快速脫碳情景下峰值升溫的數(shù)據(jù)驅(qū)動(dòng)預(yù)測(cè) Data-Driven Predictions of Peak Warming Under Rapid Decarbonization

ADVANCINGEARTHAND

SPACESCIENCES

RESEARCHLETTERData‐DrivenPredictionsofPeakWarmingUnderRapid

10.1029/2024GL111832Decarbonization

KeyPoints:

?WetrainCNNstopredictpeakglobalwarminggiventhemapofrecent

annualtemperaturesandtotaladditionalCO2emissions

?Evenifnet‐zeroemissionsarereachedmid‐century,meanwarmingisvirtu-allycertaintoexceed1.5°C,withevenoddsof2°C

?Thereishighlikelihoodofindividualyearsthatareatleast0.5°Chotterthan

2023eveninthemostambitiousdecarbonizationscenario

SupportingInformation:

SupportingInformationmaybefoundintheonlineversionofthisarticle.

Correspondenceto:

N.S.Diffenbaugh,

diffenbaugh@

Citation:

Diffenbaugh,N.S.,&Barnes,E.A.

(2024).Data‐drivenpredictionsofpeakwarmingunderrapiddecarbonization.

GeophysicalResearchLetters,51,

e2024GL111832.

/10.1029/

2024GL111832

Received7AUG2024Accepted5NOV2024

AuthorContributions:

Conceptualization:NoahS.Diffenbaugh,ElizabethA.Barnes

Datacuration:NoahS.Diffenbaugh,ElizabethA.Barnes

Formalanalysis:ElizabethA.BarnesMethodology:NoahS.Diffenbaugh,ElizabethA.Barnes

Software:ElizabethA.Barnes

Visualization:NoahS.Diffenbaugh,ElizabethA.Barnes

Writing–originaldraft:NoahS.Diffenbaugh

Writing–review&editing:Elizabeth

A.Barnes

?2024.TheAuthor(s).

Thisisanopenaccessarticleunderthetermsofthe

CreativeCommons

Attribution‐NonCommercial‐NoDerivs

License,whichpermitsuseand

distributioninanymedium,providedtheoriginalworkisproperlycited,theuseisnon‐commercialandnomodificationsoradaptationsaremade.

NoahS.Diffenbaugh

1

andElizabethA.Barnes2

1DoerrSchoolofSustainability,StanfordUniversity,Stanford,CA,USA,2DepartmentofAtmosphericScience,ColoradoStateUniversity,FortCollins,CO,USA

AbstractThesevereimpactsassociatedwithrecentrecord‐settingannualglobaltemperatureselevatetheneedtoaccuratelypredictthehottestconditionsthatcouldoccurevenifthemostambitiousdecarbonizationgoalsareachieved.Weuseconvolutionalneuralnetworks(CNNs)topredictpeakglobalwarmingfromrecentobservedtemperaturemapsandfuturecumulativeCO2emissions.FortheSSP1‐1.9decarbonizationscenariothereis>99%probabilitythatmeanglobalwarmingexceeds1.5°C,approximatelyevenoddsthatitreaches2°C,and~90%probabilitythatthehottestyeargloballyexceeds2023byatleast0.5°C.Further,fortheSSP2‐4.5decarbonizationscenario,thereis>90%probabilitythatthehottestannualglobaltemperatureanomalyistwicethe2023anomaly.Thatourframeworkmakeshighlyaccurateout‐of‐samplepredictionsofthehottesthistoricalyearprovidesconfidenceinthepredictedfutureprobabilities,suggestingsubstantialrisksfromtheextremelocalconditionsthatarelikelytoresultfromgloballyhotyearsduringrapiddecarbonization.

PlainLanguageSummaryCalendaryear2023wasthehottestyearonrecordglobally,reaching~1.5°Cabovethepre‐industrial.Manynational,sub‐nationalandnon‐stateactorshavearticulatedambitiousdecarbonizationgoalstostabilizetheglobaltemperature.However,theintensifyingimpactsasindividualyearshaveapproached1.5°Chaveheightenedtheneedtomoreaccuratelypredictnotjustthemeanwarmingbutalso

thehottestyearsthatcouldoccureveninthecontextofrapiddecarbonization.Wetrainneuralnetworksonan

ensembleofglobalclimatemodelsandthenusehistoricalobservationsasinputtothetrainednetworks,thusconstrainingtheuncertaintyinclimatemodelprojectionsbyusingthecurrentstateoftheclimatesystemasthebasisforatrulyout‐of‐sampleprediction.Forthehistoricalperiod,wefindthat,despiteawiderangeofclimatesensitivitiesacrossglobalclimatemodels,theneuralnetworksmakehighlyaccuratepredictionsofthehottesthistoricalyearwhengivenobservedclimatepatternsasout‐of‐sampleinputs.Predictingfuturewarmingunderdifferentcumulativeemissions,wefindthatevenifnet‐zeroemissionsareachievedmid‐century,mean

warmingisvirtuallycertaintoexceed1.5°Candhasevenoddsofreaching2°C,withhighlikelihoodofindividualyearsthatareatleast0.5°Chotterthan2023.

1.Introduction

TheUnitedNationsParisAgreementcodifiedtheinternationalgoalsof“holdingtheincreaseintheglobalaveragetemperaturetowellbelow2°Cabovepre‐industriallevelsandpursuingeffortstolimitthetemperatureincreaseto1.5°Cabovepre‐industriallevels”(UNFCCC,

2015

),leadingtoambitiousdecarbonizationgoalsfocusedonamid‐21st‐centurynet‐zeroemissionstimehorizon(Matthewsetal.,

2020

;UNEP,

2023

).Whileglobalemissionscontinuetoincrease,therateofincreasehasslowed(Friedlingsteinetal.,

2023

).Asaresult,althoughagapremainswithmid‐centurynet‐zerogoals,theglobalemissionstrajectoryiscurrentlyconsistentwithscenariosinwhichemissionsdecreaseduringthe21stcentury,andiswellbelowthoseinwhichemissionscontinuetoincrease(Friedlingsteinetal.,

2023

;Giddenetal.,

2019

;Riahietal.,

2017

).Thesedevelopmentshavegeneratedsubstantialinterestinquantifyinguncertaintyintheglobalwarmingtrajectoryunderdecarbonization,includingtheprobabilitythatmeanglobalwarmingexceeds1.5°Cor2.0°C(e.g.,IPCC,

2018

;Leeetal.,

2021

;Rogeljetal.,

2017

).

LiketheParisAgreement,muchoftheexistingresearchisframedaroundtherelationshipbetweencumulativeemissionsandglobaltemperaturechange.However,althoughthefirst‐orderrelationshipisrobust(Leeetal.,

2021

;Matthews&Caldeira,

2008

;Matthewsetal.,

2009

;Matthews&Wynes,

2022

),anumberofun-certaintiesremain,includingintheexactamountofwarmingforagivenlevelofcumulativeemissions(e.g.,Matthews&Wynes,

2022

),theexactcarbonbudgetforagivenlevelofwarming(e.g.,Lambolletal.,

2023

;

DIFFENBAUGHANDBARNES1of10

19448007,2024,23,

10.1029/2024GL111832

ADVANCINGEARTHANDSPACESCIENCES

GeophysicalResearchLetters

Downloadedfrom

/doi/10.1029/2024GL111832byCochraneChina

,WileyOnlineLibraryon[30/12/2024].SeetheTermsandConditions(

/terms-and-conditions

)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsLicense

industrial(R.Rohde,

2024

;Schmidt,

2024

;Figure1)—haveheightenedtheneedtoaccuratelypredictnotjusttheforcedglobaltemperatureresponsebutalsothehottestindividualyearsthatcouldoccurevenifthemostambitiousdecarbonizationgoalsareachieved.

Withthatmotivation,weuseconvolutionalneuralnetworks(CNNs)toquantifytheprobabilityofpeakglobalwarmingunderdifferentdecarbonizationscenarios.Webuildonrecentworkusingneuralnetworkstopredictthetimeuntilagivenforcedtemperatureresponseisreached(Diffenbaugh&Barnes,

2023

).Akeyadvantageofthisframeworkisthecapacitytoconstrainclimatemodeluncertaintyusingtheobservedclimateasthebasisforanout‐of‐sampleprediction.Inthecurrentstudy,wetrainCNNstopredictthedistributionofpeakforcedtem-peratureresponseandmaximumannualtemperaturefrom(a)mapsofannualtemperaturesimulatedbyanensembleofglobalclimatemodels(GCMs)and(b)remainingcumulativeCO2emissions.Toquantifyuncer-tainty,theCNNarchitectureisconstructedtopredicttheparametersofasinh‐arcsinhconditionaldistribution(SHASH(Barnesetal.,

2023

;C.Jones&Pewsey,

2019

;M.C.Jones&Pewsey,

2009

)),andwetrainmultipleCNNsusingdifferentinitialrandomseedstoproducederiveddistributionsthatcapturemultiplesourcesofuncertainty.OncetheCNNsaretrained,weuseobservedmapsofrecentannualtemperatureanomaliestopredictthedistributionofpeakwarmingunderdifferentlevelsoffuturecumulativeemissions.

2.MaterialsandMethods

2.1.DataSets

TotraintheCNNsacrossarangeofdecarbonizationpathways,wepoolresultsfrommultiplerealizationsofmultipleGCMsintheSSP1‐1.9,SSP1‐2.6andSSP2‐4.5decarbonizationscenariosarchivedinPhase6oftheCoupledModelIntercomparisonProject(CMIP6(Eyringetal.,

2016

);Figure1,TableS1inSupportingInfor-mationS1).AsinDiffenbaughandBarnes(

2023

),weselectthoseGCMsthathavearchived≥5realizationsinordertocapturetheinfluenceofinternalclimatevariability,butonlyinclude5realizationsfromeachinordertoweightthoseGCMsevenly.OnelimitationisthatonlyafewGCMshavearchivedatleast5realizationsintheCMIP6scenariosthatreachnet‐zeroduringthe21stcentury(SSP1‐1.9andSSP1‐1.6;TableS1inSupportingInformationS1).Giventheimportanceofincludingabroadrangeofclimatesensitivities,weexpandthetrainingdatasetbyalsoincludingonerealizationofeachavailableGCM,asintheIPCC's“onemodel,onevote”approach.

Withinthis“grandensemble”,wedefinethepeakforcedresponseforeachensemblememberseparately.Wefiteachrealization'sglobalmeantemperaturetimeseriesfrom2000to2100withacubicfunction,andthemaximumofthiscubicfitdefinesthepeakforcedresponse(FigureS2AinSupportingInformationS1).Themaximumannualtemperatureissimilarlydefinedseparatelyforeachrealization,exceptthatnocubicfitisapplied(FigureS2BinSupportingInformationS1).

WeusetheBerkeleyEarthgriddedtemperatureobservationsasourprimaryobservationaldataset.Althoughtheglobaltemperaturesynthesisisveryconsistentbetweenobservationaldatasets(Gulevetal.,

2021

),therecanbedifferencesintheexactspatialpatternsinagivenyear(e.g.,NCEI,

2024

;R.Rohde,

2024

).However,theneuralnetworkpredictionsoftheevolutionoftheglobaltemperatureresponseareconsistentbetweengriddeddatasets(Diffenbaugh&Barnes,

2023

),andwefindverycloseagreementinthepredictionofthehistoricalmaximumannualtemperatureusinggriddeddatasetsfromBerkeleyEarth(R.A.Rohde&Hausfather,

2020

)andNASA(GISTEMP,

2021

)(Figure1).

Forallobservedandsimulateddatasets,weinterpolatetheannualtemperaturedatatoacommon2.5°×2.5°grid(IPCC,

2013

).Globalmeantemperatureanomalies(thepredictand)arecomputedasdeviationsfromthe1850–1899meanofeachrealization.Surfacetemperatureanomalies(thepredictors)forboththeGCMsandobser-vationsarecomputedasthedifferencefromeachrealization's1975–2004mean.

Forcalculatingcumulativeemissions,wefirstconcatenatethehistoricalemissionsdatafrom(Friedlingsteinetal.,

2022

)withtheSSPemissionsdatafortheremainderofthe21stcentury(Giddenetal.,

2019

;IIASA,

2018

;Riahietal.,

2017

),andthenfollowtheSSPextensionsof(Meinshausenetal.,

2020

)bylinearlyinterpolatingatanannualresolutionoutto2250(FigureS7AinSupportingInformationS1).Foragivenyear,thecumulativeemissionsremainingarethencomputedasthesumoverallpositiveemissionsfrom1850to2250minusthecumulativesum‐to‐date.SinceSSP1‐1.9andSSP1‐2.6bothreachnet‐zerooverthe21stcentury,thisdefinitionleadstonegativeemissionsremainingpriorto2100(FigureS7binSupportingInformationS1).

DIFFENBAUGHANDBARNES3of10

Downloadedfrom

/doi/10.1029/2024GL111832byCochraneChina

,WileyOnlineLibraryon[30/12/2024].SeetheTermsandConditions(

/terms-and-conditions

)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsLicense

2.2.ConvolutionalNeuralNetworks(CNNs)

2.2.1.ArchitectureandTrainingoftheCNN

WetrainaCNNtotakesinglemapsofannual‐meannear‐surfacetemperature(“tas”)andtheadditionalcu-mulativeCO2emissionsremaining,andpredictthepeakwarming(asananomalyfrom1850to1899)thatwilloccurby2100(FigureS1inSupportingInformationS1).Followingpreviouswork,wetraintheCNNtopredictthefourparametersofasinh‐arcsinh(SHASH)distribution(Barnesetal.,

2023

;C.Jones&Pewsey,

2019

;M.C.Jones&Pewsey,

2009

):mu;sigma;gamma;andtau.Thus,foreachinputintothenetwork,thenetworkpredictsafullconditionaldistributionofthewarmingremainingthrough2100(FigureS3inSupportingInformationS1).Fromthis,wecomputethepredictedpeakwarmingasthesumofthecurrentyear'stemperatureanomalyplusthepredictedremainingwarming.AlthoughtheSHASHtakesfourparameters,wehavefoundthattrainingtheCNNtopredictonlythefirstthree(andfreezingtau=1.0)producessimilarresultstotrainingallfour(seeFigureS5inSupportingInformationS1).

TheCNNhasaninitialcircularpaddinglayerthataccountsforthediscontinuityinlongitudeatthePrimeMeridian.Fromthere,theinputmapisfedthroughthreeconvolutionallayerswith32kernelseachofsize5×5,3×3and3×3,respectively.Eachconvolutionallayerisfollowedbyamaxpoolinglayerwithkernelsize2×2andstrideof2toreducethedimensionality.ThefinalmaxpoolinglayerisfollowedbyaflattenlayertowhichtheinputcumulativeCO2emissionsremainingscalarisappended.WeseparatethepredictionsoftheSHASHpa-rameters,therebyallowingthenetworktolearndifferentstrategiesifnecessary.Eachparameterpathwayconsistsofthreedenselayersof10unitseach.TheinputcumulativeCO2emissionsremainingareonceagainconcate-natedtotheoutputofthedenseblockandtheresultingtensorisfedintoafinaldenseblockof5units.Theoutputisprocessedthrougharescalinglayerpriortopredictingeachparameter.Weusetherectifiedlinearunit(ReLu)forallactivationsexceptforthefinaloutput,whichusesahyperbolictangent(Tanh)activation.ThefullCNNsetupdepictedinFigureS1inSupportingInformationS1containsatotalof273,992trainableparameters.However,wefreezethefourthparameter(tau)asitsinclusiondidnotimproveourresults(comparethe“default”to“hp9”and“hp10”inFigureS5inSupportingInformationS1).Thus,thefinalCNNarchitecturecontains210,326trainableparameters.

ThetrainingsetiscreatedbyrandomlyselectingtworealizationsfromeachGCMthathasatleast5realizations(“multi‐memberGCMs”),andrandomlyselecting80%ofthesingle‐memberrealizations(whichincludethefirstrealizationfromthemulti‐memberGCMs).Thevalidationset,whichisusedforearlystopping,iscreatedbyrandomlyselectingoneoftheremainingmembersofthemulti‐memberGCMsand10%oftheremainingsingle‐memberrealizations.Thetestingsetisthencreatedbycombiningalloftheremainingrealizations,whichamountstoonefromthemulti‐memberGCMsand10%ofthesingle‐memberrealizations.TheCNNsaretrainedtominimizethenegativelog‐likelihoodlossusingtheAdamoptimizerwithabatchsizeof32,learningrateof1.0e?05,weightdecayofzero,andepsparameterof1.0e?07.Weemployearlystoppingwhenthevalidationlossnolongerdecreasesafterapatienceof20epochstohelpavoidoverfitting.

WeconductanumberofteststoexplorethesensitivityofourresultstotheCNNarchitecturechoicesbytraining5differentCNNsforeachnetworkhyperparameterconfiguration:“hp1‐10”(FigureS5inSupportingInforma-tionS1).ThefiveCNNsforeachconfigurationdifferonlybytheirrandomseed,whichsetsboththetraining/validation/testingsplitaswellastherandominitializationoftheweights.Thearchitectureusedhere(“default”)resultsinlowperformancemetricssuchasmeanabsoluteerrorandloss,althoughitisreassuringthatotherhyperparameterchoicesleadtosimilarpredictionsfortheobservations(FigureS5inSupportingInformationS1).

2.2.2.UncertaintyQuantification

ThenetworkpredictsaSHASHdistributionoftheremainingwarmingforeachinput,providingameasureofuncertaintyconditionedontheinput.However,trainingdifferentCNNsthatvaryonlyintherandomseedusedtosplitthedataandinitializethenetworkcanresultinarangeofobservationalpredictions(FiguresS4,S5inSupportingInformationS1).Thus,thereisanadditionalsourceofuncertaintynotdirectlycapturedbyeachpredictedSHASH(Barnesetal.,

2023

;Haynesetal.,

2023

).Toaccountforthis,wetrain15+CNNswithdifferingseedsforeachpeakwarmingdefinition(i.e.,peakforcedresponseandmaximumannualtemperature)andselectthefirst15CNNswhosetestinglossislessthan?0.1(FigureS5inSupportingInformationS1providesmetricsforeachofthe15+individualSHASHdistributionpredictions.).Wethencreateaderiveddistributionby

DIFFENBAUGHANDBARNES4of10

19448007,2024,23,Downloadedfrom

/doi/10.1029/2024GL111832byCochraneChina

,WileyOnlineLibraryon[30/12/2024].SeetheTermsandConditions(

/terms-and-conditions

)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsLicense

GeophysicalResearchLetters10.1029/2024GL111832

Figure2.Probabilityofpeakwarminginresponsetocumulativeemissions.(a)Probabilitythatthepeakforcedtemperatureresponse(comparedtothe1850–1899pre‐industrialbaseline)remainsbelowarangeofthresholdsasafunctionofremainingcumulativeemissions,predictedbyCNNsinitializedusingobservationsin2023.Verticaldottedlinesdenotetheremainingemissionsunderthreedecarbonizationscenarios.(b)TheCNN‐predictedderiveddistributionsforthepeakforcedresponseby2100,correspondingtothecross‐sectionsshownasverticaldottedlinesin(a).(c)Asin(a)butforthemaximumannualtemperatureby2100.(d)Asin(b),butforthemaximumannualtemperatureby2100(i.e.,correspondingtothecross‐

sectionsshownasverticaldottedlinesin(c)).

combiningthepredictionsacrossall15CNNs.Specifically,foreachpredictionontheobservations(i.e.,eachsetofinputs)weobtain15SHASHdistributionsoftheremainingwarming.Wethentake50,000randomdrawsfromeachpredictedSHASHdistributionresultingin15×50,000=750,000randomvalues.These750,000valuesthuscompriseourpredictedderiveddistributionfromwhichwecanplottheresultingprobabilitydensityfunctionandcalculateprobabilities(Figure

2

).Allresultsdiscussedinthetextarequantifiedfromthisderiveddistribution.IndividualSHASHpredictionsandmetricsarepresentedinFiguresS3–S5inSupportingInformationS1.

Ourdecisiontoemployaderiveddistributionforquantifyinguncertaintyissupportedbythefactthatthetestinguncertainty(anderror)introducedbythechoiceintraining/validation/testingsplitismuchlargerthantheun-certaintyduetotherandominitializationofthenetworkweights(FigureS6inSupportingInformationS1).Sinceweemploya“one‐model,one‐vote”approachfortheGCMensemble,everytraining/validation/testingsplitisequallyvalidpresumingthatthenetworkdoesareasonablejobtraining(i.e.,meetsourcriteriathatthelossbelessthan?0.1).Becauseeachofthe15CNNsareequallyvalid,wechoosetocombineall15predictionsintoasinglederiveddistributiontocapturetheuncertaintyfromthedatasplitratherthanchooseoneCNN,whichwouldfailtocapturethisadditionalsourceofclimatemodeluncertainty.

DIFFENBAUGHANDBARNES

5of10

19448007,2024,23,

10.1029/2024GL111832

ADVANCINGEARTHANDSPACESCIENCES

GeophysicalResearchLetters

Downloadedfrom

/doi/10.1029/2024GL111832byCochraneChina

,WileyOnlineLibraryon[30/12/2024].SeetheTermsandConditions(

/terms-and-conditions

)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsLicense

3.Results

Wefirsttesttheaccuracyofourpredictionframeworkduringthehistoricalperiod.WetrainmultipleCNNsontheCMIP6Historicalsimulations(extendingto2023usingtheSSP1‐1.9simulations).Then,foreachhistoricalyear,weusetheobservedmapsofannualtemperatureanomaliesandremainingcumulativeCO2emissions(through2023)topredictthemaximumannualtemperaturethatwillbereachedby2023.Theobservations‐basedpre-dictionisquitestablethroughtime,andisverysimilarwhenusinginputsfromNASA(“GISTEMP”;Figure1c)orBerkeleyEarth(“BEST”;Figure1d).Further,theactualhottesthistoricalyearfallsveryclosetothemeanofCNNpredictionsmadeusinginputsfrom1980to1999,andwithinthe“l(fā)ikely”range(66%intheIPCCun-certaintyguidance(Mastrandreaetal.,

2011

))forbothobservedtemperaturedatasets.

WenextusetheCNNstrainedacrosstheSSPdecarbonizationscenariostopredictthedistributionofpeakwarminggiven2023temperatureobservationsandvaryinglevelsofcumulativeemissions(Figure

2

).WefindthatthetimeseriesofpeakwarmingpredictedbytheCNNsisrobusttotheinitializationyear,particularlyforthepasttwodecades(Figures1eand1f).Further,theuncertaintiesintheobservations‐basedCNNpredictionsaresubstantiallysmallerthantheuncertaintyintheGCMsimulationsonwhichtheCNNsaretrained,particularlyfortherapiddecarbonizationscenarios(Figures1e,1f,S8inSupportingInformationS1).

Althoughthemedianpredictionisgenerallyproportionaltothecumulativeemissions,thedistributionsbecomemorepositivelyskewedascumulativeemissionsincrease,withtheuppertailincreasingatafasterratethanthelowertail(Figures

2a

and

2c

).ForSSP1‐1.9,whichreachesnet‐zeroCO2emissionsinthemid‐2050sfollowedbydeepnegativeemissions,theCNN‐predictedpeakforcedresponseby2100is“virtuallycertain”(Mastrandreaetal.,

2011

)toexceed1.60°C(medianprediction:1.96°C),andthemaximumannualtemperatureis“virtuallycertain”toexceed1.80°C(medianprediction:2.18°C).ForSSP1‐2.6,whichreachesnet‐zeroCO2emissionsinthemid‐2070s,thepeakedforcedresponseis“virtuallycertain”toexceed1.83°C(medianprediction:2.22°C),andthemaximumannualtemperatureis“virtuallycertain”toexceed2.01°C(medianprediction:2.43°C).ForSSP2‐4.5,whichdoesnotreachnet‐zerointhe21stcentury,thepeakforcedresponseis“virtuallycertain”toexceed2.69°C(medianprediction:3.33°C),andthemaximumannualtemperatureis“virtuallycertain”toexceed2.76°C(medianprediction:3.46°C).

Themagnitudeofpredictedpeakwarmingsuggeststhat,evenifrapiddecarbonizationisachieved,regionalconditionsarelikelytobeconsiderablymoreseverethanwhatpeopleandecosystemshaveexperiencedtodate.Forexample,foranannualglobaltemperatureanomalyof~2°C,regionalconditionsarehighlylikelytobehotterthanin2023,includingastrongpossibilityofregionaltemperatureanomaliesthatareatleastdoublethe2023regionalanomaly(Figure3a).Likewise,forasingleclimaterealizationwhosemaximumannualglobaltem-peratureanomalyreaches~2°C,themaximumannualdailyprecipitationinthehottestyearsexceedsthebaselinemaximumovermostoftheglobe,includingwidespreadincreasesofatleast25%overcontinentalareasandevenlargerincreasesovermuchofthetropics(Figure3c).However,suchrapiddecarbonizationisfarfromguaranteed(UNEP,

2023

),andthereisveryhighlikelihoodthatthemaximumannualglobaltemperatureanomalyreaches2.5°Cforadditionalcumulativeemissionsof2,000GtCO2and3.0°Cforadditionalcumulativeemissionsof3,500GtCO2(Figure

2c

).Foranannualglobaltemperatureanomalyof~3°C,regionalconditionsarevirtuallycertaintobehotterthanin2023overmostillustrativeregions,includingastrongpossibilityofregionaltem-peratureanomaliesthatareatleasttriplethe2023regionalanomaly(Figure3a).

4.Discussion

Ourresultssuggestvirtualcertainty(>99%intheIPCCuncertaintyguidance(Mastrandreaetal.,

2011

))thatreachingnet‐zeroCO2emissionsbymid‐centurywillnotholdtheglobalforcedtemperatureresponsebelow1.5°C.Whilethereareavarietyofdefinitionsforglobaltemperaturethresholds(e.g.,Bettsetal.,

2023

),theforcedresponseisamongthemostconservative.Hence,ourresultsraisethelikelihoodofabroadsuiteofimpactsthatareexpectedtooccurbeyondthe1.5°Cthreshold,includingonterrestrialandmarineecosystems,humanhealth,livelihoods,economicgrowth,andfoodandwatersecurity(IPCC,

2018

).

Inaddition,themedianCNN‐predictedpeakforcedresponseby2100forSSP1‐1.9emissionsis1.96°C(Figure1),andourresultssuggestthatotherambitiousdecarbonizationpathwaysthatholdfuturecumulativeemissionsto~500GtCO2arealsoonly~50%likelytoavoidapeakforcedresponseof2°C(Figure

2a

).Absentadaptation,thislevelofwarmingposes“high”to“veryhigh”riskstonaturalandhumansystemsacrossthe

DIFFENBAUGHANDBARNES6of10

Downloadedfrom

/doi/10.1029/2024GL111832byCochraneChina

,WileyOnlineLibraryon[30/12/2024].SeetheTermsandConditions(

/terms-and-conditions

)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsLicense

comprehensiveareasassessedbytheIPCC(IPCC,

2022

).Further,ourresultssuggestthatthepeakforcedresponseisverylikelytoexceed2°Cifadditionalcumulativeemissionsexceed1,000GtCO2,andislikelytoexceed3°Cifadditionalcumulativeemissionsexceed3,500GtCO2(Figure

2a

).Absentadaptation,3°Cofmeanglobalwarmingposes“veryhigh”risks(IPCC,

2022

).

Giventheimportanceofextremeeventsforclimatechangeimpacts(e.g.,IPCC,

2012

),themagnitudeofindi-vidual