Comparison of Matlantis and VASP bulk formation and surface energies in metal hydrides, carbides, nitrides, oxides, and sulfides

Generic neural network potentials without forcing users to train potentials could result in significantly acceleration of total energy calculations. Takamoto et al. [Nat. Commun. (2022), 13, 2991] developed such a deep neural network potential (NNP) and made it available in their Matlantis package. We compared the Matlantis bulk formation, surface, and surface O vacancy formation energies of metal hydrides, carbides, nitrides, oxides, and sulfides with our previously calculated VASP values obtained from first-principles with the PBEsol(+U) functional. Matlantis bulk formation energies were consistently ~0.1 eV/atom larger and the surface energies were typically ~10 meV/{\AA}^2 smaller than the VASP counterpart. Surface O vacancy formation energies were generally underestimated within ~0.8 eV. These results suggest that Matlantis energies could serve as a relatively good descriptor of the VASP bulk formation and surface energies.