Berlin 2018 – wissenschaftliches Programm
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 61.1: Topical Talk
Donnerstag, 15. März 2018, 15:45–16:15, H 0107
Big Data of Materials Science: Interpretability of Machine Learning — Luca M. Ghiringhelli1, •Jan Vybiral2, Sergey V. Levchenko1, Claudia Draxl3, and Matthias Scheffler1 — 1Fritz-Haber-Institut der Max-Planck-Gesellschaft, Berlin-Dahlem, Germany — 2Czech Technical University, Dept. of Mathematics FNSPE, Prague, Czech Republic — 3Humboldt-Universität zu Berlin, Institut für Physik and IRIS Adlershof, Berlin, Germany
An important part of every machine learning approach to statistical learning is the representation of the input data. This representation usually assumes an implicit (and largely unchallenged) choice of right descriptors. To allow for human interpretability of learned structures in the data, their choice is crucial. Trustful prediction of new promising materials, identification of anomalies, and scientific advancement are doubtful if the connection between the descriptor and the actuating mechanisms is unclear.
We use the techniques of compressed sensing and feature selection to analyze this issue, define requirements for useful descriptors, and propose a practical algorithm for their identification. It selects suitable descriptors from a large set of physically meaningful quantities, which is created in a human-guided process.
For a classical example, the energy difference of zincblende/wurtzite and rocksalt semiconductors, we demonstrate how a meaningful descriptor can be found systematically.