Machine Learning Interatomic Potentials (MLIPs) enable first-principles accuracy in modeling complex catalytic interfaces at reduced cost. The Open Catalyst 2020 (OC20) and Open Materials 2024 (OMat24) datasets consist of density functional theory (DFT) data for adsorbates on metals and inorganic bulk materials, respectively. In this work, we benchmarked the Universal Model for Atoms (UMA) trained on the OC20 and OMat24 datasets with DFT calculations for adsorption energies on Pt and \(\textrm{IrO}_2\) , respectively. We evaluated adsorption energies of common adsorbates such as *H, *O, *CO, and *NO on Pt, and reaction intermediates in the water splitting reaction, such as *O, *OH, and *H, on \(\textrm{IrO}_2\) . UMA-OC20, trained using non-spin polarized DFT-RPBE energies with no van der Waals (vdW) corrections on metallic surfaces, reproduced the adsorption energies from the same flavor of DFT within 0.1 eV. However, the energy difference between UMA-OC20 and spin-polarized DFT was \(\sim \) 1.2 eV, with the difference mainly arising from the reference gases. UMA-OMat24, trained using spin-polarized DFT-PBE with vdW corrections, predicted adsorption energies within 0.3 eV of the energies from the same flavor of DFT on \(\textrm{IrO}_2\) . When both the DFT flavor and geometry were kept fixed between the MLIPs and DFT, UMA-OC20 and UMA-OMat24 predicted adsorption energies within 0.1 and 0.2 eV of their corresponding DFT adsorption energies, respectively. The norm of coordinate differences between UMA-OC20 and DFT was \(\sim \) 0.1 Å for a system size of \(\sim \) 20 atoms, while it was < 0.6 Å for a system size of \(\sim \) 60 atoms, between UMA-OMat24 and DFT. MLIP requires considerably fewer ( \(\sim 10^3-10^4\) times less) CPU-hours compared to DFT for geometry optimization. Thus, we recommend opting for an MLIP-followed-by-DFT procedure, as it not only reduces the total compute time but also ensures accurate geometry and energies.
