If two materials share the same atomic species in the same positions, they will exhibit the same phyiscal properties. Extending this idea further, quantifying how similar different materials are would allow us to predict their properties, given enough data to compare. Machine learning (ML) models can be used to make these predictions, but unfortunately crystal structures are not well-suited to existing ML methods.
Here, we have developed a "grouped representation of interatomic distances" (GRID) as a way to feed crystallographic structure to machine learning algorithms. Although simple, when combined with Earth mover's distance this representation has the ability to accurately predict physical properties of materials, and reflects chemical intuition. Importantly, GRID can be applied to both crystalline and disordered materials without modification.
R.-Z..Zhang, S. Seth and J. Cumby, Grouped representation of interatomic distances as a similarity measure for crystal structures, Digital Discovery, 2, 2023, 81-90.