Electronic Configurations of 3d Transition-Metal Compounds Using Local Structure and Neural Networks.

Machine learning (ML) methods extract statistical relationships between inputs and results. When the inputs are solid-state crystal structures, structure-property relationships can be obtained. In this work, we investigate whether a simple neural network is able to learn the 3d orbital occupations for the transition-metal (TM) centers in crystalline inorganic solid-state compounds using only the local structure around the transition-metal centers described by rotationally invariant fingerprints based on spherical harmonics and one-hot elemental encoding. A multilayer neural network trained on density functional theory (DFT) results of about 1800 samples was developed and showed good performance in predicting the TM orbital occupations (for both spin channels). We study in detail how the local structure affects the predictions of the local properties and how they provide physical insights for the design of a future machine learning model for materials chemistry. The proposed ML method is illustrated in practical application by predicting local magnetic moments of the transition-metal atoms in a full set of inorganic structures with large unit cells. Although less accurate compared to the experimental data, the ML results compared well with the DFT results, suggesting the feasibility of electronic property prediction based only on structure input.