Expanding Materials Embeddings for More Expressive Machine Learning Models

High-dimensional representations of the elements have become common within the field of materials informatics to build useful, structure-agnostic models for the chemistry of materials. (1,2) These representations are often pooled to form composition-based feature vectors to represent materials. Beyond their utility for property prediction, element representations also have applications for defining the chemical similarity of compounds for structure substitution approaches. (3,4) However, the characteristics of elements change when they adopt a given oxidation state, with distinct structural preferences and physical properties.<br/>Here, we propose SkipSpecies, a method of learning distributed representations of ions, which is an adaptation of SkipAtom, a method for learning distributed representations of atoms. (5) Clustering these learned representations of ionic species in low-dimensional space reproduces expected chemical heuristics, in particular the separation of cations from anions. We show that these representations have enhanced expressive power for property prediction tasks involving inorganic compounds. We expect that ionic representations, necessary for the description of mixed valence and complex magnetic systems, will support more powerful machine learning models for materials.<br/><br/>1. R. E. A. Goodall, A. A. Lee, Nat. Commun. 11, 6280 (2020).<br/>2. A. Y.-T. Wang, S. K. Kauwe, R. J. Murdock, T. D. Sparks, Npj Comput. Mater. 7, 77 (2021).<br/>3. A. Onwuli, A. V. Hegde, K. V. T. Nguyen, K. T. Butler, A. Walsh, Digit. Discov. (2023), doi:10.1039/D3DD00121K.<br/>4. M. Kusaba, C. Liu, R. Yoshida, Comput. Mater. Sci. 211, 111496 (2022).<br/>5. L. M. Antunes, R. Grau-Crespo, K. T. Butler, Npj Comput. Mater. 8, 1–9 (2022).