Multilevel Approaches within the Local Pair Natural Orbital Framework.

The linear-scaling local coupled cluster method DLPNO-CCSD(T) allows calculations on systems containing hundreds of atoms to be performed while reproducing canonical CCSD(T) energies typically with chemical accuracy (<1 kcal/mol). The accuracy of the method is determined by two main truncation thresholds that control the number of electron pairs included in the CCSD iterations and the size of the pair natural orbital virtual space for each electron pair, respectively. While the results of DLPNO-CCSD(T) calculations converge smoothly toward their canonical counterparts as the thresholds are tightened, the improved accuracy is accompanied by a fairly steep increase of the computational cost. Many applications study events that are confined to a relatively small region of the system of interest. Hence, it is viable to develop methods that allow the user to treat different parts of a large system at various levels of accuracy. In this work we present an extension to the native DLPNO method that fragments the system into preselected molecular parts and uses different thresholds or even different levels of theory for the interaction between individual fragments. Thereby chemical intuition can be used to focus computational resources on a more accurate evaluation of the properties at the center of interest, while permitting a less demanding description of the surrounding moieties. The strategy was implemented within the DLPNO-CCSD(T) framework. We tested the scheme for a series of realistic quantum chemical applications such as the calculation of the dimerization energies, potential energy surfaces, enantiomeric excess in organometallic catalysis, and the binding energy of the anticancer drug ellipticine to DNA. This work demonstrates the power of the approach and offers guidance to its setup.