AI Magazine - Spring 2018

positional variation in the shift parameter λ, demonstrating alloying within phases 3, 5, and 6. Incorporating both an alloying-based peak shifting model and the Gibbs phase constraint resulted in basis patterns and phase concentration maps that are physically meaningful, which is emblematic of a general strategy for AI-enhanced scientific discovery — injecting scientific knowledge where possible has trickle-down effects that result in scientifically meaningful solutions.

In addition to the broader implications of our
algorithms for scientific discovery, the solutions in
figures 5 and 6 are also emblematic of broader trends
in materials science. As new technologies are con-
ceived, new materials are needed, and often these
materials are required to simultaneously exhibit a
variety of properties and perform a variety of func-
tions. At JCAP, researchers are pursuing the discovery
of several materials, including photoanodes, which
are materials that must absorb sunlight and harness
its energy to oxidize water into oxygen, freeing pro-
tons and electrons to be utilized in fuel synthesis.
Metal oxides, such as the compositions in the Nb-V-
Mn oxide composition library discussed earlier, are
excellent photoanode candidates because of their
generally good stability under these conditions.
Among the challenges in identifying and tailoring
metal oxides to be effective photoanodes is the gen-
eral difficulty in tuning their optical properties. The
band gap energy of a given material dictates the
range of sunlight that can be utilized by the materi-
al, and although materials with a variety of band
gaps are available in the photovoltaic and light-emit-
ting diode fields, such band gap tuning is quite rare
in metal oxides. Pattern 4 in figure 5 corresponds to
the MnV2O6 crystal structure, and as indicated by the
activation map for this phase, it exists over a range of
compositions where the peaks shift due to alloying.
By combining this data with band gap measure-
ments, figure 6 shows that within this single phase,
the band gap can be tuned from approximately 1. 9 to
2. 1 eV. At the higher band gap range, the materials do
not absorb orange or red light, but lowering the band
gap enables these absorptions and thus increases the
potential efficiency of the material. This discovery of
alloying-based band gap tuning in this crystal struc-
ture is one component of the much broader portfo-
lio of materials science needed to design and create
photoanode materials. More generally, the discovery
of a material for new technology is typically the cul-
mination of a suite of smaller discoveries, and with
AI algorithms such as those provided by Phase-Map-
per, these discoveries are being accelerated and com-
piled to create a more comprehensive understanding
of the underlying science, thus changing the arc of
scientific discovery.

Conclusion
In this article, we show that the combination of
high-throughput experimentation, AI problem solv-
ing, and human intelligence can yield rich scientific
discoveries, with an application in materials science.
A major, critically missing component of the high-
SPRING 2018 23

Figure 5. Solutions for the Phase Map of 317 XRD Patterns in the Nb-V-Mn Oxide Composition Space.

The 317 XRD patterns measured in the Nb-V-Mn oxide composition space were analyzed to produce phase-mapping solutions. On top, the six basis patterns obtained using AgileFD (blue) are shown along with the peak pattern (red sticks) for the phases identified by materials scientists. For pattern 2, this phase could be identified only after applying the Gibbs phase constraint, which, even though applied only to the basis pattern activations, results in the procurement of more meaningful (phase-pure) basis patterns. The middle row shows the six basis patterns obtained with AgileFD using Gibbs phase constraints (black), with the bottom row showing the composition map of activations of each basis pattern. There is a point for each of the 317 composition samples, with point size corresponding to the phase concentration and color corresponding to the alloying-based peak shifting. Nonshifted (nonalloyed) samples show as black.