Calculation of referenced energies for error eval of heterogeneous catalyst systems in UMA

[bwander-sbaq]

Hello! The UMA paper states

Since our model predicts total energy, we make a total energy prediction on the same structure,
and another on a clean surface without the adsorbate. These two values are subtracted (along with the
energy of the adsorbate in isolation) to predict the adsorption energy.

It is clear that the model is used to perform energy evaluations on the slab and the adsorbate-slab, but its not clear what the state of the clean surface is (unrelaxed v relaxed) or how it arrived there (DFT v. MLP). I assume relaxed + DFT but I think worth clarifying.

Thanks for the help!

Brook

[mshuaibii]

Correct, for the S2EF evals the relaxed DFT surface is used. Since we only want to evaluate S2EF performance we don't want to introduce ML relaxations here to complicate that evaluation.

[bwander-sbaq]

Makes sense. Thank you!