Foundation Force Fields Face Lunar Regolith Transferability Test
A v1 arXiv preprint submitted on July 10, 2026 benchmarks six universal machine-learning force fields on four minerals relevant to lunar regolith. The authors ran all 24 model-mineral combinations for 1,000 molecular-dynamics steps at 300 K, a trajectory of 1 picosecond, and compared local structures with crystallographic references. Every run remained stable, but iron-oxygen and titanium-oxygen environments varied more across models than silicon, magnesium, aluminum, or calcium coordination. On one NVIDIA RTX 4090, SevenNet-0 delivered the shortest author-reported runtimes, while MatterSim used the least peak memory. The result is a useful screening benchmark, not validation for lunar reaction chemistry: it lacks direct DFT force or energy errors, long trajectories, defects, amorphous regolith, and independent reproduction.
A v1 arXiv preprint submitted on July 10, 2026 benchmarks six universal machine-learning force fields on four minerals relevant to lunar regolith. The work tests whether models trained broadly across materials can preserve plausible local structures when transferred to forsterite, fayalite, ilmenite, and anorthite without domain-specific retraining.
What the authors tested
The authors ran all 24 model-mineral combinations for 1,000 molecular-dynamics steps at 300 K, a trajectory of 1 picosecond, and compared local structures with crystallographic references. The six models were MACE-MH, MatterSim, SevenNet-0, UPET, UMA, and NequIP-OAM-L. Each crystal was first relaxed, then simulated at fixed volume with a Langevin thermostat. Separate hydroxylated surface runs examined whether oxygen-hydrogen bonds remained stable.
All combinations completed the short trajectories without atom ejection, rapid collapse, or divergent bond lengths. Silicon-oxygen, magnesium-oxygen, aluminum-oxygen, and calcium-oxygen local environments were comparatively consistent, while iron-oxygen and titanium-oxygen coordination showed broader distributions and larger short-timescale fluctuations. The paper treats that difference as a reason for more ground-truth validation before studying redox processes in iron- and titanium-bearing phases.
| Benchmark layer | What was measured | Practical meaning |
|---|---|---|
| Short-run stability | Temperature, atom retention, and bond behavior | Finds immediate catastrophic transfer failures |
| Local structure | Bond distances, angles, and radial distributions | Checks agreement with crystal references |
| Hydroxylated surfaces | Oxygen-hydrogen distance distributions | Screens stable initial surface configurations |
| Runtime | Wall time and peak GPU allocation | Compares operational cost on one workstation GPU |
Performance results
On one NVIDIA RTX 4090, SevenNet-0 completed the tested trajectories in 33.2 to 59.8 seconds, depending on mineral size. MatterSim took 40.1 to 71.9 seconds and used the least peak memory in the largest tested cell. The ranking is useful for this software and hardware configuration, but it is not a general cost benchmark: compiler versions, neighbor lists, precision, batch strategy, and model implementations can change throughput.
The public repository includes input structures, benchmark and memory-profiling scripts, environment setup, model-loading workflows, and analysis notebooks. It does not include the generated trajectories, intermediate tables, or figures, although the authors say those outputs can be regenerated. Separate environments are required because the model stacks have conflicting dependencies.
What the benchmark does not prove
A stable picosecond trajectory is a smoke test, not evidence that a force field predicts lunar chemistry accurately. The study holds crystal volumes fixed and compares structural distributions with crystallographic references, but it does not report direct DFT force or energy errors for the tested states. It does not test long diffusion times, defects, amorphous grains, irradiation damage, micrometeoroid impact conditions, proton transfer, water formation, desorption, or redox reactions. The preprint is also an unreviewed first version with no independent reproduction.
The cross-model agreement on oxygen-hydrogen distances is encouraging for stable initial configurations, but agreement among models trained on overlapping materials data is not independent ground truth. Reactive pathways can fail even when equilibrium bond lengths look plausible.
LDS analysis
The benchmark's most valuable output is a map of where transferability needs stronger tests, not a declaration that universal force fields are lunar-ready. For practitioners, the next validation layer should pair representative snapshots with DFT energies, forces, and stresses; include disordered and defect-rich structures; test temperatures and pressures relevant to actual processes; and report uncertainty or disagreement across models.
That workflow turns a quick foundation-model screen into a defensible model-selection process. Until those checks exist, SevenNet-0's speed and the broad structural stability are useful engineering signals, while detailed lunar volatile and reaction predictions remain hypotheses.
Key Points
- 1Six universal force fields remained stable across short simulations of four lunar-relevant minerals without domain-specific retraining.
- 2Iron-oxygen and titanium-oxygen coordination varied more across models, identifying the strongest need for domain-specific ground-truth validation.
- 3The open scripts improve reproducibility, but short fixed-volume runs cannot validate reaction chemistry, defects, or long-timescale transport.
Scoring Rationale
The open benchmark is useful to scientific-ML practitioners evaluating transferability and compute cost, while its short trajectories, missing direct DFT errors, and unreviewed status limit scientific conclusions.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
