spe-227546 PRISM：用于快速地震解释的预训练基础模型.pdfVIP

SPE-227546-MS

PRISM:APre-TrainedFoundationModelforRapidSeismicInterpretation

H.DiandA.Sharma,SLB

Copyright2025,SocietyofPetroleumEngineersDOI10.2118/227546-MS

ThispaperwaspreparedforpresentationattheMiddleEastOil,GasandGeosciencesShow(MEOSGEO),Manama,Bahrain,16–18September2025.

ThispaperwasselectedforpresentationbyanSPEprogramcommitteefollowingreviewofinformationcontainedinanabstractsubmittedbytheauthor(s).Contents

ofthepaperhavenotbeenreviewedbytheSocietyofPetroleumEngineersandaresubjecttocorrectionbytheauthor(s).Thematerialdoesnotnecessarilyreflect

anypositionoftheSocietyofPetroleumEngineers,itsofficers,ormembers.Electronicreproduction,distribution,orstorageofanypartofthispaperwithoutthewritten

consentoftheSocietyofPetroleumEngineersisprohibited.Permissiontoreproduceinprintisrestrictedtoanabstractofnotmorethan300words;illustrationsmay

notbecopied.TheabstractmustcontainconspicuousacknowledgmentofSPEcopyright.

Abstract

Whilerobustseismicinterpretationandmodelingisessentialinsuccessfulsubsurfacemappingand

characterization,itremainschallengingduetoitsstrongdependencyondataquality,domainknowledge,and

expertsupervision.Inrealscenariosofnoisyseismic,correspondingly,manualeffortscouldbeintensive,

andthussignificantprogressisanticipatedinautomationandefficiencyimprovement.Therecentsuccess

ofdeeplearning(DL)innaturalimageprocessingandpatternrecognitionhasledtothedevelopment

ofmanymore(semi-)automatedsolutionstoresolvevariousseismicinterpretationchallenges,suchas

saltbodydetection,faultpickingandhorizontracking.However,eachofthemisdefinedasastandalone

processforcompletingoneorafewspecifictasksonlyandrequiresrepeatingmodeltrainingincaseof

anychangesindataand/orlab