Benchmarking Universal Machine Learning Interatomic Potentials on Elemental Systems

TL;DR

This paper introduces a benchmarking framework to evaluate the generalization and performance of state-of-the-art universal Machine Learning Interatomic Potentials across elemental systems, revealing strengths and gaps in their accuracy and efficiency.

Contribution

It provides a systematic evaluation method for uMLIPs on elemental systems, highlighting their strengths and limitations in both equilibrium and far-from-equilibrium scenarios.

Findings

Most models accurately predict equilibrium volumes for transition metals.

Significant performance gaps exist for alkali and alkaline earth metals.

Smoother PESs do not always lead to more accurate energetic landscapes.

Abstract

The rapid emergence of universal Machine Learning Interatomic Potentials (uMLIPs) has transformed materials modeling. However, 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 three MACE-based models, MatterSim, and PET-MAD. Our assessment utilizes Equation-of-State (EOS) tests to evaluate near-equilibrium properties, such as bulk moduli and equilibrium volumes, alongside extensive Minima Hopping (MH) structural searches to probe the global Potential Energy Surface (PES). Here, we assess universality within the fundamental limit of unary (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. 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.