frobnitzem
commented
3 months ago @allaffa Comparing energies is complicated because QM methods have many settings which modify their energy function:
- density functional (e.g. LDA, PBE or PBE0 or HSE06 or B3LYP)
- pseudopotential (e.g Ultrasoft, Norm-Conserving, or PAW)
- "U" parameter applied to metal d- or f- electrons
- basis set / planewave cutoff
And there are 2 different ways to reference energies between different QM methods (to make comparison of the "energy" across methods valid). Both subtract off the energy of all isolated atoms computed with the same QM method. One (simple) way treats isolated atoms as single, uncharged atoms in gas phase. The other (physical) way is to treat isolated atoms as part of their equilibrium material (e.g. N2 gas for nitrogen, solid graphite for carbon, liquid for bromine, etc).
Among the QM7-x properties, the most useful energies for comparison are:
- 'ePBE0': PBE0 energy [eV] (1)
- 'eAT': PBE0 atomization energy [eV] (1)
The first one is the total energy output by DFT with the PBE0 functional. The second one is the PBE0 energy after subtracting the energies of the isolated atoms using the first (simple) method. Ideally we should fit that second one. see atomic energies in Table S3
OpenCatalyst provides "raw DFT" energies output from VASP with DFT+U as well as (since OC22) reference energies using a non-standard version of the second (physical) method, - see also Table 17 on p. 42 of the OC22 paper. However, they don't seem to have included similar references for the slab atoms. The VASP forum notes that the "raw DFT" energies are usually close to atomization energies, but this isn't guaranteed. Calculations on single atoms are difficult with VASP.
So we need to adjust the referenced energies of OC or QM7x to compare. Despite being non-physical, I favor the simple atomization energies because their reference QM calculations are more exact (single atom vs. liquid, for example).
allaffa
@frobnitzem among all the energies that can be retrieved from the QM7x dataset, may you help me identify the one that corresponds to the energy returned by OC2020 and OC2022?