Computational Evaluation of Molecular Binding on Metal Surfaces in Temperature Programmed Desorption: Accuracy of <scp>DFT</scp> Functionals and Neural Network Potentials

Understanding molecular adsorption and desorption on metal surfaces is crucial for heterogeneous catalysis and surface science. In this work, we systematically evaluate the performance of various density functional theory (DFT) exchange-correlation functionals-PBE, PBE-D3, revPBE-D3, optB88-vdW, BEEF-vdW, and SCAN+rVV10-and a neural network potential for modeling molecular binding relevant to temperature programmed desorption (TPD). Using well-characterized systems, we compare computed bulk lattice constants, surface energies, and binding energies of representative adsorbates (CO, CO₂, methanol, and benzene) on transition metal surfaces (Ni, Cu, Ru, Rh, Pd, Ag, Pt, Au) against experimental TPD data. Our results reveal that dispersion-corrected functionals like PBE-D3 and SCAN+rVV10 yield accurate bulk lattice constants, while BEEF-vdW tends to overestimate them. optB88-vdW and SCAN+rVV10 accurately reproduce surface energies, while PBE-D3 and revPBE-D3 often overestimate them. On the other hand, BEEF-vdW provides better agreement for binding energies but at the cost of less accurate bulk properties. We further demonstrate how machine-learned potentials can efficiently reproduce DFT-level energetics and enable molecular dynamics simulations to extract more realistic kinetic parameters, including pre-exponential factors and potential of mean force (PMF) profiles. This study demonstrates the accuracy of modeling molecular adsorption using DFT and neural network potentials, elucidating the trade-offs associated with functional selection in surface science and offering practical guidance for choosing appropriate computational methods for accurately simulating desorption processes.