Nuclear magnetic resonance (NMR) chemical shifts are one of the most useful spectroscopic observables in chemistry. In combination with X-ray diffraction experiments and chemical shift prediction models, Ångstrom resolution crystal structures can be routinely generated. However, there is often a significant computational cost to evaluate NMR predictions for multiple candidate structures or even time-averaged models. While the popular density functional theory (DFT) based gauge including projector augmented wave (GIPAW) is the de facto method for chemical shift prediction, it often generates errors that are larger than the experimental uncertainty for NMR experiments. In addition, popular machine learning (ML) models to circumvent this computational outlay are often based on GIPAW training data, thus creating additional prediction errors. To overcome the limits of accuracy for GIPAW, we present a simple molecular correction to GIPAW that improves chemical shift predictions for molecular crystals that is also insensitive to basis set size, ensuring that the correction is added at a minimal computational cost. We first benchmark our method on a curated crystal dataset with experimental shifts achieving mean absolute of errors for 13C, 15N, and 17O of 0.8, 2.8, and 4.3 ppm, respectively. We also present examples using the simple molecular correction on a number of real-world cases that improve or correct chemical shift predictions. Lastly, we present a Δ-machine learning (Δ-ML) approach as an alternative to other ML shift prediction models that significantly outperforms the accuracy of state-of-the art ML models. Notably, this work shows how one does not need to sacrifice higher accuracy DFT predictions while being 1–2 orders of magnitude cheaper. Furthermore, we show Δ-ML models can be used as an ensemble to evaluate the quality of predictions, which other ML shift predictions often lack. This model is finally tested on nine real world solution-phase 13C NMR spectra to demonstrate it’s predictive power and other possible further applications.

Improving the accuracy and efficiency of nuclear magnetic resonance chemical shift predictions for molecular crystals

ACS Spring Meeting 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.