Learning atoms from crystal structure

Understanding of chemical elements underlies physical sciences, revealing the fundamental role of elements in the formation and properties of materials. Recent advances offer insight into chemical elements through multi-dimensional descriptors, enabling materials modelling and functional property prediction. The predictive power of such modelling is strongly affected by availability of crystal structure information, often inaccessible in exploratory studies; however, existing elemental descriptors lack direct access to structural insights. In this study, we introduce Local Environment-induced Atomic Features (LEAFs), incorporating structural information into elemental descriptors. By deducing atomic properties from local structural environments, LEAFs offer a valuable insight into composition-structure-property relationships. In the crystal structure of a material, each atomic site can be quantitatively described by similarity to common structural motifs, producing a set of descriptors for atoms in crystal structures. By combining these unique identifiers from the local structures of chemical elements across the experimentally verified inorganic materials, LEAFs formulate a comprehensive set of elemental descriptors. We demonstrate the versatility of LEAFs in addressing critical challenges in materials science, including structure-informed interpretation of property predictions based only on compositional information, quantitative mapping of chemical space in structural terms, and efficient evaluation of elemental substitutions for materials design.