Regensburg 2019 – scientific programme

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

MM: Fachverband Metall- und Materialphysik

MM 37: Topical session (Symposium MM): Big Data Analytics in Materials Science

MM 37.4: Talk

Thursday, April 4, 2019, 16:00–16:15, H43

Information-theoretic Feature Selection and its Applications in Materials Science — •Benjamin Regler, Matthias Scheffler, and Luca M. Ghiringhelli — Fritz Haber Institute of the Max Planck Society, Berlin, Germany

Feature selection is a technique for proposing subsets of relevant properties (features), along with a measure which scores the different subsets.

In this talk, we give an overview of feature selection methods applied to materials science problems and discuss how to identify relationships between fundamental properties at the atomistic scale and materials properties at the macroscopic scale. In particular, we focus on the complexity of machine-learning models and highlight the advantages of using a systematic feature selection prior to making predictions (i.e., building machine-learning models).

Moreover, we propose a parameter-free, deterministic information-theoretic feature-selection framework for identifying approximate functional relationships between properties of interest. Importantly, the framework detects redundant and irrelevant features by performing nonlinear correlation analysis.

As showcase, we apply our approach to crystal-structure and other properties prediction in a restricted class of materials such as functionalized or octet binaries.

We conclude that our approach reduces the complexity of machine-learning models, extracts the most informative set of features, and supplements the analyses and identification of relevant properties.