Every organic chemist has been there: staring at an NMR spectrum at 2 AM, coffee going cold, trying to figure out what molecule is producing that infuriating cluster of peaks between 7.2 and 7.4 ppm. You know it's aromatic. You know there's substitution happening. But which substitution pattern? And why do these peaks overlap like rush-hour traffic on a single-lane highway? Manual NMR structure elucidation can eat up eight hours of your life per compound - sometimes weeks for complex molecules - and it demands the kind of deep expertise that takes years to build and seconds to second-guess. For something so fundamental to organic chemistry, the process has remained stubbornly manual for decades.

A team led by Yongqi Jin and colleagues just dropped a paper in Nature Communications that might finally change the math on this problem. Meet NMR-Solver: a framework that automates molecular structure determination from plain old 1D proton and carbon-13 NMR spectra.

How Does This Thing Actually Work?

NMR-Solver runs four modules in a closed loop, like a well-organized lab group where everyone actually does their job:

Module 1 - NMRNet: An SE(3)-equivariant Transformer (fancy geometry-aware neural network) that predicts chemical shifts with mean absolute errors of 0.181 ppm for proton and 1.098 ppm for carbon-13 NMR. That's roughly on par with density functional theory calculations, except it runs several orders of magnitude faster. DFT is the gold standard, but NMRNet is the gold standard that doesn't need a supercomputer and a long weekend.

Module 2 - The Database: They built a spectral library from approximately 106 million molecules pulled from PubChem. Each one gets NMRNet-predicted chemical shifts, then gets indexed with FAISS vector search for sub-second retrieval. Think of it as Shazam, but for molecules instead of songs.

Module 3 - Fragment Optimization (FB-MO): This is the clever bit. When the database doesn't cough up an exact match, FB-MO runs a directed evolutionary strategy - chopping candidate molecules into fragments and recombining them, guided by atomic-level spectrum correlations. Each iteration can generate up to a billion candidate combinations, filtered through spectral matching. It's molecular Lego, but the instruction manual is written in NMR peaks.

Module 4 - Scenario Adaptation: Users can feed in constraints like known reactants, molecular formulas, or scaffold hints. Because sometimes you do know your starting material, and telling the algorithm "it definitely has a benzene ring" saves everyone a lot of time.

The Numbers (Reviewer 2 Would Approve)

On simulated benchmarks combining proton and carbon-13 NMR with molecular formula constraints, NMR-Solver hit 66.9% top-1 accuracy and 89.9% top-10. Respectable, but the real test is messier data. On 450 experimental reactant-product pairs from JACS 2024 papers, NMR-Solver scored 60.2% top-1 recall when given reactant information - compared to a previous method (NMR-to-Structure) that managed just 14.4%. That's not an incremental improvement; that's a different zip code.

Perhaps more impressive: the team used NMR-Solver to independently identify and correct two misassigned structures from peer-reviewed publications. The algorithm literally fact-checked published papers. Somewhere, Reviewer 2 is smiling.

Why Should You Care?

The combinatorial explosion here is staggering. For molecules with just 17 heavy atoms (carbon, nitrogen, oxygen, sulfur, halogens), there are roughly 166 billion possible structures. By 21 heavy atoms, that number exceeds 20 trillion. No human brain navigates that search space efficiently, no matter how many years of training or how much coffee. This is exactly the kind of problem where computational methods shine - pattern matching at inhuman scale.

The timing matters too. Automated chemistry labs are becoming real. Robotic synthesis platforms can run reactions around the clock, but they generate NMR data faster than any human can interpret it. NMR-Solver slots into that pipeline as the analytical bottleneck-buster, offering both fully automated analysis and a human-in-the-loop mode for when you want to sanity-check the machine's suggestions. If you've ever wished you could outsource the "stare at the spectrum" part of your workflow to something tireless and fast, this is that something.

It's worth noting the limitations honestly: strained ring systems gave NMRNet trouble, and the method still works best when molecular formula information is available. This isn't replacing spectroscopists overnight - it's more like giving them a very capable research assistant who never sleeps and has memorized 106 million molecules.

The Bigger Picture

NMR-Solver joins a growing wave of AI-for-spectroscopy tools, including multi-modal approaches combining IR and NMR and transformer-based methods for direct structure prediction. The field is converging on something genuinely useful: spectroscopy that interprets itself. For anyone who's spent a career squinting at peak tables and coupling constants, that future can't arrive fast enough.

The code is available on GitHub, and the team provides a web application for interactive use. The authors performed ablation studies, because apparently one experiment is never enough for Reviewer 2.

References

1. Jin, Y., Wang, J.-J., Xu, F., Ji, X., Gao, Z., Zhang, L., Ke, G., Zhu, R., & E, W. (2026). NMR-Solver: automated structure elucidation via large-scale spectral matching and physics-guided fragment optimization. Nature Communications. DOI: 10.1038/s41467-026-71315-0. arXiv: 2509.00640

2. Howarth, A., Ermanis, K., & Goodman, J.M. (2024). Accurate and efficient structure elucidation from routine one-dimensional NMR spectra using multitask machine learning. ACS Central Science, 10(12). DOI: 10.1021/acscentsci.4c01132. PMCID: PMC11613330

3. Li, J., et al. (2025). NMRexp: A database of 3.3 million experimental NMR spectra. Scientific Data. DOI: 10.1038/s41597-025-06245-5

4. Zhang, X., et al. (2025). Artificial intelligence in spectroscopy: advancing chemistry from prediction to generation and beyond. IJCAI 2025. arXiv: 2502.09897

5. Chen, Y., et al. (2026). NMRTrans: Structure elucidation from experimental NMR spectra via set transformers. arXiv: 2602.10158

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.