Deep-Learning-Based Denoising for Improved Phase Precision in Electron Holography of Electromagnetic Fields in Nanoscale Materials.

Electron holography is a powerful tool for the quantitative mapping of electric and magnetic fields at the nanoscale, with direct applications in semiconductor devices, magnetic nanostructures, and functional nanomaterials. However, its sensitivity is fundamentally limited by shot noise. At low dose conditions, shot noise overwhelms subtle fringe shifts that carry weak signals. Recent advances in deep learning have shown remarkable success in image denoising. Here we present HoloDenoiser, an unsupervised deep neural network that exploits the localization of phase information at sideband peaks in Fourier space as an explicit physics-guided prior. The network achieves 3-4 fold improvement in phase precision across both Linear and Counting detector modes, outperforming conventional denoisers and state-of-the-art deep learning baselines. We demonstrate the method experimentally on zero-phase vacuum holograms and on built-in potential mapping across a GaAs p-n junction. Using simulated holograms of a magnetic nano-disk, we further recover weak magnetic phase signals with a ∼25-fold reduction in the required electron dose. We anticipate that this work will provide a useful tool for the broader nanoscience community working on quantitative phase imaging at the nanoscale.