Comparison of Explicitly Correlated Methods for Computing High-Accuracy Benchmark Energies for Noncovalent Interactions.

The reliability of explicitly correlated methods for providing benchmark-quality noncovalent interaction energies was tested at various levels of theory and compared to estimates of the complete basis set (CBS) limit. For all systems of the A24 test set, computations were performed using both aug-cc-pVXZ (aXZ; X = D, T, Q, 5) basis sets and specialized cc-pVXZ-F12 (XZ-F12; X = D, T, Q, 5) basis sets paired with explicitly correlated coupled cluster singles and doubles [CCSD-F12n (n = a, b, c)] with triple excitations treated by the canonical perturbative method and scaled to compensate for their lack of explicit correlation [(T**)]. Results show that aXZ basis sets produce smaller errors versus the CBS limit than XZ-F12 basis sets. The F12b ansatz results in the lowest average errors for aTZ and larger basis sets, while F12a is best for double-ζ basis sets. When using aXZ basis sets (X ≥ 3), convergence is achieved from above for F12b and F12c ansatzë and from below for F12a. The CCSD(T**)-F12b/aXZ approach converges quicker with respect to basis than any other combination, although the performance of CCSD(T**)-F12c/aXZ is very similar. Both CCSD(T**)-F12b/aTZ and focal point schemes employing density-fitted, frozen natural orbital [DF-FNO] CCSD(T)/aTZ exhibit similar accuracy and computational cost, and both are much more computationally efficient than large-basis conventional CCSD(T) computations of similar accuracy.