Accurate prediction of barrier heights and reaction energies is paramount for understanding chemical reactivity and kinetics. The RDB7 dataset created by Spiekermann et al in Sci Data 9, 417 (2022) is a large and diverse data set that covers a large variety of reaction templates with approximately 12000 reactions in total. Despite the breadth of barrier heights and reaction energies that this work offers, it also states that comparing the barrier heights calculated at wB97X-D3/def2-TZVP relative to the CCSD(T)-F12a/cc-pVDZ-F12 reference yields a mean absolute error (MAE) of 3.5 kcal/mol and a root mean squared error (RMSE) of 5.0 kcal mol. These values are significantly larger than the reported values in other benchmarks, with the primary error sources remaining ambiguous.
In this work, we revisited the RDB7 dataset to identify and address potential sources of error. By rigorously analyzing the stability of orbitals at the self-consistent field (SCF) level and kappa regularized orbital-optimized second-order Moller-Plesset perturbation theory (k-OOMP2) level, we categorized all the reactions based on the nature and degree of spin-symmetry breaking into easy, intermediate, and difficult subsets. We did new benchmarks with new functionals wB97M-V, MN15, and wB97M(2), and the performances on easy and intermediate subsets are comparable to other benchmarks. For the difficult subset where the k-OOMP2 orbitals are also not stable, we recalculated the reference level with UHF-based UCCSD(T)-F12 and Yamaguchi spin projection on the 100 reactions with the largest deviations between the old reference and previously published wB97M-V values. Our findings underscore the critical role of spin-unpolarized DFT orbitals in accurately computing barrier heights and reaction energies and highlight the need for comprehensive stability analysis and spin square expectation value assessments to detect strong correlation effects.

Barrier height error distribution for the "easy" and "intermediate" subsets

Angiras Menon

Appearances