Benchmarking Universal Machine Learning Interatomic Potentials on Elemental Systems

TL;DR

The paper introduces a two‑axis benchmarking framework for unary elemental uMLIPs, combining near‑equilibrium EOS tests with far‑from‑equilibrium MH global optimizations to assess both local accuracy and global structure exploration. It evaluates five state‑of‑the‑art potentials (three MACE variants, MatterSim, PET‑MAD) on a large unary dataset, revealing strong performance in transition metals but notable gaps in alkali and alkaline‑earth metals, and a notable decoupling between PES smoothness and the ability to recover correct low‑energy minima. The work provides a rigorous, interpretable protocol for universality in interatomic potentials, identifies specific strengths and failure modes across chemical groups, and highlights the need for data coverage and architecture improvements to achieve reliable, broad‑range PES representations. Overall, the framework and findings guide future development of foundation interatomic potentials toward more consistent accuracy across diverse elemental chemistries.

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.

Figures (9)

• Figure 1: Periodic table map showing the element-wise equilibrium-volume error ($\Delta V_0$) for the MACE-Unary model. Colors indicate the percent deviation relative to DFT reference equilibrium volumes.
• Figure 2: Periodic table map showing the element-wise equilibrium-volume error ($\Delta V_0$) for the MACE-MATPES model. Colors indicate the percent deviation relative to DFT reference equilibrium volumes.
• Figure 3: Periodic table map showing the element-wise equilibrium-volume error ($\Delta V_0$) for the MACE-OMAT model. Colors indicate the percent deviation relative to DFT reference equilibrium volumes.
• Figure 4: Periodic table map showing the element-wise equilibrium-volume error ($\Delta V_0$) for the MatterSim model. Colors indicate the percent deviation relative to DFT reference equilibrium volumes.
• Figure 5: Periodic table map showing the element-wise equilibrium-volume error ($\Delta V_0$) for the PET-MAD model. Colors indicate the percent deviation relative to DFT reference equilibrium volumes.