Coordinate-free representation of crystals for accelerated materials discovery using machine learning.

Machine-learning (ML) is emerging as a useful technique to create very efficient computational models of the relationships between the structure and properties of a material. The ability to uncover such relations is highly dependent on the choice of representation of the material. Many representations rely either on the description of the composition, which fails to capture the uniqueness of the material (i.e., does not differentiate polymorphs) or on the accurate spatial description via atomic coordinates, which are not readily available when screening hypothetical materials. In this talk, we present a crystal symmetry-based coordinate-free representation¹. It describes the structure via the occupied Wyckoff positions, which captures enough spatial information, but coarse-grains the precise placements of the atoms into an enumerable search space. This descriptor allows us to train an ML model with an exceptionally high hit ratio for finding new stable materials. We demonstrate this by showing that our model identifies 1,558 materials below the known convex hull of previously calculated materials from just 5,675 ab-initio calculations.

1.  arXiv:2106.11132v2 [cond-mat.mtrl-sci]