Bridging accuracy and efficiency: assessing universal ML potentials for niobium-oxygen clusters

Abstract

Machine Learning Interatomic Potentials (MLIPs) promise to combine the accuracy of DFT with the speed of classical force fields. However, their reliability for complex, multi-component systems requires rigorous validation. Here, we perform a targeted evaluation of three leading universal MLIPs using niobium oxide clusters (Nb_nO_m, n≤6, m≤6) as a challenging test case. The Nb‑O system is very well suited for this task due to its complex electronic interactions, manifested in existence of the bulk phase with 25% vacancy-ordered lattice and, at nanoscale, by a diverse range of non-stoichiometric clusters. We employ a dataset of global minima structures identified via DFT-based evolutionary search as a strict reference. A comparative analysis is then performed by executing evolutionary searches with the MLIPs. By directly comparing predicted structures, energies, and relative stability, we provide a comprehensive assessment of the accuracy and limitations of current universal potentials for modeling complex nanoscale oxides.