Benchmarking universal machine learning interatomic potentials on elemental systems

Abstract The rapid emergence of universal machine learning interatomic potentials (uMLIPs) has transformed materials modeling. Nevertheless, a comprehensive understanding of their generalization behavior across configurational space remains an open challenge. In this work, we introduce a benchmarking framework to evaluate both the equilibrium and far-from-equilibrium performance of state-of-the-art uMLIPs, including two MACE-based models, two PET-based models, MatterSim, and a custom MACE model trained exclusively on elemental data. Our assessment utilizes Equation-of-State (EOS) tests to evaluate near-equilibrium properties, such as equilibrium volumes and bulk moduli, alongside extensive Minima Hopping (MH) structural searches to probe the Potential Energy Surface (PES). Here, we assess universality within the fundamental limit of elemental systems, which serve as a necessary baseline for broader chemical generalization and provide a framework that can be systematically extended to multicomponent materials. We find that while most models exhibit high accuracy in reproducing equilibrium volumes for transition metals, significant performance gaps emerge in alkali and alkaline earth metal groups as well as reactive non-metals. Crucially, our MH results reveal a decoupling between search efficiency and structural fidelity, highlighting that smoother learned PESs do not necessarily yield more accurate energetic landscapes.