Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Application of pretrained universal machine-learning interatomic potential for physicochemical simulation of liquid electrolytes in Li-ion battery

T0 review · 2 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read SevenNet-0, trained mostly on inorganic crystals, predicts solvation and ion transport in battery electrolytes with a systematic density bias that a few hours of fine-tuning removes.

desk verdict A careful, honest benchmark of SevenNet-0 for liquid electrolytes; the density-error diagnosis is credible but hinges on an unvalidated cross-code D3 equivalence, and the abstract overclaims fine-tuning benefits beyond the demonstrated subset. read the letter →

arxiv 2501.05211 v1 pith:7OZJ7HZD submitted 2025-01-09 cond-mat.mtrl-sci

Suyeon Ju , Jinmu You , Gijin Kim , Yutack Park , Hyungmin An , Seungwu Han This is my paper · ORCID
classification cond-mat.mtrl-sci
keywords machine-learninginteratomicpotentialliquidelectrolytelithium-ionbatterymoleculardynamicsdensitypredictionsolvationstructurefine-tuningSevenNet-0
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 a general-purpose machine-learning interatomic potential trained mostly on inorganic crystals can be used directly to simulate the liquid electrolytes of lithium-ion batteries. The authors test SevenNet-0 on 20 solvents spanning cyclic carbonates, linear carbonates, ethers, and esters, with lithium hexafluorophosphate ($\mathrm{LiPF_6}$) and LiFSI salts, comparing single-molecule geometries, liquid densities, solvation-shell structures, ion dissociation, and diffusivities against DFT, ab initio molecular dynamics, and experiment. They report broad agreement for solvation and transport but a systematic overestimation of solvent density, which they attribute to softened forces and overbinding in the model rather than to the DFT reference. Fine-tuning on a small dimethyl carbonate dataset for only a few hours reduces the stress error from 2.78 to 0.57 kbar and brings linear-carbonate densities close to experiment at a fraction of the cost of building a bespoke potential. The stake is practical: if the claims hold, a public pretrained MLIP plus cheap fine-tuning is a viable screening platform for electrolyte formulations.

What carries the argument

The load-bearing object is SevenNet-0, an equivariant message-passing neural network (NequIP-style) pretrained on the MPtrj trajectory database, whose predictions for energy, forces, and stresses come from the gradient of a learned total energy. To bring the model to the PBE-D3 level used by the DFT references, the authors attach an in-house CUDA implementation of Grimme's D3 dispersion correction with Becke-Johnson damping inside LAMMPS. The performance argument is carried by layered benchmarks—single molecules, dimers, pure solvents, dilute and concentrated electrolytes—and by a fine-tuning protocol that scales lattice parameters by 0.9 and 1.1 to expose the model to volume variations and increases the stress-loss weight so that density-related errors are explicitly trained out.

What would settle it

Run the same NVT pressure test on 30-molecule cells of DFEC, FEC, DMC, and PC at experimental density using SevenNet with an independent, publicly verified D3 implementation instead of the in-house CUDA code; if the negative mean pressures and density overestimates persist, the model-error attribution survives, and if they largely disappear, the correction code was the culprit. A cheaper complementary check is to compare SevenNet liquid densities with the D3 correction switched off, isolating the model's intrinsic overbinding.

Watch

Extended reading notes

Core claim

The paper's central claim is that SevenNet-0, an equivariant graph-neural-network interatomic potential trained predominantly on inorganic crystal data, nevertheless captures the physics that controls liquid electrolytes: Li–O radial and angular distributions in solvation shells match AIMD results, the preference for ethylene carbonate over dimethyl carbonate in the Li solvation shell and the degree of ion dissociation track experiment across solvent compositions, and ion diffusivities fall in line with measurements. The one prominent failure is liquid density, which SevenNet overestimates by 3–7% for cyclic solvents and 9–15% for linear solvents, an error the authors trace to a softened potential-energy surface and too-strong intermolecular binding that produce negative pressures at experimental volumes. A short fine-tuning run on 150 DFT single points for DMC, with the stress loss weight raised from 0.01 to 1.0, repairs most of the density error for linear carbonates. Analysis of the training set supports the interpretation that the model generalizes through latent-space interpolation rather than memorization: of the 20 solvents only DME appears in the training data, yet absent chemical moieties such as fluorinated carbon groups are placed between related trained moieties in descriptor space.

Load-bearing premise

The diagnosis that the density error belongs to the model rather than to the reference calculation assumes that the in-house CUDA D3 dispersion correction used with SevenNet in LAMMPS is numerically identical to VASP's D3 correction with the same parameters; the paper states this consistency but shows no cross-code validation.

Editorial extensions

If this is right

  • SevenNet-0 can be used without retraining to rank solvents and salt-solvent combinations by solvation-shell geometry and ion-dissociation trends, with the caveat that densities must be taken from experiment or corrected.
  • A few hours of fine-tuning on roughly 150 DFT single points, including compressed and expanded volumes, is enough to bring stress prediction from 2.78 kbar to 0.57 kbar MAE and fix linear-carbonate densities, making pretrained-MLIP screening much cheaper than bespoke potential development.
  • Because density errors dominate transport errors, diffusivities computed at the model's own equilibrium density are underestimated by 50–70%; using experimental densities restores good agreement, so density must be treated as a first-order correction in any screening pipeline.
  • The latent-space interpolation result implies that pretrained potentials can be extended to chemistries absent from their training set, such as fluorinated carbonates, by fine-tuning on a small set of targeted molecules rather than retraining from scratch.

Reading between the lines

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

  • A natural test of the paper's diagnosis is whether other inorganic-trained universal potentials show the same pattern of good solvation and transport but overestimated liquid density; if they do, the density bias is a general property of this model class and the fine-tuning recipe may transfer.
  • The few-kbar D3-equivalence assumption could be checked directly by running SevenNet with an independent D3 implementation on the same 30-molecule NVT cells; if pressures shift by more than the reported spread, part of the 'model error' attribution would need to be reassigned to the correction.
  • Because force softening is most pronounced for fluorine-containing local environments, a screening campaign for fluorinated electrolyte solvents should deliberately include fluorinated molecules in the fine-tuning set; otherwise the density improvements demonstrated for DMC may not transfer.
  • The volume-scaling fine-tuning strategy points toward a cheaper recipe than AIMD generation: pretrain on crystals, then add only a few dozen DFT single points per target molecule instead of building a full bespoke training set.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper evaluates the pretrained universal machine-learning interatomic potential SevenNet-0 for simulating liquid electrolytes relevant to Li-ion batteries. It benchmarks single-molecule energies and geometries, pure-solvent densities, solvation shell structures, ion dissociation, and diffusion coefficients for 20 solvents and two salts against experimental data and ab initio MD from other groups. The authors find generally good agreement, with a systematic tendency for SevenNet to overestimate liquid densities. They attribute this overestimation to imperfect stress learning by the model, based on pressure distributions from 30-molecule NVT simulations at experimental densities. They then fine-tune SevenNet on DMC with an increased stress-loss weight, obtaining improved density prediction for linear carbonates at modest computational cost. The paper also analyzes the MPtrj training set and latent-space structure to argue that the model generalizes across chemical space rather than memorizing specific configurations.

Significance. If the conclusions hold, the paper provides a valuable systematic benchmark of a state-of-the-art pretrained MLIP for a strongly out-of-distribution class of systems, and demonstrates that targeted fine-tuning can repair systematic errors with only a few hours of compute. The use of external experimental and AIMD references, rather than in-house fitted data, lends credibility to the benchmarks. The training-set coverage analysis and latent-space interpolation evidence are informative for understanding MLIP generalization. However, the central causal diagnosis of the density error depends on a cross-code D3 equivalence that is not demonstrated, and the fine-tuning transferability is limited to the trained solvent class, so the strength of the claims is contingent on these points.

major comments (2)
  1. [Section 2.2 and Section 4.2] The attribution of SevenNet's density overestimation to 'imperfect learning of stress' rests on the pressure comparison in Fig. 3b, in which SevenNet pressures are computed in LAMMPS with an in-house CUDA implementation of Grimme's D3 correction, while the DFT reference pressures use VASP's D3. Section 4.2 states that the two implementations are identical, but no cross-code validation is provided. Because the D3 correction contributes to the virial, any discrepancy between the implementations—for example, in periodic boundary handling or in the stress derivative—would shift the SevenNet pressures by an amount that could be comparable to the observed few-kbar differences, especially given the short 15-ps runs on 30-molecule cells. The paper should provide a quantitative comparison of energies, forces, and stresses from both codes on a representative set of configurations, or alternatively present a sensitivity analysis (e.g., remove the D3 correction from both sides) to show that the pressure difference persists. Without this, the conclusion that the density error is a model-intrinsic stress-learning defect is not established, and the fine-tuning motivation is weakened.
  2. [Section 2.2, Fig. 3b] The pressure distributions in Fig. 3b are based on a single trajectory of 30 molecules with 15 ps of production sampling per system. The text reports 1500 instantaneous pressure samples from this single run, but no error bars or block-averaging estimates are given for the mean pressures, and no independent simulations are presented. Given that the density errors are on the order of several percent and the inferred pressure differences are a few kbar, the statistical significance of the differences is unclear. The authors should report uncertainty estimates (e.g., standard errors from block averages) and, if feasible, a larger-cell check to rule out finite-size effects. This is directly relevant to the strength of the diagnostic underlying the main negative finding.
minor comments (6)
  1. [Abstract and Section 2.5] The abstract and conclusion claim that fine-tuning 'improved accuracy' without qualification; the results in Fig. 3a and the text show that SevenNet-FT improves densities for linear carbonates but underestimates densities of cyclic carbonates and fluorinated solvents. Please qualify the claim to reflect the demonstrated scope.
  2. [Section 2.2, near Fig. 3b] The text reads 'The temperatures were set to 298 K (DEFC, DMC, and PC)' but the compound is DFEC (difluoroethylene carbonate); please correct the typo.
  3. [Fig. 3a and Table S5] The computed densities in Fig. 3a are plotted as single points without error bars; please add error bars from the NPT trajectory or state the standard deviation of the density average.
  4. [Section 2.2 and Section 4.2] The use of a tritium mass (3 a.u.) for hydrogen atoms in all MD simulations is a notable approximation; the paper does not comment on its potential effect on the computed diffusivities and densities. Please add a brief discussion or a test showing that the reported transport properties are insensitive to this mass choice.
  5. [Section 2.4] The sentence 'the model learned large parts of the PES by generalizing across the chemical space, facilitated by deep learning and learnable atomic embeddings' is somewhat vague; the supporting evidence from PCA/UMAP is qualitative and could be described more precisely.
  6. [Section 4.2] The in-house CUDA D3 implementation is not made available; including it in the ESI or a public repository would facilitate reproducibility and independent validation of the equivalence claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper benchmarks an externally pretrained MLIP against experimental and independent AIMD data, and its fine-tuning step is a genuine transfer test.

full rationale

The paper's central claims are empirical evaluations of SevenNet-0 on liquid electrolytes, benchmarked against experimental densities and diffusivities and against AIMD results from other groups (refs 28 and 29). The fine-tuning procedure uses DFT single-point energies, forces, and stresses on 150 DMC-derived structures (50 snapshots plus lattice scalings of 0.9 and 1.1), not the target experimental densities or diffusivities, and the fine-tuned model is then tested on other solvents and compositions. This is a supervised transfer-learning test rather than a fitted parameter renamed as a prediction. Citations to the authors' own prior development of SevenNet-0 (refs 56 and 69) describe the model and its training data, but they do not supply the electrolyte properties that are compared with experiment, so the self-citations are not load-bearing in a circular sense. The assumption that the in-house CUDA D3 correction in LAMMPS is numerically equivalent to VASP's D3 is an implementation-equivalence assumption used to attribute density errors to the model; if that assumption failed, the diagnostic would be weakened, but the argument would not reduce to its own inputs by definition. No equation in the paper equates a predicted quantity with a fitted quantity, and no external benchmark is constructed from the model's own outputs. Hence the derivation chain is self-contained with respect to the claims actually made, and the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central benchmark adds no free parameters to SevenNet-0 itself; the listed items are hand-chosen simulation and fine-tuning settings that affect the quantitative results. The main load-bearing assumptions are reference-level (PBE-D3 accuracy and D3 cross-code consistency) and ergodicity assumptions for the MD runs.

free parameters (4)
  • Tritium hydrogen mass (3 a.u.) = 3 a.u.
    Hydrogen atoms were assigned a mass of 3 a.u. to allow a 2 fs timestep and stable NPT simulations (Section 2.2). This choice modifies molecular mass and vibrational dynamics, and the reported diffusivities are compared with experiment without a mass correction.
  • Stress loss weight in fine-tuning = 1.0 (raised from 0.01)
    The stress loss weight was increased from 0.01 to 1.0 during fine-tuning (Section 2.5) to improve density; the resulting SevenNet-FT density accuracy depends on this hand-chosen value.
  • Fine-tuning learning rate schedule = 10^-4 to 10^-6 over 600 epochs
    Chosen for fine-tuning (Section 2.5); no ablation is reported, so the outcome depends on this schedule.
  • Number of fine-tuning structures and scaling factors = 150 structures; lattice scaled by 0.9 and 1.1
    The training set for fine-tuning is 50 snapshots plus two scaled copies (Section 2.5); this choice determines the improvement shown for DMC and the degradation elsewhere.
assumptions (5)
  • domain assumption PBE-D3 is an adequate reference for liquid electrolyte energetics, pressures, and solvation.
    Invoked throughout Sections 2.1-2.3 and 4.1; the paper cites PBE0-D3 agreement for dimers and organic crystal densities, but the density-error attribution relies on PBE-D3 alone.
  • domain assumption MPtrj-trained SevenNet-0, with added D3, can generalize to out-of-distribution organic liquids.
    The transferability claim in Sections 2.4-5 assumes local chemical moieties (C=O, C-O, C-H, Li-O) are sufficiently represented in the inorganic training set for interpolation; the paper provides supporting latent-space analysis but not a proof.
  • domain assumption 1 ns NPT simulations of ~1000 atoms reach equilibrium density, and 1-7 ns production runs yield converged diffusivities.
    No convergence tests versus simulation length or system size are provided (Sections 2.2, 2.3.3); the tritium mass trick and the 3H mass affect the dynamics.
  • ad hoc to paper The in-house CUDA D3 implementation is numerically equivalent to VASP's D3 with the same parameters.
    Stated in Section 4.2 without a direct validation of the implementation; the pressure comparison in Fig. 3b depends on this equivalence.
  • domain assumption Dispersion interactions were excluded in the dilute solvation benchmark to match ref. 28.
    Section 2.3.1 excludes dispersion for the solvation-shell comparison with AIMD from ref. 28, while density and transport runs include D3; this inconsistency could affect the consistency of the model evaluation across properties.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Application of pretrained universal machine-learning interatomic potential for physicochemical simulation of liquid electrolytes in Li-ion battery." pith.science (2026). https://pith.science/paper/7OZJ7HZD

@misc{pith2026250105211,
  author       = {Pith},
  title        = {Pith review of: Application of pretrained universal machine-learning interatomic potential for physicochemical simulation of liquid electrolytes in Li-ion battery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7OZJ7HZD}},
  note         = {Machine review of arXiv:2501.05211}
}
abstract

Achieving higher operational voltages, faster charging, and broader temperature ranges for Li-ion batteries necessitates advancements in electrolyte engineering. However, the complexity of optimizing combinations of solvents, salts, and additives has limited the effectiveness of both experimental and computational screening methods for liquid electrolytes. Recently, pretrained universal machine-learning interatomic potentials (MLIPs) have emerged as promising tools for computational exploration of complex chemical spaces with high accuracy and efficiency. In this study, we evaluated the performance of the state-of-the-art equivariant pretrained MLIP, SevenNet-0, in predicting key properties of liquid electrolytes, including solvation behavior, density, and ion transport. To assess its suitability for extensive material screening, we considered a dataset comprising 20 solvents. Although SevenNet-0 was predominantly trained on inorganic compounds, its predictions for the properties of liquid electrolytes showed good agreement with experimental and $\textit{ab initio}$ data. However, systematic errors were identified, particularly in the predicted density of liquid electrolytes. To address this limitation, we fine-tuned SevenNet-0, achieving improved accuracy at a significantly reduced computational cost compared to developing bespoke models. Analysis of the training set suggested that the model achieved its accuracy by generalizing across the chemical space rather than memorizing specific configurations. This work highlights the potential of SevenNet-0 as a powerful tool for future engineering of liquid electrolyte systems.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

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

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

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

Reference graph

Works this paper leans on

12 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Manthiram, ACS Cent

    1 A. Manthiram, ACS Cent. Sci. , 2017, 3, 1063–1069. 2 Z. Zhu, T. Jiang, M. Ali, Y. Meng, Y. Jin, Y. Cui and W. Chen, Chem. Rev., 2022, 122, 16610–16751. 3 Y. S. Meng, V. Srinivasan and K. Xu, Science, 2022, 378, eabq3750. 4 Z. Yu, H. Wang, X. Kong, W. Huang, Y. Tsao, D. G. Mackanic, K. Wang, X. Wang, W. Huang, S. Choudhury , Y. Zheng, C. V. Amanchukwu, S...

    work page 2017

  2. [11]

    109 J. Kim, J. Kim, J. Kim, J. Lee, Y. Park, Y. Kang and S. Han, J. Am. Chem. Soc. , 2024, DOI: 10.1021/jacs.4c14455. 110 G. Kresse and D. Joubert, Phys. Rev. B, 1999, 59, 1758–1775. 111 J. P. Perdew, K. Burke and M. Ernzerhof, Phys. Rev. Lett. , 1996, 77, 3865–3868. 112 S. Grimme, J. Antony , S. Ehrlich and H. Krieg,J. Chem. Phys., 2010, 132, 154104. 113...

    doi:10.1021/jacs.4c14455 2024

  3. [27]

    97 C. M. Burba and R. Frech, J. Phys. Chem. B , 2005, 109, 15161–15164. 98 K. Hayamizu, J. Chem. Eng. Data , 2012, 57, 2012–2017. 99 S. Miyoshi, H. Nagano, T. Fukuda, T. Kurihara, M. Watan- abe, S. Ida and T. Ishihara, J. Electrochem. Soc. , 2016, 163, A1206. 100 M. Takeuchi, Y. Kameda, Y. Umebayashi, S. Ogawa, T. Son- oda, S. ichi Ishiguro, M. Fujita and...

    work page 2005

  4. [146]

    A predictive machine learning force field framework for liquid electrolyte development

    47 F. Wang and J. Cheng, Chin. J. Struct. Chem. , 2023, 42, 100061. 48 D. Zhu, L. Sheng, T. Hu, S. Chen, M. Shi, H. Hua, K. Yang, J. Wang, Y. Tang, X. He and H. Xu,J. Phys. Chem. Lett., 2024, 15, 4024–4030. 49 O. Shayestehpour and S. Zahn, J. Chem. Phys. , 2024, 161, 134505. 50 S. Gong, Y. Zhang, Z. Mu, Z. Pu, H. Wang, Z. Yu, M. Chen, T. Zheng, Z. Wang, L...

    work page Pith review arXiv 2023

  5. [1968]

    9 R. Fong, U. Von Sacken and J. R. Dahn, J. Electrochem. Soc. , 1990, 137,

    work page 1990

  6. [2009]

    Xu, Chem

    10 K. Xu, Chem. Rev., 2004, 104, 4303–4418. 11 J. Tarascon and D. Guyomard, Solid State Ionics , 1994, 69, 293–305. 12 S. Chen, F. Dai and M. Cai, ACS Energy Lett. , 2020, 5, 3140–

    work page 2004

  7. [2017]

    Miyoshi, H

    105 S. Miyoshi, H. Nagano, T. Fukuda, T. Kurihara, M. Watan- abe, S. Ida and T. Ishihara, J. Electrochem. Soc. , 2016, 163, A1206. 106 L. McInnes, J. Healy and J. Melville, arXiv, 2020, preprint, arXiv:1802.03426, https://arxiv.org/abs/1802.03426. 107 SevenNet-UMAP, https://doi.org/10.5281/zenodo. 14613413, (accessed Jan 2025). 108 P. Eastman, P. K. Behar...

    arXiv 2016

  8. [2453]

    Batatia, D

    64 I. Batatia, D. P. Kovács, G. N. C. Simm, C. Ortner and G. Csányi, arXiv, 2023, preprint, arXiv:2206.07697, https://arxiv.org/abs/2206.07697. 65 J. Kim, Y. Park, S. Hwang and S. Han, Modell. Simul. Mater . Sci. Eng., 2024, DOI: 10.1088/1361-651X/ad9d63, (Section 11 – Pretrained universal machine learning potential). 66 A. Jain, S. P. Ong, G. Hautier, W....

    arXiv 2023

Show all 12 references
  1. [2991]

    Merchant, S

    54 A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon and E. D. Cubuk, Nature, 2023, 624, 80–85. 55 I. Batatia, P. Benner, Y. Chiang, A. M. Elena, D. P. Kovács, J. Riebesell, X. R. Advincula, M. Asta, M. Avaylon, W. J. Baldwin, F. Berger, N. Bernstein, A. Bhowmik, S...

    2023 arXiv

  2. [3151]

    Masias, J

    13 A. Masias, J. Marcicki and W. A. Paxton, ACS Energy Lett. , 2021, 6, 621–630. 14 J. Neubauer, A. Pesaran, C. Bae, R. Elder and B. Cunning- ham, J. Power Sources, 2014, 271, 614–621. 15 K. Guo, S. Qi, H. Wang, J. Huang, M. Wu, Y. Yang, X. Li, Y. Ren and J. Ma, Small Sci., 20...

    2021

  3. [3441]

    35 M. I. Chaudhari, J. R. Nair, L. R. Pratt, F. A. Soto, P. B. Bal- buena and S. B. Rempe, J. Chem. Theory Comput. , 2016, 12, 5709–5718. 36 J. Self, K. D. Fong and K. A. Persson, ACS Energy Lett. , 2019, 4, 2843–2849. 37 Z. Luo, S. A. Burrows, S. K. Smoukov, X. Fan and E. S. ...

    2016

  4. [4074]

    75 R. L. Hurle and L. A. Woolf, J. Chem. Soc., Faraday Trans. 1 , 1982, 78, 2233–2238. 76 Molview (v2.4), https://molview.org, (accessed August 2024). 77 L. Martínez, R. Andrade, E. G. Birgin and J. M. Martínez, J. Comput. Chem., 2009, 30, 2157–2164. 78 M. Mantina, A. C. Chamb...

    1982

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.