Bimodal Learning from Imperfect NMR Data (BLIND): A chemistry-ignorant model for autonomous structure elucidation from NMR

<jats:p>Automatically inferring molecular structure from analytical measurements is a key step toward closing the loop in autonomous chemical research. Yet it remains challenging, because structurally relevant information is distributed across multiple channels, affected by experimental imperfections, and often insufficient to uniquely define a molecular structure. Here, we present BLIND (Bimodal Learning from Imperfect NMR Data), a bimodal transformer that translates 1H and/or 13C NMR data into molecular structures without any chemical priors, such as molecular formula, elemental composition, or possible fragments. Unlike previous approaches, BLIND is trained exclusively on more than 5.5 million experimental spectra extracted from the chemical literature, learning from the inherent heterogeneity of real-world NMR data. Under fully unconstrained, chemistry-ignorant conditions, BLIND correctly infers the molecular structure, including stereochemistry, in more than half of the chemically diverse test cases and otherwise narrows the solution to only a few plausible candidates. On an independent expert-curated dataset acquired under controlled experimental conditions, the prediction accuracy exceeds 80%. BLIND handles various molecular sizes and structural complexity (up to 130 non-hydrogen atoms) and generalizes across a broad chemical space, including compounds with less common elements, multiple stereocentres, and multiple chemical species. These findings demonstrate that experimental data are a valuable source of structural information that should not be blindly replaced by idealized simulated datasets. Trained directly on such data, BLIND enables end-to-end structure elucidation from routine NMR measurements, allowing inferred structures to be fed back into real-time synthetic decisions in autonomous chemical research.</jats:p>