While significant advances have been made in accelerating chemical discovery with machine learning, the use of these methods for crystalline materials usually requires restraints on the input structure or manually constructed feature vectors. The arbitrary size and periodicity of crystalline systems pose challenges as these systems need to be represented with a fixed dimensionality and retain translational, rotational, and permutation invariance. We present the use of crystallographically inspired transformations that are amenable to deep learning algorithms. By calculating a modified structure factor, a suite of representations are developed that are of fixed size irrelevant of the input dimensionality. The ability for these representations to capture the periodic structural information is demonstrated through classification and regression problems related to crystalline materials. This approach has direct impact in crystal structure prediciton and the development of energy-structure-function maps for autonomous discovery.

Deep learning from crystallographic representations of periodic systems

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.