Mining Knowledge from Crystal Structures: Oxidation States of Oxygen-Coordinated Metal Atoms in Ionic and Coordination Compounds.

We propose a universal scheme for predicting the oxidation states of metal atoms in ionic and coordination compounds with a small set of structural descriptors, which include the parameters of atomic Voronoi polyhedra. The scheme has been trained and checked with more than 35,000 crystal structures containing more than 90,000 metal atoms in the oxygen environment. The accuracy of the prediction exceeded 95%; we have detected a number of wrong oxidation states and incorrect chemical compositions in the crystallographic databases using this scheme. The scheme is easily extendable to any kind of atomic environment and can be used to search for correlations between geometrical and physical properties of crystal structures.