Exploring chemical space: Molecular Space Shuttle

Transcription

1 Exploring chemical space: Molecular Space Shuttle Alán Aspuru-Guzik Professor Harvard University Twitter: A_Aspuru_Guzik

2 10 82 atoms in the observable universe

3 medium-size molecules

4 Long-term Goal: Mapping chemical space What is the nature of chemical space in terms of functionality What are the molecules reachable by certain synthetic rules and mechanisms Can we learn to automate quantum chemistry prediction for chemical space

5 Molecular screening for organic materials How good is this molecule as a battery material? Quantum Mechanics Machine Learning

6 From to 6 10 to 10 Initial library Computational screening Synthesis and testing Computational cost Molecules most likely to be of interest

7 Organic materials in the larger context US Materials Genome Initiative Large Medium Small Inorganic Materials Organic Materials Shared Features Timescale is important Automated techniques Data-driven discovery Computational funnel Number of descriptors Size of search space Level of approximation Organic Pharmaceuticals Pyzer-Knapp, et al. Ann Rev. Mat Sci. (2015)

8 My research group s explorations of chemical space The Harvard Clean Energy Project Generating renewable Blue Organic LED For your next gadget or TV energy Organic flow batteries Storing renewable energy Origins of life How life may have come about?

9 Project chronology and screening methodology improvements 1.0 Harvard Clean Energy Project Distributed highthroughput Experimental calibration + + MolecularSpace database 2.0 Organic flow batteries Tight experimental feedback cycle Stability screening + + Synthesizability at forefront 3.0 Organic Light Emitting diodes Super-flexible molecular builder Machine learning screening + + Infinite libraries 4.0 Harvard Clean Energy Project Thompson sampling + + Several experimental collaborations at once Molecular crystal ML and prediction

10 Harvard Clean Energy Project:Organic solar cells 10

11 The Harvard Clean Energy Project The Largest quantum chemistry screening project to date 30,000+ CPU years have led to more than 35,000 high-performance organic photovoltaic candidates. Collaborators: Juan Hindo (IBM) Zhenan Bao (Stanford), Johannes Hachmann (Buffalo), Alejandro Briseño (UMass), Carlos Amador (UNAM), and Hachmann, et. Al, J. Phys. Chem Lett. (2011), Energ. Env. Sci. (2014) 11

12 Energy Levels and Efficiency Scharber M, Mühlbacher D, Koppe M, et al., Advanced Materials (2006)

13 Sifting through 2.3 million molecules % ~35000 molecules (1.5% of sample space) Candidates Energy and Environmental Science, 7, 698 (2014)

14 The Clean Energy Project gets an artificial intelligence boost! 4.0 Machine learning Learn Calculate Prioritize Bayesian calibration Smart Screening Using machine learning Easy to synthesize libraries Collaborator: Ryan Adams (Harvard) Neural 14 Fingerprints E. O. Pyzer-Knapp, et al. Advanced Functional Materials 2015 E. O. Pyzer-Knapp, et al. arxiv: D. Duvenaud, arxiv: NIPS 2015

15 Organic Flow Batteries The Harvard Clean Energy Project Generating renewable Blue Organic LED For your next gadget or TV energy Organic flow batteries Storing renewable energy Chemical networks Origins of life Organic reactions Chemical autoencoders

16 Organics for storing clean energy Organic flow batteries Suh, et al., Chem. Sci., 6, 2015, p. 885 Huskinson, et al., Nature, 505, 2014, p. 195 Lin, et al., Science, 349, 2015, p Collaborators: Mike Aziz and Roy Gordon (Harvard)

17 Search space for redox potentials Handbook of Electrochemistry Ed. C.G. Zoski E (V vs. Saturated Calomel Electrode) Estimated potential range of organic functional group@ 25 o C

18 1,2-BenzoQuinones Choice for combinatorial library: 1R and fully substituted cases only 1,4-BenzoQuinones 1. N(CH 3 ) 2 2. NH 2 3. OCH 3 4. OH 5. SH 6. CH 3 7. SiH 3 8. F 9. Cl 10. C 2 H CHO 12. COOCH CF CN 15. COOH 16. PO 3 H SO 3 H 18. NO 2 2.0

19 Current'Status'(systema.c'explora.on)' 2,68naphtho' 2,38naphtho' 1,78naphtho' 1,58naphtho' 1,28naphtho' 1,48naphtho' Naphtoquinones 1,28benzo' 1,48benzo'

20 1,5%anthra* 2,9%anthra* 1,10%anthra* 1,4%anthra* Anthraquinones 2,6%anthra* 2,3%anthra* 1,7%anthra* 1,2%anthra* 9,10%ahthra*

21 Theoretical calibration of quinone redox potentials 1.5 GEN-4 model with inclusion of GEN-4 molecules (ADS and PDS) and NQPS Quinones used for GEN-1,2 and 3 models Added quinones for GEN-4 model ASA (GEN-3 molecule) (V vs. SHE) 1.0 QSA (GEN-3 molecule) NQPS 1,2-BQ 1,4-BQ E 0 exp 0.5 1,2-NQ 1,4-NQ 0.0 R 2 : RMSD: V 9,10-PQ AQDS (GEN-1 molecule) DHAQDS (GEN-2 molecule) 9,10-AQ PDS (GEN-4 molecule) AQS (GEN-1 molecule) ADS (GEN-4 molecujle) 1,5-DHAQDS (GEN-2 molecule) ΔH (ev) QSA (GEN-3 molecule) ASA (GEN-3 molecule)

22 > 300 new candidate quinones predicted S. Er., C. Suh, M. P. Marshak, A. Aspuru-Guzik, Chemical Science (2015)

23 Our metal-free aqueous flow battery 2.0 b V 3+ /V 2+ VO 2+ /V 3+ VO 2 + /VO 2+ Br 2 /Br E 0 (V vs SHE) ,5-AQ 2,3-AQ 2,6-AQ 1,7-AQ 2,9-AQ 1,10-AQ 1,2-AQ 1,4-AQ 9,10-AQ 2,3-NQ 2,6-NQ 1,7-NQ 1,5-NQ 1,4-NQ 1,2-NQ 1,4-BQ 1,2-BQ Computational screening of 10,000 quinone molecules HO 3 S Intense design cycle O Synthesize molecules SO 3 H Test in flow battery O Selected molecule

24 Theory-experiment collaboration Michael Aziz Engineering Roy Gordon Chemistry Alán Aspuru-Guzik Chemistry Nature, 505, 2014, p. 195

25 17 COOCH 3 CHO CHO COOCH 3 COOCH 3 COOCH 3 COOCH 3 COOH COOH CHO COOCH 3 CHO CHO COOCH 3 CHO CN COOCH 3 16 CN COOH COOCH 3 CN CN COOH COOH COOCH 3 CN COOCH 3 CN CF 3 COOCH 3 CHO COOCH 3 COOCH 3 CN 15 COOH CN Molecular COOH CF 3 COOH Flow CN CHO Battery CN COOCH 3 Data COOH CHO View CN SO 3 H CN CN COOH COOH 14 CF 3 COOCH 3 CN COOH CF 3 SO 3 H CN CF 3 CF 3 CN COOH COOCH 3 CN COOH CF 3 CHO OCH 3 13 CHO PO 3 H 2 SO 3 H SO 3 H PO 3 H 2 CF 3 CF 3 CHO SO 3 H CF 3 CF 3 COOH CF 3 CF 3 COOH CF 3 CF 3 12 Cl CF 3 Blue: CF 3 Cl Stable Cl CHO molecule SO 3 H Cl CHO SO 3 H SO 3 H PO 3 H 2 COOH PO 3 H 2 SO 3 H PO 3 H 2 F 11 F Cl PO 3 H 2 F SO 3 H PO 3 H 2 Cl PO 3 H 2 Cl PO 3 H 2 PO 3 H 2 Cl PO 3 H 2 Cl PO 3 H 2 Cl Cl Rank of ΔE SO 3 H PO 3 H 2 OCH 3 SO 3 H F C 2 H 3 Red: Unstable molecule Cl CHO F Cl F F F SiH 3 PO 3 H 2 CHO F PO 3 H 2 SO 3 H SiH 3 F OCH 3 OCH 3 OCH 3 OCH 3 SiH 3 PO 3 H 2 Cl Cl F F SiH 3 SiH 3 F SiH 3 SO 3 H Cl F Cl SO 3 H F SO 3 H F F SiH 3 SiH 3 SiH 3 SiH 3 CHO N(CH 3 ) 2 C 2 H 3 7 C 2 H 3 SiH 3 X axis: Redox Potential C 2 H 3 C 2 H 3 C 2 H 3 SiH 3 SiH 3 OCH 3 C 2 H 3 C 2 H 3 C 2 H 3 C 2 H 3 OCH 3 C 2 H 3 C 2 H 3 OCH 3 SO 3 H 6 SiH 3 CH 3 Y CH 3 axis: SiH 3 Free SiH 3 C 2 H energy 3 C 2 H 3 C 2 H 3 of CH 3 Solvation CH 3 CH 3 SH C 2 H 3 SH OCH 3 C 2 H 3 PO 3 H 2 5 CH 3 N(CH 3 ) 2 OCH 3 SH N(CH 3 ) 2 CH 3 CH 3 CH 3 OCH 3 OCH 3 OCH 3 OH CH 3 CH 3 CH 3 CH 3 SiH 3 Lowering E N(CH 3 ) 2 SH NH 2 OH OCH 3 SH CH 3 CH 3 SH SH N(CH 3 ) 2 SH SH SH SH ~ OH 100,000 OH SH N(CH molecules 3 ) 2 N(CH 3 ) 2 SH N(CH shown 3 ) 2 N(CH 3 ) 2 N(CH 3 ) 2 OH N(CH 3 ) 2 NH 2 OH OH OH OH NH 2 OH OH NH 2 NH 2 N(CH 3 ) 2 NH 2 NH 2 NH 2 NH 2 OH NH 2 NH 2 CH 3 OCH 3 NH 2 N(CH 3 ) 2 SH OCH 3 SH N(CH 3 ) 2 OH N(CH 3 ) 2 N(CH 3 ) 2 SH N(CH 3 ) 2 OH OH OH NH 2 NH 2 NH 2 NH 2 CH 3 SH OH NH 2 1,10-AQ 1,2-AQ 1,2-BQ 1,2-NQ 1,4-AQ 1,4-BQ 1,4-NQ 1,5-AQ 1,5-NQ 1,7-AQ 1,7-NQ 2,3-AQ 2,3-NQ 2,6-AQ 2,6-NQ 2,9-AQ 9,10-AQ

26 Molecular Flow Battery Data View Filtering the data view

27 Molecular Flow Battery Data View Baseball card view

28 High-throughput materials discovery process and tools Moelcular Space Shuttle: advanced molecular discovery platform Info Cards Voting Interface Detailed Tables Plotting Data Access Bubble Plot Web tools and critically enable partner communication and successful molecular discovery 28

29 Feedback tool Database-backed web system tracks: ~1,000,000 machine-generated molecules ~1,500 (8000 including oxidation, decomposition and dissociation products) 29

30 : Complex quinone redox pathways Highly reduced Highly oxidized Interactive data manipulation via web 30

31 Additional current-theory work 1: quinone stability Screening procedures excluding potential Michael addition 1 st screening with Michael addition 9,866 couples 22,364 couples 2 nd screening: Fewer than two R-groups 2,290 couples 3 rd screening: Fewer than two R-groups and good solubility (ΔG 0 solv < ev ) 2,052 couples X-axis: Quinone redox potential (E 0 1) / Y-axis: Stability (K hyd ) Warmer colors represent higher density of molecules

32 Additional current-theory work 2: Second-oxidation quinones Screening Procedures with consideration of Michael addition 1 st screening with Michael addition 8,252 couples 762 couples 2 nd screening: Fewer than two R-groups 98 couples 3 rd screening: Fewer than two R-groups and good solubility (ΔG 0 solv < ev ) 84 couples

33 Molecular baseball cards including stability

34 Beyond quinones Sulfolobus archaebacteria Pineda-Flores, et al. J. Phys. Chem. C (2015)

35 Long-lasting blue organic LED The Harvard Clean Energy Project Generating renewable Blue Organic LED For your next gadget or TV energy Organic flow batteries Storing renewable energy Origins of life How life may have come about?

36 Harvard-MIT collaboration Harvard MIT Ryan Adams Machine Learning Tim Swager Stephen Buchwald Synthetic Chemistry Marc Baldo Alán Aspuru-Guzik High-throughput quantum chemistry Device Engineering Troy Van Voorhis Microscopic theory >450,000 molecules screened so far! ~25 synthesized and tested

37 Speedy screening

38 Machine Learning Supervised learning algorithms Neural networks for ultrafast predictions leveraging thousands of data-points. Result in 10x speedup by discarding poor candidates Role of dimensionality Chemical space is sparse but libraries are dense. Powerful interpolation Explore-exploit strategy 5/9/16 38

39 Selecting molecules is like dating.

40 Organic LED Screening Synthetic accessibility voting tool

41 Neural Net Training Workflow

42 Data mining 500,000 quantum calculations 4CzIpn R. Gomez-Bombarelli, et al. Submitted (2015) 5/9/16 42

43 Batches Batches Selection of molecules for experimentalists to browse in a contained way. Usually explore some chemical family, using ancestry from database. Need to confirm novelty post hoc: sometimes re-discover known molecules. 5/9/16 43

44 Synthesis of compounds Molecule Sample Baseball card Not the true structure 22.5% High performance devices Device 16.7% EQE!! Rafael Gómez-Bombarelli, Jorge Aguilera-Iparaguirre, Tim Hirzel Martin BloodAdams, Baldo, Swager groups, Samsung IT

45 R. Gomez-Bombarelli, et al. Submitted (2015) Key breakthroughs in efficiency: Strength Name S 0 splitting T 1 splitting S 0 strength T 1 strength EQE(%) 4CzIPN Foxtrot Hotel Julie Lima * A small gap is crucial for TADF behavior We need also need a big fluorescence We have managed to control both for great overall efficiency

46 Screening billions of molecules: takes the driver s seat

47 To design something really well you have to get it. You have to really grok what it s all about. It takes a passionate commitment to really thoroughly understand something. Chew it up, not quickly swallow it. Most people don t take time to do that.

48 Aspuru-Guzik group Sponsors: DOE BES, ARPA-E, Samsung, NSF, ARO, ONR, AFOSR, Samsung, Sloan Foundation, Camille and Henry Dreyfus Foundation, DTRA, DARPA Twitter: A_Aspuru_Guzik