FusionofNeutronsandNeuralNets:Energizingenergyfrontiers.

WeiHUANG

Director,DivisionofPlanningInformationandKnowledgeManagement

DepartmentofNuclearEnergy,IAEA

29-04-2025

GlobalAIlandscape

Rapidexpansioninallsectorsdrivenbyadvancementsinmachinelearning,bigdata,andcomputingpower.

GrowinginvestmentfromgovernmentandtechgiantstoraceAIharness.

RegulatoryeffortstoensureethicalandsafeAIuse.

Enormouspotentialtoacceleratenucleartechnological

development,fromsciencetoenergytomedicineandothers

NuclearPowerOpportunitiesfromAIexpansion..

HarnessAIfornucleartoaddresspressingglobalchallenges

AIinHumanhealthAIinFoodandAgricultureAIinWaterandtheenvironment

AIinNuclearFusionAIinNuclearSecurityAIinSafeguards

AIinNuclearPower

NuclearPowerempoweredbyAI

?Design:supportingadvancednuclearreactordesignandanalysiswithAIandML,e.g.moreaccuratemodelsinfewerchanges,improved

characterizationofsafetymarginandefficientandfasteranalysis.

?Construction:Identifyingpotentialrisks(scheduledelays,costoverrunsandsupplychaindisruptions)forenhancedreal-timedecision-making

?Operations:Detectinganomalies,forecastingpotentialequipmentfailures,reducinghumanerrorandimprovingreliabilityofoperation.

?Maintenance:EnablingpredictivemaintenanceandreducingunplanneddowntimewithAIpoweredpredictionandprognostics.

?Hazardousenvironment:Reducinghumanexposuretohazardouswithdrones,robots,andAI-drivenmachinesandautomation.

?Knowledgeandcompetency:preservingcriticalknowledgeforsustainableoperationandmaintainingHRcompetency.

SpecificapplicationsinNPPoperation

In-ServiceInspections:AIautomatesanalysisofnon-destructivetestingdata(e.g.,ultrasonic,radiographicinspections).

ControlRodDriveMechanism(CRDM)Monitoring:MLdetectsanomaliesincoilcurrents(~96%accuracy),preventingroddrops.

AnomalyDetection:AIidentifiesearlywarningsignsofequipmentfailures

beforehumanoperatorsnoticethem.

WorkProcessAutomation:Droneswithcomputervisionreducetheneedformanualplantinspections.

Human-ComputerInteraction:Speech/gesturerecognitionincontrolroomsimprovesoperatorefficiency.

NaturalLanguageProcessing(NLP):Automatesreportanalysisanddecision-makingfromeventlogs.

FuelManagement&CoreDesign:AIoptimizesfuelloadingpatternsandreactorcontrolstrategies.

lAIconcernsanduncertainties

Regulatory&SafetyConcerns:LackofstandardsforAIinsafety-criticalnuclearapplications.

Explainability&Trust:AIdecisionsmustbeinterpretableforregulatorsandoperators.

CybersecurityRisks:AIsystemsarevulnerabletoadversarialattacksanddatamanipulation.

DataLimitations:Needforhigh-quality,labelledtrainingdata(oftenrestrictedduetonuclearsecrecy).

Toaddresschallenge,mitigateriskandreleasepotentialAIfor

sustainablenuclearenergydevelopment,Internationalcooperationis

akey.

IAEA’sRoleinAIapplication

DataSharing&Benchmarking:

CreatesyntheticdatasetsforAItrainingwhileensuringprivacy.

Facilitateanonymizeddataexchangeamongmemberstates.

Collaboration&Standardization:

FormAIworkinggroupsandnetworktoalignAIapplicationswithnuclearsafetyandindustrystandards.

StrengthentieswithISO/IECforAIregulatoryguidelines.

Researchandcapacitybuilding:

OrganizeTM,CM,WS,WebinaronAIreliabilityandexplainability.

ConductCoordinatedResearchProjects(CRPs).

AIfocusedactivitiesofIAEA

?TMonAIfornucleartechnologyapplication(2021)

?WebinaronAIforNuclearEnergy(2022)

?CMontheuseofanAIinassessingtheuncertaintiesinsevereaccidentmodelingandsimulationunderoneIAEACoordinatedResearchProject(2023).

?Sideevent(GC67)ArtificialInteligenceandtheIAEAhighlightedtheIAEA'scrucialroleinpromotingtheunderstandingofAIinnuclearscience(2023).

?IAEACollaboratingCentregrantedtotheCenterforScienceofInformation,PurdueUniversity,USAonAI:modellingandsimulation,validationofAIconceptsforNPPs,educationandtraining.

?TMonAIforAdvancingReactorDesignDevelopment(2024)andTMDeploymentofArtificialIntelligenceApplicationsfortheNuclearPowerIndustry:ConsiderationsandGuidance(2024)

?OnejointIAEA-ICTPWorkshoponAIandMLinAdvancingNuclearEngineeringandTechnology:fromTheorytoPractice(2025)

IAIWorkingGroup

NADIAproject

4.85million

knowledgeproducts

NADIA-NuclearArtificialintelligenceforDocumentIndexingandAnalysis

NADIA-Advantages

Consistentandreliable

tagging–90%iswelldone

Mechanismstodetect

inaccuratetaggingand

routethosetohumans

(~10%)

10am–12pm

Automationfreesupexpertstoperformothertasks

UpdatestoINISThesaurus

arepossiblewithoutmodel

re-training

m

RunsonCPUs

NuclearforAIpowerhunger

Withtheelectricity

consumptionofdata,centres,artificialintelligence(AI)andcryptocurrenciessettogrowinthecomingyears,major

techcompaniesareactivelylookingt