Universal ML Force Fields Reproduce Lunar Silicate Structures but Struggle with Fe-Ti Coordination in Regolith Minerals

The arXiv preprint benchmarks transferability of universal machine-learning interatomic potentials to lunar-relevant phases using 10-nanosecond simulations per mineral. Structural metrics such as partial radial distribution functions and bond-angle distributions were compared against crystallographic references, with throughput measured on a single RTX 4090 GPU. SevenNet-0, MatterSim, and UPET delivered the highest simulation rates while MACE-MH offered a practical cost-accuracy balance. Hydroxylated surface tests showed consistent O-H bond lengths across models, indicating utility for initial volatile-stability screening. The work directly addresses In-Situ Resource Utilization needs for Artemis-era oxygen extraction from ilmenite, where accurate Fe and Ti coordination governs reduction kinetics. Prior lunar MD studies relied on classical potentials that systematically under-estimate Fe-O bond flexibility; the present comparison quantifies how far foundation models close that gap without mineral-specific training. Key limitation remains absence of ab initio reference trajectories for Fe- and Ti-bearing compositions, leaving short-timescale fluctuations unvalidated. Future fine-tuning on DFT data for ilmenite and armalcolite would strengthen predictive power for space-weathering and polar-volatile models. Integration with upcoming Chang'e-6 and Artemis sample suites could supply the missing ground-truth spectra within 18 months.