NMR spectroscopy undoubtedly plays a central role in
determining
molecular structures across different chemical disciplines, and the
accurate computational prediction of NMR parameters is highly desirable.
In this work, a new Δ-machine learning approach is presented
to correct DFT-computed NMR chemical shifts using input features from
the calculation and in addition highly accurate reference data at
the CCSD(T)/pcSseg-2 level of theory with a basis set extrapolation
scheme. The model is trained on a data set containing 1000 optimized
and geometrically distorted structures of small organic molecules
comprising most elements of the first three periods and containing
data for 7090 1H and 4230 13C NMR chemical shifts.
Applied to the PBE0/pcSseg-2 method, the mean absolute deviation (MAD)
on the internal NMR shift test set is reduced by 81% for 1H and 92% for 13C at virtually no additional computational
cost. For 12 different DFT functional and basis set combinations,
the MAD of the ML-corrected NMR shifts ranges from 0.021 to 0.039
ppm (1H) and from 0.38 to 1.07 ppm (13C). Importantly,
the new method consistently outperforms the simple and widely used
linear regression correction technique. This behavior is reproduced
on three different external benchmark sets, confirming the generality
and robustness of the correction scheme, which can easily be applied
in DFT-based spectral simulations.