Model Predicts Molecular Energy Gap From NMR Data

Yale researchers built a machine learning tool that estimates a key molecular property using nuclear magnetic resonance spectra.

Updated on Nov. 2, 2026 in Chemistry

A tall cylindrical NMR magnet housing with an open bore stands in a quiet laboratory, with cold vapor rising from its base and sample tubes in the foreground. — Yale researchers built a machine learning model that predicts the Homo-Lumo energy gap of molecules from nuclear magnetic resonance spectra, trained on more than 100,000 small molecules. AI Illustration. Upload story photo >

Yale University researchers developed a machine learning model that predicts the Homo-Lumo energy gap of molecules from NMR spectral data. The model was trained on NMR data from more than 100,000 small molecules.

Why it matters

The Homo-Lumo energy gap is a key determinant of chemical behaviour and physical properties such as optical absorption and conductivity. Because chemical shift data reflect local electron density and wider molecular interactions, the approach could offer a faster route to estimating these properties.

The model was trained on over 100,000 small molecules using eight features derived from chemical shift data, while some comparison models rely on about 400 features. The reported performance drops for sulfur-, boron-containing molecules and zwitterions, and the model was principally trained on computational rather than experimental NMR data.

The players

Yale University

A private Ivy League research university in New Haven, Connecticut, whose chemistry researchers developed the model.

Kevin Jablonka

A researcher at Friedrich Schiller University Jena in Germany who commented on the work.

The details

The researchers split chemical shift data into features such as the number of peaks and the mean, maximum and minimum shift values, then fed these into a mathematical model to predict the Homo-Lumo gap. The model separated molecules that absorb visible light from those that absorb ultraviolet light, and the most downfield carbon peak played a large role in predicting reactivity.

The Big Picture

The work extends the use of computational NMR chemical shift datasets beyond structure identification toward predicting electronic properties such as the Homo-Lumo gap. Its reduced performance on sulfur-, boron-containing molecules and zwitterions marks the current limits of this extension.

If the approach matures, researchers could estimate molecular energy gaps from routine NMR measurements rather than relying solely on costly calculations. The team has said it plans user-friendly, chemistry-adjacent applications, though no timeline has been given.

The takeaway

Combining routine NMR spectra with machine learning could make electronic-property estimates faster for chemists. Its current gaps with certain elements and molecule types mean it is a tool still in development.