Pith. sign in

REVIEW 2 major objections 6 minor 30 references

Six universal ML force fields keep lunar minerals stable in short MD runs, but Fe and Ti environments need more care.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 01:03 UTC pith:JWRTJLMC

load-bearing objection Useful first head-to-head of six foundation MLIPs on the four main lunar minerals, with honest scope and a public repo; short fixed-cell NVT vs static refs is the real limit, not a hidden flaw. the 2 major comments →

arxiv 2607.09005 v1 pith:JWRTJLMC submitted 2026-07-10 physics.geo-ph

Benchmarking Universal Machine Learning Force Fields for Molecular Dynamics of Lunar Regolith Minerals

Ziyu Huang , Ken-ichi Nomura This is my paper
classification physics.geo-ph
keywords foundation modelsmachine-learning interatomic potentialslunar regolithmolecular dynamicsforsteritefayaliteilmeniteanorthite
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Universal machine-learning force fields promise fast atomic simulations without hand-tuned parameters for every mineral, yet no one had checked whether they actually work for the silicates and oxides that make up lunar regolith. This paper runs short constant-temperature molecular-dynamics trajectories of forsterite, fayalite, ilmenite and anorthite with six foundation models and measures how well bond lengths, angles and radial distributions match crystallographic references. The models keep the crystals intact and reproduce Si–O, Mg–O, Al–O and Ca–O environments reasonably well; Fe–O and Ti–O shells are systematically broader and more fluctuating. Hydroxylated surfaces give nearly identical O–H distances across all models, and the authors also report GPU timings so users can choose speed versus coverage. The result is a practical baseline for using these models as starting points for lunar volatile, space-weathering and ISRU studies, while flagging Fe- and Ti-bearing phases as the clear next target for fine-tuning.

Core claim

Across six foundation models, short NVT MD trajectories of four representative lunar minerals remain stable at 300 K; Si–O, Mg–O, Al–O and Ca–O first-shell distances and angles stay close to crystallographic ranges, while Fe–O and Ti–O distributions are systematically broader, and hydroxylated surfaces yield consistent O–H distances, establishing an initial transferability baseline for lunar-mineral simulations.

What carries the argument

A uniform six-model NVT MD protocol on fixed crystallographic supercells of forsterite, fayalite, ilmenite and anorthite, scored by temperature stability, first-shell bond statistics, bond-angle distributions and partial radial distribution functions against static crystallographic references, plus hydroxylated-surface O–H checks and single-GPU throughput/memory profiling.

Load-bearing premise

That one-picosecond fixed-cell runs at room temperature, compared only to static crystal structures rather than finite-temperature first-principles dynamics or experiment, are enough to judge model transferability and to attribute the extra Fe and Ti fluctuations to artificial softening of the potentials.

What would settle it

Run the same four minerals with each model for nanoseconds or longer, or against finite-temperature AIMD/DFT references that include magnetic and correlation effects for Fe and Ti; if the Fe–O and Ti–O shells then collapse to the same width as the Si–O shells, or if the crystals become unstable, the present transferability claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Ordered Mg- and Ca-bearing lunar silicates can already be screened with these foundation models without material-specific reparameterization.
  • Fe- and Ti-bearing phases require targeted fine-tuning on lunar-relevant ground-truth data before redox or space-weathering simulations are trusted.
  • Consistent O–H distances make the models usable starting points for high-throughput hydroxyl-stability and volatile-retention screens.
  • SevenNet-0, MatterSim and UPET currently offer the highest practical throughput for short MD screening on a single high-end GPU.
  • The same protocol supplies a reusable baseline for later extensions to defects, amorphous surfaces, impacts and polar-sample chemistry.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the broader Fe–O/Ti–O fluctuations really come from missing spin or correlation physics, foundation-model developers may need explicit magnetic or DFT+U training data before multi-element lunar regolith models become reliable.
  • The clear performance ranking suggests that mixed-model workflows—fast screening with SevenNet-0/MatterSim followed by higher-cost refinement—could become standard for large-scale ISRU or space-weathering campaigns.
  • Because O–H consistency holds even on Fe- and Ti-bearing surfaces, the next decisive test is whether the same models preserve proton-transfer barriers and water-formation pathways under solar-wind conditions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. This manuscript benchmarks six universal machine-learning interatomic potentials (MACE-MH, MatterSim, SevenNet-0, UPET, UMA, NequIP-OAM-L) on short NVT molecular-dynamics trajectories of four lunar-relevant minerals (forsterite, fayalite, ilmenite, anorthite) plus simple hydroxylated surfaces. Structural fidelity is assessed via temperature stability, first-shell bond distances and angles, and partial RDFs against static crystallographic ranges (Table 1, Figs. 2–5); O–H distances on hydroxylated slabs are also compared (Fig. 6); and wall-clock time and peak GPU memory are reported on an RTX 4090 (Table 2, Fig. 7). All 24 model–mineral combinations remain stable over 1 ps at 300 K. Si–O, Mg–O, Al–O and Ca–O environments are reproduced reasonably well, while Fe–O and Ti–O distributions are broader; O–H distances are consistent across models. The authors frame the work as an initial transferability baseline and explicitly call for ab initio validation and fine-tuning before redox or volatile-reaction applications.

Significance. If the reported stability, structural statistics and performance ranking hold, the paper supplies a timely, reproducible baseline for applying foundation MLIPs to lunar regolith mineralogy—an area previously limited by sparse reactive force-field parameterizations. The multi-model head-to-head design, public repository of structures and scripts, and explicit GPU throughput/memory numbers are concrete strengths that lower the barrier for subsequent fine-tuning and larger-scale space-weathering or ISRU simulations. The scoped claim (short-timescale ordered crystals plus simple OH surfaces) is appropriate for an initial benchmark and is useful to the planetary-materials and computational-materials communities even without new reaction pathways.

major comments (2)
  1. §3.3 and Table 1: The attribution of broader Fe–O/Ti–O standard deviations to “artificial softening” of the potential around transition metals (lack of spin-dependent features, shallower minima relative to DFT+U) is presented as the likely explanation, yet the only references are static crystallographic first-shell ranges, not finite-temperature AIMD or experiment. The paper already notes this limitation in §4; the Discussion should either (i) soften the causal language to “consistent with possible softening, pending AIMD comparison” or (ii) add a short AIMD or literature thermal-broadening comparison for at least one Fe/Ti phase so the claim is not left as an untested inference.
  2. Methods 2.3–2.4 and §3.1: The central stability/transferability claim rests on 1 ps fixed-cell NVT trajectories after FIRE relaxation to 0.05 eV/Å. That protocol is adequate for an initial screen, but the manuscript should state more explicitly in the Abstract/Conclusions that the benchmark does not yet constrain thermal expansion, equation of state, defect energetics, or reaction barriers—quantities that matter for the space-weathering and ISRU applications listed in the Abstract. A single clarifying sentence would prevent over-reading of the present results.
minor comments (6)
  1. Abstract and Introduction: “remains elucidated” appears to be a wording error; “remains to be elucidated” or “remains unclear” is intended.
  2. Figure 1 caption lists forsterite/fayalite order inconsistently with the panel labels (a–d); align caption order with the figure layout.
  3. Table 1 vs. Methods 2.4: Bond-distance cutoffs used for analysis (e.g., Mg/Fe–O 2.8 Å in text, 2.7 Å in Fig. 4 caption) should be made fully consistent across Methods, figure captions and table footnotes.
  4. Figure 3 caption states bin width 0.03 Å while Methods 2.4 states 0.02 Å; reconcile.
  5. §2.2: Brief one-line notes on the training-data coverage of UMA and NequIP-OAM-L (analogous to the MPtrj/OMAT notes for the other models) would help readers interpret the Fe/Ti performance differences.
  6. Performance section: Peak-memory entries that are “unavailable” for some backends are mentioned in Methods but not flagged in Table 2; a footnote would avoid confusion.

Circularity Check

0 steps flagged

Empirical head-to-head MLIP benchmark against external crystallographic structures and GPU timings; no fitted inputs renamed as predictions and no load-bearing self-citation chain.

full rationale

The paper is a transferability and performance benchmark, not a derivation of a physical law or a fitted predictive model. Six foundation MLIPs are applied as black-box force calculators to four lunar mineral supercells (and simple hydroxylated slabs) under fixed-cell NVT MD at 300 K. Structural fidelity is assessed by comparing MD bond distances, angles, and partial RDFs to static first-shell ranges recomputed from independent crystallographic supercells (Hazen, Smyth, Wechsler & Prewitt, Wainwright & Starkey) and to Materials Project structures. Stability is reported as the absence of atom ejection or bond divergence over 1 ps; O–H distributions are reported as observed consistency across models; wall-clock time and peak GPU memory are measured on an RTX 4090. No free parameters are fitted to the lunar-mineral data and then re-presented as predictions. Self-citations to the authors’ prior ReaxFF lunar MD papers appear only as motivation for the application domain and do not underwrite the numerical results. The authors themselves scope the work as an “initial benchmark” and flag the need for ab initio validation and fine-tuning of Fe/Ti environments. Consequently the derivation chain is self-contained against external references and contains no circular reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper is an empirical benchmark; its claims rest on standard MD protocol choices, literature crystallographic references, and the assumption that short fixed-cell NVT runs plus static first-shell ranges are informative. No new physical entities are postulated. Free parameters are the usual simulation cutoffs and thresholds chosen by the authors.

free parameters (4)
  • species-pair bond cutoffs = Si–O 2.0 Å; Mg/Fe–O 2.8 Å; Ti–O 2.6 Å; Al–O 2.4 Å; Ca–O 3.2 Å
    Fixed cutoffs (Si–O 2.0 Å, Mg/Fe–O 2.8 Å, Ti–O 2.6 Å, Al–O 2.4 Å, Ca–O 3.2 Å) are chosen by hand to define first-shell statistics; results near the boundaries are sensitive to these values.
  • FIRE force convergence threshold = 0.05 eV/Å
    Geometry relaxation stops at max force < 0.05 eV Å⁻¹; this hand-chosen tolerance affects the starting structures for MD.
  • NVT trajectory length and timestep = Δt = 1.0 fs, 1000 steps
    1 fs timestep for 1000 steps (1 ps) is a short, hand-selected protocol used both for equilibration and for all structural statistics.
  • surface protonation geometry = 0.98 Å initial O–H; 1.0 Å height cutoff
    Initial O–H distance of 0.98 Å and selection of non-bridging surface oxygens within 1.0 Å of the topmost O are author-chosen construction rules for the hydroxylated slabs.
axioms (4)
  • domain assumption Static crystallographic first-shell distance ranges are valid external references against which finite-temperature MD means and widths can be judged.
    Invoked throughout §3.3 and Table 1; the paper itself notes these ranges describe site inequivalence, not thermal broadening.
  • ad hoc to paper Successful completion of 1 ps NVT without atom ejection or bond divergence constitutes a meaningful stability and transferability check for lunar compositions.
    Stated in Methods 2.3 and Results 3.1; longer or NPT runs, defects, or high-T conditions are left for future work.
  • domain assumption Broader Fe–O and Ti–O distributions relative to Si–O primarily reflect artificial softening of the ML potential energy surface rather than physical thermal or magnetic effects.
    Advanced in §3.3 with citation to Deng et al.; no spin-polarized DFT or AIMD comparison is performed in the present work.
  • domain assumption The six chosen foundation models (MACE-MH, MatterSim, SevenNet-0, UPET, UMA, NequIP-OAM-L) are representative of the current generation of universal MLIPs.
    Selection justified in §2.2; other recent models are omitted.

pith-pipeline@v1.1.0-grok45 · 18075 in / 3269 out tokens · 36821 ms · 2026-07-13T01:03:59.671658+00:00 · methodology

0 comments
read the original abstract

Universal machine-learning interatomic potentials provide a promising route for accelerating molecular dynamics simulations of materials, but their transferability to lunar regolith-relevant silicates, oxides, and hydrogen-bearing surface species remains elucidated. Here, we benchmark six foundation models, MACE-MH, MatterSim, SevenNet-0, UPET, UMA, and NequIP-OAM-L, using NVT molecular dynamics simulations of four representative lunar minerals: forsterite, fayalite, ilmenite, and anorthite. Structural fidelity is evaluated using temperature stability, bond-distance statistics, bond-angle distributions, and partial radial distribution functions, with comparison to crystallographic reference data. The models reproduce Si--O, Mg--O, Al--O, and Ca--O local environments reasonably well, while Fe--O and Ti--O coordination environments show broader distributions and larger short-timescale fluctuations, highlighting the need for further validation and fine tuning with additional ground truth data for Fe- and Ti-bearing lunar phases. Hydroxylated surface tests show consistent O--H bond-distance distributions across models and minerals, suggesting that these foundation models may provide useful starting points for screening surface hydroxyl stability and volatile-related processes. Performance benchmarks on a single NVIDIA RTX 4090 show that SevenNet-0, MatterSim, and UPET provide the highest throughput among the six tested models, MACE-MH remains practical at intermediate cost, and UMA and NequIP-OAM-L extend the comparison to newer foundation potentials at higher runtime cost and memory demand. These results provide an initial benchmark for applying universal foundation models to lunar mineral simulations and identify key directions for future ab initio validation, model fine-tuning, and applications to lunar volatile evolution, space weathering, ISRU, and polar sample return studies.

Figures

Figures reproduced from arXiv: 2607.09005 by Ken-ichi Nomura, Ziyu Huang.

Figure 1
Figure 1. Figure 1: Crystal structures of the four lunar regolith minerals studied. Atomic positions are shown as spheres coloured by element (Mg: green; Fe: brown; Ti: purple; Ca: Cyan; Al: orange; Si: gold; O: red). From left to right: anorthite (CaAl2Si2O8, 208 atoms), forsterite (Mg2SiO4, 336 atoms), fayalite (Fe2SiO4, 336 atoms), ilmenite (TiFeO3, 360 atoms). Supercell boundaries are indicated by black lines. In addition… view at source ↗
Figure 2
Figure 2. Figure 2: Temperature (top row) and per-atom potential energy (bottom row) as a function of simulation time for all six foundation models on each mineral. The dashed grey line marks the target temperature of 300 K. All models reach thermal equilibrium within 100–200 fs and maintain stable dynamics throughout the 1 ps trajectory. Model colours: SevenNet-0 (blue), MatterSim (green), UPET (orange), MACE-MH (red), UMA (… view at source ↗
Figure 3
Figure 3. Figure 3: Partial radial distribution functions gαβ(r) computed from the last 500 frames of each trajectory using a bin width of 0.03,˚A. Each panel corresponds to one mineral, and columns within each panel show chemically distinct pair types. The dashed horizontal line at g(r) = 1 marks the ideal-gas uncorrelated limit. Sharp first-shell peaks are observed for Si–O and Al–O pairs, confirming the integrity of tetrah… view at source ↗
Figure 4
Figure 4. Figure 4: Bond-distance probability distributions from the final 500 frames for all six models and minerals. Distances are sampled every 5 frames using species-pair cutoffs to select first-shell pairs only: Si–O, 2.0,˚A; Mg/Fe–O, 2.7,˚A; Ti–O, 2.4,˚A; Al–O, 2.2,˚A; and Ca–O, 3.0,˚A. The Si–O and Al–O distributions are tightly peaked, while the Fe–O and Ti–O distributions are substantially broader, consistent with th… view at source ↗
Figure 5
Figure 5. Figure 5: Bond angle probability distributions from the final 500 frames. Columns within each panel correspond to chemically distinct angle types within each mineral. All six models reproduce O–Si–O and O–Al–O tetrahedral geometry within ±0.8 ◦ of each other. The broader O–Ti–O and O–Fe–O distributions in ilmenite reflect the distorted octahedral environment of the corundum-type structure. Top to bottom: forsterite,… view at source ↗
Figure 6
Figure 6. Figure 6: O–H bond-distance distributions for hydroxylated surfaces of representative lunar minerals predicted by the six foundation models. The distributions show strong consistency across models and mineral surfaces, indicating stable surface hydroxyl groups during the short NVT trajectories. Top to bottom: forsterite (Mg2SiO4), fayalite (Fe2SiO4), ilmenite (TiFeO3), and anorthite (CaAl2Si2O8). 3.6 Computational p… view at source ↗
Figure 7
Figure 7. Figure 7: Wall-clock timing comparison on an NVIDIA GeForce RTX 4090 (24 GB). (a) Total wall-clock time (s) for 1000 NVT-MD steps per mineral and model, shown as grouped bars. (b) Mean time per MD step (ms/step) averaged across all four mineral supercells. SevenNet-0 and MatterSim achieve the highest throughput, followed by UPET and MACE-MH, while UMA and NequIP-OAM-L provide broader new-model coverage at higher wal… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 2 linked inside Pith

  1. [1]

    Batatia et al

    I. Batatia et al. A foundation model for atomistic materials chemistry.arXiv preprint arXiv:2401.00096, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Batzner, A

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials.Nat. Commun., 13:2453, 2022

    2022

  3. [3]

    Chen and S

    C. Chen and S. P. Ong. A universal graph deep learning interatomic potential for the periodic table.Nat. Comput. Sci., 2:718–728, 2022

    2022

  4. [4]

    B. Deng, Y. Choi, P. Zhong, J. Riebesell, S. Anand, Z. Li, K. Jun, K. A. Persson, and G. Ceder. Systematic softening in universal machine learning interatomic potentials.npj Computational Materials, 11(1):9, 2025

    2025

  5. [5]

    V. L. Deringer, A. P. Bart´ ok, N. Bernstein, D. M. Wilkins, M. Ceriotti, and G. Cs´ anyi. Gaussian process regression for materials and molecules.Chem. Rev., 121:10073–10141, 2021

    2021

  6. [6]

    Georgiou, Z

    A. Georgiou, Z. Huang, L. H. Yeo, W. Farrell, S. Verkercke, J. R. Lewis, C. Dong, and L. S. Morrissey. Effect of solar wind and micrometeoroid impact on the lunar water cycle: A molecular dynamics study.Journal of Geophysical Research: Planets, 130(4):e2024JE008687, 2025

    2025

  7. [7]

    Gu´ enol´ e, W

    J. Gu´ enol´ e, W. G. N¨ ohring, A. Vaid, F. Houll´ e, Z. Xie, A. Prakash, and E. Bitzek. Assessment and optimization of the fast inertial relaxation engine (fire) for energy minimization in atomistic simulations and its implementation in lammps.Computational Materials Science, 175:109584, 2020

    2020

  8. [8]

    R. M. Hazen. Effects of temperature and pressure on the crystal structure of forsterite.Am. Mineral., 61:1280–1293, 1976

    1976

  9. [9]

    Huang and M

    Z. Huang and M. Hirabayashi. Coupled space weathering: Nanophase iron formation by micrometeoroid impact and solar wind sputtering.Geophysical Research Letters, 53(7):e2025GL118718, 2026

    2026

  10. [10]

    Huang and M

    Z. Huang and M. Hirabayashi. Revealing exotic nanophase iron in lunar samples through impact-driven spatial fingerprints.The Planetary Science Journal, 7(3):62, 2026

    2026

  11. [11]

    Huang, M

    Z. Huang, M. Hirabayashi, and T. M. Orlando. Micrometeoroid impacts: Dual pathways for iron reduction and oxidation on lunar and asteroidal surfaces.The Astrophysical Journal, 994(2):240, 2025

    2025

  12. [12]

    Huang, K.-i

    Z. Huang, K.-i. Nomura, L. S. Morrissey, and J. Wang. Molecular dynamics simulation of solar wind implantation in the permanently shadowed regions on the lunar surface.Geophysical Research Letters, 49(18):e2022GL099333, 2022

    2022

  13. [13]

    Huang, K.-i

    Z. Huang, K.-i. Nomura, A. Nakano, and J. Wang. Molecular dynamics simulations of dielectric breakdown of lunar regolith: Implications for water ice formation on lunar surface. Geophysical Research Letters, 48(3):e2020GL091681, 2021

    2021

  14. [14]

    Huang, K.-i

    Z. Huang, K.-i. Nomura, and J. Wang. Molecular dynamics simulations of water formation and retention by micrometeoroid impact on lunar surface.Geophysical Research Letters, 48(15):e2021GL093509, 2021. 13 IOP PublishingJournalvv(yyyy) aaaaaa Authoret al

    2021

  15. [15]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, and K. A. Persson. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation.APL Mater., 1:011002, 2013

    2013

  16. [16]

    S. R. Kavanagh and M. G. . Harvard. Nequip & allegro foundation potentials, Aug. 2025

    2025

  17. [17]

    W. S. Kiefer, R. J. Macke, D. T. Britt, A. J. Irving, and G. J. Consolmagno. The density and porosity of lunar rocks.Geophys. Res. Lett., 39:L07201, 2012

    2012

  18. [18]

    A. H. Larsen et al. The atomic simulation environment — a Python library for working with atoms.J. Phys.: Condens. Matter, 29:273002, 2017

    2017

  19. [19]

    Mazitov, F

    A. Mazitov, F. Bigi, M. Kellner, P. Pegolo, D. Tisi, G. Fraux, S. Pozdnyakov, P. Loche, and M. Ceriotti. Pet-mad as a lightweight universal interatomic potential for advanced materials modeling.Nature Communications, 16(1):10653, 2025

    2025

  20. [20]

    Y. Park, J. Kim, S. Hwang, and S. Han. Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations.J. Chem. Theory Comput., 20:4857–4872, 2024

    2024

  21. [21]

    D. Shoji. Reactive molecular dynamics simulations of iron reduction on lunar and asteroidal surfaces by micrometeoroids and solar wind.The Astrophysical Journal, 985(1):143, 2025

    2025

  22. [22]

    D. Shoji. Molecular dynamics simulations of solar-wind induced h2o formation and retention in vesicles of lunar soil.Scientific Reports, 16(1):4531, 2026

    2026

  23. [23]

    J. R. Smyth. High temperature crystal chemistry of fayalite.Am. Mineral., 60:1092–1097, 1975

    1975

  24. [24]

    C. W. Tan, M. L. Descoteaux, M. Kotak, G. de Miranda Nascimento, S. R. Kavanagh, L. Zichi, M. Wang, A. Saluja, Y. R. Hu, T. Smidt, et al. High-performance training and inference for deep equivariant interatomic potentials.Digital Discovery, 5(4):1558–1567, 2026

    2026

  25. [25]

    A. C. Van Duin, S. Dasgupta, F. Lorant, and W. A. Goddard. Reaxff: a reactive force field for hydrocarbons.The Journal of Physical Chemistry A, 105(41):9396–9409, 2001

    2001

  26. [26]

    J. E. Wainwright and J. Starkey. A refinement of the structure of anorthite.Z. Kristallogr., 133:75–84, 1971

    1971

  27. [27]

    B. A. Wechsler and C. T. Prewitt. Crystal structure of ilmenite (FeTiO 3) at high temperature and high pressure.Am. Mineral., 69:176–185, 1984

    1984

  28. [28]

    B. Wood, M. Dzamba, X. Fu, M. Gao, M. Shuaibi, L. Barroso-Luque, K. Abdelmaqsoud, V. Gharakhanyan, J. Kitchin, D. Levine, et al. Uma: A family of universal models for atoms. Advances in Neural Information Processing Systems, 38:129391–129427, 2026

    2026

  29. [29]

    H. Yang, C. Hu, Y. Zhou, X. Liu, Y. Shi, J. Li, G. Li, Z. Chen, S. Chen, C. Zeni, M. Horton, R. Pinsler, A. Fowler, J. Zhu, T. E. Markland, J. Gillan, and Z. Liu. Mattersim: A deep learning atomistic model across elements, temperatures and pressures.arXiv preprint arXiv:2405.04967, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Zhang, W

    P. Zhang, W. Dai, R. Niu, G. Zhang, G. Liu, X. Liu, Z. Bo, Z. Wang, H. Zheng, C. Liu, et al. Overview of the lunar in situ resource utilization techniques for future lunar missions.Space: Science & Technology, 3:0037, 2023. 14

    2023