Enhancing atomic-resolution in electron microscopy: A frequency-domain deep learning denoiser
Atomic resolution electron microscopy, particularly high-angle annular dark-field scanning transmission electron microscopy, has become an essential tool for many scientific fields, when direct visualization of atomic arrangements and defects are needed, as they dictate the material's functional and mechanical behavior. However, achieving this precision is often hindered by noise, arising from electron microscopy acquisition limitations, particularly when imaging beam-sensitive materials or light atoms. In this work, we present a deep learning-based denoising approach that operates in the frequency domain using a convolutional neural network U-Net trained on simulated data. To generate the training dataset, we simulate FFT patterns for various materials, crystallographic orientations, and imaging conditions, introducing noise and drift artifacts to accurately mimic experimental scenarios. The model is trained to identify relevant frequency components, which are then used to enhance experimental images by applying element-wise multiplication in the frequency domain. The model enhances experimental images by identifying and amplifying relevant frequency components, significantly improving signal-to-noise ratio while preserving structural integrity. Applied to both Ge quantum wells and WS2 monolayers, the method facilitates more accurate strain quantitative analyses, critical for assessing functional device performance (e.g. quantum properties in SiGe quantum wells), and enables the clear identification of light atoms in beam sensitive materials. Our results demonstrate the potential of automated frequency-based deep learning denoising as a useful tool for atomic-resolution nano-materials analysis.