Regensburg 2016 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 22: Topical session: Integrated computational materials engineering for design of new materials V

MM 22.4: Vortrag

Dienstag, 8. März 2016, 12:30–12:45, H39

Machine Learning of the (Meta-)Stability of Octet Binaries — •Emre Ahmetcik, Christian Carbogno, Luca M. Ghiringhelli, and Matthias Scheffler — Fritz-Haber-Institut der Max-Planck-Gesellschaft, 14195 Berlin-Dahlem, Germany

Statistical learning is regarded as the most promising technique to accelerate and systematically facilitate insights into computational material science. For instance, this has been recently successfully demonstrated by using compressed sensing techniques to predict the relative stability of zincblende versus rocksalt octet binary materials from the properties of the atomic constituents alone [1]. For an application in practical material science, it is however uncertain to which extent these approaches can be generalized, e.g., to predict metastable polymorphs or thermodynamic properties. To clarify this question, we have computed the relative stability of octet binaries for several different (meta-)stable crystal structures. We discuss the applicability of statistical learning for this question, how the approach can be generalized to predict thermodynamic properties such as transition pressures, and demonstrate its robustness with respect to numerical parameters. In particular, we critically discuss to which extent further data, e.g., dimeric and/or thermodynamic descriptors, are necessary.

[1] L. M. Ghiringhelli et. al., Phys. Rev. Lett. 114, 105503 (2015).