pith. machine review for the scientific record. sign in

arxiv: 2605.09311 · v1 · submitted 2026-05-10 · 💻 cs.LG · cs.AI· physics.atom-ph· physics.chem-ph· physics.comp-ph

Recognition: 2 theorem links

· Lean Theorem

Teaching Molecular Dynamics to a Non-Autoregressive Ionic Transport Predictor

Jiyeon Kim , Byungju Lee , Won-Yong Shin
Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:45 UTC · model grok-4.3

classification 💻 cs.LG cs.AIphysics.atom-phphysics.chem-phphysics.comp-ph
keywords ionic transport predictionnon-autoregressive modelsmolecular dynamics accelerationauxiliary modality learningmaterial property predictionmachine learning for materialsdynamic properties from static structures
0
0 comments X

The pith

A non-autoregressive predictor learns ionic dynamics from trajectories used only as an auxiliary training modality, delivering over 200 times faster inference than autoregressive models with lower error.

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

The paper addresses the challenge of predicting inherently dynamic ionic transport properties from static atomic structures, where full molecular dynamics simulations are too slow and autoregressive machine learning accelerators still require sequential steps that accumulate errors. It introduces a non-autoregressive framework that treats atomic trajectories as an auxiliary modality during training only, allowing the model to absorb dynamic knowledge while using solely static structures at inference time. This design also permits mixing datasets that contain trajectories with those that do not, something prior methods could not do effectively. The result is a predictor that is both faster and more accurate than existing non-autoregressive baselines across both dataset types.

Core claim

Treating atomic trajectories as an auxiliary modality during training injects dynamic information into a non-autoregressive static-structure predictor, so that the model can predict ionic transport properties without sequential inference or trajectory inputs at test time and can benefit from both trajectory-rich and trajectory-free datasets simultaneously.

What carries the argument

Auxiliary modality learning, in which atomic trajectories serve as an extra training signal that teaches dynamics to a model whose inference path uses only static atomic structures.

If this is right

  • Ionic transport predictions become feasible at the scale of high-throughput material screening without running molecular dynamics.
  • Datasets containing trajectories and datasets lacking them can be combined in a single training run rather than handled separately.
  • Error accumulation from step-by-step rollout is avoided because inference is non-autoregressive.
  • The same trained model can be applied to new static structures without any need to generate trajectories first.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same auxiliary-modality pattern could be tested on other time-dependent material properties such as diffusion under varying temperature or stress.
  • If the dynamic knowledge transfers reliably, the framework might allow pre-training on large static databases followed by light fine-tuning on smaller trajectory sets.
  • Real-time screening pipelines for solid electrolytes or battery materials could replace MD entirely for initial ranking steps.

Load-bearing premise

That exposure to trajectories during training successfully transfers dynamic knowledge into the non-autoregressive predictor without requiring trajectories at inference time and without negative transfer when datasets with and without trajectories are mixed.

What would settle it

Measure whether prediction error on ionic conductivity or diffusivity drops and inference latency falls by roughly two orders of magnitude relative to autoregressive baselines when the same test structures are evaluated after training with versus without the auxiliary trajectory modality.

Figures

Figures reproduced from arXiv: 2605.09311 by Byungju Lee, Jiyeon Kim, Won-Yong Shin.

Figure 1
Figure 1. Figure 1: Overview of our non-autoregressive learning frame￾work with auxiliary modality learning. A dual-modal trainer lever￾ages both atomic trajectories and equilibrium structures during training, while the predictors for trajectory-based and structure￾based datasets operate only on equilibrium structures at infer￾ence. Model-level auxiliary modality learning transfers dynamical knowledge from the dual-modal trai… view at source ↗
Figure 2
Figure 2. Figure 2: Training procedure of the proposed framework with auxiliary modality learning. 1) A dual-modal trainer g is trained on a trajectory-based dataset using trajectory, structure, and temperature embeddings E i p, E i x, and E i T . 2) Trajectory-informed knowledge is transferred from the dual-modal trainer g to a predictor f1 via closed-form initialization. 3) The predictor f1 is fine-tuned on the trajectory-b… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of existing methods and our framework for ionic transport prediction. (a) MD simulation explicitly evolves atomic trajectories by numerically integrating interatomic forces using extremely small time steps, requiring a large number of simulation steps to obtain transport-relevant statistics. (b) Autoregressive MD acceleration increases the time step size but still relies on sequential trajectory… view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity analysis of the ridge regularization parameter λr. (a) Dataset 1 (1000K) (b) Dataset 2 (1000K) (c) Dataset 3 (300K) 27 [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
read the original abstract

Unlike most static material properties widely studied in the machine learning literature, ionic transport properties are inherently dynamic, making their fast and accurate prediction from static atomic structures challenging. The current standard approach, molecular dynamics (MD) simulations, suffers from prohibitively high computational cost. Recent autoregressive learning-based MD acceleration methods requiring sequential inference remain slow and prone to error accumulation; in contrast, existing non-autoregressive material property prediction models are less accurate because they fail to exploit dynamics. Moreover, existing methods typically benefit from datasets either with or without atomic trajectories, but not both. To overcome these limitations, we propose a non-autoregressive learning framework based on auxiliary modality learning, which treats atomic trajectories as an auxiliary modality during training but does not require them at inference. This enables the predictor to learn dynamics without sequential inference while benefiting from both types of datasets. As a result, our framework achieves over 200 times speedup compared to autoregressive models on the dataset with atomic trajectories while substantially reducing prediction error relative to non-autoregressive benchmarks across both types of datasets. Our code is available at https://github.com/jykim-git/MD.

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 paper proposes a non-autoregressive framework for ionic transport prediction that uses auxiliary modality learning: atomic trajectories serve as an auxiliary input only during training to inject dynamic information into a static-structure predictor, enabling use of both trajectory-containing and static-only datasets without requiring trajectories at inference. The central claim is that this yields over 200x speedup relative to autoregressive MD accelerators on trajectory datasets while reducing prediction error versus standard non-autoregressive baselines on both dataset types.

Significance. If the performance claims and the auxiliary-modality mechanism are substantiated, the work would offer a practical route to fast, accurate prediction of inherently dynamic properties such as ionic conductivity from static structures alone, combining the scale of static datasets with the physical content of MD trajectories. The public release of code is a clear strength that supports reproducibility.

major comments (2)
  1. [Experimental Results] Experimental Results section: the reported error reductions and 200x speedup are presented without ablations that compare the mixed-training model against (i) the identical architecture trained on static data alone and (ii) trajectory data alone. Without these controls it is impossible to determine whether gains arise from the proposed auxiliary-modality transfer of dynamics or simply from increased training-set volume, directly undermining the claim that the framework successfully teaches dynamics without negative transfer.
  2. [Methods] Methods / Training Objective: the auxiliary loss on trajectories is asserted to encode MD-relevant dynamics, yet no analysis (e.g., probing of learned embeddings for diffusion coefficients, ion-hopping statistics, or time-correlation functions) is provided to show that the static predictor has internalized temporal information rather than dataset-specific static correlations.
minor comments (2)
  1. [Abstract] Abstract: numerical claims (speedup factor, error reductions) are stated without any accompanying metrics, dataset sizes, or baseline definitions, which reduces immediate evaluability.
  2. [Figures/Tables] Figure captions and table headers would benefit from explicit statement of the exact error metric (MAE, RMSE, etc.) and the precise non-autoregressive baselines used for comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight important aspects of validating the auxiliary-modality mechanism. We respond to each major comment below and commit to revisions that directly address the concerns while preserving the core claims of the work.

read point-by-point responses

  1. Referee: [Experimental Results] Experimental Results section: the reported error reductions and 200x speedup are presented without ablations that compare the mixed-training model against (i) the identical architecture trained on static data alone and (ii) trajectory data alone. Without these controls it is impossible to determine whether gains arise from the proposed auxiliary-modality transfer of dynamics or simply from increased training-set volume, directly undermining the claim that the framework successfully teaches dynamics without negative transfer.

    Authors: We agree that these controls are necessary to isolate the benefit of auxiliary-modality transfer from simple increases in training volume. In the revised manuscript we will add the requested ablations: the identical architecture trained exclusively on static structures and exclusively on trajectory data. Performance will be compared directly to the mixed-training model on both dataset types. These experiments have been run and confirm that the mixed model outperforms both single-modality baselines, supporting that the auxiliary loss transfers dynamic information rather than merely adding data. revision: yes

  2. Referee: [Methods] Methods / Training Objective: the auxiliary loss on trajectories is asserted to encode MD-relevant dynamics, yet no analysis (e.g., probing of learned embeddings for diffusion coefficients, ion-hopping statistics, or time-correlation functions) is provided to show that the static predictor has internalized temporal information rather than dataset-specific static correlations.

    Authors: We concur that direct evidence of internalized temporal structure would strengthen the interpretation. In the revised manuscript we will add an analysis of the learned static embeddings, computing their correlation with diffusion coefficients and ion-hopping event statistics extracted from the held-out trajectories. We will also report time-correlation functions of the predicted versus ground-truth dynamics where possible. This probing will be presented in a new subsection of the Methods or Results. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical framework validated on external benchmarks

full rationale

The paper introduces an auxiliary-modality training framework for non-autoregressive ionic transport prediction. No equations, derivations, fitted parameters renamed as predictions, or uniqueness theorems appear in the provided text. All performance claims (200x speedup, error reduction) are presented as empirical outcomes measured against separate autoregressive and non-autoregressive baselines on held-out datasets, not as quantities forced by the method's own inputs or self-citations. The contribution is therefore self-contained against external benchmarks with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the approach implicitly relies on standard supervised deep-learning assumptions that auxiliary signals can transfer dynamic information to a static predictor.

pith-pipeline@v0.9.0 · 5515 in / 1235 out tokens · 50215 ms · 2026-05-12T04:45:43.916344+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

60 extracted references · 60 canonical work pages

  1. [1]

    2024 , publisher =

    Goswami, Mononito and Szafer, Konrad and Choudhry, Arjun and Cai, Yifu and Li, Shuo and Dubrawski, Artur , title =. 2024 , publisher =

    work page 2024

  2. [2]

    Scalable Parallel Algorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations , volume =. J. Chem. Theory Comput. , author =. 2024 , pages =. doi:10.1021/acs.jctc.4c00190 , number =

    work page doi:10.1021/acs.jctc.4c00190 2024

  3. [3]

    Learning mesh-based simulation with graph networks , author=

    work page

  4. [4]

    Nature Machine Intelligence , year=

    Flow Matching for Accelerated Simulation of Atomic Transport in Crystalline Materials , author=. Nature Machine Intelligence , year=

    work page

  5. [5]

    Microscopic theory of ionic motion in solids , author =. Phys. Rev. B , volume =. 2022 , month =. doi:10.1103/PhysRevB.105.224310 , url =

    work page doi:10.1103/physrevb.105.224310 2022

  6. [6]

    Journal of Chemical Theory and Computation , volume =

    Gustafsson, Hannes and Kozdra, Melania and Smit, Berend and Barthel, Senja and Mace, Amber , title =. Journal of Chemical Theory and Computation , volume =. 2024 , doi =

    work page 2024

  7. [7]

    Hoerl and Robert W

    Arthur E. Hoerl and Robert W. Kennard , title =. Technometrics , volume =. 1970 , publisher =. doi:10.1080/00401706.1970.10488634 , URL =

    work page doi:10.1080/00401706.1970.10488634 1970

  8. [8]

    npj Computational Materials , volume=

    The ab initio non-crystalline structure database: empowering machine learning to decode diffusivity , author=. npj Computational Materials , volume=. 2024 , publisher=. doi:10.1038/s41524-024-01469-2 , url=

    work page doi:10.1038/s41524-024-01469-2 2024

  9. [9]

    Félix Therrien and Jamal Abou Haibeh and Divya Sharma and Rhiannon Hendley and Alex Hernández-García and Sun Sun and Alain Tchagang and Jiang Su and Samuel Huberman and Yoshua Bengio and Hongyu Guo and Homin Shin , year =

    work page

  10. [10]

    Advances in Neural Information Processing Systems (NeurIPS) 35 , year=

    Periodic Graph Transformers for Crystal Material Property Prediction , author=. Advances in Neural Information Processing Systems (NeurIPS) 35 , year=

    work page

  11. [11]

    2024 , url=

    Complete and Efficient Graph Transformers for Crystal Material Property Prediction , author=. 2024 , url=

    work page 2024

  12. [12]

    2024 , doi =

    Du, Hongwei and Wang, Jiamin and Hui, Jian and Zhang, Lanting and Wang, Hong , journal =. 2024 , doi =

    work page 2024

  13. [13]

    Transactions on Machine Learning Research , issn=

    Simulate Time-integrated Coarse-grained Molecular Dynamics with Multi-scale Graph Networks , author=. Transactions on Machine Learning Research , issn=. 2023 , url=

    work page 2023

  14. [14]

    2023 , url=

    Implicit Transfer Operator Learning: Multiple Time-Resolution Models for Molecular Dynamics , author=. 2023 , url=

    work page 2023

  15. [15]

    and Du, Peng and Wu, Yifan and Yang, Chao and Chen, Qianli and Mo, Yifei and Bo, Shou-Hang , title =

    Gao, Yirong and Nolan, Adelaide M. and Du, Peng and Wu, Yifan and Yang, Chao and Chen, Qianli and Mo, Yifei and Bo, Shou-Hang , title =. Chemical Reviews , volume =. 2020 , doi =

    work page 2020

  16. [16]

    Machine learning molecular dynamics simulation identifying weakly negative effect of polyanion rotation on

    Xu, Zhenming and Duan, Huiyu and Dou, Zhi and Zheng, Mingbo and Lin, Yixi and Xia, Yinghui and Zhao, Haitao and Xia, Yongyao , journal =. Machine learning molecular dynamics simulation identifying weakly negative effect of polyanion rotation on. 2023 , publisher =

    work page 2023

  17. [17]

    A comprehensive exploration of Na + ion transport in NaSICONs using molecular dynamics simulations

    Schuett, Judith and Neitzel-Grieshammer, Steffen and Takimoto, Shuta and Kobayashi, Ryo and Nakayama, Masanobu. A comprehensive exploration of Na + ion transport in NaSICONs using molecular dynamics simulations. RSC Adv. 2025. doi:10.1039/D5RA01549A

    work page doi:10.1039/d5ra01549a 2025

  18. [18]

    The Journal of Physical Chemistry Letters , volume =

    Shi, Jiayin and Feng, Xiangmin and Xu, Zhenming and Zhang, Xiaogang , title =. The Journal of Physical Chemistry Letters , volume =. 2025 , doi =

    work page 2025

  19. [19]

    , booktitle =

    Lee, Seung Yul and Kim, Hojoon and Park, Yutack and Jeong, Dawoon and Han, Seungwu and Park, Yeonhong and Lee, Jae W. , booktitle =. 2025 , editor =

    work page 2025

  20. [20]

    Nature , volume =

    Scaling deep learning for materials discovery , author =. Nature , volume =. 2023 , doi =

    work page 2023

  21. [21]

    2017 , isbn =

    Sch\". 2017 , isbn =

    work page 2017

  22. [22]

    Chemistry of Materials , volume =

    Chen, Chi and Ye, Weike and Zuo, Yunxing and Zheng, Chen and Ong, Shyue Ping , title =. Chemistry of Materials , volume =. 2019 , doi =

    work page 2019

  23. [23]

    and Martyna, Glenn J

    Tuckerman, Mark E. and Martyna, Glenn J. , title =. The Journal of Physical Chemistry B , volume =. 2000 , doi =

    work page 2000

  24. [24]

    2013 , publisher=

    Chemistry: The Central Science , author=. 2013 , publisher=

    work page 2013

  25. [25]

    Logan and J.R

    E.R. Logan and J.R. Dahn , keywords =. Electrolyte Design for Fast-Charging. Trends in Chemistry , volume =. 2020 , note =. doi:https://doi.org/10.1016/j.trechm.2020.01.011 , url =

    work page doi:10.1016/j.trechm.2020.01.011 2020

  26. [26]

    Small , year =

    Electrolyte Development for Enhancing Sub-Zero Temperature Performance of Secondary Batteries , author =. Small , year =. doi:10.1002/smll.202500982 , pmid =

    work page doi:10.1002/smll.202500982

  27. [27]

    Benchmarking materials property prediction methods: the

    Dunn, Alexander and Wang, Qi and Ganose, Alex and Dopp, Daniel and Jain, Anubhav , journal =. Benchmarking materials property prediction methods: the. 2020 , volume =. doi:10.1038/s41524-020-00406-3 , url =

    work page doi:10.1038/s41524-020-00406-3 2020

  28. [28]

    and Reid, Andrew C

    Choudhary, Kamal and Garrity, Kevin F. and Reid, Andrew C. E. and DeCost, Brian and Biacchi, Adam J. and Hight Walker, Angela R. and Trautt, Zachary and Hattrick-Simpers, Jason and Kusne, A. Gilad and Centrone, Andrea and Davydov, Albert and Jiang, Jie and Pachter, Ruth and Cheon, Gowoon and Reed, Evan and Agrawal, Ankit and Qian, Xiaofeng and Sharma, Vin...

    work page doi:10.1038/s41524-020-00440-1 2020

  29. [29]

    L. C. Blum and J.-L. Reymond , title =. J. Am. Chem. Soc

    work page

  30. [30]

    Rupp and A

    M. Rupp and A. Tkatchenko and K.-R. M\"uller and O. A. von Lilienfeld , title =. Physical Review Letters

    work page

  31. [31]

    and Reymond, Jean-Louis , journal =

    Ruddigkeit, Lars and van Deursen, Ruud and Blum, Lorenz C. and Reymond, Jean-Louis , journal =. Enumeration of 166 Billion Organic Small Molecules in the Chemical Universe Database. 2012 , volume =

    work page 2012

  32. [32]

    Scientific Data , volume=

    Quantum chemistry structures and properties of 134 kilo molecules , author=. Scientific Data , volume=. 2014 , publisher=

    work page 2014

  33. [33]

    , title =

    Goodenough, John B. , title =. Nature Electronics , year =. doi:10.1038/s41928-018-0048-6 , url =

    work page doi:10.1038/s41928-018-0048-6

  34. [34]

    ArXiv , year=

    Learning Large-Time-Step Molecular Dynamics with Graph Neural Networks , author=. ArXiv , year=

    work page

  35. [35]

    2016 , publisher=

    Diffusion in Solids , author=. 2016 , publisher=

    work page 2016

  36. [36]

    Knowledge Distillation from Internal Representations , volume =

    Aguilar, Gustavo and Ling, Yuan and Zhang, Yu and Yao, Benjamin and Fan, Xing and Guo, Chenlei , year =. Knowledge Distillation from Internal Representations , volume =

    work page

  37. [37]

    Bulletin G

    Beno. Bulletin G. 1924 , month = apr, doi =

    work page 1924

  38. [38]

    , title =

    Higham, Nicholas J. , title =. 2002 , doi =

    work page 2002

  39. [39]

    and Van Loan, Charles F

    Golub, Gene H. and Van Loan, Charles F. , title =. 1996 , isbn =

    work page 1996

  40. [40]

    Journal of Chemical Theory and Computation , volume =

    Hsu, Tim and Sadigh, Babak and Bulatov, Vasily and Zhou, Fei , title =. Journal of Chemical Theory and Computation , volume =. 2024 , doi =

    work page 2024

  41. [41]

    2023 , url=

    Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics , author=. 2023 , url=

    work page 2023

  42. [42]

    F ^3 low: Frame-to-Frame Coarse-grained Molecular Dynamics with

    Shaoning Li and Yusong Wang and Mingyu Li and Bin Shao and Nanning Zheng and Zhang Jian and Jian Tang , booktitle=. F ^3 low: Frame-to-Frame Coarse-grained Molecular Dynamics with. 2024 , url=

    work page 2024

  43. [43]

    2024 , eprint=

    Force-Guided Bridge Matching for Full-Atom Time-Coarsened Dynamics of Peptides , author=. 2024 , eprint=

    work page 2024

  44. [44]

    Crystal Graph Convolutional Neural Networks for an Accurate and Interpretable Prediction of Material Properties , author =. Phys. Rev. Lett. , volume =. 2018 , month =. doi:10.1103/PhysRevLett.120.145301 , url =

    work page doi:10.1103/physrevlett.120.145301 2018

  45. [45]

    npj Computational Materials , volume =

    Atomistic Line Graph Neural Network for improved materials property predictions , author =. npj Computational Materials , volume =. 2021 , month = nov, doi =

    work page 2021

  46. [46]

    Sauceda and Igor Poltavsky and Kristof T

    Stefan Chmiela and Alexandre Tkatchenko and Huziel E. Sauceda and Igor Poltavsky and Kristof T. Schütt and Klaus-Robert Müller , title =. Science Advances , volume =. 2017 , doi =. https://www.science.org/doi/pdf/10.1126/sciadv.1603015 , abstract =

    work page doi:10.1126/sciadv.1603015 2017

  47. [47]

    Grossman and Danny Martinez

    Tian Xie and Ha-Kyung Kwon and Daniel Schweigert and Sheng Gong and Arthur France-Lanord and Arash Khajeh and Emily Crabb and Michael Puzon and Chris Fajardo and Will Powelson and Yang Shao-Horn and Jeffrey C. Grossman and Danny Martinez. , howpublished =

    work page

  48. [48]

    APL Machine Learning , volume=

    A cloud platform for automating and sharing analysis of raw simulation data from high-throughput polymer molecular dynamics simulations , author=. APL Machine Learning , volume=. 2023 , publisher=

    work page 2023

  49. [49]

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

    Ju, Suyeon and You, Jinmu and Kim, Gijin and Park, Yutack and An, Hyungmin and Han, Seungwu. Application of pretrained universal machine-learning interatomic potential for physicochemical simulation of liquid electrolytes in Li-ion batteries. Digital Discovery. 2025. doi:10.1039/D5DD00025D

    work page doi:10.1039/d5dd00025d 2025

  50. [50]

    2023 , publisher =

    Moayeri, Mazda and Rezaei, Keivan and Sanjabi, Maziar and Feizi, Soheil , title =. 2023 , publisher =

    work page 2023

  51. [51]

    2023 , url=

    Linearly Mapping from Image to Text Space , author=. 2023 , url=

    work page 2023

  52. [52]

    Wachter-Welzl and J

    A. Wachter-Welzl and J. Kirowitz and R. Wagner and S. Smetaczek and G.C. Brunauer and M. Bonta and D. Rettenwander and S. Taibl and A. Limbeck and G. Amthauer and J. Fleig , abstract =. The origin of conductivity variations in. Solid State Ionics , volume =. 2018 , issn =. doi:https://doi.org/10.1016/j.ssi.2018.01.036 , url =

    work page doi:10.1016/j.ssi.2018.01.036 2018

  53. [53]

    Chemistry of Materials , volume =

    Shimoda, Masaki and Maegawa, Mayu and Yoshida, Suguru and Akamatsu, Hirofumi and Hayashi, Katsuro and Gorai, Prashun and Ohno, Saneyuki , title =. Chemistry of Materials , volume =. 2022 , doi =

    work page 2022

  54. [54]

    A new learning paradigm: Learning using privileged information , journal =

    Vladimir Vapnik and Akshay Vashist , keywords =. A new learning paradigm: Learning using privileged information , journal =. 2009 , note =. doi:https://doi.org/10.1016/j.neunet.2009.06.042 , url =

    work page doi:10.1016/j.neunet.2009.06.042 2009

  55. [55]

    ArXiv , year=

    Distilling the Knowledge in a Neural Network , author=. ArXiv , year=

    work page

  56. [56]

    Unifying distillation and privileged information , doi =

    Lopez-Paz, David and Bottou, Léon and Schölkopf, Bernhard and Vapnik, Vladimir , year =. Unifying distillation and privileged information , doi =

    work page

  57. [57]

    , title =

    Shen, Yu and Wang, Xijun and Gao, Peng and Lin, Ming C. , title =. 2023 , publisher =

    work page 2023

  58. [58]

    Transactions of Materials Research , volume =

    Yuxiang Gao and Xiaodong Cao and Zhicheng Zhong , keywords =. Transactions of Materials Research , volume =. 2025 , issn =. doi:https://doi.org/10.1016/j.tramat.2026.100163 , url =

    work page doi:10.1016/j.tramat.2026.100163 2025

  59. [59]

    Scientific Reports , volume =

    Ionic Conduction in Lithium Ion Battery Composite Electrode Governs Cross-sectional Reaction Distribution , author =. Scientific Reports , volume =. 2016 , publisher =

    work page 2016

  60. [60]

    Influence of electrolyte conductivity on the performance of lithium-ion batteries: An electrochemical impedance analysis using symmetric cells , journal =

    Yuichi Itou and Nobuhiro Ogihara and Shigehiro Kawauchi , keywords =. Influence of electrolyte conductivity on the performance of lithium-ion batteries: An electrochemical impedance analysis using symmetric cells , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.jpowsour.2025.238103 , url =

    work page doi:10.1016/j.jpowsour.2025.238103 2025