New crystal structures are frequently derived by performing ionic substitutions on known crystal structures. These derived structures are then used in further experimental analysis, or as the initial guess for structural optimization in electronic structure calculations, both of which usually require a reasonable guess of the lattice parameters. In this work, we propose two lattice prediction schemes to improve the initial guess of a candidate crystal structure. The first scheme relies on a one-to-one mapping of species in the candidate crystal structure to a known crystal structure, while the second scheme relies on data-mined minimum atom pair distances to predict the crystal volume of the candidate crystal structure and does not require a reference structure. We demonstrate that the two schemes can effectively predict the volumes within mean absolute errors (MAE) as low as 3.8% and 8.2%. We also discuss the various factors that may impact the performance of the schemes. Implementations for both schemes are available in the open-source pymatgen software.
BT - Computational Materials Science DA - 04/2018 DO - 10.1016/j.commatsci.2018.01.040 LA - eng N2 -New crystal structures are frequently derived by performing ionic substitutions on known crystal structures. These derived structures are then used in further experimental analysis, or as the initial guess for structural optimization in electronic structure calculations, both of which usually require a reasonable guess of the lattice parameters. In this work, we propose two lattice prediction schemes to improve the initial guess of a candidate crystal structure. The first scheme relies on a one-to-one mapping of species in the candidate crystal structure to a known crystal structure, while the second scheme relies on data-mined minimum atom pair distances to predict the crystal volume of the candidate crystal structure and does not require a reference structure. We demonstrate that the two schemes can effectively predict the volumes within mean absolute errors (MAE) as low as 3.8% and 8.2%. We also discuss the various factors that may impact the performance of the schemes. Implementations for both schemes are available in the open-source pymatgen software.
PY - 2018 SP - 184 EP - 192 ST - Computational Materials Science T2 - Computational Materials Science TI - Predicting the volumes of crystals VL - 146 SN - 09270256 ER -