Troubles in the systematic prediction of transition metal thermochemistry with contemporary out-of-the-box methods

The recently developed DLPNO-CCSD(T) method and 7 popular DFT functionals (B3LYP, M06, M06L, PBE, PBE0, TPSS and TPSSh) with and without an empirical dispersion term have been tested to reproduce 111 gas phase reaction enthalpies involving 11 different transition metals. Our calculations, corrected for both relativistic effects and basis set incompleteness, indicate that most of the methods applied with default settings perform with acceptable accuracy on average. Nevertheless, our calculations also evidenced unexpected and non systematic large deviations for specific cases. For group 12 metals (Zn, Cd, Hg) most of the methods provided mean unsigned errors (MUE) less than 5.0 kcal/mol, with DLPNO-CCSD(T) and PBE methods performing excellently (MUE lower 2.0 kcal/mol). Problems started with group 4 metals (Ti and Zr). Best performer for Zr complexes with a MUE of 1.8 kcal/mol, PBE0-D3, provides a MUE larger than 8 kcal/mol for Ti. DLPNO-CCSD(T) provides a reasonable MUE of 3.3 kcal/mol for Ti reactions, but gives MUE a larger than 14.4 kcal/mol for Zr complexes, with all the larger deviations for reactions involving ZrF4. Large and non-systematic errors have been obtained for group 6 metals (Mo and W), for 8 reactions containing Fe, Cu, Nb and Re complexes. Finally, for the whole set of 111 reactions, the DLPNO-CCSD(T), B3LYP-D3 and PBE0-D3 methods turned out to be the best performers, both providing MUE below 5.0 kcal/mol. Since DFT results cannot be systematically improved and large non-systematic deviations of 20-30 kcal/mol were obtained even for best performers, our results indicates that current DFT methods are still unable to provide robust predictions in transition metal thermochemistry, at least for the functionals explored in this work. The same conclusion holds for both DLPNO-CCSD(T) and canonical CCSD(T) methods when used entirely as out-of-the-box. However if careful investigation core correlation is performed, relativistic effects are properly included and the quality of the reference wave function is properly checked, CCSD(T) methods can still provide good quality results that might be even used to validate DFT methods, due to paucity of accurate thermodynamic data for realistic-size transition metal complexes.