FDNS21: Revealing the Full Spectrum of 2D Materials with Superhuman Predictive Abilities

2021.01.20
Evan Reed, Stanford University, Stanford, CA
This talk is part of FDNS21: Future Directions in Nanomaterial Synthesis: From Rational Design to Data-Driven Manufacturing workshop sponsored by Nanomanufacturing (nanoMFG) Node at the University of Illinois at Urbana-Champaign. Presentations for this workshop can be found on nanoHUB at nanohub.org/resources/35001
Table of Contents:
00:00 Revealing the full spectrum of 2D materials with super-human predictive abilities
01:34 We Discover New 2D Materials
04:17 Our data mining identifies van der Waals-bounded 2D or 1D crystals
05:17 We develop an algorithm for identifying 2D or 1D crystals from database of bulk materials
07:20 We compile a genome of 1173 2D materials
10:21 We compile a genome of 1173 2D materials
12:07 We find a wide spectrum of 2D material band gaps
13:16 We seek all possible 2D synthesizable chemical formulas with physics-based machine learning
18:23 How to determine unsynthesizable materials: a key challenge
19:45 Support Vector Machine (SVM) attempts to separate layer and unlayered data with a line (hyperplane)
21:05 Model Performance (Precision) is 10X Better than Random Guessing
23:48 Semi-supervised learning: an alternate definition of non-layered materials
27:10 Our machine learning model predicts 2D materials better than (most) humans
28:18 Human versus algorithm comparison
29:30 Our machine learning model predicts 2D materials better than (most) humans
31:51 The model predicts 2D materials 5x better than expert practicioners
32:08 Predictions sorted by predicted bandgap
33:31 Our DFT indicates 13/16 success rate for select preductions
34:22 ML identifies families for which there are no known chemical analogs
36:00 Acknowledgements
36:41 We Discover New 2D Materials
37:55 Predictions sorted by predicted bandgap