Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Foundation ML interatomic potentials need finite-temperature MD checks, not just force errors, to prove they reproduce real material structure and dynamics.

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

2026-07-12 02:30 UTC pith:KUDSBS3P

load-bearing objection Solid, reusable finite-T MD benchmark of 15 foundation MLIPs against DFT trajectories; force RMSE tracks RDF/VDOS on average, pressure does not, and the Pareto ranking is immediately useful. the 3 major comments →

arxiv 2607.03433 v1 pith:KUDSBS3P submitted 2026-07-03 cond-mat.mtrl-sci physics.chem-ph

Dyna-Mat: End-to-end benchmarking of foundation machine learning interatomic potentials in finite-temperature ensembles

Miko{\l}aj J. Gawkowski , Nongnuch Artrith , Silvia Bonfanti , Abhijeet Sadashiv Gangan , Hendrik H. Heenen , Joseph Kioseoglou , Ivor Lon\v{c}ari\'c , Hemanadhan Myneni
show 9 more authors
Janosh Riebesell Mariana Rossi Matthias Rupp Jonathan Schmidt Shubham Sharma Benjamin X. Shi Antoni Wadowski Lukas H\"ormann Venkat Kapil
This is my paper
classification cond-mat.mtrl-sci physics.chem-ph
keywords foundation MLIPsmolecular dynamicsfinite-temperature ensemblesbenchmark datasetradial distribution functionvibrational density of statespressurePareto frontier
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.

Foundation machine-learning interatomic potentials are sold as drop-in replacements for quantum calculations in materials design, yet almost all public tests stop at static energies, forces, or harmonic phonons. This paper supplies the missing ground truth: Dyna-Mat-v1.0, seventeen first-principles molecular-dynamics trajectories spanning metals, alloys, dichalcogenides, perovskites and molecular crystals. Fifteen foundation models are driven with identical settings and scored on radial distribution functions, pressure distributions and vibrational densities of states. On average, lower force error on the reference snapshots tracks lower structural and dynamical error, but individual systems still fail qualitatively and pressure remains systematically poor—traced to noisy stress labels in the large training sets. An accuracy-versus-cost Pareto front shows the newest multi-dataset models are near-optimal for production MD. The practical message is that end-to-end finite-temperature validation is now both necessary and feasible for choosing which foundation potential to trust.

Core claim

On average, foundation MLIPs whose single-point force RMSE is lower on the Dyna-Mat-v1.0 first-principles configurations also produce lower errors on radial distribution functions and vibrational densities of states extracted from matched NVT trajectories; pressure remains poorly described across most models, and the latest cross-trained models sit close to the accuracy–cost Pareto frontier for molecular dynamics.

What carries the argument

Dyna-Mat-v1.0—seventeen condensed-phase first-principles NVT trajectories used both as single-point energy/force test sets and as reference ensembles against which MLIP-driven trajectories are scored via normalised RDF, pressure-histogram and VDOS errors, then combined into an aggregate accuracy metric for Pareto analysis.

Load-bearing premise

The seventeen short, fixed-cell trajectories (one per system) are assumed to be a representative and statistically converged ground-truth ensemble for ranking foundation models and for treating force RMSE as a reliable diagnostic of finite-temperature behaviour.

What would settle it

Generate additional independent first-principles trajectories for the same systems (different initial conditions or longer runs) or for new chemistries; if model rankings by force RMSE or by RDF/VDOS error reverse, or if pressure errors collapse once stress labels are recomputed with tighter DFT settings, the claimed diagnostic value of the present benchmark fails.

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

If this is right

  • Force RMSE on finite-temperature configurations becomes a cheap, useful screen for structural and dynamical reliability before expensive MD campaigns.
  • Pressure (and therefore density) cannot be trusted without separate validation or better-converged stress labels in future training sets.
  • Cross-trained multi-dataset models are the practical choice when both accuracy and wall-clock cost matter for production molecular dynamics.
  • Community leaderboards that rely only on hull classification or harmonic thermal conductivity will miss failures that appear only under finite-temperature sampling.

Where Pith is reading between the lines

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

  • Future large-scale datasets will need tighter electronic and stress convergence criteria if NPT or free-energy work is to become trustworthy with foundation models.
  • The same end-to-end protocol can be extended to variable-cell trajectories and multi-trajectory statistics, turning Dyna-Mat into a living finite-temperature leaderboard.
  • Systems where low force error still yields qualitative structural failure point to missing long-range or anharmonic physics that local message-passing architectures may systematically under-represent.

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

3 major / 5 minor

Summary. The manuscript introduces Dyna-Mat-v1.0, a set of 17 first-principles NVT MD trajectories (pure metals, TMDs, alloys, perovskites, molecular crystals; ~20 ps; 293–1500 K) generated with carefully converged DFT settings. Fifteen foundation MLIPs spanning four training-set tiers are evaluated both by single-point energy/force RMSE on the reference configurations and by end-to-end MLIP-driven MD, comparing RDFs, pressure distributions and VDOS via explicitly defined normalised error metrics (Eqs. 4, 9, 10). On average, lower force RMSE correlates with lower RDF and VDOS errors (and with Matbench-discovery F1), while pressure remains poorly predicted for most models; an accuracy–cost Pareto frontier identifies several recent cross-trained models as near-optimal for MD.

Significance. The work fills a genuine gap: community MLIP benchmarks have largely been static or harmonic, whereas materials design increasingly relies on finite-temperature MD. Providing an independent DFT-MD reference set, transparent error definitions, correlation analyses against existing leaderboard metrics, and a practical Pareto ranking is a useful community contribution. Strengths include documented DFT convergence (Fig. 8, Table I), explicit metric definitions, system-resolved failure modes (e.g. UMA on Pt3Co), and candid limitations. If the averages and the pressure-decoupling observation hold under broader sampling, the dataset and protocol will become a reusable standard for assessing foundation MLIPs beyond energy/force tables.

major comments (3)
  1. Sec. III and Methods: the central claim that force RMSE on Dyna-Mat configurations is a useful diagnostic of finite-T reliability rests on 17 fixed-cell NVT trajectories (~20 ps, one per system). VDOS errors are already acknowledged to be statistically noisy at this length; RDF and pressure averages may likewise be sensitive to initial conditions and trajectory length. The paper should either (i) quantify trajectory-to-trajectory variability for a subset of systems (multiple independent seeds or longer runs) or (ii) more sharply qualify the ranking and correlation statements as provisional averages over this limited ensemble, rather than leaving the caveat only in the final limitations paragraph.
  2. Fig. 4, Sec. II.C.2 and the Pt3Co discussion: pressure MAE is dominated by a few large outliers (especially UMA on experimental-cell Pt3Co). The manuscript correctly notes that cell relaxation recovers reasonable pressures (Fig. S5), yet the main-text tier medians and the claim that “pressure remains poorly described across most models” still mix model error with out-of-equilibrium cell choice. A clearer separation—e.g. reporting pressure errors both on experimental cells and after model-consistent zero-pressure relaxation—would make the stress-label diagnosis more robust and avoid over-weighting a single protocol failure.
  3. Sec. II.D / Fig. 6: the strong anti-correlation of RDF/VDOS errors with Matbench F1 (and the weaker correlation with κSRME) is presented as evidence that energy-landscape improvements transfer to finite-T observables. Because Tier-4 models are excluded from the F1/κSRME panels (different DFT settings) and the remaining points largely follow the tier progression, the correlation may partly reflect a common generational trend rather than a mechanistic link. A leave-one-tier-out or residual analysis after regressing out force RMSE would strengthen (or qualify) this interpretation.
minor comments (5)
  1. Abstract and Sec. II.B: energy RMSE for molecular crystals is referenced to isolated-atom limits; EquiformerV2 is omitted because of empty neighbour graphs. A short note in the main text (not only SI) on how absolute energy offsets are handled would help readers interpret Fig. 2(a).
  2. Eqs. (4), (9), (10): the ideal-gas / histogram-overlap normalisations are reasonable but can saturate at 100 %. Mentioning the saturation behaviour when the reference RDF is nearly flat (or when pressure distributions have no overlap) would clarify why some systems appear equally “bad”.
  3. Fig. 7 and Methods: Pareto costs are measured with ASE, float32, no acceleration libraries. Stating the hardware (V100) and the deliberate choice to disable kernels is good; a one-line caveat that production timings with vendor kernels may reorder the frontier would be useful.
  4. Data availability: “will be made available upon publication / Matbench-discovery” is acceptable for review, but a permanent DOI or repository link should be supplied before final acceptance so that the benchmark can be reused.
  5. Minor typos / consistency: “H However” (Sec. II.C.1); UMA-M-P1 vs UMA-M-P2 labelling in Fig. 5 caption; occasional “mean” vs “median” wording when describing tier dashed lines.

Circularity Check

0 steps flagged

No significant circularity: external DFT-MD ground truth vs pre-trained MLIPs; self-citations are secondary only.

full rationale

Dyna-Mat-v1.0 is an independent first-principles NVT trajectory set generated for this work (VASP/FHI-aims, converged settings distinct from most training corpora). The 15 foundation MLIPs are evaluated zero-shot: single-point energy/force RMSEs on the reference snapshots, then full MLIP-driven MD with matched settings, followed by comparison of RDF, pressure histograms and VDOS via explicitly defined normalised error metrics (Eqs. 1–12). None of these quantities is fitted to the target observables or defined in terms of them. Tier labels are conventional groupings by training-set provenance; the aggregate ERPV and Pareto frontier are post-hoc summaries of the same independent errors plus measured wall-clock cost. The only self-citation of note is Matbench-discovery (Riebesell co-author) used solely for secondary Pearson correlations of F1/κSRME with the paper’s own force/RDF/VDOS errors; those correlations are not load-bearing for the primary claims (force RMSE tracks structural/dynamical accuracy on average; pressure remains poorly described; latest multi-head models lie near the accuracy–cost frontier). No uniqueness theorem, ansatz, or fitted parameter is smuggled in as a prediction. The evaluation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

The paper is an empirical benchmarking study. Its claims rest on standard DFT-as-reference assumptions, conventional MD ensembles, and a small set of hand-chosen error normalisations and model tiers; no new physical entities or free parameters are fitted to produce the central ranking.

free parameters (3)
  • RDF error normalisation (ideal-gas denominator)
    The percentage error in Eq. (4) is scaled by the L1 deviation of the reference RDF from 1; the precise binning (500 bins) and Rmax choice affect numerical values but not the ranking narrative.
  • Model-tier boundaries
    Four tiers are defined by training-set composition (MPTrj, +Alexandria, OMat24-style, multi-head). The grouping is useful but not uniquely determined by the data.
  • Aggregate ERPV / LRPV weights
    Equal average (Eq. 11) or Lehmer mean (Eq. 12) of RDF, pressure and VDOS errors; alternative weightings shift which models sit on the Pareto front.
axioms (3)
  • domain assumption PBE(+D3) DFT MD trajectories with the stated convergence settings constitute an adequate ground-truth ensemble for ranking foundation MLIPs.
    Invoked throughout Results and Methods; the entire benchmark is relative to these trajectories.
  • domain assumption Fixed-cell NVT dynamics with matched thermostats is a sufficient test of finite-temperature fidelity for the observables considered.
    All production runs and comparisons are performed in NVT; NPT or variable-cell behaviour is left for future work (Sec. III).
  • ad hoc to paper A single ~20 ps trajectory per system is statistically representative for RDF, pressure and VDOS comparisons.
    Acknowledged as a limitation; VDOS errors are noted to contain non-negligible statistical uncertainty.
invented entities (2)
  • Dyna-Mat-v1.0 dataset no independent evidence
    purpose: Provide condensed-phase first-principles MD trajectories as a reusable finite-temperature benchmark for foundation MLIPs.
    New collection of 17 trajectories; independent evidence will exist once the public repository is fully released and used by others.
  • Normalised RDF / pressure-histogram / VDOS error metrics (Eqs. 4, 9, 10) no independent evidence
    purpose: Convert distributional observables into single-valued percentage errors for ranking and Pareto analysis.
    Paper-specific normalisations; useful but not uniquely mandated by physics.

pith-pipeline@v1.1.0-grok45 · 37290 in / 2776 out tokens · 26824 ms · 2026-07-12T02:30:17.228689+00:00 · methodology

0 comments
read the original abstract

Foundation machine learning interatomic potentials (MLIPs) are increasingly being used as drop-in replacements for first-principles calculations, enabling simulations of materials at length and time scales that were previously inaccessible. However, due to lack of ground truth data, their accuracy on structural and dynamical observables in finite thermodynamic ensembles is yet to be established. Here, we introduce Dyna-Mat-v1.0, a benchmark dataset of condensed-phase first-principles molecular dynamics trajectories designed to test foundation MLIPs at realistic finite-temperature conditions. Using this dataset, we evaluate 15 foundation MLIPs across four model tiers by comparing both single-point energy and force errors on first-principles configurations and observables generated from MLIP-driven trajectories. We find that "on average" models with lower single-point force errors also yield lower errors for structural and dynamical observables. However, there are individual systems for which low force errors lead to qualitative failures in the predicted structure. Pressure remains poorly described across most models, pointing to limitations in the density functional theory stress labels available in current large-scale training datasets. Finally, we construct an accuracy-cost Pareto frontier to identify the best trade-offs for molecular dynamics with foundation MLIPs, finding that the latest generation of cross-trained models is close to Pareto-optimal according to the accuracy metrics considered here. Overall, Dyna-Mat-v1.0 shows that end-to-end finite-temperature validation is essential for quantifying the predictive behaviour of foundation MLIPs, and provides a simple, scalable route for assessing them beyond static and harmonic benchmarks relevant to materials design.

Figures

Figures reproduced from arXiv: 2607.03433 by Abhijeet Sadashiv Gangan, Antoni Wadowski, Benjamin X. Shi, Hemanadhan Myneni, Hendrik H. Heenen, Ivor Lon\v{c}ari\'c, Janosh Riebesell, Jonathan Schmidt, Joseph Kioseoglou, Lukas H\"ormann, Mariana Rossi, Matthias Rupp, Miko{\l}aj J. Gawkowski, Nongnuch Artrith, Shubham Sharma, Silvia Bonfanti, Venkat Kapil.

Figure 1
Figure 1. Figure 1: FIG. 1. Schematic of the dataset generation and benchmarking workflow. Starting from a consistent [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. (a) Mean energy RMSE and (b) mean force RMSE for each model averaged over all 17 sys [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Radial distribution function [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Unit-cell pressure distribution panel. (a): pressure MAE for the MLIPs grouped in four tiers. The [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Vibrational density of states panel. (a): VDOS error for the MLIPs grouped in four tiers. The [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. (a–c) RDF error, (d–f) pressure histogram error, and (g–i) VDOS error correlation plots. In each [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Pareto front of the combined error metric and the mean time per MD step using ASE Calculators [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Convergence tests on our dataset; a) Convergence of the total energy against the plane-wave cutoff [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Edge Cluster Expansion with Radial Rotary Attention for Interatomic Potentials

    stat.ML 2026-07 conditional novelty 6.0

    Edge Cluster Expansion plus Radial Rotary Complex Attention yield TECE-OAM-RRA-1.0, a Matbench Discovery SOTA SO(2) MLIP trained on OMat24, sAlex, and MPTrj.

Reference graph

Works this paper leans on

55 extracted references · 8 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    author author J. Behler ,\ title title Four Generations of High - Dimensional Neural Network Potentials , \ 10.1021/acs.chemrev.0c00868 journal journal Chemical Reviews \ ( year 2021 ),\ 10.1021/acs.chemrev.0c00868 NoStop

    doi:10.1021/acs.chemrev.0c00868 2021

  2. [2]

    author author V. L. \ Deringer , author A. P. \ Bartók , author N. Bernstein , author D. M. \ Wilkins , author M. Ceriotti , \ and\ author G. Csányi ,\ title title Gaussian Process Regression for Materials and Molecules , \ 10.1021/acs.chemrev.1c00022 journal journal Chemical Reviews \ volume 121 ,\ pages 10073--10141 ( year 2021 ) NoStop

    doi:10.1021/acs.chemrev.1c00022 2021

  3. [3]

    Jacobs , author D

    author author R. Jacobs , author D. Morgan , author S. Attarian , author J. Meng , author C. Shen , author Z. Wu , author C. Y. \ Xie , author J. H. \ Yang , author N. Artrith , author B. Blaiszik , author G. Ceder , author K. Choudhary , author G. Csanyi , author E. D. \ Cubuk , author B. Deng , author R. Drautz , author X. Fu , author J. Godwin , author...

    doi:10.1016/j.cossms.2025.101214 2025

  4. [4]

    author author E. C.-Y. \ Yuan , author Y. Liu , author J. Chen , author P. Zhong , author S. Raja , author T. Kreiman , author S. Vargas , author W. Xu , author M. Head-Gordon , author C. Yang , author S. M. \ Blau , author B. Cheng , author A. Krishnapriyan , \ and\ author T. Head-Gordon ,\ title en title Foundation models for atomistic simulation of che...

    doi:10.1038/s41570-025-00793-5 2026

  5. [5]

    Creed , author T

    author author I. Creed , author T. Rein , author I. Vitenburgs , author W. G. \ Stark , author V. Ellingsson , author A. Y. \ Ismail , author G. Liu , author Y. Lou , author B. A. A. \ Martin , author C. Bone , author M. A. H. \ Walker , author M. Taj , author S. Wang , author K. Wong , author R. Wu , author P. Kayastha , author B. Cheng , author A. Krish...

    Pith/arXiv arXiv doi:10.48550/arxiv.2606.07327 2026

  6. [6]

    Kaur , author F

    author author H. Kaur , author F. D. \ Pia , author I. Batatia , author X. R. \ Advincula , author B. X. \ Shi , author J. Lan , author G. Csányi , author A. Michaelides , \ and\ author V. Kapil ,\ https://arxiv.org/abs/2405.20217 title Data-efficient fine-tuning of foundational models for first-principles quality sublimation enthalpies , \ ( year 2024 ),...

    Pith/arXiv arXiv 2024

  7. [7]

    Fast and Fourier Features for Transfer Learning of Interatomic Potentials

    author author P. Novelli , author G. Meanti , author P. J. \ Buigues , author L. Rosasco , author M. Parrinello , author M. Pontil , \ and\ author L. Bonati ,\ 10.48550/arXiv.2505.05652 title Fast and Fourier Features for Transfer Learning of Interatomic Potentials , \ ( year 2025 ),\ note arXiv:2505.05652 [physics] NoStop

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.05652 2025

  8. [8]

    Radova , author W

    author author M. Radova , author W. G. \ Stark , author C. S. \ Allen , author R. J. \ Maurer , \ and\ author A. P. \ Bartók ,\ 10.48550/arXiv.2502.15582 title Fine-tuning foundation models of materials interatomic potentials with frozen transfer learning , \ ( year 2025 ),\ note arXiv:2502.15582 [cond-mat] NoStop

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.15582 2025

  9. [9]

    Hänseroth , author A

    author author J. Hänseroth , author A. Flötotto , author M. N. \ Qaisrani , \ and\ author C. Dreßler ,\ title title Fine- Tuning Unifies Foundational Machine - Learned Interatomic Potential Architectures at ab initio Accuracy , \ 10.1021/acs.jpclett.5c03801 journal journal The Journal of Physical Chemistry Letters \ volume 17 ,\ pages 3152--3162 ( year 20...

    work page doi:10.1021/acs.jpclett.5c03801 2026

  10. [10]

    author author T. L. \ Tompa , author E. Varga-Umbrich , author I. Batatia , author A. M. \ Elena , author N. Bernstein , \ and\ author G. Csányi ,\ 10.48550/arXiv.2606.12704 title Fine-tuning MLIP foundation models: strategies for accuracy and transferability , \ ( year 2026 ),\ note arXiv:2606.12704 [physics.chem-ph] NoStop

    Pith/arXiv arXiv doi:10.48550/arxiv.2606.12704 2026

  11. [11]

    Merchant , author S

    author author A. Merchant , author S. Batzner , author S. S. \ Schoenholz , author M. Aykol , author G. Cheon , \ and\ author E. D. \ Cubuk ,\ title title Scaling deep learning for materials discovery , \ 10.1038/s41586-023-06735-9 journal journal Nature \ volume 624 ,\ pages 80--85 ( year 2023 ) NoStop

    doi:10.1038/s41586-023-06735-9 2023

  12. [12]

    Probing the Limit of Heat Transfer in Inorganic Crystals with Deep Learning

    author author J. Li , author Z. Chen , author Q. Wang , author H. Yang , author Z. Lu , author G. Li , author S. Chen , author Y. Zhu , author X. Liu , author J. Tan , author M. Tang , author Y. Zhou , author C. Zeni , author A. Fowler , author D. Zügner , author R. Pinsler , author M. Horton , author T. Xie , author T.-Y. \ Liu , author H. Liu , author T...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.11568 2025

  13. [13]

    author author T. J. \ Inizan , author S. Yang , author A. Kaplan , author Y.-h. \ Lin , author J. Yin , author S. Mirzaei , author M. Abdelgaid , author A. H. \ Alawadhi , author K. Cho , author Z. Zheng , author E. D. \ Cubuk , author C. Borgs , author J. T. \ Chayes , author K. A. \ Persson , \ and\ author O. M. \ Yaghi ,\ 10.48550/arXiv.2504.14110 titl...

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.14110 2025

  14. [14]

    Riebesell , author R

    author author J. Riebesell , author R. E. A. \ Goodall , author P. Benner , author Y. Chiang , author B. Deng , author G. Ceder , author M. Asta , author A. A. \ Lee , author A. Jain , \ and\ author K. A. \ Persson ,\ title en title A framework to evaluate machine learning crystal stability predictions , \ 10.1038/s42256-025-01055-1 journal journal Nature...

    doi:10.1038/s42256-025-01055-1 2025

  15. [15]

    Chiang , author T

    author author Y. Chiang , author T. Kreiman , author E. Weaver , author I. Amin , author M. Kuner , author C. Zhang , author A. Kaplan , author D. Chrzan , author S. M. \ Blau , author A. S. \ Krishnapriyan , \ and\ author M. Asta ,\ title en title MLIP Arena : Advancing Fairness and Transparency in Machine Learning Interatomic Potentials through an Open ...

    2025

  16. [16]

    Choudhary , author D

    author author K. Choudhary , author D. Wines , author K. Li , author K. F. \ Garrity , author V. Gupta , author A. H. \ Romero , author J. T. \ Krogel , author K. Saritas , author A. Fuhr , author P. Ganesh , author P. R. C. \ Kent , author K. Yan , author Y. Lin , author S. Ji , author B. Blaiszik , author P. Reiser , author P. Friederich , author A. Agr...

    work page doi:10.1038/s41524-024-01259-w 2024

  17. [17]

    Peng , author C

    author author A. Peng , author C. Cai , author M. Guo , author D. Zhang , author C. Zhang , author W. Jiang , author Y. Wang , author A. Loew , author C. Wu , author W. E , author L. Zhang , \ and\ author H. Wang ,\ title en title LAMBench : a benchmark for large atomistic models , \ 10.1038/s41524-025-01929-3 journal journal npj Computational Materials \...

    doi:10.1038/s41524-025-01929-3 2026

  18. [18]

    P \'o ta , author P

    author author B. P \'o ta , author P. Ahlawat , author G. Cs \'a nyi , \ and\ author M. Simoncelli ,\ title title Thermal conductivity predictions with foundation atomistic models , \ @noop journal journal arXiv preprint arXiv:2408.00755 \ ( year 2024 ) NoStop

    Pith/arXiv arXiv 2024

  19. [19]

    Moon , author U

    author author J. Moon , author U. Jeon , author S. Choung , \ and\ author J. W. \ Han ,\ title title CatBench framework for benchmarking machine learning interatomic potentials in adsorption energy predictions for heterogeneous catalysis , \ 10.1016/j.xcrp.2025.102968 journal journal Cell Reports Physical Science \ volume 6 ,\ pages 102968 ( year 2025 ) NoStop

    work page doi:10.1016/j.xcrp.2025.102968 2025

  20. [20]

    Loew , author J

    author author A. Loew , author J. Schmidt , author S. Botti , \ and\ author M. A. \ Marques ,\ title title Universal machine learning potentials under pressure , \ @noop journal journal Journal of Physics: Materials \ volume 9 ,\ pages 015010 ( year 2026 ) NoStop

    2026

  21. [21]

    Kraß , author J

    author author H. Kraß , author J. Huang , \ and\ author S. M. \ Moosavi ,\ title en title MOFSimBench : evaluating universal machine learning interatomic potentials in metal-organic framework molecular modeling , \ 10.1038/s41524-025-01872-3 journal journal npj Computational Materials \ volume 12 ,\ pages 4 ( year 2025 ) NoStop

    work page doi:10.1038/s41524-025-01872-3 2025

  22. [22]

    Focassio , author L

    author author B. Focassio , author L. P. \ M. Freitas , \ and\ author G. R. \ Schleder ,\ title title Performance Assessment of Universal Machine Learning Interatomic Potentials : Challenges and Directions for Materials ’ Surfaces , \ 10.1021/acsami.4c03815 journal journal ACS Applied Materials & Interfaces \ volume 17 ,\ pages 13111--13121 ( year 2025 ) NoStop

    doi:10.1021/acsami.4c03815 2025

  23. [23]

    Deng , author Y

    author author B. Deng , author Y. Choi , author P. Zhong , author J. Riebesell , author S. Anand , author Z. Li , author K. Jun , author K. A. \ Persson , \ and\ author G. Ceder ,\ title en title Systematic softening in universal machine learning interatomic potentials , \ 10.1038/s41524-024-01500-6 journal journal npj Computational Materials \ volume 11 ...

    doi:10.1038/s41524-024-01500-6 2025

  24. [24]

    Wines \ and\ author K

    author author D. Wines \ and\ author K. Choudhary ,\ title title CHIPS - FF : Evaluating Universal Machine Learning Force Fields for Material Properties , \ 10.1021/acsmaterialslett.5c00093 journal journal ACS Materials Letters \ volume 7 ,\ pages 2105--2114 ( year 2025 ) NoStop

    work page doi:10.1021/acsmaterialslett.5c00093 2025

  25. [25]

    Deng , author P

    author author B. Deng , author P. Zhong , author K. Jun , author J. Riebesell , author K. Han , author C. J. \ Bartel , \ and\ author G. Ceder ,\ title title Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling , \ 10.1038/s42256-023-00716-3 journal journal Nature Machine Intelligence \ volume 5 ,\ pages 1031--...

    doi:10.1038/s42256-023-00716-3 2023

  26. [26]

    Schmidt , author T

    author author J. Schmidt , author T. F. \ Cerqueira , author A. H. \ Romero , author A. Loew , author F. Jäger , author H.-C. \ Wang , author S. Botti , \ and\ author M. A. \ Marques ,\ title title Improving machine-learning models in materials science through large datasets , \ https://doi.org/10.1016/j.mtphys.2024.101560 journal journal Materials Today ...

    doi:10.1016/j.mtphys.2024.101560 2024

  27. [27]

    Yang , author C

    author author H. Yang , author C. Hu , author Y. Zhou , author X. Liu , author Y. Shi , author J. Li , author G. Li , author Z. Chen , author S. Chen , author C. Zeni , author M. Horton , author R. Pinsler , author A. Fowler , author D. Zügner , author T. Xie , author J. Smith , author L. Sun , author Q. Wang , author L. Kong , author C. Liu , author H. H...

    Pith/arXiv arXiv 2024

  28. [28]

    Barroso-Luque , author M

    author author L. Barroso-Luque , author M. Shuaibi , author X. Fu , author B. M. \ Wood , author M. Dzamba , author M. Gao , author A. Rizvi , author C. L. \ Zitnick , \ and\ author Z. W. \ Ulissi ,\ https://arxiv.org/abs/2410.12771 title Open materials 2024 (omat24) inorganic materials dataset and models , \ ( year 2024 ),\ http://arxiv.org/abs/2410.1277...

    Pith/arXiv arXiv 2024

  29. [29]

    Malosso , author F

    author author C. Malosso , author F. Bigi , author P. Pegolo , author J. W. \ Abbott , author P. Loche , author M. Rossi , author M. Ceriotti , \ and\ author A. Mazitov ,\ 10.48550/arXiv.2603.02089 title High-quality, high-information datasets for universal atomistic machine learning , \ ( year 2026 ),\ note arXiv:2603.02089 [cond-mat.mtrl-sci] NoStop

    doi:10.48550/arxiv.2603.02089 2026

  30. [30]

    author author A. D. \ Kaplan , author R. Liu , author J. Qi , author T. W. \ Ko , author B. Deng , author J. Riebesell , author G. Ceder , author K. A. \ Persson , \ and\ author S. P. \ Ong ,\ 10.48550/arXiv.2503.04070 title A Foundational Potential Energy Surface Dataset for Materials , \ ( year 2025 ),\ note arXiv:2503.04070 [cond-mat.mtrl-sci] NoStop

    doi:10.48550/arxiv.2503.04070 2025

  31. [31]

    author author N. L. \ Fragapane \ and\ author V. L. \ Deringer ,\ title title Li– P – S Electrolyte Materials as a Benchmark for Machine - Learned Interatomic Potentials , \ 10.1021/acs.jctc.5c02006 journal journal Journal of Chemical Theory and Computation \ ( year 2026 ),\ 10.1021/acs.jctc.5c02006 NoStop

    work page doi:10.1021/acs.jctc.5c02006 2026

  32. [32]

    Bigi , author P

    author author F. Bigi , author P. Pegolo , author A. Mazitov , \ and\ author M. Ceriotti ,\ https://arxiv.org/abs/2601.16195 title Pushing the limits of unconstrained machine-learned interatomic potentials , \ ( year 2026 ),\ http://arxiv.org/abs/2601.16195 arXiv:2601.16195 [physics.chem-ph] NoStop

    Pith/arXiv arXiv 2026

  33. [33]

    author author J. Hafner ,\ title en title Ab-initio simulations of materials using VASP : Density -functional theory and beyond , \ 10.1002/jcc.21057 journal journal Journal of Computational Chemistry \ volume 29 ,\ pages 2044--2078 ( year 2008 ) NoStop

    doi:10.1002/jcc.21057 2044

  34. [34]

    Blum , author R

    author author V. Blum , author R. Gehrke , author F. Hanke , author P. Havu , author V. Havu , author X. Ren , author K. Reuter , \ and\ author M. Scheffler ,\ title title Ab initio molecular simulations with numeric atom-centered orbitals , \ 10.1016/j.cpc.2009.06.022 journal journal Computer Physics Communications \ volume 180 ,\ pages 2175--2196 ( year...

    doi:10.1016/j.cpc.2009.06.022 2009

  35. [35]

    Litman , author V

    author author Y. Litman , author V. Kapil , author Y. M. Y. \ Feldman , author D. Tisi , author T. Begušić , author K. Fidanyan , author G. Fraux , author J. Higer , author M. Kellner , author T. E. \ Li , author E. S. \ Pós , author E. Stocco , author G. Trenins , author B. Hirshberg , author M. Rossi , \ and\ author M. Ceriotti ,\ title title i- PI 3.0:...

    doi:10.1063/5.0215869 2024

  36. [36]

    author author S. P. \ Ong , author W. D. \ Richards , author A. Jain , author G. Hautier , author M. Kocher , author S. Cholia , author D. Gunter , author V. L. \ Chevrier , author K. A. \ Persson , \ and\ author G. Ceder ,\ title title Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis , \ 10.1016/j.commatsc...

    doi:10.1016/j.commatsci.2012.10.028 2012

  37. [37]

    author author A. M. \ Ganose , author H. Sahasrabuddhe , author M. Asta , author K. Beck , author T. Biswas , author A. Bonkowski , author J. Bustamante , author X. Chen , author Y. Chiang , author D. C. \ Chrzan , author J. Clary , author O. A. \ Cohen , author C. Ertural , author M. C. \ Gallant , author J. George , author S. Gerits , author R. E. A. \ ...

    doi:10.1039/d5dd00019j 1944

  38. [38]

    Hjorth Larsen , author J

    author author A. Hjorth Larsen , author J. J rgen Mortensen , author J. Blomqvist , author I. E. \ Castelli , author R. Christensen , author M. Du ak , author J. Friis , author M. N. \ Groves , author B. Hammer , author C. Hargus , author E. D. \ Hermes , author P. C. \ Jennings , author P. Bjerre Jensen , author J. Kermode , author J. R. \ Kitchin , auth...

    2017

  39. [39]

    Batatia , author P

    author author I. Batatia , author P. Benner , author Y. Chiang , author A. M. \ Elena , author D. P. \ Kovács , author J. Riebesell , author X. R. \ Advincula , author M. Asta , author M. Avaylon , author W. J. \ Baldwin , author F. Berger , author N. Bernstein , author A. Bhowmik , author F. Bigi , author S. M. \ Blau , author V. Cărare , author M. Cerio...

    doi:10.1063/5.0297006 2025

  40. [40]

    Bochkarev , author Y

    author author A. Bochkarev , author Y. Lysogorskiy , \ and\ author R. Drautz ,\ title title Graph atomic cluster expansion for semilocal interactions beyond equivariant message passing , \ 10.1103/PhysRevX.14.021036 journal journal Phys. Rev. X \ volume 14 ,\ pages 021036 ( year 2024 ) NoStop

    doi:10.1103/physrevx.14.021036 2024

  41. [41]

    Neumann , author J

    author author M. Neumann , author J. Gin , author B. Rhodes , author S. Bennett , author Z. Li , author H. Choubisa , author A. Hussey , \ and\ author J. Godwin ,\ https://arxiv.org/abs/2410.22570 title Orb: A fast, scalable neural network potential , \ ( year 2024 ),\ http://arxiv.org/abs/2410.22570 arXiv:2410.22570 [cond-mat.mtrl-sci] NoStop

    Pith/arXiv arXiv 2024

  42. [42]

    Lysogorskiy , author A

    author author Y. Lysogorskiy , author A. Bochkarev , \ and\ author R. Drautz ,\ https://arxiv.org/abs/2508.17936 title Graph atomic cluster expansion for foundational machine learning interatomic potentials , \ ( year 2026 ),\ http://arxiv.org/abs/2508.17936 arXiv:2508.17936 [cond-mat.mtrl-sci] NoStop

    arXiv 2026

  43. [43]

    Rhodes , author S

    author author B. Rhodes , author S. Vandenhaute , author V. Šimkus , author J. Gin , author J. Godwin , author T. Duignan , \ and\ author M. Neumann ,\ https://arxiv.org/abs/2504.06231 title Orb-v3: atomistic simulation at scale , \ ( year 2025 ),\ http://arxiv.org/abs/2504.06231 arXiv:2504.06231 [cond-mat.mtrl-sci] NoStop

    Pith/arXiv arXiv 2025

  44. [44]

    Fu , author B

    author author X. Fu , author B. M. \ Wood , author L. Barroso-Luque , author D. S. \ Levine , author M. Gao , author M. Dzamba , \ and\ author C. L. \ Zitnick ,\ https://arxiv.org/abs/2502.12147 title Learning smooth and expressive interatomic potentials for physical property prediction , \ ( year 2025 ),\ http://arxiv.org/abs/2502.12147 arXiv:2502.12147 ...

    Pith/arXiv arXiv 2025

  45. [45]

    author author C. W. \ Tan , author M. L. \ Descoteaux , author M. Kotak , author G. de Miranda Nascimento , author S. R. \ Kavanagh , author L. Zichi , author M. Wang , author A. Saluja , author Y. R. \ Hu , author T. Smidt , author A. Johansson , author W. C. \ Witt , author B. Kozinsky , \ and\ author A. Musaelian ,\ https://arxiv.org/abs/2504.16068 tit...

    Pith/arXiv arXiv 2025

  46. [46]

    \ Liao , author B

    author author Y.-L. \ Liao , author B. Wood , author A. Das , \ and\ author T. Smidt ,\ 10.48550/arXiv.2306.12059 title EquiformerV2 : Improved Equivariant Transformer for Scaling to Higher - Degree Representations , \ ( year 2024 ),\ note arXiv:2306.12059 [cs] NoStop

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.12059 2024

  47. [47]

    Batatia , author C

    author author I. Batatia , author C. Lin , author J. Hart , author E. Kasoar , author A. M. \ Elena , author S. W. \ Norwood , author T. Wolf , \ and\ author G. Csányi ,\ https://arxiv.org/abs/2510.25380 title Cross learning between electronic structure theories for unifying molecular, surface, and inorganic crystal foundation force fields , \ ( year 2025...

    arXiv 2025

  48. [48]

    author author B. M. \ Wood , author M. Dzamba , author X. Fu , author M. Gao , author M. Shuaibi , author L. Barroso-Luque , author K. Abdelmaqsoud , author V. Gharakhanyan , author J. R. \ Kitchin , author D. S. \ Levine , author K. Michel , author A. Sriram , author T. Cohen , author A. Das , author A. Rizvi , author S. J. \ Sahoo , author Z. W. \ Uliss...

    arXiv 2026

  49. [49]

    Tuckerman ,\ @noop en title Statistical Mechanics : Theory and Molecular Simulation \ ( publisher OUP Oxford ,\ year 2010 )\ note google-Books-ID: Lo3Jqc0pgrcC NoStop

    author author M. Tuckerman ,\ @noop en title Statistical Mechanics : Theory and Molecular Simulation \ ( publisher OUP Oxford ,\ year 2010 )\ note google-Books-ID: Lo3Jqc0pgrcC NoStop

    2010

  50. [50]

    Kubo ,\ title title Statistical- Mechanical Theory of Irreversible Processes

    author author R. Kubo ,\ title title Statistical- Mechanical Theory of Irreversible Processes . I . General Theory and Simple Applications to Magnetic and Conduction Problems , \ 10.1143/JPSJ.12.570 journal journal Journal of the Physical Society of Japan \ volume 12 ,\ pages 570--586 ( year 1957 ) NoStop

    doi:10.1143/jpsj.12.570 1957

  51. [51]

    author author M. S. \ Green ,\ title title Markoff Random Processes and the Statistical Mechanics of Time ‐ Dependent Phenomena . II . Irreversible Processes in Fluids , \ 10.1063/1.1740082 journal journal The Journal of Chemical Physics \ volume 22 ,\ pages 398--413 ( year 1954 ) NoStop

    doi:10.1063/1.1740082 1954

  52. [52]

    author author J. P. \ Perdew , author K. Burke , \ and\ author M. Ernzerhof ,\ title title Generalized Gradient Approximation Made Simple , \ 10.1103/PhysRevLett.77.3865 journal journal Physical Review Letters \ volume 77 ,\ pages 3865--3868 ( year 1996 ) NoStop

    doi:10.1103/physrevlett.77.3865 1996

  53. [53]

    Grimme , author A

    author author S. Grimme , author A. Hansen , author J. G. \ Brandenburg , \ and\ author C. Bannwarth ,\ title title Dispersion-corrected mean-field electronic structure methods , \ 10.1021/acs.chemrev.5b00533 journal journal Chemical Reviews \ volume 116 ,\ pages 5105--5154 ( year 2016 ) NoStop

    doi:10.1021/acs.chemrev.5b00533 2016

  54. [54]

    o rkman , author P. Blaha , author S. Bl \

    author author K. Lejaeghere , author G. Bihlmayer , author T. Bj \"o rkman , author P. Blaha , author S. Bl \"u gel , author V. Blum , author D. Caliste , author I. E. \ Castelli , author S. J. \ Clark , author A. Dal Corso , et al. ,\ title title Reproducibility in density functional theory calculations of solids , \ @noop journal journal Science \ volum...

    2016

  55. [55]

    author author R. T. \ McGibbon , author K. A. \ Beauchamp , author M. P. \ Harrigan , author C. Klein , author J. M. \ Swails , author C. X. \ Hern \'a ndez , author C. R. \ Schwantes , author L.-P. \ Wang , author T. J. \ Lane , \ and\ author V. S. \ Pande ,\ title title Mdtraj: A modern open library for the analysis of molecular dynamics trajectories , ...

    doi:10.1016/j.bpj.2015.08.015 2015