Improved correlation energy extrapolation schemes based on local pair natural orbital methods

It is well-known that the basis set limit is difficult to reach in correlated post Hartree-Fock ab initio calculations. One possible route forward is to employ basis set extrapolation schemes. In order to avoid prohibitively expensive calculations, the highest level calculation (typically based on the "gold standard" coupled cluster theory with single, double, and perturbative triple excitations, CCSD(T)) is only performed with the smallest basis set, and the remaining basis set incompleteness is estimated at a lower level of theory, typically second-order M?ller-Plesset perturbation theory (MP2). In this work, we provide a comprehensive investigation of alternative schemes where the MP2 extrapolation is replaced by the coupled-electron pair approximation, version 1 (CEPA/1) or the local pair natural orbital version of this method (LPNO-CEPA/1). It is shown that the MP2 method achieves apparent accuracy only due to error cancellation. Systematically more accurate results at small additional computational cost are obtained if the MP2 step is replaced by LPNO-CEPA/1. The errors of LPNO-CEPA/1 relative to canonical CEPA/1 are negligible. Owing to the highly systematic nature of the deviations between canonical and LPNO methods, basis set extrapolation reduces the LPNO errors in the total energies by 1 order of magnitude (~0.2 kcal/mol) and errors in energy differences to essentially zero. Using the CCSD(T)/LPNO-CEPA/1-based extrapolation scheme, new reference values are proposed for the recently published S66 set of interaction energies. The deviations between the new values and the original interactions energies are mostly very small but reach values up to 0.3 kcal/mol.