Data-Driven Smart Proxy for Computational Fluid Dynamics - PowerPoint Presentation

1. Data-Driven Smart Proxy for Computational Fluid DynamicsShahab D. Mohaghegh1,2, Mehrdad Shahnam3. Ayodeji Aboaba1, Yvon Martinez1, Chris Guenther3, Young Liu3, Anthony Morrow1, & Ashley Konya11West Virginia University - WVU2Intelligent Solutions, Inc. - ISI 3National Energy Technology Laboratory - NETL Laboratory for theEngineeringApplication ofDataScienceWVU-LEADS

2. OutlineSummaryObjectiveDefinitionsArtificial Intelligence & Machine Learning (AI&ML)Engineering Application of AI&MLSmart Proxy vs ROMResultsConclusion

3. SummaryA Smart Proxy was built for a CFD (FLUENT) model that includes 9.3 million cellsThe CFD Model computational footprint is: 24 hours on a 40 Core HPCThe Smart Proxy was built (trained, calibrated and validated) using 8 CFD runs.The Smart Proxy was validated using two new CFD Runs that were made after the completion of the project.Results of the Two Blind Validation runs are presented here.The Smart Proxy computational footprint is: few minutes on a laptop or PC workstation

4. Project ObjectiveSet up a ANSYS FLUENT model and validate with B-6 experimental combustion

5. Project ObjectiveProvide data for reduced order model developmentConduct any additional simulations of the B-6 experiment as required by the ROM developmentExtend the B-6 modeling efforts to an industrial scale boiler.

6. Definitions: Artificial Intelligence & Machine LearningThe Technology that Mimics Human Brain in Analysis, Modeling, and Decision Making Using Open Computer Algorithms to Learn from Data instead of Explicit Programming. Artificial IntelligenceMachine Learning

7. Definitions: Engineering Application of Artificial Intelligence & Machine LearningModeling Physics and Engineering related problem using Artificial Intelligence & Machine LearningIt requires:Domain ExpertiseExpertise in Artificial Intelligence & Machine Learning PracticesKnowledge of the Foundation of Artificial Intelligence & Machine Learning

8. Definitions: ROM vs. Smart ProxyReduced Order Models (ROM)Smart Proxy Modeling (SPM)ROM: Reduces computational footprint through reducing Physics & ResolutionSPM: Reduces computational footprint Without reducing Physics & ResolutionIt learns the behavior of the Numerical Model through experience (observation, data)It replicates the behavior of the Numerical Model with high accuracy, without reducing the physics or resolutionIt reduces the PhysicsIt reduces the Resolution in Time &/or Space

9. ResultsTwo ScenariosOne scenarios uses natural gas as the fuel (marked red).One scenario uses blended propane with natural gas as fuel (marked blue).Each scenario is repeated 4 times.The air is preheated and the fuel is at room temperature

10. Cross-SectionsFull3/41/41/2

11. Pressure

12. BASE Blind CFD run

13. BASE Blind CFD run

14. Cross-Section = 3/4BASE Blind CFD run

15. Cross-Section = 1/2BASE Blind CFD run

16. Cross-Section = 1/4BASE Blind CFD run

17. BLENDED Blind CFD run

18. BLENDED Blind CFD run

19. Cross-Section = 3/4BLENDED Blind CFD run

20. Cross-Section = 1/2BLENDED Blind CFD run

21. Cross-Section = 1/4BLENDED Blind CFD run

22. Temperature

23. BASE Blind CFD run

24. BASE Blind CFD run

25. Cross-Section = 3/4BASE Blind CFD run

26. Cross-Section = 1/2BASE Blind CFD run

27. Cross-Section = 1/4BASE Blind CFD run

28. BLENDED Blind CFD run

29. BLENDED Blind CFD run

30. Cross-Section = 3/4BLENDED Blind CFD run

31. Cross-Section = 1/2BLENDED Blind CFD run

32. Cross-Section = 1/4BLENDED Blind CFD run

33. Nitrogen

34. BASE Blind CFD run

35. BASE Blind CFD run

36. Cross-Section = 3/4BASE Blind CFD run

37. Cross-Section = 1/2BASE Blind CFD run

38. Cross-Section = 1/4BASE Blind CFD run

39. BLENDED Blind CFD run

40. BLENDED Blind CFD run

41. Cross-Section = 3/4BLENDED Blind CFD run

42. Cross-Section = 1/2BLENDED Blind CFD run

43. Cross-Section = 1/4BLENDED Blind CFD run

44. Oxygen

45. BASE Blind CFD run

46. BASE Blind CFD run

47. Cross-Section = 3/4BASE Blind CFD run

48. Cross-Section = 1/2BASE Blind CFD run

49. Cross-Section = 1/4BASE Blind CFD run

50. BLENDED Blind CFD run

51. BLENDED Blind CFD run

52. Cross-Section = 3/4BLENDED Blind CFD run

53. Cross-Section = 1/2BLENDED Blind CFD run

54. Cross-Section = 1/4BLENDED Blind CFD run

55. CO2

56. BASE Blind CFD run

57. BASE Blind CFD runBASE Blind CFD run

58. Cross-Section = 3/4BASE Blind CFD run

59. Cross-Section = 1/2BASE Blind CFD run

60. Cross-Section = 1/4BASE Blind CFD run

61. BLENDED Blind CFD run

62. BLENDED Blind CFD run

63. Cross-Section = 3/4BLENDED Blind CFD run

64. Cross-Section = 1/2BLENDED Blind CFD run

65. Cross-Section = 1/4BLENDED Blind CFD run

66. ConclusionsUnlike ROM, Smart Proxy Modeling does not Reduces the Physics or the Resolution in Time and Space in order to minimize the CFD Computational FootprintSmart Proxy is much faster and much more accurate than ROMSmart Proxy is a solid demonstration of the quality of the Engineering Application of Artificial Intelligence & Machine LearningQuestion: If you can build sensors in order to generate data from physical phenomenon that you are trying to model, Would You Still Need CFD?

67. Thank you

68. Impact of Domain Expertise in AI & MLImage Recognition

69. DOGCATImpact of Domain Expertise in AI & MLImage Recognition

70. Impact of Domain Expertise in AI & MLEngineering Application of Data ScienceSeparate the black and the yellow Circles by drawing a line.How much expertise do you need to do that?The line will be like a circle.It will be non-linear

71. Impact of Domain Expertise in AI & MLEngineering Application of Data ScienceSeparate the black and the yellow Circles by drawing a line.What if we require that: “THE LINE MUST BE LINEAR”Do you need any type of expertise to do this? (i.e. Mathematics)

72. Impact of Domain Expertise in AI & MLEngineering Application of Data Science 

73. Q1Q2H2INLETCOMBUSTOREXHAUSTQ1Q2H2Exhaust

74. Human BrainThe most powerful pattern recognition engine in the universe.74

75. How Does Human Brain Manages Learning?Human Brain is 10 million times slower than the computer chips in your smart phone.Yet, it has capabilities to learn and implement complex activities that is very hard for computers to mimic. This has to do with HOW human brain processes information.75

76. How do we Learn?Does “Human Brian” follows the engineering approach to problem solving?Using first principle physics to build a model of the physical phenomenon in the form of mathematical equations [second order, non-linear, partial differential equation - Diffusivity Equation].Using applied mathematics to solve the mathematical equations (analytical solutions [well testing], or numerical solutions [numerical reservoir simulation]).76

77. How do we Learn?Artificial Intelligence & Machine Learning is when “Computers” are used to mimic the Human Brian to solve problem:Strengthening a specific set and pattern of synaptic connections in the part of the brain that is responsible for the specific action. Observe,Collect data, Trial and Error, Discover Pattern,Practice and get better and better.77

78. Complex Example – Range of TrajectoryThrowing an object to be caught by another person.78Engineering ApproachHuman Brain ApproachqV0gRange 

79. Complex Example – Inverted PendulumBalancing a stick on the palm of your hand, or on your fingers.79Engineering ApproachHuman Brain ApproachPivotqyL

80. Correlation & CausationOne of the most contentious issues that are usually brought up by engineers and scientist when presented by statistical analyses, is the relationship between “correlation” and “causation” and the fact that they are not necessarily the same. Are the engineers and scientists correct?

81. Correlation vs. Causation

82. Correlation vs. Causation

83. Correlation vs. Causation

84. Correlation & CausationWhen it comes to data-driven reservoir modeling, it is not enough to show correlations. Unlike many other problems and industries that have been the environment where data analytics has thrived using mainly statistical approaches (such as social media and Consumer Relation Management CRM), overwhelming number of problems related to oil and gas industry are very much physics-based.

85. Correlation & CausationThis is how data-driven reservoir modeling is distinguished from its competitor empirical technologies that rely purely on statistics and mathematics without regards for physics and geology. Data-driven modeling has been referred to as “black box” technology that usually stems from a lack of understanding (or may be a superficial understanding) of the machine learning technology.

86. Correlation & CausationIf the term “black box” is used to emphasize the fact that there is not a deterministic mathematical formulation that can fully and comprehensively explain the behavior of a data-driven model, then this term is correct. If it means that the functionality of the model cannot be understood or verified (which many times is the reason why the critics use this terminology), then it is a miss use of the term. All interactions between different parameters in a trained data driven reservoir model can be fully investigated in order to make sure that it makes physical and geological sense. From this point of view there are nothing in a data-driven reservoir model that makes it a “black box”.