A deep learning model has reconstructed molecular geometries and identified several chemical elements directly from simulated tip-enhanced Raman spectroscopy images.
The model, named SMARTERS, converts hyperspectral TERS data into two-dimensional maps of atomic positions. It could eventually reduce reliance on manual interpretation and computationally intensive quantum-chemistry comparisons, although tests with experimental data exposed a substantial gap between simulated and real measurements.
TERS combines Raman spectroscopy with a scanning probe tip that concentrates the electromagnetic field into a nanoscale region. The technique can provide chemical information with spatial resolution beyond the diffraction limit, but its datasets are difficult to interpret because image contrast depends on factors including tip geometry, molecular orientation, substrate interactions, and vibrational selection effects.
In a study published in PRX Intelligence, researchers developed an Attention U-Net encoder-decoder model to translate TERS hyperspectral image cubes into atomic maps.
The training data were generated from density functional theory calculations and TERS simulations. From an initial collection of 28,570 small organic molecules, the researchers selected 1,840 planar molecules containing up to 40 atoms. The simulated local plasmonic field had a full width at half maximum of 5 Å.
Because different molecules have different numbers of vibrational modes, the team binned spectral data into a fixed number of frequency channels. The dataset was then divided into training, validation, and test sets using an 80:10:10 split.
For atomic-position prediction, SMARTERS achieved a mean Dice similarity coefficient of 0.842 on the test set. Using a separate coordinate-extraction procedure, the model 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 for correctly detected atoms.
A second model predicted both atomic positions and elemental identities for hydrogen, carbon, nitrogen, and oxygen. Its mean test-set Dice score was 0.810. Hydrogen and carbon were identified more reliably than nitrogen and oxygen, reflecting an imbalance in the composition of the training dataset.
Performance declined as molecules deviated from planarity because atoms farther from the tip contributed weaker signals. The current approach is therefore best suited to flat molecules on weakly interacting substrates.
The principal limitation emerged when the researchers applied SMARTERS to experimental TERS images of iron phthalocyanine on silver. The predicted atomic positions did not resemble the molecular structure, even after image denoising.
The researchers attributed the failure to differences between the simplified simulations and experimental conditions, including tip geometry, substrate effects, restricted spectral coverage, and changes to the tip during acquisition. Iron was also absent from the training data.
Newsletters
Receive the latest analytical science news, personalities, education, and career development – weekly to your inbox.
