Database, Features, and Machine Learning Model to Identify Thermally Driven Metal–Insulator Transition Compounds

Abstract

Metal-insulator transition (MIT) compounds are materials that may exhibit\ninsulating or metallic behavior, depending on the physical conditions, and are\nof immense fundamental interest owing to their potential applications in\nemerging microelectronics. There is a dearth of thermally-driven MIT materials,\nhowever, which makes delineating these compounds from those that are\nexclusively insulating or metallic challenging. Here we report a material\ndatabase comprising temperature-controlled MITs (and metals and insulators with\nsimilar chemical composition and stoichiometries to the MIT compounds) from\nhigh quality experimental literature, built through a combination of\nmaterials-domain knowledge and natural language processing. We featurize the\ndataset using compositional, structural, and energetic descriptors, including\ntwo MIT relevant energy scales, an estimated Hubbard interaction and the charge\ntransfer energy, as well as the structure-bond-stress metric referred to as the\nglobal-instability index (GII). We then perform supervised classification,\nconstructing three electronic-state classifiers: metal vs non-metal (M),\ninsulator vs non-insulator (I), and MIT vs non-MIT (T). We identify two\nimportant descriptors that separate metals, insulators, and MIT materials in a\n2D feature space: the average deviation of the covalent radius and the range of\nthe Mendeleev number. We further elaborate on other important features (GII and\nEwald energy), and examine how they affect classification of binary vanadium\nand titanium oxides. We discuss the relationship of these atomic features to\nthe physical interactions underlying MITs in the rare-earth nickelate family.\nLast, we implement an online version of the classifiers, enabling quick\nprobabilistic class predictions by uploading a crystallographic structure file.\n