| A novel computational approach is developed for the study of chemical crystallography in materials science using the tools of information theory and data science. By integrating the information derived from phase homologies, electronic structure calculations and known crystal structure data, a high-dimensional data space is created. From this data space, we seek to extract statistically robust and yet physically meaningful relationships linking structure with chemistry and property in the form of chemical design rules that identifies the exact role of key structure governing factors without any a priori assumptions. The powerful role of data dimensionality reduction, clustering analysis and data mining methods for both classification and prediction of structure-property relationships in materials is discussed. Using examples from two different crystal chemistry families, apatites and perovskite oxides, it is shown how high-dimensional data can be used to assess patterns of behavior as well as establish predictive Quantitative Structure-Activity Relationships (QSARs). A library of potentially new "virtual" stoichiometric and non-stoichiometric apatites and high TC piezoelectric perovskite materials chemistries were identified, thereby reinforcing the value of statistical learning methods in rational materials design. |
