Automated interpretation of characterization data

Analyzing characterization data is often a challenging and time-consuming process that requires domain expertise. For instance, identifying crystalline materials from X-ray diffraction (XRD) patterns is a common task for materials scientists, but remains difficult when dealing with multi-phase mixtures that are highly prevalent in exploratory syntheses. To automate this approach, we have developed a machine learning algorithm that uses convolutional neural networks to detect crystalline materials from XRD patterns. It can do so with high accuracy while also providing a measure of prediction confidence to alert the user when it is uncertain about a given phase. These results were recently published in Chemistry of Materials.

We have also extended this machine learning algorithm to actively guide experimental XRD measurements as they are performed. By analyzing the data as it is collected, our approach determines which features in the XRD pattern are most useful for accurate phase identification. It uses this information to steer the diffractometer so it spends more time scanning the most important features. This work was recently published in npj Computational Materials.

Optimization of solid-state synthesis procedures

To aid in the synthesis of novel inorganic materials, we have developed an active learning algorithm that can optimize the selection of precursors used on solid-state synthesis experiments. This algorithm combines ab-initio computed formation energies with insights gained from experimental synthesis outcomes to construct a database of known reactions, from which new experiments are proposed. We recently used this algorithm to find improved synthesis procedures for three materials previously reported in the literature, achieving higher yields with less time. These findings are showcased in our recent Nature Communications publication.

Applications in energy storage

We are currently using our understanding of materials synthesis to improve upon the traditional approaches to lithium refinement. This element is critical for the increased adoption of clean energy, being used in the synthesis of lithium-ion batteries. Yet it remains costly and energy-intensive to refine lithium from its available ores. By leveraging computational techniques to model reaction pathways, we are designing and implementing safe and low-cost routes to extract lithium into a form suitable for battery production. Early progress toward these goals is demonstrated in our recent Inorganic Chemistry paper.