Accelerating Materials Development via Automation, Machine Learning, and High-Performance Computing

Abstract

Successful materials innovations can transform society. However, materials research often involves long timelines and low success probabilities, dissuading investors who have expectations of shorter times from bench to business. A combination of emergent technologies could accelerate the pace of novel materials development by 10x or more, aligning the timelines of stakeholders (investors and researchers), markets, and the environment, while increasing return-on-investment. First, tool automation enables rapid experimental testing of candidate materials. Second, high-throughput computing (HPC) concentrates experimental bandwidth on promising compounds by predicting and inferring bulk, interface, and defect-related properties. Third, machine learning connects the former two, where experimental outputs automatically refine theory and help define next experiments. We describe state-of-the-art attempts to realize this vision and identify resource gaps. We posit that over the coming decade, this combination of tools will transform the way we perform materials research. There are considerable first-mover advantages at stake, especially for grand challenges in energy and related fields, including computing, healthcare, urbanization, water, food, and the environment.

Full text

1
Accelerating Materials Development via Automation, Machine Learning, and High-
Performance Computing
Juan Pablo Correa-Baena1, Kedar Hippalgaonkar2, Jeroen van Duren3, Shaffiq Jaffer4, Vijay R.
Chandrasekhar5, Vladan Stevanovic6, Cyrus Wadia7, Supratik Guha8, Tonio Buonassisi1*
1Massachusetts Institute of Technology, Cambridge, MA 02139, USA
2Institute of Materials Research and Engineering (IMRE), A*STAR (Agency for Science, Technology
and Research), Innovis, Singapore
3Intermolecular Inc., San Jose, CA 95134, USA
4TOTAL American Services, Inc., 82 South Street, Hopkington, MA 01748, USA
5Institute for Infocomm Research (I2R), A*STAR (Agency for Science, Technology and Research), #21-
01 Connexis (South Tower), Singapore
6Colorado School of Mines, Golden, CO 80401, USA
7Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA
8Center for Nanoscale Materials, Argonne National Laboratory, Argonne, IL 60439, USA
*Corresponding author: Tonio Buonassisi, buonassisi@mac.com
Successful materials innovations can transform society. However, materials research often
involves long timelines and low success probabilities, dissuading investors who have
expectations of shorter times from bench to business. A combination of emergent
technologies could accelerate the pace of novel materials development by 10x or more,
aligning the timelines of stakeholders (investors and researchers), markets, and the
environment, while increasing return-on-investment. First, tool automation enables rapid
experimental testing of candidate materials. Second, high-throughput computing (HPC)
concentrates experimental bandwidth on promising compounds by predicting and inferring

2
bulk, interface, and defect-related properties. Third, machine learning connects the former
two, where experimental outputs automatically refine theory and help define next
experiments. We describe state-of-the-art attempts to realize this vision and identify resource
gaps. We posit that over the coming decade, this combination of tools will transform the way
we perform materials research. There are considerable first-mover advantages at stake,
especially for grand challenges in energy and related fields, including computing, healthcare,
urbanization, water, food, and the environment.

3
The development of novel materials has long been stymied by a mismatch of time constants
(Figure 1). Materials development typically occurs over a 15–25-year time horizon, sometimes
requiring synthesis and characterization of millions of samples. However, corporate and
government funders desire tangible results within the residency time of their leadership, typically
2–5 years. The residency time for postdocs and students in a research laboratory is usually 2–5
years; when a project outlasts the residency of a single individual, seamless continuity of
motivation and intellectual property is often the exception, not the rule. Market drivers of novel
materials development, informed by business competition and environmental considerations, often
demand solutions within a shorter time horizon. This mismatch in time constants results in a
historically poor return-on-investment of energy-materials (cleantech) research relative to
comparable investments in medical or software development.1
Figure 1. Timelines for materials discovery and development. Timelines of examples of certain
technologies (blue area), typical academic funding grants (orange), development capacity (green)
and deployment of sustainable energy (i.e., via solar cells) to fulfill the 2030 climate targets.

4
To bridge this mismatch in time horizons and increase the success rate of materials research, both
public- and private-sector actors endeavor to develop new paradigms for materials development.
The U.S. Materials Genome Initiative focused on three “missing links”: computational tools to
focus experimental efforts in the most promising directions, data repositories to aggregate
learnings and identify trends, and higher-throughput experimental tools.2 This call to action was
mirrored in industry and by university- and laboratory-led consortia, many focused on simulation-
based inverse design and discovery and properties databases. As these tools matured, the
throughput of materials prediction often vastly outstripped experimentalists’ ability to screen for
materials with low rates of false negatives.
Today, a new paradigm is emerging for experimental materials research, which promises
to enable more rapid discovery of novel materials.3,4 Figure 2 illustrates one such prototypical
vision, entitled “accelerated materials development and manufacturing.” Rapid, automated
feedback loops are guided by machine learning, and an emphasis on value creation through end-
product and industry transfer. There is a unique opportunity today to develop these capabilities in
testbed fashion, with considerable improvements in research productivity and first-mover
advantages at stake.

5
Figure 2. Schematic of accelerated materials discovery process. The automated feedback loop,
driven by machine learning, drives process improvement. The theory, synthesis, and device
processes take advantage of high-performance computing and materials databases. Icons from Ref.
31.
As is often the case with convergent technologies, one observes significant advances in individual
“silos” before the leveraged ensemble effect bears its full impact. A historical example is three-
dimensional printing, wherein 3D computer-aided design (CAD), computer-to-hardware interface
protocols, and ink-jet printing technologies evolved individually, before being combined by Prof.
Ely Sachs and his MIT team into the first 3D printer. The ability to observe emergent technologies
within individual silos, and assemble them into an ensemble that is greater than the sum of its
parts, mirrors the challenge in novel materials development today. The following paragraphs
describe the discrete, emergent innovations in “siloed” domains that are presently converging, and
promise to enable this paradigm shift within the next decade.
10x faster
material
syn thesis
10x faster
device
assembly
10–100x
faster
diagnostics
Automated Feedback Loop $
Theory Product
Automated Feedback Loop

6
Theory: Today, the rate of theoretical prediction vastly outstrips the rate of experimental synthesis,
characterization, and validation.5 This emergence is enabled by three trends: faster computation,
more efficient and accurate theoretical approaches and simulation tools, and the ability to screen
large databases quickly, such as MaterialsProject.org. To bridge the growing gap between theory
and experiment, researchers are increasingly focusing efforts on predictive materials synthesis
routes, especially synthesis routes that consider environmental factors (e.g., humidity), reaction-
energy barriers, and kinetic limitations (so-called “non-equilibrium” synthesis).17 In parallel,
theorists seek to rationally design materials with combinations of properties — first, by predicting
combinations of properties (e.g., chemical, microstructural, interface, surface…) in one simulation
framework and/or database, then connecting material predictions with device performance &
reliability predictions, then extending this framework to both known and not-yet-discovered
compounds, and ultimately, solving the inverse problem.
High-Throughput Materials, Device, and Systems Synthesis: Historically, slow vacuum-based
deposition methods inhibit materials development. Modern vacuum-based tools, including
combinatorial approaches and large-scale, fast serial deposition/reactions, enable meaningful rate
increases for materials and device synthesis.29,30 Variants of existing deposition methods (e.g.,
close-space sublimation) offer higher growth rates, point-defect control, and precise stoichiometry
and impurity control for process-compatible materials. Solution synthesis has gained acceptance
with the emergence of higher-quality precursors and materials, including CdS quantum dots,
polymer solar cells, and lead-halide perovskites.5,6 The growing diversity of precursors (from
molecular to nanoparticle), synthesis control (including solvent engineering), and thin-film
synthesis methods (lab-based spin-coating to industrially-compatible large-area printing) makes

7
this a powerful and flexible platform to deposit a range of new materials. Emergence of 3D printed
materials provides another ubiquitous alternative. At laboratory scale, throughputs for such rapid
synthesis routes5,7 can be up to an order of magnitude greater than vacuum-based techniques, and
remain to be explored for multinary materials with novel microstructures. With declining
component costs and greater adoption of standards, the ability to rapidly combine discrete devices
into components and systems in a modular and flexible manner is emerging.
Defect Tolerance & Engineering: Often, theoretical predictions are made for “ideal” materials
systems. However, real samples contain defects (impurities, structural defects…), which can harm
(or, occasionally, benefit) bulk and interface properties. To mitigate the risk of defect-induced
false negatives during high-throughput materials screening, it is desirable to identify classes of
materials less adversely affected by defects (so-called “defect tolerant” 8,9), and rapidly diagnose
& decouple the effects of defects on material performance. A notable recent example is the
serendipitous discovery of lead-halide perovskites for optoelectronic applications.6,7 In addition to
being amenable to high-throughput solution-phase deposition, lead-halide perovskites also
required orders of magnitude less research effort to achieve similar performance improvements to
traditional inorganic thin-film materials (Figure 3). It is suspected that part of the facility to
improve performance is owed to increased defect tolerance of lead-halide perovskites, resulting in
improved bulk-transport properties. Determining the underlying physics of and developing design
rules for defect tolerance may inform screening criteria for new materials, especially with new
computational tools such as General Adversarial Networks (GANs) that are state-of-the-art in
anomaly detection.22,23 The next step lies in focusing experimental effort on candidates capable
of rapid performance improvements during early screening and development, and wider process

8
tolerance in manufacturing. In relation to the beneficial aspects of defects and impurities, recent
theory advancements15 in combination with computational tools to rapidly assess and predict
solubility and electrical properties of defects16 allows high-throughput screening of materials for
applications where the desired functionality is enabled by the defects and/or dopants (e.g.,
thermoelectrics, transparent electronics…).
Figure 3. A case study of fast materials development based on photovoltaic applications. a.
certified power conversion efficiency (PCE) over time for CdTe and perovskite solar cells. b.
Number of J-V sweeps measured divided by the increase in percentage point achieved during the
device development of CdTe and perovskite solar cells. Three orders of magnitude fewer J-V
sweeps per percentage efficiency improvement were needed to advance perovskite efficiencies
relative to traditional thin-film solar cell materials. We hypothesize that this difference is partially
due to greater “defect tolerance” of perovskites, enabling a faster and more economical materials
development process.
High-Throughput Diagnosis: Characterization tools have also benefitted from high-throughput
computing, automation, and machine learning. For instance, one high-resolution X-ray
photoelectron spectroscopy spectrum could take an entire day with technology from the 1970’s,
while the same measurement today requires less than an hour. Today, advanced statistics and

9
machine learning promises to further accelerate the rate of learning. Tools now exist that can
acquire multiple XPS spectra on a single sample (e.g., with composition gradients), and automated
spectral analysis of large datasets is now possible, enabling estimation of unknown materials in a
compositional map. Others seek to replace spectroscopy with rapid non-destructive testing; several
bulk and interface properties can be simultaneously diagnosed by using Bayesian inference in
combination with non-destructive device testing, enabling ≥10x faster (and in certain cases, more
precise) diagnosis vis a vis traditional characterization tools.10 This kind of parameter estimation
can be applied to finished components, devices, and systems, and has the potential to not only
enable faster troubleshooting, but also to accurately estimate ultimate performance potential, thus
informing the decision to pursue or abandon further investment in a given candidate material even
at early stages of materials screening.
Machine Learning comprises a broad class of approaches, which may play several different roles
in the future materials-development cycle. First, a common application of machine learning is for
materials selection, in which historical experimental observations are used to inform predictions
of future properties (attributes) of unknown compounds, or discover new ones.24 Such an approach
has been realized to help discover novel active layers in organic solar cells11 and light-emitting
diodes12, and metal alloys13,27, among many others.28 Second, machine learning tools can help
extract greater and more accurate information from diagnosis, as detailed in the previous section.
Third, machine learning tools may help close the automation loop between diagnosis and synthesis,
shown in Figure 2, by reducing the degree of human intervention and reliance on heuristics. For
example, when relationships between experimental inputs and diagnosis outputs can be inferred
by neural networks, detailed process and device models may no longer be needed to predict

10
outcomes and optimize processes. All three applications of machine learning to the materials
development cycle benefit from the availability of more data, to train and sharpen the predictive
capacity of such tools.
Achieving predictability without losing physical insights is an emergent challenge and
research opportunity. Such methods may also increase learning from diagnosis, by consolidating
research output in singular databases, drawing automated inferences from the data, and in the
future perhaps aggregating the experience and knowledge base via natural language processing of
existing research papers and materials property databases.
Envisioning the “Hardware Cloud”: Materials synthesis equipment today is becoming
increasingly remotely operable—enabling research and operation by an investigator who is not in
proximal presence to the deposition equipment. This opens up two related opportunities with far-
reaching consequences. Large, expensive, synthesis equipment can be grouped together with
massively parallel characterization equipment to form synthesis centers of the future, which are
operated by remote users and researchers and managed by an on-site professional staff. Akin in
concept to the Software Cloud concept, where one’s computing and data is stored across machines
worldwide in a seamless manner, a Hardware Cloud would enable a user to deposit, measure and
carry out research (with real time feedback through in-situ characterization tools) across a number
of networked materials processing systems distributed nationally or internationally in a seamless
manner. This also leads to the second opportunity: to be able to store, curate, access, process and
diagnose all data gathered in these networked experiments in Public or Private Clouds. (Protocols
and formats for such science data collectives will be discussed in the following paragraphs.) This
will greatly facilitate two emerging issues: (a) increasing the efficient availability of data across a

11
wide number of experiments and experimental platforms for post-analysis; and (b) making
available for analysis data that indicates “what did not work” — this is not easily available but is
instrumental in the learning process, and has its own value in increasing the collective efficiency
of research progress.
Infrastructure Investments Toward Accelerated Materials Development and
Manufacturing: Realizing the vision shown in Figure 2 requires a sustained commitment over
several years to develop software, hardware, and human resources, and to connect these new
capabilities in testbed fashion.
Investments in applied machine learning: Supported by ample investments into machine-learning
methods development, a pressing challenge is how to down-select and apply the most appropriate
machine-learning methods to enable the “automated feedback loop” shown in Figure 2. Compared
to other widely recognized applications of machine learning today (e.g., vision recognition,
natural-language processing, and board gaming), materials research often involves sparse data sets
(e.g., small sample sizes and number of experimental inputs & outputs, for training and fitting)
and less well-constrained “rules” (e.g., complex physics and chemistry, non-binary inputs and
outputs, large experimental errors, uncontrolled input variables, and incomplete characterization
of outputs, to name a few). These realities make the typical materials-science problem (e.g., layer-
by-layer atomic assembly of a thin film) decidedly more complex and less well defined than a
match of “Go,” where the rules and playing board are constrained. Deep machine learning (DML)
appears well poised to address this complexity. Computation speed can be improved by developing

12
“pre-trained” neural networks that incorporate the underlying physics and chemistry common to
materials synthesis, performance, and defects, bringing DML within reach of commonly available
hardware and software.
A balance must be found between achieving actionable results and inferring physical
insight from “black-box” computational methods, to advance both engineering and scientific
objectives, and minimize unintended consequences. There is a need to apply “white box” (i.e.,
opposite of black box) machine learning methods to materials science problems. One possible
approach may be application of semi-supervised deep learning algorithms, which learn with lots
of unlabeled data and very little labeled data.25
Lastly, the ability of machine-learning tools to adapt to uncontrolled and changing
experimental conditions is essential. Promising developments include online deep learning, which
builds neural networks on the fly, gradually adding neurons (e.g., as baseline experimental
conditions change, or as new physics becomes dominant).26
Investment in standards governing data formatting and storage would facilitate data entry into
machine-learning software. Standards embed contextual know-how, hierarchy, rational thought.
Some communities have implemented standards governing raw and processed data, e.g.,
crystallography, genetics, and geography. However, in most materials-research communities, there
are no universally accepted and implemented data standards. Several materials databases have
been created, often specialized by material class or application, and with varying protocols for
updating information and enforcing hygiene. Furthermore, these databases often lack ability to
quickly & accurately predict device-relevant combinations of properties (e.g., chemical,
mechanical, optoelectronic, microstructural, surface, interface…). Several data standards have

13
been proposed19–21; widespread adoption may hinge on widespread adoption of data-management
systems described in the next paragraph. In the absence of data standards, it is possible that the
burden of data aggregation will shift onto natural language processors18, i.e., computer programs
designed to extract relevant data from available media (e.g., publications, reports, presentations,
and theses).
Investment in data-management tools (e.g., informatics systems) are needed to manage data
obtained from lab equipment and store records, coordinate tasks, and enforce protocols. On one
hand, such systems have been shown to be of high value for well-defined research problems and
tool sets. For early-stage materials research, data management tools require a deft balance between
flexibility and standardization, and the ability to accommodate non-standard workflows, multiple
participants, and equipment spread across multiple sites, including shared-use facilities, in an
elegant and seamless manner. When implemented well, data-management systems can increase
the quality, uniformity, and accessibility of data serving as inputs into machine-learning tools;
when implemented too inflexibly, data-management systems can cause frictions to researcher
workflow and stimulate their resistance. It is possible that, as suggested by Rafael Jaramillo (MIT),
metadata-based distributed data-management systems may warrant strong consideration for early-
stage materials research; a challenge will be, how to capture metadata in an automated, accurate,
thorough, and comprehensive manner.
Investments in infrastructure are needed, to increase throughput of synthesis, device-fabrication,
and diagnosis tools. The potential of automation must be realized, without sacrificing material
quality and offsetting the advantages of higher throughput with an increase in false negatives. The

14
emergence of multi-parameter estimation methodologies, including Bayesian inference and
Design of Experiments (DoE) algorithms, invites the invention new non-destructive diagnostic
apparatus designed to take full advantage of these new methodologies.
There are significant challenges associated with producing and analyzing large quantities
of data. New tools being developed by machine learning specialists invite the possibility of
modifying hardware design to take advantage of machine-learning tools, rather than the other way
around.
Revised policies at institution, funding agency, and government levels may accelerate or
stymie the required ongoing investments at levels large and small, and invites considering how
export control laws, import duties, grant purchasing restrictions, overhead rates, auditing, and
claw-back clauses affect required equipment investments to enable this transformation.
Human-Capital Investments Toward Accelerated Materials Development and
Manufacturing:
Investments in human capital are required to prepare researchers to leverage these new tools. The
transition from being “data-poor” to being “data-rich” invites changes in how we think, how we
incentivize, and how we teach.
How we think: In a “data-poor” world, the time and cost of conducting each experiment is relatively
large, and a risk-adverse mindset is advantageous. In a “data-rich” world, a larger number of
unique experiments can be conducted per unit time, meaning that failure of any given experiment
will have lesser negative impact on a researcher’s milestones and publication record. This will

15
enable researchers to experiment with greater creativity and risk-taking. This has three important
implications for “how we think”: First, a greater premium will be placed on experimental concept
and design, as researchers who design experiments amenable to new tools will be rewarded.
Second, a decreasing cost-per-experiment may result in reduced barriers for junior researchers to
establish themselves, decreasing the premium of initial investment, prompting new as well as
established researchers to explore new fields.
Third, an accelerated materials development framework invites a system-level
perspective14 that mirrors the new tools. Greater experimental throughput suggests that devices
and systems may increasingly be analyzed holistically in lieu of isolated sub-components, test
structures, and proxies. A “data-rich” world will allows us to analyze complex systems more
directly, with lesser need to break into sub-components or impose a priori simplifications even
without complete visibility into each sub-component. Wielding these new computer-based tools
to greatest effect requires that researchers learn to “think” like machine-learning algorithms,
appreciating the nuances and trade-offs of different approaches, requiring a mindset change
providing an opportunity to identify weak links faster, focusing effort on those parameters with
highest returns on investment.
Incentives: Encouraging the mindset change and transitions mentioned in the previous section will
be complemented with a “constant of friction” governed in part by professional incentives of
decades-old institutions. Young researchers will be encouraged to take proactive steps if they are
rewarded by hiring committees, promotion committees, fellowship & awards committees, journal
editors, and conference committees. Funding agencies could encourage open-source development
of equipment that enables integration of high throughput synthesis of materials with data

16
management. Industries may see value in funding solution-driven system-level approaches to
accelerate their development timelines.
Community: Realizing this future requires merging domain expertise currently resident in robotics,
software, computer science, electronics, materials, and design silos, each with their own language
/ acronyms, and academic conferences. The learning curve to become even a generalist in these
different domains remains very steep. Reducing barriers to communication and achieving
percolation of ideas across domains may be facilitated via cross-cutting conferences, workshops,
and creation of funded research centers. Adoption of best practices across various fields can be
encouraged via these percolation pathways of ideas.
Education and Up-Skilling: Public opinion (read: support or opposition) to ML/AI is influenced
by whether or not citizens can envision a hopeful future that includes their employment and
empowers society. First, these transformations require individuals at all levels and employment
types to be willing to up-skill. Educators at all levels have an opportunity to revamp their curricula,
considering both technical and societal impacts. Online tools and courses for machine-learning /
artificial intelligence are growing in availability, but direct applications to materials science and
systems engineering are needed. Second, we are invited to consider how we teach reflects the most
suitable skills and mindsets to harness the full potential of accelerated materials development &
manufacturing platforms. Domain expertise in supporting fields, including advanced statistics, will
increase in utility with the mainstreaming of system-level design of experiments. Third, the
scientific method will still be valid, and the premium will only increase for asking the right
questions, designing good experiments, and disseminating results well.

17
Conclusions
The convergence of high-performance computing, automation, and machine learning promises to
accelerate the rate of materials discovery, better aligning investor and stakeholder timelines. These
new tools are set to become an indispensable part of the scientific process. >10x faster synthesis,
device fabrication, diagnostics in a (semi-)automated feedback loop are distinctly possible in the
near future. Discrete advances in theory, high-throughput materials, device, systems synthesis,
diagnostics, the understanding of defects and defect tolerance, and machine learning are enabling
this transition. There are several infrastructure and human-capital needs to enable this future,
including greater emphasis on appropriate applications of existing methods to materials-relevant
problems, adoption of data and metadata standards, data-management tools, and laboratory
infrastructure, including both decentralized and centralized facilities. To integrate these tools into
the R&D ecosystems depends in part on several human elements — namely, the time needed to
evolve incentive structures, community support, education & up-skilling offerings, and researcher
mindsets, as our field transitions from thinking “data poor” to thinking “data rich.” We envision a
scientific laboratory where the process of materials discovery continues without disruptions, aided
by computational power augmenting the human mind, and freeing the latter to perform research
closer to the speed of imagination, addressing societal challenges in market-relevant timeframes.
Acknowledgements: The ideas represented herein evolved in discussion with numerous
individuals, including but not limited to Riley Brandt, Danny Ren, Felipe Oviedo, Daniil Kitchaev,
Rachel Kurchin, I. Marius Peters, Shijing Sun, Rafael Jaramillo, Gang Chen, and Anantha
Chandrakasan of MIT/SMART; Benjamin Gaddy of Clean Energy Trust; Rolf Stangl, Chaobin

18
He, and Anthony Cheetham of NUS; Dirk Weiss and Raffi Garabedian of First Solar; BJ Stanbery
of Siva Power; Karthik Kumar, Sir John O’Reilly, Pavitra Krishnaswamy, Alfred Huan, and Cedric
Troadec of A*STAR; Andrij Zakutayev, Stephan Lany, Dave Ginley, Greg Wilson, and William
Tumas of NREL; and Lydia Wong, Shuzhou Li, Tim White, and Subbu Venkatraman of NTU,
among many others.
References:
1. B.E. Gaddy, V. Sivaram, T.B. Jones, L. Wayman, “Venture capital and cleantech: The wrong
model for energy innovation,” Energy Policy 102, 385–395 (2017).
2. A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D.
Skinner, G. Ceder and K. A. Persson, “Commentary: The Materials Project: A materials
genome approach to accelerating materials innovation,” APL Materials 1, 011002 (2013).
3. N. Nosengo, “The Material Code: Machine-learning techniques could revolutionize how
materials science is done,” Nature 533, 22 (2016).
4. P. De Luna, J. Wei, Y. Bengio, A. Aspuru- Guzik, E. Sargent, “Use machine learning to find
energy materials,” Nature Materials 552, 23 (2017).
5. E. O. Pyzer-Knapp, C. Suh, R. Gómez-Bombarelli, J. Aguilera-Iparraguirre and A. Aspuru-
Guzik, “What is high-throughput virtual screening? A perspective from organic materials
discovery,” Annual Review of Materials Research 45, 195–216 (2015).
6. A. Kojima, K. Teshima, Y. Shirai and T. Miyasaka, “Organometal halide perovskites as
visible-light sensitizers for photovoltaic cells,” Journal of the American Chemical Society 131,
6050–6051 (2009).

19
7. M. Graetzel, R. A. J. Janssen, D. B. Mitzi and E. H. Sargent, “Materials interface engineering
for solution-processed photovoltaics,” Nature 488, 304–312 (2012).
8. W.-J. Yin, T. Shi and Y. Yan, “Unusual defect physics in CH3NH3PbI3 perovskite solar cell
absorber,” Applied Physics Letters 104, 063903 (2014)
9. R. E. Brandt, V. Stevanović, D. S. Ginley and T. Buonassisi, “Identifying defect-tolerant
semiconductors with high minority-carrier lifetimes: beyond hybrid lead halide perovskites,”
MRS Communications 5, 265–275 (2015).
10. R. E. Brandt, R. C. Kurchin, V. Steinmann, D. Kitchaev, C. Roat, S. Levcenco, G. Ceder, T.
Unold and T. Buonassisi, “Rapid photovoltaic device characterization through Bayesian
parameter estimation,” Joule 1, 843-856 (2017).
11. S. A. Lopez, B. Sanchez-Lengeling, J. de Goes Soares and A. Aspuru-Guzik, “Design
principles and top non-fullerene acceptor candidates for organic photovoltaics,” Joule 1, 857-
870 (2017).
12. R. Gómez-Bombarelli, J. Aguilera-Iparraguirre, T.D. Hirzel, D. Duvenaud, D. Maclaurin,
M.A. Blood-Forsythe, H.S. Chae, M. Einzinger, D.-G. Ha, T. Wu, G. Markopoulos, S. Jeon,
H. Kang, H. Miyazaki, M. Numata, S. Kim, W. Huang, S.I. Hong, M. Baldo, R.P. Adams, A.
Aspuru-Guzik, “Design of efficient molecular organic light-emitting diodes by a high-
throughput virtual screening and experimental approach,” Nature Materials 15, 1120–1127
(2016).
13. B.D. Conduit, N.G. Jones, H.J. Stone, G.J. Conduit, “Design of a nickel-base superalloy using
a neural network,” Materials & Design 131, 358–365 (2017).

20
14. C.R. Cox, J.Z. Lee, D.G. Nocera, T. Buonassisi, “Ten-percent solar-to-fuel conversion with
nonprecious materials,” Proceedings of the National Academy of Sciences 111, 14057–14061
(2014)
15. C. Freysoldt, B. Grabowski, T. Hickel, J. Neugebauer, G. Kresse, A. Janotti, and C. G. Van de
Walle, “First-principles calculations for point defects in solids” Reviews of Modern Physics
86, 253 (2014).
16. A. Goyal, P. Gorai, H. Peng, S. Lany, and V. Stevanovic, “A Computational Framework for
Automation of Point Defect Calculations”, Computational Materials Science 130, 1 (2017).
17. C.L. Phillips and P. Littlewood, “Preface: Special Topic on Materials Genome,” APL Materials
4, 053001 (2016).
18. E. Kim, K. Huang, A. Saunders, A. McCallum, G. Ceder, E. Olivetti, “Materials Synthesis
Insights from Scientific Literature via Text Extraction and Machine Learning,” Chemistry of
Materials 29, 9436–9444 (2017).
19. J. Hill, A. Mannodi-Kanakkithodi, R. Ramprasad, B. Meredig, “Materials Data Infrastructure
and Materials Informatics” in D. Shin, J. Saal (eds) Computational Materials System Design.
Springer (2017). ISBN 978-3-319-68280-8
20. R. Ananthakrishnan, K. Chard, I. Foster, S. Tuecke, “Globus platform-as-a-service for
collaborative science applications,” Concurrency and Computation.: Practice and Experience
27, 290–305 (2015)
21. K. Chard, E. Dart, I. Foster, D. Shifflett, S. Tuecke, J. Williams, “The Modern Research Data
Portal: a design pattern for networked, data-intensive science,” PeerJ Computer Science 4,
e144 (2018).

21
22. H. Zenati, C. Foo, B. Lecouat, G. Manek, V. Chandrasekhar, “Efficient GAN-based Anomaly
Detection,” arxiv
23. G.L. Guimaraes, B. Sanchez-Lengeling, C. Outeiral, P.L.C. Farias, A. Aspuru-Guzik,
“Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation
Models”, arxiv (2017)
24. L. Ward, A. Agrawal, A. Choudhary, C. Wolverton, “A general-purpose machine learning
framework for predicting properties of inorganic materials,” npj Computational Materials 2,
16028 (2016).
25. B. Lecouat, C. Foo, H. Zenati, V. Chandrasekhar, “Semi-supervised Deep Learning with
GANs: Revisiting Manifold Regularization,” arxiv
26. S. Ramasamy, K. Rajaraman, P. Krishnaswamy, V. Chandrasekhar, “Online Deep Learning:
Growing RBM on the fly,” arxiv
27. O.N. Senkov, J.D. Miller, D.B. Miracle, C. Woodward, “Accelerated exploration of multi-
principal element alloys with solid solution phases,” Nature Communications 6, 6529 (2015).
28. M. L. Green, C. L. Choi, J. R. Hattrick-Simpers, A. M. Joshi, I. Takeuchi, S. C. Barron, E.
Campo, T. Chiang, S. Empedocles, J. M. Gregoire, A. G. Kusne, J. Martin, A. Mehta, K.
Persson, Z. Trautt, J. Van Duren, and A. Zakutayev, “Fulfilling the promise of the materials
genome initiative with high-throughput experimental methodologies,” Applied Physics
Reviews 4, 011105 (2017).
29. J. Eid, H. Liang, I. Gereige, S. Lee, J. Van Duren, “Combinatorial study of NaF addition in
CIGSe films for high efficiency solar cells,” Progress in Photovoltaics 23, 269–280 (2015).

22
30. M.K. Jeon, J.S. Cooper, P.J. McGinn," “Investigation of PtCoCr/C catalysts for methanol
electro-oxidation identified by a thin film combinatorial method,” Journal of Power Sources
192, 391–395 (2009).
31. Icons in Figure 2 are freeware, and were made by Freepik from www.flaticon.com