Supplementary Data: Multi-Modal Contrastive Learning for Chemical Structure Elucidation with VibraCLIP

VibraCLIP is a multi-modal machine learning framework designed to align molecular graph representations with vibrational spectra, such as IR and Raman spectra, using contrastive learning strategies. By learning a shared embedding space for molecular structures and their corresponding spectral signatures, VibraCLIP enables accurate retrieval, analysis, and understanding of vibrational data for chemical systems.

This supplementary data repository contains the datasets, model checkpoints, and trained embeddings associated with the VibraCLIP project. The main objectives are:

To map molecular graphs and vibrational spectra into a common latent space.

To enhance molecule-spectrum retrieval and cross-modal understanding.

To explore the relationships between molecular structure and vibrational signatures using deep learning.

Contents

Datasets: Pre-processed molecular graphs and corresponding IR and Raman spectra. These are in different formats such as pickle and LMDB files.

Model Checkpoints: Trained model weights for reproducibility and further fine-tuning.

Retrieval Accuracy Files: Pickle files from the retrieval accuracy callbacks of the different experiments for reproducibility and visualization with the provided notebooks in the GitHub repository.