(777h) Local Pattern Discovery for Uncovering Structure-Property Relationships of Materials

Machine learning applications to data from atomistic material simulations typically focus on the inference of a global prediction model for some physical or chemical property of interest, such as activation barriers or binding energies.  Due to their global prediction-based objective, these models are not well suited for the discovery of physically interpretable characterizations describing sub-groups of material configurations that share common properties.  In contrast to predictive global modeling, local pattern discovery techniques directly aim to uncover specific local properties of the data.  Moreover, they describe these properties through models that are built from discrete symbols corresponding to meaningful notions of the discovery domain, e.g., find sub-groups of materials with large band gaps and low formation energies. Therefore, they constitute a complimentary part of the data mining toolbox for the automatic analysis of materials science data repositories, such as for the Novel Materials Discovery Repository.  As an exemplary application of local pattern discovery for materials science, we consider the automatic discovery of structure-property relationships of gold clusters (size 7-14).  Local patterns between geometrical (e.g., local bond coordination, cluster size, and radius of gyration) and physicochemical (e.g., HOMO-LUMO gap and intramolecular van der Waals energy) properties are found.