Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
KFM: Fachverband Kristalline Festkörper und deren Mikrostruktur
KFM 7: Focus Session: Defects and Interfaces in Multiferroics 2
KFM 7.4: Vortrag
Montag, 5. September 2022, 16:25–16:45, H5
Deep learning evaluation of conductive atomic force microscopy data — Lorenz Glück1,2, •Manuel Zahn1,3, Lukas Puntigam1, Donald M. Evans1, Somnath Ghara1, Michael Heider2, and Stephan Krohns1 — 1Experimental Physics V, University of Augsburg, 86159 Augsburg — 2Organic Computing Group, University of Augsburg, 86159 Augsburg — 3Institut für Angewandte Physik, Technische Universität Dresden, 01069 Dresden
Machine learning has gained an enormous interest in the past decade to boost data evaluation in many fields of applied physics. For example, feature recognition in high dimensional datasets in scanning probe microscopy (SPM) can be improved and hidden effects resolved. However, the physical relevance of resolved features is normally still determined by humans.
In this work, we investigate if the regularization of a deep learning (DL) neuronal network, composed of long-short term memory and temporal convolutional network based layers inside an autoencoder architecture, can be utilized to characterize physical significance. We do this on a conductive atomic force microscopy dataset, collected on ferroelectric GaV4S8, as the general properties have already been identified and there are emergent traits at the domain walls. The resolved features from the DL approach are compared to those derived from classical clustering algorithms and classically resolved local material properties. This set up is the first steps to automatic evaluation of physically significant properties in GaV4S8, and is expected to be applicable to other ferroelectric systems.