Search for crystal structures with generative models

Abstract

Recently, prediction of material properties and search for crystal structures with machine learning have been extensively explored. This study addresses the issue of generating crystal structures with generative models. So far, an algorithm called CrystalGAN has been proposed. This algorithm generates crystal structures in a fashion of "A-H-B" (A, B: metal, H: hydrogen) with DiscoGAN, a generative model across different domains. CrystalGAN is a simple algorithm to generate crystal structures. However, on the other hand, since it builds a feature by combining the lattice vectors and the coordinates of hydrogen and metals, it is not sufficient to consider the geometric structure of the crystals. We propose an algorithm to generate crystal structures by representing crystals with graph structures to consider those geometric structures. There are three key ideas of the proposed algorithm: (1) usage of crystal graph as a feature, (2) usage of generative models for graph structures such as GraphGAN, (3) conversion of crystal graph into an interpretable format such as POSCAR.