PREPRINT (not peer-reviewed). A transformer generated the correct molecular structure among 15 candidates for 34.8% of spectrum pairs in a held-out test, rising to 56.6% for molecules with up to 19 heavy atoms. It uses one-dimensional proton and carbon NMR peak lists plus molecular weight. On a separate benchmark of 86 experimental spectra, it reached 20.6% accuracy with 1,000 candidates versus 4.8% for NMR2Struct. That comparison does not isolate the effect of training data: the models also differ in inputs, architecture, stereochemistry handling and preprocessing. Accuracy fell with molecular size, and the authors note that the benchmark is small and published peak lists may already be curated alongside assigned structures. Find the paper on ChemRxiv by searching its full title: A generative neural network architecture for structure elucidation.
12:39:39 AM · Nov 23, 2026
Sign in to reply and join the conversation.
