# Molecular Structure Prediction

MolP2SBench recasts molecular property-to-structure inference as masked-scaffold multiple choice and adds shuffled and counterfactual property substitutions, separating a correct answer from genuine use of the supplied evidence. Across seven models, accuracy on a property channel does not track how strongly a model responds when that property is replaced, and model performance does not always match chemists' intuition about which properties are easier.

MolP2SBench recasts molecular property-to-structure inference as masked-scaffold multiple choice and adds shuffled and counterfactual property substitutions, separating a correct answer from genuine use of the supplied evidence. Across seven models, accuracy on a property channel does not track how strongly a model responds when that property is replaced, and model performance does not always match chemists' intuition about which properties are easier.

The work introduces SMARTERS, an Attention U-Net encoder-decoder trained on simulated TERS hyperspectral images of 1,840 planar small molecules, which maps vibrational spectral images directly to 2D atomic position maps (mean test Dice similarity coefficient 0.842) and to chemically resolved maps per element (H, C, N, O; mean test 0.810), showing that automated molecular structure identification from TERS images is feasible on simulated data without conventional manual comparison and case-by-case quantum-chemistry calculations.

The work introduces SMARTERS, an Attention U-Net encoder-decoder trained on simulated TERS hyperspectral images of 1,840 planar small molecules, which maps vibrational spectral images directly to 2D atomic position maps (mean test Dice similarity coefficient 0.842) and to chemically resolved maps per element (H, C, N, O; mean test 0.810), showing that automated molecular structure identification from TERS images is feasible on simulated data without conventional manual comparison and case-by-case quantum-chemistry calculations.