Resolution Enhancement of Scanning Electron Micrographs using AI

Microscopic imaging of materials often requires the examination of large sample areas at high magnification to identify and analyse rare structural features. High-resolution imaging in scanning electron microscopy is particularly time-intensive, as images are acquired through sequential scanning of an electron beam across the sample surface. This presents a critical challenge, as researchers must balance imaging speed and resolution while ensuring statistically meaningful observations of sparsely distributed features. To address these challenges, we present a novel resolution enhancement method for electron microscopy based on artificial intelligence. Suitable reference images are selected using vector embeddings and processed by a texture-transformer network. Using a tailored dataset of dual-phase steel micrographs, we demonstrate that our trained network outperforms traditional interpolation methods in both quantitative similarity metrics and crucial material-specific features, such as phase boundaries and microstructural voids. The method’s transferability is validated using micrographs from a 16MnCrS5 case-hardening steel sample. We achieve a 16-fold acceleration through resolution trade-offs between target and recording and additionally propose a scan-enhance-rescan workflow where resolution-enhanced micrographs guide the identification of regions of interest for targeted high-resolution rescanning. We provide quantitative estimates of expected time savings, offering a practical framework for efficient high-resolution microscopy across large areas. Here we provide the code necessary for training and dataset generation together with the datasets used in the publication. `data` contains all image triplets of high-resolution low-resolution and assigned reference images. In `model` the best performing network weights are supplied. `code` contains the code for network training and evaluation and dataset preparation.