An
extended theoretical benchmark set, NS372, for light main-group
nuclear shieldings and NMR shifts has been constructed based on high-level
GIAO-CCSD(T)/pcSseg-3//CCSD(T)/cc-pVQZ reference data. After removal
of the large static-correlation cases O3, F3
–, and BH from the statistical evaluations for
the 17O, 19F, and 11B subsets, the
benchmark comprises overall 372 shielding values in 117 molecules
with a wide range of electronic-structure situations, containing 124 1H, 14 11B, 93 13C, 43 15N,
31 17O, 47 19F, 14 31P, and 6 33S shielding constants. The CCSD(T)/pcSseg-3 data are shown
to be close to the basis-set and method limit and thus provide an
excellent benchmark to evaluate more approximate methods, such as
density functional approaches. This dataset has been used to evaluate
Hartree–Fock (HF) and MP2, and a wide range of exchange–correlation
functionals from local density approximation (LDA) to generalized
gradient approximations (GGAs) and meta-GGAs (focusing on their current-density
functional implementations), as well as global hybrid, range-separated
hybrid, local hybrid, and double-hybrid functionals. Starting with
absolute shielding constants, the DSD-PBEP86 double hybrid is confirmed
to provide the highest accuracy, with an aggregate relative mean absolute
error (rel. MAE) of only 0.9%, followed by MP2 (1.1%). MP2 and double
hybrids only show larger errors for a few systems with the largest
static-correlation effects. The double-hybrid B2GP-PLYP, the two local
hybrids cLH12ct-SsirPW92 and cLH12ct-SsifPW92, and the current-density
functional meta-GGA cB97M-V follow closely behind (all 1.5%), as do
some further functionals, cLH20t and cMN15-L (both 1.6%), as well
as B2PLYP and KT3 (both 2.0%). Functionals on the lower rungs of the
usual ladder offer the advantage of lower computational cost and access
to larger molecules. Closer examination also reveals the best-performing
methods for individual nuclei in the test set. Different ways of treating
τ-dependent functionals are evaluated. When moving from absolute
shielding constants to chemical shifts, some of the methods can benefit
from systematic error compensation, and the overall error range somewhat
narrows. Further methods now achieve the 2% threshold of relative
MAEs, including functionals based on TPSS (TPSSh, cmPSTS).