Inverse materials design using predictive and generative machine learning

Abstract

With the recent increase in demand for renewable and sustainable energy resources, material development has become increasingly important. In traditional approach, a candidate material is specified first using intuition or by slightly changing the existing materials. Furthermore, as various material databases have been accumulated with the development of computational science, studies have been proposed to discover new materials by exploring the database directly. In addition, inverse design techniques have been presented to design materials with desired properties by leveraging structure-property relation through machine learning. From this point of view, this dissertation seeks to propose an method for inorganic crystal design utilizing various inverse design techniques. In particular, we shown the machine learning-driven high-throughput screening framework (augmented with the uncertainty quantification) can be effectively used to accelerated materials design. Also, we demonstrate that the generative framework can effectively used to explore entirely new chemical space. Lastly, we proposed an novel generative framework for exploration of synthesizable molecular chemical space.