Regensburg 2022 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 22: Data Driven Materials Science: Experimental Data Treatment and Machine Learning
MM 22.5: Topical Talk
Wednesday, September 7, 2022, 11:45–12:15, H46
Physics guided machine learning tools in analytical transmission electron microscopy — •Cecile Hebert1,3, Hui Chen1, and Adrien Teurtrie1,2 — 1LSME - IPHYS Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland — 2Unité Matériaux et Transformations, Université de Lille, France — 3Institut de Matériaux, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland
Modern transmission electron microscopes are capable of recording large datasets containing both structural and chemical information on a scale ranging from sub-micrometer to atomic resolution. Operated in scanning TEM mode, two kind of chemical information can be obtained: either via energy dispersive X-Ray spectroscopy of via electron energy loss spectroscopy. On modern instruments, both signals can be acquired at the same time. Turning this huge amount of information (datasets can weight up to several Gb) into segmented quantitative information representing the different phases of the specimen is a real challenge. Pure statistical analysis like principal component analysis generally fails because of two main reasons: artifacts linked to the detection chain and/or non uniqueness of a statistical decomposition. The task is generally complicated by the overlap of phases in the specimen thickness and the presence of the same elements in different phases
With the introduction of physical constraints, like a modelling of the spectrum based on prior knowledge, both on the specimen and on the physical process leading to the spectra, it is possible to obtain a physically meaningful spatial segmentation of the data and to proceed with chemical analysis.