Automatic orbital pair selection for multi-level local coupled-cluster based on orbital maps

We present an automatic, orbital-map based orbital pair selection scheme for multi-level local coupled-cluster approaches that exploits the locality of chemical reactions and concentrates on the part of the molecule directly involved in the reaction. The previously introduced pair-selected multi-level extension to domain-based local pair natural orbital coupled-cluster with singles, doubles, and semi-canonical perturbative triples [DLPNO-CCSD(T0)] partitions the orbital pairs according to relative changes in pair correlation energies [J. Chem. Phys. 157, 064102 (2022)]. To this end, maps between localized orbitals are required which in turn require maps between the atoms of structures along reaction paths. So far, these atom maps have been manually determined, which can be a (human) time-consuming procedure. In order to automatize this procedure, we present an atom mapping algorithm based on the principle of minimum chemical distance that incorporates the orientation through dihedral angles. Within the methodology of this algorithm, linear assignment problems are addressed and explicitly solved, giving rise to the idea of considering the orbital mapping problem as such. This led to a new strategy to obtain orbital maps that proves advantageous over the previously used direct orbital selection. Along with a modified orbital pair prescreening, this results in an improved variant of the pair-selected multi-level DLPNO-CCSD(T0) method. For given combinations of pair selection thresholds, the performance of this approach is demonstrated for various reaction types showing a significant efficiency gain and accurate results due to beneficial, systematic error cancellation. The presented method operates in a black-box manner due to its fully-automatized algorithms with only the need to specify a single target-accuracy parameter. Core orbitals can be included in the correlation treatment in a computationally cheap way such that there is no need for the frozen-core approximation. Additionally, we demonstrate that basis set extrapolation techniques can be applied. In this context, the approach shows deficiencies for the use of large basis sets, especially with diffuse basis functions, which can be traced back to the semi-canonical triples correction.