pith. sign in

arxiv: 2604.07077 · v1 · submitted 2026-04-08 · ❄️ cond-mat.mtrl-sci

Unveiling Mechanisms of SEI Formation and Sodium Loss in Sodium Batteries via Interface Reactor Sampling

Zhoulin Liu , Ziliang Wang , Zherui Chen , Jianchun Sha , Fengzijun Pan , Pingyang Zhang , Yinghe Zhang This is my paper

Pith reviewed 2026-05-10 17:18 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci
keywords SEI formationsodium-metal batteriesmachine learning potentialsinterface simulationselectrolyte solventsNaF crystallizationsodium ion storagemetadynamics
0
0 comments X

The pith

A new sampling method for machine learning potentials reveals that carbonate electrolytes form mixed organic-inorganic SEI layers while ether electrolytes create dense NaF barriers in sodium batteries.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that early-stage solid electrolyte interphase formation proceeds by two distinct routes depending on the electrolyte solvent. Carbonate-based systems produce heterogeneous organic-inorganic matrices through simultaneous deposition of both components, whereas ether-based systems produce compact inorganic films by surface-energy-controlled crystallization of NaF. These structural differences directly control sodium-ion behavior at the interface: NaF-rich barriers permit reversible metallic deposition, while carbonate-rich layers trap sodium ions and drive ongoing electrolyte breakdown. The authors introduce an Interface Reactor sampling strategy to train a charge-aware neuroevolution potential that remains stable over hundred-nanosecond timescales, enabling the atomic-resolution simulations that expose these mechanisms.

Core claim

The Interface Reactor sampling strategy constructs a charge-aware neuroevolution potential that supports stable molecular dynamics of electrode-electrolyte interfaces. Simulations performed with this potential show that carbonate electrolytes generate heterogeneous organic-inorganic SEI matrices via mixed co-formation, while ether electrolytes generate dense, self-limiting inorganic barriers via surface-energy-controlled NaF crystallization. Metadynamics further demonstrates that NaF-rich SEIs promote efficient sodium deposition whereas carbonate-dominated interphases cause irreversible sodium trapping and continuous electrolyte decomposition.

What carries the argument

The Interface Reactor sampling strategy, which generates training data for a charge-aware neuroevolution potential (qNEP) that remains stable for long-timescale interface simulations.

Load-bearing premise

The charge-aware neuroevolution potential built from Interface Reactor sampling remains first-principles accurate and free of the usual instabilities for complex electrode-electrolyte interfaces over hundred-nanosecond timescales.

What would settle it

Experimental imaging or spectroscopy of the early-stage SEI that shows identical mixed organic-inorganic composition in both carbonate and ether electrolytes, or that shows NaF crystallization not limited by surface energy in ether systems.

Figures

Figures reproduced from arXiv: 2604.07077 by Fengzijun Pan, Jianchun Sha, Pingyang Zhang, Yinghe Zhang, Zherui Chen, Zhoulin Liu, Ziliang Wang.

Figure 1
Figure 1. Figure 1: (a) Radial descriptors; (b) Angular descriptors; (c) Relationship between temperature and [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Loss function during NEP-MLP training; (b) Energy parity plot for the NEP model; (c) Force parity plot; (d) PCA projection of the training set; (e) 1 ns sampling test for energy; (f) 1 ns sampling test for forces; (g) Comparison of computational speed between AIMD, ReaxFF-MD, and NEP-MD; (h) Predicted charges for the training set and their corresponding atomic environments. Distinct Decomposition Pathw… view at source ↗
Figure 3
Figure 3. Figure 3: (a) Schematic of the model for 1M NaPF6 in EC on a Na electrode; (b) Schematic of the model for 1M NaPF6 in DME on a Na electrode. Species evolution diagrams: (c) Na-EC system at 350 K; (d) Na-EC system at 450 K; (e) Na-DME system at 350 K; (f) Na-DME system at 450 K (some curves are slightly offset to prevent overlap). Species transport diagrams at 450 K: (g) Na-EC [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Small-system results at 350 K: (a) Reaction snapshots of the Na-EC system from 0–60 ns; (b) Reaction snapshots of the Na-DME system from 0–60 ns; (c) Reaction snapshot of the Na-EC system at 100 ns; (d) Reaction snapshot of the Na-DME system at 100 ns; Large-system results at 350 K: (e) Models of the Na-EC system and Na-DME system; (f) NaF crystal; (g) NaF crystal growth in the large system; (h) Wulff cons… view at source ↗
Figure 5
Figure 5. Figure 5: Experimental validation of interfacial decomposition products. (a) Gaseous product com￾position in distinct electrolytes; (b) Temperature dependence of CO2/CO ratios in EC system; (c) [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MD simulations of EC electrolytes: (a) Initial models and MD snapshots (3 ns) for 0.5 M and 2 M NaPF6 in EC systems; (b) Schematic of sodium ion solvation structures; (c) RDFs for 1 M NaPF6 in EC systems at 300 K; (d) Na–O Coordination numbers (CNs) for different electrolytes; At￾oms are represented in a ball-and-stick model (C: grey, F: blue, H: white, Na: pink, O: red, P: or￾ange). Dynamic Evolution of t… view at source ↗
read the original abstract

The solid electrolyte interphase SEI critically dictates the cyclability and Coulombic efficiency of sodium-metal batteries, yet its dynamic formation mechanisms and atomic-scale evolution during electrochemical cycling remain elusive due to the spatiotemporal limitations of existing techniques. Here, an "Interface Reactor" sampling strategy is proposed to construct a charge-aware neuroevolution potential (qNEP). This approach overcomes the instability bottlenecks of conventional machine learning potentials, enabling stable, first-principles-accurate molecular dynamics simulations of complex electrode-electrolyte interfaces on the hundred-nanosecond scale. Fundamentally distinct SEI formation mechanisms are revealed during the early stage: carbonate-based electrolytes form heterogeneous organic-inorganic matrices via "mixed co-formation," whereas ether-based electrolytes generate dense, self-limiting inorganic barriers through "surface-energy-controlled" NaF crystallization. Metadynamics simulations further elucidate that these compositional disparities govern sodium-ion storage dynamics: NaF-rich SEIs facilitate efficient metallic deposition, while carbonate-dominated interphases induce irreversible sodium trapping and continuous electrolyte decomposition. These findings establish a comprehensive atomic-scale framework linking solvation structure, interfacial reaction networks, and electrochemical performance, providing mechanistic guidelines for rational SEI engineering in next-generation alkali-metal batteries. Crucially, a general and robust computational framework is established for simulating complex interfacial reactions in electrochemical systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes an 'Interface Reactor' sampling strategy to construct a charge-aware neuroevolution potential (qNEP) that enables stable, long-timescale (hundred-nanosecond) molecular dynamics simulations of electrode-electrolyte interfaces in sodium-metal batteries. It reports that this approach overcomes instabilities in conventional ML potentials and reveals distinct early-stage SEI formation mechanisms: heterogeneous organic-inorganic matrices via mixed co-formation in carbonate electrolytes versus dense, self-limiting inorganic NaF barriers via surface-energy-controlled crystallization in ether electrolytes. Metadynamics simulations are used to connect these compositional differences to sodium-ion storage dynamics, with NaF-rich SEIs favoring efficient deposition and carbonate-rich interphases causing trapping and decomposition. A general framework for interfacial reaction simulations is claimed.

Significance. If the qNEP validation and stability claims hold, the work offers atomic-scale mechanistic insights into SEI formation that could guide electrolyte engineering for improved sodium battery cyclability and efficiency. The computational framework for complex interfaces on extended timescales addresses a recognized challenge in the field and, if reproducible, would be a useful methodological contribution.

major comments (2)
  1. [Abstract and Methods] Abstract and (presumed) Methods/Results sections: The central claims of distinct SEI mechanisms and their link to Na storage dynamics depend on the qNEP being first-principles accurate and stable on 100 ns scales without charge-transfer artifacts. No tabulated force/energy RMSE on held-out interface configurations, no short-trajectory DFT comparisons for Na/electrolyte slabs, and no explicit demonstration that the charge-aware correction prevents instabilities during NaF nucleation or organic decomposition are provided. This validation gap is load-bearing for the mechanistic distinctions reported.
  2. [Results] Results on SEI formation (early-stage simulations): The reported 'mixed co-formation' versus 'surface-energy-controlled NaF crystallization' distinction requires that the Interface Reactor sampling produces configurations independent enough from training data to support predictive claims. The abstract labels outputs as 'predictions' but provides no clarification on held-out testing or avoidance of circularity with the sampled training set, undermining confidence in the physical versus artifactual nature of the observed mechanisms.
minor comments (2)
  1. [Abstract] Abstract: The phrase 'first-principles-accurate' should be qualified with specific error metrics or cross-validation results once added.
  2. [Throughout] Notation and figures: Define 'qNEP' and 'Interface Reactor' explicitly on first use and ensure any accompanying figures clearly label the sampling workflow and charge correction terms.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights important aspects of validation and sampling independence that strengthen the manuscript. We address each major comment below and have incorporated revisions to provide explicit evidence for the qNEP accuracy and mechanism claims.

read point-by-point responses

  1. Referee: [Abstract and Methods] Abstract and (presumed) Methods/Results sections: The central claims of distinct SEI mechanisms and their link to Na storage dynamics depend on the qNEP being first-principles accurate and stable on 100 ns scales without charge-transfer artifacts. No tabulated force/energy RMSE on held-out interface configurations, no short-trajectory DFT comparisons for Na/electrolyte slabs, and no explicit demonstration that the charge-aware correction prevents instabilities during NaF nucleation or organic decomposition are provided. This validation gap is load-bearing for the mechanistic distinctions reported.

    Authors: We agree that explicit tabulated validation on held-out interface data is essential to support the stability and accuracy claims. The original submission included overall training RMSE and some stability tests but lacked dedicated held-out interface metrics and direct DFT comparisons at nucleation events. In the revised manuscript, we have added a new table (Table S2) reporting force and energy RMSE on 500 held-out interface configurations extracted from independent short DFT-MD runs. We also include 10-ps DFT reference trajectories for Na/electrolyte slabs with direct force-error plots versus qNEP. A supplementary figure now shows side-by-side qNEP versus DFT forces during NaF nucleation and organic decomposition steps, confirming that the charge-aware correction eliminates the charge-transfer instabilities observed in the non-charge-aware baseline. These additions make the validation load-bearing evidence explicit. revision: yes

  2. Referee: [Results] Results on SEI formation (early-stage simulations): The reported 'mixed co-formation' versus 'surface-energy-controlled NaF crystallization' distinction requires that the Interface Reactor sampling produces configurations independent enough from training data to support predictive claims. The abstract labels outputs as 'predictions' but provides no clarification on held-out testing or avoidance of circularity with the sampled training set, undermining confidence in the physical versus artifactual nature of the observed mechanisms.

    Authors: The Interface Reactor procedure is designed to generate diverse configurations via iterative active learning, but we acknowledge that the original text did not explicitly quantify independence or report held-out performance for the long-timescale trajectories. In revision, the Methods section now details an 80/20 train/test split on the final dataset, with uncertainty-based selection ensuring that the 100-ns production runs start from configurations whose local environments were not directly present in training. We have added a paragraph clarifying that the reported mechanisms are extrapolative predictions beyond the training distribution, supported by low uncertainty estimates throughout the trajectories. This removes any ambiguity regarding circularity while preserving the mechanistic distinctions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation chain is self-contained

full rationale

The paper proposes an Interface Reactor sampling strategy to generate first-principles data for training a charge-aware neuroevolution potential (qNEP), then deploys the trained potential for extended MD and metadynamics simulations that reveal emergent SEI formation mechanisms. These mechanisms (mixed co-formation in carbonates vs. surface-energy-controlled NaF crystallization in ethers) arise from the dynamics on 100 ns scales, which exceed direct DFT reach, rather than being equivalent to the training inputs by construction. No equations or claims reduce a prediction to a fitted parameter, no self-citation is load-bearing for the central mechanistic distinction, and no uniqueness theorem or ansatz is imported from prior author work. The approach is therefore not circular; any concerns about validation metrics or potential artifacts belong to correctness rather than circularity analysis.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

The central claim depends on the new sampling method producing a stable, accurate qNEP and on metadynamics faithfully modeling Na storage; these rest on domain assumptions about ML potential transferability and simulation timescales rather than independent external benchmarks.

free parameters (1)
  • qNEP fitting parameters
    Neuroevolution potential parameters are trained on Interface Reactor sampled data; specific values and fitting protocol not provided in abstract.
axioms (1)
  • domain assumption qNEP delivers first-principles-accurate and stable dynamics for complex electrode-electrolyte interfaces
    Invoked to justify long-timescale simulations and mechanistic revelations.
invented entities (2)
  • Interface Reactor sampling strategy no independent evidence
    purpose: Construct stable qNEP for interfacial reaction networks
    New method introduced to overcome conventional ML potential instability.
  • qNEP (charge-aware neuroevolution potential) no independent evidence
    purpose: Enable hundred-nanosecond MD of SEI formation
    Developed via the new sampling approach in this work.

pith-pipeline@v0.9.0 · 5552 in / 1464 out tokens · 79310 ms · 2026-05-10T17:18:46.159069+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

  1. [1]

    Young, Martin Lebrat, and Markus Greiner

    (1) Yang, C.; Singh, A.; Pu, X.; Mallarapu, A.; Smith, K.; Keyser, M.; Haberman, M. R.; Khani, H.; Misztal, P.; Spray, R.; Ezekoye, O. A.; Finegan, D. P. Addressing the Safety of Ne xt- 27 Generation Batteries. Nature 2025, 645 (8081), 603–613. https://doi.org/10.1038/s41586-025- 09358-4. (2) Li, M.; Sun, C.; Zhang, R.; Qi, M.; Wu, Z.; Zhang, X.; Zhang, Y...

    work page doi:10.1038/s41586-025- 2025

  2. [2]

    (15) Liu, Q.; Mu, D.; Wu, B.; Wang, L.; Gai, L.; Wu, F

    https://doi.org/10.1038/s41467-021-23368-6. (15) Liu, Q.; Mu, D.; Wu, B.; Wang, L.; Gai, L.; Wu, F. Density Functional Theory Research into the Reducti on Mechanism for the Solvent/Additive in a Sodium -Ion Battery. ChemSusChem 2017, 10 (4), 786–796. https://doi.org/10.1002/cssc.201601356. (16) Young, J.; Kulick, P. M.; Juran, T. R.; Smeu, M. Comparative ...

    work page doi:10.1038/s41467-021-23368-6 2017

  3. [3]

    (22) Wu, S.; Yang, X.; Zhao, X.; Li, Z.; Lu, M.; Xie, X.; Yan, J

    https://doi.org/10.1038/s41524-025-01615-4. (22) Wu, S.; Yang, X.; Zhao, X.; Li, Z.; Lu, M.; Xie, X.; Yan, J. Applications and Advances in Ma- chine Learning Force Fields. J. Chem. Inf. Model. 2023, 63 (22), 6972 –6985. https://doi.org/10.1021/acs.jcim.3c00889. (23) Unke, O. T.; Chmiela, S.; Sauceda, H . E.; Gastegger, M.; Poltavsky, I.; Schü tt, K. T.; T...

    work page doi:10.1038/s41524-025-01615-4 2023

  4. [4]

    (25) Zhang, D.; Yi, P.; Lai, X.; Peng, L.; Li, H

    https://doi.org/10.1038/s41467-025-67982-0. (25) Zhang, D.; Yi, P.; Lai, X.; Peng, L.; Li, H. Active Machine Learn ing Model for the Dynamic Simulation and Growth Mechanisms of Carbon on Metal Surface. Nat. Commun. 2024, 15 (1),

    work page doi:10.1038/s41467-025-67982-0 2024

  5. [5]

    https://doi.org/10.1038/s41467-023-44525-z. (26) Song, K.; Zhao, R.; Liu, J.; Wang, Y.; Lindgren, E.; Wang, Y.; Chen, S.; Xu, K.; Liang, T.; Ying, P.; Xu, N.; Zhao, Z.; Shi, J.; Wang, J.; Lyu, S.; Zeng, Z.; Liang, S.; Dong, H.; Sun, L.; Chen, Y.; Zhang, Z.; Guo, W.; Qian, P.; Sun, J.; Erhart, P.; Ala -Nissila, T.; Su, Y.; Fan, Z. General-Purpose Machine -...

    work page doi:10.1038/s41467-023-44525-z 2024

  6. [6]

    (29) Xu, L.; Shao, W.; Jin, H.; Wang, Q

    https://doi.org/10.1021/jacs.5c21439. (29) Xu, L.; Shao, W.; Jin, H.; Wang, Q. Data Efficient and Stability Indicated Sampling for Devel- oping Reactive Machine Learning Potential to Achieve Ultralong Simulation in Lithium -Metal Batteries. J. Phys . Chem. C 2023, 127 (50), 24106 –24117. https://doi.org/10.1021/acs.jpcc.3c05522. (30) Zhang, S. Exploring t...

    work page doi:10.1021/jacs.5c21439 2023

  7. [7]

    (32) Niu, Y.-L.; Chen, X.; Zhang, T

    https://doi.org/10.1038/s41524-025-01747-7. (32) Niu, Y.-L.; Chen, X.; Zhang, T. -C.; Gao, Y.-C.; Chen, Y.-P.; Yao, N.; Fu, Z. -H.; Zhang, Q. A Large-Scale on-the-Fly Machine Learning Molecular Dynamics Simulation to Explore Lithium Metal Battery Interfaces. J. Energy Chem. 2025, 110, 356 –362. https://doi.org/10.1016/j.jechem.2025.06.072. (33) Xu, B.; Ba...

    work page doi:10.1038/s41524-025-01747-7 2025

  8. [8]

    31 (35) Zhang, K.; Ji, Y.; Wu, Q.; Nabavizadeh, S

    https://doi.org/10.1038/s41467-023-42212-7. 31 (35) Zhang, K.; Ji, Y.; Wu, Q.; Nabavizadeh, S. A.; Qi, Y.; Chen, L.-Q. Simulating Solid Electrolyte Interphase Formation Spanning 108 Time Scales with Atomically Informed Phase-Field Model. Energy Environ. Sci

    work page doi:10.1038/s41467-023-42212-7

  9. [9]

    https://doi.org/10.1039/D5EE01030F. (36) Xu, K.; Bu, H.; Pan, S.; Lindgren, E.; Wu, Y.; Wang, Y.; Liu, J.; Song, K.; Xu, B.; Li, Y.; Hainer, T.; Svensson, L.; Wiktor, J.; Zhao, R.; Huang, H.; Qian, C.; Zhang, S.; Zeng, Z.; Zhang, B.; Tang, B.; Xiao, Y.; Yan, Z.; Shi, J.; Liang, Z.; Wang, J.; Liang, T.; Cao, S.; Wang, Y.; Ying, P.; Xu, N.; Chen, C.; Zhang,...

    work page doi:10.1039/d5ee01030f 2025

  10. [10]

    (38) Pan, F.; Li, G.; Xu, J.; Wang, Z.; Fan, Z.; Zhang, A.; Wang, D

    https://doi.org/10.48550/arXiv.2601.19034. (38) Pan, F.; Li, G.; Xu, J.; Wang, Z.; Fan, Z.; Zhang, A.; Wang, D. CeNEP: A General-Purpose Machine Learning Interatomic Potential for Cementitious Materials. Chem. Eng. J. 2025, 526, 170639. https://doi.org/10.1016/j.cej.2025.170639. (39) Yan, Z.; Zhu, Y. Impact of Lithium Nonstoichiometry on Ionic Diffusion i...

    work page doi:10.48550/arxiv.2601.19034 2025

  11. [11]

    32 (41) Kresse, G.; Furthmü ller, J

    https://doi.org/10.48550/arXiv.2504.15925. 32 (41) Kresse, G.; Furthmü ller, J. Efficient Iterative Schemes for Ab Initio Total-Energy Calculations Using a Plane -Wave Basis Set. Phys. Rev. B 1996, 54 (16), 11169 –11186. https://doi.org/10.1103/PhysRevB.54.11169. (42) Perdew, J. P.; Burke, K.; Ernzerhof, M. Generalized Gradient Approximation Made Simple. ...

    work page doi:10.48550/arxiv.2504.15925 1996

  12. [12]

    (44) Kresse, G.; Joubert, D

    https://doi.org/10.1038/s41467-025-62824-5. (44) Kresse, G.; Joubert, D. From Ultrasoft Pseudopotentials to the Projector Augmented -Wave Method. Phys. Rev. B 1999, 59 (3), 1758–1775. https://doi.org/10.1103/PhysRevB.59.1758. (45) Monkhorst, H. J.; Pack, J. D. Special Points for Brillouin-Zone Integrations. Phys. Rev. B 1976, 13 (12), 5188–5192. https://d...

    work page doi:10.1038/s41467-025-62824-5 1999

  13. [13]

    Liang, K

    https://doi.org/10.48550/ARXIV.2504.21286. (50) Chen, H. ReaxTools: A High Performance ReaxF F/AIMD/MLP-MD Post -Process Code (Computer Software),

    work page doi:10.48550/arxiv.2504.21286

  14. [14]

    33 (51) Islam, M

    https://github.com/tgraphite/reax_tools (accessed 2025-05-01). 33 (51) Islam, M. M.; Van Duin, A. C. T. Reductive Decomposition Reactions of Ethylene Carbonate by Explicit Electron Transfer from Lithium: An eReaxFF Molecular Dynamics Study. J. Phys. Chem. C 2016, 120 (48), 27128–27134. https://doi.org/10.1021/acs.jpcc.6b08688. (52) Young, J.; Smeu, M. Eth...

    work page doi:10.1021/acs.jpcc.6b08688 2025

  15. [15]

    Kalikmanov,Nucleation Theory, Lecture Notes in Physics, Volume 860, Springer Science+Business Media, Dordrecht, NL (2013), 10.1007/978-90-481-3643-8

    https://doi.org/10.1007/978-90-481-3643-8. (54) Zhang, L.; Yang, M.; Zhang, S.; Niu, H. Unveiling the Crystallization Mechanism of Cadmium Selenide via Molecular Dynamics Simulation with Machine-Learning-Based Deep Potential. J. Mater. Sci. Technol. 2024, 185, 23–31. https://doi.org/10.1016/j.jmst.2023.09.059. (55) Warren, J. A.; Pusztai, T.; Kö rnyei, L....

    work page doi:10.1007/978-90-481-3643-8 2024