Accuracy and Limitations of the Pair-Selected Multilevel Approach for DLPNO Coupled Cluster: Extensive Benchmark for...

Reliable approximations to coupled-cluster (CC) methods are highly desirable for accurate yet efficient computations of barrier heights, reaction energies, and other molecular properties. Among these methods, domain-based local pair natural orbital CC with singles, doubles, and perturbative triple excitations [DLPNO-CCSD(T)] is widely used due to its formal linear scaling with the system size. However, since DLPNO-CCSD(T) remains costly, the extension to multilevel (ML) variants becomes an obvious route. This strategy can be made even more economic with the pair-selected ML ansatz [M. Bensberg and J. Neugebauer, J. Chem. Phys. 157, 064102 (2022)] to DLPNO-CCSD(T₀) with a semi-canonical (SC) perturbative triples correction. This ansatz uses an automatic partitioning of orbital pairs according to their contribution to the overall correlation energy change in a chemical reaction. Herein, the advantages of this approach are demonstrated for closed-shell organic reactions of the BH9 test set. The errors are nearly always within chemical accuracy (4 kJ mol^-1) along with a significant time benefit. In rare cases, larger errors are observed. These are analyzed by comparison of SC and iterative perturbative triples, of different ML thresholds, and of ML and single-level schemes. A beneficial error cancelation between DLPNO and ML contributions is also observed in several cases.