Ichiro Takeuchi University of Maryland

Size: px

Start display at page:

Download "Ichiro Takeuchi University of Maryland"

Jonah Black
6 years ago
Views:

1 High-throughput Experimentation and Machine Learning for Materials Discovery 55 Å 45 Å 35Å Ferroelectric library t s (Å) 25 Å 20 Å 15 Å 10 Å 5 Å No impurity Ti (3 Å) Ti (6 Å) Ti (9 Å) Cu (3 Å) Cu (6Å) Cu (9 Å) t i (Å) Permanent magnet library Superconductor library Ichiro Takeuchi University of Maryland

2 Outline Combinatorial approach to materials discovery: mapping of complex energy and property landscapes Search for new superconductors and supervised machine learning of experimental database Structural phase distribution mapping: unsupervised machine learning of diffraction data Active learning is the future of high-throughput experimentation

3 Acknowledgement Univ. of Maryland V. Stanev T. Gao Y. Iwasaki R. Greene NIST A. G. Kusne SLAC A. Mehta Duke University S. Curtarolo Support ONR, AFOSR, DOE

4 Chemical & Engineering News, August 2001

6 Combinatorial Libraries of Inorganic Materials Luminescent materials libraries, Science 279, 1712 (1998) Semiconductor gas sensor library, electronic nose, Appl. Phys. Lett. 83, 1255 (2003) Magnetic shape memory alloy library, Nature Materials 2, 180 (2003)

7 Various combinatorial experimental designs: discrete libraries vs composition spreads Composition A B A B C Composition spreads allow continuous mapping of physical properties and phase boundaries Run to run variation in ordinary experiments is removed

8 Exploration of Interesting Phase Boundaries Temperature Ferromagnetic phase Low symmetry Ferroelectric phase ferromagnetic compound composition ferroelectric compound Shape memory alloys, piezoelectric materials, multiferroic materials, magnetostrictive materials, etc.

9 Composition Spreads of Ternary Metallic Alloy Systems 3 spread wafer Mn Mn Co-sputtering scheme Ni Al Al Ni Phase diagram Composition is mapped using an electron probe (WDS) Review article: Green et al., JAP 113, (2013)

temperature samples are measured 0 0 13 25 38 50 63 75 col -2.

10 Rapid mapping of magnetic properties: scanning SQUID Raw data 80 Mn 60 row Ni Ga SQUID assembly inside vacuum Room temperature samples are measured col -2.50e e e+007 rho1_25 Nature Materials 2, 180 (2003)

11 Rapid mapping of magnetic properties: scanning SQUID Mn Raw data 80 Mn 60 row 40 Ni Ga 20 Ni Ga M (emu/cc) col -2.50e e e+007 rho1_25 Nature Materials 2, 180 (2003)

12 Rapid mapping of magnetic properties: constructing functional phase diagram Mn Ni Ga M (emu/cc) Nature Materials 2, 180 (2003)

13 Rapid mapping of magnetic properties: comparison with phase diagram C. Wedel and K. Itagaki, Journal of Phase Equilibria 22, 324 (2001) Nature Materials 2, 180 (2003)

14 Targeting superconductors predicted by theory Prediction: FeB 4 is a superconductor with T c ~ K

15 Fe-B composition spread: FeB x (x =2-4), 16 spots on one 1 cm 2 chip resistance temperature 4.2 K 300 K Middle region: FeB 2 FeB 4 Ch 3 Ch 11 Ch 13 more Fe more B

16 FeBx: it looks like a real superconductor Partial R drop Susceptibility shows diamagnetism ~ 10 K? B c2 (T)= B c2 (0)[1-(T/T c ) 2 ]/ [1+(T/T c ) 2 ] gives B c2 (0) = 2 T -> Type II BCS superconductor Superconducting phase was detected in 2 spread wafers APL Materials 1, (2013)

17 Mining superconducting databases Superconductors grouped by 3 golden coordinates Predictions of new compounds Based on ~ 600 superconductors (1988)

Visualization and mining of NIMS SuperCon database 29000 entries (Cuprates: more than 10000) Solution: create our ow from literature Removing

18 Visualization and mining of NIMS SuperCon database entries (Cuprates: more than 10000) Solution: create our ow from literature Removing misentries, duplicates, etc. results in entries Data mining via supervised machine learning using Magpie descriptors entries in MatNavi (2014)

1400 entries Random forest solves the problem of including

19 Machine learning modeling of superconducting critical temperature (random forest) Models a Models are built based on 1400 entries Random forest solves the problem of including nonsuperconducting materials Stanev, et al. arxiv preprint arxiv:

vs low T c Predicted T c (log) Low T c vs Fe-based Cuprates vs Fe-based

20 Machine learning modeling of superconducting critical temperature Overall prediction accuracy is pretty good Low T c vs cuprates Cuprates vs low T c Predicted T c (log) Low T c vs Fe-based Cuprates vs Fe-based Experimental T c (log) Different classes of superconductors can be distinguished

21 Machine learning modeling of superconducting critical temperature Different chemical properties/trends are revealed BCS superconductors ~

22 Machine learning modeling of superconducting critical temperature Combining SuperCon (experimental database), machine learning, ICSD, and AFLOW, we predict possible new superconductors DOS of predicted superconductors show unusual structure Stanev, et al. arxiv:

23 Rapid mapping of structural phase diagram C. Wedel and K. Itagaki, Journal of Phase Equilibria 22, 324 (2001) Nature Materials 2, 180 (2003)

$diffraction set up at SSRL The$ be measured in 2 hrs Transmission

24 Structural property mapping of combinatorial wafers Synchrotron diffraction set up at SSRL The entire 3 wafer (300 spots) can now be measured in 2 hrs Transmission set up Reflection set up XRF carried out simultaneously Each wafer produces: 300 MB to 2 GB of image data w/ A. Mehta

25 Intensity (Arb. Unit) Hundreds of XRD Patterns are difficult to analyze by hand 2 (Degrees) The same is true for any spectral data (Raman, FTIR, etc.)

$First step is visualization Peak Angle and Intensity Al Ni Mn Intensity Cross Section Mn Diffraction Intensity$

26 First step is visualization Peak Angle and Intensity Al Ni Mn Intensity Cross Section Mn Diffraction Intensity Ni Austenite Phase Al Martensite Phase Al Mn Ni I. Takeuchi et al., Rev. Sci. Instrum. 76, (2005)

Spectrum Number Fe 40 Pd 60 Adjust the dendrogram C. J. Long, et al. Rev. Sci.

27 Analyze all spectra together using cluster analysis: Look for similarities between spectra Group To Group Distance Multi Dimensional Data Scaling Dendrogram XRD Spectrum Number Fe 40 Pd 60 Adjust the dendrogram C. J. Long, et al. Rev. Sci. Instrum. 78, (2007) Fe Initial construction of structural phase diagram Fe 40 Ga 60

28 Comparing XRD patterns r xy = 1 D=0 r xy = 0 D=.5 r xy = -1 D=1 D = distance

Different metrics ( distance between data points (vectors)) <Euclidean> d n k x k y k 2 d 1

is often used in image recognition field < 1 Cos> D= 1 C <Pearson> D=(1 <Spearman> D=(1 C)/2

the cost matrix Known as normal correlation coefficient (parametric) C cor x x y y 2 x x y y 2 i

1 m n min fij i 1 j 1 f is minimum transportation cost This is non parametric index.

29 Different metrics ( distance between data points (vectors)) <Euclidean> d n k x k y k 2 d 1 <Manhattan> n x k y k <fast DTW> This is often used in speech recognition field <fast EMD> This is often used in image recognition field < 1 Cos> D= 1 C <Pearson> D=(1 <Spearman> D=(1 C)/2 C)/2 C k xi yi 2 xi sin 2 yi This is often used in text mining field To find a minimum pass in the cost matrix Known as normal correlation coefficient (parametric) C cor x x y y 2 x x y y 2 i i i i This is just transportation problem between two data series EMD P, Q m n min dij fij i 1 j 1 m n min fij i 1 j 1 f is minimum transportation cost This is non parametric index. d i is difference between ranking of data C spearman 6 1 N i 3 d i N Iwasaki et al., npj Computational Materials 3 (1), 4

30 Different metrics calculated and clustered using hierarchical clustering: XRD patterns from Fe Co Ni composition spread <Euclidean> <Manhattan> <fast DTW> <fast EMD> < Correct answer > Fe <1 Cosine> <Pearson> D=(1 C)/2 <Spearman> D=(1 C)/2 Co hcp Ni Y.K. Yoo et al. / Intermetallics 14 (2006) Iwasaki et al., npj Computational Materials 3 (1), 4

31 Another analysis method: Non-Negative Matrix Factorization (NMF): (The Basic Idea) S = W * B + E Experimental Spectra Weights Basis Spectra Residual Error Experimental Spectra are Deconvolved

32 Working toward a structural phase diagram Orthorhombic Fe and Pd Silicides Hexagonal FeGa FCC FePd Long et al., Not in Database Rev. Sci. Instrum. Not in Database 80, (2009) BCC Fe FCC Fe using NMF

Integrating ICSD with combi XRD data Most ternary phase diagrams are not known Each point on the ternary phase diagram is one X ray spectrum (expt)

33 Integrating ICSD with combi XRD data Most ternary phase diagrams are not known Each point on the ternary phase diagram is one X ray spectrum (expt) Points on binary lines are simulated spectra from ICSD They are rapidly mined/analyzed together mean shift theory Scientific Reports 4, 6367 (2014)

Each point on the ternary phase diagram is one X ray spectrum (expt) Points on

34 Fe 0.6 Pd 0.4 Integrating ICSD with combi XRD data Most ternary phase diagrams are not known Each point on the ternary phase diagram is one X ray spectrum (expt) Points on binary lines are simulated spectra from ICSD They are rapidly mined/analyzed together Fe mean shift theory Fe 0.6 Ga 0.4 Scientific Reports 4, 6367 (2014)

$Diffraction data (integrated Pilatus images) plotted for different compositions as they are taken$ X ray fluorescence data are also processed real time Hierarchical clustering is used to group

X ray fluorescence data are also processed real time Hierarchical clustering is used to group

35 Real time analysis of combinatorial library data (synchrotron XRD); Integration with ICSD Diffraction data (integrated Pilatus images) plotted for different compositions as they are taken X ray fluorescence data are also processed real time Hierarchical clustering is used to group similar spectra together Scientific Reports 4, 6367 (2014) Spectra are mapped on ternary phase diagram real time

36 Phase mapping based on active learning Semi Supervised Learning: Propagate knowledge from measured/known samples to unmeasured samples Quantify Certainty of Prediction Graph based phase separation algorithm used Mn Mn Ni Coarse Measurement Measured (Active) Not Measured Ge Ni Ge Sample to Query Certainty of Cluster Label (0 1) Runs Live at SLAC X Ray Beamline

Active Learning: Flow Chart Measure Composition Combinatorial Library Graph Coarse Structure Measurement Measured Samples (black) Machine Learning: Phase

37 Active Learning: Flow Chart Measure Composition Combinatorial Library Graph Coarse Structure Measurement Measured Samples (black) Machine Learning: Phase Diagram Determination Active Learning: Optimal Query Active Learning Implementation: Select query that minimizes risk Risk : estimated expected classification error

6 Pd 0.4 Previous results: non negative matrix factorization: Rev. Sci. Instrum.

38 Active Learning: Example Ternary Spread Demonstration : Fe Ga Pd Coarse Measurement Measured (Active) Not Measured Sample to Query Certainty of Cluster Label (0 1) Fe 0.6 Pd 0.4 Previous results: non negative matrix factorization: Rev. Sci. Instrum. 80, (2009) Hierarchical clustering: Rev. Sci. Instrum. 78, (2007) Fe Fe 0.6 Ga 0.4

39 Active learning: algorithm finds the boundaries for you Risk: Estimated expected classification error Iteration

40 Active Learning Phase Mapping Ni Mn Ge Mn Mn Mn Ni Ge Ni Ge Ni Ge Data from ICSD Data from AFLOW Initial measurements Mn (16) Combined initial information Ni Ge

information F measure used for accuracy 10% of samples

41 Active Learning Phase Mapping Ni Mn Ge Final mapping Mn Manual phase mapping Mn Ni Ge Ni Ge Mn Combined initial information F measure used for accuracy 10% of samples required for measurements to get 80% accuracy for the whole Ni Ge

Evolution of the combinatorial strategy (and a future forecast) Circa

samples fast # of samples: x 100-1000 How to measure large number of

large amount of data Machine learning to control and reduce number of

42 Evolution of the combinatorial strategy (and a future forecast) Circa and beyond Challenges: How to make hundreds of samples fast # of samples: x How to measure large number of samples: speed, number and quantitative accuracy How to quickly analyze large amount of data Machine learning to control and reduce number of samples # of samples: x # of samples: x Reducing the number of data points

experimentation Annual Machine Learning for Materials Research Workshop and Summer

43 Summary Combinatorial (high throughput) experimentation + theory can speed up discovery of materials Machine learning can greatly enhance combinatorial experimentation Annual Machine Learning for Materials Research Workshop and Summer Camp, Univ. of Maryland Sponsored by UMD and NIST June, 2017

Rapid structural mapping of ternary metallic alloy systems using the combinatorial approach and cluster analysis

REVIEW OF SCIENTIFIC INSTRUMENTS 78, 072217 2007 Rapid structural mapping of ternary metallic alloy systems using the combinatorial approach and cluster analysis C. J. Long, J. Hattrick-Simpers, M. Murakami,