Pith. sign in

REVIEW 3 major objections 5 minor 92 references

A Study on the Fine-Tuning Performance of Universal Machine-Learned Interatomic Potentials (U-MLIPs)

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Fine-tuned MACE foundation models can beat training from scratch.

desk verdict Useful multi-system benchmark, but the fine-tuning-vs-scratch claim is unverified because the scratch baselines are never specified or shown to be converged. read the letter →

arxiv 2506.07401 v1 pith:452CSEZR submitted 2025-06-09 physics.comp-ph

Xiaoqing Liu , Kehan Zeng , Yangshuai Wang , Teng Zhao This is my paper
classification physics.comp-ph
keywords machine-learnedinteratomicpotentialsfine-tuningMACEfoundationmodelsdatasetselectionactivelearningmoleculardynamicsatomisticsimulation
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

The paper's central claim is that universal machine-learned interatomic potentials work best when treated as starting points, not finished models: fine-tuning MACE-MP-0 and MACE-MP-0b on task-specific datasets improves their accuracy for atomistic simulations, and in several cases lets the fine-tuned model match or slightly exceed a model trained from scratch. The paper also claims that fine-tuned models converge faster because the foundation model provides a strong initial guess, and that dataset selection is the decisive factor: a curated subset of physically relevant configurations outperforms a large unfiltered database. If true, this means researchers can avoid expensive from-scratch training of interatomic potentials and instead adapt a general-purpose potential to a target material with less data and compute. The evidence spans elemental metals, silicon, BCC iron and tungsten dislocations, a five-element high-entropy alloy, and molten LiCl, and the same recipe transfers across these very different systems.

What carries the argument

The load-bearing machinery is multi-head fine-tuning of a shared MACE backbone. MACE is an equivariant message-passing graph tensor network built on the atomic cluster expansion, and the two foundation models are pre-trained versions of it; fine-tuning attaches task-specific output heads to the shared backbone and updates the backbone parameters on energy, force, and stress labels under a 1:10:100 MSE loss. The second piece of machinery is dataset construction, handled through three strategies: using open-source configurations directly (FT-1), manually filtering a database to defect-rich and low-temperature states (FT-2), and applying uncertainty-based filtering to a self-constructed candidate set (FT-3). The interaction of these two components—a strong pretrained initial predictor and a well-chosen task-specific dataset—is what produces the reported accuracy gains and faster convergence.

What would settle it

Re-run the elemental-metal and dislocation benchmarks with fresh DFT reference values computed using a different exchange-correlation functional and pseudopotentials, and check whether fine-tuned MACE models still match or beat scratch-trained ACE models for vacancy formation energies and Peierls barriers; if the ranking flips, the central claim depends on the choice of reference data.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the known weaknesses of the MACE foundation models—large errors in elastic constants, defect formation energies, and Peierls barriers—are largely fixable by fine-tuning. The fine-tuned models recover DFT-level accuracy for most tested properties, match or slightly beat MACE and ACE models trained from scratch in several benchmarks, and do so with faster convergence. A second, equally central finding is that the fine-tuning dataset matters more than its size: for silicon, the full 531,710-point database produces poor convergence, while a defect-focused, low-temperature subset gives final energy errors near 7 meV per atom and force errors near 60 meV/Å. The same fine-tuning recipe works for metals, a semiconductor, dislocation cores in BCC metals, a high-entropy alloy, and an ionic melt, and it does not slow down molecular dynamics simulations.

Load-bearing premise

Every accuracy comparison in the paper is measured against DFT or QM/MM reference values taken from earlier published datasets, so the conclusions inherit the accuracy and consistency of those references; if any reference is biased for a tested property, the ranking between fine-tuned and scratch-trained models could change.

Editorial extensions

If this is right

  • Fine-tuned foundation models can serve as accurate force fields for defect and dislocation simulations in BCC metals, reproducing Peierls barrier shapes and heights that untuned foundation models miss.
  • Users can replace expensive from-scratch training with fine-tuning when a suitable task-specific dataset is available, because fine-tuning converges faster and needs less data.
  • Dataset selection should be treated as part of the modeling workflow: filtering a large database to physically relevant configurations can outperform using the full database.
  • Fine-tuning improves accuracy without changing the computational cost of molecular dynamics, so the scaling of large-scale simulations is preserved; further gains would require model distillation.
  • The fine-tuning recipe transfers across metals, semiconductors, high-entropy alloys, and ionic melts, suggesting it is not specific to one chemistry or bonding type.

Reading between the lines

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

  • The paper's results imply that for a new target material, the highest-leverage investment is curating the fine-tuning dataset rather than choosing between foundation models or architectures: the same two backbones succeed across all tested systems once the data are right.
  • The uncertainty-filtering result (FT-3) suggests a testable workflow the paper does not close: run an active-learning loop that generates candidate configurations, ranks them by model uncertainty, fine-tunes, and measures accuracy gain; the paper only tests one round of filtering, not the loop.
  • Because every benchmark is measured against DFT or QM/MM references from earlier literature, the general claim that fine-tuning beats scratch training is conditional on those references; users targeting different properties should revalidate the ranking against their own reference calculations.
  • Multi-head fine-tuning, mentioned in the paper but not systematically benchmarked, could support multi-task potentials where one backbone serves several elements or defect types; the data here suggest the backbone has enough shared representation to make that feasible.
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

3 major / 5 minor

Summary. The paper evaluates fine-tuning of two MACE-based universal machine-learned interatomic potentials, MACE-MP-0 and MACE-MP-0b, across elemental metals, silicon, BCC dislocations, a high-entropy alloy, and molten LiCl. For each system, the authors compare the base foundation models, their fine-tuned versions, and MACE/ACE models trained from scratch against DFT or QM/MM reference values from the literature. The main claims are that fine-tuning improves task-specific accuracy relative to the base models, can sometimes match or slightly surpass training from scratch, converges faster because of the foundation-model initialization, and is strongly affected by dataset selection. The paper also discusses dataset filtering strategies and future directions, including model distillation and integration with the RBMD package.

Significance. If the central claims hold, the paper provides practically useful guidance for users of universal MLIPs: targeted fine-tuning on curated datasets may replace expensive scratch training while retaining or improving accuracy. The study has concrete strengths: it benchmarks against external DFT and QM/MM references rather than circular self-consistency checks; it covers several chemically and structurally diverse systems; it distinguishes three dataset-selection strategies; and it reports numerical tables and MD results that are straightforward to compare against future work. The limitations are equally concrete: scratch-trained baselines are not specified in sufficient detail, the faster-convergence claim is asserted without a supporting learning-curve comparison, and no repeated-seed statistics or error bars are provided for any reported value. These gaps directly affect the headline comparison between fine-tuning and training from scratch, so the manuscript currently supports a conditional rather than definitive conclusion.

major comments (3)
  1. [Section 5.1; Tables 2-6] The scratch-trained baselines used for the central comparison are not specified. Section 5.1 gives a detailed fine-tuning protocol for MACE-MP-0 and MACE-MP-0b, but nowhere does the paper state the training data, data splits, loss weights, epochs, convergence criteria, or hyperparameters used for the 'MACE-scratch' and 'ACE-scratch' models that appear in Tables 2-6 and Figures 3-4. Without this information, the Abstract's claim that fine-tuning 'in some cases, outperforms models trained from scratch' cannot be distinguished from a comparison against undertrained or differently trained baselines. This is load-bearing because the fine-tuning-versus-scratch comparison is the main practical rationale of the paper. The authors should either supply the missing training details or clearly label the scratch models as taken from prior work with citations.
  2. [Section 3.3.2; Figure 1] The claim that fine-tuned models benefit from faster convergence is asserted but not demonstrated. The only convergence plot, Figure 1, compares two fine-tuning dataset strategies (full database vs subset), not fine-tuning versus training from scratch. The text in Section 3.3.2 states that fine-tuned foundation models 'achieve faster convergence' and cites the strong initial guess from the foundation model, but no learning curve, epoch count, or computational-cost comparison for scratch training is shown. To support this claim, the paper should provide at least one direct comparison of error versus training epoch (or wall-clock time) for a fine-tuned model and a matched scratch-trained model on the same data and hardware.
  3. [Tables 2-6; Figures 3-6] No repeated-seed statistics, error bars, or uncertainty estimates are reported for any numerical result. Many of the differences that support the conclusions are small: for example, Table 2 shows Cu lattice parameters varying by less than 0.5% across methods, and Table 5 shows fine-tuned models at 13.8 meV/atom versus MACE-scratch at 16.4 meV/atom. Single deterministic runs cannot establish whether these differences are significant, especially given the stochastic training procedure (Adam, random data shuffling, EMA). The authors should report results over multiple seeds or provide ensemble/uncertainty estimates for the central comparisons, or alternatively soften claims such as 'consistently outperform' in Section 3.4.
minor comments (5)
  1. [Section 3.4, Table 5] The table header reads 'MAE Energy' and 'MAE Force', but the surrounding text refers to 'Testing RMSE error'; the metric name should be made consistent and the redundant phrase 'RMSE error' should be replaced with 'RMSE'.
  2. [Section 3.2, Figure 1] The y-axis labels in Figure 1 say 'MAE Energy' and 'MAE Force', while the text reports 'final RMSE values of 7 meV per atom for energy and 60 meV/Å for forces'; please use the same metric name in the figure and text.
  3. [Section 5.1] The term 'MPtraj samples' is used without definition; please clarify what an MPtraj sample is and how the fixed value of 400 was chosen.
  4. [Figures 3-4] The abbreviations 'FT-0b' and 'FT-0' are used in the figure legends but are not defined in the captions; they should be defined either in the captions or in the main text before first use.
  5. [Section 3.3, Figure 3] The dual-scale y-axes in Figure 3, with one scale for MACE-MP-0b and another for all other models, make direct visual comparison difficult; consider overlaying all curves on a single axis or adding an inset at the smaller scale.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's fine-tuning claims are measured against external DFT, QM/MM, and experimental references, and no derivation reduces to its own inputs.

full rationale

This paper is an empirical benchmark study rather than a derivation. Its central claims, that fine-tuning universal MLIPs improves accuracy, can match or exceed training from scratch, and benefits from data selection, are supported by comparing predicted material properties against external references: DFT values from Zou et al. [64], Bartok et al. [66], and Byggmastar et al. [71], QM/MM reference barriers, and experimental data for LiCl from Sivaraman et al. [77]. No fitted parameter is relabeled as a prediction: the fine-tuned models are trained on energies, forces, and stresses and then evaluated on separately reported physical properties. The foundation models and datasets are external to this paper, and the authors' own RBMD package is presented as an implementation tool rather than as evidence for the accuracy rankings. Some evaluation settings reuse the same source dataset used for fine-tuning, which could raise benchmark-leakage or reproducibility concerns, and the scratch-training protocol is underspecified in Section 5.1; these are verification gaps, not circular reductions. The paper also explicitly acknowledges that a full active learning strategy with uncertainty estimation is beyond its scope. There is no self-definitional step, no fitted input called a prediction, no load-bearing self-citation chain, no imported uniqueness theorem, no ansatz smuggled in via citation, and no renaming of a known result. The derivation chain, such as it is, is self-contained as an empirical comparison.

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

The paper introduces no new physical entities and fits no scientific constants. Its conclusions instead rest on the validity of DFT/QM-MM reference data, the representativeness of the two MACE models, the suitability of public training datasets, and the reliability of an auxiliary uncertainty filter. These are domain assumptions, listed above.

assumptions (4)
  • domain assumption DFT references from Zou et al. [64], Bartók et al. [66], and related works are accurate ground truth for energies, forces, and mechanical properties.
    All reported errors are deviations from these references; biased references would alter the model ranking. Invoked throughout Section 3 tables.
  • domain assumption The MACE-MP-0 and MACE-MP-0b foundation models, in the medium L=1 variant, are representative of universal MLIPs and provide a fair basis for fine-tuning comparisons.
    The paper generalizes its conclusions to U-MLIPs from these two models, so their representativeness is load-bearing. Invoked in Sections 1-2.
  • domain assumption The public datasets used for fine-tuning match the target domain of each application (e.g., the Fe active-learning dataset captures crack-tip deformation, the W dataset contains dislocation configurations).
    The FT-1 strategy assumes off-the-shelf datasets are suitable; the silicon section explicitly shows this can fail, so this assumption is load-bearing for Sections 3.1, 3.3, 3.5.
  • domain assumption The uncertainty estimate from the auxiliary ACE model used in FT-3 reliably identifies configurations worth adding to the training set.
    FT-3 filtering (Section 2.3, Section 3.4) relies on this uncertainty criterion; if it is miscalibrated, the HEA fine-tuning conclusions change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Study on the Fine-Tuning Performance of Universal Machine-Learned Interatomic Potentials (U-MLIPs)." pith.science (2026). https://pith.science/paper/452CSEZR

@misc{pith2026250607401,
  author       = {Pith},
  title        = {Pith review of: A Study on the Fine-Tuning Performance of Universal Machine-Learned Interatomic Potentials (U-MLIPs)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/452CSEZR}},
  note         = {Machine review of arXiv:2506.07401}
}
read the original abstract

Universal machine-learned interatomic potentials (U-MLIPs) have demonstrated effectiveness across diverse atomistic systems but often require fine-tuning for task-specific accuracy. We investigate the fine-tuning of two MACE-based foundation models, MACE-MP-0 and its variant MACE-MP-0b, and identify key insights. Fine-tuning on task-specific datasets enhances accuracy and, in some cases, outperforms models trained from scratch. Additionally, fine-tuned models benefit from faster convergence due to the strong initial predictions provided by the foundation model. The success of fine-tuning also depends on careful dataset selection, which can be optimized through filtering or active learning. We further discuss practical strategies for achieving better fine-tuning foundation models in atomistic simulations and explore future directions for their development and applications.

Figures

Figures reproduced from arXiv: 2506.07401 by the authors.

Figure 1
Figure 1. Testing MAE versus number of epochs during fine-tuning of the foundation model [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Relative error in defect formation energy of silicon for various MLIPs, including the pre-trained model [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Dislocation glide energy barriers in Fe calculated using various MLIPs, including two foundation models [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Dislocation glide energy barriers in W calculated using various MLIPs, including two pre-trained models [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Left: Illustration of the NbAlMoTiZr high-entropy alloy, colored by element using Ovito software [74]; [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Left: Density of molten LiCl at various temperatures, comparing results from [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Radial distribution function (RDF) of molten LiCl at 900 K, comparing results from [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

92 extracted references · 57 canonical work pages

  1. [1]

    Behler, Perspective: Machine learning potentials for atomistic simulations, The Journal of chemical physics 145 (17) (2016)

    J. Behler, Perspective: Machine learning potentials for atomistic simulations, The Journal of chemical physics 145 (17) (2016). 14 DFTMACE-MP-0bFine-tuning-0bMACE-MP-0Fine-tuning-0 MACE-scratch ACE-scratch Mo a0 3.168 3.177 (+0.3%) 3.16 (-0.3%) 3.170 (+0.1%)3.168 (+0.0%)3.156 (-0.4%)3.168 (+0.0%) C11 472 290 (-38.6%) 483 (+2.3%) 343 (-27.3%) 498 (+5.5%) 5...

    2016

  2. [2]

    Vamathevan, D

    J. Vamathevan, D. Clark, P. Czodrowski, I. Dunham, E. Ferran, G. Lee, B. Li, A. Madab- hushi, P. Shah, M. Spitzer, et al., Applications of machine learning in drug discovery and development, Nature reviews Drug discovery 18 (6) (2019) 463–477

    2019

  3. [3]

    Raccuglia, K

    P. Raccuglia, K. C. Elbert, P. D. Adler, C. Falk, M. B. Wenny, A. Mollo, M. Zeller, S. A. Friedler, J. Schrier, A. J. Norquist, Machine-learning-assisted materials discovery using failed experiments, Nature 533 (7601) (2016) 73–76

    2016

  4. [4]

    D. S. Sholl, J. A. Steckel, Density functional theory: a practical introduction, John Wiley & Sons, 2022

    2022

  5. [5]

    W. Koch, M. C. Holthausen, A chemist’s guide to density functional theory, John Wiley & Sons, 2015

    2015

  6. [6]

    J. W. Ponder, D. A. Case, Force fields for protein simulations, Advances in protein chemistry 66 (2003) 27–85

    2003

  7. [7]

    Altona, D

    C. Altona, D. H. Faber, Empirical force field calculations: A tool in structural organic chem- istry, in: Dynamic Chemistry, Springer, 2005, pp. 1–38

    2005

  8. [8]

    B.-J. Lee, M. I. Baskes, Second nearest-neighbor modified embedded-atom-method potential, Physical Review B 62 (13) (2000) 8564

    2000

Show all 92 references
  1. [9]

    Behler, M

    J. Behler, M. Parrinello, Generalized neural-network representation of high-dimensional potential-energy surfaces, Physical review letters 98 (14) (2007) 146401

    2007

  2. [10]

    A. P. Bartók, M. C. Payne, R. Kondor, G. Csányi, Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons, Physical review letters 104 (13) (2010) 136403

    2010

  3. [11]

    W. C. Witt, C. van der Oord, E. Gelžinyt˙ e, T. Järvinen, A. Ross, J. P. Darby, C. H. Ho, W. J. Baldwin, M. Sachs, J. Kermode, et al., Acepotentials. jl: A julia implementation of the atomic cluster expansion, The Journal of Chemical Physics 159 (16) (2023)

    2023

  4. [12]

    Batatia, D

    I. Batatia, D. P. Kovacs, G. Simm, C. Ortner, G. Csányi, Mace: Higher order equivariant mes- sage passing neural networks for fast and accurate force fields, Advances in Neural Information Processing Systems 35 (2022) 11423–11436

    2022

  5. [13]

    A. V. Shapeev, Moment tensor potentials: A class of systematically improvable interatomic potentials, Multiscale Modeling & Simulation 14 (3) (2016) 1153–1173

    2016

  6. [14]

    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 (2019) 014104

    2019

  7. [15]

    Dusson, M

    G. Dusson, M. Bachmayr, G. Csányi, R. Drautz, S. Etter, C. van der Oord, C. Ortner, Atomic cluster expansion: Completeness, efficiency and stability, J. Comput. Phys. 454 (2022) 110946

    2022

  8. [16]

    S. Luo, T. Chen, A. S. Krishnapriyan, Enabling efficient equivariant operations in the fourier basis via gaunt tensor products, arXiv preprint arXiv:2401.10216 (2024). 16

    2024 arXiv

  9. [17]

    J. P. Darby, D. P. Kovács, I. Batatia, M. A. Caro, G. L. Hart, C. Ortner, G. Csányi, Tensor- reduced atomic density representations, Physical Review Letters 131 (2) (2023) 028001

    2023

  10. [18]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, G. E. Dahl, Neural message passing for quantum chemistry, in: International conference on machine learning, PMLR, 2017, pp. 1263–1272

    2017

  11. [19]

    Gilmer, S

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, G. E. Dahl, Message passing neural networks, in: Machine learning meets quantum physics, Springer, 2020, pp. 199–214

    2020

  12. [20]

    A. J. Thirunavukarasu, D. S. J. Ting, K. Elangovan, L. Gutierrez, T. F. Tan, D. S. W. Ting, Large language models in medicine, Nature medicine 29 (8) (2023) 1930–1940

    2023

  13. [21]

    Kasneci, K

    E. Kasneci, K. Seßler, S. Küchemann, M. Bannert, D. Dementieva, F. Fischer, U. Gasser, G. Groh, S. Günnemann, E. Hüllermeier, et al., Chatgpt for good? on opportunities and challenges of large language models for education, Learning and individual differences 103 (2023) 102274

    2023

  14. [22]

    Chanussot, A

    L. Chanussot, A. Das, S. Goyal, T. Lavril, M. Shuaibi, M. Riviere, K. Tran, J. Heras-Domingo, C. Ho, W. Hu, et al., Open catalyst 2020 (oc20) dataset and community challenges, Acs Catalysis 11 (10) (2021) 6059–6072

    2021

  15. [23]

    J. M. Bowman, C. Qu, R. Conte, A. Nandi, P. L. Houston, Q. Yu, The md17 datasets from the perspective of datasets for gas-phase “small” molecule potentials, The Journal of Chemical Physics 156 (24) (2022)

    2022

  16. [24]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, et al., Commentary: The materials project: A materials genome ap- proach to accelerating materials innovation, APL materials 1 (1) (2013)

    2013

  17. [25]

    Bommasani, D

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, et al., On the opportunities and risks of foundation models, arXiv preprint arXiv:2108.07258 (2021)

    2021 arXiv

  18. [26]

    Batatia, P

    I. Batatia, P. Benner, Y. Chiang, A. M. Elena, D. P. Kovács, J. Riebesell, X. R. Advincula, M. Asta, W. J. Baldwin, N. Bernstein, et al., A foundation model for atomistic materials chemistry, arXiv preprint arXiv:2401.00096 (2023)

    2023 arXiv

  19. [27]

    B.Deng, P.Zhong, K.Jun, J.Riebesell, K.Han, C.J.Bartel, G.Ceder, Chgnetasapretrained universal neural network potential for charge-informed atomistic modelling, Nature Machine Intelligence 5 (9) (2023) 1031–1041

    2023

  20. [28]

    Merchant, S

    A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, E. D. Cubuk, Scaling deep learning for materials discovery, Nature 624 (7990) (2023) 80–85

    2023

  21. [29]

    Zhang, X

    D. Zhang, X. Liu, X. Zhang, C. Zhang, C. Cai, H. Bi, Y. Du, X. Qin, J. Huang, B. Li, et al., Dpa-2: Towards a universal large atomic model for molecular and material simulation, arXiv preprint arXiv:2312.15492 (2023). 17

    2023 arXiv

  22. [30]

    Choudhary, B

    K. Choudhary, B. DeCost, L. Major, K. Butler, J. Thiyagalingam, F. Tavazza, Unified graph neural network force-field for the periodic table: solid state applications, Digital Discovery 2 (2) (2023) 346–355

    2023

  23. [31]

    C. Chen, S. P. Ong, A universal graph deep learning interatomic potential for the periodic table, Nature Computational Science 2 (11) (2022) 718–728

    2022

  24. [32]

    Z. Li, T. Wen, Y. Zhang, X. Liu, C. Zhang, A. Pattamatta, X. Gong, B. Ye, H. Wang, L. Zhang, et al., An extendable cloud-native alloy property explorer, arXiv preprint arXiv:2404.17330 (2024)

    2024 arXiv

  25. [33]

    Focassio, L

    B. Focassio, L. P. M. Freitas, G. R. Schleder, Performance assessment of universal machine learning interatomic potentials: Challenges and directions for materials’ surfaces, ACS Ap- plied Materials & Interfaces (2024)

    2024

  26. [34]

    B. Deng, Y. Choi, P. Zhong, J. Riebesell, S. Anand, Z. Li, K. Jun, K. A. Persson, G. Ceder, Overcoming systematic softening in universal machine learning interatomic potentials by fine- tuning, arXiv preprint arXiv:2405.07105 (2024)

    2024 arXiv

  27. [35]

    S. F. Alavi, Y. Chen, Y.-F. Hou, F. Ge, P. Zheng, P. O. Dral, Towards accurate and efficient anharmonic vibrational frequencies with the universal interatomic potential ani-1ccx-gelu and its fine-tuning (2024)

    2024

  28. [36]

    H. Yu, M. Giantomassi, G. Materzanini, J. Wang, G.-M. Rignanese, Systematic assessment of various universal machine-learning interatomic potentials, Materials Genome Engineering Advances 2 (3) (2024) e58

    2024

  29. [37]

    E.O.Pyzer-Knapp, M.Manica, P.Staar, L.Morin, P.Ruch, T.Laino, J.R.Smith, A.Curioni, Foundation models for materials discovery–current state and future directions, npj Computa- tional Materials 11 (1) (2025) 61

    2025

  30. [38]

    Shuang, Z

    F. Shuang, Z. Wei, K. Liu, W. Gao, P. Dey, Universal machine learning interatomic potentials poised to supplant dft in modeling general defects in metals and random alloys, arXiv preprint arXiv:2502.03578 (2025)

    2025 arXiv

  31. [39]

    H. Du, J. Hui, L. Zhang, H. Wang, Universal machine learning interatomic potentials are ready for solid ion conductors, arXiv preprint arXiv:2502.09970 (2025)

    2025 arXiv

  32. [40]

    H. Lee, V. I. Hegde, C. Wolverton, Y. Xia, Accelerating high-throughput phonon calculations via machine learning universal potentials, Materials Today Physics (2025) 101688

    2025

  33. [41]

    S. P. Niblett, P. Kourtis, I.-B. Magdău, C. P. Grey, G. Csányi, Transferability of datasets between machine-learning interaction potentials, arXiv preprint arXiv:2409.05590 (2024)

    2024 arXiv

  34. [42]

    Casillas-Trujillo, A

    L. Casillas-Trujillo, A. S. Parackal, R. Armiento, B. Alling, Evaluating and improving the predictive accuracy of mixing enthalpies and volumes in disordered alloys from universal pretrained machine learning potentials, Physical Review Materials 8 (11) (2024) 113803

    2024

  35. [43]

    L. Wang, T. He, B. Ouyang, The impact of domain knowledge on universal machine learning models, in: arXiv, 2024. 18

    2024

  36. [44]

    S. Jin, L. Li, J.-G. Liu, Random batch methods (rbm) for interacting particle systems, Journal of Computational Physics 400 (2020) 108877

    2020

  37. [45]

    S. Jin, L. Li, Z. Xu, Y. Zhao, A random batch ewald method for particle systems with coulomb interactions, SIAM Journal on Scientific Computing 43 (4) (2021) B937–B960

    2021

  38. [46]

    Liang, P

    J. Liang, P. Tan, Y. Zhao, L. Li, S. Jin, L. Hong, Z. Xu, Superscalability of the random batch ewald method, The Journal of Chemical Physics 156 (1) (2022)

    2022

  39. [47]

    Liang, Z

    J. Liang, Z. Xu, Y. Zhao, Improved random batch ewald method in molecular dynamics simulations, The Journal of Physical Chemistry A 126 (22) (2022) 3583–3593

    2022

  40. [48]

    W. Gao, T. Zhao, Y. Guo, J. Liang, H. Liu, M. Luo, Z. Luo, W. Qin, Y. Wang, Q. Zhou, et al., Rbmd: A molecular dynamics package enabling to simulate 10 million all-atom particles in a single graphics processing unit, arXiv preprint arXiv:2407.09315 (2024)

    2024

  41. [49]

    C. H. Ho, T. S. Gutleb, C. Ortner, Atomic cluster expansion without self-interaction, arXiv preprint arXiv:2401.01550 (2024)

    2024 arXiv

  42. [50]

    Zhang, B

    L. Zhang, B. Onat, G. Dusson, A. McSloy, G. Anand, R. J. Maurer, C. Ortner, J. R. Kermode, Equivariant analytical mapping of first principles Hamiltonians to accurate and transferable materials models, npj Comput. Mater. 8 (1) (2022) 158

    2022

  43. [51]

    Y. Wang, S. Patel, C. Ortner, A theoretical case study of the generalization of machine-learned potentials, Computer Methods in Applied Mechanics and Engineering 422 (2024) 116831

    2024

  44. [52]

    Torabi, Y

    T. Torabi, Y. Wang, C. Ortner, Surrogate models for vibrational entropy based on a spatial decomposition, arXiv preprint arXiv:2402.12744 (2024)

    2024 arXiv

  45. [53]

    Y. Wang, G. Csanyi, C. Ortner, Many-body coarse-grained molecular dynamics with the atomic cluster expansion, arXiv preprint arXiv:2502.04661 (2025)

    2025 arXiv

  46. [54]

    Vignac, A

    C. Vignac, A. Loukas, P. Frossard, Building powerful and equivariant graph neural networks with structural message-passing, Advances in neural information processing systems 33 (2020) 14143–14155

    2020

  47. [55]

    Maron, H

    H. Maron, H. Ben-Hamu, N. Shamir, Y. Lipman, Invariant and equivariant graph networks, arXiv preprint arXiv:1812.09902 (2018)

    2018 arXiv

  48. [56]

    Tajbakhsh, J

    N. Tajbakhsh, J. Y. Shin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B. Gotway, J. Liang, Convolutional neural networks for medical image analysis: Full training or fine tuning?, IEEE transactions on medical imaging 35 (5) (2016) 1299–1312

    2016

  49. [57]

    M. I. Jordan, T. M. Mitchell, Machine learning: Trends, perspectives, and prospects, Science 349 (6245) (2015) 255–260

    2015

  50. [58]

    Zhang, G

    H. Zhang, G. Li, J. Li, Z. Zhang, Y. Zhu, Z. Jin, Fine-tuning pre-trained language models effectively by optimizing subnetworks adaptively, Advances in Neural Information Processing Systems 35 (2022) 21442–21454. 19

    2022

  51. [59]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional trans- formers for language understanding, in: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologie...

    2019

  52. [60]

    J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C. H. So, J. Kang, Biobert: a pre-trained biomedical language model for biomedical text mining, Bioinformatics 36 (4) (2020) 1234–1240

    2020

  53. [61]

    C. Oord, M. Sachs, D. Kovacs, C. Ortner, G. Csányi, Hyperactive learning for data-driven interatomic potentials, npj Computational Materials 9 (09 2023)

    2023

  54. [62]

    Voita, D

    E. Voita, D. Talbot, F. Moiseev, R. Sennrich, I. Titov, Analyzing multi-head self-attention: Specializedheads dotheheavylifting, therest canbepruned, arXivpreprintarXiv:1905.09418 (2019)

    2019 arXiv

  55. [63]

    S. Kim, H. Yang, Y. Kim, Y. Hong, E. Park, Hydra: Multi-head low-rank adaptation for parameter efficient fine-tuning, Neural Networks (2024) 106414

    2024

  56. [64]

    Y. Zuo, C. Chen, X. Li, Z. Deng, Y. Chen, J. Behler, G. Csányi, A. V. Shapeev, A. P. Thompson, M. A. Wood, S. P. Ong, Performance and cost assessment of machine learning interatomic potentials, J. Phys. Chem. A 124 (4) (2020) 731–745, pMID: 31916773

    2020

  57. [65]

    Henkelman, B

    G. Henkelman, B. P. Uberuaga, H. Jónsson, A climbing image nudged elastic band method for finding saddle points and minimum energy paths, The Journal of chemical physics 113 (22) (2000) 9901–9904

    2000

  58. [66]

    A. P. Bartók, J. Kermode, N. Bernstein, G. Csányi, Machine learning a general-purpose interatomic potential for silicon, Physical Review X 8 (4) (2018) 041048

    2018

  59. [67]

    D.F.Toit, Y.Zhou, V.L.Deringer, Hyperparameteroptimizationforatomicclusterexpansion potentials, arXiv preprint arXiv:2408.00656 (2024)

    2024 arXiv

  60. [68]

    Zhang, G

    L. Zhang, G. Csányi, E. Van Der Giessen, F. Maresca, Atomistic fracture in bcc iron revealed by active learning of gaussian approximation potential, npj Computational Materials 9 (1) (2023) 217

    2023

  61. [69]

    Guénolé, W

    J. Guénolé, W. G. Nöhring, A. Vaid, F. Houllé, Z. Xie, A. Prakash, 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 (2020) 109584

    2020

  62. [70]

    Grigorev, L

    P. Grigorev, L. Frérot, F. Birks, A. Gola, J. Golebiowski, J. Grießer, J. L. Hörmann, A. Kle- menz, G. Moras, W. G. Nöhring, et al., matscipy: materials science at the atomic scale with python, Journal of Open Source Software 9 (93) (2024)

    2024

  63. [71]

    Byggmästar, A

    J. Byggmästar, A. Hamedani, K. Nordlund, F. Djurabekova, Machine-learning interatomic potential for radiation damage and defects in tungsten, Physical Review B 100 (14) (2019) 144105

    2019

  64. [72]

    D. B. Miracle, O. N. Senkov, A critical review of high entropy alloys and related concepts, Acta materialia 122 (2017) 448–511. 20

    2017

  65. [73]

    E. P. George, D. Raabe, R. O. Ritchie, High-entropy alloys, Nature reviews materials 4 (8) (2019) 515–534

    2019

  66. [74]

    A. Stukowski, Visualization and analysis of atomistic simulation data with ovito–the open visualization tool, Modelling and simulation in materials science and engineering 18 (1) (2009) 015012

    2009

  67. [75]

    J. Gou, B. Yu, S. J. Maybank, D. Tao, Knowledge distillation: A survey, International Journal of Computer Vision 129 (6) (2021) 1789–1819

    2021

  68. [76]

    Alkhulaifi, F

    A. Alkhulaifi, F. Alsahli, I. Ahmad, Knowledge distillation in deep learning and its applica- tions, PeerJ Computer Science 7 (2021) e474

    2021

  69. [77]

    Sivaraman, J

    G. Sivaraman, J. Guo, L. Ward, N. Hoyt, M. Williamson, I. Foster, C. Benmore, N. Jackson, Automated development of molten salt machine learning potentials: application to licl, The Journal of Physical Chemistry Letters 12 (17) (2021) 4278–4285

    2021

  70. [78]

    Liu, C.-Y

    S.-Y. Liu, C.-Y. Wang, H. Yin, P. Molchanov, Y.-C. F. Wang, K.-T. Cheng, M.-H. Chen, Dora: Weight-decomposed low-rank adaptation, arXiv preprint arXiv:2402.09353 (2024)

    2024 arXiv

  71. [79]

    A. E. Allen, N. Lubbers, S. Matin, J. Smith, R. Messerly, S. Tretiak, K. Barros, Learning together: Towards foundation models for machine learning interatomic potentials with meta- learning, npj computational materials 10 (1) (2024) 154

    2024

  72. [80]

    Matin, A

    S. Matin, A. Allen, E. Shinkle, A. Pachalieva, G. T. Craven, B. Nebgen, J. Smith, R. Messerly, Y. W. Li, S. Tretiak, et al., Teacher-student training improves accuracy and efficiency of machine learning inter-atomic potentials, arXiv preprint arXiv:2502.05379 (2025)

    2025 arXiv

  73. [81]

    H. Chen, C. Ortner, Y. Wang, Qm/mm methods for crystalline defects. part 3: Machine- learned mm models, Multiscale Modeling & Simulation 20 (4) (2022) 1490–1518

    2022

  74. [82]

    Y. Wang, H. Chen, M. Liao, C. Ortner, H. Wang, L. Zhang, A posteriori error estimates for adaptive qm/mm coupling methods, SIAM Journal on Scientific Computing 43 (4) (2021) A2785–A2808

    2021

  75. [83]

    K. Fu, M. Liao, Y. Wang, J. Chen, L. Zhang, Meshac: A 3d mesh generation and adaptation package for multiscale coupling methods, Computer Physics Communications (2025) 109523

    2025

  76. [84]

    K. Fu, M. Liao, Y. Wang, J. Chen, L. Zhang, Adaptive multigrid strategy for geometry optimization of large-scale three dimensional molecular mechanics, Journal of Computational Physics 485 (2023) 112113

    2023

  77. [85]

    Y. Wang, A posteriori analysis and adaptive algorithms for blended type atomistic-to- continuum coupling with higher-order finite elements, Computer Physics Communications (2025) 109533

    2025

  78. [86]

    Siddiqui, N

    A. Siddiqui, N. D. Hine, Machine-learned interatomic potentials for transition metal dichalco- genide mo1- x w x s2- 2 y se2 y alloys, npj Computational Materials 10 (1) (2024) 169. 21

    2024

  79. [87]

    Vandenhaute, M

    S. Vandenhaute, M. Cools-Ceuppens, S. DeKeyser, T. Verstraelen, V. Van Speybroeck, Ma- chinelearningpotentialsformetal-organicframeworksusinganincrementallearningapproach, npj Computational Materials 9 (1) (2023) 1–8

    2023

  80. [88]

    A. F. Harper, M. L. Evans, J. P. Darby, B. Karasulu, C. P. Koçer, J. R. Nelson, A. J. Morris, Abinitiostructurepredictionmethodsforbatterymaterials: Areviewofrecentcomputational effortstopredicttheatomiclevelstructureandbondinginmaterialsforrechargeablebatteries, Johnson Matth...

    2020

  81. [89]

    J. Wang, C. Jiang, W. Li, X. Xiao, Anisotropic low-dimensional materials for polarization- sensitive photodetectors: from materials to devices, Advanced Optical Materials 10 (6) (2022) 2102436

    2022

  82. [90]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N.Gimelshein, L.Antiga, etal., Pytorch: Animperativestyle, high-performancedeeplearning library, Advances in neural information processing systems 32 (2019)

    2019

  83. [91]

    Geiger, T

    M. Geiger, T. Smidt, e3nn: Euclidean neural networks, arXiv preprint arXiv:2207.09453 (2022)

    2022 arXiv

  84. [92]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014). 22

    2014 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.