Reconstruction of nontrivial magnetization textures from magnetic field images using neural networks

Spatial imaging of magnetic stray fields from magnetic materials is a useful tool for identifying a material's underlying magnetic configurations; however, transforming a magnetic image into a magnetization image is an ill-posed problem, and this can result in artefacts that limit the inferences that can be drawn from the material under investigation. In this work, we develop a neural-network fitting approach that approximates this transformation, reducing these artefacts. In addition, we demonstrate that the approach allows the inclusion of additional models and bounds that are not possible with traditional reconstruction methods. These advantages allow for the reconstruction of nontrivial magnetization textures with varying magnetization directions in thin-film magnets, which was not possible previously. We demonstrate this capability by performing magnetization reconstructions on a variety of topological spin textures.