Keypoint-Guided Three-Stage Framework with an Interpretable Geometry-Based Decision Layer for Recognizing Structurally Similar Metallosupramolecular Isomers in Scanning Tunneling Microscopy Images

Abstract Automated and accurate recognition of molecular images obtained by scanning tunneling microscopy (STM) is a fundamental prerequisite for high-throughput surface science investigations. However, the detection and discrimination of these metallosupramolecules in STM images are further challenged by the presence of structurally similar isomers and the intrinsic noise inherent to STM imaging. Here, we developed a keypoint-guided three-stage recognition framework for three metallosupramolecular isomers (C2, C3, and C6) formed by coordination-driven self-assembly of a tetratopic terpyridine ligand with Zn(II). Stage I used the YOLOv8m object detector (You Only Look Once, version 8 medium) only for target detection, localizing the broad class of six-membered ring molecules in large-area STM images without performing C2/C3/C6 subtype classification. Stage II used Lite-HRNet-30 to predict six ordered keypoints from each cropped single-molecule image. Stage III converted the predicted keypoints into explicit edge-length, diagonal, angular, and axial-symmetry descriptors and then applied predefined geometric rules to determine the final isomer category. Stage I achieved an [email protected] of 0.956 under five-fold cross-validation. Of the 292 held-out Stage I candidates, 194 had defensible C2/C3/C6 reference labels and were therefore eligible for Stage II-to-Stage III evaluation, corresponding to a label-eligible coverage of 66.44%. On this held-out subset, the complete Stage II-to-Stage III path achieved an overall accuracy of 81.44%, with class-wise recalls of 84.29%, 73.44%, and 86.67% for C2, C3, and C6, respectively. The workflow, therefore, automates expert geometric analysis while limiting direct interpretability to the final geometry-based decision layer.