Training Dataset for SMARTERS: A Machine Learning Model for Automated Structure Discovery in Tip-Enhanced Raman Spectroscopy

The repository contains the data used to train SMARTERS, an encoder-decoder model for molecular structure prediction from tip-enhanced Raman spectroscopy (TERS) hyperspectral images. It includes both the original quantum-chemistry outputs and the derived simulated TERS data. Quantum-chemistry calculations were performed using the Gaussian software package, and the resulting Gaussian formatted checkpoint files (.fchk) contain the molecular geometry, electronic structure, and vibrational information used to generate the TERS images. The dataset is distributed as archived .fchk files and processed NumPy compressed archives (.npz). The processed samples contain rasterised TERS intensity maps and structural metadata, and were filtered to retain only near-planar molecules with a root-mean-square deviation (RMSD) below 0.5 Å from the best-fit molecular plane.