What is the energy reference of OCPCalculator ?

[MSJavaScript]

What is the energy given by get_potential_energy() when I use OCPCalculator as a calculator ?
If it is an adsorption energy, what is the reference ?
If I want to fine tune a model, what kind of energy should I use ? If I just use the adsorption energy, then the energy reference will be different across different surfaces.

[zulissimeta]

Hi - the adsorption energies in OC20/OC22 are chosen from a linear combination of N2/CO/H2O/H2 and a relaxed structure for the bare surface. We have some details on this in the Demo FAQ https://open-catalyst.metademolab.com/faq and it should also be in the respective OC20/OC22 papers.

Note that as long you are consistent with whatever your favorite functional is (BEEF/RPBE/etc) and do the referencing yourself, you should be able to fine-tune the models on your new dataset. Obviously do not try to add new structures with a new functional (eg BEEF) and re-use the RPBE gas reference energies from the OC20 paper.

I'm not sure what you mean by If I want to fine tune a model, what kind of energy should I use ? If I just use the adsorption energy, then the energy reference will be different across different surfaces. - for a new surface, if you want to be consistent with OC20 you should prepare adsorption energies of the form "relaxed_surface + gas_references -> relaxed_slab_adsorbate".

[MSJavaScript]

@zulissimeta Thanks for your reply.
then the energy reference will be different across different surfaces I meant that, for example, if you calculate the adsorption energy of N* on two different surfaces: a flat $4\times 4\times 3$ Ag(111) surface and a re-constructed one, the adsorption energies are both calculated by
$E_{ads} = E_{tot} - E_{N*} - E_{slab}$, but the energies of $E_{slab}$ s are different for a flat surface and a re-constructed one, so the $E_{ads}$ s will have different energy references.

[mshuaibii]

Great question - Models were trained directly on adsorption energy where the slab reference, for the most part, is consistent with the adsorbate+slab system. That is - If you construct a system with N* on a flat Ag(111), you can expect the energy returned to be referenced to flat Ag(111).

Models in this scenario are implicitly assuming that the reference is consistent with your original adsorbate+slab system because that's how they were trained. Total energy models found here - https://fair-chem.github.io/core/model_checkpoints.html#s2ef-total-models alleviate this issue because you can use this model to independently predict the adsorbate+slab system and slab system, and subtract them directly.

[brunosamp-usp]

@mshuaibii I think that this is what you meant, but i just want to make sure. I am using the checkpoint file of the "EquiformerV2-153M-S2EF-OC20-All+MD" model and calling OCPCalculator.
If i construct a Ni slab adsorbed by a CO molecule, for example, and define

calc = OCPCalculatator(checkpoint_path=checkpoint_path, cpu=False)

atoms.set_calculator(calc)
opt = BFGS(atoms, trajectory='optimization.traj')
slab = fcc111("Ni", (3, 3, 4), vacuum=20)
slab_e = slab.get_potential_energy()
print(slab_e)

The energy that was printed is the adsorption energy, right? So, i don't need to calculate the energy of the clean slab, and of the isolated CO molecule, to do Eads = E(Complex) - E(Clean Slab) - E(Isolated CO)

[MSJavaScript]

@mshuaibii Thanks for your help. 🤝

[MSJavaScript]

@brunosamp-usp Yes, you're right, it is just the Eads you write below. So I think it is problematic to use EquiformerV2-153M-S2EF-OC20-All+MD to do NEB or any other tasks in which you can not properly define a clean slab.