How accurate predicted NMR shifts are

On data the model never saw, predicted 1H shifts are typically within 0.08 ppm and 13C shifts within 0.7 ppm. Each shift comes with a 90% range. Multiplicities, J values and carbon types are worked out from the structure, not predicted.

The note under the prediction: the model, its typical error on held-out data, how multiplicities are estimated, and the training data's source and license.The note under the prediction: the model, its typical error on held-out data, how multiplicities are estimated, and the training data's source and license.

The numbers

Nucleus Model Typical error on held-out data
¹H Athanor Resonantia H 0.08 ppm
¹³C Athanor Resonantia 0.7 ppm

"Held-out" means molecules kept out of training and used only to measure the error, so the figure is what to expect on a structure the model has not seen.

The 90% range

Every shift in the table has a 90% range: the span the true shift falls in nine times out of ten. A narrow range means the model is confident; a wide one means treat the value with care.

What is not predicted

  • Multiplicity and J for ¹H are Athanor's first-order estimate from typical couplings (vicinal 7 Hz, aromatic ortho 8 and meta 2, vinyl trans 16 and cis 10). Real spectra can show second-order effects these do not.
  • CH3, CH2, CH and C types for ¹³C come from the drawn structure.
  • Solvent: the model is trained on CDCl₃. Other solvents are listed but not predicted yet.

The data behind it

The model is trained on NMRexp (Wang and Zhu, Zenodo 10.5281/zenodo.17296666), licensed CC BY 4.0, as the note in the window says.

Prediction

A prediction is a guide for assigning and checking spectra, not a substitute for measuring one. Everything predicted is labeled as predicted, in the window, in the journal text and on the inserted spectrum.

Is this helpful?

Last updated Nov 12, 2026