Quantitative characterization of in situ Lorentz transmission electron microscopy (LTEM) data for nanoscale magnetic spin textures is challenging. This paper presents an AI-enabled phase-retrieval method integrating a generative deep image prior with an LTEM image formation forward model. Using a single out-of-focus image, this approach (SIPRAD) achieves higher accuracy and noise robustness than existing methods, isolating sample heterogeneities from magnetic contrast. SIPRAD enables quantitative phase reconstruction of in situ data and potentially near real-time quantitative magnetic imaging.

Arthur R. C. McCray, Tao Zhou, Saugat Kandel, Amanda Petford-Long, Mathew J. Cherukara, Charudatta Phatak