Our paper Contrastive Image-Metadata Pre-Training for Materials Transmission Electron Microscopy has been accepted at the AI4Physics workshop at ICML 2026 in Seoul. In it we release the first paired HAADF-STEM image and metadata dataset, 7,330 micrographs annotated with their acquisition parameters, and use it to train a CLIP-style encoder that links an image to the settings it was recorded with. It reaches 84.4% top-1 cross-modal retrieval. The learned style representation also turns out to be useful for denoising: in blind comparisons, microscopists preferred our output over the state of the art in 70.2% of cases. Preprint on arXiv.