-- tutorial on machine learning and hand-on coding training
TRANSCRIPT
2019CTPLWorkshoponMachineLearning
1
GateofIndustry4.0,EmbracingArtificialIntelligence
--TutorialonMachineLearningandHand-onCodingTraining
TheCTPLgroup(ComputationalTransportPhenomenaLaboratory,KAUST)
is pleased to announce a workshop themed “Gate of Industry 4.0, EmbracingArtificialIntelligence--TutorialonMachineLearningandHand-onCodingTraining”togatherourguestsfromXi’anJiaotongUniversity,VirginiaPolytechnicInstituteand State University, King Fahd University of Petroleum and Minerals, ChinaUniversityofPetroleum,TaifUniversityandXiamenUniversity,togetherwithourCTPLmemberstostepintotheareaofmachinelearning,especiallydeeplearning.A distinguished young scholar, Yu Li, from Computational Bioscience ResearchCentre(CBRC)willgiveusatutorialtalkonthebasicstructureandprocedureaswellas instructionsonthecodingusingTensorflowandMatlab.Tao,YitengandJingfa from CTPL will discuss on our current application of deep learning onoptimizingengineeringcomputations.
Location:Room4214,Building1,KAUST,SaudiArabiaTime:12:00pm–16:00pm,4th,March,2019WorkshopChair:ShuyuSunWorkshopSecretary:TaoZhang
2019CTPLWorkshoponMachineLearning
2
Agenda
Time Speaker Topic
12:00pm—12:30pm Grouplunchandfree
discussion12:30pm—12:40pm ShuyuSun Openingandwelcome
12:40pm—13:10pm YuLiIntroductionontheMachineLearningandDeepLearning
13:10pm—13:40pm TaoZhangApplicationofDeepLearning
onacceleratingflashcalculations
13:40pm—14:10pm YitengLiFurtherapplicationofdeeplearningonmorecomplexphaseequilibriumproblems
14:10pm—14:30pm Coffeebreakandfree
discussion
14:30pm—15:00pm JingfaLi
Areviewonmachinelearningapproachforefficientuncertaintyquantificationusingmultiscalemethods
15:00pm—15:15pm HaiyiWuPredictingEffectiveDiffusioninPorousMediafromTheirimagesbyDeepLearning
15:15pm—16:00pm YuLi Hand-ontrainingoncoding
2019CTPLWorkshoponMachineLearning
3
*AllpresentationswillincludeashortQ&Aattheend.**Lunchwillbeservedat12:00pm,andjuice,coffeeandwaterwillbeprovidedallthetime.***Ashortintroductiontoeachspeakerandabstractofeachtalkareattachedinthefollowingpages.****ThemapofKAUSTInnandDiscoverySquare,aswellasofCampus,areillustratedonthelastpagetoguideyoutotheworkshopmeetingroomandtouraroundouruniversity.*****ItishighlysuggestedthattheattendeescouldsetupTensorflowandMatlabinadvance.******ItisrecommendedthattheattendeescouldfinishtheMatlabOnrampCourseathttps://www.mathworks.com/learn/tutorials/matlab-onramp.html*******ThetimeisshowninSaudiArabiatimezone(GMT+3)
2019CTPLWorkshoponMachineLearning
4
GuestAttendees
Name Affiliation
Prof.RuiQiaoVirginiaPolytechnicInstituteandStateUniversity,theUSA
Prof.BoYuBeijingInstituteof
PetrochemicalTechnology,China
Prof.JieChenXi’anJiaotongUniversity,
China
Prof.LiangGongChinaUniversityof
Petroleum,Qingdao,ChinaProf.HuangxinChen XiamenUniversity,China
Prof.ManalAlotibi TaifUniversity,SaudiArabia
Prof.XianbinLuo GuizhouUniversity,China
Dr.GangLeiKingFahdUniversityofPetroleumandMinerals,
SaudiArabia
2019CTPLWorkshoponMachineLearning
5
IntroductionontheMachineLearningandDeepLearning
Name:YuLiAffiliation: King Abdullah University of Science andTechnology,KAUST,SaudiArabiaBiography:Mr.LiisnowaPhDstudentatKAUST,majoringinComputer Science. He is a member of Structural andFunctionalBioinformatics(SFB)Group,whichisledbyDr.XinGao.HegotMasterdegreeinComputerScienceatKAUSTinDecember 2016. Before that, he got Bachelor degree inBiosciencesatUniversityofScienceandTechnologyofChina(USTC)in2015.HisresearchinterestsareDeepLearning,BioinformaticsandMachineLearning.Abstract:Machine learningalgorithms,especiallydeep learning,haveachievegreatsuccesses inrecentyearsacrossdifferentfields.Inthistalk,Iwillgiveabriefintroductiononmachinelearninganddeep learning. Because of the impressive potential of artificial intelligence, the relatedterminologies,suchas‘artificialintelligence(AI)’,‘machinelearning(ML)’,‘deeplearning(DL)’,have been abused in media, which can inevitably cause incorrect usages and people’smisunderstanding,whoarenotinthefield.So,Iwillstartfromclarifyingthebasicconceptsof‘AI’, ‘ML’ and ‘DL’. After that, I will introduce themain tasks ofmachine learning, includingclassification, regression, clustering and dimensionality reduction. Then, I will give a briefintroduction to neural networks, touching some interesting models, including convolutionalneuralnetworks(CNN),recurrentneuralnetworks(RNN),graphconvolutionalneuralnetworks(GCN)andgenerativemodels,suchasgenerativeadversarialnetworks(GAN).
2019CTPLWorkshoponMachineLearning
6
ApplicationofDeepLearningonacceleratingflashcalculations
Name:TaoZhangAffiliation:KingAbdullahUniversityofScienceandTechnology,KAUST,SaudiArabiaBiography:Mr.ZhangisnowaPhDstudentinthedepartmentof Earth Science and Engineering, joining our group sinceAugust 2016. In 2013, he got his Bachelor degree in ChinaUniversity of Petroleum, Beijing with the major Oil & GasStorage and Transportation. In June 2016, he obtained hismaster degree in the same major and university. Currently his research areaincludesmultiphaseflowsimulationindifferentcontinuumscales.Abstract:In the past two decades, researchers have made remarkable progress in accelerating flashcalculation,whichisveryusefulinavarietyofengineeringprocesses.Inthistalk,generalphasesplittingproblemstatementsandflashcalculationproceduresusingtheSuccessiveSubstitutionMethodare reviewed,while themain shortagesarepointedout. Twoaccelerationmethods,Newton'smethodandtheSparseGridsMethodarepresentedafterwardsasacomparisonwiththedeep learningmodelproposed in thispresentation.Adetailed introduction fromartificialneural networks to deep learningmethods is provided herewith the authors' own remarks.Factors inthedeep learningmodelare investigatedtoshowtheireffectonthefinalresult.Aselectedmodelbasedonthathasbeenusedinaflashcalculationpredictorwithcomparisonwithothermethodsmentioned above. It is shown that results from the optimized deep learningmodelmeet the experimental datawell with the shortest CPU time.More comparisonwithexperimentaldatahasbeenconductedtoshowtherobustnessofourmodel.
2019CTPLWorkshoponMachineLearning
7
Isothermal-isochoric phase equilibrium calculation inconventionalandunconventionalreservoirsbyadynamicmodel
Name:YitengLiAffiliation:KingAbdullahUniversityofScienceandTechnology,KAUST,SaudiArabiaBiography:Mr.Li isnowaPhDstudent inthedepartmentofEarthScienceandEngineering,joiningourgroupsinceJanuary2016.In2012,hegothisBachelordegreeinOceanUniversityofChina,QingdaowiththemajorMaterialChemistry.InMay2014,heobtainedhismasterdegreeinpetroleumengineeringinUniversityofSouthernCalifornia,LosAngeles.Currentlyhisresearch area includes phase equilibrium calculation and multiphase flow insubsurfacereservoirs.Abstract: Phase equilibrium calculation has various applications in petroleumengineering,notonlyasastandalonecalculationforseparationprocessbutalsoanintegral component of the compositional reservoir simulation. Previously, amajority of studies focus on the safety rather than the efficiency. It has beenreportedthattheequation-of-statebasedflashcalculationconsumesanenormousamountofcomputationaltime,upto70%,incompositionalsimulation,makingitthebottleneckfortheextensiveapplicationofcompositionalsimulators.Asaresult,the acceleration of flash calculations without much compromise in accuracybecomes an active research topic in the last two decades. In today’s topic, thephaseequilibriaproblemismodeledbasedontheNVTflashformulation.BasedonthelawsofthermodynamicsandOnsager’sreciprocalprinciple,moleandvolumeevolutionaryequationsareconstructedtodescribethedynamicprocessfromanynon-equilibriumstatetotheequilibriumstate.Benefitingfromitsgreatcapacity,weincorporatestabilitytestandphasesplittingcalculationtogether,bothofwhichconstitutethephaseequilibriumcalculation,andaccomplishthembyasingledeeplearningmodel. Itdiffers fromtheconventional flash frameworkwherestabilitytesting precedes phase splitting calculation, as most of the researches in flashspeedupbymachinelearningfollow.Anumberofsimulationresultsarepresentedtoshowtheaccuracyandefficiencyoftheproposeddeepneuralnetworkmodel.
2019CTPLWorkshoponMachineLearning
8
Areviewonmachinelearningapproachforefficientuncertaintyquantificationusingmultiscalemethods
Name:JingfaLiAffiliation:KingAbdullahUniversityofScienceandTechnology,KAUST,SaudiArabiaBiography:JingfaLicurrentlyworksasapostdoctoralfellowatComputational Transport Phenomena Laboratory (CTPL) atKAUST. He received the Bachelor degree from SouthwestPetroleumUniversityin2012andthePhDdegreefromChinaUniversity of Petroleum (Beijing) in 2017 both in PetroleumStorage& Transportation Engineering.His research interestsfocusonFVM,modelreductionmethod,anduncertaintyqualification.Abstract:Inthistalk,Iwillreviewamachinelearningapproachfortheestimationofcoarsescalebasisfunctionsinmultiscalefinitevolumemethod.Inthisapproach,a neural network predictor fitted using a set of solution samples fromwhich itlearnstogeneratesubsequentbasisfunctionsatalowercomputationalcostthansolving the local problems is developed. The computational advantage of thisapproachisrealizedforuncertaintyquantificationtaskswherealargenumberofrealizationshas tobeevaluated.Theproposedapproach isevaluatedonellipticproblemsyieldingverypromisingresults.
2019CTPLWorkshoponMachineLearning
9
PredictingEffectiveDiffusioninPorousMediafromTheir
imagesbyDeepLearning
Name:HaiyiWuAffiliation:VirginiaPolytechnicInstituteandstateUniversity,theUSABiography:Mr.Wu is now a PhD student at Virginia Tech,majoring in Mechanical Engineering. He is a member inLaboratory of Transport Phenomena for AdvancedTechnology,whichis ledbyProf.RuiQiao.HegotBachelordegree inModernMechanics at University of Science andTechnologyofChinain2015.Hisresearchinterestsaremultiscaleandmultiphasetransportphenomenainporousmedia.Abstract:Thisworkaimstotesttheapplicationofmachinelearningforpredictingthe effective diffusion of a porous medium from its geometry and to improvemachine learningperformanceby combing itwith field knowledge. The generalframework includes two parts. In the first part, a dataset for the diffusioncoefficient inporousmediaisgeneratedusingtheLatticeBoltzmannmethod.Inthe second part, machine learning method is used to process the datasetsgeneratedinthefirststepandtopredictthediffusioncoefficientofgivenporousstructure.Theconvolutionalneuralnetworks (CNN) isusedto train thedataset.ComparisonwiththegroundtruthsindicatestheexcellentpredictiveperformancebyCNNinawidevarietyofporousstructureswithacomputationalcostofseveralorders of magnitude smaller. The CNN results improves the prediction byconventionalBruggemanequationbyalmost2ordersofmagnitude,especiallyforthosesampleswithsmallporositiesandstrongheterogeneity.
2019CTPLWorkshoponMachineLearning
10
KaustInn&DiscoverySquare
Campus