The next generation of Lars Goerigk Stefan Grimme’s massive and very diverse GMTKN (general main group thermochemistry, kinetics, and noncovalent interactions) benchmark for DFT has just been published. It entails almost 2,500 items grouped in 55 chemically distinct problem sets. [L. Goerigk, A. Hansen, C. A. Bauer, S. Ehrlich, A. Najibi, and S. Grimme, “ A Look at the Density Functional Theory Zoo with the Advanced GMTKN55 Database for General Main Group Thermochemistry, Kinetics and Noncovalent Interactions”, Phys. Chem. Chem. Phys. Early View, (2017); http://dx.doi.org/10.1039/C7CP04913G ] Full supporting data (and input structures in Turbomole format) are available for viewing and download at https://www.chemie.uni-bonn.de/pctc/mulliken-center/software/GMTKN/gmtkn55
Below is a graph of the weighted performance statistics over the dataset. The top 3 performers on each rung of Perdew's "Jacob's Ladder" of functionals have been marked in red.
Highlights for overall performance:
* top 3 performers: DSD-PBEP86-D3BJ, DSD-BLYP-D3BJ, and B2GPPLYP-D3BJ, all double hybrids developed with Sebastian Kozuch and Amir Karton when they were in my group. A heartfelt mazal tov and kol ha-kavod to both of you!
* best range-separated hybrid: Mardirossian and Head-Gordon’s ωB97X-V, where the "V" stands for Vydrov-Van Voorhis (VV10) nonlocal dispersion. (ωB97M-V is likely to perform slightly better still.) Its predecessor ωB97X-D3, Zhao and Truhlar's M06-2X, and the latter's predecessor M05-2X are essentially tied for second.
* best global hybrid without elaborate parametrization: Zhao and Truhlar’s PW6B95 with added D3BJ dispersion.
* best meta-GGA: Perdew’s SCAN with D3BJ dispersion
* best GGA: revPBE with D3BJ dispersion, followed by Grimme’s B97D3
* the advantage of meta-GGAs over GGAs is not readily apparent. (It might be more in a periodic DFT benchmark for crystal structures.)
More detailed comparisons can be found inside the paper, which is a veritable treasure trove.
The GMTKN55 authors also note that some of the most popular “choices by inertia” such as B3LYP are not part of their recommendations in any category.
All of the GMTKN55 reference data were obtained by high-level ab initio methods near the basis set limit. A number of the 55 data subsets were taken in whole or in part from published studies by our group or Amir Karton’s; yet others were calculated by Lars Goerigk et al. themselves using Weizmann-n or Weizmann-n-F12 like protocols.
Obviously, this benchmark does not sample excited states, transition metal complexes, or derivative properties. Also, periodic solid-state calculations might show a different picture, notably for the difference between GGAs and meta-GGAs.Still, we are quite gratified by the performance of our double hybrids.