基于物理的機器學習和生成式 AI 用于科學和工業中的替代建模.pdf

《基于物理的機器學習和生成式 AI 用于科學和工業中的替代建模.pdf》由會員分享，可在線閱讀，更多相關《基于物理的機器學習和生成式 AI 用于科學和工業中的替代建模.pdf（24頁珍藏版）》請在三個皮匠報告上搜索。

1、PHYSICS-INFORMED MACHINE LEARNING AND GENERATIVE AI FOR SURROGATE MODELING IN SCIENCE AND INDUSTRYMIKE OKEEFFE,SENIOR SOLUTIONS ARCHITECT,NVIDIANEW YORK SCIENTIFIC DATA SUMMIT 2024:ADDRESSING DATA CHALLENGES IN DIGITAL TWINSSCIENTIFIC COMPUTING IS EVOLVINGThought Leaders Map Out the Opportunity and

2、Constraints Given Current Technology and Market RealityBusiness-as-usual will not be adequate TRANSITION TO POST EXASCALE ERAFEATURETERA THROUGH EXASCALEPOST EXASCALEUSAGEBATCH&MOSTLY LOCAL TO A SITE INTERACTIVE&DISTRIBUTED WITH MULTIPLE SITESWORKLOAD SINGLE SIMULATION/ENSEMBLEWORKFLOW COMPRISED OF

3、SIMULATION ENSEMBLES,AI TRAINING AND INFERENCE,LIVE DATA ANALYTICSEXPERIMENTSOFFLINE DATA ANALYSIS FOR EXPERIMENTSMIX OF REAL-TIME ANALYSIS TIGHTLY COUPLED WITH OFFLINEDIGITAL TWINSIN-SITU VISUALIZATION OFFLNEINTERACTIVE VISUAL MODEL COUPLED WITH PHYSICAL ASSETQUANTUM COMPUTINGSIMULATIONSIMULATION P

4、REPARING FOR A HYBRID MODEL PROGRAMMING MODELSFORTRAN,C+,MPI,OPENMPSTANDARD PARALLELISM SUPPORT IN FORTRAN,C+,MPI,OPENMP,OPENACC,PYTHON,JULIA,PYTORCH,TENSORFLOWSYSTEM CONFIGURATIONMONOLITHICMODULARCLOUDGRIDBURST CAPABILITIES,FASTER REFRESH CYCLE,ACCESS TO LATEST TECHNOLOGY AT SCALEDIGITAL TWIN FOR S

5、CIENCEA Relatively New Modeling Concept That is Being Adopted by the HPC Community in the Post Moore EraThe concept was first publicly introduced in 2002 by Michael Grieves,at a Society of Manufacturing Engineers conference as the conceptual model underlying Product Life Cycle ManagementA digital tw

6、in is a set of virtual information constructs that mimics the structure,context,and behavior of a natural,engineered,or social system(or system-of-systems),is dynamically updated with data from its physical twin,has a predictive capability,and informs decisions that realize value.The bidirectional i

7、nteraction between the virtual and the physical is central to the digital twin.CHAPTER:DIGITAL TWIN PAST,PRESENT AND FUTUREMichael GrievesExecutive Director Chief Scientist Digital Twin Institutehttps:/ AI ENABLES DIGITAL TWINS FOR SCIENCEQuantum Accuracy at Cost for Physical Scale ModelsErrorTight-

8、bindingDFTCCSD(T)Log computational cost Exact solution Classical AI ENABLES DIGITAL TWINS FOR SCIENCEQuantum Accuracy at Cost for Physical Scale ModelsErrorTight-bindingDFTCCSD(T)Log computational cost Exact solution AI SurrogateClassical AI INTRODUCES NEW USE CASES FOR SCIENCE AND ENGINEERINGAI Bri

9、dges the Gap Between Simulation and Real-TimeErrorTight-bindingDFTCCSD(T)Log computational cost Exact solution Classical AI Surrogates and Control logicInstrument DIGITAL TWIN POC SCIENCE EXAMPLESCollaborate with the Global Research Community to Pursue Science Discovery that Benefits MankindKubota F

10、or Earth For Life Earth 2Destination EarthTowards Real time Fusion Reactor Design Multi-Messenger Neutrino Detection Real Time Multi-Messenger Astrophysics Covid is Airborne Genome Scale LLMs for Covid Earthquake Model with Machine Learning Earthquake Early Warning SCEC Generative AI to Predict Disr

11、uption Earth-2 ProgramBuild the technology needed to create the digital twin of the earths weather and climate systemsAccelerating NWP codes on GPUOperationalize using Cloud ServicesAI research and collaboration with the science communityInteractive Visualization Digital TwinsAI ALGORITHMS EVOLVING

12、AT UNPRECEDENTED PACEREGIONAL FORECASTING VIA KM-SCALE SUPER-RESOLUTIONGenerative AI Diffusion models Case Study:Super-resolve 25-km AI weather models(SFNO,GraphCast,Pangu Weather)12.5x super-resolution&channel synthesis using CorrDiff over Taiwan(25km-2 km)Sample diversity from Gen-AI of equivalent

13、ly plausible fine-scale atmospheric conditions.1000 x Faster,3000 x more Energy Efficient,200 x Data Compression relative to WRF on CPU(Numerical Model)EARTH-2Connecting Complex Simulation,Data and AI WorkflowsRendererEarth-2 PlatformTrainingInferenceGPU accelerated SimulationInteractive Visualizati

14、on ModulusEarth Digital TwinOmniverse Data SourcesAI EnterpriseNVIDIA MODULUSOpen-Source Platform for Developing Physics-Based Machine Learning Training Neural Networks Using Both Data And The Governing EquationsWith generative AI using diffusion models,you can enhance engineering simulations and ge

15、nerate higher-fidelity data for scalable,responsive designs.Advancing Scientific Discovery With ModulusAneurysmClimate Change45,000X Faster extreme weather prediction with FourCastNetHealthcareHigh-fidelity results faster for blood flow in inter-cranial aneurysmDigital TwinsKinetic Vision:Design opt

16、imization using parameterized modelsRenewable EnergySiemens Gamesa:4000X Faster wind turbine wake optimizationIndustrial HPC NETL:10,000X Faster build of high-fidelity surrogate models Science and Engineering Teaching Kit available now.Generative AI making headway into Biology and Drug DiscoverySour

17、ce:arXiv.org Q-bio:AI,ML,DL,NNESMAlphaFoldCASP13AlphaFold2CASP14ESM2EquiFoldDiffDockGenSLMsProteinMPNNDNABERTAI Published PapersLab Automation:Sensors&RoboticsIn Silico Drug Discovery:AI&ComputingDRUG DISCOVERY IS AT AN INFLECTION POINTComputer Aided Drug Discovery is Expanding ExponentiallyCHARMMDF

18、T&Force FieldsX-Ray Protein StructuresLarge ScaleSimulationHigh Throughput ScreeningVirtualScreeningGenomicsMulti-SystemSimulationCryo-Electron MicroscopyAlphaFoldAI Structure PredictionMulti-Scale OmicsGENERATIVE AI DRY LABS ARE ACCELERATING DRUG DISCOVERYGenerate PredictAnalyzeDesignMakeTestAnalyz

19、eCandidate LikelihoodYears 250%Candidate LikelihoodMonths990%3 Years Faster|100s of Millions CheaperNVIDIA BioNeMoBuild,Optimize and Deploy Foundation Models for Computer-Aided Drug DiscoveryTrainOptimizeBiomolecularGenerationSequence AnnotationTraining&FinetuningBiomolecularProperty PredictionPertu

20、rbation PredictionStructurePredictionDeployFoundation ModelData LoadersOptimized TrainingTraining FrameworkMicroservicesESM-1|ESM-2Protein LLMsNEW:DiffDock|EquiDockDocking PredictionNEW:DNABERTDNA Sequence ModelCOMING SOON:MolMIMMolecular GenerationCOMING SOON:Single Cell BERTSingle Cell Expression

21、ModelBioNeMo Framework Supports Optimized Biomolecular ModelsProteins|Small Molecules|GenomicsNEW:OpenFold3D Protein Structure PredictionMegaMolBARTGenerative Chemistry ModelProtT5Protein Sequence GenerationBuild Generative AI Virtual Screening Workflows with NVIDIA NIMUse composable NVIDIA NIMS to

22、build workflows for CADD applicationsUniDockGradient-Free OptimizationUserOptimized MoleculesProtein DatabaseESM FoldDiffDockMoIMIMNVIDIA PLATFORM EVOLVING TO MEET THE CHALLENGEGPU CPUDPUNVLINKSuperPODHGXOVXEGXAGXDGXBase CommandRivermaxMagnum I/OHPC SDKCUDA-QCuQuantumModulusHoloscanOmniversePARTICLE ACCELERATORSTELESCOPESLIGHT SOURCESMICROSCOPESQUANTUM SIMULATIONVASPGEANT4QCDLAMMPSPINNBERTGNNDIGITAL TWINSHPC*AIVIRTUAL DESIGNAND CONTROLEXPERIMENTS/SENSORSQUANTUM COMPUTINGSIMULATIONJAXRAPIDSJETSONLLMMegatronNeMoRAPIDSModulusBioNeMoIsaacSimPhysXTHANK YOU