arXiv Paper Benchmarks 6 Universal ML Interatomic Potentials, Evaluating Structural Fidelity and Performance for Lunar Regolith MD Simulations

July 19, 2026

arXiv International

Overview

Six universal machine-learning interatomic potentials (MLIPs)—MACE-MH, MatterSim, SevenNet-0, UPET, UMA, and NequIP-OAM-L—were comprehensively benchmarked for molecular dynamics (MD) simulations of lunar regolith minerals. The study evaluated each MLIP’s structural fidelity and computational performance, identifying models with the highest throughput and pinpointing areas requiring further validation, especially for Fe- and Ti-containing lunar phases. This is expected to enhance the precision of materials science research for future lunar activities.

In Depth

Key Findings

Six leading universal machine learning interatomic potentials (MLIPs) used in molecular dynamics (MD) simulations of lunar regolith minerals were comprehensively benchmarked, evaluating their structural fidelity and computational performance. This study identified specific MLIPs that offer excellent throughput for complex lunar regolith environment simulations, while also highlighting the need for additional validation for certain modeling aspects of iron (Fe) and titanium (Ti)-containing lunar phases.

Technical / Clinical Details

This paper applied state-of-the-art MLIPs—MACE-MH, MatterSim, SevenNet-0, UPET, UMA, and NequIP-OAM-L—to systems of lunar regolith minerals (e.g., olivine, pyroxene, plagioclase, ilmenite). The evaluation focused on how accurately each potential could reproduce physical properties such as crystal structure stability, diffusion coefficients, and thermal conductivity in MD simulations. Specifically, structural fidelity (e.g., interatomic distances, bond angles, density) was quantitatively assessed by comparing with results from Density Functional Theory (DFT) calculations. The results showed that some MLIPs could achieve both excellent computational efficiency and high structural fidelity for key lunar regolith minerals. However, for lunar phases containing transition metals like iron and titanium, particularly complex oxides like ilmenite, existing universal MLIPs were noted to have limitations in accurately capturing their electronic structures and magnetic properties. This is likely due to the complex electron correlations and spin states associated with these elements not being sufficiently learned in existing datasets.

Background & Context

Lunar regolith is an indispensable material for future lunar activities such as constructing lunar bases, extracting resources (water, helium-3), and space agriculture. Accurately understanding its physical and chemical properties at the atomic level is fundamental to developing these technologies. MD simulations are powerful tools for predicting regolith properties and evaluating material durability, but traditional classical force fields have limitations in accuracy, and DFT is too computationally expensive. MLIPs are expected to bridge this gap, but their generality and applicability to specific material systems required detailed validation. This study provides this validation and offers specific guidance for lunar regolith research.

Strategic Significance & Outlook

The results of this benchmark study will significantly influence the selection of MLIPs for MD simulations of lunar regolith materials and future development directions. MLIPs offering the highest throughput and sufficient structural fidelity can be immediately utilized for material design in support of future lunar exploration missions and infrastructure development. On the other hand, the challenges regarding Fe- and Ti-containing lunar phases suggest the need for fine-tuning MLIPs specifically for these elements and building more diverse datasets. In the future, these improvements are expected to enable more accurate prediction of material behavior in the lunar environment, strengthening the materials science foundation for sustainable lunar activities.

Source: https://arxiv.org/abs/2607.09005

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.