Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MA: Fachverband Magnetismus
MA 58: Computational Magnetism II
MA 58.7: Talk
Friday, March 13, 2026, 11:15–11:30, POT/0151
Automated Defect Detection in Magnetic Imaging Data Using Latent Measures and U-Net Segmentation — •Ross Knapman1,2, Nasim Bazazzadeh1,3, Kübra Kalkan1, Atreya Majumdar1, and Karin Everschor-Sitte1 — 1Faculty of Physics and Center for Nanointegration Duisburg-Essen (CENIDE), University of Duisburg-Essen, 47057 Duisburg, Germany — 2Institute of Mechanics, Faculty of Engineering, University of Duisburg-Essen, 45141 Essen, Germany — 3Radiology Department, Massachusetts General Hospital, 175 Cambridge St., Boston, MA 02114, USA
Detecting and characterising local inhomogeneities is essential for understanding and optimising magnetic materials. We present an automated framework that combines physics-based latent measures with deep convolutional segmentation for robust defect identification in magnetic imaging data. Time-resolved micromagnetic simulations with randomly distributed defects are used to compute three per-pixel descriptors: temporal mean, temporal standard deviation, and a latent-entropy measure that quantifies local dynamical complexity [1,2]. Each measure serves as input to a U-Net architecture trained for pixel-level segmentation of defect regions. Performance is evaluated under additive and multiplicative noise to test robustness. This approach demonstrates how integrating physics-motivated feature construction with deep learning enables reliable automated analysis of magnetic textures and defect landscapes in simulation and experiment.
[1] Horenko, I. et al., Comm. App. Math. and Comp. Sci 16, 267 (2021). [2] Rodrigues, D. R. et al., iScience 24, 102171 (2021).
Keywords: Machine Learning; Magnetic Imaging; Neural Network; Image Segementation; Micromagnetic Simulation