Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

Universal Machine Learning Potentials under Pressure

T0 review · 4 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Universal machine-learning potentials lose accuracy under pressure; fine-tuning restores it

desk verdict Useful high-pressure benchmark and dataset, but the headline causal claim outruns the evidence; the pressure trend would be cleaner with matched-set analysis. read the letter →

arxiv 2508.17792 v1 pith:K7GWA2JN submitted 2025-08-25 cond-mat.mtrl-sci

Antoine Loew , Jonathan Schmidt , Silvana Botti , Miguel A. L. Marques This is my paper
classification cond-mat.mtrl-sci
keywords machinelearninginteratomicpotentialsuniversalMLIPshighpressureDFTdatasetfine-tuningextrapolationmaterialsdiscovery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Universal machine-learning interatomic potentials (uMLIPs) aim to give density-functional-theory accuracy at a tiny fraction of the cost, but this paper argues their reliability stops at ambient pressure. Benchmarking eight leading uMLIPs on a new dataset of about 190,000 materials and 32 million density-functional calculations, the authors find energy errors that grow from a few meV per atom at 0 GPa to tens or hundreds of meV per atom at 150 GPa, with the worst degradation in models trained only on near-equilibrium ambient structures. The same data, used to fine-tune two of the models, brings the 150 GPa error back down to about 31 meV per atom. The paper concludes that the pressure blind spot is a training-data problem, not an algorithmic one, and that compressed configurations belong in the next generation of universal potentials.

What carries the argument

The load-bearing object is a new pressure dataset built by taking about 190,000 structures from the base DFT database, re-relaxing them with the same PBE settings at 0, 25, 50, 75, 100, 125, and 150 GPa, and recording equilibrium structures, total energies, forces, stress tensors, and relaxation-path configurations—32 million single-point calculations in all. The benchmark protocol then runs each uMLIP through the same relaxations and compares final relaxed energies and volumes to the PBE reference, while the 90–5–5 material-level split prevents leakage between training, validation, and test sets.

What would settle it

Take the materials whose high-pressure relaxations did not converge, re-relax them with a more robust DFT protocol, and measure the energy and volume errors of the fine-tuned models on that held-out set; if the energy error is substantially larger than 31 meV/atom, the reported pressure recovery is incomplete for the hardest systems.

Watch

Extended reading notes

Core claim

The central claim is that uMLIP accuracy under pressure is governed by whether compressed atomic environments appear in the training distribution. Across pressures from 0 to 150 GPa, the benchmark shows a systematic decline: the least-affected model goes from 4.1 to 41.7 meV/atom mean absolute energy error, while an earlier ambient-trained model rises from 33.8 to 346.8 meV/atom. Models that saw high-pressure or non-equilibrium structures during training degrade less, and targeted fine-tuning on the pressure dataset reduces the best 150 GPa error to about 31 meV/atom at the cost of a modest loss at ambient pressure. The authors read this as evidence that the bottleneck is data coverage, not

Load-bearing premise

The benchmark and fine-tuning results exclude materials whose pressure relaxations did not converge in DFT (roughly one to three thousand of 190,000 per pressure), and the paper assumes that missing those cases does not bias the measured errors.

Editorial extensions

If this is right

  • At pressures above 25 GPa, untuned uMLIP predictions of relaxed energies and volumes should be treated as unvalidated; errors can exceed 100 meV/atom.
  • Fine-tuning on high-pressure configurations is a cheap and effective fix; two models dropped their 150 GPa energy error to roughly 31 meV/atom.
  • Ambient-pressure rankings do not predict pressure behavior; the most accurate ambient model degrades fastest under compression.
  • Training strategy matters: models exposed to high-pressure molecular dynamics or non-equilibrium denoising extrapolate markedly better.
  • The dataset doubles as a reusable benchmark for measuring progress in out-of-distribution generalization of future universal potentials.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported fine-tuned errors are computed on materials whose high-pressure relaxations converged; if the roughly one to three thousand non-converged materials per pressure are the hardest cases, real-world errors on those systems could be larger—a direct test would be to relax them with a more robust protocol and remeasure.
  • The same fine-tuning recipe was demonstrated on only two model families; applying it to the other benchmarked uMLIPs would likely shrink their high-pressure errors too, since their degradation follows the same data-coverage pattern.
  • The result suggests a design rule for the next generation: sample the compressed structural manifold explicitly rather than relying on ambient datasets augmented by rattling or volume scaling.
  • Nothing in the data establishes behavior beyond 150 GPa; extending to terapascal conditions, relevant for planetary interiors, would require new training data in that regime.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper introduces a large DFT dataset of roughly 190k materials relaxed with PBE at pressures from 0 to 150 GPa, and uses it to benchmark eight universal machine learning interatomic potentials (uMLIPs) by comparing PBE reference volumes and energies against model relaxations. The authors report that most uMLIPs become substantially less accurate with increasing pressure, and show that fine-tuning two representative models (MatterSim-v1 and eSEN-30M-OAM) on high-pressure configurations largely recovers accuracy. The central claim is that the pressure-induced degradation stems from limitations in the training data rather than algorithmic constraints, and that targeted fine-tuning is an effective remedy.

Significance. If the central claim holds, this work provides a valuable resource: a consistent, large-scale high-pressure DFT dataset, a multi-model benchmark of practical relevance, and a concrete demonstration that fine-tuning on high-pressure configurations improves transferability. The study uses consistent DFT settings, trajectory-level split to avoid leakage, and a broad set of contemporary uMLIPs. The dataset and code are promised to be released. However, the quantitative headline is currently supported mainly by aggregate MAEs from a single split, with no matched-set control across pressures and no analysis of the discarded non-converged calculations, so several load-bearing components of the claim need additional evidence before the conclusions can be accepted at face value.

major comments (4)
  1. [Section II.A, Table I vs Tables II/III] The pressure trend is computed on non-identical material pools. Table I shows 162k materials at 0 GPa versus 187-190k at every elevated pressure, with the 0 GPa deficit explicitly attributed to missing relaxation paths, not only to non-convergence. Because the 90/5/5 split is applied at the level of the full 190k set, a material absent at 0 GPa can nevertheless be in the test set at 25-150 GPa. The monotonic MAE increases in Tables II and III could therefore be partly caused by a change in test composition rather than pressure. Please provide a matched-set analysis, e.g., restricting to materials with converged data at all pressures, and report the test-set compositions; if the trend persists, this concern is fully resolved.
  2. [Table I caption and Section II.A] The manuscript drops calculations that did not converge under pressure without further analysis. Table I states that missing materials under pressure 'concern calculations that did not converge,' but there is no discussion of how many materials are affected per pressure, why they failed, or whether they are systematically the hardest high-pressure systems. If non-converged cases are preferentially difficult, the reported errors are optimistic and fine-tuning gains may not transfer to the most challenging materials. Please quantify the convergence failures and perform a sensitivity analysis, e.g., comparing models on the converged subset against bounds that include the non-converged systems.
  3. [Tables II/III and Section II.B] All quantitative claims rest on a single 90/5/5 split with no uncertainty estimates. The differences between some models and the fine-tuning improvements are small relative to the spread of values (e.g., eSEN at 150 GPa changes from 41.7 to 32.4 meV/atom; several models differ by only a few meV/atom at low pressure). Without repeated splits, bootstrap confidence intervals, or multiple seeds, the reported ranking and the magnitude of the fine-tuning benefit are not statistically grounded. Please provide uncertainty estimates for the MAEs in Tables II and III, at least for the test-set metric.
  4. [Abstract and Section III] The attribution 'originates from fundamental limitations in the training data rather than algorithmic constraints' is stronger than the evidence supports. Fine-tuning on high-pressure data improves the two tested models, which is consistent with a data-coverage explanation, but it does not rule out algorithmic constraints: eSEN-30M-OAM, without explicit high-pressure fine-tuning, is already the best model under pressure, suggesting that training strategy or architecture has a large effect. A cleaner test would be to train or fine-tune the same architecture on datasets with and without high-pressure configurations, or at least to soften the causal claim and limit it to 'the models' original training distributions lack sufficient high-pressure coverage.'
minor comments (7)
  1. [Table I caption] Typo: 'dataset develop in this work' should be 'dataset developed in this work.'
  2. [Section I] Duplicate phrase: 'materials materials science' near the end of the introductory paragraph.
  3. [Section II.A vs Section III] The dataset size is given as 32 million single-point calculations in Section II.A and 30 million in Section III. Please make this consistent.
  4. [Section IV.A] The code is referred to as 'v asp' with missing spacing; should be 'VASP.'
  5. [Figures 3/4] Axis labels use 'A3' and 'A' without proper superscripts or math formatting; this should be fixed for clarity.
  6. [Section II.B] The reasons for MACE-MPA-0's 'density renormalization' and its effect on high-pressure performance are mentioned but not explained or quantitatively assessed. A brief explanation or citation would help.
  7. [Section IV.B] The exact model checkpoints and versions (e.g., MatterSim-v1.0.0-5M) are listed, but for full reproducibility the commit hashes or release IDs of the pretrained checkpoints and libraries should be provided.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark rests on independent PBE DFT reference and held-out evaluation.

full rationale

The paper's central empirical claims are evaluated against freshly computed PBE DFT reference data (Section II.A, Table I) that is independent of the uMLIP predictions. The uMLIPs are external pretrained models (M3GNet, MACE-MPA-0, SevenNet, DPA3, GRACE, ORB, MatterSim, eSEN); their errors are measured as differences to DFT relaxed structures and energies (Tables II–III). No target quantity is defined in terms of model outputs, and no prediction is a fitted parameter renamed as a result. The fine-tuning experiments use a 90/5/5 split at the material level, so test materials are held out from fine-tuning; improved errors on those held-out materials are genuine out-of-sample results. The only self-referential element is use of the authors' Alexandria database as the starting point for the new pressure dataset; this is provenance, not a load-bearing assumption, and the benchmark would be meaningful even if Alexandria had been built by others. The paper itself flags two limitations: (i) 0 GPa lacks ~28k older trajectories and high-pressure non-converged calculations are excluded (Table I caption), and (ii) possible leakage at 0 GPa for models pretrained on Alexandria (Section II.B). These affect comparability across pressures and could make the 0 GPa baseline look better, but they do not make the pressure trend or fine-tuning gains equivalent to the paper's assumptions by construction. No equation is self-referential and no result is imported via self-citation.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central benchmark rests on the choice of PBE as the reference, the representativeness of the Alexandria structures, the completeness of converged calculations, and the chosen metrics. The fine-tuning comparison also depends on several ad hoc hyperparameters and filters. There are no newly postulated physical entities.

free parameters (7)
  • Subsampling energy threshold = 10 meV/atom
    Used to select configurations from merged relaxation paths in Section IV B; chosen by the authors, affects the training/benchmark composition.
  • Outlier energy bounds = -18 to 10 eV/atom
    Ad-hoc filter in Section IV B that removes 0.3% of data; shifting bounds changes the fine-tuning dataset.
  • Outlier force/stress thresholds = max force < 100 eV/A, min stress > -200 eV/A^3
    Ad-hoc filters in Section IV B, no sensitivity analysis.
  • eSEN fine-tuning hyperparameters = lr=4e-4, wd=0.001, 100 epochs, loss weights 20/200/200
    Section IV B; no ablation or sensitivity analysis.
  • MatterSim fine-tuning hyperparameters = lr=2e-4, loss schedule 1-100-0.1, 1-1-0.5, 1-10-1
    Section IV B; ad hoc schedule based on epochs 20 and 61.
  • Pressure grid = 0, 25, 50, 75, 100, 125, 150 GPa
    Chosen pressure points; results could differ with a finer or extended grid.
  • Train/val/test split = 90/5/5
    Single split; no repeated splits or confidence intervals reported.
assumptions (4)
  • domain assumption PBE DFT is the ground truth for high-pressure structural and energetic properties.
    All benchmark references and fine-tuning labels come from PBE relaxations (Section IV A); no comparison to experiment or higher-level theory.
  • domain assumption The Alexandria-derived 190k structures are representative of the materials space relevant to uMLIPs.
    The pressure dataset is built by re-relaxing Alexandria structures (Section II A); biases in Alexandria propagate to the benchmark.
  • domain assumption Excluding non-converged DFT calculations does not bias benchmark conclusions.
    Table I notes materials were dropped when calculations did not converge; no analysis of their properties is provided.
  • domain assumption Relaxed end-state volumes and energies are the appropriate metrics for pressure transferability.
    Errors are computed at relaxed geometries (Section II B); other observables such as phase transitions or kinetic barriers are not tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Universal Machine Learning Potentials under Pressure." pith.science (2026). https://pith.science/paper/K7GWA2JN

@misc{pith2026250817792,
  author       = {Pith},
  title        = {Pith review of: Universal Machine Learning Potentials under Pressure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K7GWA2JN}},
  note         = {Machine review of arXiv:2508.17792}
}
read the original abstract

Universal machine learning interatomic potentials (uMLIPs) represent arguably the most successful application of machine learning to materials science, demonstrating remarkable performance across diverse applications. However, critical blind spots in their reliability persist. Here, we address one such significant gap by systematically investigating the accuracy of uMLIPs under extreme pressure conditions from 0 to 150 GPa. Our benchmark reveals that while these models excel at standard pressure, their predictive accuracy deteriorates considerably as pressure increases. This decline in performance originates from fundamental limitations in the training data rather than in algorithmic constraints. In fact, we show that through targeted fine-tuning on high-pressure configurations, the robustness of the models can be easily increased. These findings underscore the importance of identifying and addressing overlooked regimes in the development of the next generation of truly universal interatomic potentials.

Figures

Figures reproduced from arXiv: 2508.17792 by the authors.

Figure 1
Figure 1. FIG. 1. Violin plot showing the distribution of first neighbor [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Violin plot showing the distribution of unit cell vol [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Volume error after structural optimization: compar [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Error in predicted energy/atom after structural opti [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Active learning and explicit electrostatics enable accurate modeling of electrolytes

    physics.chem-ph 2025-10 conditional novelty 6.0 of 10

    Automated active learning produces transferable MTPs for EC/EMC/LiPF6 electrolytes, while explicit electrostatics (QRd) matches accuracy with fewer parameters only where its MD remains stable.

  2. VASP Plugins: Linking the Vienna ab-initio Simulation Package with Python

    cond-mat.mtrl-sci 2026-07 accept novelty 5.5 of 10

    A C++/pybind11 shared-memory plugin layer exposes VASP SCF and ionic data as NumPy arrays so Python can modify structure, forces, local potential, and occupancies in place.

Reference graph

Works this paper leans on

52 extracted references · 37 canonical work pages · cited by 2 Pith papers

  1. [1]

    P. F. McMillan, New materials from high-pressure exper- iments, Nat. Mater.1, 19–25 (2002)

    work page 2002

  2. [2]

    Zhang, Y

    L. Zhang, Y. Wang, J. Lv, and Y. Ma, Materials discov- ery at high pressures, Nat. Rev. Mater.2, 17005 (2017)

    work page 2017

  3. [3]

    M. I. McMahon and R. J. Nelmes, High-pressure struc- tures and phase transformations in elemental metals, Chem. Soc. Rev.35, 943 (2006)

    work page 2006

  4. [4]

    Loubeyre, F

    P. Loubeyre, F. Occelli, and P. Dumas, Synchrotron in- frared spectroscopic evidence of the probable transition to metal hydrogen, Nature577, 631–635 (2020)

    work page 2020

  5. [5]

    M. Moog, F. Pietrucci, and A. M. Saitta, Carbon dioxide under earth mantle conditions: From a molecular liquid through a reactive fluid to polymeric regimes, J. Phys. Chem. A125, 5863–5869 (2021)

    work page 2021

  6. [6]

    Grochala, R

    W. Grochala, R. Hoffmann, J. Feng, and N. Ashcroft, The chemical imagination at work in very tight places, Angew. Chem., Int. Ed.46, 3620–3642 (2007)

    work page 2007

  7. [7]

    Popov, M.Kyotani, R.J.Nemanich,andY.Koga,Su- perhardphasecomposedofsingle-wallcarbonnanotubes, Phys

    M. Popov, M.Kyotani, R.J.Nemanich,andY.Koga,Su- perhardphasecomposedofsingle-wallcarbonnanotubes, Phys. Rev. B65, 33408 (2002)

    work page 2002

  8. [8]

    M. I. Eremets, A. G. Gavriliuk, I. A. Trojan, D. A. Dzivenko, and R. Boehler, Single-bonded cubic form of nitrogen, Nat. Mater.3, 558–563 (2004)

    work page 2004

Show all 52 references
  1. [9]

    Tang, H.-w

    J. Tang, H.-w. Yang, and G.-b. Cheng, Nitrogen-rich tetracyclic-based heterocyclic energetic materials, Energ. Mater. Front.4, 110–122 (2023)

    2023

  2. [10]

    M. I. Eremets and I. A. Troyan, Conductive dense hy- drogen, Nat. Mater.10, 927–931 (2011)

    2011

  3. [11]

    V. Iota, C. S. Yoo, and H. Cynn, Quartzlike carbon diox- ide: An optically nonlinear extended solid at high pres- sures and temperatures, Science283, 1510–1513 (1999)

    1999

  4. [12]

    V. M. Giordano and F. Datchi, Molecular carbon dioxide at high pressure and high temperature, Europhys. Lett. 77, 46002 (2007)

    2007

  5. [13]

    F. Li, J. Xu, Y. Wang, H. Zheng, and K. Li, Pressure- induced polymerization: Addition and condensation re- actions, Molecules26, 7581 (2021)

    2021

  6. [14]

    D. Liu, W. Lei, Y. Li, Y. Ma, J. Hao, X. Chen, Y. Jin, D. Liu, S. Yu, Q. Cui, and G. Zou, High-pressure struc- tural transitions of sc2o3 by x-ray diffraction, raman spectra, and ab initio calculations, Inorg. Chem. 48, 7 8251–8256 (2009)

    2009

  7. [15]

    Sun, H.-T

    J. Sun, H.-T. Wang, J. He, and Y. Tian, Ab initio inves- tigations of optical properties of the high-pressure phases of zno, Phys. Rev. B71, 125132 (2005)

    2005

  8. [16]

    Napiórkowska, K

    E. Napiórkowska, K. Milcarz, and L. Szeleszczuk, Review ofapplicationsofdensityfunctionaltheory(dft)quantum mechanical calculations to study the high-pressure poly- morphs of organic crystalline materials, Int. J. Mol. Sci. 24, 14155 (2023)

    2023

  9. [17]

    Zeller, C

    F. Zeller, C. Hsieh, W. Dononelli, and T. Neudecker, Computational high-pressure chemistry: Ab initio sim- ulations of atoms, molecules, and extended materials in the gigapascal regime, Wiley Interdiscip. Rev.:Comput. Mol. Sci.14, e1708 (2024)

    2024

  10. [18]

    D. A. Rychkov, A short review of current computational concepts for high-pressure phase transition studies in molecular crystals, Crystals10, 81 (2020)

    2020

  11. [19]

    Behler, Perspective: Machine learning potentials for atomistic simulations, J

    J. Behler, Perspective: Machine learning potentials for atomistic simulations, J. Chem. Phys.145, 170901 (2016)

    2016

  12. [20]

    Thompson, L

    A. Thompson, L. Swiler, C. Trott, S. Foiles, and G. Tucker, Spectral neighbor analysis method for auto- mated generation of quantum-accurate interatomic po- tentials, J. Comput. Phys.285, 316–330 (2015)

    2015

  13. [21]

    Graser, S

    J. Graser, S. K. Kauwe, and T. D. Sparks, Machine learning and energy minimization approaches for crystal structure predictions: A review and new horizons, Chem. Mater. 30, 3601–3612 (2018)

    2018

  14. [22]

    Schmidt, M

    J. Schmidt, M. R. G. Marques, S. Botti, and M. A. L. Marques, Recent advances and applications of machine learning in solid-state materials science, npj Comput. Mater. 5, 83 (2019)

    2019

  15. [23]

    O. T. Unke, S. Chmiela, H. E. Sauceda, M. Gastegger, I. Poltavsky, K. T. Schütt, A. Tkatchenko, and K.-R. Müller, Machine learning force fields, Chem. Rev.121, 10142–10186 (2021)

    2021

  16. [24]

    K. T. Schütt, P. Kessel, M. Gastegger, K. A. Nicoli, A. Tkatchenko, and K.-R. Müller, SchNetPack: A Deep Learning Toolbox For Atomistic Systems, Journal of Chemical Theory and Computation 15, 448 (2019), https://doi.org/10.1021/acs.jctc.8b00908

    2019 doi

  17. [25]

    H. Wang, L. Zhang, J. Han, and W. E, Deepmd-kit: A deep learning package for many-body potential energy representation and molecular dynamics, Comput. Phys. Commun. 228, 178–184 (2018)

    2018

  18. [26]

    P.Ong, G.Hautier, W.Chen, W.D.Richards, S

    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

  19. [27]

    Kirklin, J

    S. Kirklin, J. E. Saal, B. Meredig, A. Thompson, J. W. Doak, M. Aykol, S. Rühl, and C. Wolverton, The open quantum materials database (OQMD): assessing the ac- curacy of DFT formation energies, npj Comput. Mater. 1, 15010 (2015)

    2015

  20. [28]

    Curtarolo, W

    S. Curtarolo, W. Setyawan, S. Wang, J. Xue, K. Yang, R. H. Taylor, L. J. Nelson, G. L. Hart, S. San- vito, M. Buongiorno-Nardelli, N. Mingo, and O. Levy, AFLOWLIB.ORG: A distributed materials properties repository from high-throughput ab initio calculations, Comput. Mater. Sci...

    2012

  21. [29]

    Schmidt, T

    J. Schmidt, T. F. Cerqueira, A. H. Romero, A. Loew, F. Jäger, H.-C. Wang, S. Botti, and M. A. Marques, Improving machine-learning models in materials science through large datasets, Mater. Today Phys.48, 101560 (2024)

    2024

  22. [30]

    Scheidgen, L

    M. Scheidgen, L. Himanen, A. N. Ladines, D. Sikter, M. Nakhaee, A. Fekete, T. Chang, A. Golparvar, J. A. Márquez, S. Brockhauser, S. Brückner, L. M. Ghir- inghelli, F. Dietrich, D. Lehmberg, T. Denell, A. Al- bino, H. Näsström, S. Shabih, F. Dobener, M. Küh- bach, R. Mozumder,...

    2023

  23. [31]

    Chen and S

    C. Chen and S. P. Ong, A universal graph deep learning interatomic potential for the periodic table, Nat. Com- put. Sci.2, 718–728 (2022)

    2022

  24. [32]

    Neumann, J

    M. Neumann, J. Gin, B. Rhodes, S. Bennett, Z. Li, H. Choubisa, A. Hussey, and J. Godwin, Orb: A fast, scalable neural network potential, arXiv , 2410.22570 (2024)

    2024 arXiv

  25. [33]

    Y. Park, J. Kim, S. Hwang, and S. Han, Scalable parallel algorithm for graph neural network interatomic poten- tialsinmoleculardynamicssimulations,J.Chem.Theory Comput. 20, 4857–4868 (2024)

    2024

  26. [34]

    Y.-L. Liao, B. M. Wood, A. Das, and T. Smidt, EquiformerV2: Improved equivariant transformer for scaling to higher-degree representations, inThe Twelfth International Conference on Learning Representations (2024)

    2024

  27. [35]

    B. Deng, P. Zhong, K. Jun, J. Riebesell, K. Han, C. J. Bartel, and G. Ceder, CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling, Nat. Mach. Intell.5, 1031–1041 (2023)

    2023

  28. [36]

    Choudhary and B

    K. Choudhary and B. DeCost, Atomistic line graph neu- ral network for improved materials property predictions, npj Comput. Mater.7, 185 (2021)

    2021

  29. [37]

    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, D. Zügner, T. Xie, J. Smith, L. Sun, Q. Wang, L. Kong, C. Liu, H. Hao, and Z. Lu, Mattersim: A deep learning atomistic model across elements, temperatures and pr...

    2024 arXiv

  30. [38]

    Barroso-Luque, M

    L. Barroso-Luque, M. Shuaibi, X. Fu, B. M. Wood, M. Dzamba, M. Gao, A. Rizvi, C. L. Zitnick, and Z. W. Ulissi, Open materials 2024 (OMat24) inorganic materi- als dataset and models, arXiv , 2410.12771 (2024)

    2024 arXiv

  31. [39]

    A. Peng, X. Liu, M.-Y. Guo, L. Zhang, and H. Wang, The openlam challenges, arXiv , 2501.16358 (2025)

    2025 arXiv

  32. [40]

    J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized gradient approximation made simple, Phys. Rev. Lett. 77, 3865 (1996)

    1996

  33. [41]

    Riebesell, R

    J. Riebesell, R. E. A. Goodall, P. Benner, Y. Chi- ang, B. Deng, A. A. Lee, A. Jain, and K. A. Persson, Matbench discovery – a framework to evaluate machine learning crystal stability predictions, arXiv , 2308.14920 (2023)

    2023 arXiv

  34. [42]

    Batatia, D

    I. Batatia, D. P. Kovács, G. N. C. Simm, C. Ortner, and G. Csányi, MACE: Higher order equivariant message passing neural networks for fast and accurate force fields, arXiv , 2206.07697 (2022)

    2022 arXiv

  35. [43]

    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). 8

    2022

  36. [44]

    Zhang, A

    D. Zhang, A. Peng, C. Cai, W. Li, Y. Zhou, J. Zeng, M. Guo, C. Zhang, B. Li, H. Jiang, T. Zhu, W. Jia, L. Zhang, and H. Wang, A graph neural network for the era of large atomistic models, arXiv , 2506.01686 (2025)

    2025

  37. [45]

    Bochkarev, Y

    A. Bochkarev, Y. Lysogorskiy, and R. Drautz, Graph atomic cluster expansion for semilocal interactions be- yond equivariant message passing, Phys. Rev. X 14, 021036 (2024)

    2024

  38. [46]

    Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Phys

    R. Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Phys. Rev. B 99, 014104 (2019)

    2019

  39. [47]

    Rhodes, S

    B. Rhodes, S. Vandenhaute, V. Šimkus, J. Gin, J. God- win, T. Duignan, and M. Neumann, Orb-v3: Atomistic simulation at scale, arXiv , 2504.06231 (2025)

    2025 arXiv

  40. [48]

    X. Fu, B. M. Wood, L. Barroso-Luque, D. S. Levine, M. Gao, M. Dzamba, and C. L. Zitnick, Learning smooth and expressive interatomic potentials for physical prop- erty prediction, arXiv , 2502.12147 (2025)

    2025 arXiv

  41. [49]

    Y.-L. Liao, T. Smidt, M. Shuaibi, and A. Das, Gener- alizing denoising to non-equilibrium structures improves equivariant force fields, arXiv preprint arXiv:2403.09549 (2024)

    2024 arXiv

  42. [50]

    Kresse and J

    G. Kresse and J. Furthmüller, Efficiency of ab-initio to- tal energy calculations for metals and semiconductors us- ing a plane-wave basis set, Comput. Mater. Sci.6, 15–50 (1996)

    1996

  43. [51]

    Kresse and J

    G. Kresse and J. Furthmüller, Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set, Phys. Rev. B54, 11169 (1996)

    1996

  44. [52]

    (facebookresearch), fairchem: Fair chemistry’s li- brary of machine learning methods for chemistry,https: //github.com/facebookresearch/fairchem (2025), ac- cessed: 2025-08-17

    F. (facebookresearch), fairchem: Fair chemistry’s li- brary of machine learning methods for chemistry,https: //github.com/facebookresearch/fairchem (2025), ac- cessed: 2025-08-17

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.