Place: PSI-k 2015, San Sebastian, Spain
The prediction of the crystal structure of a material from only its chemical composition is one of the key challenges in materials design. We use a cluster analysis of experimentally observed crystal structures and derive structure maps with a predictive power comparable to standard density-functional theory calculations. In particular, we present a three-dimensional structure map for compounds that contain sp-block elements and transition metals in arbitrary composition. The structure map predicts the correct crystal structure with a probability of 86% and has a confidence of 98% that the correct crystal structure is among three predicted crystal structures. The three descriptors that span the structure map are physically intuitive functions of the number of valence electrons, the atomic volume and the electro-negativity of the constituent elements. We demonstrate that the identified descriptors remain valid for off-stoichiometric compounds and that they separate binary and ternary crystal-structure prototypes. We test the structure map against standard density-functional theory calculations for 1:1 sp-d-valent compounds and demonstrate that our three-parameter model has comparable predictive power.