GMTKN55 (General Main-group Thermochemistry, Kinetics, and Noncovalent interactions) is the broadest public benchmark suite for density functional approximations on molecular properties. Published by Grimme and coworkers, it contains 55 test sets spanning 2462 reference data points covering thermochemistry, barrier heights, conformational energies, and noncovalent interactions. For anyone selecting a DFT functional for production use, GMTKN55 is the primary reference.
We ran the Qchemvyx production functional stack against the full GMTKN55 suite to characterize where each functional we support performs well and where users should apply caution. This post covers the barrier height subsets (most critical for catalyst design) and the noncovalent interaction subsets (most critical for binding energy prediction) in detail, with summary numbers for the thermochemistry subsets.
Calculation Setup and Reference Values
All calculations were run at the def2-TZVP basis set with D4 dispersion (or the native dispersion parameterization for ωB97X-D series). Reference geometries are the GMTKN55 published geometries — we did not reoptimize. Reference energies are CCSD(T)/CBS or W2X composites depending on the subset (as specified in the original GMTKN55 publication).
Functionals tested: B3LYP-D4, PBE0-D4, ωB97X-D3, M06-2X, r²SCAN-D4, TPSS-D4, PBE-D4. We excluded BLYP-D4 from the primary comparison table as it is not in our production stack — its barrier height performance is insufficient for catalyst design applications regardless of the ease of access.
Barrier Heights: BH76 and BHPERI
BH76 covers 76 forward and reverse barriers for 38 reactions: hydrogen atom transfer (18), heavy-atom transfer (7), nucleophilic substitution (6), unimolecular and association reactions (7). BHPERI covers 26 barrier heights for pericyclic reactions (electrocyclic, sigmatropic, cycloadditions). These are the subsets most directly relevant to reaction rate prediction in organic and organometallic synthesis.
BH76 MAE results (kcal/mol), def2-TZVP + D4:
- ωB97X-D3: 1.62 (best single-functional performance)
- M06-2X: 1.88
- r²SCAN-D4: 2.74
- PBE0-D4: 3.24
- B3LYP-D4: 3.87
- TPSS-D4: 6.21
- PBE-D4: 7.89
BHPERI MAE results (kcal/mol):
- ωB97X-D3: 1.41
- M06-2X: 1.73
- r²SCAN-D4: 2.96
- PBE0-D4: 2.88
- B3LYP-D4: 3.41
- TPSS-D4: 5.18
The consistent pattern: functionals with higher Hartree–Fock exchange fraction perform better on barrier heights. ωB97X-D3 (range-separated, approaches 100% HF exchange at long range) and M06-2X (54% HF exchange) lead across both subsets. Pure GGA functionals (PBE, TPSS) systematically underestimate barriers by 6–10 kcal/mol and are not appropriate for kinetics work.
The B3LYP Underestimation Bias
B3LYP's BH76 MAE of 3.87 kcal/mol includes a systematic directional component — it underestimates barriers. The mean signed error (MSE) for B3LYP-D4 on BH76 is −3.2 kcal/mol (consistent with the widely cited ~4 kcal/mol underestimation from earlier studies). This systematic bias is problematic for absolute rate predictions. For relative ranking of catalysts within a homologous series, the cancellation of systematic errors partially mitigates the problem — but not fully, particularly when the structural variation changes the charge-transfer character of the TS.
We're not saying B3LYP is inappropriate for all DFT work — it remains highly accurate for ground-state geometries, thermochemistry of closed-shell organic molecules (W4-11 subset MAE of 2.1 kcal/mol), and qualitative analysis. We're saying it should not be the functional of choice when barrier heights are the primary output and accuracy below 2 kcal/mol is required.
Thermochemistry Subsets: W4-11 and ISOL24
W4-11 contains 140 total atomization energies. ISOL24 contains 24 isomerization energies for large organic molecules — particularly sensitive to intramolecular dispersion.
W4-11 MAE results (kcal/mol):
- ωB97X-D3: 2.14
- M06-2X: 1.87
- PBE0-D4: 3.12
- r²SCAN-D4: 2.98
- B3LYP-D4: 2.09
- TPSS-D4: 4.51
The thermochemistry ranking differs from the barrier height ranking. B3LYP performs comparably to ωB97X-D3 on W4-11 — its failure mode is specifically in kinetics, not thermochemistry. M06-2X shows the best thermochemistry performance in our test and the second-best barrier height performance, which explains its popularity for broad-scope DFT studies.
ISOL24 (isomerization, large molecules): r²SCAN-D4 (MAE 1.4 kcal/mol) and ωB97X-D3 (MAE 1.8 kcal/mol) lead; M06-2X (3.1 kcal/mol) underperforms expectations here, reflecting known issues with M06-2X for extended conformational energies of molecules with many rotatable bonds.
Noncovalent Interactions: S22, S66, WATER27
S22 contains 22 noncovalent complexes (hydrogen-bonded, dispersion-dominated, mixed). S66 extends this to 66 complexes with better structural diversity. WATER27 contains 27 water cluster binding energies (monomer through hexamer).
S66 MAE results (kcal/mol):
- ωB97X-D3: 0.22
- r²SCAN-D4: 0.28
- PBE0-D4: 0.31
- B3LYP-D4: 0.41
- M06-2X: 0.38
- TPSS-D4: 0.52
WATER27 MAE results (kcal/mol):
- ωB97X-D3: 0.66
- r²SCAN-D4: 0.71
- PBE0-D4: 0.94
- M06-2X: 0.81
- B3LYP-D4: 1.02
For noncovalent interactions, the performance spread across functionals is smaller than for barrier heights — all tested functionals with D4 dispersion give MAE below 1.1 kcal/mol on S66. This reflects that modern dispersion corrections do most of the heavy lifting for these interactions. The choice of functional matters less than the choice of dispersion correction for noncovalent binding predictions.
Weighted WTMAD-2 Summary
GMTKN55 provides a weighted total mean absolute deviation (WTMAD-2) that weights subsets by difficulty to give a single-number performance summary:
- ωB97X-D3: WTMAD-2 = 2.19 kcal/mol
- M06-2X: WTMAD-2 = 2.73 kcal/mol
- r²SCAN-D4: WTMAD-2 = 3.41 kcal/mol
- PBE0-D4: WTMAD-2 = 3.88 kcal/mol
- B3LYP-D4: WTMAD-2 = 4.12 kcal/mol
- TPSS-D4: WTMAD-2 = 6.73 kcal/mol
ωB97X-D3 leads the overall ranking, consistent with its consistent top-2 performance across most subsets. The WTMAD-2 summary is useful for general-purpose functional selection but can be misleading for domain-specific work: if your application is exclusively barrier heights, M06-2X is competitive with ωB97X-D3. If your application is exclusively noncovalent binding, r²SCAN-D4 is an excellent cost-efficient choice.
Implications for Qchemvyx Default Settings
Based on this analysis, the Qchemvyx default functional for reaction pathway calculations (transition state search, activation barrier prediction) is ωB97X-D3/def2-TZVP. For catalyst binding energy prediction where cost is a constraint and the substrate is an organic molecule (no transition metals), r²SCAN-D4 is offered as a cost-efficient alternative with good S66 and WATER27 performance. B3LYP-D4 is supported for literature reproducibility and for geometry optimization, but is not recommended for barrier height prediction in new studies.