Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MA: Fachverband Magnetismus
MA 6: Magnetic Imaging Techniques I
MA 6.7: Vortrag
Montag, 9. März 2026, 11:15–11:30, POT/0361
Hamiltonian reverse engineering from magnetic skyrmion images via deep learning surrogates — •Moritz Winterott1,2 and Samir Lounis3 — 1Peter Grünberg Institut, Forschungszentrum Jülich & JARA, Germany — 2Faculty of Physics, University of Duisburg-Essen, Germany — 3Institut für Physik and Halle-Berlin-Regensburg Cluster of Excellence CCE, Martin-Luther Universität Halle-Wittenberg, Germany
The extraction of physical parameters from experimental observations is a central inverse problem in condensed matter physics. Complex magnetic textures such as skyrmions are imaged via Scanning Tunneling Microscopy (STM), which probes the local density of states (LDOS). Inferring the underlying interactions by matching theoretical models to experimental LDOS is often computationally expensive. In this work, we introduce a deep-learning framework that serves as a fast and accurate surrogate for this modeling process. Our approach employs a novel neural network architecture that integrates modern Transformers[1,2] with Convolutional Neural Networks (CNNs) to extract spatial features from energy-resolved LDOS images and to learn the highly nonlinear mapping to the parameters of an effective tight-binding Hamiltonian. These parameters encode the skyrmion texture, hopping amplitudes, and spin-orbit coupling strengths. Also, on-the-fly noise augmentation during training enhances robustness, enabling the model to maintain high accuracy even for noisy experimental data.
[1] Vaswani et al. NeurIPS '17; [2] Devlin et al. Proc. NAACL-HLT '19.
Keywords: Transformer; STM; Neural Network; Skyrmions