Abstract
Reconstructing the underlying magnetic structure from stray- field measurements is a central inverse problem across spintronics\, mater ials science\, medicine\, and geology. However the problem is notoriously ill posed — multiple solutions can fit a particular experimentally measu red stray field profile. Classical Fourier inversion imposes restrictive a ssumptions on the sample geometry and magnifies noise\, limiting its appli cability for many advanced applications. This has led to more advanced mac hine-learning- and optimization-based algorithms in recent years to addres s the issue. Using physics-informed optimization\, we investigate the dege neracy of the forward operator with respect to topology. In particular we show how topologically distinct textures such as skyrmions and merons can give rise to near-identical stray fields under realistic magnetometry setu ps\, challenging the widespread use of the topological charge as a strong regularizer term in modern reconstruction methods. Next we incorporate unc ertainty quantification into an Untrained Physics-Informed Neural Network (uPINN) approach by Dubois et al\, an enhancement that instead of a single deterministic solution\, tells us how confident the model is about the re construction\, a feature previously absent in the literature. Finally\, we apply a state-of-the-art differentiable micromagnetics package NeuralMag (Abert et al) to perform full 3D magnetization reconstruction of a complex synthetic antiferromagnetic layered structure\, from NV vector-field magn etometry data. We conclude with a roadmap of future projects and exciting directions in this frontier.
Zoom : 633 2801 2201 (https://tu-dresden.zoom-x.de/j/633280122 01?pwd=jFjgmPWVLgobCkrIMab0jFS2jLXcWg.1)\, Passcode: IFMP2025-6
DTSTAMP:20260727T095414Z CREATED:20260709T053718Z LAST-MODIFIED:20260720T053841Z END:VEVENT END:VCALENDAR