instance generators, and developed several algorithms with theoretical and practical guarantees. We
have also continuously developed tools to share
experimental instance data, results, and solution
visualizations with our collaborators throughout (Le
Bras et al. 2011; Ermon et al. 2012; Le Bras et al. 2014;
Ermon et al. 2015; Xue et al. 2015). Phase-Mapper is
our most successful tool to date in this area: it
removes many of the practical barriers to the use of
previous methods, including better scalability, run-times suitable for interactive use, and ease of access.
Phase-Mapper has been used at the Department of
Energy’s Joint Center for Artificial Photosynthesis
(JCAP) to run hundreds of phase-mapping solutions
in the JCAP materials discovery pipeline. Prior to
Phase-Mapper, the difficulty of interpreting X-ray diffraction data limited JCAP scientists’ ability to take
full advantage of resources to conduct high-throughput experiments. Since the deployment of Phase-Mapper, thousands of X-ray diffraction patterns have
been processed and the results are yielding discovery
of new materials for energy applications. These are
exemplified by the discovery of a new family of metal oxide light absorbers in the previously unsolved
Nb-Mn-V oxide system, which is provided here as a
case study and is an illustrative example of the
importance of encoding physical constraints to
obtain physically meaningful phase diagram solutions. We believe Phase-Mapper will lead to further
developments in high-throughput materials discovery by providing rapid and critical insights into the
phase behavior of new materials.
Phase-Mapper: AI for
Materials Discovery
An experimentation pipeline for rapidly synthesiz-
ing, characterizing, and identifying new materials is
referred to as high-throughput materials discovery or
combinatorial materials discovery. In this pipeline, a
handful of elements are deposited together on a two-
dimensional substrate, so that different locations on
the substrate receive varying proportions of the ele-
ments. This smooth variation in elemental composi-
tion across the substrate gives rise to the forming of
a discrete set of materials, each of which is present in
particular regions of the substrate.
The deposition process is analogous to atomic
spray paint, as mentioned earlier. Imagine red, green,
and blue spray paint being simultaneously sprayed
onto a surface (or wafer) with each color source
placed at the vertex of an equilateral triangle. Near
these vertices, the deposited color appears simply
red, green, or blue, and throughout the area of the
triangle a continuum of the possible colors are
obtained, where each color on the spectrum exists at
a unique point on the wafer. In the same manner, the
deposited materials “library” contains a broad spec-
trum of compositions (given the starting elements),
and the atoms in different composition regions may
arrange in a unique way to form a unique “phase”
whose properties differ from other materials, even
other compositions and phases formed from the
same elements. It is the hope that one of these new
materials will have a composition and phase that
exhibit the desired properties, and to fully under-
stand the composition-phase-property relationships,
the full phase map must be solved.
In the libraries being studied, the new materials are
typically crystalline, meaning that at the atomic scale
atoms are arranged in particular lattice structures,
and the phase noted earlier is described by the sym-
metry and composition of the lattice structure. On a
larger length scale, typically 5 to 500 nm, the lattice
structure may alternate between two or three differ-
ent structures, constituting a mixed-phase material.
Each phase and phase mixture can exhibit unique
properties, creating the need for materials scientists
to understand, for each material library, how to cate-
gorize each material in the library (on the wafer) in
term of its phase mixture
What data should materials scientists look at to
determine the crystal structures? An indirect way of
probing the microscopic structure is through X-ray
diffraction. When X-rays are directed against a crys-
tal, atomic layers will reflect the light; and for specific
angles determined by the spacing between atomic
layers, this reflected light will interfere constructive-
ly, giving rise to a strong signal. Thus, by scanning
through all angles and measuring the reflected light,
materials scientists are able to infer the structure of
Figure 1. The Phase-Mapper Platform.
The Phase-Mapper platform integrates experimentation, AI problem solving, and human feedback into a platform for high-throughput materials discovery.
Phase-Mapper
ctionData Experim
High
Throughput
Experiments
AI
Solver
Human
Intelligence
and Feedback