VQM24 transition-state benchmark¶
The VQM24 dataset contains 835,947 converged structures calculated with PSI4 1.7 at the omegaB97X-D3/cc-pVDZ level. Its chemical space includes C, N, O, F, Si, P, S, Cl, and Br. The associated Scientific Data paper provides geometries, harmonic frequencies, and thermochemical properties for minima and saddle structures.
Run the transition-state benchmark from a source checkout:
python scripts/validate_vqm24.py
The command downloads the 112 MB DFT_saddles.npz archive, verifies its MD5
checksum, and keeps it outside the repository. It scans all 51,072 saddle
records and recomputes thermochemistry for a deterministic sample of 1,000
first-order transition states.
Method¶
A first-order transition state must contain exactly one imaginary frequency.
The full scan finds 26,445 such records. It excludes 21 records with 3N-3
stored modes because they do not represent a projected vibrational spectrum.
One near-linear C2NPSi record contains 3N-5 stored modes while its
published thermal corrections use three rotational degrees of freedom. The
extra 5.0625 cm-1 mode makes the source representation internally
inconsistent, so the benchmark reports and excludes the record instead of
changing the thermochemistry model to reproduce it.
The sample combines the first 100 comparable records, 450 evenly spaced
records, a pinned near-linear C-Si-C regression, and a seeded random remainder. The
electronic baseline is reconstructed as U0 - ZPVE so the comparison
isolates thermochemistry from inconsistencies between independent source
fields. Calculations use 298.15 K, 101325 Pa, a singlet electronic state, and
rotational symmetry number one.
Reference result¶
Energy differences are reported in microhartree.
Property |
Mean absolute error |
Maximum absolute error |
|---|---|---|
|
0.002560 microhartree |
0.006189 microhartree |
|
0.005651 microhartree |
0.020289 microhartree |
|
0.005880 microhartree |
0.019318 microhartree |
|
7.164475 microhartree |
25.527762 microhartree |
|
0.015089 cal/(mol K) |
0.053738 cal/(mol K) |
|
0.000256 cal/(mol K) |
0.000517 cal/(mol K) |
The maximum Gibbs-energy difference is approximately 0.016 kcal/mol.
Near-linear geometry¶
VQM24 array index 2 is a nonlinear C-Si-C transition state with three
projected vibrational modes. Its smallest principal moment is only about
4.7e-10 of the largest, below a practical geometry-only linearity cutoff.
An exact 3N-6 supplied mode count therefore resolves the geometry as
nonlinear within the near-linear ambiguity band. Strictly linear geometries,
full Cartesian spectra, and 3N-5 spectra continue to use the geometry-based
classification.
Interpretation¶
This benchmark independently validates transition-state mode handling and ideal-gas rigid-rotor harmonic-oscillator thermochemistry against PSI4 over a broader elemental range than QM9. It does not evaluate the predictive accuracy of the density-functional method or the physical validity of the underlying saddle structures.