To develop a deep learning model that reconstructs molecular geometries and identifies chemical elements from simulated tip-enhanced Raman spectroscopy (TERS) images.
Approach:
Model Development: The model, named SMARTERS, uses an Attention U-Net encoder-decoder architecture to convert TERS hyperspectral image cubes into atomic maps.
Data Generation: Training data were generated from density functional theory calculations and TERS simulations, focusing on 1,840 planar molecules from an initial set of 28,570.
Performance Metrics: SMARTERS achieved a mean Dice similarity coefficient of 0.842 on the test set for atomic-position prediction.
Key Findings:
SMARTERS recorded precision and recall of 0.98, a mean atom-count error of 0.38, and a coordinate root mean square deviation of 0.097 Å.
A second model predicted both atomic positions and elemental identities, achieving a mean test-set Dice score of 0.810.
Performance declined for non-planar molecules due to weaker signals from atoms farther from the tip.
Interpretation:
SMARTERS shows variability in performance based on molecular geometry and training data composition.
Limitations:
SMARTERS performed poorly on experimental TERS images, with predicted atomic positions not resembling the actual molecular structure.
Differences between simplified simulations and experimental conditions, such as tip geometry and substrate effects, contributed to the limitations.
Conclusion:
Significant gaps were identified when applying the current approach to experimental data.
