科学人工智能报告

Australia’sNationalScienceAgency Artificialintelligenceforscience Adoptiontrendsandfuturedevelopmentpathways November2022 Citation HajkowiczS,NaughtinC,SandersonC,SchleigerE,KarimiS,BratanovaA,BednarzT(2022).Artificialintelligenceforscience–Adoptiontrendsandfuture developmentpathways.CSIROData61,Brisbane,Australia. Copyright ©CommonwealthScientificandIndustrialResearchOrganisation2022.Totheextentpermittedbylaw,allrightsarereserved,andnopartofthispublicationcoveredbycopyrightmaybereproducedorcopiedinanyformorbyanymeansexceptwiththewrittenpermissionofCSIRO. Importantdisclaimer CSIROadvisesthattheinformationcontainedinthispublicationcomprisesgeneralstatementsbasedonscientificresearch.Thereaderisadvisedandneedstobeawarethatsuchinformationmaybeincompleteorunabletobeusedinanyspecificsituation.Norelianceoractionsmustthereforebemadeonthatinformationwithoutseekingpriorexpertprofessional,scientificandtechnicaladvice.Totheextentpermittedby law,CSIRO(includingitsemployeesandconsultants)excludesallliabilitytoanypersonforanyconsequences,includingbutnotlimitedtoalllosses,damages,costs,expensesandanyothercompensation,arisingdirectlyorindirectlyfromusingthispublication(inpartorinwhole)andanyinformationormaterialcontainedinit. Acknowledgements TheauthorsofthisstudywouldliketothankthemanyartificialintelligenceexpertswithinCSIROandbeyondwhogenerouslysharedtheirtimeandknowledgethroughconsultativeinterviews.TheauthorswouldalsoliketothankstaffatTheLens(lens.org)whoprovidedsupport,adviceandguidanceonourbibliometricanalysis. UnitsandNotation Unlessotherwisespecified,dollarvaluesinthisreportaregiveninAustraliandollarsincurrentpricesandthisreportusestheInternationalSystemofUnits(SI). Accessibility CSIROiscommittedtoprovidingwebaccessiblecontentwhereverpossible.Ifyouarehavingdifficultieswithaccessingthisdocumentpleasecontactcsiro.au/contact. Coverimage:3DSLAM(SimultaneousLocalisationandMapping)mappingsoftwareWildcatcreatesadetailedscanofFishLaneinBrisbanetoactasamapforanautonomousrobot. Executivesummary Keyfindings 123 InthewetlandsofKakadu,rangersareusingAIandIndigenousKnowledgetocareforcountry. TogetherwiththeKakaduRangersandMicrosoft,wedevelopedtheHealthyCountryDashboard. Image:MichaelDouglas i Weareamidstthelargestsurgeinhistoryinthedevelopmentandadoptionofartificialintelligence(AI)forscientificresearchinalldisciplines ofnaturalscience,physicalscience,socialscienceandtheartsandhumanities.Thiswillimpacttheresearchsector,researchorganisationsandindividualresearchcareers. ThefutureimpactsofAIrangefromenablementtotransformation.Enablingmeansimprovingthespeed,safety,cost‑effectivenessandqualityofresearch.It’salso possiblethatAIcouldtransformknowledgediscovery,allowingscientiststosolveproblemshithertobeyondreach. WeidentifysixfuturedevelopmentpathwaysforresearchorganisationsseekingtoupgradeAIcapabilityforthefuturesotheycanharnessthebenefitsandmitigate therisksofAItechnologies.Weprovidedetailedguidanceoneachpathway. Whywewrotethisreport Thisreportaimstoinformresearchersandresearchorganisationswithinthespheresofgovernment,industry,communityandacademiaseekingtodevelopimprovedAIcapabilities.ThereportisfocusedontheuseofAIforscience,anditdescribesAIadoptiontrendsinthephysical,naturalandsocialsciencefields.Usingabibliometricanalysisofpeer‑reviewedpublishingtrendsover63years(1960–2022),thereportdemonstratesasurgeinAIadoptionacrossallfieldsoverthepastseveral years.Thereportexaminesfuturedevelopmentpathwaysandexploresimplicationsforscienceorganisations. Artificialintelligenceadoptiontrends ThereisarecentsurgeinpublicationsofAIadoptionbyscientists.BySeptember2022approximately 5.7%ofallpeer‑reviewedresearchworldwidewas onthetopicofAI.Thisisupfrom3.1%in2017and1.2%in2000.In2020,beforeresearchhadmigratedtoCOVID‑19topics,five(ofseven)ofthemostinfluentialresearchpapersonGoogleScholar(basedoncitationmetrics)wereonthetopicofAI[1]. AIisbeingusedinincreasinglydiverseresearchdomains.Inthefieldofcomputerscience,30%ofallpeer‑reviewedpapersarenowonthetopicofAI.However,theuseof AIextendswellbeyondcomputerscience.Practicallyall fieldsofnatural,physicalandsocialscienceandtheartsandhumanitiesarerapidlyabsorbingAItechnology. PublicandprivatesectorAIresearchanddevelopmentinvestmentisrising.Theaverageadvanced‑economyspendonresearchanddevelopment(R&D)hasrisenfrom2.1% ofGDPin2000to2.5%in2019[2].WithintheR&Dspend, resourcesdevotedtoAIareincreasing.Since2017over700AIpolicyandstrategyinitiativeshavebeendevelopedacross60countriesandterritorialjurisdictions[3,4]. Thereisevidenceofenhancedproductivitythroughcasestudies.InthisreportweidentifyandreviewnumerouscasestudieswhereAIhasimprovedtheefficiencyandeffectivenessofresearch.Forexample,aCSIROteamdevelopedanautomatedroboticsystemthattests 12,000solarcellsin24hours.Previouslytheresearchteamcouldmanuallytestupto20solarcellsperday. TheAIapproachimprovedproductivityby600times[5].Inanothercase,CSIROresearchersusedmachinelearningtofinddiseasegenes.Thisapproachprocesses10milliongenomicvariantsin15hourscomparedtotraditionalapproach