Machine learning in materials design, oil exploration, and beyond

Transcription

1 Machine learning in materials design, oil exploration, and beyond Gareth Conduit & Bartomeu Monserrat EP ; US 2014/177578; GB EP ; US 2014/223465; GB EP ; US 2014/224885; GB EP ; amendment to US 2013/ A1; GB Acta Materialia 61, 3378 (2013) Intermetallics 48, 62 (2014) Theory of Condensed Matter Group, Department of Physics

2 Stone age: 3.4 million BC 2000 BC 1.9 million BC Stone age

3 Bronze age: 2000 BC 1000 BC 1.9 million BC Stone age 1200 BC Bronze age

4 Iron age: 1000 BC 1850 AD 1.9 million BC Stone age 300 BC Iron age 1200 BC Bronze age

5 Steel age: 1850 AD 1930 AD 1.9 million BC Stone age 1200 BC Bronze age 300 BC Iron age 1906 Steel age

6 Scientific age 1930s Plastics 1940s Semiconductors

7 Scientific age 1930s Plastics 1990s High temperature superconductors 1940s Semiconductors 2000s Graphene

8 Jet engine

9 Jet engine

10 Designing a new alloy what is required? Yield strength Processibility Cost Fracture toughness Corrosion resistance Oxidation resistance Required properties for new alloy Creep Density Fatigue life

11 Materials pipeline Cr Co Mo W Ta Nb Al Ti Fe Mn Si C B Zr Cu N P V Hf Mg Ni and 4 different manufacturing processes

12 Materials pipeline Experimental databases Materials characterization Computational

13 Two new tools Experimental databases Neural network fitting Materials optimization Computational

14 Hardness Neural network fitting & optimization Target Aluminum

15 Hardness Neural network fitting & optimization Target Aluminum

16 Hardness Neural network fitting & optimization Target Aluminum

17 Optimizing the likelihood P P Cost Hardness Likelihood of alloy satisfying all targets P P Fatigue Weldability EP ; US 2014/177578; GB

18 Ni-base superalloy Amendment to US 2013/ A1; EP ; GB

19 Ni-base superalloy Amendment to US 2013/ A1; EP ; GB

20 Alloys discovered Discovery algorithm EP US 2014/ GB Mo-Hf forging alloy EP US 2014/ GB RR1000 grain growth Acta Materialia, 61, 3378 Ni disc alloy EP US 2013/ A2 GB Mo-Nb forging alloy EP US 2014/ GB Cr-Cr2Ta alloys Intermetallics 48, 62

21 Two new tools Experimental databases Neural network fitting Materials optimization Computational

22 Light emitting diodes Cost Efficiency Color Band gap

23 Light emitting diodes Cost Efficiency Color Band gap

24 Computer simulations

25 Band gap

26 Computational challenges Inevitable approximations behind first principles simulations Reducing number of simulations performed

27 Band gap expt Correlations between properties Aluminum

28 Band gap expt Correlations between properties Aluminum

29 Band gap expt Band gap DFT Correlations between properties Aluminum Aluminum

30 Aluminum Band gap expt Band gap expt Band gap DFT Correlations between properties Aluminum Aluminum

31 Correlations between properties Experimental database Neural network weights

32 Correlations between properties Experimental database Composition Neural network weights Prediction of band gap

33 Correlations between properties Experimental database Composition Neural network weights Prediction of band gap Composition Neural network weights Prediction of band gap DFT database

34 Correlations between properties Experimental database Composition Neural network weights Prediction of band gap Composition Neural network weights Prediction of band gap DFT database

35 Correlations between properties Cost Experimental band gap DFT carrier density Experimental carrier density Composition Predictions DFT band gap Experimental efficiency Experimental efficacy DFT efficacy

36 Correlations between properties Cost Experimental band gap DFT carrier density Experimental carrier density Composition Predictions DFT band gap Experimental efficiency Experimental efficacy DFT efficacy

37 Recursive learning Calculate material property Generate neural network models Band gap Search for optimal solution Target Aluminum

38 Recursive learning Generate neural network models Large uncertainty Calculate material property Band gap Search for optimal solution Target Aluminum

39 Case study: III-V InGaN-base semiconductors

40 Case study: III-V InGaN-base semiconductors Combine

41 Case study: III-V InGaN-base semiconductors

42 Three new tools Experimental databases Ni-based alloy EP / A1 GB Neural network fitting Property correlations Computational Mo-Hf alloy EP US 2014/ GB Mo-Nb alloy EP US 2014/ GB InGaN-based LED

43 Search for oil

44 Search for oil

45 Search for oil

46 Seismic survey

47 Seismic survey

48 Seismic survey

49 Prospects in the future Three tools in machine analysis to maximize information Maximum likelihood Correlations between properties Recursive learning Concurrent materials design