From Conference Electronic Materials and Applications 2019 (EMA 2019), Abstract # 3065049, EMA-S13-013-2019 Symposium: S13: From Basic Science to Agile Design of Functional Materials: Aligned Computational and Experimental Approaches and Materials Informatics, January 24, 2019
1. 1
High-Throughput Ab Initio and Machine
Learning for Screening Perovskites
Dane Morgan, John Booske, Ryan Jacobs, Wei Li,
Lin Lin, Guangfu Luo, Tianyu Ma
University of Wisconsin, Madison
Abstract # 3065049, EMA-S13-013-2019
Symposium: S13: From Basic Science to Agile Design of
Functional Materials: Aligned Computational and
Experimental Approaches and Materials Informatics
January 24, 2019
2. National Energy Research
Scientific Computing Center
Oak Ridge National
Laboratory
NSF Extreme Science and
Engineering Discovery
Environment
DOE BES
Materials
Chemistry
Financial Support Computing Support
Center for
Nanoscale Materials
2
NSF grant
1148011.
INVEST
Program
Award #: 1720415
3. COMPUTATIONAL MATERIALS GROUP
Faculty
* Izabela Szlufarska * Dane Morgan
Staff Scientists
* Ryan Jacobs * Hubin Luo
Postdocs
* Ajay Annamareddy * Jianqi Xi
* Senlin Cui * Hongliang Zhang
Graduate Students
* Amy Kaczmarowski * Benjamin Afflerbach
* Chaiyapat Tangpatjaroen * Cheng Liu
* Dongzheng Chen * Lane Schultz
* Lin Lin * Tianyu Ma
* Dongzheng Chen * Shuguang Wei
* Shuxiang Zhou * Vrishank Jambur
* Yeqi Shi * Yipeng Cao
* Yu-chen Liu * Zhuohan Li
Undergraduate Students
* 35+ students involved in Informatics Skunkworks
https://skunkworks.engr.wisc.edu/
9. Key Cathode Property
Oxygen Surface Exchange (k*)
• A measure of intrinsic
ability to catalyze oxygen
reduction.
• Correlates strongly with
area specific resistance of
cathode and SOFC
performance
• Good database of values
for analysis from isotope
exchange/SIMS and
impedance spectroscopy.
9
Cathode
Material
O2
But DFT cannot predict k*! Need correlating descriptor!
10. O p-Band as k* Descriptor
10
PO2=0.2-1atm, T≈1000K
Good linear correlations for k* (went from 9 values in 2011 to
21 values in 2018).
11. High-Throughput Screening Process
MAST, VASP: DFT-GGA+U simulation of ≈2000 perovskites
Set of promising
high T ORR
materials
Elimination criteria (Stability):
Remove unstable compounds > 40
meV/formula unit above hull
Elimination criteria (Activity): Remove
compounds with predicted k* < LSCF
Elimination criteria
(Conduction):
Remove insulating
compounds
11
pypi.python.org/pypi/MAST
Jacobs, et al., Adv Energy Mat ‘18
17. Stability and Predicted Activity
17
• Predict many promising compounds. Best predicted k* for BaFe0.75Nb0.25O3. Similar to
BaCo0.625Fe0.25Nb0.125O3, which gives confidence in the prediction.
• Initial tests with infiltration of most promising materials at NETL (Harry Abernathy, Shiwoo
Lee) are encouraging (and show limitations).
19. Thermionic Electron Emitters
• Emit electrons when hot for high
frequency, high power vacuum
electronic devices (magnetrons,
klystrons, traveling wave tubes,
thermionic energy converters, etc.).
• Present materials are typically metals
with dipoles on surfaces (e.g., Os/W
with BaO). Volatile species, non-
uniform coatings are issues.
• Single-phase non-volatile low work
function oxide emitter cathodes are
promising alternative.
19
https://en.wikipedia.org/wiki/Hot_cathode; Cahen, David, and Antoine Kahn. Advanced Materials 15.4 (2003): 271-277;
Kirkwood, David M., et al. "IEEE Transactions on Electron Devices 65.6 (2018): 2061-2071; Yamamoto, Shigehiko. Reports
on Progress in Physics 69.1 (2005): 181.
Want low work function, stable, conductive materials
20. Getting Work Function Values
DFT predictions use slab geometries
and are slow, particularly for accurate
hybrid DFT methods.
20
Correlation of low work function
(001)AO surface to p-band used for
screening, followed up by careful hybrid
calculations for most promising systems
21. High-Throughput Screening Process
MAST, VASP: DFT-GGA+U simulation of ≈2000 perovskites
Set of promising
electron emitter
materials
Elimination criteria (Stability):
Remove unstable compounds > 50
meV/formula unit above hull
Elimination criteria (Work Function): Remove
compounds with work function > 2 eV
Elimination criteria
(Conductivity):
Remove > 0.5 eV
gap
pypi.python.org/pypi/MAST
22. Electron Emitter Screening Results
22
[1] Jacobs, et al., Adv Func Mat ‘16
XXXXXX
Removed for
Public Distribution
24. Halide Perovskites for Solar Cells
• ABX3 (X=Cl,Br,I) perovskites
are some of the most
promising new materials for
solar applications (tunable
direct bandgap, excellent light
absorption, long electron and
hole diffusion lengths, few
recombination centers).
• High efficiency, low-cost, easy
synthesis.
• Best performing materials,
e.g,. MAPbI3, tend to contain
toxic Pb and be unstable,
particularly in presence of
water.
24
Want no Pb, high stability, targeted gap, good efficiency
Jacobs, et al., Adv. Funct. Mat. ‘19
25. Band Gap Determination
• Standard LDA/GGA band
gaps tend to be quite
wrong.
• Hybrid (HSE) band gaps are
modestly accurate (~0.25
eV errors) but ~20x slower
• Use “multifidelity machine
learning approach”
approach to estimate HSE
values from GGA and first
screen on that.
• Then use HSE for final band
gaps and Shockley-Queisser
efficiency limit calculations.
25
Data from Kim, et al., Scientific Data ’17 and Pilania, et al., Comp. Mat. Sci. ‘17
MAE = 0.11 eV
30. Machine Learning to Accelerate Slower
Screening Steps
• Most screenings require stability
check that is “intermediate” but
time consuming.
• Opportunity to accelerate
screening with machine learning
30
k* from p-band Work function from p-band
HSE gap from GGA gap
31. Predicting Perovskite Stability
• Y = Stability above convex hull (1926 oxide
perovskites)
• X = Properties of elements in compound (e.g.,
atomic size) – using expanded MAGPIE database
[1]
• Assume Y=F(X) and determine F with fitting to
initial database. F = logistic regression, SVM,
Decision trees, Neural networks, Kernel ridge
regression.
• We use scikit-learn package with MAST-ML
wrapper [2]
31[1] L. Ward, et al. Comp. Mat. ‘16; https://bitbucket.org/wolverton/magpie, http://oqmd.org/static/analytics/magpie/doc/
[2] https://github.com/uw-cmg/MAST-ML
32. 10-fold Cross-Validation Performance
• Encouraging results suggests we can predict stability accurately.
• But what really is domain of the model?
32
Classification: F1=0.88 Regression: RMSE=28meV/atom
Li, et al., Comp. Mat. Sci. ‘18
33. Domain Assessment
33
CV in this region:
RMSE < 30 meV/atom
CV in this region:
RMSE = 73 meV/atom
Need to establish rigorous
bounds and ways to
manage errors for best
screening.
34. • High-throughput ab initio screening is
effective for discovering perovskites for
myriad applications (e.g., fuel cells,
thermionic electrodes, solar cells).
• A critical step is typically some
correlation to accelerate a slow step
(e.g., hybrid calculations, surfaces).
• Machine learning models can be
integrated to greatly expand scope,
efficiency of searchers, but clear
model domain/uncertainty is a key
challenge.
34
Conclusions
Thank You - Any Questions?