Reproduce “Unraveling Molecular Structure: A Multimodal Spectroscopic Dataset for Chemistry” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Marvin Alberts, Oliver Schilter, Federico Zipoli, Nina Hartrampf, Teodoro Laino
The paper introduces the first large-scale, multimodal spectroscopic dataset for chemistry, comprising simulated IR, NMR (¹H, ¹³C, HSQC), and Mass (positive/negative) spectra for over 790,000 unique molecules. This dataset aims to enable the development of AI models capable of integrating diverse spectroscopic information for automated structural elucidation, demonstrating initial benchmarks with up to 73.38% Top-1 accuracy for combined ¹H- and ¹³C-NMR in structure elucidation.
Key insights
- Simulated multimodal spectroscopic data, even with inherent 'sim-to-real' distribution shifts, can be used to train AI models that capture the complementary nature of different techniques.
- Transformer models, when provided with structured text representations of spectra, show strong performance in tasks like structure elucidation, functional group prediction, and spectra prediction.
- Combining information from multiple modalities (e.g., ¹H- and ¹³C-NMR) significantly improves structural elucidation accuracy, mirroring human expert workflows.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.