UniMR: A Plug‐and‐Play Framework of Automated Molecular Recognition for Scanning Tunneling Microscopy

Abstract Scanning Tunneling Microscopy (STM) has become an indispensable tool for molecular surface chemistry, where recognizing molecules in STM images is a fundamental task. Human annotation requires a great deal of domain knowledge and is laborious, while current automated methods usually require training and focus on particular molecular systems. In this work, a training‐free universal molecular recognizer (UniMR) is developed that is widely applicable to various molecular systems and is tolerant of low‐resolution images. UniMR integrates off‐the‐shelf computer vision tools with a feature adaptive selection module. First, input images are normalized to ensure consistent molecular brightness and orientation. To address the varying complexity of molecular differentiation across different systems, an adaptive feature selection module is introduced. Specifically, each molecular image is represented by a pixel matrix for shallow structural features and CLIP embeddings for deep semantic features. A Gaussian Mixture Model (GMM) is followed to dynamically select the optimal representation. Finally, cosine similarity identifies molecules of the same type. UniMR is evaluated on five molecular systems with diverse imaging resolutions. The results demonstrate a significant improvement over previous methods, achieving an average F1‐score . This framework can serve as a versatile auxiliary tool for STM, advancing microscopic surface chemistry research.