* This dataset supports online use.Click here to jump.

Multimodal Spectroscopic (Chemical Multimodal Spectroscopy) was created in 2024 by a research team from IBM Research, University of Zurich, EPFL and NCCR Catalysis. The related paper results are "Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry", which has been accepted by NeurIPS.

The dataset contains simulated 1H-NMR, 13C-NMR, HSQC-NMR, infrared and mass spectrometry (positive and negative ion modes) spectral data of 790,000 molecules extracted from chemical reactions in patent data. The core value of this dataset lies in its ability to integrate information from multiple spectral modalities and simulate the method of human experts analyzing molecular structures, which is expected to automate structural analysis and simplify the molecular discovery process from synthesis to structure determination.

The dataset was constructed taking into account the complementarity between different spectroscopic techniques, such as nuclear magnetic resonance (NMR), infrared spectroscopy, and mass spectrometry, which can provide different perspectives on the molecular structure, including the presence or absence of functional groups. By combining this information, researchers can gain a deeper understanding, which is critical for developing AI/ML models that can integrate information from multiple spectral modalities.

In addition, the Multimodal Spectroscopic dataset also provides benchmarks for evaluating single modality tasks, such as structure elucidation, spectral prediction of target molecules, and functional group prediction. These benchmarks not only help evaluate the performance of models, but also provide clear directions for future research.