Breaking the Limits of Scanning Tunneling Microscopy Using Image Super Resolution

A Scanning Tunneling Microscope (STM) is a type of microscope that harnesses quantum tunneling to process images of surfaces at the atomic level. STMs are used to study particles at nanoscale, enabling detailed investigations at the level of individual atoms. However, the resolutions of STM images are often limited by realistic experimental conditions. In this paper, we investigate image super-resolution methods empowered by deep learning to break the experimental limits of STM to obtain high-resolution STM images. We first train an object detection model to detect and segment target molecules from the raw STM images. Then, an STM Super-Resolution (STM-SR) program based on Super-Resolution Generative Adversarial Networks (SRGAN) is developed and trained on a high-resolution STM dataset to learn to convert the low-resolution STM images into high-resolution ones. Our benchmark results on the test STM set show that STM-SR leads to significant resolution improvement measured by Peak Signal to Noise Ratio (PSNR), compared to the bicubic interpolation method. This super-resolution breaks the experimental barrier and reaches a higher resolution level. Finally, a Localization Scanning Tunneling Microscopy (Localization-STM) method is developed to reconstruct and further enhance the resolution of the STM image beyond the STM tip-radius limitation. As a case study, we apply our machine learning approach to generate super-resolution STM images that resemble the supramolecule, which matches its theoretical chemical structure well. The machine learning-based super-resolution approach enables STM images to reach a similar quality under less restrictive conditions and brings new insight into atomic-level physics. In particular, this method can be applied to the STM study of large, complex molecular structures, which requires stringent experimental conditions.