Two benchmark sets of dissociation energies of diatomic transition-metal complexes have been assembled from previously published data, with selected reference values reassessed or updated where necessary, and used to evaluate a variety of density functionals. The BDE41 set of 41 3d complexes based on both experimental data and phaseless auxiliary-field quantum Monte Carlo reference data is the statistically more robust of the two sets. A number of local hybrid functionals top the list with the smallest deviations, including the strong-correlation (sc) corrected scLH22t and scLH22ta, and the range-separated variant ωLH23tP, followed by the previous top-performing global hybrids B97 and B97-1. While global hybrids and local hybrids without sc-corrections tend to underbind on average, sc-corrections may shift this to moderate overbinding. Nitride and oxide complexes account for much of the underbinding. Functionals based on neural-network local mixing functions underbind more clearly also in other cases, confirming that for these very flexible constructions a lack of transition-metal training data is more harmful to transferability than for less flexible functionals with heuristic mixing functions. The smaller and somewhat less reliable HBDE17 set of 4d/5d complexes reinforces several of the observations but appears to disfavor large exact-exchange admixtures even more strongly. In this case carbide and oxide complexes provide the most pronounced challenges. Nevertheless, several functionals perform not much worse than the most sophisticated post-Hartree-Fock reference calculations currently available based on CCSD(T) with multi-reference corrections. These benchmark sets may also prove useful for training and evaluating new approximate methods, an application we intend to explore in future work.
No takes yet. Share an insight, caveat, or question.
Weber et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: