| |
[摘要]：The performance of 250 different computational protocols (combinations of density functionals, basis sets and methods) was assessed on a set of 165 well-established experimental H-1-H-1 nuclear coupling constants (J(H-H)) from 65 molecules spanning a wide range of "chemical space". Thereby we found that, if one uses core-augmented basis sets and allows for linear scaling of the raw results, calculations of only the Fermi contact term yield more accurate predictions than calculations where all four terms that contribute to J(H-H) are evaluated. It turns out that B3LYP/6-31G(d,p)u+1s is the best (and, in addition, one of the most economical) of all tested methods, yielding predictions of J(H-H) with a root-mean-square deviation from experiment of less than 0.5 Hz for our test set. Another method that does similarly well, without the need for additional Is basis functions, is B3LYP/cc-pVTZ, which is, however, ca. 8 times more "expensive" in terms of CPU time. A selection of the better methods was tested on a probe set comprising 61 J(H-H) values from 37 molecules. In this set we also included five molecules where conformational averaging is required. The rms deviations were better than or equal to those with the training set, which indicates that the method we recommend is generally applicable for organic molecules. We give instructions on how to carry out calculations of H-1 chemical shifts and J(H-H) most economically and provide scripts to extract the relevant information from the outputs of calculations with the Gaussian program in clearly arranged form, e.g., to feed them into programs for simulating entire H-1 NMR spectra. |
|
