Machine Learning for Tip-Enhanced Raman Spectroscopy

Interpreting TERS Maps of π-Conjugated Molecules

Quick links:
• Portfolio Summary (PDF)
• All Figures Bundle (PDF)

---

Introduction

Tip-Enhanced Raman Spectroscopy (TERS) provides chemically specific imaging at (near) atomic resolution, but interpreting spectra/maps is difficult due to mode mixing, broadening, probe–molecule coupling, and substrate effects.
This repo explores machine learning (ML) to automate interpretation for π-conjugated molecules (benzene, naphthalene, anthracene, pyrene, coronene, small graphene nanoflakes). We pair 1D spectral classification with 2D TERS-map recognition, aiming to bridge spectroscopy and AI.

---

Data (synthetic & reproducible)

• 1D spectra (Raman-like): start from discrete “sticks” (frequency–intensity), apply Gaussian broadening, then min–max normalize.
• 2D TERS-like maps: synthesize spatial hotspots (Gaussian fields) + weak background + noise; include flips/rotations/blur for augmentation.
• Splits: stratified train/val/test; all indices and metadata saved in results/tables/.
• Generators live in src/dataset.py; examples in notebooks/01_data_exploration.ipynb and 02_generate_ters_maps.ipynb.

See statistics.md for all the equations we use (broadening, normalization, attention, losses, metrics, etc.).

---

Methods

We implement two complementary pipelines:

1. Spectral fingerprints (1D)

   • Input: synthetic Raman-like spectra (vectors).
   • Model: lightweight CNN-1D.
   • Task: molecular classification.

2. TERS imaging (2D)

   • Input: simulated TERS-like intensity maps (images).
   • Model: Vision Transformer (ViT) (tiny, patch-16).
   • Task: molecular classification from TERS maps.

Physics-aware option: a small Raman-tensor symmetry penalty can be added to the loss (see statistics.md).

---

Current Results (baseline)

From the saved artifacts (see PDFs above):

• CNN-1D (spectra): Accuracy 0.174, Macro-F1 0.049
• ViT (TERS maps): Accuracy 0.164, Macro-F1 0.113

Artifacts include training curves, confusion matrices, and per-class tables in results/tables/ and figures/.

These are baseline synthetic settings (small models, simple augmentations); metrics are expected to rise with more data, stronger aug, and tuning.

---

Figures

Key figures live in figures/ (see the All-Figures PDF for a full index). Typical outputs:

• Loss/Accuracy/F1 curves for CNN & ViT.
• Confusion matrices for test splits.
• Sample spectra overlays and TERS-map montages.

---

How to Run

Colab (recommended):

Notebook order:

1. 00_colab_setup.ipynb — environment & project scaffold
2. 01_data_exploration.ipynb — build & inspect 1D spectra dataset
3. 02_generate_ters_maps.ipynb — synthesize 2D TERS-like maps
4. 03_train_cnn.ipynb — train CNN-1D on spectra
5. 04_train_vit.ipynb — train ViT on TERS maps
6. 05_results_visualization.ipynb — re-plot curves/CM; merge metrics
7. 06_summary_report.ipynb — export Portfolio Summary (PDF/HTML/MD)

Local (optional):

pip install -r requirements.txt
python -c "import torch, timm; print('OK', torch.__version__)"
# then run notebooks in order with Jupyter/VSCode

---

Discussion & Next Steps

• Baselines show the ViT is already learning spatial cues; CNN-1D excels when peaks are sharp and well separated.
• To improve: increase samples/class, diversify hotspot geometry/backgrounds, add RandAugment, tune LR/weight decay/DropPath.
• Add attention/Grad-CAM visualizations and ablation (noise, patch size, #hotspots).
• Medium-term: domain adaptation to experimental TERS; incorporate GNNs for structure-aware features; uncertainty estimation.

---

References (selection)

• O. J. Silveira et al., Raman tensors in low-symmetry 2D materials, J. Raman Spectrosc. (2021).
• O. J. Silveira et al., Local probe-induced structural isomerization, Nat. Commun. (2023).
• O. J. Silveira et al., Frustration-Induced Many-Body Degeneracy in Spin-1/2 Molecular Quantum Rings, JACS (2025).
• T. Schütt et al., SchNet: Deep Learning for Quantum Chemistry, JCTC (2018).

---

Repo Map

src/                 # dataset, models, metrics, viz
notebooks/           # 00..06 workflow (Colab-first)
figures/             # spectra, ters_maps, curves, confusion_matrices, ...
results/             # tables, checkpoints, reports
docs/                # portfolio-summary.pdf, all-figures-portfolio.pdf
scripts/             # helpers (e.g., export_all_figures_pdf.py)
statistics.md        # equations (math/physics/stats used)
paperwork            #draft paper from this research

---

License

MIT License

Paperwork

(Paperwork TersML (PDF))