TERS-ABNet: deep learning for single-molecule reconstruction with atomic precision from TERS mapping

Abstract

Determining the chemical structure of a single molecule on a surface from spectroscopic data constitutes a challenging high-dimensional inverse problem. While tip-enhanced Raman spectroscopy (TERS) enables chemically specific imaging of single molecules with sub-nanometer spatial resolution, reconstructing complete molecular structures from TERS maps remains difficult due to ambiguous vibrational signatures and heavy reliance on expert interpretation. Here, we introduce TERS-ABNet, a deep-learning framework that reformulates single-molecule structure determination as an image-to-graph inference task. Leveraging a two-track architecture, the model simultaneously predicts probabilistic maps for predefined atom and bond/group classes, enabling the direct construction of explicit atom-bond graphs without dependence on conventional vibrational-mode assignment. Trained on simulated datasets, TERS-ABNet achieves ~94% accuracy in atom-type classification (mean coordinate error: ~0.23 Å), enabling reliable recovery of molecular connectivity and full reconstruction of single-molecule structures from TERS maps. The framework generalizes across diverse spatial resolutions and structural complexities through transfer learning, and successfully retrieves partial atom- and bond-resolved features of a single porphyrin molecule from experimental TERS data. This work establishes a generalizable deep-learning paradigm for inferring explicit atom-bond graph representations from high-dimensional spectroscopic imaging data, offering a transformative pathway towards automated molecular structure determination in nanoscale characterization.