A Machine Learning Model and Database for The Identification of New Metal-Insulator Transition Compounds

One of the main bottlenecks in the discovery of new thermally-driven metal-insulator transition (MIT) materials is the lack of a clear database of MIT compounds, and their relevant features. We have built a database of MIT and stoichiometrically related materials, and trained a machine learning model designed to classify whether a material is an MIT material or not [1]. Our easily interpretable model allows us to identify new features, such as the Average Deviation of the Covalent Radius and its interplay with the Range Mendeleev Number, as well as others. We also built an online pipeline where one can upload their own structures, and obtain a prediction on whether the material is a metal, an insulator, or an MIT material and tested it on previously identified materials [2].

[1] https://arxiv.org/abs/2010.13306
[2] https://arxiv.org/abs/2004.07365