1. AI at Scale
for Materials
(and Chemistry)
IAN FOSTER
Director, Data Science and Learning Division
foster@anl.gov
AI for Science Town Hall, Argonne National Laboratory, July 22, 2019

2. Materials innovation is central to
meeting energy challenges
Crosscuts the energy technology portfolio,
from energy generation and storage to
delivery and end use
Foundation of every clean energy innovation:
advanced batteries, solar cells, low-energy
semiconductors, thermal storage, coatings,
and catalysts for the conversion, capture,
and use of CO2, etc.

3. Space of potential materials is
inordinately large
Key properties such as stability,
synthesizability, and toxicity are
hard to determine
Quantum mechanics cannot be
computed precisely and efficiently
Both experiments and simulations
are expensive
Experiments generate large data
that cannot easily be interpreted
But materials innovation is difficult

4. A grand challenge for science:
Building a Materials Acceleration Platform
4
“[In the clean energy context] perhaps of greatest interest to the
theoretical physics, physical chemistry, and materials science
communities that are working alongside the machine learning,
robotics, and next-generation computing communities is the
challenge of developing clean energy materials. The goal is
to provide an integrated end-to-end materials innovation
approach, or platform, to deliver the mix of solutions … [We
must] spur the interest of the creative community in what is now
the most urgent series of demands facing humanity. The
opportunities are, simply, immense.” – Sir David King*
* David King, “Global clean energy in 2017,” Science 355 (6321), 111, DOI: 10.1126/science.aam7088

5. Six elements of a Materials Acceleration Platform
1) “Self-driving laboratories” that design, perform,
and interpret experiments in an automated way
2) The development of specific forms of AI for
materials discovery
3) Modular materials robotics platforms
that can be assemblies of modular building
blocks for synthesis and characterization
4) Research into computational methods
for inverse design
5) New methodologies for bridging the
length and timescales associated with
materials simulation
6) Sophisticated data infrastructure and
interchange platforms
5Report: Materials Acceleration Platform, 2017 – http://bit.ly/2Z7kp9Q

6. Example opportunities
AI-assisted multi-fidelity scale bridging
– Accelerate numerical simulation by using AI to
learn low-cost models of finer-grained behavior
AI assistant to understand and design
complex measurements
– Cross-modal image learning across scales,
integrated/in-situ AI within experimental
processes, automated anomaly detection
AI-driven materials design
– Integrate data at scales not feasible for human
experts, to construct expert assistants for
material designers—or even fully automated
design pipelines
DOI:10.1021/acs.jpcc.8b09917
DOI: 10.1126/sciadv.aaq1566

7. AI-assisted multi-fidelity scale bridging
“The underlying physical laws necessary for the mathematical theory of a large part
of physics and the whole of chemistry are thus completely known, and the difficulty is
only that the exact application of these laws leads to equations much too complicated
to be soluble. It therefore becomes desirable that approximate practical methods
of applying quantum mechanics should be developed.” -- Paul Dirac, 1929
7
Hierarchy of numerical approximations to Schrödinger’s equation.
https://doi.org/10.1002/qua.24954
~25% of NERSC time is DFT;
another 20% is MD and QC.

8. Ultra-fast simulation of complex systems
Use massive data, AI, and exascale computing to define task-specific
approximations to quantum mechanics
H. Chan et al., 2019, DOI:10.1021/acs.jpcc.8b09917

9. Kinetic
model
Sample
N
Y
Fix model
New sample
Products
Adjust
model
Match?
Species
detected
Species
predicted
Pyrolysis reactor
Credit: Kirill Prozument and AI Task Force
AI-driven understanding of flame spray pyrolyis

10. 10
Human
intuition
Pure or applied
problem
Extant
knowledge
“Design a material
that does X and that is
synthesizable, cheap,
non-toxic, stable,
biodegradable, …”
Materials science and chemistry as search (in a huge space)
Design Synthesis
Character-
ization

11. 11
Expand knowledge: e.g., machine reading
Automated navigation of chemical,
functional, synthesis spaces
Active/reinforcement learning to choose
next experiment or simulation
AI for materials characterization: choose
measurement(s), infer properties
Ren et al. Sci Adv. (2017) eaaq1566
A new era of discovery based on high-throughput, AI-guided search
High-throughput
experiment / simulation
Eli Rotenberg, https://bit.ly/2BJ81Vv

12. Use exascale computing and automated experiments to identify promising candidates
Designing new electrolytes
Trillions of
candidates
Autoencoder to expand and
navigate materials space
Continuously updated
predictive models of
materials properties
Supercomputer
simulations
Laboratory experiments
Active learning
used to direct
simulation and
experiment
Important properties:
redox potential, stability,
toxicity, solubility,
synthesizability, …

13. Materials data: diffuse, sparse, heterogeneous
arXiv:1907.05644

14. 14
Chart Source and Method: https://github.com/blaiszik/ml_publication_charts
Exponential growth in machine learning publications

15. • Data: Capture, organize, analyze,
share move and deliver data
• Compute: Accessible at many
scales, types, and locations
• Models: Discoverable, described,
and portable to wherever data
and/or computer are located
• Workflows: Easily discovered,
adapted, composed, scaled, and
reused
Needed:
A cohesive infrastructure for AI-driven (materials) science
Encompassing:
Models and
Workflows
Data
Computations
&
Experiments
materialsdatafacility.org

16. These are just a few examples of the many
opportunities for AI in materials and chemistry
Progress requires much work in applications, learning systems,
foundations, hardware.
Resources:
Materials Acceleration Platform
workshop http://bit.ly/2Z7kp9Q
Computing Community Consortium
AI Roadmap -- http://bit.ly/2JK8OZ9