Discovery of superionic conductors by ensemble-scope descriptor

Abstract Machine learning accelerates virtual screening in which material candidates are selected from existing databases, facilitating materials discovery in a broad chemical search space. Machine learning models quickly predict a target property from explanatory material features called descriptors. However, a major bottleneck of the machine learning model is an insufficient amount of training data in materials science, especially data with non-equilibrium properties. Here, we develop an alternative virtual-screening process via ensemble-based machine learning with one handcrafted and two generic descriptors to maximize the inference ability even using a small training dataset. A joint representation with the three descriptors translates the physical and chemical properties of a material as well as its underlying short- and long-range atomic structures to describe a multifaceted perspective of the material. As an application, the ensemble-scope descriptor learning model was trained with only 29 entries in the training dataset, and it selected potential oxygen-ion conductors from 13,384 oxides in the inorganic crystal structure database. The experiments confirmed that we successfully discovered five compounds that have not been reported, to the best of our knowledge, as oxygen-ion conductors.