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

Researchers introduced UltraNMR, a foundation model trained on 158 million simulated nuclear magnetic resonance spectra that successfully bridges the gap between simulation and real-world molecular analysis. The model demonstrates state-of-the-art performance on experimental NMR tasks and has been applied to identify previously unknown natural products from Chinese herbal medicines, suggesting large-scale simulation pre-training can enable robust generalization in spectroscopy.

UltraNMR represents a significant advancement in applying foundation models to scientific instrumentation and molecular analysis. The research tackles a persistent challenge in computational chemistry: the scarcity of experimental NMR datasets limits deep learning applications to narrow, task-specific problems. By leveraging 158 million simulated spectra paired with domain-specific pre-training objectives, the researchers created a model capable of capturing both intra-spectral and inter-spectral dependencies, effectively solving the simulation-to-real adaptation problem that has constrained prior approaches.

The foundation model approach mirrors recent successes in large language models and vision transformers, translating the scaling paradigm to scientific domains. The authors constructed a 94-million-molecule spectral vector library, enabling structure-aware retrieval and positioning UltraNMR as infrastructure for downstream molecular discovery applications. This library extends the model's utility beyond direct spectral interpretation toward drug discovery, materials science, and natural product research workflows.

The real-world validation through structural elucidation of unknown natural products demonstrates practical value beyond benchmark performance. For the pharmaceutical, chemistry, and materials science industries, this suggests automation of structure determination—historically a labor-intensive bottleneck—becomes feasible at scale. The approach also establishes a template for other spectroscopic techniques and analytical methods constrained by limited experimental datasets.

Future development should focus on extending UltraNMR to multi-modal spectroscopy integration and addressing edge cases in real experimental conditions. The scalability of simulation-based pre-training opens opportunities for similar foundation models across chromatography, mass spectrometry, and other analytical instrumentation, potentially reshaping how structural chemistry research is conducted.

• →UltraNMR achieves state-of-the-art performance on experimental NMR tasks using simulation-based pre-training on 158 million spectral pairs.
• →The model enables rapid structural elucidation of unknown natural products, demonstrating practical pharmaceutical and materials science applications.
• →A 94-million-molecule spectral vector library enables structure-aware molecular retrieval, creating infrastructure for downstream discovery workflows.
• →Simulation-to-real adaptation successfully bridges the experimental data scarcity problem that previously limited deep learning in spectroscopy.
• →The foundation model approach establishes a scalable template for other analytical techniques constrained by limited experimental datasets.