A large-scale foundation model enables simulation-to-real adaptation for nuclear magnetic resonance-based molecular structure analysis

UltraNMR, a large-scale foundation model for Nuclear Magnetic Resonance (NMR) spectroscopy, has been introduced to enhance molecular structure analysis by leveraging 158 million paired simulated $^{1}$H and $^{13}$C NMR spectra for training. It employs multiple domain-specific pre-training objectives to learn generalizable spectral representations, enabling effective simulation-to-real adaptation and achieving state-of-the-art performance across various experimental NMR tasks. This model also supports the construction of a comprehensive NMR spectral vector library, facilitating structure-aware retrieval for 94 million unique molecules, which is significant for practitioners aiming to improve the robustness and generalization of AI applications in molecular analysis.