantially biased ( [[#Addor--2016|Addor et al., 2016]] ). Adjusting the number of wet days can artificially deteriorate the spell-length distribution ( [[#Maraun--2017|Maraun et al., 2017]] ). In the presence of location biases of circulation patterns, bias adjustment may introduce physically implausible solutions ( [[#Maraun--2017|Maraun et al., 2017]] ). Bias adjusting the location of circulation features ( [[#Levy--2013|Levy et al., 2013]] ) may introduce inconsistencies with the model orography, land–sea contrasts, and SSTs ( [[#Maraun--2017|Maraun et al., 2017]] ).

There is ''medium confidence'' that the selection of climate models with low biases in the frequency, persistence and location of large-scale atmospheric circulation can reduce negative impacts of bias adjustment.

'''Using bias adjustment for statistical downscaling'''

Bias adjustment is often used to downscale climate model results from grid box data to finer resolution or point scale. It is sometimes even directly applied to coarse-resolution global model output to avoid an intermediate dynamical downscaling step ( [[#Johnson--2012|Johnson and Sharma, 2012]] ; [[#Stoner--2013|Stoner et al., 2013]] ). But bias adjustment does not add any information about the processes acting on unresolved scales and is therefore by construction not capable of bridging substantial scale gaps ( [[#Maraun--2013a|Maraun, 2013a]] ; [[#Maraun--2017|Maraun et al., 2017]] ). Using bias adjustment for downscaling has been shown to artificially modify long-term trends, misrepresent the spatial characteristics of extreme events, and misrepresent local weather phenomena such as temperature inversions ( [[#Maraun--2013a|Maraun, 2013a]] ; [[#Gutmann--2014|Gutmann et al., 2014]] ; [[#Maraun--2017|Maraun et al., 2017]] ). Crucially, sub-grid influences on the local climate change signal are not represented. For instance, if a mountain chain is not resolved in the driving model, the snow–albedo feedback is not represented by the bias adjustment such that local temperature trends in high altitudes are under-represented (Cross-Chapter Box 10.2, Figure 1; [[#Maraun--2017|Maraun et al., 2017]] ). It has therefore been suggested to account for local random variability by combining bias adjustment with stochastic downscaling ( [[#Volosciuk--2017|Volosciuk et al., 2017]] ; [[#Lange--2019|Lange, 2019]] ), although this approach still does not account for local modifications of the climate change signal. Two approaches have been proposed to represent these local changes: dynamical downscaling with high-resolution RCMs ( [[#Maraun--2017|Maraun et al., 2017]] ) or statistical emulators of such ( [[#Walton--2015|Walton et al., 2015]] ). Sections 10.3.3.4–10.3.3.6 and 10.3.3.9 discuss other examples where RCMs improve the representation of regional phenomena and regional climate change.

[[File:9db913b9421a3849f5ab7fe73b1841dd IPCC_AR6_WGI_CCBox_10_2_Figure_1.png]]

'''Cross-Chapter'''

'''Box 10.2, Figure 1 |'''

'''Boreal spring (March to May) daily mean surface air temperature in the Sierra Nevada region in California. (a)''' Present climate (1981–2000 average, in °C) in the GFDL-CM3 GCM, interpolated to 8 km (left), GCM bias adjusted (using quantile mapping) to observations at 8 km resolution (middle) and WRF RCM at 3 km horizontal resolution (right). '''(b)''' Climate change signal (2081–2100 average minus 1981–2000 average according to RCP8.5, in °C) in the GCM (left), the bias adjusted GCM (middle) and the RCM (right). Further details on data sources and processing are available in the chapter data table (Table 10.SM.11). Figure adapted from [[#Maraun--2017|Maraun et al. (2017)]] .
Overall, there is ''high confidence'' that the use of bias adjustment for statistical downscaling, in particular to downscale coarse resolution global models, has severe limitations.

'''Bias adjustment of multiple variables'''

Impact models, as well as indices of climatic impact-drivers, often require input of several meteorological variables (Chapter 12). In several situations, for example, if the dependence between the variables is not well-simulated, univariate bias adjustment of the individual variables may increase biases in the resulting indicator ( [[#Zscheischler--2019|Zscheischler et al., 2019]] ). A simple alternative would be a bias adjustment of the indicator, but such a procedure may substantially alter the climate change signal, in particular for extreme events ( [[#Casanueva--2018|Casanueva et al., 2018]] ). In principle, multivariate bias adjustment methods are good to adjust all statistical aspects of the multivariate distribution that they intend to adjust. Depending on the method, this includes the correlation structure or even broader aspects of the dependence ( [[#Cannon--2016|Cannon, 2016]] , 2018; [[#Vrac--2018|Vrac, 2018]] ; [[#François--2020|François et al., 2020]] ). If multivariate adjustment includes a spatial dimension, then spatial dependence is adjusted well ( [[#Vrac--2018|Vrac, 2018]] ), but care is needed when applied across large areas ( [[#François--2020|François et al., 2020]] ). Adjustment of multivariate dependence necessarily modifies the temporal sequencing of the driving model ( [[#Cannon--2016|Cannon, 2016]] ; [[#Maraun--2016|Maraun, 2016]] ). The extent of the modification depends on the chosen method and the number of variables to adjust ( [[#Vrac--2015|Vrac and Friederichs, 2015]] ; [[#Cannon--2016|Cannon, 2016]] ; [[#Vrac--2018|Vrac, 2018]] ; [[#François--2020|François et al., 2020]] ).

'''Bias adjustment in the presence of observational uncertainty and internal variability'''

Observational uncertainties and internal variability introduce uncertainty in the estimation of biases and thus in the calibration of bias-adjustment methods. [[#Dobor--2019|Dobor and Hlásny (2019)]] found a considerable influence of the choice of the observational dataset and calibration period on the adjustment for some regions. RCM biases are typically larger than observational uncertainties, but in some regions, and in particular for wet-day frequencies, spatial patterns and the intensity distribution of daily precipitation, the situation may reverse ( [[#Kotlarski--2019|Kotlarski et al., 2019]] ). [[#Switanek--2017|Switanek et al. (2017)]] found a strong influence of internal variability and thus of the choice of calibration period on the calibration of quantile mapping and on the modification of the climate change signal.

Bias adjustment is typically evaluated using cross-validation, that is, by calibrating the adjustment function to one period of the observational record, and by evaluating it on a different one. [[#Maraun--2017|Maraun et al. (2017)]] and [[#Maraun--2018a|Maraun and Widmann (2018a)]] demonstrated that, in the presence of multi-decadal internal variability, cross-validation may lead to a rejection of a valid bias adjustment or even lead to a positive evaluation of an invalid adjustment. The authors therefore argued that, in the presence of substantial internal variability, the evaluation of bias adjustment requires to consider aspects that have not been adjusted, such as temporal, spatial, or multivariable dependence.

There is ''high confidence'' that observational uncertainty and internal variability adversely affect bias adjustment and introduce uncertainties in bias-adjusted future projections.

'''Overall assessment and new avenues'''

In the light of these issues, several authors dismiss the use of bias adjustment for climate change studies ( [[#Vannitsem--2011|Vannitsem, 2011]] ; [[#Ehret--2012|Ehret et al., 2012]] ). [[#Ehret--2012|Ehret et al. (2012)]] and [[#IPCC--2015|IPCC (2015)]] propose to at least provide the raw model output alongside the adjusted data. [[#Maraun--2017|Maraun et al. (2017)]] argue that the target resolution should be similar to the model resolution to avoid downscaling issues. [[#IPCC--2015|IPCC (2015)]] and [[#Maraun--2017|Maraun et al. (2017)]] highlighted the relevance of understanding model biases and the misrepresentations of the underlying physical processes prior to any adjustment. Together with [[#Galmarini--2019|Galmarini et al. (2019)]] , they point out the need for collaboration between bias adjustment users, experts in climate modelling and experts in the considered regional climate. As new research avenues, development of process-oriented bias adjustment methods ( [[#Addor--2016|Addor et al., 2016]] ; [[#Verfaillie--2017|Verfaillie et al., 2017]] ; [[#Manzanas--2019|Manzanas and Gutiérrez, 2019]] ) or run-time bias adjustment integrated into the climate simulation, for example, to reduce circulation errors ( [[#Guldberg--2005|Guldberg et al., 2005]] ; [[#Kharin--2012|Kharin et al., 2012]] ; [[#Krinner--2019|Krinner et al., 2019]] , 2020) are proposed.

<div id="10.4" class="h1-container"></div>

<span id="interplay-between-anthropogenic-change-and-internal-variability-at-regional-scales"></span>