Dresden 2020 – wissenschaftliches Programm
Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 103: Topical Session: Data Driven Materials Science - Machine Learning for Materials Characterization (joint session MM/CPP)
CPP 103.1: Topical Talk
Donnerstag, 19. März 2020, 15:45–16:15, BAR 205
Machine learning tools in analyticat transmission electron microscopy — •Cécile Hébert and Hui Chen — LSME, Institute of Physics, Ecole Polytechnique Fédérale de Lausanne, Switzerland
Analytical scanning transmission electron microscopy probes the chemistry of the investigated sample by recording spectral information as a function of electron probe position. The acquired spectra can consist of X-Ray photons emitted by the sample after the incoming electron probe has excited it (EDX) and/or an analysis of the energy lost by the incoming electron when it excites the sample (EELS). Both EELS and EDX spectra can be recorded on a scanned area consisting of 1000x1000 pixels of even more, leading to a so called *hyperspectral datacube* of up to several 106 spectra. Such a vast amount of data calls for machine learning tools belonging to the family of multivariate statistical analysis (MSA). Such methods have been implemented and used since the mid-ninties, however, there are still many challenges related to their application. MSA methods are very sensitive to detector artifacts, they deliver components, which do not necessarily bear a physical meaning, they might discard small and very localized signal, etc. In this presentation, I will review the use of unsupervised machine learning in analytical TEM, and present some new results based on a dictionary learning approach where we implement knowledge we have about the shape of the spectral components.