Uncertainties in regional climate change - National Centre for Physics

Projecting future climate change given futureconcentrations - temperaturePage 3

Projecting future climate change given futureconcentrations - precipitationPage 4

UKCIP ‘02• Based on the state-of-the-artat the time - HadCM3,HadAM3H time-slice, 50kmHadRM3 experiments• Used by many private andpublic-sector organisations tomake decisions and spendmoney• “Scenario” based with noquantification of uncertainties(although plenty of caveatspointing this out)Page 5

UKCIP ‘022080temperaturechange (K)2080precipitationchange (%)We can produce very detailed projections of climatechange with no idea of how reliable they might bePage 6

2041-602061-80AOGCMProjected changein summermonsoon rainfall,relative to 1961-90IPCC TARPage 7

UKCIPnext – Aims and Objectives• To provide joint probability distribution functions (pdfs)of projected changes in a selection of key UK climatevariables at 25km resolution for each decade duringthe 21st century• Results will be presented for each variable by monthand summarised as quantiles indicating both meanand extreme outcomes• The set of climate variables will be determined inconsultation with stakeholders• We aim to deliver the final report and the pdfs duringthe first half of 2008Page 8

Modelling impactsPage 9

GJJ1999Modelling impactsEMISSIONSCONCENTRATIONSCO 2, methane, etc.HEATING EFFECT‘Climate Forcing’.CLIMATE CHANGETemp, rain, sea-level, etc.IMPACTSFlooding, food supply, etc.FeedbacksFeedbacksScenarios frompopulation, energy,economics modelsCarbon cycle andchemistry modelsGas propertiesCoupled climatemodelsImpacts modelsPage 10

Development of climate modelsFrom IPCC TAR, 2001Page 11

Assessment of model uncertainties• Uncertainty in representing physical andbiological processes in climate models1. Uncertainties due to different representations ofprocesses (structural)2. Uncertainties in key parameters in models3. Omitted processesPage 12

The HadCM3 perturbed physics ensemble• Assessment of the uncertainty due to the parametrization• Models built using different values of parameters whose rangesare ill-constrained• 256 member ensemble with multiple parameter perturbations –atmosphere/slab ocean model, 2xCO 2• Assessment of model quality and global mean climateresponse (i.e. climate sensitivity)• Run a 17 member ensemble of coupled model transientexperiments using “high quality” models with range of climatesensitivities – 1860-2100, A1B (flux-adjusted HadCM3)• Additional perturbed model are emulated (Sexton and Rougier,in preparation)• Transient response generated by pattern-scaling (Harris et al,2005, EBM based)Page 13

The slab ensemble picture: frequencies ofprojected rainfall changes (DJF)Amazon basinMexico CityBangkok404040Relative likelihood (%)3020103020103020100-50 0 +50Rainfall change (%)0-50 0 +50Rainfall change (%)0-50 0 +50Rainfall change (%)Blue – mean change: Red – top 95%ile changePage 14

Simple uncertainty estimatesChanges in percentiles for T(max) JJAfor a) Czech republic, b) Great Lakes(USA), c) eastern China and d) SouthernFrance.10-90% spread of the 53 membersensemblesfrom Clark et al, J. Clim, 2006Page 15

Use of obs constraints: Bayesian framework•Perform a limited ensemble of GCMexperiments with perturbed input parameters•Introduce an emulator which can estimate theGCM output at untried parameter values•Introduce a discrepancy term derived from theoutput of other climate models to representstructural uncertainties•Produce prior predictive distributions of climatevariables•Use observations to produce a likelihoodfunction and posterior predictive distributionsPage 16

The first example“emulated”priorpredictivedistributionposteriorpredictivedistributionlikelihoodweighting viacomparisonwith real worldhistogram of“perturbed physics”ensembleNo discrepancyMurphy et al., 2004, Nature, 430, 768-772 Page 17

Climate prediction index (CPI) to select AOGCMsThe choice of themodels:• a reasonablespread in climatesensitivity• a reasonablygood CPI• a spread inmodel parameterspacePage 18

Global mean temperature response fromHadCM3 ensemblePage 19

Multi-model (MM) and parameter perturbation(PP) ensemble rangesResults for 17member HadCM3PP ensemble and21 member IPCCAR4 MM ensembleDFJ precipitation change in mm/day5% 5% Average Average 95% 95%MME PPE MME PPE MME PPECAM -0.76 -0.54 -0.27 -0.15 0.23 0.26JJA precipitation change in mm/dayPrecipitation andtemperaturechange for CentralAmerica (CAM)5% 5% Average Average 95% 95%MME PPE MME PPE MME PPECAM -0.99 -1.59 -0.44 -0.59 0.11 0.37Annual temperature change deg C5% 5% Average Average 95% 95%MME PPE MME PPE MME PPECAM 1.8 2.46 3.2 3.58 4.6 5.02Page 20

Application of HadCM3 ensemble to climatescenario generation• For UKCIP08: run 17-member HadRM3 ensemble at 25kmdriven by HadCM3 ensemble and pattern-scale results forthe 256 slab ensemble• For ENSEMBLES (EU project for Europe): run HadRM3and other RCMs driven by HadCM3.0 and high and lowsensitivity members of the full ensemble (i.e. 3experiments) – ensemble augmented with other GCMsand RCMs• Application from other groupsPage 21

ENSEMBLES-EU perturbed physics GCMsMSLP bias DJF (hPa)T bias DJF (K)“unperturbed” HadCM3high sensitivitylow sensitivityPage 22

QUMP predictions at RCM scale• 17 RCM 150-years transient simulations at 25km horizontalresolution• RCMs driven from QUMP GCMs (RCM parameters consistentwith GCM)• Variables from other GCMs will be obtained by statisticaldownscaling from GCM to RCM scale• Transferability of methodsPage 23

Statistical Downscaling GCM -> RCMK. Brown (unpublished)• SDSM (Wilby et al, 2002)• Daily temperature distribution for UK• HadAM3P/HadRM3P simulations(1960-1990 and 2070-2100 A2)• 70% RCM variance explained by GCMtemperature• Preliminary results on precipitation notso goodPage 24

“Local scaling” approachE. Kennett (unpublished)TemperaturePrecipitation50 th percent95 th percent99 th percentPage 25

“Local scaling” approach: effect of GCM resolutionon precipitation - E. Kennett (unpublished)HadCM2 (~300km)HadAM3P (~150km)50 th percent.95 th percent.99 th percent.Page 26

Summary• Methods to estimate uncertainty in climate projectionare under development.• Possible problems with the users, impact models notready for pdfs of climate variables.• Ensembles of climate simulations require models withlow computational costs to maximise the number ofperturbed models involved. Sophisticated samplingstrategies are also needed.Page 27

Thanks!Erasmo.Buonomo@metoffice.gov.ukPage 28