Show HN: Rose – reusable foundation embeddings for industrial 1H NMR (github.com)

A new project named Rose introduces a pretrained ¹H NMR foundation model aimed at enhancing the analysis of nuclear magnetic resonance (NMR) spectra. This model facilitates various tasks, including denoising noisy spectra, peak prediction, and structure retrieval based on spectral data. By utilizing a standardized approach with frozen embeddings and offering flexible adaptation protocols, Rose empowers researchers to leverage transfer learning effectively across different domains of NMR analysis. Key features include the ability to clean spectra through zero-shot learning and the implementation of contrastive retrieval processes translating spectral data into chemical structures. The significance of Rose lies in its potential to streamline analytical workflows in chemical research and industrial applications, where precise spectral interpretation is crucial. The model, consisting of 7.8 million parameters trained on 3.2 million spectra, demonstrates impressive performance metrics—such as an 87.8% top-1 accuracy for spectrum pairing and a balanced accuracy of 98.8% for edible oil analysis. By providing flexible access to pretrained weights on platforms like Hugging Face and simplifying user interactions through straightforward coding protocols, Rose not only advances the technical capabilities of NMR analysis but also fosters wider adoption and innovation in AI-driven chemical research.