Paper deep dive
Interpretable AI predicts a 2026 summer dry anomaly in central China
Anran Wang, Wen Shi, Yong Luo, Jianbin Huang, Lijuan Chen, Junhu Zhao, Weixin Jin, Huihui Yuan
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/22/2026, 1:43:12 AM
Summary
The paper presents an interpretable deep learning framework that predicts a significant summer 2026 dry anomaly in central China by translating dynamical atmospheric circulation predictions into precipitation estimates. The model identifies northerly winds associated with a cyclonic circulation anomaly over the western North Pacific-South China Sea as the dominant driver, supported by retrospective evaluations of historical analogue years featuring central equatorial Pacific warming. Layer-wise relevance propagation (LRP) and perturbation tests confirm the physical interpretability and faithfulness of these AI-derived explanations.
Entities (8)
Relation Signals (6)
Deep Learning Model → predicts → Central China
confidence 95% · Here, we employ a deep learning model that translates dynamical circulation predictions into precipitation estimates. Predictions initialized from March to May consistently indicate a dry anomaly over central China in summer 2026.
Layer-wise Relevance Propagation (LRP) → identifies → Northerly Winds
confidence 94% · Supporting this mechanism, layer-wise relevance propagation (LRP) independently identifies these northerly winds as the dominant driver of the prediction among all model inputs.
Northerly Winds → suppresses → Rainfall
confidence 94% · ...northerly winds and moisture divergence that jointly suppress rainfall over central China.
Cyclonic Circulation → induces → Northerly Winds
confidence 93% · ...which induces northerly winds and moisture divergence that jointly suppress rainfall over central China.
Northerly Winds → isdominantdriverof → Dry Anomaly Prediction
confidence 93% · LRP independently identifies these northerly winds as the dominant driver of the prediction among all model inputs.
Central Equatorial Pacific Warming → favors → Cyclonic Circulation
confidence 92% · This warming favors an anomalous cyclonic circulation over the western North Pacific-South China Sea-South China region...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Seasonal precipitation anomalies are largely regulated by atmospheric circulation, which dynamical models predict with greater reliability than precipitation itself. Here, we employ a deep learning model that translates dynamical circulation predictions into precipitation estimates. Predictions initialized from March to May consistently indicate a dry anomaly over central China in summer 2026. Retrospective evaluations revealed higher predictive skill in the analogue years, which also tended to feature central equatorial Pacific warming persisting from the preceding winter into summer. This warming favors an anomalous cyclonic circulation over the western North Pacific-South China Sea-South China region, which induces northerly winds and moisture divergence that jointly suppress rainfall over central China. Supporting this mechanism, layer-wise relevance propagation (LRP) independently identifies these northerly winds as the dominant driver of the prediction among all model inputs. Perturbation tests supported this attribution: removing LRP-identified features effectively eliminates the dry anomaly. Our framework thus provides physically interpretable explanations for AI-derived regional climate projections, facilitating evidence-based assessment before observational data become available.
Tags
Links
- Source: https://arxiv.org/abs/2608.19163v1
- Canonical: https://arxiv.org/abs/2608.19163v1
Trouble viewing inline? Open PDF directly →
Full Text
47,763 characters extracted from source content.
Expand or collapse full text
InterpretableAIpredictsa2026summerdryanomalyin centralChina AnranWANG 1† ,WenSHI 1† ,YongLUO 1 ,JianbinHUANG 2,3,4 ,LijuanCHEN 5,6 ,Junhu ZHAO 6 ,WeixinJIN 7 ,andHuihuiYUAN 1 1DepartmentofEarthSystemScience,TsinghuaUniversity,Beijing100084 2CollegeofResourcesandEnvironment,UniversityofChineseAcademyofSciences, Beijing101408 3BeijingYanshanEarthCriticalZoneNationalResearchStation,UniversityofChinese AcademyofSciences,Beijing101408 4CollegeofResourcesandEnvironment,UniversityofChineseAcademyofSciences, Beijing100049 5StateKeyLaboratoryofClimateSystemPredictionandRiskManagement,National ClimateCenter,ChinaMeteorologicalAdministration,Beijing100081 6ChinaMeteorologicalAdministrationKeyLaboratoryforClimatePredictionStudies, NationalClimateCenter,ChinaMeteorologicalAdministration,Beijing100081 7Microsoft,Beijing100080 Abstract Seasonalprecipitationanomaliesarelargelyregulatedbyatmosphericcirculation, whichdynamicalmodelspredictwithgreaterreliabilitythanprecipitationitself.Here, weemployadeeplearningmodelthattranslatesdynamicalcirculationpredictions intoprecipitationestimates.PredictionsinitializedfromMarchtoMayconsistently indicateadryanomalyovercentralChinainsummer2026.Retrospectiveevaluations revealedhigherpredictiveskillintheanalogueyears,whichalsotendedto featurecentralequatorialPacificwarmingpersistingfromtheprecedingwinterinto summer.Thiswarmingfavorsananomalouscycloniccirculationoverthewestern NorthPacific–SouthChinaSea–SouthChinaregion,whichinducesnortherlywinds andmoisturedivergencethatjointlysuppressrainfallovercentralChina.Supporting † Theseauthorscontributedequally:AnranWangandWenShi. Correspondenceto:YongLuo(yongluo@tsinghua.edu.cn) thismechanism,layer-wiserelevancepropagation(LRP)independentlyidentifies thesenortherlywindsasthedominantdriverofthepredictionamongallmodelinputs. Perturbationtestssupportedthisattribution:removingLRP-identifiedfeatures effectivelyeliminatesthedryanomaly.Ourframeworkthusprovidesphysically interpretableexplanationsforAI-derivedregionalclimateprojections,facilitating evidence-basedassessmentbeforeobservationaldatabecomeavailable. Introduction Credibleandskillfulprecipitationpredictionsissuedaheadofthesummerflood seasoninChinacouldprovideactionableleadtimeforwatermanagement, agriculturalplanning,anddroughtandfloodpreparedness 1,2 .Yetregionalsummer precipitationremainsoneofthemostchallengingtargetsofseasonalpredictionowing tocomplexinteractionsamongatmospheric,oceanicandland-surfaceprocesses 3–6 . Seasonalprecipitationskillisgenerallymodestandvariessubstantiallywiththe prevailingclimatestate 7 ,creatingintermittentwindowsofopportunityformore skillfulprediction 8–11 .Averagingskilloverahindcastperiodcanobscurethis state-dependentpredictabilityandprovideonlylimitedguidanceonthecredibilityof anindividualreal-timeprediction 12 .Therefore,tracingapredictiontophysically meaningfulpredictivesignalssupportedbythehistoricalrecordcanhelpassess whetheritprovidesacrediblebasisforissuinganearlywarning. Suchpredictivesignalscanbeinvestigatedthroughhistoricalclimatediagnostics andanalysesofmodelbehavior.Historicalanalogueanalysesidentifypastcases resemblingthetargetanomaly,andcompositeanalysesassesswhethertheyshare recurrentpatternsofatmosphericcirculation,moisturetransport,oroceanic conditions 13–15 .Atthemodellevel,featureattributionmethodsinexplainableartificial intelligence(XAI)identifytheinputregionsandvariablesthatcontributetoa particularprediction 9,16–22 .Becauseattributionresultscanvaryamongmethods,their reliabilityiscommonlyassessedthroughcross-methodcomparisons 23,24 .Perturbation testsexaminewhetherchangingtheidentifiedfeaturesaltersthepredictionas expected 25,26 .Evidenceforthecredibilityofanindividualpredictionisstrengthened whenhistoricaldiagnosticsandfeatureattributionconvergeonthesamephysically interpretablesignals. Dynamical–statisticalbridgingprovidesasuitableframeworkinwhichthese formsofevidencecanbebroughttogether.Suchapproachestranslatedynamical predictionsofrelativelypredictablelarge-scaleclimatemodesintoregional precipitationanomaliesthroughstatisticalrelationships 27–30 .Morerecently, deep-learningmodelshaveexpandedthisframeworkbylearningnonlinear relationshipsbetweenmultivariablecirculationfieldsandregionalprecipitation 31–33 . Thisexplicitcirculation-to-precipitationmappingallowsfeatureattributiontoidentify physicallyinterpretablecirculationsignalsunderlyinganindividualprediction.Yet thisinterpretivecapabilityhasrarelybeenusedtosupportphysicallygrounded assessmentwhenseasonalprecipitationpredictionsareissuedinrealtime 11,34 . Here,weuseacirculation-to-precipitationbridgingmodeltogeneratereal-time predictionsofsummer2026precipitationoverChina.PredictionsinitializedinMarch, AprilandMayconsistentlyindicateadryanomalyovercentralChina.Weassessthe credibilityofthepredictedanomalybyevaluatingmodelskillinhistoricalanalogue yearsandexaminingwhethertheclimateconditionsassociatedwiththeseyearsare alsoevidentin2026.Wethenusefeatureattributiontodeterminewhetherthemodel highlightscirculationsignalsconsistentwiththeseclimatefeatures.Weevaluatethe stabilityoftheattributionsacrossmethodsanduseLRP-guidedperturbationteststo assesstherelevanceoftheidentifiedfeaturestothepredictedanomaly.Together, theseanalysesestablishaframeworkforassessingAI-basedseasonalprecipitation predictionsthroughconvergenthistorical,physicalandmodel-centeredevidence availableaheadofthetargetseason. Results 1.ConsistentpredictionspointtoacentralChinadryanomalyinsummer2026 Wegeneratedreal-timepredictionsofsummer2026precipitationoverChinaby applyingthecirculation-to-precipitationbridgingmodeltodynamicalcirculation predictionsinitializedinMarch,AprilandMay.Allthreepredictionsshoweda broadlyconsistentspatialpattern,withbelow-normalprecipitationacrosscentral Chinaandabove-normalprecipitationoverSouthChinaandpartsofNorthChina(Fig. 1a–c).ThecentralChinadryanomalywasthemostprominentregionalfeature,with regionalmeansof−34%,−29%and−38%inthethreepredictions,respectively. AlthoughcentralChinashowedthelargestanomalymagnitude,itsinter-initialization rangewascomparativelysmall,whereaslargerrangesoccurredoverpartsofSouth andNortheastChina(Fig.1d).Thelocation,spatialextentandmagnitudeofthe centralChinadryanomalythereforeremainedrelativelystableastheinitialization approachedthetargetseason,indicatingtheanomalywasnotspecifictoaparticular initializationmonth. Cross-validationover1993–2025showedcomparableskillbetweenthe bridging-modelanddynamicalmulti-modelensembles(MMEs;Fig.1e,f).The bridging-modelMMEyieldedanomalycorrelationcoefficient(ACC)valuesof 0.140–0.161andpredictionscores(Ps)of73.27–74.25%acrossthethreeinitialization months.RelativetothedynamicalMME,itslargestgainsoccurredfortheMarch initialization,whereasthetwoMMEsperformedsimilarlyinMay.Theseresults motivateacase-specificcredibilityassessmentofthepredicteddryanomalyover centralChina. Fig.1:Summer2026precipitationpredictionsoverChinaandcross-validatedhindcast skill.a–c,Precipitationanomalies(%)relativetothe1991–2020climatologypredictedbythe bridging-modelmulti-modelensemble(MME)forJune–August2026,basedontheMarch, AprilandMayinitializations,respectively.d,Inter-initializationrangeateachstation, calculatedasthedifferencebetweenthemaximumandminimumanomaliesamongthethree predictions.Redboxesina–ddelineatecentralChina(28°–36°N，103°–113°E),andthe valuesina–cindicatethecorrespondingregional-meanprecipitationanomalies.e,f, Historicalcross-validatedanomalycorrelationcoefficient(ACC;e)andpredictionscore(Ps;f) forsummerprecipitationpredictionsinitializedinMarch,AprilandMayduring1993–2025. GrayandhatchedbarsrepresenttherawdynamicalMMEandthebridging-modelMME, respectively.Errorbarsindicatebootstrap95%confidenceintervalsforthemean cross-validatedscores.Valuesabovethebarsindicatethedifferencesbetweenthe bridging-modelandrawdynamicalMMEs. 2.Higherpredictionskillinhistoricaldry-anomalyyears Thehistoricalcasesmostcloselyresemblingthepredicted2026precipitation patternwerehighlyconsistentacrossinitializations.Foreachinitialization,the observedsummersduring1993–2025wererankedseparatelyusingfourmetricsof similaritytothecorrespondingprediction.Thefourrankswerethensummedforeach year,andthesevenyearswiththelowestranksumswereretainedforanalysis (MethodsandSupplementaryTable1).Sixyears—1994,1997,2001,2006,2011and 2015—werecommontoallthreesets.TheAprilandMaypredictionsyielded identicalsets,additionallyincluding2002,whereastheMarchpredictionincluded 2022instead. Wenextcomparedcross-validatedpredictionskillbetweentheselected dry-anomalyyearsandtheremaining26yearstoexaminehowthemodelhad performedinhistoricalcasesresemblingthe2026prediction(Fig.2).Acrossallthree initializationsandbothyeargroups,ERA5-drivenoutputsattainedmarkedlyhigher ACCandPsthanpredictionsdrivenbydynamical-modelcirculation,suggestingthat errorsinthepredictedcirculationpartlyconstrainedreal-timepredictionskill.Despite theirdifferentoverallskilllevels,bothinputsettingsshowedacommonconditional pattern:medianskillwashigherinthedry-anomalyyearsineverycomparisonexcept ACCfortheMay-initialized,dynamicallydrivenpredictions,forwhichthedry-year medianwasslightlylowerandthetwodistributionsoverlappedsubstantially.These comparisonssuggestthatthemodelmayperformbetterforprecipitationstates resemblingthepredicted2026dryanomalythanitsaveragehistoricalskillwould imply.Becausethebridging-modelpredictionsarederivedentirelyfromatmospheric circulation,thehigherskillinthedry-anomalyyearssuggeststhatarecurrent circulationpatternmayunderliethemodel’spredictionofcentralChinadrying. Fig.2:Predictionskillinanalogueyearsforthepredicted2026dryanomaly.a–c,ACC ofsummerprecipitationoutputsfortheMarch,AprilandMayinitializationsduring 1993–2025,groupedintothe7dry-anomalyanalogueyearsselectedforeachinitialization andtheremaining26years.d–f,correspondingPs.Grayboxplotsrepresentbridging-model outputsdrivenbyERA5circulation,whereasblueboxplotsrepresentbridging-model predictionsdrivenbydynamical-modelcirculation.Withineachinputsetting,darkerand lightershadesdenoteanalogueyearsandotheryears,respectively. 3.ClimateconditionsconducivetoadryanomalyovercentralChina Wecompositedthesummercirculationacrossthesevenanalogueyearsselected fortheMay-initializedbridging-modelpredictionandcomparedthemwiththe May-initializeddynamicalMMEcirculationpredictionforsummer2026(Fig.3).The March-andApril-initializedpredictionsshowedsimilarspatialpatterns (SupplementaryFigs.1and2). Intheuppertroposphere,theanaloguecompositeshowedanorthward displacementoftheEastAsiansubtropicalwesterlyjet(EASWJ),withstrengthened 200-hPazonalwindsnorthoftheclimatologicaljetaxisandweakenedwindstoits south(Fig.3e).TheMay-initializedpredictionreproducedthismeridionaldipole(Fig. 3i).Previousobservationalstudieshaveassociatedapoleward-displacedEASWJwith reducedsummerrainfallovertheYangtze–HuaiheRiverValleyandaredistribution ofrainfalltowardSouthandNortheastChina 35–37 . Inthemiddletroposphere,thecompositefeaturedapronouncedpositive500-hPa geopotential-heightanomalyextendingfromMongoliatosouthofLakeBaikal(Fig. 3f).Thistypeofcontinentalhighanomalyhasbeenassociatedwithenhanced subsidencesouthofLakeBaikal,anomalouslow-levelnortherlywindsovereastern China,andaweakenedEastAsiansummermonsoonwithreducedmoisture transport 38 .TheMay-initializedpredictionreproducedthepositive-heightanomaly butplaceditscenterfartherwest(Fig.3j). At700hPa,boththedry-anomalycompositeandthe2026predictionfeatureda cycloniccirculationanomalyoverthewesternNorthPacificthatextendedwestward intotheSouthChinaSea–SouthChinaregion(Fig.3g,k).CentralChinalayonthe northwesternflankofthiscirculation,whereanomalousnortheasterlyflowwas strongerinthe2026predictionthaninthehistoricalcomposite.Consistentwiththis circulation,the850–500-hPalayer-integratedmoisture-fluxanomalyopposedthe climatologicalmoisturetransportintocentralChina,andthetargetregionshowed anomalousmoisture-fluxdivergence(Fig.3d,h,l).The2026predictionalsoshowed anomalousmoisture-fluxconvergenceoverSouthChina,consistentwiththe above-normalprecipitationpredictedthere(Figs.3land1). Amongthesecirculationanomalies,thewesternNorthPacific–SouthChinaSea –SouthChinaregioncyclonicanomalymostdirectlyfavoredcentralChinadryingby weakeningmoisturetransportintothetargetregion,withthestrongeranomalous windsassociatedwiththiscirculationinthe2026predictionsuggestingthatitsdrying influencemaybestrongerthaninthecompositeofthesevenanalogyears. Fig.3:CirculationpatternsassociatedwiththepredictedcentralChinadryanomaly. a–d,ERA5JJAclimatologyfor1993–2016.e–h,Compositeanomaliesfortheseven historicaldry-anomalyyearsselectedfortheMayinitialization.i–l,May-initialized dynamicalMMEanomaliesforJJA2026.Columnsshow200-hPazonalwind(U₂₀; shadingandcontours;ms⁻¹),500-hPageopotentialheight(Z₅₀;shadingandcontours; gpm),700-hPaspecifichumidity(q₇₀;shading;gkg⁻¹)andhorizontalwinds(vectors;m s⁻¹),and850–500-hPaintegratedvaportransportcalculatedfromthe850-,700-and500-hPa levels(IVT;vectors;kgm⁻¹s⁻¹)anditsdivergence(div.;shading;10⁻⁵kgm⁻²s⁻¹).Red boxesdelineatethecentralChinatargetregion.Blue5880-gpmcontoursinb,fandj delineatethewesternNorthPacificsubtropicalhigh.Ind,IVTvectorsareshownonlywhere theirmagnitudeis≥50kgm⁻¹s⁻¹.Inh,vectorsareretainedwheretheirmagnitudeis≥6kg m⁻¹s⁻¹andatleastonehorizontalcomponentdiffersfromzeroinatwo-sidedone-sample t-test(p<0.20);divergenceshadingisrestrictedtogridpointssatisfyingthesame significancethreshold.Inl,vectorsareshownwheretheirmagnitudeis≥6kgm⁻¹s⁻¹,and divergenceshadingisretainedwhereatleastsixofthesevenmodelsagreeonthesignofthe MME-meandivergenceanomaly. Therecurrentcirculationpatternwasaccompaniedbyatendencytowardswarmer conditionsinthecentralequatorialPacific.Niño-4SSTincreasedmarkedlyfromthe precedingwinterintosummerinsixofthesevenhistoricaldry-anomalyyears, whereasthecentralequatorialPacificwasalreadyanomalouslywarminthepreceding winterandremainedsothroughout2015(Fig.4).In2026,Niño-4rosecontinuously fromanegativeanomalyinDecember2025toabove1 °CbyMay2026.Recent climatemonitoringindicatesthatElNiñoconditionshavedevelopedinthetropical Pacific 39–41 ,andforecastsfurthersuggestthatthedevelopingElNiñointensitycould becomeexceptionallystrong 42 .Althoughthestrongestwarmingin2026waslocated farthereast,thewarmingextendedintotheNiño-4region.Previousstudiesfoundthat therelationshipbetweendevelopingElNiñoandreducedsummerrainfallovercentral Chinabecamesignificantafterthelate1980s,owingtotheincreasedfrequencyof central-PacificElNiñoevents 43 .WhenSSTwarmingiscenteredoverthecentral equatorialPacific,theassociatedWalkercirculationanomalyshiftswestward, extendingitsdescendingbranchtotheMaritimeContinentandfavoringa lower-troposphericcycloniccirculationanomalyoverthewesternNorth Pacific–SouthChinaSea 44 .Thecycloniccirculationweakensnorthwardmonsoon moisturetransport 43 .ThepronouncedNiño-4warmingin2026isconsistentwiththis mechanismandthepredictedreductioninsummerrainfallovercentralChina. Fig.4:Niño-4SSTanomalyevolutionin2026andhistoricalanalogueyears.Monthly Niño-4SSTanomalies(°C)fromDecemberoftheprecedingyeartoAugustofeachofthe sevenhistoricaldry-anomalyyearsselectedfortheMayinitialization.The2026seriesspans December2025–May2026.GrayshadingmarkstheJJAtargetseason. 4.PhysicallycoherentpredictionsignalsidentifiedbyLRP Toexaminewhetherthebridgingmodeldrewonthecirculationfeatures highlightedintheprevioussection,weappliedLRPtotheMay-initializedprediction forsummer2026(Fig.5).Wedefinedtheattributiontargetasthepredicted regional-meanprecipitationanomalyacrossthecentralChinastations.Thelinear mappingfromthepredictedprincipalcomponent(PC)coefficientsthroughempirical orthogonalfunction(EOF)reconstructionandregionalaveragingwasincorporatedas thefinaltargetlayerforrelevancepropagation,witheachmodeweightedbyits contributiontotheregionalanomaly.Thisformulationtiestheresultingattribution directlytotheregionalpredictionofinterest,avoidingseparateinterpretationand subsequentcombinationofmode-specificattributionmaps. Aphysicallycoherentattributionpatternwasevidentinthemeridionalwind fieldsat500,700and850 hPa(Fig.5d,g,j,n).Atallthreelevels,dry-supporting relevancewasconcentratedonthenortherlyanomaliesalongthenorthwesternflank ofthewesternNorthPacific–SouthChinaSea–SouthChinaregioncyclonic circulation,withaparticularlypronouncedsignalat700 hPa.Together,thethree meridional-windpredictorsaccountedfor37.5%ofthearea-weighteddry-supporting relevance.Thesenortherlyanomaliesopposetheclimatologicalmoisturetransport intocentralChina.Thepredictedfieldsshowedwidespreadpositivespecific-humidity anomaliesat500,700and850 hPa,whilethecorrespondingpredictorsreceivedlittle dry-supportingrelevance(Fig.5e,h,k,n).Theseresultsindicatethatthemodel associatedthepredictedrainfalldeficitprimarilywithcirculation-inducedweakening ofmoisturetransportintocentralChinaratherthanwithlimitedmoistureavailability. Thespatialdistributionofthisattributioncloselymatchedthecirculationpathway identifiedindependentlyfromtheanaloguecompositesandthe2026dynamical prediction. Beyondthemeridional-windsignals,dry-supportingrelevancein2-mair temperature(T2m)wasconcentratedoverananomalouslywarmregionupstreamof centralChina(Fig.5m,n).Thisspatialcorrespondenceindicatesthatthemodel associatedthetemperaturesignalwiththepredicteddryanomaly.The500-hPa geopotentialheight(Z500)fieldalsoreceiveddry-supportingrelevance(Fig.5b,n), concentratedmainlyalongstronggeopotential-heightgradientsontheWNPSH periphery.However,thealignmentofitsedgeswiththeVisionTransformer(ViT) patchgridsuggeststhatsomeofthespatialstructuremayarisefromtokenization artifacts. IntegratedGradients(IG)andGuidedIntegratedGradients(GuidedIG)broadly reproducedtheMay-initializedwind-fieldattributionpatterns(SupplementaryFigs.3 and4).Cross-methodagreementinpredictorrankingsremainedhighacrossallthree initializations(SupplementaryTable2). Fig.5:LRP-identifiedcirculationsignalssupportingthepredictedsummer2026dry anomalyovercentralChina.a–m,Spatialdistributionsofdry-supportingnegativeLRP relevanceforthe13circulationpredictors,calculatedfortheregional-meanprecipitation anomalyovercentralChinafromtheMayinitialization.Highnegativerelevanceindicatesa strongercontributiontowardthepredicteddryanomaly;onlynegative,dry-supporting relevanceisshown,andthefieldsaresmoothedusingaGaussianfilterwithσ=1gridcell. Contoursdenotethecorrespondingpredictedanomaliesforscalarfields,andvectorsshow horizontalwindanomaliesat500,700and850 hPainthepaireduandvpanels.Theredbox marksthetargetregion.n,Relativecontributionofeachpredictor,calculatedfromthe spatiallyintegratednegativerelevance.Here,u,v,qandzdenotezonalwind,meridional wind,specifichumidityandgeopotentialheight,withpressurelevelsindicatedbythe subscripts;mslandt2mdenotemeansea-levelpressureand2-mairtemperature. 5.PerturbationtestssupportthefaithfulnessofLRPexplanations TotestwhetherthecirculationsignalshighlightedbyLRPmateriallyinfluenced thepredicteddryanomaly,weconductedLRP-guidedRemovalandRetentiontests fortheMayinitialization(Fig.6).Perturbationresponseswereevaluatedthroughthe regionalLRPtarget, Φ � ,whichisequivalenttothemeanstandardizedprecipitation anomalyacrosscentralChinastations;positivevaluesindicatewetanomaliesand negativevaluesindicatedryanomalies.For2026anditssevenanalogyears, case-specificmaskswereconstructedfromthevariable–grid-cellfeatureswiththetop 5%ofdry-supportingrelevanceandtheirneighboringgridcells(Fig.6a).Themasked featureswerezeroedoutintheRemovaltestandretainedexclusivelyintheRetention test.Eachtestwasrepeatedwith100randommasksperturbingthesamefractionof theinput.Theprespecifiedone-sidedcomparisonstestedwhetherRemovalyieldeda higher Φ � andRetentionalower Φ � thanthecorrespondingrandom-mask controls. LRP-guidedRemovalreversedthesignof Φ � fromnegativetopositivefor 2026andallsevenanalogyears,whereasrandomRemovalleft Φ � closetothe baselines(Fig.6b).Conversely,LRP-guidedRetentionpreservednegative Φ � and producedmorenegativevaluesthanthecorrespondingbaselinesineverycase,while randomRetentionyieldedvaluesclusterednearzero(Fig.6c).Noneofthe100 randomperturbationsproducedaresponsemoreextremethantheLRP-guidedresult intheprespecifieddirectionforanyoftheeightcases(one-sidedempiricaltests,all �<0.01).ThedisappearanceofthedrysignalafterRemovalanditspersistence underRetentionindicatethatthecirculationsignalssupportingthepredicteddry anomalywereconcentratedintheinputfeatureshighlightedbyLRP,providing evidencethattheattributionfaithfullyreflectedthemodel’sdependenceonthese inputs. Fig.6:LRP-guidedperturbationtestsofcentralChinadry-anomalypredictions.a, Schematicofthezero-outandkeep-onlymasksappliedacrossthe13circulationinputfields fortheMayinitialization.TheLRP-guidedmaskscomprisethevariable–grid-cellfeatures withthetop5%ofdry-supportingrelevanceandtheirneighboringgridcells;theredbox delineatescentralChina.b,c, Φ � for2026anditssevenanalogyearsundertheRemoval(b) andRetention(c)tests.Openredcirclesshowthebaselines,coloreddotstheLRP-guided results,andgrayboxplotsthedistributionsfrom100random-maskcontrolsperturbingthe samefractionoftheinput. Φ � denotestheregionalLRPtarget. Discussion Thisstudyprovidesaprospective,case-specificassessmentofthepredicted summer2026dryanomalyovercentralChinausingonlyinformationavailableat eachpredictionstarttime.Historicalevaluationshowedgenerallyhigherskillinthe selectedanalogueyearsthaninotheryears.Thepredictedcirculationalsoreproduced featuresassociatedwithsuppressedprecipitationovercentralChina,includinga lower-troposphericcyclonicanomalyoverthewesternNorthPacific–SouthChinaSea –SouthChinaregion,northerlyanomaliesonitsnorthwesternflank,weakened monsoonmoisturetransport,andmoisture-fluxdivergenceovercentralChina.LRP indicatedthatthemodelreliedstronglyonthesenortherlyanomalies.Inperturbation tests,removingtheLRP-selectedfeaturesreversedthepredictedanomalyfromdryto wet,whereasretainingthesefeaturesaloneintensifiedthedryanomaly,supporting thefaithfulnessoftheLRPattribution. Inthisstudy,wedefinedtheLRPtargetasthepredictedregional-mean precipitationanomalyovercentralChina.Thenetworkoutputs512PCcoefficients, andtheregionalanomalyisreconstructedfromall512associatedEOFmodes. AttributingeachPCseparatelywouldnotdirectlyexplainwhycentralChinawas predictedtobedryandwouldrequirecombiningmanymode-specificmaps.Thus,we includedthelinearreconstructionfromall512PCstotheregionalmeaninthe relevance-propagationpathway.Theresultingmapintegratesallmodesintoasingle region-specificattributionthatcanbecompareddirectlywiththecirculationpatterns. Thisformulationmayalsobeusefulforregionalinterpretationofotherspatial predictionmodelswithreduced-orderoutputs. LRPattributesthemodeloutputtoinputfeaturesbutdoesnotestablishcausal relationshipsintheclimatesystem.Evidencewasstrongestforthe mid-to-lower-troposphericnortherlysignal,whichagreedwithclimatediagnostics andwasbroadlyreproducedbyIGandGuidedIG.Secondarysignalsrequirecaution. Inparticular,T2misarapidlyrespondingcomponentofthecoupledland–atmosphere system.Reducedprecipitationandcloudcover,togetherwithdiminishedsoil moistureandevaporativecooling,canincreasenear-surfacetemperature,producinga coupledwarm–drysurfacestate.Thebridgingmodelcanexploitthiscovariation,and thedry-supportingT2mrelevanceovernorthwesternChinamayindicatethatsucha coupledwarm–drystatecontributedpredictiveinformation.Itdoesnotdemonstrate thatwarmingovernorthwesternChinadirectlycausedthepredictedprecipitation deficitovercentralChina.Thepatch-alignedZ500relevancealsosuggeststhatits detailedspatialpatternmaypartlyreflectViTtokenization.Theperturbationtests confirmedthejointimportanceoftheselectedfeaturestothemodelprediction,but didnotestablishthemascausaldriversintheclimatesystem. Theframeworkexamineswhethersimilarhistoricalcasesweremorepredictable, whetherthepredictedcirculationsupportsaplausiblephysicalpathway,andwhether themodelreliesonfeaturesassociatedwiththatpathway.Forsummer2026,allthree analysesprovidedsupportiveevidence.Thevalueofthisframeworkliesinmaking theevidentialbasisofapredictionexplicitbeforeobservationsbecomeavailable, ratherthanguaranteeingthatthepredictionwillverify.Thelarge-scaleconditions capturedbytheframeworkmaymodulatesubseasonalrainfallvariabilityandthe likelihoodofextremeevents,butthemodeldoesnotexplicitlyresolvetheoccurrence, timingorintensityofindividualevents.Thepredicteddryanomalytherefore representsaseasonaltendencytowarddrierconditionsanddoesnotruleoutthe possibilityofheavyrainfallthatcouldsubstantiallyaffecttheobservedsummer precipitationtotal.Morebroadly,itprovidesphysicallygroundedexplanationsfor seasonalpredictionsofpotentiallyhigh-impactclimateanomaliesandsupportsthe transparentandevidence-baseduseofAIinpre-seasonclimate-riskassessment. Method 1.Data Monthlystationprecipitationobservationsfor1961–2025werecalculatedfrom theChinaPrecipitationDailyDataset(V3.0) 45 ,providedbytheNational MeteorologicalInformationCenteroftheChinaMeteorologicalAdministration. Theseobservationsservedasthepredictandfortransferlearning,thereferencefor cross-validationevaluation,andthebasisforidentifyinghistoricalanalogueyears. MonthlyNiño-4indexwereobtainedfromtheNOAAPhysicalSciencesLaboratory climate-indexarchive.TheindexisproducedbytheNOAAClimatePrediction CenterfromERSSTv5 46 andrepresentsthearea-meanSSTanomalyover5° S–5° N, 160° E–150° W. Monthlymeanatmosphericfieldsfor1961–2025wereobtainedfromERA5 47 . Thefieldswereinterpolatedonto1° × 1°gridcovering64.5°–159.5° Eand 3.5° S–59.5° N.Weused13variables:zonalwindat200 hPa;geopotentialheight, zonalandmeridionalwinds,andspecifichumidityat500 hPa;zonalandmeridional windsandspecifichumidityat700and850 hPa;meansea-levelpressure;and2-mair temperature.ERA5circulationfieldsservedasmodelinputsduringtransferlearning andretrospectiveevaluation.Theywerealsousedtodiagnosecirculationand moisturetransportintheanalogueyears. Seasonalpredictiondatawereobtainedfromthemonthlypressure-leveland single-leveldatasetsintheCopernicusClimateChangeService(C3S)ClimateData Store 48,49 .Thehindcastfor1993–2025comprised27forecast-systemversionsfrom eightforecastingcenters;thesystemsandtheirensemblesizesarelistedinTable1, withensemblesizesgiveninparentheses.Thesame13atmosphericvariablesandthe correspondingprecipitationformedpairedinputsandtargetsforpretraining.Because theoperationalsystemcompositionchangedduringspring2026,theMarch initializationusedsevensystems(BOM-2,CMCC-4,DWD-22,ECCC-5, ECMWF-51,Météo-France-9andUKMO-605),whereastheAprilandMay initializationsusedeight(BOM-2,CMCC-4,DWD-22,ECCC-5,ECMWF-51, JMA-4,Météo-France-9andUKMO-610).Eachsystemwasprocessedseparately, andtheresultingprecipitationpredictionswereaveragedwithequalsystemweightsto formtheMMEprediction.Allmonthlydatawereconvertedtooverlapping three-monthmeans,yielding12seasonalsamplesperyear. forecastingcentermodel ECMWF SEAS5(25)，SEAS5.1(25) MFMétéo-FranceSystem6(25),7(25),8(25),9(31) UKMO GloSea5-GC212(28),13(28),GloSea5-GC2-LI14(28), 15(28),GloSea6600(28),601(28),602(28),603(28),604 (28) CMCCCMCC-SPS3(40),3.5(40),4(30) DWDGCFS2.0(30),2.1(30),2.2(30) ECCC GEM5-NEMO(10),CanESM5.1p1bc(20), GEM5.2-NEMO(20) JMAJMA-CPS2(10),3(10) BOMACCESS-S2(27) Table1.ForecastingcentersandC3Sseasonalpredictionsystemsusedformodel pretraining. 2.Bridging-modeldevelopmentandevaluation Thecirculation-to-precipitationbridgingmodelmaps13seasonal circulation-anomalyfieldstostationprecipitationanomalies(Fig.7).Themodelisan updatedversionofthecirculation-to-precipitationbridgingmodeldevelopedbyJinet al. 32 ,withaViTarchitecturereplacingtheoriginalconvolutionalbackbone.The multichannelfieldsweredividedintopatchesandtransformedintospatialtokens throughconvolutionalpatchembeddingandbottleneckprojection.Thetarget three-monthseasonwasembeddedasaclassification(CLS)token,whichwas prependedtothespatial-tokensequenceandusedtoconditionfeatureextraction throughadaptivelayernormalization.Afterpositionalembeddingswereadded,the sequencewasprocessedbysixViTblocks.TheupdatedCLStokenwaspassed throughrootmeansquarenormalization,dropoutandtwofullyconnectedlayersto predictthePCcoefficientsoftheleading512EOFmodes.Stationprecipitation anomalieswerethenreconstructedfromthepredictedcoefficientsandtheEOFbasis. Becauseeachthree-monthwindowyieldsonlyoneseasonalsample,the observationdatasetisrelativelysmall.Themodelwasthereforepretrainedusing pairedcirculationandprecipitationanomaliesfromdynamicalmodels.During transferlearning,thefinaltwolinearlayerswerefittedbyridgeregressionusing ERA5circulationanomaliesandobservedprecipitationanomalies. Fig.7:Architectureofthecirculation-to-precipitationbridgingmodel.Spatial informationisintegratedwithineachViTblockviaMulti-HeadSelf-Attention(MHA),while target-seasoninformationfromtheCLStokenmodulatesthemainbranchoftheneural networkthroughshifting,scaling,andgating.PurpletrapezoidsindicatetheEOFprojection usedtoderivethePCtargetsandreconstructstationprecipitationanomalies.� � ,positional embedding;CLS,classification;PC,principalcomponent. Cross-validationfor1993–2025usedsixcontiguousyearblocks.Ineachfold, oneblockwasheldoutfortesting,theprecedingblockincyclicorderforvalidation, andtheremainingfourfortraining.Thesamesplitswereappliedduringpretraining andtransferlearning.Predictionsforthesixtestblocksformedthecomplete cross-validatedrecord. PredictionperformancewasevaluatedusingthespatialACCand� � .For� stations,ACCwascalculatedas 퐴�= � =1 � (� � −� )(� � −� ) � =1 � (� � −� ) 2 � =1 � (� � −� ) 2 where� � and� � arethepredictedandobservedprecipitationanomalypercentages atstation�,andtheoverbarsdenotespatialmeans.FollowingtheNationalClimate Centerspecification 50 ,the � � isanempiricaloperationalmetricthatcombines anomaly-gradeaccuracywithgrade-dependentweightingofprecipitation anomalies: � � = 2×� 0 +2× � 1 +4× � 2 �+� 0 + 2×� 1 +4× � 2 +M ×100% Here,�isthetotalnumberofstations;� 0 denotescorrectanomaly-sign predictions,� 1 and� 2 denotecorrectlypredictedlevel-1andlevel-2anomalies, and�denotesthenumberofstationsatwhichtheobservedprecipitationanomaly wasatleast100%,butthepredictedanomalyfailedtoreachthesecond-level positive-anomalycategory.Level-2andlevel-1anomaliescorrespondto 20%≤ ∣Δ�∣<50%and∣Δ�∣≥50%. 3.Identificationofhistoricalanalogueyears Foreachinitialization,observedJJAprecipitationanomaliesduring1993–2025 werecomparedwiththecorresponding2026predictionusingfourmetrics:all-station ACC( ACC all ),CentralChinaACC( ACC C ),drycoverage( DC ),anddry-intensity distance( DID ).Let � � denotetheobservedprecipitationanomalypercentageat CentralChinastationsinyear � , � thecorrespondingpredictionfor2026. DC and DID werecalculatedas DC y = � � − � � ×100% , DID y =� � −� where � � − isthenumberofCentralChinastationswithnegativeanomaliesand � � isthetotalnumberofstationsintheregion.Yearswererankedseparatelyfor eachmetric,withrank1assignedtotheclosestmatch:largervaluesforthetwoACCs and DC ,butsmallervaluesfor DID .Theoverallranksumwas � � =� � ACC all +� � ACC C +� � DC+� � (DID) Thesevenyearswiththesmallest� � wereretainedashistoricalanalogues,closely correspondingtothetop20%ofranked33years.Tieswereresolvedbysmaller DID , followedbyhigherCentralChinaandall-stationACCs. 4.Layer-wiserelevancepropagationandperturbationtests LRPattributesascalarneural-networkoutputtoitsinputfeaturesby redistributingrelevancebackwardthroughthenetworkusinglayer-specificrules 51 . Thisredistributionapproximatelyconservestotalrelevancebetweenadjacentlayers, yieldingsignedscoresthatindicatefeaturessupportingoropposingtheselected output.Inclimatescience,LRPhasbeenusedtointerpretthespatialpatterns underlyingneural-networkpredictions,identifyindicatorsofexternallyforcedchange, anddiagnoseclimatestatesassociatedwithenhancedpredictability 52–54,9 . Here,weusedLRPtotracethecirculationsignalscontributingtothepredicted centralChinadryanomaly.ToaccommodatetheViTarchitecture,weadoptedthe CP-LRPrelevance-propagationscheme 55 .Withineachself-attentionlayer,attention weightswereheldfixedandrelevancewaspropagatedthroughthevaluepath, withoutredistributionthroughthequery–keyscoresorsoftmaxoperation. Normalizationandpositionalencodingweretreatedaspass-throughoperations.We usedaγ-rulefortheconvolutionalpatch-embeddinglayersandanε-ruleforlinear layers.LRPrequiresascalarexplanationtarget,whereasthebridgingmodelpredicts �=512precipitationPCcoefficients.Wethereforedefinedthetargetasthe predictedmeanstandardizedprecipitationanomalyoverthecentralChinaregion�: Φ � �= 1 � � �∈� � � (�)= � =1 � � �,� � � (x) � �,� = 1 � � �∈� 퐸� �,� where � � isthenumberofstationswithin � , � denotesthecirculationinput, � � isthepredictedcoefficientofthe�thEOFmode,and� � , � isthemeanloadingof thatmodeacrossthe � � stationsin � .Thisreconstructionwasimplementedasa fixedlinearlayer,allowingrelevancetopropagatefromtheregionaltargetthrough thePCoutputstothecirculationinputs.Forvisualization,onlynegativerelevance valueswereretained,becausetheycontributedtowardamorenegative Φ � .The relevancefieldforeachpredictorwassmoothedindependentlyusinga two-dimensionalGaussianfilterwith�=1푔�cell.Predictorcontributionswere calculatedbycosine-latitude-weightedintegrationofthedry-supportingrelevanceand normalizedacrossthe13predictors. LRP-guidedperturbationtestswereconductedfortheMay-initialized2026 predictionandthesevenanalogueyears.Dry-supportingrelevancevalueswere pooledacrossallgridcellsandall13predictors,andthestrongest5%wereselected usingasingleglobalthreshold.Theresultingmaskwasexpandedbyonegridcellin eachspatialdirection.IntheRemovaltest,theselectedinputelementsweresetto zero;intheRetentiontest,onlytheselectedelementswereretained.Theincreasein Φ � afterRemovalandtheretained Φ � intheRetentiontestwerecomparedwith resultsfrom �=100 randommaskscontainingthesamenumberofinputelements astheexpandedLRPmask.One-sidedempirical � valueswerecalculatedas �=(�+1)/(�+1) FortheRemovaltest,�wasthenumberofrandommasksproducinganequalor largerincreasein Φ � thantheLRP-guidedmask;fortheRetentiontest,itwasthe numberproducinganequalorlowerretainedΦ � .Thesmallestattainable�value wastherefore1/101(<0.01). 5.IntegratedGradientsandGuidedIntegratedGradients LikeLRP,IGandGuidedIGproducesignedfeature-levelattributionsfora specifiedscalaroutput.Theywerethereforeusedtoassesswhetherthemain dry-supportingsignalsidentifiedbyLRPwerereproducedbymethodsbasedon differentattributionprinciples.IGintegratesgradientsalongastraightpathfroma referencestatetotheinput 56 .GuidedIGfollowsanadaptivepathdesignedtoreduce theaccumulationofnoisygradients 57 .BothmethodswereappliedtothefrozenViT usingtheregionalattributiontarget Φ � .Becausetheinputswerestandardized anomalies,azerofieldrepresentedtheclimatologyofallpredictors.Thisfieldserved asthecommonbaselineforevaluatinghowinputanomalieschangedtheregional prediction.Thesamebaselineandattributionprocedureswereusedforpredictions initializedinMarch,AprilandMay.Negativevaluesindicatedry-supporting contributions.TheIGandGuidedIGattributionfieldsinSupplementaryFigs.3and4 weresmoothedforvisualizationinthesamemannerastheLRPfields.Cross-method agreementinpredictorrankingsandattributionpatternsisshowninSupplementary Table2. Reference 1.Golding,N.etal.Co-developmentofaseasonalrainfallforecastservice: SupportingfloodriskmanagementfortheYangtzeRiverbasin.Clim.RiskManag. 23,43–49(2019). 2.BrunoSoares,M.,Daly,M.&Dessai,S.Assessingthevalueofseasonalclimate forecastsfordecision-making.WileyInterdiscip.Rev.Clim.Change9,e523 (2018). 3.Shi,P.etal.SignificantlandcontributionstointerannualpredictabilityofEast Asiansummermonsoonrainfall.EarthsFuture9,e2020EF001762(2021). 4.Wang,B.etal.AdvancingAsianmonsoonclimatepredictionunderglobal change:Progress,challenges,andoutlook.Adv.Atmos.Sci.43,1–29(2026). 5.Ma,J.etal.SkillfulseasonalpredictionsofcontinentalEast-Asiansummer rainfallbyintegratingitsspatio-temporalevolution.Nat.Commun.16,273(2025). 6.He,C.etal.HowmuchoftheinterannualvariabilityofEastAsiansummer rainfallisforcedbySST?Clim.Dyn.47,555–565(2016). 7.Pegion,K.&Kumar,A.DoesanENSO-conditionalskillmaskimproveseasonal predictions?Mon.WeatherRev.141,4515–4533(2013). 8.Mariotti,A.etal.Windowsofopportunityforskillfulforecastssubseasonalto seasonalandbeyond.Bull.Am.Meteorol.Soc.101,E608–E625(2020). 9.Mayer,K.J.&Barnes,E.A.Subseasonalforecastsofopportunityidentifiedby anexplainableneuralnetwork.Geophys.Res.Lett.48,e2020GL092092(2021). 10.Arcodia,M.C.etal.Assessingdecadalvariabilityofsubseasonalforecasts ofopportunityusingexplainableAI.Environ.Res.Clim.2,045002(2023). 11.Dunstone,N.etal.Windowsofopportunityforpredictingseasonalclimate extremeshighlightedbythePakistanfloodsof2022.Nat.Commun.14,6544 (2023). 12.Borchert,L.F.,Düsterhus,A.,Brune,S.,Müller,W.A.&Baehr,J. Forecast-OrientedAssessmentofDecadalHindcastSkillforNorthAtlanticSST. Geophys.Res.Lett.46,11444–11454(2019). 13.Duan,S.,Ullrich,P.&Boos,W.R.MeteorologicalDriversofNorth AmericanMonsoonExtremePrecipitationEvents.J.Geophys.Res.Atmos.129, e2023JD040535(2024). 14.Li,L.&Dolman,A.J.Onthereliabilityofcompositeanalysis:anexample ofwetsummersinNorthChina.Atmos.Res.292,106881(2023). 15.Yang,Y.,Zhai,P.,Li,J.&Wang,Q.RainbeltPropertiesofPersistentHeavy PrecipitationovertheYangtzeRiverBasinandAssociatedThree-Dimensional Circulations.WeatherForecast.40,689–702(2025). 16.Pegion,K.,Becker,E.J.&Kirtman,B.P.UnderstandingPredictabilityof DailySoutheastU.S.PrecipitationUsingExplainableMachineLearning.Artif. Intell.EarthSyst.1,e220011(2022). 17.Kalashnikov,D.A.etal.PredictingCloud-To-GroundLightninginthe WesternUnitedStatesFromtheLarge-ScaleEnvironmentUsingExplainable NeuralNetworks.J.Geophys.Res.Atmos.129,e2024JD042147(2024). 18.Camps-Valls,G.etal.Artificialintelligenceformodelingandunderstanding extremeweatherandclimateevents.Nat.Commun.16,1919(2025). 19.Straaten,C.van,Whan,K.,Coumou,D.,Hurk,B.vanden&Schmeits,M. CorrectingSubseasonalForecastErrorswithanExplainableANNtoUnderstand MisrepresentedSourcesofPredictabilityofEuropeanSummerTemperatures.Artif. Intell.EarthSyst.2,e220047(2023). 20.Mamalakis,A.UnravelingWinterPrecipitationPredictabilityoverCONUS viaDeepLearningandExplainableArtificialIntelligence.Artif.Intell.EarthSyst. 5,250105(2026). 21.Liu,Q.etal.Deep-learningpost-processingofshort-termstation precipitationbasedonNWPforecasts.Atmos.Res.295,107032(2023). 22.Martin,Z.K.,Barnes,E.A.&Maloney,E.UsingSimple,Explainable NeuralNetworkstoPredicttheMadden-JulianOscillation.J.Adv.Model.Earth Syst.14,e2021MS002774(2022). 23.Krell,E.,Mamalakis,A.,King,S.A.,Tissot,P.&Ebert-Uphoff,I.The influenceofcorrelatedfeaturesonneuralnetworkattributionmethodsin geoscience.Environ.DataSci.4,e29(2025). 24.Mamalakis,A.,Barnes,E.A.&Ebert-Uphoff,I.InvestigatingtheFidelityof ExplainableArtificialIntelligenceMethodsforApplicationsofConvolutional NeuralNetworksinGeoscience.Artif.Intell.EarthSyst.1,e220012(2022). 25.Bommer,P.L.,Kretschmer,M.,Hedström,A.,Bareeva,D.&Höhne,M. M.-C.FindingtheRightXAIMethod—AGuidefortheEvaluationandRankingof ExplainableAIMethodsinClimateScience.Artif.Intell.EarthSyst.3,e230074 (2024). 26.Fong,R.C.&Vedaldi,A.InterpretableExplanationsofBlackBoxesby MeaningfulPerturbation.in3429–3437(2017). 27.Prein,A.F.etal.Sub-SeasonalPredictabilityofNorthAmericanMonsoon Precipitation.Geophys.Res.Lett.49,e2021GL095602(2022). 28.Strazzo,S.etal.ApplicationofaHybridStatistical–DynamicalSystemto SeasonalPredictionofNorthAmericanTemperatureandPrecipitation.Mon. WeatherRev.147,607–625(2019). 29.Peng,Z.etal.StatisticalcalibrationandbridgingofECMWFSystem4 outputsforforecastingseasonalprecipitationoverChina.J.Geophys.Res.Atmos. 119,7116–7135(2014). 30.Li,Y.,Xü,K.,Wu,Z.,Zhu,Z.&Wang,Q.J.Astatistical–dynamical approachforprobabilisticpredictionofsub-seasonalprecipitationanomaliesover 17hydroclimaticregionsinChina.Hydrol.EarthSyst.Sci.27,4187–4203(2023). 31.Lyu,Y.etal.Improvingsubseasonal-to-seasonalpredictionofsummer extremeprecipitationoversouthernChinabasedonadeeplearningmethod. Geophys.Res.Lett.50,e2023GL106245(2023). 32.Jin,W.etal.DeeplearningforseasonalprecipitationpredictionoverChina. J.Meteorol.Res.36,271–281(2022). 33.Gibson,P.B.etal.Trainingmachinelearningmodelsonclimatemodel outputyieldsskillfulinterpretableseasonalprecipitationforecasts.Commun.Earth Environ.2,159(2021). 34.Yang,R.etal.Interpretablemachinelearningforweatherandclimate prediction:Areview.Atmos.Environ.338,120797(2024). 35.Wang,S.,Zuo,H.,Zhao,S.,Zhang,J.&Lu,S.HowEastAsianwesterly jet’smeridionalpositionaffectsthesummerrainfallinYangtze-HuaiheRiver Valley?Clim.Dyn.51,4109–4121(2018). 36.Yan,Y.,Li,C.&Lu,R.MeridionalDisplacementoftheEastAsian Upper-troposphericWesterlyJetandItsRelationshipwiththeEastAsianSummer RainfallinCMIP5Simulations.Adv.Atmos.Sci.36,1203–1216(2019). 37.Ling,S.,Lu,R.,Liu,H.&Yang,Y.InterannualMeridionalDisplacementof theUpper-TroposphericWesterlyJetoverWesternEastAsiainSummer.Adv. Atmos.Sci.40,1298–1308(2023). 38.TAN,G.,SUN,Z.,LIN,Z.&JIA,J.LandhighoverareasouthtoLake BaikalanditsrelationwithEastAsiansummermonsoonandclimateanomaliesof China.Clim.Environ.Res.13,791–799(2008). 39.ElNiñoisforecasttointensify,increasinglikelihoodofextremeweather. WorldMeteorologicalOrganization https://wmo.int/news/media-centre/el-nino-forecast-intensify-increasing-likelihood -of-extreme-weather(2026). 40.ClimatePredictionCenter:ENSODiagnosticDiscussion. https://w.cpc.ncep.noaa.gov/products/analysis_monitoring/enso_advisory/ensod isc.shtml?utm_source=chatgpt.com. 41.ElNinoMonitoringandOutlook/TCC. https://ds.data.jma.go.jp/tcc/tcc/products/elnino/outlook.html?utm_source=chatgpt. com. 42.Dinneen,J.ThisElNiñoissettobethelargestonrecordbya‘mind-blowing margin’.Natured41586-026-02293-y(2026)doi:10.1038/d41586-026-02293-y. 43.Chen,L.etal.InterdecadalchangeintheinfluenceofElNiñointhe developingstageonthecentralChinasummerprecipitation.Clim.Dyn.59, 1265–1282(2022). 44.Wang,H.&Wang,C.Large-ScaleAnomalousCycloneintheWestern NorthPacific.J.Clim.36,5895–5906(2023). 45.Zhihua,R.,Yu,Y.,Fengling,Z.&Yan,X.Qualitydetectionofsurface historicalbasicmeteorologicaldata.J.Appl.Meteorol.Sci.23,739–747(2012). 46.Huang,B.etal.ExtendedReconstructedSeaSurfaceTemperature,Version5 (ERSSTv5):Upgrades,Validations,andIntercomparisons.J.Clim.30,8179–8205 (2017). 47.Hersbach,H.etal.TheERA5globalreanalysis.Q.J.R.Meteorol.Soc.146, 1999–2049(2020). 48.CopernicusClimateChangeService,ClimateDataStore.Seasonalforecast monthlystatisticsonsinglelevels.https://doi.org/10.24381/cds.68d14c3(2018). 49.CopernicusClimateChangeService,ClimateDataStore.Seasonalforecast monthlystatisticsonpressurelevels.https://doi.org/10.24381/cds.0b79e7c5 (2018). 50.NationalClimateCenter,ChinaMeteorologicalAdministration.Forecast evaluationmethodsandparameters. 51.Bach,S.etal.Onpixel-wiseexplanationsfornon-linearclassifierdecisions bylayer-wiserelevancepropagation.PloSOne10,e0130140(2015). 52.Martin,Z.K.,Barnes,E.A.&Maloney,E.UsingSimple,Explainable NeuralNetworkstoPredicttheMadden-JulianOscillation.J.Adv.Model.Earth Syst.14,e2021MS002774(2022). 53.Toms,B.A.,Barnes,E.A.&Ebert-Uphoff,I.PhysicallyInterpretable NeuralNetworksfortheGeosciences:ApplicationstoEarthSystemVariability.J. Adv.Model.EarthSyst.12,e2019MS002002(2020). 54.Barnes,E.A.etal.IndicatorPatternsofForcedChangeLearnedbyan ArtificialNeuralNetwork.J.Adv.Model.EarthSyst.12,e2020MS002195(2020). 55.Ali,A.etal.XAIfortransformers:Betterexplanationsthroughconservative propagation.inInternationalconferenceonmachinelearning435–451(PMLR, 2022). 56.Sundararajan,M.,Taly,A.&Yan,Q.Axiomaticattributionfordeep networks.inInternationalconferenceonmachinelearning3319–3328(PMLR, 2017). 57.Kapishnikov,A.etal.Guidedintegratedgradients:Anadaptivepathmethod forremovingnoise.in2021IEEE/CVFconferenceoncomputervisionandpattern recognition(CVPR)5048–5056(IEEE,2021). InterpretableAIpredictsa2026summerdryanomalyin centralChina AnranWANG 1† ,WenSHI 1† ,YongLUO 1 ,JianbinHUANG 2,3,4 ,LijuanCHEN 5,6 ,Junhu ZHAO 6 ,WeixinJIN 7 ,andHuihuiYUAN 1 1DepartmentofEarthSystemScience,TsinghuaUniversity,Beijing100084 2CollegeofResourcesandEnvironment,UniversityofChineseAcademyofSciences, Beijing101408 3BeijingYanshanEarthCriticalZoneNationalResearchStation,UniversityofChinese AcademyofSciences,Beijing101408 4CollegeofResourcesandEnvironment,UniversityofChineseAcademyofSciences, Beijing100049 5StateKeyLaboratoryofClimateSystemPredictionandRiskManagement,National ClimateCentre,ChinaMeteorologicalAdministration,Beijing100081 6ChinaMeteorologicalAdministrationKeyLaboratoryforClimatePredictionStudies, NationalClimateCenter,ChinaMeteorologicalAdministration,Beijing100081 7Microsoft,Beijing100080 Contentsofthisfile SupplementaryTables1to2 SupplementaryFigs.1to4 † Theseauthorscontributedequally:AnranWangandWenShi. Correspondenceto:YongLuo(yongluo@tsinghua.edu.cn) SupplementaryTable1.Historicalanalogueyearsselectedforthesummer2026 predictionsinitializedinMarch,AprilandMay.Foreachinitialization,33 summersduring1993–2025wererankedbyall-stationACC(ACC all ),centralChina ACC(ACC C ),drycoverage(DC)anddry-intensitydistance(DID).Thesevenyears withthesmallestsumofmetricranks(S y )wereretained.DCisreportedasa percentageandDIDinpercentagepoints.Sixyearswerecommontoallthree selections;AprilandMayyieldedidenticalsets,whereasMarchreplaced2002with 2022. InitializationYearACC all ACC C DC(%)DIDS y Final rank March 19970.470.86931.541 20060.380.578110.392 20010.300.597916.0123 20150.180.517719.4214 20220.150.427415.5255 20110.200.377519.0256 19940.110.547719.7267 April 19970.510.88936.451 20060.350.49815.4122 20010.380.587911.1123 20150.220.597714.5174 19940.180.507714.8235 20020.270.506918.4246 20110.090.357514.1307 May 19970.570.87933.241 20060.380.558115.092 20010.340.607920.7123 20150.230.547724.1204 19940.160.547724.4245 20020.260.456928.0266 20110.160.407523.7277 SupplementaryFig.1.CirculationpatternsassociatedwiththeMarch-initialized prediction.SameasFig.3,butforthesevenhistoricalanalogueyearsselectedforthe MarchinitializationandtheMarch-initializeddynamicalMMEpredictionforJJA 2026.Theanaloguecompositeand2026predictionbothshowawesternNorth Pacific–SouthChinaSea–SouthChinacycloniccirculationanomalyandanomalous moisture-fluxdivergenceovercentralChina. SupplementaryFig.2.CirculationpatternsassociatedwiththeApril-initialized prediction.SameasSupplementaryFig.1,butforthesevenhistoricalanalogueyears selectedfortheAprilinitializationandtheApril-initializeddynamicalMME predictionforJJA2026. SupplementaryFig.3.IntegratedGradientsattributionfortheMay-initialized prediction.SameasFig.5,butshowingdry-supportingnegativeattributionsfrom IntegratedGradients(IG).IGbroadlyreproducesthedry-supportingmeridional-wind signalsat500,700and850hPaidentifiedbyLRP,whileassigningthelargest predictorcontributiontoT2m. SupplementaryFig.4.GuidedIntegratedGradientsattributionforthe May-initializedprediction.SameasSupplementaryFig.3,butusingGuided IntegratedGradients(GuidedIG).Thedry-supportingmeridional-windsignalsat500, 700and850hPaareagainevident,whileT2mhasthelargestpredictorcontribution. SupplementaryTable2.Cross-methodagreementindry-supportingattributions. Spearmanrankcorrelationassessesagreementintherankingofarea-weighted dry-supportingcontributionsacrossthe13predictors.Area-weightedfull-tensor cosinesimilaritymeasuresagreementintheunsmoothedattributionpatternsacrossall predictorsandgridcells.Forbothmetrics,valuescloserto1indicatestronger agreement.Thehighrankcorrelations(0.88–0.98)indicatethattheoverallpredictor hierarchywasconsistentacrossmethods.Full-tensorsimilaritywaslowerbetween LRPandthetwogradient-basedmethods,indicatinggreatermethoddependencein thedetailedspatialpatterns.AgreementwithLRPwasstrongestfortheMay initializationexaminedinthemaintext.LRP,layer-wiserelevancepropagation;IG, IntegratedGradients;GuidedIG,GuidedIntegratedGradients. InitializationMethodpair Spearmanrank correlation Area-weighted full-tensor cosinesimilarity March LRPvsIG0.910.47 LRPvsGuidedIG0.880.40 IGvsGuidedIG0.950.92 April LRPvsIG0.920.48 LRPvsGuidedIG0.900.43 IGvsGuidedIG0.960.92 May LRPvsIG0.950.64 LRPvsGuidedIG0.980.56 IGvsGuidedIG0.970.90