Predicting inorganic crystal structures using adversarial learning

Abstract

The discovery of new functional materials has been continuously demanded in recent decades and has brought significant advancements across various industries. Among various materials, particularly in the energy-related field that has attracted considerable attention in recent years, inorganic materials play a key role and their diverse properties are determined by the crystal structure of the inorganic materials. In the past, predicting the crystal structure of a material with desired properties involved using chemical intuition or making slight modifications to existing material structures. Subsequently, their properties were evaluated experimentally or computationally. However, a more efficient approach is needed to accelerate the discovery of advanced inorganic crystals with desired properties. Recently, with the development of computational science, various databases of inorganic crystal structures have been accumulated, and research on material exploration using these databases, specifically high-throughput virtual screening (HTVS), is actively conducted. At the same time, the advancement of machine learning has led to the proposal of various machine learning methodologies to aid in material exploration, contributing to the acceleration of inorganic material design. In this paper, we propose two deep generative model-based crystal structure prediction models, trained through adversarial learning, to overcome the limitations of previous material exploration research and increase search efficiency. The first model is a crystal structure prediction model based on conditional generative adversarial networks. We overcome the limitations of existing HTVS methods by utilizing this generative model-based HTVS model to create a new Mg-Mn-O polymorph promising as photoanode that does not exist in the existing database. The second model is a domain transformation model based on the pix2pix model. To reduce the significant computational cost associated with density functional theory (DFT) relaxation, which has been considered a major bottleneck in existing HTVS methods, we propose a new data-driven crystal structure relaxation model.