Accuracy and validation
What the numbers have been checked against, how closely they agree, and what is deliberately not supported.
Every quantity below has been compared against an established reference program at matched settings, on a set of molecules spanning closed and open shells, neutral and charged species and several basis sets. The agreement is quoted as the largest difference observed, not the average.
Measured agreement
| Quantity | Agreement | Conditions |
|---|---|---|
| Hartree-Fock total energy | Better than 1e-8 hartree | Matched grids and thresholds |
| DFT total energy, local, gradient-corrected, meta and hybrid functionals | Better than 1e-6 hartree | Converged integration grid |
| Valence orbital energies | Better than 1e-5 hartree | |
| Nuclear gradients | Better than 1e-6 hartree per bohr | |
| Optimised bond lengths | Better than 0.0001 angstrom | |
| Harmonic frequencies | Better than 1 reciprocal centimetre | |
| Dipole moment | Better than 0.0001 debye | |
| Mulliken and Loewdin charges | Better than 0.0001 electrons | |
| CHELPG charges | Better than 0.02 electrons | Same radii and grid on both sides |
| Electrostatic potential on the density surface | Better than 0.002 kcal/mol |
Reproducibility
The same input gives bit-identical results on any number of processor cores, from one to twenty. This is not automatic: it requires that every parallel sum be accumulated in a fixed order rather than in whatever order the threads happen to finish. It means a number you publish can be reproduced exactly on a different machine.
Internal consistency checks
- Analytic gradients agree with numerical differentiation of the energy.
- The integration grid integrates the electron density to the exact electron count.
- Two distant fragments give the sum of their separate energies, to the accuracy of the physical interaction between them.
- The energy does not change when the molecule is rotated or translated.
- Different starting guesses reach the same converged solution.
- A closed-shell system gives the same answer restricted and unrestricted.
Settings matter
Not supported, and why
| Item | Status |
|---|---|
| Minnesota functionals | Available, but the agreement with the reference does not reach 1e-6 hartree. Treat results as indicative. |
| wB97X-D | Refused. Its definition includes a dispersion term that is not implemented, so a result under that name would not be that functional. Use wB97X or CAM-B3LYP. |
| Double hybrids | Single-point energies only. There is no analytic gradient, so they cannot drive an optimisation or a frequency calculation. |
| ROHF and ROKS | Not supported. Use the unrestricted forms for open-shell systems; the spin contamination is reported. |
| Hirshfeld and CHELPG charges compared across programs | The definitions differ between programs. Our values are internally consistent and documented; comparisons across programs need matched definitions. |
Dispersion
The dispersion correction supplied with GeinDFT is parameterised for B3LYP, PBE0, PBE, BLYP, TPSS and Hartree-Fock, and covers elements from hydrogen to krypton. On a halogen benchmark set that played no part in fitting it, the mean absolute error is between 0.26 and 0.40 kcal/mol, against 0.92 to 2.29 kcal/mol with no dispersion correction at all.