Regensburg 2016 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 55: Methods in Computational Materials Modelling III: Machine learning and statistics

MM 55.2: Talk

Thursday, March 10, 2016, 12:00–12:15, H53

Local Pattern Discovery for Uncovering Physically Interpretable Characterizations of Material Configurations — •Mario Boley, Bryan R. Goldsmith, Luca M. Ghiringhelli, and Matthias Scheffler — Fritz-Haber-Institut der MPG, Berlin.

Currently, machine learning applications to data from atomistic material simulations mostly focus on the inference of a global prediction model for some quantities of interest, such as (relative) energies. However, due to their global prediction-based objective, these models are usually not well suited for the discovery of physically interpretable characterizations of local groups of subsets of material, in particular when aiming at partitioning the configurational space of a system at finite temperature.

In contrast to predictive global modelling, 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. Therefore, they constitute an important complimentary part of the data mining toolbox for the automatic analysis of materials-science data repositories. As an exemplary application of local pattern discovery for materials science, we consider the automatic discovery of configurational basins of gold clusters and their characterization in terms of interpretable features such as their coordination histogram and their moments of inertia, or (non-linear) functions of those features.