Improving "Silver-Standard" Benchmark Interaction Energies with Bond Functions.

We investigate the effect of adding midbond basis functions on the performance of various conventional and explicitly correlated (F12) estimates of complete basis set limit coupled-cluster (CCSD(T)/CBS) noncovalent interaction energies. In particular, we search for an improved "silver standard" of interaction energy calculations for systems where the CCSD(T) computation is feasible in a double-ζ basis but not in a triple-ζ one. We follow a recent study ( Sirianni J. Chem. Theory Comput. 2017 , 13 , 86 ) of different CCSD(T)-F12 variants in midbondless bases over the A24 and S22 benchmark interaction energy databases, and extend Dunning's correlation-consistent basis sets with three different midbond sets. The addition of bond functions is highly beneficial for conventional CCSD(T) and most CCSD(T)-F12 variants, improving both the CCSD part and the unscaled triples contribution. However, the commonly used scaling of triples by the ratio of the MP2-F12 and MP2 correlation energies usually overshoots: as a result, the scaled triples term gets worse upon the addition of bond functions. In contrast, a milder triples scaling by the ratio of the CCSD-F12b and CCSD correlation energies ( Brauer , Phys. Chem. Chem. Phys. 2016 , 18 , 20905 ) leads to the most accurate estimates of this term as long as bond functions are included. The combination of the triples term scaled in this way with the CCSD-F12b interaction energy leads to the CCSD(Tbb)-F12b approach that provides consistent high accuracy when a (3 s3 p2 d2 f) set of midbond functions is added to the aug-cc-pVDZ atom-centered basis set. The combination of midbond functions and the composite MP2/CBS+δ(CCSD(T)) treatment is able to make up for the deficiencies in the atom-centered part of the basis set, in particular, for a partial (or even complete) lack of diffuse functions. Considering both the A24 and S22 accuracy and the computational efficiency, we propose several new "silver standard" approaches improving upon the currently established midbondless levels of theory, ranging from the most consistent CCSD(Tbb)-F12b/aug-cc-pVDZ+(3 s3 p2 d2 f) variant (with mean unsigned errors of 0.010 and 0.042 kcal/mol for the A24 and S22 databases, respectively) to the significantly cheaper MP2/CBS+δ(CCSD(T))/cc-pVDZ+(3 s3 p2 d2 f) approach (mean unsigned errors of 0.039 and 0.096 kcal/mol for A24 and S22, respectively).