NMIRacle: Multi-modal Generative Molecular Elucidation from IR and NMR Spectra

---

NMIRacle (NMR-IR oracle) is a multi-modal generative framework for de novo molecular structure elucidation from spectroscopic data. The model learns to generate molecular structures (SMILES) directly from raw IR, ¹H-NMR, and ¹³C-NMR spectra through a two-stage training approach that combines count-aware fragment encodings with multi-spectral conditioning.

Installation

📦 Data Preparation

NMIRacle expects data in the following structure:

datasets/
├── pretrain/                          # Stage 1: Fragment pre-training
│   ├── smiles.npy                     # Molecular SMILES strings
│   ├── substructure_counts.h5         # Fragment compositions
│   ├── split_indices.p                # Train/val/test splits
│
└── multispectra/                      # Stage 2: Spectra fine-tuning
    ├── smiles.npy                     # Molecular SMILES strings
    ├── spectra.h5                     # Multi-modal spectra (IR, ¹H, ¹³C)
    ├── substructure_counts.h5         # Fragment compositions
    └── split_indices.p                # Train/val/test splits

The spectra.h5 contains concatenated (in order):

• ir: IR spectra (1,800 features)
• hnmr: ¹H-NMR spectra (10,000 features)
• cnmr: ¹³C-NMR spectra (10,000 features)

Dataset files are available at the following link.

🎓 Training

Stage 1: fragments-to-molecule pre-training

Train the model to reconstruct molecules from count-aware fragment representations:

python -m nmiracle.train \
  training_stage=sub2struct \
  data/dataset=pretrain_dataset \
  trainer.max_epochs=500 \
  logger.wandb.enabled=true

Stage 2: spectra-to-molecule fine-tuning

Fine-tune with spectral conditioning (requires Stage 1 checkpoint):

python -m nmiracle.train \
  training_stage=spec2struct \
  data/dataset=multispectra_dataset \
  data.use_ir=true \
  data.use_hnmr=true \
  data.use_cnmr=false \
  model.pretrained_structure_model_path=/path/to/stage1/checkpoint.ckpt \
  trainer.max_epochs=300 \
  logger.wandb.enabled=true

🧪 Inference

Testing on fragments-to-molecule task

python -m nmiracle.test_sub2struct \
  --model_path nmiracle/ckpts/sub2struct \
  --checkpoint best.ckpt \
  --temperature 1.0 \
  --top_k 5 \
  --num_sequences 15

Testing on spectra-to-molecule task

python -m nmiracle.test_spec2struct \
  --model_path nmiracle/ckpts/spec2struct_ir_hnmr_cnmr \
  --checkpoint epoch=295-val_loss=0.15.ckpt \
  --temperature 1.0 \
  --top_k 5 \
  --num_sequences 15

We provide checkpoints of trained models at the following link.

---

🤝 Citation

If you use NMIRacle in your research, please cite:

@misc{ottomano2025nmiraclemultimodalgenerativemolecular,
      title={NMIRacle: Multi-modal Generative Molecular Elucidation from IR and NMR Spectra}, 
      author={Federico Ottomano and Yingzhen Li and Alex M. Ganose},
      year={2025},
      eprint={2512.19733},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2512.19733}, 
}

---

🙏 Acknowledgments

• Built with PyTorch and PyTorch Lightning
• Inspired by NMR2Struct and related work.

⭐ If you find NMIRacle useful, please star the repository!