Implementation, Development and Assessment of Local Hybrid Density Functionals

In order to describe complex molecular systems theoretically, an efficient and reliable solution to the underlying quantum mechanical equations of motion is required. Density functional theory (DFT) represents in most cases the best compromise between accuracy and efficiency for the treatment of electronic interactions. In Kohn-Sham DFT, the non-classical contribution to electron-electron interactions is gathered in the exchange-correlation functional, which has to be approximated in practice. While a large number of exchange-correlation functionals are of semi-empirical nature, some have been derived from physical considerations exclusively. In so-called global hybrid functionals a constant amount of the integrated DFT exchange-energy density is replaced by the exact-exchange energy from Hartree-Fock theory. The most popular functional, B3LYP, contains 20 % exact exchange and several empirical parameters. It has been discovered that the optimal amount of exact exchange depends to a large extent on the molecular property to be computed. A possible solution to this problem is to use local hybrid functionals. Therein, the admixture of exact exchange is controlled by a position-dependent local mixing function (LMF), leading to molecule-specific amounts of exact exchange. In this work a semi-empirical approach is pursued for the development of new local hybrid functionals. Parameterized LMFs are introduced in the exchange-energy density integrals, for which the DFT contributions are taken from established approximations to the exchange-correlation functional. The LMFs developed here contain at least one empirical parameter and a variable that depends on the ratio of the von-Weizsacker single-particle kinetic energy density to the correlated kinetic-energy density (the so-called t-LMFs), or on the reduced density gradient (referred to as s-LMF). Additional LMFs are obtained by inclusion of the spin polarization. All parameters are fitted to atomization energies and reaction barriers of well-established test sets. Visualization of the LMFs provides an additional tool for analyzing their physical and chemical behavior, potentially leading to further developments. As a general trend, an increasing exact-exchange admixture is observed upon bond stretching for all LMFs, with a more pronounced effect for t-LMFs. This observation correlates with a better performance for reaction barriers of t -LMF-based local hybrid functionals. Most of the local hybrid functionals discussed in this work are based on the exchange and correlation functional from the local spin density approximation (LSDA) and contain therefore no gradient correction such as in the generalized gradient approximation (GGA). The new functionals were initially implemented non-self-consistently into a development version of the quantum chemical Turbomole program package. That is, only the total energy is calculated for a given set of molecular orbitals or electron density, respectively. This is a reliable approximation that allows for significant time savings especially during parameter optimizations. In order to calculate orbital-dependent molecular properties, the local hybrid potential corresponding to the local hybrid energy is required as well. It is obtained as a functional derivative of the exchange-correlation energy with respect to the orbitals. Some of the resulting integrals contain the LMF-weighted non-local exact-exchange potential. These terms as well as the exact-exchange energy density itself cannot be calculated analytically. Following a well-established approach, they have been approximated using a basis set expansion of the exact-exchange potential. For simplicity, the underlying atomic basis set is employed in this resolution of the identity (RI) approximation. For comparison and in view of the optimization of auxiliary basis sets, the optional calculation of the potential by numerical integration has also been implemented in this work. The computational cost of local hybrid calculations for a given basis set, using the RI approximation is comparable to the one of gobal hybrid functionals: a slightly larger prefactor applies to a calculation with a local hybrid functional as compared to a meta-GGA global hybrid, while the scaling of computational effort as a function of system size is the same. Several molecular test sets including atomization energies, barrier heights, dissociation energies and equilibrium distances have been considered for the assessment. Some of them represent particular challenges for current density functional approximations. All of the discussed local hybrid functionals yield significantly better results for the 223 atomization energies of the G3 test set than the B3LYP functional. Especially local hybrid functionals with spin-polarized t -LMFs gives impressively small mean absolute errors for the G3 set. Most of our functionals are in addition significantly superior to B3LYP for the calculation of barrier heights. Some other global hybrid functionals perform even better than our functionals for barriers, but their intrinsic amount of exact exchange is inappropriately high for thermochemical property calculations. For the first time, LSDA-based local hybrid functional have thus been presented that gives accurate results for thermochemistry and reaction barriers simultaneously. The dissociation behavior of symmetric radical cations remains a challenge for the local hybrid functionals presented here. Dissociation energies are significantly overestimated, and the equilibrium distances are too short. The results are overall only slightly better than those obtained using the B3LYP functional. A larger amount of exact exchange is most likely needed for these systems to reduce self-interaction errors. Additionally, the performance of local hybrid functionals for 3d transition metal dimers and monohydrides has been studied. An accurate description of dynamical and nondynamical correlation is essential for the former. The poor performance of most exchange-correlation functionals for transition metal monohydrides can be attributed to self-interaction errors. Our local hybrid functionals perform similarly to B3LYP for the dimers and marginally better for the monohydrides. They do not provide any improvement for the atomic s-d transfer energies of 3d metals. The most suitable local hybrid functional for this particular property uses a s-LMF in the exchange functional and the LYP correlation functional. It yields, however, only average-quality results for thermochemistry and kinetics. Satisfactory results similar to B3LYP are obtained for the isotropic hyperfine coupling constants (HFCCs) of small main group compounds with a t-LMF-based local hybrid functional. The RI approximation to the local hybrid potential has been validated by comparing it to the numerically exact potential for the calculation of total energies, isotropic HFCCs and orbital energies. The error in total energies due to the RI approximation is comparatively small considering the rather large deviations from experimental values. Comparison of mean absolute errors from experimental values of the 26 isotropic HFCCs reveals only small differences between the RI and the numerically exact local hybrid potential. Further analysis shows that inaccuracies in the RI potential may have a larger impact on the isotropic HFCCs or the orbital energies of a particular molecule, especially if only small or medium-sized basis sets are employed. Several of the local hybrid functionals are suitable for the calculation of thermochemical and kinetic properties. Different functionals yield also results similar to other commonly used functionals for isotropic HFCCs of small main group compounds, as well as for the dissociation energies and equilibrium distances of 3d transition metal dimers and monohydrides. The local hybrid functionals studied in this work represent therefore an important step towards the development of universal approximations to the exchange-correlation functional. For a more accurate description of certain transition metal properties and the dissociation behavior of symmetric radical cations while maintaining a good performance for thermochemistry and kinetics, more complex LMFs will have to be considered. Ultimately a local hybrid functional with meta-GGA exchange and correlation energy densities that fulfills more exact constraints is desirable. Therefore further studies on the different gauges of the exchange energy densities are necessary. Another possibility would be the development of a specifically designed correlation functional to be combined with a local hybrid exchange functional based on the LSDA. More detailed studies on the quality of the RI approximation are recommended. Possible properties for this purpose include, e.g., ionization energies and electron affinities. Auxiliary basis sets should be implemented and optimized for the expansion of the exact-exchange potential in order to avoid additional deviations due to the RI-approximation or even fortuitously good results in the assessment of local hybrid functionals with normally contracted basis sets. Since density functional methods are applied extensively for structure optimizations, the gradient of the local hybrid energy with respect to the nuclear coordinates should be implemented to enable this feature in future versions of the code.