Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 47: Poster Session II
MM 47.51: Poster
Mittwoch, 14. März 2018, 18:30–19:45, Poster C
Compressed-sensing-based feature selection strategies in materials science: Defining the "best model" — •Benedikt Hoock1,2, Santiago Rigamonti1, Luca Ghiringhelli2, Matthias Scheffler1,2, and Claudia Draxl1,2 — 1Humboldt-Universität zu Berlin, Berlin, DE — 2Fritz-Haber-Institut der MPG, Berlin,DE
Machine Learning (ML) methods are being currently established in materials science in order to find best models that help to better understand existing data or to identify even new materials. In this context, one needs to define what the term "best model" means. Up to now, this is far from being precisely defined, even if one confines oneself to a certain model class. We compare several approaches for cross validation (CV) based model selection strategies [1]. These differently balance between fitting accuracy and generalizability by using either training, or average training or test errors as selection criterion, respectively, and hence lead to different definitions for the "best model". We apply these strategies to a set of ab-initio calculated group-IV ternaries to predict lattice constants and energies of mixing. This is achieved by adapting a LASSO-based ML method [2] to find best descriptors constructed from simple atomic, dimer and tetrahedron data.
[1]: B. Hoock, S. Rigamonti, L. M. Ghiringhelli, M. Scheffler, and C. Draxl, "Predicting lattice constants and energies of mixing of group-IV ternary materials by compressed sensing", in preparation. [2]: L. M. Ghiringhelli, J. Vybiral, E. Ahmetcik, R. Ouyang, S. V. Levchenko, C. Draxl, New Journal of Physics 19.2, 023017 (2017).