Machine-learning-assisted search for functional materials over extended chemical space

Abstract

Materials discovery is a grand challenge for modern materials science. In particular, inverse materials design is aimed at the accelerated search for materials with human-defined target properties. Unfortunately, this is associated with various obstacles, such as incremental improvements of known compounds, unreported properties of synthesized materials, and chemically plausible “missing compounds.” A machine-learning-based approach using unified compositional–structural representations is proposed to overcome the issues mentioned above. The validity of the proposed method has been approved by searching for functional materials—some previously known phases were “re-discovered.” In addition to well-known superhard compounds, unconventional structures that have never been considered in this context were also presented. Analysis of the generated populations provided insights into the underlying quantitative structure–property relationships. This data-driven approach can be successfully applied to discover materials with arbitrary functionalities given a reliable experimental/computational database for the target property.