An active-learning framework for the discovery of new crystalline materials

A challenging problem in materials science is that the space of possible crystal structures is so vast that it is impossible to sample with brute-force screening approaches. However, electronic structure methods combined with machine-learning (ML) techniques have a huge potential to speed up the search [1][2], which hold a great promise for the discovery of novel materials and/or catalysts for energy applications.
Here, we present a ML based active-learning framework to search for stable and meta-stable inorganic materials. It has already been applied to the discovery of new stable polymorphs of IrO, using a search space of experimentally observed crystal prototypes with element substitutions. We extend this search to hypothetical crystal structures that are generated with a crystal prototype enumeration scheme combined with a ML aided search for appropriate lattice parameters. Using a graph-theory based distinction, we identify a finite number of geometries to span a highly diverse material sub-space, which is used for the active-learning exploration.

[1] J. Noh et al. Matter (2019)
[2] L. Ward et al. Physical Review B 96.2 (2017): 024104.