You are on page 1of 8

Computational Materials Science 161 (2019) 143–150

Contents lists available at ScienceDirect

Computational Materials Science


journal homepage: www.elsevier.com/locate/commatsci

Accelerating materials science with high-throughput computations and T


machine learning
Shyue Ping Ong
Materials Virtual Lab, Department of NanoEngineering, University of California San Diego, 9500 Gilman Dr, Mail Code 0448, La Jolla, CA 92093-0448, United States

ARTICLE INFO ABSTRACT

Keywords: With unprecedented amounts of materials data generated from experiments as well as high-throughput density
Machine learning functional theory calculations, machine learning techniques has the potential to greatly accelerate materials
High-throughput discovery and design. Here, we review our efforts in the Materials Virtual Lab to integrate software automation,
Materials discovery data generation and curation and machine learning to (i) design and optimize technological materials for energy
Materials design
storage, energy efficiency and high-temperature alloys; (ii) develop scalable quantum-accurate models, and (iii)
Multi-scale models
enhance the speed and accuracy in interpreting characterization spectra.

1. Introduction human performance in a variety of tasks, from playing board games


such as chess and Go [15] to interpretation of medical images [16] to
Computational methods are today an important complement to autonomous driving. It should be noted that most large computed
experiments in the study and design of materials. Computational materials databases (and the ML models on which they are based) are
techniques provide the means to precisely control the simulation con- still constructed using the Perdew-Berke-Ernzerhof (PBE) [17] gen-
ditions (e.g., modification of a single parameter or set of parameters), eralized gradient approximation (GGA) functional for reasons of com-
something which is difficult to truly achieve under experimental con- putational efficiency, which has well-known failures in certain patho-
ditions. Such controlled “virtual experiments” can often lead to fun- logical systems, such as those where strong electron correlation and van
damental insights on the factors determining a particular material der Waals interactions play a significant role [18–21]. Nevertheless,
property or set of material properties, enabling rational materials de- this is a limitation that we believe will be gradually addressed by
sign. Among computational techniques, first principles electronic methodological advances, for example, such as the SCAN functional
structure methods, which start from the basic laws of physics with [22], in the future.
minimal assumptions, are particularly useful due to their chemical In the Materials Virtual Lab, we leverage on our core expertise in
agnosticity. software automation [3–5], data generation and curation [23–27], and
In the recent decade, electronic structure codes have matured and machine learning in three ways: (i) to optimize existing materials or
computing power has advanced to a degree that first principles calcu- discover new technological materials, (ii) to develop models to access
lations, especially those based on more cost-effective approximations beyond-DFT scales at near-DFT accuracy, and (iii) to accelerate and
such as Kohn-Sham density functional theory (DFT) [1,2], can be re- enhance the accuracy of materials characterization.
liably automated [3–5] for high-throughput property prediction across
vast number of materials. High-throughput DFT calculations have been 2. Materials optimization and discovery
used in successful materials design efforts for applications as varied as
alkali-ion batteries [6–8], catalysts [9], topological insulators [10], and 2.1. Rechargeable alkali-ion batteries
organic semiconductors [11], and led to the development of large, high-
quality open databases of computed materials properties such as the As rechargeable alkali-ion batteries move beyond consumer elec-
Materials Project [12], Open Quantum Materials Database [13] and the tronic applications into large-scale applications such as transportation
AFLOW repository [14]. This explosion in the quantity of computed and grid storage, there is an urgent need for novel materials and battery
materials data has also begun to fuel interest in the application of architectures that not only deliver higher gravimetric and volumetric
machine learning to accelerate the study and design of materials. Ma- densities, but are also safer. To this end, our research efforts on this
chine learning (ML) has emerged as a means to mimic, or even surpass, front are focused on two parallel efforts: (i) solid electrolytes for safer,

E-mail address: ongsp@eng.ucsd.edu.

https://doi.org/10.1016/j.commatsci.2019.01.013
Received 27 October 2018; Received in revised form 10 January 2019; Accepted 10 January 2019
0927-0256/ © 2019 Elsevier B.V. All rights reserved.
S.P. Ong Computational Materials Science 161 (2019) 143–150

more energy dense all-solid-state alkali-ion batteries; and (ii) multi- solid electrolytes (see Fig. 1) [41]. We have also identified several
electron cathodes for high-capacity batteries. binary oxides (Sc2O3, SiO2, TiO2, ZrO2 and HfO2) to be chemically inert
with respect to electrodes and SEs and therefore are potential candidate
2.1.1. All-solid-state batteries buffer layers in addition to the commonly used Al2O3 [41]. Finally, we
There has been a surge in interest in all-solid-state rechargeable have demonstrated that AIMD simulations of an explicit electrode/SE
alkali-ion batteries (SSABs) in recent years [28–32]. By replacing interface model predict that the oxidation of S2− to form SxOy may be
kinetically preferred over the formation of PO34 , in contrast to the
flammable organic solvent electrolytes with nonflammable solid elec-
thermodynamic predictions [41]. This hierarchy of models provide a
trolytes (SE), SSABs have the potential to be safer and to provide gains
spectrum of computational approaches to study interfacial stability, and
in system-level energy densities through stacking or enabling new
provide much-needed guidance for the future development of practical
electrode chemistries (e.g., high-voltage cathodes and alkali metal an-
SSABs.
odes. To be an effective SE, a material has to have extremely high ionic
conductivities (> 0.1 mS/cm), i.e., be a superionic conductor, and
exhibit good chemical and mechanical compatibility with the elec- 2.1.2. Multi-electron cathodes
trodes. In the Materials Virtual Lab, we have sought to simultaneously Cathodes in commercial lithium-ion batteries today operate on the
address all these requirements using DFT computations [26,33–42]. basis of a single redox per transition metal on average. For example,
Using nudged elastic band [43] calculations coupled with a bond per- LiCoO2 and LiFePO4 operate on the Co3+/4+ and Fe2+/3+ redox cou-
colation model, we have proposed rational composition optimization ples, respectively. To achieve higher capacities, one approach is to
strategies for the lithium-rich anti-perovskite [44] family of superionic develop multi-electron cathodes that can reversibly cycle more than
conductors [33]. We have also used ab initio molecular dynamics one electron per transition metal [45]. A particularly promising family
(AIMD) simulations to propose cation and anion doping strategies for of multi-electron cathodes are the vandayl phosphates with formula
Na3PS4 [34,35], a highly promising SE for all-solid-state Na-ion bat- VOPO4. Utilizing the V3+/4+/5+ couple, VOPO4 has high theoretical
teries [31]. These works have led to the successful development of Cl- capacities of 308 mAh/g [46] and relatively high voltages due to the
doped Na3PS4 with experimentally-measured room temperature ionic inductive effect of the PO4 polyanion [47]. Through a highly colla-
conductivity exceeding 1 mS/cm, and stable cycling demonstrated over borative effort integrating DFT calculations, in operando characteriza-
more than 10 cycles in a TiS2/Cl-Na3PS4/Na cell [35]. On a broader tion and electrochemical measurements, we have provided crucial in-
level, our work has provided insights into the crucial role of defects and sights into the thermodynamics, phase evolution and kinetics of the
alkali concentration play in optimizing ionic conductivity [33–38], and polymorphs of VOPO4 upon lithiation [48–51]. We have successfully
the intrinsic mechanical properties [26] and electrochemical stability identified the structures of the intermediate phases -Li1.5VOPO4 and
[35–38] of various classes of superionic conductors. -Li1.75VOPO4, and showed that nanosizing and carbon coating are
Besides bulk ionic conductivity, a critical bottleneck to the practical critical to achieving good electrochemical performance due to the poor
application of SSABs is the high resistance and poor stability of elec- electronic conductivity and 1D diffusion topology of -LixVOPO4 [48].
trode/solid electrolyte interface. Using various thermodynamic ap- More recently, we have performed a comprehensive DFT study of the
proximations with DFT calculations, we have shown that the formation , and -I polymorphs of VOPO4, which suggested the phase may
of energetically favored PO34 through an exchange reaction is the main yield superior kinetics for Li diffusion, but only the layered -I phase is
cause of instability between layered oxide cathodes and thiophosphate predicted to have sufficiently low migration barriers for Na [49].

(a) (b)
Fig. 1. Reaction energies and volume changes for (a) discharged electrode/solid electrolyte pairs and (b) potential buffer layer materials and active materials, in all-
solid-state sodium-ion batteries. Reproduced with permission from Ref. [41].

144
S.P. Ong Computational Materials Science 161 (2019) 143–150

Furthermore, other VOPO4 derived from bigger alkali cations, e.g., earths or expensive elements in large quantities. There is therefore a
KVOPO4, may also yield low migration barriers for Na+ and high critical need to find new earth-abundant phosphors as well as to further
achievable capacities and energy densities [52]. optimize the performance of existing phosphors.
In the Materials Virtual Lab, our focus has been primarily on
phosphors utilizing Eu2+ and Ce3+ activators. One of our major
2.2. Phosphors for next-generation solid-state lighting achievements has been the discovery of the Sr2LiAlO4 phosphor host
(see Fig. 2) [55]. Sr2LiAlO4, the first known quaternary crystal in the Sr-
Phosphor-converted white-light-emitting diodes (pc-WLEDs) are an Li-Al-O system, was identified to be a thermodynamically and dyna-
energy efficient and environmentally friendly next generation lighting mically stable crystal by a machine-learned structure prediction algo-
technology [53,54]. For white light for general illumination, the key rithm [56] and DFT screening effort, guided by statistical analysis of
performance metrics are high quantum efficiency, resistance to thermal known phosphors in the Inorganic Crystal Structure Database (ICSD)
quenching, and high color quality, i.e., a low color-correlated tem- [57]. Experimental collaborators have successfully synthesized
perature (CCT) of <3000 K and a high color-rendering index (CRI) of > Sr2LiAlO4 and confirmed that the Eu2+- and Ce3+-activated phosphors
85. Even today, commercial WLEDs such as those based on blue LED exhibit broad emissions at 512 nm (green-yellow) and
max
chips with a yellow-emitting phosphor (Y3Al5O12:Ce3+)) have poor CRI 434 nm (blue), respectively, with excellent thermal quenching
max
<80 and high CCT >5000 K, and many contain phosphors with rare

Fig. 2. (a) Frequency at which each element appears in compounds having the word “phosphor” in the publication title in the 2017 version of ICSD. (b) DFT 0 K SrO-
Li2O-Al2O3 phase diagram. Blue circles, known stable phases in the Materials Project database [12]; red square, new stable quaternary phase, Sr2LiAlO4. (c) Unit cell
and Sr local environment (in Å) of the Sr2LiAlO4 crystal. (d, e) Measured excitation and emission spectra of (d) Sr2LiAlO4:0.005Eu2+ and (e) Sr2LiAlO4:0.005Ce3+
phosphors. Colors are indicated under the emission spectra for easy reference. Reproduced with permission from Ref. [55].

145
S.P. Ong Computational Materials Science 161 (2019) 143–150

resistance of >88% intensity at 150°C and excellent CRI exceeding 90 in in the 5 (310) tilt and 5 (100) twist Mo GBs [64]. We have demon-
a prototype WLED. This is a rare example of the successful discovery of strated that segregation preference between different GBs was due to
a technological material from computational prediction to experimental the strain effect of the dopants and segregation site. The smaller seg-
confirmation to device fabrication. We have also developed an elec- regation site of the twist GB will better facilitate segregation of smaller
tronic structure descriptor for emission bandwidth and identified po- dopants. We also demonstrated that previous empirical continuum
tential narrow-band red phosphors, which are of great interest for im- models such as the McLean [65] and Miedema [66] do not account for
proving CRI [58]. the effects of intermetallic compound formation between the dopants
Using DFT calculations, we have provided insights into the struc- and host material. Furthermore, Ta, W, Re and Os were predicted to
ture-composition-photoluminescent property relationships in strengthen the 5(310) tilt Mo GB while dopants such as Bi acted as GB
-SiAlON:Eu2+ [59], one of the most promising narrow-band green embrittlers, which is in agreement with previous experimental ob-
phosphors for high-power light-emitting diodes and liquid crystal dis- servations [67]. Using multilinear regression, we have successfully
play back-lighting with wide color gamut. We showed that reducing modelled the GB strengthening energy as a function of the dopant strain
oxygen content is a pathway to reduce red shifting and achieving nar- and cohesive energy.
rower band emissions, predictions which have independently been Most practical applications of Mo involves the use of its inter-
confirmed by experiments [60]. metallic compounds such as MoSi2, which has a melting temperature up
to 2030 °C and excellent oxidation resistance [68]. Despite its out-
2.3. Molybdenum and its alloys standing refractory properties, MoSi2 is brittle at intermediate tem-
peratures (400–600 °C) due to the oxygen short-circuit diffusion at the
Mo and its alloys are known for their high stiffness, low coefficient GB [68,69]. Alloying with Zr has been shown to improve the ductility of
of thermal expansion and the ability to maintain excellent mechanical Mo-based host materials [70–73]. Through the use of GB and interfacial
properties under extreme temperatures [61]. This makes them well models, we have shown using DFT calculations that the most effective
suited for use in harsh environments such as those found in aerospace strengthening mechanism of Zr in MoSi2 is as “getter” nanoparticles
engineering, defense applications and nuclear power plants. However, [74]. While Zr dopants somewhat mitigate the embrittling effect of
their wider applicability is limited by their intrinsic brittleness due to oxygen interstitials, a Zr secondary phase nanoparticle has a strong
their fragile grain boundaries (GBs) under room temperature [62,63]. affinity for parasitic oxygen (see Fig. 3(a)). This getter effect sig-
Introducing dopants and secondary “getter” phases are common ap- nificantly increases the work of separation of the GB. Based on these
proaches to address these limitations. findings, we have also developed a simple and effective screening ap-
We have performed a large-scale DFT study of 29 metallic dopants proach to identify other potential getter materials for MoSi2(Fig. 3(b))

Fig. 3. (a) The relationship between the position of interstitial oxygen and relative energy. The energy of Oint in bulk-like MoSi2 is set as a zero reference. (b) Two
descriptors: silicate formation energy and oxygen exchange reaction energy, are used to assess the viability of 55 elements as getter materials. The regions are labeled
as follows: I: oxygen exchange reaction favored and silicate formation energy is lower than Zr; II: oxygen exchange reaction favored and silicate formation energy is
higher than Zr; III: oxygen exchange reaction not favored but silicate formation energy is negative. Reproduced with permission from [74].

146
S.P. Ong Computational Materials Science 161 (2019) 143–150

which can be readily extended to other systems.


Our results have provided highly useful insights into the mechan-
isms behind the enhancement of mechanical properties of Mo and its
alloys due to atomic dopants and getter materials and will serve as a
guide towards the design and optimization of similar materials.

3. Beyond DFT scale, at near-DFT accuracy

A fundamental challenge in computational materials science is the


trade-off between scale (length and time), accuracy and transferability.
Today, empirical potentials can access reasonable length (> nm) and
time ( µ s) scales but are neither transferable across chemistries nor
especially accurate. Quantum mechanical calculations, such as those
based on DFT, are transferable and accurate but are too computation-
ally demanding for large systems and/or long time scales and/or large Fig. 4. High-temperature Ni-Mo phase diagram from experiments [81], CAL-
chemical spaces. In the Materials Virtual Lab, we aim to address this PHAD [82], SNAP, and EAM models. Reproduced with permission from Ref.
“scale challenge” by developing accurate potential models that scale [77].
linearly with number of atoms, as well as property prediction models
that enable rapid screening across vast chemical spaces using machine
learning. enabler to high accuracy studies of microstructure and other phe-
nomena requiring large-scale simulations over long-time scales.
3.1. Quantum-accurate, linear scaling potentials
3.2. Property prediction
We have developed highly accurate Spectral Neighbor Analysis
Potential (SNAP) [75,76] models for various metals (Mo, Cu, Ni) and Predicting stability is usually the first step in any materials dis-
alloys (Ni-Mo) [77,78]. The starting point is a robust description of the covery, and the formation energy of a crystal is typically the metric by
local environments in a system of atoms that satisfies rigorous in- which stability is determined. Though DFT calculations have proven to
variances under translation, rotation, and permutation of homonuclear be effective in estimating formation energies [83,84], they remain far
atoms and is unique and differentiable [79]. In the SNAP model, the too computationally expensive to traverse large combinatorial chemical
energies and forces are expressed as a linear function of the coefficients spaces. It is therefore unsurprising that there have been substantial
of the bispectrum built from the four-dimensional spherical harmonic efforts in the development of machine learning models to predict for-
expansion of the local 3D atomic density (see Bartók et al. [79] for mation energies [85–89].
details). The simple formalism reduces the potential for over-fitting In the Materials Virtual Lab, we have recently developed deep
compared to more complicated models and is expected to be more neural network models [90] that are able to predict the DFT formation
generalizable. Using automated workflows developed using pymatgen energies of garnets and perovskites to within 7–34 meV/atom accuracy
[3] and fireworks [4], we have generated accurate and consistent DFT (see Fig. 5), a substantial improvement over prior efforts [91]. To
datasets for our systems of interest, which include elastically distorted overcome well-known limitations of DFT calculations in handling redox
structures, vacancy structures, surfaces and grain boundaries, and reaction energies [92], we propose the formation energy from the
snapshots from AIMD simulations [77,78]. The models are then trained binary oxides as the appropriate target for a machine-learning model.
to reproduce desired properties, e.g., elastic constants, with a feedback Starting from a chemically-intuitive hypothesis that a quantitative re-
look that optimizes model hyperparameters using a differential evolu- lationship exists between the electronegativity and ionic radius of the
tion global optimization scheme. We have developed an open-source species on each symmetrically distinct site, we trained deep neural
materials machine learning python package called veidt [80] that is networks to model that relationship using the DFT calculated formation
freely available to the research community. We have also published our energies on thousands of unmixed and mixed garnets and perovskites.
fitted SNAP models in a Github repo at https://github.com/ We have shown that these models can be used to classify stable/un-
materialsvirtuallab/snap. stable garnet and perovskite compositions with >80% accuracy. A
In general, the SNAP models for bcc Mo, fcc Ni, fcc Cu and Ni-Mo particularly important feature of our work is the extension of these
binary systems significantly outperform current state-of-the-art em- neural network models beyond simple garnets and perovskites to mixed
bedded atom method (EAM) and modified embedded atom method compositions with negligible increase in prediction error, which greatly
(MEAM) empirical potentials. Not only are the mean absolute errors expands their applicability of the vast combinatorial chemical spaces
(MAEs) lower for energy and force predictions, the SNAP models also for discovery of novel technological materials. It is our belief that si-
yield more accurate elastic constants, phonon dispersion curves, milar approaches can be applied to develop accurate machine learning
equations of states, melting points and surface energies [77,78]. An models for prediction of other functional properties (e.g., band gap,
especially stark result can be seen in Fig. 4, where the Ni-Mo SNAP mechanical properties, etc.), potentially enabling multi-property opti-
model was used in a hybrid Monte-Carlo/Molecular Dynamics (MC/ mization of materials for a targeted technological application.
MD) simulation of the Ni-Mo phase diagram. While the EAM potential
fails to achieve even qualitative agreement with the experimental phase 4. Accelerating and enhancing materials characterization
diagram, the SNAP model is able to reproduce the correct solubility
limits of Ni in Mo, the concave liquidus line and the almost-overlapping Finally, the Materials Virtual Lab aims to enhance the speed and
liquidus and solidus lines at the Ni-rich portion of the phase diagram accuracy of materials characterization by combining automated com-
near the melting point of Ni. We believe that such models, which are putations, experimental data curation and ML. Though similar ap-
about five orders of magnitude cheaper than DFT calculations and proaches have been applied in the life sciences [16], the critical bot-
linear scaling with respect to the number of atoms, will be an important tleneck to the application of similar approaches in materials science is

147
S.P. Ong Computational Materials Science 161 (2019) 143–150

a b

c d

Fig. 5. Plot of neural network-predicted formation energy E fANN against DFT-calculated formation energy EfDFT for (a) unmixed garnets and (c) unmixed perovskites,
and mean absolute errors (MAEs) in Ef in training, validation, and test sets of all (b) garnet and (d) perovskite models. Mixing is carried out on the 24c (C-mixed), 16a
(A-mixed) and 24d (D-mixed) sites of the Ia3̄d C3A2D3O12 garnet crystal, and the 4c (A-mixed) and 4d (B-mixed) sites of the Pnma ABO3 perovskite crystal. The similar
MAEs for the mixed and unmixed models indicate that the former has learned the effect of orderings on Ef . Reproduced with permission from Ref. [91].

the lack of large, high-quality datasets for the training and development Berkeley Lab and Advanced Photon Source at Argonne National Lab.
of models. We are developing ML models that can significantly outperform hu-
We have simultaneously sought to address this data gap, as well as mans in terms of both speed and accuracy in the interpretation of
push the frontiers in developing the application of machine learning materials spectra and images (e.g., X-ray, neutron, electron). It is our
techniques in materials characterization. Using the Green’s function belief that such approaches will complement continued advances in
approach [93,94], we have constructed the world’s largest reference instrumentation, especially in in operando studies and artificial in-
database of more than 500,000 computed K-edge X-ray absorption near telligence-guided design of experiments.
edge (XANES) spectra for more than 40,000 unique materials [24,25],
and are extending coverage to the extended fine structure (EXAFS) and 5. Conclusion
L-edges. We have also developed an Ensemble-Learned Spectra IdEn-
tification (ELSIE) algorithm (Fig. 6) that is able to match experimental To conclude, we believe that machine learning will have a trans-
spectra with computed ones to identify the coordination environment formative impact on the conduct of materials science, once critical is-
and oxidation state with 84% accuracy. These resources and tools sues with regards to data quality and availability are addressed. In the
have been made publicly available via the Materials Project [12], and Materials Virtual Lab, we aim to address the data gap through high-
have garnered substantial interest from users (>2000 spectra requests/ throughput DFT computations and experimental data curation, enabled
month) and industry. by the development of robust software frameworks for the automation
We are building on these successful, albeit nascent efforts in several of simulations [3–5]. By machine learning large materials datasets, we
ways. First, we are evaluating alternative approaches to XAS compu- have developed models that allow us to better predict novel materials
tations, such as the GW/Bethe-Saltpeter approach [95,96] implemented and rapidly compute their properties, to access time/length scales and
in the OCEAN package [97,98], which may yield more accurate spectra, chemical spaces beyond the current capabilities of DFT, and to out-
especially for L edges. Beyond computed data, we are developing a data perform humans in the interpretation of characterization. It is our hope
platform for the automatic collection and curation of experimental that further developments in these areas will fuel an acceleration in the
spectra through collaborations with the Advanced Light Source at study and discovery of materials for various technological applications.

148
S.P. Ong Computational Materials Science 161 (2019) 143–150

(NSSEFF) under Office of Naval Research (ONR) Grant No. N00014-


15-1-0030 (doping in Mo and MoSi2);
Identify
Absorption Database • NSF Ceramics under Grant No. 1411192 (phosphors for solid-state
lighting);
Species • ONR Young Investigator Program (YIP) under Award N00014-16-1-
2621 (interfaces in all-solid-state batteries and SNAP models for Mo,
fcc metals and Ni-Mo alloys);
• NSF Data Infrastructure Building Blocks (DIBBS) under Award
Number 1640899 (X-ray absorption spectroscopy database and
matching algorithms).

Computation resources were provided by the Triton Shared


Computing Cluster (TSCC) at the University of California, San Diego,
the National Energy Research Scientific Computing Center (NERSC),
and the Extreme Science and Engineering Discovery Environment
(XSEDE) supported by National Science Foundation under Grant No.
Learner 1 Learner N ACI-1053575.
Peak Shifting / Peak Shifting /
Alignment Alignment References
Spectra Norm. Spectra Norm.
(Optional) (Optional) [1] W. Kohn, L.J. Sham, Phys. Rev. 140 (1965) A1133.
[2] P. Hohenberg, W. Kohn, Phys. Rev. 136 (1964) B864.
[3] S.P. Ong, W.D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter,
Feature Trans. Feature Trans. V.L. Chevrier, K.A. Persson, G. Ceder, Comput. Mater. Sci. 68 (2013) 314–319.
[4] A. Jain, S.P. Ong, W. Chen, B. Medasani, X. Qu, M. Kocher, M. Brafman, G. Petretto,
G.-M. Rignanese, G. Hautier, D. Gunter, K.A. Persson, Concurr. Comput. Pract. Exp.
Intensity Norm. Intensity Norm. 27 (2015) 5037–5059.
[5] K. Mathew, et al., Comput. Mater. Sci. 139 (2017) 140–152.
Similarity Similarity [6] G. Hautier, A. Jain, S.P. Ong, B. Kang, C. Moore, R. Doe, G. Ceder, Chem. Mater. 23
Measure Measure (2011) 3495–3508.
[7] G. Hautier, A. Jain, H. Chen, C. Moore, S.P. Ong, G. Ceder, J. Mater. Chem. 21
(2011) 17147–17153.
[8] S.P. Ong, Y. Mo, W.D. Richards, L. Miara, H.S. Lee, G. Ceder, Energy Environ. Sci. 6
Rank 1 Rank N (2013) 148.
[9] J. Greeley, T.F. Jaramillo, J. Bonde, I.B. Chorkendorff, J.K. Nørskov, Nat. Mater. 5
(2006) 909–913.
[10] K. Yang, W. Setyawan, S. Wang, M. Buongiorno Nardelli, S. Curtarolo, Nat. Mater.
Combined 11 (2012) 614–619.
Rank [11] J. Hachmann, R. Olivares-Amaya, S. Atahan-Evrenk, C. Amador-Bedolla,
R.S. Sanchez-Carrera, A. Gold-Parker, L. Vogt, A.M. Brockway, A. Aspuru-Guzik, J.
Phys. Chem. Lett. 2 (2011) 2241–2251.
Prob. Each [12] A. Jain, S.P. Ong, G. Hautier, W. Chen, W.D. Richards, S. Dacek, S. Cholia,
D. Gunter, D. Skinner, G. Ceder, K.A. Persson, APL Mater. 1 (2013) 011002.
Spectrum [13] J.E. Saal, S. Kirklin, M. Aykol, B. Meredig, C. Wolverton, JOM 65 (2013)
1501–1509.
Fig. 6. The Ensemble-Learned Spectra IdEntification (ELSIE) algorithm. In the [14] S. Curtarolo, W. Setyawan, S. Wang, J. Xue, K. Yang, R.H. Taylor, L.J. Nelson,
first step, the absorption species is identified and used to narrow down the G.L. Hart, S. Sanvito, M. Buongiorno-Nardelli, N. Mingo, O. Levy, Comput. Mater.
candidate computed reference spectra. In the second step, the spectral matching Sci. 58 (2012) 227–235.
ensemble yields a rank-ordered list of computational spectra according to si- [15] D. Silver, et al., Nature 550 (2017) 354–359.
[16] E.M. Christiansen, et al., Cell 173 (2018) 792–803 e19.
milarity with respect to the target spectrum. Reproduced with permission from
[17] J.P. Perdew, K. Burke, M. Ernzerhof, Phys. Rev. Lett. 77 (1996) 3865–3868.
Ref. [25]. [18] K. Choudhary, I. Kalish, R. Beams, F. Tavazza, Sci. Rep. 7 (2017) 1–16.
[19] M. Aykol, S. Kim, C. Wolverton, J. Phys. Chem. C 119 (2015) 19053–19058.
[20] K. Choudhary, G. Cheon, E. Reed, F. Tavazza, Phys. Rev. B 98 (2018) 1–12.
CRediT authorship contribution statement [21] A. Lozano, B. Escribano, E. Akhmatskaya, J. Carrasco, Phys. Chem. Chem. Phys. 19
(2017) 10133–10139.
Shyue Ping Ong: Conceptualization, Writing - review & editing. [22] J. Sun, R.C. Remsing, Y. Zhang, Z. Sun, A. Ruzsinszky, H. Peng, Z. Yang, A. Paul,
U. Waghmare, X. Wu, M.L. Klein, J.P. Perdew, Nat. Chem. 8 (2016) 831–836.
[23] R. Tran, Z. Xu, B. Radhakrishnan, D. Winston, W. Sun, K.A. Persson, S.P. Ong, Sci.
Acknowledgement Data 3 (2016) 160080.
[24] K. Mathew, C. Zheng, D. Winston, C. Chen, A. Dozier, J.J. Rehr, S.P. Ong,
K.A. Persson, Sci. Data 5 (2018) 180151.
The author acknowledges generous support from the following [25] C. Zheng, K. Mathew, C. Chen, Y. Chen, H. Tang, A. Dozier, J.J. Kas, F.D. Vila,
sources (listed in the order in which the work was discussed in this J.J. Rehr, L.F.J. Piper, K.A. Persson, S.P. Ong, NPJ Comput. Mater. 4 (2018) 12.
review): [26] Z. Deng, Z. Wang, I.-H. Chu, J. Luo, S.P. Ong, J. Electrochem. Soc. 163 (2016)
A67–A74.

• NorthEast Center for Chemical Energy Storage (NECCES), an Energy


[27] B. Radhakrishnan, S.P. Ong, Front. Energy Res. 4 (2016).
[28] K. Takada, Acta Mater. 61 (2013) 759–770.
Frontier Research Center funded by the U.S. Department of Energy [29] P. Knauth, Solid State Ionics 180 (2009) 911–916.
(DOE), Office of Science, Basic Energy Sciences under Award # [30] N. Kamaya, K. Homma, Y. Yamakawa, M. Hirayama, R. Kanno, M. Yonemura,
T. Kamiyama, Y. Kato, S. Hama, K. Kawamoto, A. Mitsui, Nat. Mater. 10 (2011)
DESC0012583 (multi-electron cathodes); 682–686.
• National Science Foundation (NSF) Designing Materials to [31] A. Hayashi, K. Noi, A. Sakuda, M. Tatsumisago, Nat. Commun. 3 (2012) 856.
[32] Y. Wang, W.D. Richards, S.P. Ong, L.J. Miara, J.C. Kim, Y. Mo, G. Ceder, Nat. Mater.
Revolutionize and Engineer our Future (DMREF) program under
14 (2015) 1026–1031.
Grant No. 1436976 (solid-state electrolytes);

[33] Z. Deng, B. Radhakrishnan, S.P. Ong, Chem. Mater. 27 (2015) 3749–3755.
U.S. DOE, Office of Science, Basic Energy Sciences under Award [34] Z. Zhu, I.-H. Chu, Z. Deng, S.P. Ong, Chem. Mater. 27 (2015) 8318–8325.
DESC0012118 (anti-perovskites); [35] I.-H. Chu, C.S. Kompella, H. Nguyen, Z. Zhu, S. Hy, Z. Deng, Y.S. Meng, S.P. Ong,

• National Security Science and Engineering Faculty Fellowship


Sci. Rep. 6 (2016) 33733.
[36] Z. Zhu, I.-H. Chu, S.P. Ong, Chem. Mater. 29 (2017) 2474–2484.

149
S.P. Ong Computational Materials Science 161 (2019) 143–150

[37] I.-H. Chu, H. Nguyen, S. Hy, Y.-C. Lin, Z. Wang, Z. Xu, Z. Deng, Y.S. Meng, S.P. Ong, [68] Z. Yao, J. Stiglich, T. Sudarshan, J. Mater. Eng. Perform. 8 (1999) 291–304.
ACS Appl. Mater. Interfaces 8 (2016) 7843–7853. [69] T.C. Chou, T.G. Nieh, JOM 45 (1993) 15–21.
[38] Z. Deng, Z. Zhu, I.-H. Chu, S.P. Ong, Chem. Mater. 29 (2017) 281–288. [70] J.A. Lemberg, R.O. Ritchie, Adv. Mater. 24 (2012) 3445–3480.
[39] H. Nguyen, S. Hy, E. Wu, Z. Deng, M. Samiee, T. Yersak, J. Luo, S.P. Ong, Y.S. Meng, [71] M.K. Miller, E.A. Kenik, M.S. Mousa, K.F. Russell, A.J. Bryhan, Scripta Mater. 46
J. Electrochem. Soc. 163 (2016) A2165–A2171. (2002) 299–303.
[40] M. Samiee, B. Radhakrishnan, Z. Rice, Z. Deng, Y.S. Meng, S.P. Ong, J. Luo, J. [72] M.K. Miller, A.J. Bryhan, Mater. Sci. Eng. A 327 (2002) 80–83.
Power Sources 347 (2017) 229–237. [73] H. Saage, M. Kruger, D. Sturm, M. Heilmaier, J.H. Schneibel, E. George,
[41] H. Tang, Z. Deng, Z. Lin, Z. Wang, I.-H. Chu, C. Chen, Z. Zhu, C. Zheng, S.P. Ong, L. Heatherly, C. Somsen, G. Eggeler, Y. Yang, Acta Mater. 57 (2009) 3895–3901.
Chem. Mater. 30 (2018) 163–173. [74] H. Zheng, R. Tran, X.-G. Li, B. Radhakrishnan, S.P. Ong, Acta Mater. 145 (2018)
[42] Z. Zhu, Z. Deng, I.-h. Chu, B. Radhakrishnan, S.P. Ong, Computational Materials 470–476.
System Design; Springer International Publishing, Cham, 2018 pp 147–168. [75] A. Thompson, L. Swiler, C. Trott, S. Foiles, G. Tucker, J. Comput. Phys. 285 (2015)
[43] H. Jónsson, G. Mills, K.W. Jacobsen, 1998. 316–330.
[44] Y. Zhao, L.L. Daemen, J. Am. Chem. Soc. 134 (2012) 15042–15047. [76] M.A. Wood, A.P. Thompson, 2017, pp. 1–7.
[45] M.S. Whittingham, Chem. Rev. 114 (2014) 11414–11443. [77] X.-G. Li, C. Hu, C. Chen, Z. Deng, J. Luo, S.P. Ong, Phys. Rev. B 98 (2018) 094104.
[46] M.S. Whittingham, C. Siu, J. Ding, Acc. Chem. Res. 51 (2018) 258–264. [78] C. Chen, Z. Deng, R. Tran, H. Tang, I.-H. Chu, S.P. Ong, Phys. Rev. Mater. 1 (2017)
[47] A.K. Padhi, K.S. Nanjundaswamy, C. Masquelier, S. Okada, J.B. Goodenough, J. 043603.
Electrochem. Soc. 144 (1997) 1609. [79] A.P. Bartók, R. Kondor, G. Csányi, Phys. Rev. B 87 (2013) 184115.
[48] Y. Lin, et al., Chem. Mater. 28 (2016) 1794–1805. [80] Veidt, 2018. &lt;https://github.com/materialsvirtuallab/veidt&gt;.
[49] Y.-C. Lin, M.F.V. Hidalgo, I.-H. Chu, N.A. Chernova, M.S. Whittingham, S.P. Ong, J. [81] R.E.W. Casselton, 1964, 7, 212–221.
Mater. Chem. A 5 (2017) 17421–17431. [82] S.L. Chen, F. Zhang, F.Y. Xie, S. Daniel, X.Y. Yan, Y.A. Chang, R. Schmid-Fetzer,
[50] L.W. Wangoh, et al., Appl. Phys. Lett. 109 (2016) 053904. W.A. Oates, JOM 55 (2003) 48–51.
[51] B. Wen, Q. Wang, Y. Lin, N.A. Chernova, K. Karki, Y. Chung, F. Omenya, S. Sallis, [83] A. Jain, G. Hautier, S.P. Ong, C.J. Moore, C.C. Fischer, K.A. Persson, G. Ceder, Phys.
L.F.J. Piper, S.P. Ong, M.S. Whittingham, Chem. Mater. 28 (2016) 3159–3170. Rev. B 84 (2011) 045115.
[52] J. Ding, Y.-c. Lin, J. Liu, J. Rana, H. Zhang, H. Zhou, I.-h. Chu, K.M. Wiaderek, [84] G. Hautier, S.P. Ong, A. Jain, C.J. Moore, G. Ceder, Phys. Rev. B 85 (2012) 155208.
F. Omenya, N.A. Chernova, K.W. Chapman, L.F.J. Piper, S.P. Ong, [85] J. Schmidt, J. Shi, P. Borlido, L. Chen, S. Botti, M.A. Marques, Chem. Mater. 29
M.S. Whittingham, Adv. Energy Mater. 1800221 (2018) 1800221. (2017) 5090–5103.
[53] P. Pust, P.J. Schmidt, W. Schnick, Nat. Mater. 14 (2015) 454–458. [86] B. Meredig, A. Agrawal, S. Kirklin, J.E. Saal, J.W. Doak, A. Thompson, K. Zhang,
[54] J. Brodrick, Solid-State Lighting R&amp;D Plan (2016). A. Choudhary, C. Wolverton, Phys. Rev. B 89 (2014) 094104.
[55] Z. Wang, J. Ha, Y.H. Kim, W.B. Im, J. McKittrick, S.P. Ong, Joule 2 (2018) 914–926. [87] F.A. Faber, A. Lindmaa, O.A. Von Lilienfeld, R. Armiento, Phys. Rev. Lett. 117
[56] G. Hautier, C.C. Fischer, A. Jain, T. Mueller, G. Ceder, Chem. Mater. (2010). (2016) 135502.
[57] G. Bergerhoff, R. Hundt, R. Sievers, I.D. Brown, J. Chem. Inform. Comput. Sci. 23 [88] O. Isayev, C. Oses, C. Toher, E. Gossett, S. Curtarolo, A. Tropsha, Nat. Commun. 8
(1983) 66–69. (2017) 15679.
[58] Z. Wang, I.-H. Chu, F. Zhou, S.P. Ong, Chem. Mater. 28 (2016) 4024–4031. [89] W. Li, R. Jacobs, D. Morgan, Comput. Mater. Sci. 150 (2018) 454–463.
[59] Z. Wang, W. Ye, I.-H. Chu, S.P. Ong, Chem. Mater. 28 (2016) 8622–8630. [90] Y. LeCun, Y. Bengio, G. Hinton, Nature 521 (2015) 436.
[60] C. Cozzan, G. Laurita, M.W. Gaultois, M. Cohen, A.A. Mikhailovsky, [91] W. Ye, C. Chen, Z. Wang, I.-H. Chu, S.P. Ong, Nat. Commun. 9 (2018) 3800.
M. Balasubramanian, R. Seshadri, J. Mater. Chem. C 5 (2017) 10039–10046. [92] L. Wang, T. Maxisch, G. Ceder, Phys. Rev. B 73 (2006) 195107.
[61] N.S. Rasor, J.D. Mcclelland, J. Phys. Chem. Solids 15 (1960) 17–26. [93] J.J. Rehr, Rev. Mod. Phys. 72 (2000) 621–654.
[62] G. Liu, G.J. Zhang, F. Jiang, X.D. Ding, Y.J. Sun, J. Sun, E. Ma, Nat. Mater. 12 [94] J.J. Rehr, J.J. Kas, M.P. Prange, A.P. Sorini, Y. Takimoto, F. Vila, C.R. Phys. 10
(2013) 344–350. (2009) 548–559.
[63] T. Watanabe, S. Tsurekawa, Mater. Sci. Eng. A (2004) 447–455. [95] S. Lany, Phys. Rev. B Condens. Matter Mater. Phys. 87 (2013) 1–9.
[64] R. Tran, Z. Xu, N. Zhou, B. Radhakrishnan, J. Luo, S.P. Ong, Acta Mater. 117 (2016) [96] K. Choudhary, Q. Zhang, A.C. Reid, S. Chowdhury, N. Van Nguyen, Z. Trautt,
91–99. M.W. Newrock, F.Y. Congo, F. Tavazza, Sci. Data 5 (2018) 1–12.
[65] D. McLean, Grain Boundaries in Metals, Oxford University Press, New York, NY, [97] K. Gilmore, J. Vinson, E.L. Shirley, D. Prendergast, C.D. Pemmaraju, J.J. Kas,
USA, 1957 pp. 116. F.D. Vila, J.J. Rehr, Comput. Phys. Commun. 197 (2015) 109–117.
[66] A.R. Miedema, Z. Metal. 7 (1978) 455–461. [98] J. Vinson, J.J. Rehr, J.J. Kas, E.L. Shirley, Phys. Rev. B Condens. Matter Mater.
[67] G. Duscher, M.F. Chisholm, U. Alber, M. Rühle, Nat. Mater. 3 (2004) 621–626. Phys. 83 (2011).

150

You might also like