Physics-guided deep learning with adversarial domain adaptation: Applications to STM image denoising

Abstract

Image denoising is essential in materials characterization, particularly for recovering fine structural details in scanning tunneling microscopy (STM) images. While supervised denoising approaches have shown strong performance, they typically rely on large datasets of paired noisy and clean images, which are often unavailable in experimental settings. Unsupervised methods, though not requiring paired data, often rely on a collection of unpaired clean images for training—resources that are frequently unavailable in real-world STM laboratory environments. In this work, we propose PDA-Net, a physics-guided deep learning framework with adversarial domain adaptation for unsupervised STM image denoising. PDA-Net leverages a physics-based simulator to generate synthetic STM images for the surface of copper single crystals, i.e., Cu(111), serving as a proxy for the clean ground truth. Built upon a generative adversarial network architecture, the framework integrates cycle-consistency and domain adversarial modules to bridge the gap between simulated and real experimental domains in the absence of paired data. Additionally, feature alignment and weight-sharing strategies are employed to enhance knowledge transfer between domains. Experimental results demonstrate that PDA-Net significantly improves STM image quality, enabling more accurate interpretation of quantum material properties and facilitating accelerated scientific discovery.