PowerPoint 프레젠테이션

Size: px

Start display at page:

Download "PowerPoint 프레젠테이션"

본승 이
5 years ago
Views:

Deep learning for topology optimization design - 연구재단신진연구과제 (2018.3~2020.2) -KISTI 연구지원사업 (2018.1.1~6.

2 Deep learning for topology optimization design - 연구재단신진연구과제 (2018.3~2020.2) -KISTI 연구지원사업 ( ~6.30) Deep learning based bone microstructure reconstruction - 원자력연구원기관고유사업 ( ~9.20) Hall sensor 신호를이용한제어봉구동장치위치지시기개발 딥러닝 2 출처 : 네이버웹툰 ' 호랭총각 '

3 종의최적설계 (Acoustical Damping & Natural Frequency, 2011) 바이올린상판의최적설계 (Natural Frequency & Nodal Line,2010) 바이올린브릿지의위상최적설계 (Natural Frequency & 진동전달효율,2013) 3

4 4

5 LipNet ( 5

6 6

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 뇌의동작원리를꼭알아야할까요? 17

18 물질의양자상태 ( 스핀등 ) 가어떻게되는지시물레이션하는것은물질구성에좀더깊은이해를줌. 기본적으로물질은 Many-body System 인데, 이들사이의상호작용으로나타는물질의상태를시물레이션하는것은시간과컴퓨팅파워가무척이나많이필요함. Deep learning 을이용하여물질의양자상태에대한시물레이션을진행. 18

19 19

y=f (x) CNN Prediction LBM https://www.biorxiv.

20 y=f (x) CNN Prediction LBM 20

21 y=f(x)+f (x) 21

22 Heng Xiao and Jinlong Wu, Towards A Physics-Informed Machine Learning Framework for Predictive Turbulence Modeling 22

23 y=g(f (x)) 23

24 Archives of Computational Methods in Engineering (I.F > 5.0!!!) Special Issue : Machine Learning in Computational Mechanics 24

25 25 rning+in+computational+mechanics%22

27 딥러닝은엔지니어링시뮬레이션및설계혁명을일으킬것이다. - GE 리서치수석정보과학자마크에드가 (Marc Edgar) 2018 년은엔지니어링시뮬레이션및설계의혁신을시작하는한해가될것이다. 향후 3 ~ 5 년동안딥러닝은제품의기능, 성능및비용면에서혁신적인패러다임을창출하기위해수년에서수개월걸리던제품개발을수주또는수일만에가속화할것이다. 27

28 28

29 가능한적은수의변수로구조를표현할수있는가? 29

30 30

31 ? predicted? Inference 31

32 32

33 y=g(x) y=g (x) Conventional Modeling Differential equation Numerical simulation Slow, large memory Difficult non-linear modeling Difficult to optimize Data-driven modeling Functions trained with data Training time required Faster, small memory Non-linear modeling Easy Optimization 박문규, Simulation Environment, Big Data and Ai in Nuclear Enginee 33

34 Optimization Scheduler 기존최적설계 머신러닝알고리즘 34

35 35

37 골다공증성골절 골다공증 37

38 정확한골다공증진단을위한저선량 CT 사진고해상화 78 m resolution 78 m resolution 625 m resolution 38

39 39

40 40 arxiv:

41 - 골재형성과정은최소의골량으로주어진기계적자극에대해최대의기계적효율을얻는골미세구조를생성함 (Wolff s law, 1892) 기계적자극골재형성과정골미세구조 41

42 Encoder Network Decoder Network latent vector 딥러닝기반위상최적설계 융합 딥러닝기반영상고해상화 42

44 44 사람

45 Anomaly Detection Non destructive Test Health monitoring Pump LPMS, Acoustic alarm Uncertainty Evaluation Digital Twin Automation Normal condition, Emergency condition Structural Optimization Materials Science 45

46 46

47 47 김재용, AI 를이용한 HR 특성예측

48 * 전력경제신문 48

49 * 김길유, 증기발생기 U-Tube ECT 전문가시스템, Physics Informed Machine Learning 포럼, 2017, 충남대 49

50 50

51 초대형쓰나미에대한무방비 설계기준쓰나미설정 + 설계기준초과쓰나미대책 중대사고대응대책미흡 1980 년대이후잘알려진 Mark-I 격납용기의취약성보완미흡 중대사고대응대책 ( 설비, 절차서, 교육훈련등 ) 부족 지진과쓰나미에의해악화된작업환경 복구설비이동에제약 끊임없는여진문제 사고진행과정에서의부적절한대응 1 호기비상응축기작동상태오인, 3 호기고압주입계통수동중단, 격납용기배기밸브개방지연, 보고체계혼선등 원전내부상태에대한정보부족 원자로내부상태에대한부정확한이해 / 추정 중대사고가다수호기에서동시에전개 * 백원필, 원자력이용현황, 후쿠시마사고및지속이용을위한도전과제, 부산대학교세미나 51

52 52

53 Decision1 Decision2 Decision3 Decision4 사고발생 Safe or Not? 53

54 Simulation (Surrogate, Digital Twin) model Control & Monitoring Real Environment 복잡한다물리현상을빠르게모사할수있는가? 어떤데이터를생성할것인가? 실제데이터와차이는? 54

dimensionality reduction to identify the effective DOF.

55 The final goal of this study is to construct a surrogate model for the coupled Rattlesnake-BISON models The computational cost needed for the construction of surrogate models for a multi-physics model can be significantly reduced if one employs dimensionality reduction to identify the effective DOF. Another important conclusion of this study is that while fine mesh simulation is highly needed to accurately describe the multi-physics nature of system behavior, it comes at a great cost. 55

56 Direct numerical simulation of combustion systems is impossible Resolution requirement Number of equations to be solved Ex) 53 species and 325 reactions 57 strongly coupled PDE PCA offers the potential to automate the selection of an optimal basis for representing the manifolds 56

57 자율운전 감시시스템 화재및운전원감시 핵물질유출감시 (CCTV, neutron detector) 배관감육진단 57

58 증명할수있는가? 불확실도는얼마인가? 데이터부족 라벨링의비용및난이도 정상데이터에편중 시물레이션으로데이터생성시 시물레이터해석시간 시물레이터와실제상황과의차이 58

https://www.biorxiv.org/content/biorxiv/early/2017/12/30/240317.full.

non-linear modeling Difficult to optimize Data-driven modeling Functions trained with data Training

60 Conventional Modeling Differential equation Numerical simulation Slow, large memory Difficult non-linear modeling Difficult to optimize Data-driven modeling Functions trained with data Training time required Faster, small memory Non-linear modeling Easy Optimization 60 박문규, Simulation Environment, Big Data and Ai in Nuclear Engineering

61 61

62 인공지능학계 - 도메인지식없이도모든문제를잘푸는인공지능을개발했다! 응용분야 - 응그래? 그럼가져다쓰면되겠네? - 알파고제로가지고와서적용하면뭔가잘되겠지. - 잘안되잖아! ( 예전처럼 ) 사기야! 가장중요한부분을고민하고있지않는것같은

Data Science Technology Spectrum Real-world systems often combine several techniques Reinforcement Learning Unsupervised Assimilative Models Supervised ML Exploratory Parameterization Naïve Stats

63 Data Science Technology Spectrum Real-world systems often combine several techniques Reinforcement Learning Unsupervised Assimilative Models Supervised ML Exploratory Parameterization Naïve Stats Physics-Based Modeling UQ Machine Learning Data-Driven Expert Knowledge in Data + Labels Model (mostly) determined by D + L Open Exploration Model-Based Understanding Model-Driven Expert Knowledge in Model Details Data refines model parameters * Lukas Mandrake, Machine Learning & Autonomy

64 64

65 From Drew Conway 65

66 66

67 67

김기남_ATDC2016_160620_[키노트].key

김기남_ATDC2016_160620_[키노트].key metatron Enterprise Big Data SKT Metatron/Big Data Big Data Big Data... metatron Ready to Enterprise Big Data Big Data Big Data Big Data?? Data Raw. CRM SCM MES TCO Data & Store & Processing Computational