Thomas B. Winkler1, Isaac Labrie-Boulay 1, Alena Romanova1 , Daniel Franzen2, Hans Fangohr3,4 , Mathias Kläui1
1 Institut für Physik, Johannes Gutenberg University, Mainz, Rhineland Palatine, Germany.
2 Institut für Informatik, Johannes Gutenberg Universität, Mainz, Rhineland Palatine, Germany.
3 Scientific Support Unit for Computational Science, Max Planck Institute for the Structure and Dynamics of Matter, Hamburg, Germany.
4 Computational Modelling Group, Faculty of Engineering and Physical Sciences, University of Southampton, Southampton, Hampshire, United Kingdom.
Magnetic skyrmions [1] are magnetic, topologically stabilized quasi-particles, which exhibit interesting physical phenomena and are also potential candidates for efficient or unconventional computation schemes. On the micro-meter scale they are detectable with Kerr-microscopy, using the magneto-optical Kerr-effect.
However, depending on the material stack, temperature, the growth procedure and other influences, those measurement might suffer from noise, low-contrast, intensity gradients or defects, therefore manual data treatment is normally necessary to generate data that can be evaluated easily. We are using the U-Net[2], a convolutional neural network, to detect the position and shape of the skyrmions in our measurement data. This work is mainly motivated by recent publications reporting that machine learning has been successfully applied to (micro-)magnetic problems [3]. We are also tuning the UNet with various techniques, optimizing the quality of the outcome to obtain fast and reliable predictions. A well-trained neural network is shown to minimize manual treatment of data. The approach can be extended easily to other magnetic structures.
References
[1] Everschor-Sitte et al., Journal of Applied Physics 124, 240901 (2018)
[2] Ronneberger et al., MICCAI 2015. Lecture Notes in Computer Science, vol 9351. Springer, Cham. (2015)
[3] Alexander Kovacs et al., Journal of Magnetism and Magnetic Materials 491, 165548 (2019)