Inverse design for inorganic solid materials using generative adversarial neural network

Abstract

For decades, ab-initio first principle quantum calculation has contributed to discovering new materials, and due to recent drastic advance in computing calculation power high throughput screening that calculates enormous candidate material data and discovers good functional materials among them is available. This material discovery method, however, relies on chemical intuition and experience and also can discover materials only limited in existing database. In this work, we propose the new material design framework using Generative Adversarial Network (GAN) in machine learning that addresses previous material design method. Also, we developed the inorganic solid material representation that is simply invertible and applied to vanadium oxide system with GAN model. We found entirely new polymorphs for vanadium oxide. 23 structures among them are thermodynamically stable and could be considered synthesizable.