Predictive Simulation of Interphases on Li Metal Surface

Daniel C. Hannah; Jingxuan Ding; Kang Xu; Qichao Hu; Xinyu Li; Yumin Zhang

REVIEW 4 major objections 4 minor 2 references

Predictive Simulation of Interphases on Li Metal Surface

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that pairing a polarizable force field with a universal machine-learning force field predicts the solid–electrolyte interphase chemistry on lithium metal from electrolyte composition alone.

desk verdict A credible first pass at universal-MLFF SEI simulation that reproduces known trends, but the universality claim outruns the evidence; send to review with a demand for reactivity validation and uncertainty quantification. read the letter →

arxiv 2608.09791 v1 pith:FCQ6FF3G submitted 2026-08-10 cond-mat.mtrl-sci

Xinyu Li , Jingxuan Ding , Daniel C. Hannah , Yumin Zhang , Qichao Hu , Kang Xu This is my paper · ORCID

classification cond-mat.mtrl-sci

keywords solidelectrolyteinterphaselithiummetalanodemachinelearningforcefieldreactivemoleculardynamicsdecompositionLiF-richsimulationbatterydesign

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 claims that the chemistry of the solid–electrolyte interphase that forms on lithium metal can be predicted from the electrolyte composition alone, using atomistic simulation instead of trial-and-error experiments. It combines a transferable polarizable force field with a universal machine-learning force field and runs reactive molecular dynamics at a lithium–electrolyte interface for four chemically distinct electrolyte families. The simulations reproduce the established experimental pattern: fluorinated solvents such as FSA and F5DEE yield LiF-rich interphases, while carbonate-based electrolytes yield organic-rich interphases. If the claim holds, electrolyte design for lithium-metal batteries could be screened computationally before any cell is built.

What carries the argument

The load-bearing instrument is a reactive molecular-dynamics protocol in which a graph-neural-network polarizable force field equilibrates the bulk electrolyte, and a universal machine-learning interatomic potential trained on solid–liquid interface configurations propagates the interfacial dynamics without predefined reaction templates. A 10x10 lithium slab with a fixed bottom layer is placed against the equilibrated electrolyte, with a harmonic wall preventing electrolyte atoms from escaping into vacuum; 200 ps trajectories are then analyzed by counting reacted molecules, locating LiF along the interface, and clustering decomposition products. Because bond breaking and bond formation emerge from the learned potential, the same protocol can be applied to an arbitrary electrolyte without reparametrizing a reactive force field for each chemistry.

What would settle it

Take one electrolyte outside the force field's training distribution, run the same 200 ps interfacial protocol, and compare the predicted interphase composition and reaction products with depth-profiled X-ray photoelectron spectroscopy or cryogenic electron microscopy of a cell cycled under otherwise identical conditions. If the observed LiF fraction or organic-to-inorganic ratio disagrees with the simulation, the claim of universal predictive power fails.

Watch

Extended reading notes

Core claim

The central claim is that spontaneous interfacial decomposition can be captured by combining a polarizable force field for realistic bulk electrolyte packing with a universal machine-learning force field for reactive dynamics, requiring no chemistry-specific reactive potential. Across six formulations, the simulations generate distinct interphase morphologies and compositions in 200 ps of dynamics. The computed LiF fractions and thicknesses track the known experimental ordering: F5DEE gives the highest LiF fraction, FSA electrolytes form thick inorganic-rich interphases, and the carbonate system gives the thinnest, most organic-rich interphase. The simulation also resolves molecular pathways, including stepwise defluorination of F5DEE and dominance of FSA as a fluorine donor, and shows that higher salt concentration thickens the SEI by increasing anion participation while leaving LiF density nearly unchanged.

Load-bearing premise

The method stands or falls on the assumption that the machine-learning force field correctly describes bonds breaking and forming at the lithium–electrolyte interface for these particular molecules, even though it was not trained specifically on these reaction pathways, and the paper does not validate that assumption against quantum-chemistry calculations for the claimed reactions.

Editorial extensions

If this is right

If the protocol is correct, the solid–electrolyte interphase from a new electrolyte can be predicted by simulation before any electrochemical cell is assembled.
The method yields a mechanistic explanation for why fluorinated solvents, and higher salt concentrations, favor inorganic LiF-rich interphases rather than organic decomposition products.
The same simulation protocol could be applied to other electrode surfaces and interphase-forming processes without changing the force fields.
The computed reaction counts and LiF distributions provide direct molecular-level input for AI-driven electrolyte screening pipelines.

Reading between the lines

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

I infer that the authors intend this two-force-field protocol, not the specific 200 ps trajectories, as the transferable contribution; the quantitative SEI thicknesses should be read as trends rather than converged values.
An extension the paper leaves implicit is that the same combination could be pointed at other reactive interfaces, such as catalytic or electrosynthetic surfaces, by swapping the electrode slab and the electrolyte.
A testable consequence not stated in the paper is that repeating the simulation with different initial packings or longer trajectories should preserve the compositional ordering of the six electrolytes, even if absolute thicknesses drift.

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.

Desk Editor's Note

read the letter

Here is my read. The paper does something genuinely new: it pairs two general-purpose models, ByteFF-Pol for the liquid and UMA-s-1p2 for the reactive interface, and runs a single protocol across chemically different electrolytes to see what kind of SEI forms spontaneously. Reproducing the known LiF-rich versus organic-rich trend across FSA, F5DEE, carbonate, and SFL/TOL is real evidence that the approach has signal, not just noise. The reaction-pathway analysis (all five F on F5DEE ending as LiF, TOL staying inert) is chemically sensible and consistent with prior work. The protocol is described carefully enough to reproduce.

The soft spot is the scope of the claim. 'Universally predict' is supported by six formulations, one 200 ps trajectory per system, and no error bars or repeated seeds. The MLFF's ability to describe spontaneous bond-breaking at the Li interface for chemistries beyond its training distribution is assumed, not validated. There is no check against quantum chemistry for the specific pathways reported. That is the load-bearing assumption. Without it, readers cannot tell whether the LiF is real chemistry or a force-field artifact. No code or data accompanies the paper, which slows independent verification. These are fixable: run multiple seeds, compare selected pathways with QM/MM or DFTB, and test whether another universal potential gives the same qualitative ranking.

The stress-test note matches my view. The central trend is plausible; the universality claim is unproven. I would also note the SEI thickness values are likely sensitive to the 200 ps cutoff, since growth has not visibly plateaued for all systems. That does not destroy the qualitative ranking, but it does undercut quantitative statements.

Who this is for: people working on electrolyte screening and MLFF applications to battery interfaces. It is a useful proof-of-concept, not yet a validated predictive tool. I would send it to a serious referee because it deserves expert time, but the revision should require reactivity validation and uncertainty quantification. I would not cite it as evidence of universal prediction, though I might cite it as an early example of combined universal force fields for SEI simulation. Reading group: maybe, as a discussion of what counts as prediction in ML-driven simulation.

Referee Report

4 major / 4 minor

Summary. The manuscript presents a molecular dynamics protocol that combines a commercially developed polarizable force field (ByteFF-Pol) for bulk electrolyte equilibration with the UMA-s-1p2 universal machine-learned interatomic potential for reactive interfacial dynamics. Six electrolyte compositions are simulated on a Li(001) slab for 200 ps each, and the authors report SEI thickness, molecular composition, LiF density, and dominant reaction pathways for LiFSI/FSA, LiFSI/F5DEE, LiFSI/SFL/TOL, and LiFSI/EC/DMC/FEC. The central claim is that the protocol reproduces the qualitative experimental trend that fluorinated solvents form LiF-rich interphases whereas conventional carbonate solvents form organic-rich interphases, establishing a route toward universal predictive SEI simulation.

Significance. If the claims hold, the protocol would be a valuable high-throughput screening tool for electrolyte engineering, because it uses pre-trained universal potentials that are independent of the target SEI outcome, avoiding the circularity often present in chemistry-specific fitted potentials. The qualitative reproduction of the LiF-rich versus organic-rich trend across four chemically distinct electrolyte classes is encouraging, and the explicit description of the construction, equilibration, and production protocols supports reproducibility. The main significance is conditional, however: the reliability of the reactive simulation rests entirely on the extrapolative accuracy of UMA-s-1p2 for bond-breaking and bond-forming at the Li/electrolyte interface, and the quantitative SEI properties are reported without uncertainty estimates. The paper would be substantially strengthened by adding direct validation of representative reaction pathways against quantum chemistry and by reporting multiple independent trajectories.

major comments (4)

[Methods (Interfacial molecular dynamics) and Results (Molecular reaction pathways)] The load-bearing assumption that UMA-s-1p2 correctly describes spontaneous bond-breaking and bond-forming at the Li/electrolyte interface is not validated. The manuscript states that universal MLFFs 'maintain quantum level accuracy' and that OC25 was selected because it includes solid-liquid configurations, but no evidence is provided that the model is accurate for reaction pathways and barrier heights of the specific molecules studied (FSA, F5DEE, SFL, EC, DMC, FEC). Because the entire central claim depends on reactions emerging from the MLFF, the authors should add a direct test: compare the predicted dominant pathways (e.g., F5DEE defluorination order, FSA C-N bond stability, EC ring-opening) against DFT or QM/MM calculations on small model systems, or at least perform sensitivity tests with an alternative universal MLFF (e.g., MACE-MP or CHGNet) to see whether the same qualitative trend is reproduced. Without such a check, the simulated SEI compositions could reflect force-field artifacts rather than chemistry.
[Results, Table 4 and Figure 2] All SEI thicknesses, mole fractions, and LiF densities are derived from a single 200 ps trajectory per composition, with no independent seeds, no block averaging, and no reported statistical uncertainty. The paper makes quantitative claims such as 'thickness increases from 23.39 Å at 1:10 to 28.33 Å at 1:6 and 30.85 Å at 1:3.5' and compares values like 7.63, 9.77, and 9.35 Å across electrolytes; these differences may be within noise given the small system size and short timescale. The authors should provide error bars from multiple independent trajectories or time-block decomposition, and should state whether the observed differences are statistically significant. This is essential to support the 'systematic increase' and 'lowest/highest' statements that the paper makes.
[Results (Molecular reaction pathways, Figure 4)] The reaction mechanisms in Figure 4 are inferred from a single reactive trajectory, and the text presents them as the predominant pathways (e.g., 'the C–F bonds at the –CF2 end undergo reduction first, followed by those at the –CF3 end' for F5DEE; 'no C-N bond breaking is observed' for FSA). The generalizability of these mechanistic claims is unclear without either multiple independent trajectories showing reproducibility or quantum-chemical calculation of the relevant barriers. The paper should at minimum state the number of times each pathway was observed across the trajectory and how the 'predominant' pathway was selected, and ideally confirm the key pathways against DFT.
[Abstract and Conclusion] The abstract claims that the combined force fields allow the authors to 'simulate interphasial chemistry across chemically diverse electrolyte formulations' and describes a 'route toward universal and high-throughput predictive simulation of interphases.' The conclusion uses 'could lead to a universal tool,' which is more measured. Given that only six compositions are tested, all of them LiFSI salts and all lithium-metal anodes, the evidence is too narrow to support the adjective 'universal' as applied to the simulation method itself. The authors should either expand the validation to include substantially different chemistries (e.g., different salts, anode materials, or polymer electrolytes) or explicitly temper the abstract's universal claim to reflect the demonstrated domain, so that the title and abstract do not overstate the current scope.

minor comments (4)

[Figure 3 caption] The caption lists panels (a), (b), (c), (e), (f), and (g) but omits panel (d) and then describes 'representative structures' in panel (d) and (h); the panel labels in the caption and the figure should be reconciled.
[Table 4] The header 'Molecular ratio of reacted molecules' is used inconsistently: the first column of Table 4 lists the electrolyte name, then a ratio expression such as 'FSI:EC:DMC:FEC = 1:1.50:0.58:0.58'; the formatting should be unified and the units clarified, because the column mixes a ratio and a count-like quantity.
[Introduction] The abbreviation 'UMA' is used in the Introduction but the full model name 'UMA-s-1p2' and its training data (OC25) are first defined only in Methods; defining it at first use would make the paper easier for a nonspecialist to follow.
[Methods (Electrolyte construction)] The paper states that bulk equilibration used 'constrained anisotropic isothermal-isobaric ensemble (NPT)' with lateral cell vectors fixed and the z length fluctuating; the text also says 'Monte Carlo anisotropic barostat at 1 atm, 298 K.' The description is understandable but the fixed-lateral-cell variant of NPT is unusual and should be explicitly justified as appropriate for later matching to the Li slab, which is fixed in x and y.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified: the SEI trend is an emergent MD result from independently trained universal force fields, compared against external experiment.

full rationale

The paper's derivation chain is simulation-based rather than definitional. The central claim—that fluorinated solvents tend to produce LiF-rich interphases whereas carbonate-based electrolytes produce more organic-rich interphases—is not an input to the force fields. BytesFF-Pol and UMA-s-1p2 are pre-trained on quantum-mechanical data external to this work, and the SEI compositions, thicknesses, and reaction pathways are obtained from molecular dynamics trajectories without fitting any parameter to the experimental LiF/organic trend. The comparisons are made to published experimental observations, which is the appropriate external check. The self-citations present (e.g., Kang Xu's electrolyte reviews, and one author's prior arXiv paper on solid-state interface evolution) are contextual or technical and are not load-bearing for the central claim. The boundary wall is adapted from a cited prior simulation study, but it is a minor technical choice and does not encode the outcome. The only substantive concern is that the universal MLFF's training data may overlap chemically with the simulated Li/electrolyte systems, which would weaken the 'universal' extrapolation claim. However, training-set overlap is a validity and generalization risk, not a circularity of derivation: it does not make the reported SEI chemistry an identity with the model's parameters. No equation in the paper reduces to its own input, and no fitted quantity is renamed as a prediction. Therefore no circular step can be exhibited with the required specificity, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The computational protocol introduces no new physical entities. The load-bearing assumptions are the transferability and accuracy of the two universal force fields, the sufficiency of the 200 ps timescale, and the fidelity of the interface model.

free parameters (3)

Harmonic wall spring constant k = 1.0 eV Å^-2
Chosen by hand for the vacuum boundary in the interfacial cells (Methods, paragraph 'The harmonic wall acted independently...'). Controls the wall potential but is unlikely to affect the chemistry near the Li surface.
Simulation duration = 200 ps
Hand-set for all production runs (Methods, 'Molecular dynamics'). SEI thickness and composition are time-dependent outputs; the paper argues passivation plateaus justify this choice, but no convergence test is shown.
Thermostat friction coefficient = 0.1 ps^-1
Chosen Langevin friction for the locally thermostatted region; inherits the standard choice from related SEI-MLFF studies.

assumptions (5)

domain assumption UMA-s-1p2 accurately models reactive chemistry (bond breaking and formation) at Li/electrolyte interfaces.
The entire reactive MD relies on this MLFF (Methods: 'Molecular dynamics'; Results section). No DFT validation of the reaction pathways is provided.
domain assumption ByteFF-Pol accurately reproduces bulk electrolyte densities and structures for the six electrolytes.
Bulk equilibration and the initial liquid structures use ByteFF-Pol (Methods: 'Electrolyte construction and equilibration'; Table 2).
domain assumption 200 ps trajectories are sufficient to capture passivation and representative SEI composition.
The plateau in reacted-molecule counts (Figure 5) is cited as evidence, but the thickness continues evolving and no longer-timescale test is shown.
domain assumption The locally thermostatted fixed-cell NVT protocol and the harmonic wall do not perturb interfacial reaction kinetics.
The protocol is explicitly described as 'not an exact global canonical sampler' (Methods: 'Molecular dynamics').
domain assumption The BCC Li(001) slab with 27 layers represents the Li metal anode in these systems.
Used for all simulations (Methods: 'Lithium metal slab and interface construction'); no testing of other surfaces or defects.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Predictive Simulation of Interphases on Li Metal Surface." pith.science (2026). https://pith.science/paper/FCQ6FF3G

@misc{pith2026260809791,
  author       = {Pith},
  title        = {Pith review of: Predictive Simulation of Interphases on Li Metal Surface},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCQ6FF3G}},
  note         = {Machine review of arXiv:2608.09791}
}

read the original abstract

Interphases remain the least understood components in advanced batteries. Although their properties dictate whether a new battery chemistry could perform as designed, there has never been a reliable way to predict what an interphase could arise from a new electrolyte system due to the absence of atomistic level knowledge about interphasial formation process. In this work, we attempt to develop a simulation method that can universally predict interphasial chemistries formed on Li metal surface, so that the electrolyte engineering would no longer need lengthy Edisonian approaches. By combining a transferable universal polarizable force field and a universal machine learning force field, we simulate interphasial chemistry across chemically diverse electrolyte formulations, and successfully replicate the experimental observation that fluorinated solvents promote the formation of LiF-rich interphases, whereas interphases of more organic origin arise from conventional carbonate-based electrolytes. By directly capturing these spontaneous interfacial reactions behind these interphasial chemistries, our simulations establish molecular-level relationships between electrolyte chemistry, salt concentration, decomposition pathways, and SEI properties, and opens a route toward universal and high-throughput predictive simulation of interphases that is the foundation for AI-driven electrolyte discoveries.

Figures

Figures reproduced from arXiv: 2608.09791 by the authors.

**Figure 2.** Composition-dependent LiF formation at Li/electrolyte interfaces. Final configurations from molecular dynamics at 200 ps for six electrolyte formulations: (a) LiFSI/FSA 1:3.5, (b) LiFSI/FSA 1:6, (c) LiFSI/FSA 1:10, (d) LiFSI/F5DEE, (e) LiFSI/EC/DMC/FEC, and (f) LiFSI/SFL/TOL. The adjacent horizontal histograms show the distribution of LiF counts along the interfacial direction, binned between 30 and 60 Å. The Li met… view at source ↗

**Figure 3.** Compositions and component structures of SEI. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Reaction mechanisms of electrolyte solvents in the formation of SEI. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Time evolution of the number of reacted anion [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[16]

UMA: A family of universal models for atoms

Wood, Brandon, et al. "UMA: A family of universal models for atoms." Advances in Neural Information Processing Systems 38 (2026): 129391-129427. [17] Zeni, Claudio, et al. "A generative model for inorganic materials design." Nature 639.8055 (2025): 624-632. [18] Deng, Bowen, et al. "CHGNet as a pretrained universal neural network potential for charge-info...

work page 2026
[31]

Synchronized breathing in anion-derived interphases

Zehao Cui, Zhiao Yu, Hao Lyu, Zhenan Bao, Arumugam Manthiram; Resolving Electrolyte Decomposition Products in Gas, Liquid, and Solid Phases in Lithium–Metal Batteries. ACS Energy Lett. 8 August 2025; 10 (8): 3827–3833. [32] Tan, Sha, et al. "Synchronized breathing in anion-derived interphases." ACS Energy Letters 10.8 (2025): 3746-3754

work page 2025

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [16]

UMA: A family of universal models for atoms

Wood, Brandon, et al. "UMA: A family of universal models for atoms." Advances in Neural Information Processing Systems 38 (2026): 129391-129427. [17] Zeni, Claudio, et al. "A generative model for inorganic materials design." Nature 639.8055 (2025): 624-632. [18] Deng, Bowen, et al. "CHGNet as a pretrained universal neural network potential for charge-info...

work page 2026

[2] [31]

Synchronized breathing in anion-derived interphases

Zehao Cui, Zhiao Yu, Hao Lyu, Zhenan Bao, Arumugam Manthiram; Resolving Electrolyte Decomposition Products in Gas, Liquid, and Solid Phases in Lithium–Metal Batteries. ACS Energy Lett. 8 August 2025; 10 (8): 3827–3833. [32] Tan, Sha, et al. "Synchronized breathing in anion-derived interphases." ACS Energy Letters 10.8 (2025): 3746-3754

work page 2025