Diverging errors: A comparison of DFT and machine-learning predictions of NMR shieldings

Abstract

Accurate prediction of NMR parameters from first principles is essential for the structural characterization of molecular solids. Recent studies have shown that single-molecule correction schemes—based on hybrid DFT calculations—can significantly improve the accuracy of periodic DFT predictions of nuclear shieldings. Here, we evaluate the performance of this correction approach not only for periodic DFT calculations but also for ShiftML2, a machine-learning model trained on PBE-calculated NMR data. For ¹³C nuclei, the application of single-molecule PBE0 corrections to periodic PBE shieldings has reduced the root-mean-square deviation (RMSD) from 2.18 to 1.20 ppm, with negligible improvement observed for ¹H. When applied to ShiftML2 predictions, the corrections have yielded a smaller reduction in ¹³C RMSD (from 3.02 to 2.51 ppm); again, they have had minimal impact on ¹H predictions. Residual analysis has revealed weak correlation between DFT and ML errors, suggesting that while some sources of systematic deviation may be shared, others are likely distinct. These results demonstrate that DFT-specific correction schemes do not straightforwardly translate to machine-learning models, highlighting the need for ML-tailored post-processing or retraining strategies. The findings have important implications for the integration of machine learning into high-throughput NMR workflows and the development of more accurate predictive tools for solid-state spectroscopy.