Comparing Counterpoise-Corrected, Uncorrected, and Averaged Binding Energies for Benchmarking Noncovalent Interactions.

Abstract

High-quality benchmark computations are critical for the development and assessment of approximate methods to describe noncovalent interactions. Recent advances in the treatment of dispersion by density functional theory and also the development of more efficient wave function techniques to reliably address noncovalent interactions motivate new benchmark computations of increasing accuracy. This work considers focal point approximations to estimate the complete basis set limit of coupled-cluster theory through perturbative triples [CCSD(T)/CBS] and evaluates how this approach is affected by the use or absence of counterpoise (CP) correction or, as has recently gained traction, the average of those values. Current benchmark protocols for interaction energies are computed with all CP procedures and assessed against the A24 and S22B databases and also to highly converged results for formic acid, cyanogen, and benzene dimers. Whether CP correction, no correction, or the average is favored depends upon the theoretical method, basis set, and binding motif. In recent high-quality benchmark studies, interaction energies often use second-order perturbation theory with extrapolated aug-cc-pVTZ (aTZ) and aug-cc-pVQZ (aQZ) basis sets [MP2/aTQZ] combined with a “coupled-cluster correction,” δMP2CCSD(T), evaluated in an aug-cc-pVDZ basis. For such an approach, averaging CP-corrected and uncorrected values for the MP2 component and using CP-corrected δMP2CCSD(T) values offers errors more balanced among binding motifs and generally more favorable overall. Other combinations of counterpoise correction are not quite as accurate. When employing MP2/aQ5Z extrapolations and an aTZ basis for δMP2CCSD(T), using CP-corrected or averaged MP2 estimates are about equally effective (and slightly superior to uncorrected MP2 values), but the counterpoise treatment of δMP2CCSD(T) makes little difference. Focal point estimates at this level achieve benchmark quality results otherwise accessible only with CCSD(T)/aQZ or better.