Examining the performance of DFT methods in uranium chemistry: does core size matter for a pseudopotential?

We have investigated the performance of DFT in U(VI) chemistry. A large, representative selection of functionals has been tested, in combination with two ECPs developed in Stuttgart that have different-sized cores (60 and 78 electrons for U). In addition, several tests were undertaken with another 14 electron pseudopotential, which was developed in Los Alamos. The experimental database contained vibrational wavenumbers, thermochemical data, and (19)F chemical shifts for molecules of the type UF(6-n)Cl(n). For the prediction of vibrational wavenumbers, the large-core RECP (14 electrons) gives results that are at least as good as those obtained with the small-core RECP (32 electrons). GGA functionals are as successful as hybrid GGA for vibrational spectroscopy; typical errors are only a few percent with the Stuttgart pseudopotentials. For thermochemistry, hybrid versions of DFT are more successful than GGA, LDA, or meta-GGA. Marginally better results are obtained with a 32 electron ECP than with 14; since the experimental uncertainties are at least 25 kJ/mol for each reaction, the best functionals give results that are essentially indistinguishable from experiment. However, large-basis CCSD(T) results match experiment better than any DFT that we examined. Our findings for NMR spectroscopy are rather disappointing; no combination of pseudopotential, functional, and basis yields even a qualitatively correct prediction of trends in the (19)F chemical shifts of UF(6-n)Cl(n) species. Results yielded by the large-core RECP are, in general, slightly less bad than those obtained with the small core. We conclude that DFT cannot be recommended for predictions of NMR spectra in this series of compounds, though this conclusion should not be generalized. Our most important result concerns the good performance of the large-core Stuttgart pseudopotential. Given its computational efficiency, we recommend that it be used with DFT methods for the prediction of molecular geometries, vibrational frequencies, and thermochemistry of a given oxidation state. The hybrid GGA functionals MPW1PW91 and PBE0 give the best results overall.