Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Universal Machine Learning Interatomic Potentials are Ready for Solid Ion Conductors

T0 review · 4 major / 6 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 The paper claims that MatterSim, a universal machine learning interatomic potential, beats five other models on nearly every benchmark for solid electrolytes, with near-DFT accuracy in energy, forces, and lithium diffusion on 18 systems.

desk verdict A solid energy/force/elastic benchmark of six uMLIPs for solid electrolytes, but the transport ranking rests on DeepMD as the reference and lacks error bars; the 'ready' claim is stronger than the evidence. read the letter →

arxiv 2502.09970 v1 pith:AMHGGHTV submitted 2025-02-14 cond-mat.mtrl-sci cs.LG

Hongwei Du , Jian Hui , Lanting Zhang , Hong Wang This is my paper
classification cond-mat.mtrl-scics.LG
keywords solid-stateelectrolytesuniversalmachinelearninginteratomicpotentialspretrainedmodellithium-iondiffusivityionicconductivityMatterSimbenchmarkframeworkmoleculardynamics
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

This paper asks whether off-the-shelf machine learning interatomic potentials can replace density functional theory and hand-built force fields for studying solid lithium-ion electrolytes. The authors benchmark six universal potentials on energy, forces, elastic moduli, formation energies, thermodynamic stability, and lithium diffusivity across 18 halide, sulfide, and oxide systems, and report that MatterSim is the most accurate and consistent across nearly every test. The practical payoff is that a single pretrained potential can deliver near-DFT-quality energies and room-temperature transport coefficients at molecular-dynamics speeds, making high-throughput screening of solid electrolyte candidates feasible. Using MatterSim, the paper also identifies structural factors that control conductivity in Li3YCl6 and Li6PS5Cl, such as intermediate S/Cl disorder around 40-50% and Na/Li cation arrangements.

What carries the argument

The central object is the universal machine learning interatomic potential (uMLIP), a neural network that maps atomic structures to energies and forces and can serve as the calculator for relaxation, elastic constants, and molecular dynamics. The evaluation is built on a three-tier benchmark: DFT-labeled energy and force snapshots from ab initio and active-learning trajectories, reference thermodynamic and mechanical property sets, and long-time molecular dynamics with lithium-ion diffusivity extracted from the mean-square displacement via the Einstein relation and converted to conductivity through the Nernst-Einstein equation. For the mechanism studies, the resolving tools are the DeepMD reference potentials, van Hove correlation functions, and lithium probability-density isosurfaces.

What would settle it

Run density-functional-theory molecular dynamics at 300 K on a subset of the 18 electrolyte systems, extract lithium diffusivities from the mean-square displacement, and compare them with MatterSim and DeepMD; if another universal potential matches the direct reference better than MatterSim at low temperature, the paper's transport-superiority claim would be overturned.

Watch

Extended reading notes

Core claim

The central claim is that MatterSim, one of six universal machine learning interatomic potentials, is ready to serve as a high-precision calculator for solid ion conductors. In the paper's benchmark, MatterSim achieves the lowest mean absolute errors for energies and forces in both equilibrium and non-equilibrium structures, matches density functional theory on bulk and shear moduli, formation energies, and phase-stability energies, and reproduces lithium-ion diffusivities that agree with DeepMD reference values at all temperatures, especially at 300 K. The paper further claims that this accuracy transfers to mechanism discovery: in Li6PS5Cl, an S/Cl disorder level near 40-50% connects diffusion pathways and maximizes conductivity, while in NaxLi3-xYCl6, higher lithium content and specific Na/Li arrangements enlarge channels and lower migration barriers.

Load-bearing premise

The transport ranking stands or falls on the assumption that the reference DeepMD potentials faithfully reproduce true lithium diffusion in these materials, particularly at 300 K where ionic motion is slow.

Editorial extensions

If this is right

  • MatterSim can replace density functional theory and empirical force fields for routine energy, force, elastic, and thermodynamic calculations on halide, sulfide, and oxide solid electrolytes.
  • Room-temperature lithium diffusivities from MatterSim are reliable enough to rank and screen candidate solid electrolytes before expensive quantum-mechanical validation.
  • An S/Cl disorder level around 40-50% in Li6PS5Cl maximizes ionic conductivity by connecting diffusion pathways, giving a concrete target for disorder engineering.
  • Raising lithium content and optimizing Na/Li arrangements in NaxLi3-xYCl6 lowers migration barriers and improves transport, yielding design rules for halide electrolytes.
  • The benchmark framework offers a standardized protocol for evaluating future universal potentials on solid electrolytes, not just the six compared here.

Reading between the lines

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

  • Beyond the paper, the same benchmark design could be extended to sodium or magnesium solid electrolytes, where cation size and charge change the diffusion physics that a universal potential would need to reproduce.
  • Beyond the paper, the room-temperature transport ranking is the most fragile part of the claim; direct density-functional-theory molecular dynamics on a few low-temperature systems would settle whether the agreement is an artifact of the DeepMD reference.
  • Beyond the paper, if MatterSim's accuracy generalizes beyond the 18 tested systems, universal potentials could become the default first screening pass for electrolyte discovery, with density functional theory reserved for final validation of the most promising candidates.
  • Beyond the paper, the disorder-conductivity peak near 40-50% S/Cl substitution suggests an experimental synthesis target for tuning anion ordering in Li6PS5Cl.
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 / 6 minor

Summary. The paper presents a systematic evaluation of six universal machine-learning interatomic potentials (MatterSim, MACE, SevenNet, CHGNet, M3GNet, ORBFF) for solid-state electrolytes, benchmarking them on energy and force prediction against DFT, on thermodynamic and mechanical properties against Materials Project data, and on lithium-ion diffusivity against DeepMD reference simulations. The authors conclude that MatterSim is the most accurate and robust across nearly all metrics, especially for ion transport, and they use MatterSim to study S/Cl disorder in Li6PS5Cl and Na/Li substitution in Li3YCl6. The static-property benchmarks are well grounded in external DFT/MP data, but the diffusivity benchmarking relies entirely on DeepMD as the reference, which is itself a machine-learned surrogate and is not validated against DFT-MD or experiment in this work.

Significance. If the claims are correct, the paper provides a useful practical benchmark for choosing MLIPs in solid-electrolyte simulations, and the MatterSim-based mechanistic findings would be valuable. The strength of the paper is its comprehensive static-property evaluation against independent DFT and Materials Project references, which credibly supports MatterSim's accuracy for energies, forces, and elastic/thermodynamic properties. However, the central transport claim, highlighted in the title and abstract, is not established because the diffusivity benchmark uses DeepMD as the reference standard without validation, and the reported 300 K data lack statistical uncertainty. The significance of the paper as a definitive 'ready for solid ion conductors' statement is therefore conditional on additional validation of the transport benchmarks.

major comments (4)
  1. [Section 3.1 and Section 3.4] The lithium-ion diffusivity benchmark uses DeepMD as the reference standard, as stated in Section 3.1 ('with DeepMD results serving as the reference standard'). DeepMD is itself a machine-learned surrogate trained on DFT labels, and the manuscript provides no validation of these system-specific DeepMD potentials against direct AIMD or experimental conductivity for the 18 SSEs. Consequently, the reported agreement between MatterSim and DeepMD demonstrates model-model consistency only, not DFT-level or experimental accuracy. Since the title and abstract claim that uMLIPs (specifically MatterSim) are 'ready' for solid ion conductors, this load-bearing transport claim is unsupported as it stands. The authors should validate DeepMD against AIMD or experiment for at least a subset of systems, or benchmark the uMLIPs directly against AIMD or experimental data, or substantially weaken the transport conclusions.
  2. [Section 3.4 and Figure 17] The diffusion coefficients and conductivities are reported without any statistical uncertainty. Figure 17 shows that the 300 K MSD curves plateau within 1 ns, indicating very limited diffusive motion; extracting D from the slope of such curves is highly sensitive to non-diffusive oscillations and the fit window. No multiple independent seeds, block averaging, or convergence checks are reported. The room-temperature ranking of MatterSim versus MACE and SevenNet, which is a headline claim, may not be statistically significant once finite-sampling errors are accounted for. The authors should provide error bars on all reported D and conductivity values and assess the robustness of the ranking.
  3. [Tables 1 and 2, Section 4.2] The column labeled 'DFT/Deepmd' conflates two different reference types: DFT for static properties (E_hull, E_f, bulk/shear modulus) and DeepMD for the room-temperature conductivity. This label is misleading because the conductivity reference is not DFT-based. Moreover, the claim of 'excellent agreement' with DeepMD is not supported by the numbers: for Li6PS5Cl, MatterSim predicts sigma_300K = 0.4139 mS/cm while the DeepMD reference is 0.1379 mS/cm, a factor of three discrepancy. For Li3YCl6 the values are 0.4553 vs 0.501 mS/cm, which are close, but the Li6PS5Cl case shows that the agreement is not uniformly excellent. The authors should separate the DFT and DeepMD columns and discuss the factor-of-three difference, which corresponds to a nontrivial error in activation barrier at room temperature.
  4. [Section 4.3] The mechanistic conclusions about S/Cl anion disorder levels in Li6PS5Cl and Na/Li arrangements in NaxLi3-xYCl6 are based solely on MatterSim MD simulations and are compared only with DeepMD calculations. Since DeepMD is not validated as a reliable reference for these systems, these conclusions inherit the same weakness as the diffusivity benchmark. The authors should compare at least a few of their predicted conductivities or MSD trends with direct AIMD or experimental data (e.g., known experimental conductivities for Li6PS5Cl and Li3YCl6) to support the mechanistic claims, or the claims should be presented as model-based hypotheses rather than validated mechanisms.
minor comments (6)
  1. [Abstract and Section 3.3] The abstract states that MatterSim 'outperforms others in nearly all metrics,' but Section 3.3 reports that SevenNet performs best on formation energy and on E_above_hull. Please qualify the abstract to reflect the property-specific rankings rather than a blanket superiority claim.
  2. [Figure 17] The MSD curves in Figure 17 appear to plateau at 300 K, which suggests non-diffusive behavior; the authors should state explicitly that the extracted D values at 300 K are upper-bound estimates or treat them as tentative. Also, the isosurface contour levels are given as 0.002/a0^3 in one figure and 0.0002/a0^3 in another; please ensure consistent units and clearly define the threshold.
  3. [Section 2.2 and Section 3.1] The description of the datasets is confusing: Section 2.2 states that each SSE system has a final dataset of 3,326 structures (59,868 snapshots total) for DeepMD training, while Section 3.1 states that the energy/force benchmark dataset includes 1,980 snapshots. Please clarify which dataset is used for training versus benchmarking, and avoid the ambiguous phrase 'final training set' in Section 3.1.
  4. [General] No data or code availability statement is provided. Given that the paper proposes a 'benchmark framework' and reports detailed numerical results, making the dataset, workflows, and analysis scripts available would greatly aid reproducibility and community uptake.
  5. [Throughout] There are numerous typographical and formatting inconsistencies: 'Mattersim' versus 'MatterSim', 'ORB' versus 'ORBFF' in Tables 1 and 2, 'systemic' instead of 'systematic' in Sections 3.3 and 5, and a badly formatted Table 3 with unclear column alignment. These should be corrected in a final revision.
  6. [Tables 1 and 2 captions] The captions for Tables 1 and 2 state that metrics are 'compared with reference values based on DFT,' but the conductivity column is a DeepMD result. Please revise the captions to correctly identify the reference method for each property.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark claims are evaluated against external DFT and Materials Project references, and the DeepMD diffusivity baseline is an independent surrogate rather than a construction-level circular input.

full rationale

The paper's central claims about MatterSim are not derived from MatterSim's own outputs or from a self-referential definition. Energy and force accuracy are checked against DFT labels on AIMD/DeepGen-sampled structures (Section 3.1), thermodynamic and elastic properties are checked against Materials Project DFT data (Section 3.3), and the phase-stability numbers in Section 4.2 are compared with explicit DFT reference values. None of these comparisons define the target quantity in terms of the model being tested. The diffusivity benchmark in Section 3.4 uses DeepMD as the reference standard, and Section 2.2 shows that the DeepMD potentials were trained in this work on DFT labels. This makes the transport comparison a model-to-model agreement test rather than a direct DFT or experimental validation, which weakens the evidential weight of the 'ready for solid ion conductors' claim. However, this is not circularity in the required sense: DeepMD is an independently trained surrogate, not a quantity fitted to MatterSim or defined by the same equations, and no equation in the paper reduces MatterSim's diffusivity to the DeepMD values by construction. There is also no load-bearing self-citation chain; references to MatterSim [23], DeePMD [30], and the phonon-benchmark paper [31] are external prior works and are not used to force the conclusion. Under the stated hard rules, the absence of an explicit equation-level or definition-level reduction means the appropriate finding is no significant circularity.

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

The central comparison leans on DFT and Materials Project as external references for most properties. The only weakly supported reference is DeepMD for diffusivity, plus standard DFT/MD modeling assumptions that are not independently verified in the paper.

assumptions (4)
  • domain assumption PBE-DFT energies and forces computed with VASP are the ground truth for evaluating uMLIP energy and force accuracy.
    Section 3.1 describes DFT labels using PAW/PBE at 600 eV; the entire energy/force ranking assumes this functional and protocol are adequate for these electrolytes.
  • ad hoc to paper DeepMD per-system potentials provide a faithful reference for lithium-ion diffusivity.
    Section 3.1 states 'with DeepMD results serving as the reference standard' for diffusion; no DFT-MD or experimental conductivities are used to validate this reference for the 18 systems.
  • domain assumption Materials Project formation energies, elastic moduli, and convex hull data are reliable references.
    Section 3.3 builds thermodynamic and mechanical datasets from MP; inaccuracies in MP entries propagate into model rankings.
  • domain assumption MD simulation times of 1-10 ns and 3x3x3 supercells are sufficient to converge lithium-ion diffusion coefficients.
    The paper asserts long-duration MD and large supercells reduce finite-size effects, but no convergence analysis with respect to simulation length is shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Universal Machine Learning Interatomic Potentials are Ready for Solid Ion Conductors." pith.science (2026). https://pith.science/paper/AMHGGHTV

@misc{pith2026250209970,
  author       = {Pith},
  title        = {Pith review of: Universal Machine Learning Interatomic Potentials are Ready for Solid Ion Conductors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AMHGGHTV}},
  note         = {Machine review of arXiv:2502.09970}
}
read the original abstract

With the rapid development of energy storage technology, high-performance solid-state electrolytes (SSEs) have become critical for next-generation lithium-ion batteries. These materials require high ionic conductivity, excellent electrochemical stability, and good mechanical properties to meet the demands of electric vehicles and portable electronics. However, traditional methods like density functional theory (DFT) and empirical force fields face challenges such as high computational costs, poor scalability, and limited accuracy across material systems. Universal machine learning interatomic potentials (uMLIPs) offer a promising solution with their efficiency and near-DFT-level accuracy.This study systematically evaluates six advanced uMLIP models (MatterSim, MACE, SevenNet, CHGNet, M3GNet, and ORBFF) in terms of energy, forces, thermodynamic properties, elastic moduli, and lithium-ion diffusion behavior. The results show that MatterSim outperforms others in nearly all metrics, particularly in complex material systems, demonstrating superior accuracy and physical consistency. Other models exhibit significant deviations due to issues like energy inconsistency or insufficient training data coverage.Further analysis reveals that MatterSim achieves excellent agreement with reference values in lithium-ion diffusivity calculations, especially at room temperature. Studies on Li3YCl6 and Li6PS5Cl uncover how crystal structure, anion disorder levels, and Na/Li arrangements influence ionic conductivity. Appropriate S/Cl disorder levels and optimized Na/Li arrangements enhance diffusion pathway connectivity, improving overall ionic transport performance.

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. Fine-Tuning Universal Machine-Learned Interatomic Potentials: A Tutorial on Methods and Applications

    physics.comp-ph 2025-06 conditional novelty 4.0 of 10

    Fine-tuning universal MLIPs improves accuracy and data efficiency across electrolytes, defects, and interfaces, with some evidence of implicit long-range behavior that is not conclusive.

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

    physics.comp-ph 2025-06 conditional novelty 4.0 of 10

    Fine-tuning universal MACE potentials on targeted datasets generally improves accuracy and convergence speed, though data selection, not the foundation model alone, determines success.

Reference graph

Works this paper leans on

43 extracted references · 14 canonical work pages · cited by 2 Pith papers

  1. [1]

    Solid -State lithium -ion bat tery electrolytes: Revolutionizing energy density and safety,

    P. U. Nzereogu et al. , "Solid -State lithium -ion bat tery electrolytes: Revolutionizing energy density and safety," Hybrid Advances, vol. 8, p. 100339, 2025/03/01/ 2025, doi: https://doi.org/10.1016/j.hybadv.2024.100339

    arXiv 2025

  2. [2]

    Designing solid -state electrolytes for safe, energy -dense batteries,

    Q. Zhao, S. Stalin, C. -Z. Zhao, and L. A. Archer, "Designing solid -state electrolytes for safe, energy -dense batteries," Nature Reviews Materials, vol. 5, no. 3, pp. 229-252, 2020/03/01 2020, doi: 10.1038/s41578-019- 0165-5

    work page doi:10.1038/s41578-019- 2020

  3. [3]

    A solid future for batte ry development,

    J. Janek and W. G. Zeier, "A solid future for batte ry development," Nature Energy, vol. 1, no. 9, p. 16141, 2016/09/08 2016, doi: 10.1038/nenergy.2016.141

    doi:10.1038/nenergy.2016.141 2016

  4. [4]

    Challenges in speeding up solid -state battery development,

    J. Janek and W. G. Zeier, "Challenges in speeding up solid -state battery development," Nature Energy, vol. 8, no. 3, pp. 230-240, 2023/03/01 2023, doi: 10.1038/s41560-023-01208-9

    doi:10.1038/s41560-023-01208-9 2023

  5. [5]

    Fundamentals of inorganic solid - state electrolytes for batteries,

    T. Famprikis, P. Canepa, J. A. Dawson, M. S. Islam, and C. Masquelier, "Fundamentals of inorganic solid - state electrolytes for batteries," Nature Materials, vol. 18, no. 12, pp. 1278 -1291, 2019/12/01 2019, doi: 10.1038/s41563-019-0431-3

    doi:10.1038/s41563-019-0431-3 2019

  6. [6]

    Lithium superionic conductors with corner -sharing frameworks,

    K. Jun et al., "Lithium superionic conductors with corner -sharing frameworks," Nature Materials, vol. 21, no. 8, pp. 924-931, 2022/08/01 2022, doi: 10.1038/s41563-022-01222-4

    doi:10.1038/s41563-022-01222-4 2022

  7. [7]

    A lithium superionic conductor for mil limeter-thick battery electrode,

    Y. Li et al., "A lithium superionic conductor for mil limeter-thick battery electrode," Science, vol. 381, no. 6653, pp. 50-53, 2023/07/07 2023, doi: 10.1126/science.add7138

    work page doi:10.1126/science.add7138 2023

  8. [8]

    High-Voltage Superionic Halide Solid Electrolytes for All-Solid-State Li-Ion Batteries,

    K.-H. Park, K. Kaup, A. Assoud, Q. Zhang, X. Wu, and L. F. Nazar, "High-Voltage Superionic Halide Solid Electrolytes for All-Solid-State Li-Ion Batteries," ACS Energy Letters, vol. 5, no. 2, pp. 533-539, 2020/02/14 2020, doi: 10.1021/acsenergylett.9b02599

    work page doi:10.1021/acsenergylett.9b02599 2020

Show all 43 references
  1. [9]

    Prospects of halide-based all-solid-state batteries: From material design to practical application,

    C. Wang, J. Liang, J. T. Kim, and X. Sun, "Prospects of halide-based all-solid-state batteries: From material design to practical application," Science Advances, vol. 8, no. 36, p. eadc9516, doi: 10.1126/sciadv.adc9516

    doi

  2. [10]

    Carbon-free high-loading silicon anodes enabled by sulfide solid electrolytes,

    D. H. S. Tan et al., "Carbon-free high-loading silicon anodes enabled by sulfide solid electrolytes," Science, vol. 373, no. 6562, pp. 1494-1499, 2021/09/24 2021, doi: 10.1126/science.abg7217

    2021 doi

  3. [11]

    The General AMBER Force Field (GAFF) Can Accurately Predict Thermodynamic and Transport Properties of Many Ionic Liquids,

    K. G. Sprenger, V. W. Jaeger, and J. Pfaendtner, "The General AMBER Force Field (GAFF) Can Accurately Predict Thermodynamic and Transport Properties of Many Ionic Liquids," The Journal of Physical Chemistry B, vol. 119, no. 18, pp. 5882-5895, 2015/05/07 2015, doi: 10.1021/acs....

    2015 doi

  4. [12]

    CHARMM at 45: Enhancements in Accessibility, Functionality, and Speed,

    W. Hwang et al., "CHARMM at 45: Enhancements in Accessibility, Functionality, and Speed," The Journal of Physical Chemistry B, vol. 128, no. 41, pp. 9976-10042, 2024/10/17 2024, doi: 10.1021/acs.jpcb.4c04100

    2024 doi

  5. [13]

    Extension of the GROMOS 56a6CARBO/CARBO_R Force Field for Charged, Protonated, and Esterified Uronates,

    K. Panczyk, K. Gaweda, M. Drach, and W. Plazinski, "Extension of the GROMOS 56a6CARBO/CARBO_R Force Field for Charged, Protonated, and Esterified Uronates," The Journal of Physical Chemistry B, vol. 122, no. 14, pp. 3696-3710, 2018/04/12 2018, doi: 10.1021/acs.jpcb.7b11548

    2018 doi

  6. [14]

    Self-Consistent Equations Including Exchange and Correlation Effects,

    W. Kohn and L. J. Sham, "Self-Consistent Equations Including Exchange and Correlation Effects," Physical Review, vol. 140, no. 4A, pp. A1133-A1138, 11/15/ 1965, doi: 10.1103/PhysRev.140.A1133

    1965 doi

  7. [15]

    Anharmonic Molecular Mechanics: Ab Initio Based Morse Parametrizations for the Popular MM3 Force Field,

    R. J. Shannon, B. Hornung, D. P. Tew , and D. R. Glowacki, "Anharmonic Molecular Mechanics: Ab Initio Based Morse Parametrizations for the Popular MM3 Force Field," The Journal of Physical Chemistry A, vol. 123, no. 13, pp. 2991-2999, 2019/04/04 2019, doi: 10.1021/acs.jpca.8b12006

    2019 doi

  8. [16]

    Perspective: Machine learning potentials for atomistic simulations,

    J. Behler, "Perspective: Machine learning potentials for atomistic simulations," The Journal of Chemical Physics, vol. 145, no. 17, p. 170901, 2016, doi: 10.1063/1.4966192

    2016 doi

  9. [17]

    Machine Learning and Energy Minimization Approaches for Crystal Structure Predictions: A Review and New Horizons,

    J. Graser, S. K. Kauwe, and T. D. Sparks, "Machine Learning and Energy Minimization Approaches for Crystal Structure Predictions: A Review and New Horizons," Chemistry of Materials, vol. 30, no. 11, pp. 3601-3612, 2018/06/12 2018, doi: 10.1021/acs.chemmater.7b05304

    2018 doi

  10. [18]

    Recent advances and applications of machine learning in solid-state materials science,

    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 Computational Materials, vol. 5, no. 1, p. 83, 2019/08/08 2019, doi: 10.1038/s41524-019-0221-0

    2019 doi

  11. [19]

    Machine Learning Force Fields,

    O. T. Unke et al., "Machine Learning Force Fields," Chemical Reviews, vol. 121, no. 16, pp. 10142 -10186, 2021/08/25 2021, doi: 10.1021/acs.chemrev.0c01111

    2021 doi

  12. [20]

    Riebesell, R

    J. Riebesell, R. Goodall, A. Jain, P. Benner, K. Persson, and A. Lee, Matbench Discovery -- An evaluation framework for machine learning crystal stability prediction. 2023

    2023

  13. [21]

    Generalizing Denoising to Non -Equilibrium Structures Improves Equivariant Force Fields,

    Y.-L. Liao, T. E. Smidt, and A. Das, "Generalizing Denoising to Non -Equilibrium Structures Improves Equivariant Force Fields," ArXiv, vol. abs/2403.09549, 2024

    2024 arXiv

  14. [22]

    Systematic softening in universal machine learning intera tomic potentials,

    B. Deng et al. , "Systematic softening in universal machine learning intera tomic potentials," npj Computational Materials, vol. 11, no. 1, p. 9, 2025/01/10 2025, doi: 10.1038/s41524-024-01500-6

    2025 doi

  15. [23]

    MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures,

    H. Yang et al. , "MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures," 2024

    2024

  16. [24]

    Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models,

    L. Barroso-Luque et al., "Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models," ArXiv, vol. abs/2410.12771, 2024

    2024 arXiv

  17. [25]

    MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields,

    I. Batatia, D. a. P. e. Kov'acs, G. N. C. Simm, C. Ortner, and G. Csá nyi, "MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields," ArXiv, vol. abs/2206.07697, 2022

    2022 arXiv

  18. [26]

    Scalable Parallel Algorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations,

    Y. Park, J. Kim, S. Hwang, and S. Han, "Scalable Parallel Algorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations," Journal of Chemical Theory and Computation, vol. 20, no. 11, pp. 4857-4868, 2024/06/11 2024, doi: 10.1021/acs.jctc.4c00190

    2024 doi

  19. [27]

    CHGNet as a pretrained universal neural network potential for charge -informed atomistic modelling,

    B. Deng et al., "CHGNet as a pretrained universal neural network potential for charge -informed atomistic modelling," Nature Machine Intelligence, vol. 5, no. 9, pp. 1031 -1041, 2023/09/01 2023, doi: 10.1038/s42256-023-00716-3

    2023 doi

  20. [28]

    A universal graph deep learning interatomic potential for the periodic table,

    C. Chen and S. P. Ong, "A universal graph deep learning interatomic potential for the periodic table," Nature Computational Science, vol. 2, no. 11, pp. 718-728, 2022/11/01 2022, doi: 10.1038/s43588-022-00349-3

    2022 doi

  21. [29]

    Orb: A Fast, Scalable Neural Network Potential,

    M. Neumann et al., "Orb: A Fast, Scalable Neural Network Potential," ArXiv, vol. abs/2410.22570, 2024

    2024 arXiv

  22. [30]

    DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics,

    H. Wang, L. Zhang, J. Han, and E. Weinan, "DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics," ArXiv, vol. abs/1712.03641, 2017

    2017 arXiv

  23. [31]

    Universal Machine Learning Interatomic Potentials are Ready for Phonons,

    A. Loew, D. Sun, H. -C. Wang, S. Botti, and M. A. L. Marques, "Universal Machine Learning Interatomic Potentials are Ready for Phonons," 2024

    2024

  24. [32]

    Neural Message Passing for Quantum Chemistry,

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, "Neural Message Passing for Quantum Chemistry," presented at the Proceedings of the 34th International Conference on Machine Learning, Proceedings of Machine Learning Research, 2017. [Online]. Available: htt...

    2017

  25. [33]

    Generalized Neural -Network Representation of High -Dimensional Potential- Energy Surfaces,

    J. Behler and M. Parrinello, "Generalized Neural -Network Representation of High -Dimensional Potential- Energy Surfaces," Physical Review Letters, vol. 98, no. 14, p. 146401, 04/02/ 2007, doi: 10.1103/PhysRevLett.98.146401

    2007 doi

  26. [34]

    E(3) -equivariant graph neural networks for data -efficient and accurate interatomic potentials,

    S. Batzner et al. , "E(3) -equivariant graph neural networks for data -efficient and accurate interatomic potentials," Nature Communications, vol. 13, no. 1, p. 2453, 2022/05/0 4 2022, doi: 10.1038/s41467 -022- 29939-5

    2022 doi

  27. [35]

    The atomic simulation environment -a Python library for working with atoms,

    A. Hjorth Larsen et al. , "The atomic simulation environment -a Python library for working with atoms," Journal of physics. Condensed matter : an Institute of Physics journal, vol. 29 27, p. 273002, 2017

    2017

  28. [36]

    Assessment and optimization of the fast inertial relaxation engine (fire) for energy minimization in atomistic simulations and its implementation in lammps,

    J. Gué nolé et al. , "Assessment and optimization of the fast inertial relaxation engine (fire) for energy minimization in atomistic simulations and its implementation in lammps," Computational Materials Science, vol. 175, p. 109584, 2020/04/01/ 2020, doi: https://doi.org/10.1...

    2020

  29. [37]

    Commentary: The Materials Project: A materials genome approach to accelerating materials innovation,

    A. Jain et al., "Commentary: The Materials Project: A materials genome approach to accelerating materials innovation," APL Materials, vol. 1, no. 1, p. 011002, 2013, doi: 10.1063/1.4812323

    2013 doi

  30. [38]

    Active learning of uniformly accurate interatomic potentials for materials simulation,

    L. Zhang, D. -Y. Lin, H. Wang, R. Car, and W. E, "Active learning of uniformly accurate interatomic potentials for materials simulation," Physical Review Materials, vol. 3, no. 2, p. 023804, 02/25/ 201 9, doi: 10.1103/PhysRevMaterials.3.023804

    doi

  31. [39]

    Ab initio molecular dynamics: Concepts, recent developments, and future trends,

    R. Iftimie, P. Minary, and M. E. Tuckerman, "Ab initio molecular dynamics: Concepts, recent developments, and future trends," Proceedings of the National Academy of Sciences, vol. 102, no. 19, pp. 6654 -6659, 2005/05/10 2005, doi: 10.1073/pnas.0500193102

    2005 doi

  32. [40]

    Projector augmented-wave method,

    P. E. Blö chl, "Projector augmented-wave method," Physical Review B, vol. 50, no. 24, pp. 17953 -17979, 12/15/ 1994, doi: 10.1103/PhysRevB.50.17953

    1994 doi

  33. [41]

    Robust training of machine learning interat omic potentials with dimensionality reduction and stratified sampling,

    J. Qi, T. W. Ko, B. C. Wood, T. A. Pham, and S. P. Ong, "Robust training of machine learning interat omic potentials with dimensionality reduction and stratified sampling," npj Computational Materials, vol. 10, no. 1, p. 43, 2024/02/26 2024, doi: 10.1038/s41524-024-01227-4

    2024 doi

  34. [42]

    Data-Driven First-Principles Methods for the Study and Design of Alkali Superionic Conductors,

    Z. Deng, Z. Zhu, I.-H. Chu, and S. P. Ong, "Data-Driven First-Principles Methods for the Study and Design of Alkali Superionic Conductors," Chemistry of Materials, vol. 29, no. 1, pp. 281-288, 2017/01/10 2017, doi: 10.1021/acs.chemmater.6b02648

    2017 doi

  35. [43]

    Accelerating Computational Materials Discovery with Machine Learn ing and Cloud High - Performance Computing: from Large-Scale Screening to Experimental Validation,

    C. Chen et al., "Accelerating Computational Materials Discovery with Machine Learn ing and Cloud High - Performance Computing: from Large-Scale Screening to Experimental Validation," Journal of the American Chemical Society, vol. 146, no. 29, pp. 20009-20018, 2024/07/24 2024, ...

    2024 doi

Pith tools

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