Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 61: Topical Session (Symposium MM): Big Data in Materials Science - Managing and exploiting the raw material of the 21st century

MM 61.5: Vortrag

Donnerstag, 15. März 2018, 17:00–17:15, H 0107

From autonomous subspace selection of material properties to physically meaningful predictions — •Benjamin Regler, Matthias Scheffler, and Luca M. Ghiringhelli — Fritz Haber Institute of the Max Planck Society, Berlin, Germany

In data-driven materials science, the discovery of functional relationships is an inverse problem on finding subsets of materials properties (features) which relate to physical observables.

The inverse problem can be solved by statistical-learning algorithms that require manual adjustment and are thus based on ongoing experience and the knowledge being built. This makes ensuring the physical interpretability of the generated data-based models a demanding task.

Therefore, we highlight an autonomous systematic feature-subspace construction and selection method using information-theoretic concepts. We use a generalization of the Shannon entropy to select physically meaningful subsets (sharing the same notion of uncertainty as the Gibbs entropy in statistical thermodynamics) and use compressed sensing to find the best approximate models without prior knowledge.

Then, we apply the framework and highlight key problems such as stability predictions of zinc-blende vs. rock-salt octet binary semiconductors and band-gap predictions of topological insulators. Finally, we discuss the physical interpretation of the generated models and identify the strongest correlated materials properties with the actuating mechanism.