2026-09-26
An Attention U-Net maps simulated TERS hyperspectral images to atomic maps at 0.842 DSC and 0.097 Å localization, plus H/C/N/O element channels; real experimental data still fails.
In scanning probe microscopy, CO-functionalized STM and AFM resolve molecular skeletons, but asking which element a bright spot is requires spectroscopy. Tip-enhanced Raman spectroscopy (TERS) focuses a laser through an STM tip that acts as a nanoantenna, boosting the Raman signal to Ångström-scale resolution and combining topography with chemical fingerprints. The catch: interpreting a TERS hyperspectral cube traditionally takes quantum-chemistry calculations plus manual comparison. This paper, from Adam Foster's group at Aalto and published in PRX Intelligence's first volume, turns spectral-cube-to-structure into an end-to-end image translation problem.
| Task | Test performance |
| Structure prediction | DSC 0.842 (train 0.862) |
| Element-resolved prediction | mean DSC 0.810; H 0.856, C 0.827, N 0.759, O 0.799 |
| Coordinate-level (structure model) | atom-count error 0.38, precision/recall 0.98, RMSD 0.097 Å |
Sub-0.1 Å localization shows the model reads structure rather than counting blobs. Nitrogen is worst and most variable, tracking the scarcity of nitrogen-containing molecules in the dataset; detected atoms localize to about 0.09 Å for all four elements, so errors are misses and misassignments, not displacement. A planarity ablation is telling: relaxing the filter to 0.1/0.5 Å grows the dataset to roughly 7,800 molecules, yet test median DSC falls from 0.842 to 0.771 with wider spread, because out-of-plane atoms attenuate the near-field signal and make the inverse problem more ill-posed. Appendix D is the most honest part of the paper: on real TERS data of FePc on Ag(110), the raw input yields a scatter of incongruent atomic positions, and Gaussian denoising leaves a few random spots. The authors attribute the mismatch to the simulation's assumptions (sharp 5 Å tip, isolated molecules, no substrate) and state plainly that the model is not yet capable of handling experimental data.
This is the first demonstration of automated structure discovery on TERS, extending a pipeline already proven on AFM and STM to chemically specific spectroscopic imaging. Data (131 GB on Zenodo) and code (SINGROUP/SMARTERS on GitHub) are fully open, so the entry barrier is low. The bigger picture is the instrument-to-model-to-structure loop: if TERS closes its sim-to-real gap, chemically resolved atomic characterization can leave manual expert interpretation and enter automated discovery workflows.
Only isolated planar molecules are handled. The authors name three gaps themselves: non-planar systems will need 3D molecular graphs with graph neural networks; nitrogen and oxygen are underrepresented; and the sim-to-real gap is the decisive obstacle. The simulation is heavily simplified (isolated molecules, fixed sharp tip, no substrate), which is exactly what breaks on experiment, and whether augmentation can compensate is unproven. Evaluation rests on mask metrics with no direct measure of chemical correctness, and element-level atom-count errors (nitrogen off by about 3.7 atoms on average) show the output is not yet chemically usable.