TERS-ABNet: A Deep Learning Approach for Automated Single-Molecule Structure Reconstruction with Atomic Precision from TERS Mapping

This paper introduces TERS-ABNet, a deep learning framework that automates the reconstruction of single-molecule atomic structures from Tip-Enhanced Raman Spectroscopy (TERS) maps by formulating the task as an image-to-graph inference problem, achieving high accuracy in predicting atom types and coordinates and successfully demonstrating its capability on experimental porphyrin data.

The Gist

The Big Problem: The "Blind" Detective

Imagine you are a detective trying to solve a crime, but you can't see the suspect. You only have a blurry, noisy recording of their voice and a few scattered footprints. Your job is to reconstruct exactly what the suspect looks like, what they are wearing, and how their body is put together, just from that audio and those footprints.

In the world of science, this is what researchers face when trying to figure out the structure of a single molecule on a surface.

• The Suspect: A single molecule (like a tiny Lego structure).
• The Clues: A "TERS map." This is a special kind of image created by a microscope that uses a super-sharp metal tip to "feel" the molecule's vibrations. It's like listening to the molecule hum different notes as the tip moves over it.
• The Challenge: These "humming" notes are messy. A single note doesn't tell you if it's a carbon atom or a nitrogen atom. It's like hearing a chord on a piano and trying to guess exactly which keys were pressed without seeing the piano. Usually, this requires a human expert to stare at the data for hours and make educated guesses, which is slow and often wrong.

The Solution: TERS-ABNet (The AI Detective)

The authors created a new Artificial Intelligence (AI) called TERS-ABNet. Think of this AI as a super-detective who has read every chemistry textbook and memorized millions of molecular "fingerprints."

Instead of guessing, the AI looks at the blurry TERS map and instantly says: "I know exactly where every atom is, what kind of atom it is (Carbon, Nitrogen, Oxygen, Hydrogen), and how they are connected."

How It Works: The "Two-Track" System

The AI is built with two specialized brains working together, like a construction crew with two distinct roles:

1. The "Spotter" (ANet): This part of the AI looks at the map and shouts, "There's a Carbon atom here! There's a Hydrogen atom there!" It creates a map of dots, pinpointing exactly where every atom is sitting.
2. The "Connector" (BNet): This part looks at the same map and says, "Okay, that Carbon is holding hands with that Oxygen. And that Nitrogen is linked to that Hydrogen." It draws the lines (bonds) between the dots.

By working together, they don't just guess; they build a complete 3D model of the molecule from scratch, turning a blurry sound-map into a clear, sharp blueprint.

The Magic Trick: Seeing More Than the Eye Can

Usually, to see individual atoms, you need a microscope with perfect, crystal-clear vision (like seeing a single grain of sand from a mile away). If the image is blurry, you can't tell the atoms apart.

Here is the paper's breakthrough: TERS-ABNet can figure out the structure even if the microscope image is a bit blurry or "fuzzy."

• The Analogy: Imagine trying to recognize a friend's face in a foggy mirror. A human might fail. But if you have an AI that has seen that friend in thousands of different lights and angles, it can say, "Even though the nose is blurry, the shape of the jaw and the hairline match my friend perfectly."
• The AI uses context. It knows that if it sees a specific pattern of vibrations, it must be a ring of atoms, even if the individual atoms aren't perfectly sharp. This means scientists don't need the most expensive, perfect equipment to get good results; they can use slightly "fuzzier" data and let the AI clean it up.

Testing the Detective

The researchers tested this AI in three ways:

1. The Simulation Test: They fed the AI millions of fake, perfect molecular maps. The AI got it right 94% of the time, pinpointing atoms with an error smaller than the width of a human hair (about 0.23 Angstroms).
2. The "Blurry" Test: They made the images fuzzier (simulating a less perfect microscope). The AI still worked! It could still figure out the general shape and connections, proving it doesn't need perfect conditions to be useful.
3. The Real-World Test: They tried it on a real molecule (Magnesium Porphyrin) using real experimental data. It wasn't perfect (it missed a few tricky atoms), but it successfully identified the main ring structure and the connections. This proved the AI works in the real world, not just in computer simulations.

Why This Matters

Before this, figuring out a single molecule's structure was like trying to solve a puzzle with half the pieces missing and no picture on the box. You had to rely on a human expert's intuition.

TERS-ABNet changes the game:

• Automation: It turns a slow, manual art into a fast, automatic process.
• Accessibility: It allows scientists to use slightly less perfect equipment because the AI can "fill in the gaps."
• Discovery: It opens the door to understanding complex molecules (like drugs or new materials) much faster, potentially speeding up the discovery of new medicines and technologies.

In short: The authors built an AI that can look at a blurry, noisy "sound map" of a molecule and instantly draw its perfect, atomic-level blueprint, making the invisible world of single molecules much easier to see and understand.

Technical Summary

1. Problem Statement

Determining the complete chemical structure of a single molecule on a surface from spectroscopic data is a challenging high-dimensional inverse problem.

• Limitations of Current Methods: While Tip-Enhanced Raman Spectroscopy (TERS) offers sub-nanometer spatial resolution and chemical specificity, reconstructing full molecular structures (atomic positions and bonding connectivity) from TERS maps is difficult.
• Core Challenges:
  • Ambiguity: Vibrational signatures in confined plasmonic fields are complex, polarization-dependent, and often delocalized, meaning a single TERS image feature does not uniquely correspond to a specific atom or bond.
  • Expert Dependency: Current structural interpretation relies heavily on expert intuition and manual vibrational assignment, limiting scalability and robustness, especially for complex conjugated or heterocyclic systems.
  • Degeneracy: Conventional spectroscopic descriptors often lead to multiple candidate structures rather than a unique solution.
• Goal: To develop an automated, generalizable framework that infers explicit atom-bond graphs directly from high-dimensional TERS spectroscopic images without relying on predefined chemical rules.

2. Methodology: TERS-ABNet

The authors propose TERS-ABNet, a deep learning framework that treats single-molecule structure determination as an image-to-graph inference task.

A. Architecture

The model utilizes a "two-track" architecture based on 2D Attention U-Nets:

1. Atom-Prediction Network (ANet): Predicts the probability maps for atomic species (C, N, O, H).
2. Bond-Prediction Network (BNet): Predicts the probability maps for chemical bond types (e.g., C–H, C–C, C–N, etc.).

• Shared Encoder: Both networks share an encoder-decoder backbone with 12 convolutional layers.
• Attention Mechanisms: Three attention gates are integrated via skip connections to selectively emphasize spatial-spectral features relevant to specific atoms or bonds.
• Output Processing: The networks output multi-channel probability maps. A spot-detection algorithm extracts coordinates, and an algorithmic step determines connectivity to reconstruct the final molecular graph.

B. Data Generation and Training

• Dataset: Created from the QM9 and FG26 databases, containing ~5,800 planar molecules (C, N, O, H).
• Simulation: TERS spectra were simulated using first-principles calculations (Gaussian16) combined with electromagnetic near-field modeling (MNPBEM).
  • Input: 128 × 128 pixel grids covering a 2.5 × 2.5 nm² area.
  • Channels: 160 channels (integrated over 20 cm⁻¹ windows across 0–2000 and 2800–4000 cm⁻¹).
• Ground Truth: Encoded as Gaussian-shaped spots centered at true atomic/bond coordinates.
• Augmentation: Rotational augmentation (90°, 180°, 270°) and additive Gaussian noise were applied to ensure rotational equivariance and robustness.
• Transfer Learning: To handle non-planar molecules, the model was fine-tuned on a dataset including methyl/ethyl groups, expanding the BNet output to 11 channels.

3. Key Results

A. Prediction Accuracy (Simulated Data)

• Localization: Achieved a mean absolute error (MAE) of ~0.23 Å for atomic coordinates and ~0.20 Å for bond positions. 98.2% of atoms were localized within this tolerance.
• Classification:
  • Atom Type: ~94.2% accuracy (Hydrogen: highest; N/O: slightly lower due to backbone embedding).
  • Bond Type: ~92.0% accuracy.
• Completeness: The model successfully reconstructed complex motifs, including short chains, heteroaromatic rings, and fused heterocyclic systems, with high fidelity.

B. Generalization Capabilities

• Spatial Resolution Robustness: The model demonstrated zero-shot generalization across varying tip-molecule distances (simulating resolutions from 0.47 nm to 1.12 nm). Even at lower resolutions (1.12 nm), it could recover the overall molecular skeleton and principal connectivity, suggesting that machine learning can mitigate the need for strict Ångström-scale field confinement.
• Non-Planar Molecules: Through transfer learning, the model successfully reconstructed molecules with out-of-plane functional groups (e.g., methyl/ethyl substituents), maintaining accuracy despite signal attenuation and orientation-dependent selection rules.

C. Experimental Validation

• Case Study: The framework was applied to experimental TERS data of a Magnesium Porphyrin (MgP) molecule.
• Performance:
  • Successfully identified peripheral carbon atoms and hydrogen atoms.
  • Correctly reconstructed the macrocyclic ring structure and pyrrole subunits.
  • Limitations: The central Magnesium atom was misclassified (as O or C) due to its absence in the training set and strong spectral perturbation. Bond predictions for the conjugated core appeared as diffuse rings rather than localized spots, reflecting the physical reality of delocalized electronic states in large conjugated systems.

4. Key Contributions

1. Novel Framework: First deep learning approach to formulate TERS structure determination as a direct image-to-graph inference problem, bypassing manual spectral assignment.
2. Dual-Track Prediction: Simultaneous prediction of atomic species and bond types allows for the construction of explicit atom-bond graphs without predefined chemical rules.
3. Resolution Relaxation: Demonstrated that high-precision structural recovery is possible even with moderate spatial resolution (sub-nm), challenging the conventional requirement for Ångström-level near-field confinement.
4. Transfer Learning: Proved the framework's adaptability to non-planar and 3D molecular geometries, expanding the scope of TERS applicability.

5. Significance

• Automated Nanoscale Characterization: TERS-ABNet provides a pathway toward fully automated, quantitative single-molecule structure determination, removing the bottleneck of expert-dependent interpretation.
• Overcoming Physical Limits: By leveraging statistical correlations between spectroscopic signals and structural priors, the method effectively augments the information bandwidth of experimental measurements, allowing structural insights from data that would traditionally be considered too noisy or low-resolution.
• General Paradigm: The work establishes a general machine-learning paradigm for inferring explicit atom-bond representations from high-dimensional spectroscopic imaging, applicable beyond TERS to other scanning-probe-based spectroscopic platforms.
• Future Outlook: While current limitations exist for fully 3D molecules and highly delocalized systems (due to physical selection rules), this work highlights the potential of data-driven solutions to solve long-standing inverse problems in nanoscience.