Untrained Physically Informed Neural Network for Image Reconstruction of Magnetic Field Sources

Abstract

Magnetic materials are a vital resource in designing energy-efficient information technologies. To try to learn how magnetism develops in ultrathin systems, we measure, but deducing the physics afterward is an ill-posed problem. This study uses neural networks to facilitate the reconstruction of the underlying magnetic textures of thin magnets through measurements of their stray fields. The technique is surprisingly robust to experimental noise, and can reliably reconstruct magnetism in arbitrary directions. Importantly, prior training of the network is not required, and the technique is broadly applicable for solving ill-posed inverse problems when the forward problem is well defined.