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

MM: Fachverband Metall- und Materialphysik

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

MM 37.3: Vortrag

Donnerstag, 4. April 2019, 15:45–16:00, H43

Simultaneous and Reinforced Learning of Materials Properties from Incomplete Databases with Multi-Task SISSO — •Emre Ahmetcik, Runhai Ouyang, Christian Carbogno, Matthias Scheffler, and Luca M. Ghiringhelli — Fritz-Haber-Institut der Max-Planck-Gesellschaft, 14195 Berlin-Dahlem, Germany

Identifying descriptors that capture the underlying mechanisms for different materials properties is a key challenge in data-driven materials science. Recently, the sure-independence screening and sparsifying operator (SISSO) [1] has been introduced and was successfully applied to a number of key materials-science problems [1-3]. SISSO is a compressed-sensing based methodology yielding predictive and insightful models identified from an enormous space (billions or more) of candidate analytical expressions. In this work, we have extended the methodology to a ‘multi-task learning’ approach, a powerful and nontrivial generalization which identifies a single descriptor capturing multiple target materials properties at the same time. This approach is specifically suited for a heterogeneous materials database with missing data, i.e., in which not all properties are reported for all materials. As showcases, we address the relative stability of octet-binary compounds for several crystal phases and the metal/insulator classification of binary materials distributed over many crystal-prototype classes.

[1] R. Ouyang et. al., Phys. Rev. Mater. 2, 1-11 (2018).

[2] C. J. Bartel et. al., Nat Commun. 9, 4168 (2018).

[3] C. J. Bartel et. al., Sci. Adv., accepted (2018), arXiv:1801.07700.