Property Prediction of Crystalline Solids from Composition and Crystal Structure

Kernel regression as a data-driven and rigorous nonparametric statistical technique to predict properties of atomic crystals is proposed. A key feature of the proposed approach is the possibility of treating predictors not only as continuous, but also as categorical data. The latter specifically allows the predictive model to capture the discrete nature of crystals with regards to composition (number of atoms in the chemical formula) and spatial configuration (finite number of crystallographic space groups). Another important aspect of using kernel regression is the direct access to its explicit mathematical form, which can be directly embedded in optimal inverse problems to design new crystalline materials with given target properties. The property prediction approach is illustrated by training models to predict electronic properties of 746 binary metal oxides and elastic properties of 1173 crystals. As a first approach to solving the inverse problem, an exhaustive enumeration algorithm is described. (C) 2016 American Institute of Chemical Engineers