Dresden 2017 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 60: Topical session: Data driven materials design - machine learning

MM 60.3: Vortrag

Donnerstag, 23. März 2017, 12:30–12:45, BAR 205

Automatic crystal-structure classification using X-ray diffraction patterns and convolutional neural networks — •Angelo Ziletti, Matthias Scheffler, and Luca M. Ghiringhelli — Fritz Haber Institute of the Max Planck Society, Berlin, Germany

With the advent of high-throughput materials science, millions of calculations are now available to the scientific community (http://nomad-repository.eu and references therein). A reliable identification of the lattice symmetry in these calculations is a crucial first step for materials characterization and analytics. Current methods based on space-group symmetries require a user-specified threshold, and are unable to detect ``average'' symmetries for defective structures (e.g., with point defects, and/or strain). We propose a new machine-learning based approach to automatically classify periodic structures according to their Bravais lattice. First, we calculate the X-ray diffraction patterns, from which a classifying model is then learned using a convolutional neural network. This method is applied to crystal-structure classification of 3d, 4d, and 5d transition metal alloys, also containing vacancies. We show that our deep-learning model can correctly classify more than 99% of the crystal structures. Moreover, contrarily to other (common) methods, it does not require any tolerance threshold and provides a reliable probabilistic classification also for heavily defective structures. Our approach has been implemented in the NOMADsim package of the Novel Materials Discovery (NOMAD) Analytics-Toolkit (https://analytics-toolkit.nomad-coe.eu). This work received funding from the NOMAD Laboratory, a European Center of Excellence.