pith. sign in

arxiv: 2606.02335 · v1 · pith:I5RI5Z5Snew · submitted 2026-06-01 · ❄️ cond-mat.mtrl-sci · math-ph· math.MP· physics.app-ph· physics.comp-ph

Neural Spectral Element Methods for stiff multiphysics PDEs with electrochemical transport benchmarks

Conrard Giresse Tetsassi Feugmo , David Pankaczy This is my paper

Pith reviewed 2026-06-28 13:30 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci math-phmath.MPphysics.app-phphysics.comp-ph
keywords Neural Spectral Element MethodPoisson-Nernst-Planck equationselectrochemical transportspectral differentiationphysics-informed neural networksboundary layer resolutionL-BFGS optimizationKolmogorov-Arnold Networks
0
0 comments X

The pith

The Neural Spectral Element Method solves stiff electrochemical PDEs to 10^-4--10^-7 error using two orders of magnitude fewer collocation points than PINNs.

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

The paper presents the Neural Spectral Element Method as a way to solve multiphysics PDEs that are stiff due to electrochemical transport. It evaluates networks at fixed quadrature nodes and uses precomputed matrices for derivatives to create a deterministic loss that L-BFGS can optimize to very small residuals. A coordinate map addresses boundary layers, and a neural mortar couples domains. Tests on four Poisson-Nernst-Planck examples show high accuracy with far fewer points than an adaptive PINN approach, and this holds for both standard networks and Kolmogorov-Arnold networks.

Core claim

The Neural Spectral Element Method evaluates each network only at fixed Legendre-Gauss-Lobatto quadrature nodes and replaces all derivative calls with precomputed spectral differentiation matrices. The resulting deterministic loss enables limited-memory BFGS to reach residuals of 10^-9 to 10^-10. A Kosloff-Tal-Ezer coordinate map resolves electrochemical boundary layers, while a mesh-free neural mortar framework couples multi-element domains. On the four-example Poisson-Nernst-Planck benchmark, NSEM attains 10^-4 to 10^-7 relative pointwise error with two orders of magnitude fewer collocation points than the adaptive-resampling PINN baseline.

What carries the argument

Precomputed spectral differentiation matrices at Legendre-Gauss-Lobatto nodes combined with the Kosloff-Tal-Ezer coordinate map inside a neural mortar framework.

If this is right

  • Both tanh multilayer perceptrons and basis-aligned Legendre Kolmogorov-Arnold Networks achieve spectral accuracy within the same infrastructure.
  • The KAN backbone requires roughly half the Adam steps to enter the L-BFGS basin on the 1D PNP benchmark.
  • The method handles multi-element domains without a mesh through the neural mortar framework.
  • Accuracy reaches 10^-4 to 10^-7 relative pointwise error on the PNP benchmarks.

Where Pith is reading between the lines

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

  • Similar spectral-element ideas could reduce training costs for other stiff PDE problems in materials science.
  • The deterministic loss might allow integration with traditional solvers for hybrid approaches.
  • Extending the mortar framework could enable simulations of larger electrochemical systems like batteries.

Load-bearing premise

The precomputed spectral differentiation matrices and Kosloff-Tal-Ezer map resolve the electrochemical boundary layers accurately enough that the loss function has no quadrature or mapping errors preventing L-BFGS from reaching 10^-9 residuals.

What would settle it

Compare NSEM pointwise errors and required collocation points against a high-resolution traditional spectral element solver or an adaptive finite element method on the same four PNP examples.

Figures

Figures reproduced from arXiv: 2606.02335 by Conrard Giresse Tetsassi Feugmo, David Pankaczy.

Figure 1
Figure 1. Figure 1: Architecture of the NSEM framework. (a) The N Legendre–Gauss–Lobatto nodes on the reference element [−1, 1] with their quadrature weights wj drawn as bar heights; the network is evaluated only at these fixed nodes, making the loss in equation (1) deterministic. (b) The spectral first-derivative matrix D(1) assembled in barycentric form (25; equation (2)); D(2) = D(1)D(1) replaces a full autodiff Laplacian … view at source ↗
Figure 2
Figure 2. Figure 2: Spectral convergence of NSEM on three canonical benchmarks as the number of LGL nodes per element [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Stiff convection–diffusion benchmark −εu′′ + u ′ = 0 with ε = 10−2 , solved with a single KTE-mapped element (N = 32, α = 0.85). Left: NSEM solution (markers) overlaid on the exact boundary-layer profile (solid). Centre: pointwise error |uNSEM − uexact|, peaking at 5.4×10−3 in the centre of the layer. Right: training-loss convergence with the Adam-to-L-BFGS handover marked. Without the KTE map (α = 0) at t… view at source ↗
Figure 4
Figure 4. Figure 4: Multi-scale electrochemical benchmarks. Top: electric double layer next to a neutral surface (concentration BC at infinity). Bottom: charged wall with a fixed-charge BC at x = 0 and the extreme 5000:1 wall-to-bulk domain ratio that exposed the per-element-normalised loss-weight requirement. Both runs use a two-element KTE-stretched/uniform decomposition; analytic profiles overlaid for reference. Peak error… view at source ↗
Figure 5
Figure 5. Figure 5: 1D steady Poisson–Nernst–Planck: stiff leading coefficients [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 3D steady Poisson–Nernst–Planck: cubic domain [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 1D time-dependent PNP. Top two rows: cp, cn; NSEM solution (left), exact (centre), absolute error (right) at the final training loss. Bottom left: φ space–time map. Bottom right: loss landscape in the two leading PCA directions — a smooth bowl enabling L-BFGS convergence. 6 Convergence and ablation studies Sec. 5 demonstrated NSEM on the target electrochemical application, where the dominant design choices… view at source ↗
Figure 8
Figure 8. Figure 8: 2D time-dependent PNP problem on [−1, 1]2 × [0, 1], single tensor-product element, Nx = Ny = Nt = 12, 1728 collocation points total. Top row: cp (left) and cn (right) at t = 0.5. Bottom left: electric potential φ at t = 0.5. Bottom right: N-convergence — max error for all three fields vs the number of spatial nodes per axis, confirming spectral decay on the 2D+time problem. 6.2 KTE stretching ablation The … view at source ↗
Figure 9
Figure 9. Figure 9: Debye-length robustness of the 1D steady PNP solver. Peak pointwise errors for [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Spectral N-convergence on the Helmholtz problem at k = 10. Left: max |uNSEM − uexact| and relative L 2 error as functions of N, on a semilog axis. Right: log max |err| vs N, with an exponential fit valid for N > 2k = 20. The error rises slightly at N = 32 due to floating-point conditioning of the spectral D matrix. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Spectral N-convergence on the stiff convection–diffusion problem at ε = 10−2 , α = 0.85. Left: max and relative L 2 error. Right: log-linear view — empirical exponential rate σ≈0.20 per added node from N = 16 onwards [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: KTE coordinate-stretching ablation on the stiff convection–diffusion problem at [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Spectral N-convergence on the charged-wall benchmark (α = 0.95, Stern + diffuse layer, 5000:1 domain ratio). Exponential decay holds until N ≈ 48; saturation above N = 48 is numerical rounding from the ill-conditioned D-matrix, not a discretisation error [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Helmholtz wavenumber sweep k ∈ {5, 10, 20, 40}, NSEM versus a matched-parameter vanilla collocation PINN. Left: final maximum pointwise error. Right: wall-clock training time. NSEM achieves two to four orders of magnitude lower error at every k; the PINN baseline saturates at O(1) for k ≥ 20, consistent with the spectral-bias failure mode documented by [5] and [3]. confirm the consequence: PINN loss stops… view at source ↗
Figure 15
Figure 15. Figure 15: NSEM versus a vanilla collocation PINN [ [PITH_FULL_IMAGE:figures/full_fig_p017_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Allen–Cahn ε-sweep at fixed N = 48. Left: final maximum pointwise error. Right: final training loss. The loss remains near its floor for every ε, but the pointwise error grows once the layer width O(ε) drops below the LGL grid spacing, consistent with the predicted resolution limit at ε ∼ 1/N2 . 6.6 Allen–Cahn vs Cahn–Hilliard landscape Finally, we contrast the loss landscape of the Allen–Cahn benchmark a… view at source ↗
Figure 17
Figure 17. Figure 17: Loss landscape comparison between Allen–Cahn (second order, top row) and Cahn–Hilliard (fourth order, bottom row) at [PITH_FULL_IMAGE:figures/full_fig_p019_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Backbone comparison on the 1D steady PNP problem ( [PITH_FULL_IMAGE:figures/full_fig_p021_18.png] view at source ↗
read the original abstract

The Neural Spectral Element Method (NSEM) evaluates each network only at fixed Legendre-Gauss-Lobatto quadrature nodes and replaces all derivative calls with precomputed spectral differentiation matrices. The resulting deterministic loss enables limited-memory BFGS (L-BFGS) to reach residuals of 10^-9 to 10^-10. A Kosloff-Tal-Ezer coordinate map resolves electrochemical boundary layers, while a mesh-free neural mortar framework couples multi-element domains. On the four-example Poisson-Nernst-Planck (PNP) benchmark of Huang and co-workers, NSEM attains 10^-4 to 10^-7 relative pointwise error with two orders of magnitude fewer collocation points than the adaptive-resampling PINN baseline. Both a tanh multilayer perceptron (MLP) and a basis-aligned Legendre Kolmogorov-Arnold Network (KAN) backbone attain spectral accuracy within the same NSEM infrastructure, with the KAN requiring roughly half the Adam steps to enter the L-BFGS basin of attraction on the 1D PNP benchmark.

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 introduces the Neural Spectral Element Method (NSEM) for stiff multiphysics PDEs, with focus on Poisson-Nernst-Planck (PNP) electrochemical transport. NSEM evaluates networks only at fixed Legendre-Gauss-Lobatto (LGL) nodes, replaces derivatives with precomputed spectral differentiation matrices, applies a Kosloff-Tal-Ezer coordinate map to cluster points near boundary layers, and employs a mesh-free neural mortar framework for multi-element coupling. On four PNP benchmark problems from Huang et al., it reports 10^{-4} to 10^{-7} relative pointwise errors using two orders of magnitude fewer collocation points than an adaptive-resampling PINN baseline, with L-BFGS driving the deterministic loss to 10^{-9}–10^{-10} residuals. Both tanh-MLP and basis-aligned Legendre KAN backbones are shown to reach spectral accuracy within the same framework.

Significance. If the central numerical claims hold after verification of the mapped operators, NSEM would constitute a meaningful advance by delivering deterministic, high-accuracy neural solvers that combine spectral-element efficiency with neural flexibility for problems featuring sharp electrochemical layers. The explicit use of precomputed differentiation matrices enabling reliable L-BFGS convergence, together with the direct benchmark comparison showing substantial reduction in degrees of freedom, are concrete strengths that could influence subsequent work on multiphysics PINN variants.

major comments (2)
  1. [§2 (NSEM construction, coordinate-map paragraph)] §2 (NSEM construction, coordinate-map paragraph): the nonlinear Kosloff-Tal-Ezer map transforms the differential operators through the chain rule, producing variable-coefficient terms whose exact representation at the original fixed LGL nodes is not guaranteed by the unmapped precomputed differentiation matrices. The manuscript must demonstrate (via explicit error analysis or numerical verification) that any resulting quadrature or aliasing error lies below the claimed 10^{-9}–10^{-10} residual level; otherwise the deterministic-loss advantage and the reported pointwise accuracy cannot be taken as established.
  2. [Results section (PNP benchmark tables/figures)] Results section (PNP benchmark tables/figures): the two-order-of-magnitude reduction in collocation points relative to the adaptive PINN baseline is load-bearing for the efficiency claim, yet the manuscript does not report the precise number of points used by each method on each of the four examples or confirm that the pointwise error is evaluated on an identical, sufficiently dense reference grid independent of the training nodes.
minor comments (2)
  1. [§2.4] Notation for the neural mortar coupling should be introduced with an explicit equation showing how interface conditions are enforced in the loss; the current description is terse.
  2. [Figure captions] Figure captions for the error plots should state the exact norm and evaluation grid used for the reported relative pointwise errors.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and valuable feedback on our manuscript. We address the major comments point by point below, and we will incorporate revisions as indicated to improve the clarity and rigor of the presentation.

read point-by-point responses

  1. Referee: [§2 (NSEM construction, coordinate-map paragraph)] §2 (NSEM construction, coordinate-map paragraph): the nonlinear Kosloff-Tal-Ezer map transforms the differential operators through the chain rule, producing variable-coefficient terms whose exact representation at the original fixed LGL nodes is not guaranteed by the unmapped precomputed differentiation matrices. The manuscript must demonstrate (via explicit error analysis or numerical verification) that any resulting quadrature or aliasing error lies below the claimed 10^{-9}–10^{-10} residual level; otherwise the deterministic-loss advantage and the reported pointwise accuracy cannot be taken as established.

    Authors: We agree that an explicit demonstration is warranted for the mapped operators. In the revised manuscript we will add to §2 a short numerical verification subsection. This will compare the action of the chain-rule transformed differentiation matrices (evaluated at the fixed LGL nodes) against a reference spectral computation on a much finer grid for the specific Kosloff-Tal-Ezer stretching parameters used in the PNP benchmarks, confirming that the resulting operator error remains below 10^{-12} and therefore does not compromise the reported residual levels. revision: yes

  2. Referee: [Results section (PNP benchmark tables/figures)] Results section (PNP benchmark tables/figures): the two-order-of-magnitude reduction in collocation points relative to the adaptive PINN baseline is load-bearing for the efficiency claim, yet the manuscript does not report the precise number of points used by each method on each of the four examples or confirm that the pointwise error is evaluated on an identical, sufficiently dense reference grid independent of the training nodes.

    Authors: We concur that precise reporting is required to substantiate the efficiency claim. The revised manuscript will add a table in the Results section that lists, for each of the four PNP benchmarks, the exact number of collocation points employed by NSEM and by the adaptive-resampling PINN baseline. We will also state explicitly that all pointwise errors are computed on a fixed, independent reference grid of 100 000 uniformly spaced points that is denser than any training discretization used, and we will include pseudocode for the error-evaluation procedure. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on direct construction rather than self-referential fits or citations

full rationale

The paper describes NSEM via precomputed LGL differentiation matrices, a Kosloff-Tal-Ezer map, and neural mortar coupling as an explicit algorithmic construction. Performance figures (10^-9–10^-10 residuals, 10^-4–10^-7 pointwise error) are presented as empirical outcomes on external PNP benchmarks, not as quantities algebraically forced by the method's own parameters or loss terms. No self-citation chains, uniqueness theorems, or ansatzes are invoked to justify core choices. The derivation therefore remains self-contained against external benchmarks, consistent with the default non-circular outcome.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities beyond standard neural-network training and spectral quadrature; the Kosloff-Tal-Ezer map and neural mortar are described as standard tools repurposed rather than newly postulated.

pith-pipeline@v0.9.1-grok · 5733 in / 1249 out tokens · 22471 ms · 2026-06-28T13:30:59.205975+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

46 extracted references · 3 linked inside Pith

  1. [1]

    Raissi M, Perdikaris P and Karniadakis G E 2019Journal of Computational Physics378686–707

  2. [2]

    Karniadakis G E, Kevrekidis I G, Lu L, Perdikaris P, Wang S and Yang L 2021Nature Reviews Physics3422–440

  3. [3]

    Krishnapriyan A S, Gholami A, Zhe S, Kirby R and Mahoney M W 2021 Characterizing possible failure modes in physics-informed neural networksAdvances in Neural Information Processing Systems (NeurIPS)vol 34 pp 26548–26560 (Preprint2109.01050)

    arXiv 2021

  4. [4]

    Tancik M, Srinivasan P P, Mildenhall B, Fridovich-Keil S, Raghavan N, Singhal U, Ramamoorthi R, Barron J T and Ng R 2020 Fourier features let networks learn high-frequency functions in low-dimensional domains Advances in Neural Information Processing Systems (NeurIPS)vol 33 pp 7537–7547

    2020

  5. [5]

    Rahaman N, Baratin A, Arpit D, Draxler F, Lin M, Hamprecht F A, Bengio Y and Courville A 2019 On the spectral bias of neural networksInternational Conference on Machine Learning (ICML)(Preprint 1806.08734)

    Pith/arXiv arXiv 2019

  6. [6]

    Mishra S and Molinaro R 2023IMA Journal of Numerical Analysis431–43 (Preprint2006.16144)

    arXiv

  7. [7]

    Mishra S and Molinaro R 2022IMA Journal of Numerical Analysis42981–1022 (Preprint2007.01138) 23 Neural Spectral Element Methods for stiff multiphysics PDEs with electrochemical transport benchmarksA PREPRINT

    arXiv

  8. [8]

    Liu Z, Wang Y , Vaidya S, Ruehle F, Halverson J, Soljaˇci´c M, Hou T Y and Tegmark M 2024Transactions on Machine Learning ResearchISSN 2835-8856 (Preprint2404.19756)

    Pith/arXiv arXiv

  9. [9]

    Patera A T 1984Journal of Computational Physics54468–488

  10. [10]

    Karniadakis G E and Sherwin S J 2005Spectral/hp Element Methods for Computational Fluid Dynamics2nd ed (Oxford University Press)

  11. [11]

    Komatitsch D and Tromp J 1999Geophysical Journal International139806–822

  12. [12]

    Jagtap A D and Karniadakis G E 2020Communications in Computational Physics282002–2041 (Preprint 2004.02518)

    arXiv 2041

  13. [13]

    Moseley B, Markham A and Nissen-Meyer T 2023Advances in Computational Mathematics49(Preprint 2107.07871)

    arXiv

  14. [14]

    Kharazmi E, Zhang Z and Karniadakis G E 2021Computer Methods in Applied Mechanics and Engineering374 113547 (Preprint2003.05385)

    arXiv

  15. [15]

    Anandh T, Ghose D, Jain H and Ganesan S 2024SIAM Journal on Scientific Computing46A3881–A3908 (Preprint2404.12063)

    arXiv

  16. [16]

    Du Y , Chalapathi N and Krishnapriyan A 2024 Neural spectral methods: Self-supervised learning in the spectral domainInternational Conference on Learning Representations (ICLR)(Preprint2312.05225)

    arXiv 2024

  17. [17]

    Yu T, Qi Y , Oseledets I and Chen S 2025Journal of Computational and Applied Mathematics(Preprint 2408. 16414)

  18. [18]

    Shukla K, Zou Z, Chan C H, Pandey A, Wang Z and Karniadakis G E 2024Computer Methods in Applied Mechanics and Engineering433117498 (Preprint2407.21217)

    arXiv

  19. [19]

    Anagnostopoulos S J, Toscano J D, Stergiopulos N and Karniadakis G E 2024Computer Methods in Applied Mechanics and Engineering421116805 (Preprint2307.00379)

    arXiv

  20. [20]

    Wang S, Sankaran S and Perdikaris P 2024Computer Methods in Applied Mechanics and Engineering421116813 (Preprint2203.07404)

    arXiv

  21. [21]

    Wang Y , Sun J, Bai J, Anitescu C, Eshaghi M S, Zhuang X, Rabczuk T and Liu Y 2024Computer Methods in Applied Mechanics and Engineering(Preprint2406.11045)

    arXiv

  22. [22]

    Zhang Z, Xiong X, Zhang S, Wang W, Zhong Y , Yang C and Yang X 2025Expert Systems with Applications (Preprint2406.08992)

    Pith/arXiv arXiv

  23. [23]

    SS S and R G 2024 Chebyshev polynomial-based Kolmogorov–Arnold networks: An efficient architecture for nonlinear function approximation arXiv preprint arXiv:2405.07200 (Preprint2405.07200)

    arXiv 2024

  24. [24]

    Guo C, Sun L, Li S, Yuan Z and Wang C 2024 Physics-informed Kolmogorov–Arnold network with Chebyshev polynomials for fluid mechanics arXiv preprint arXiv:2411.04516 (Preprint2411.04516)

    arXiv 2024

  25. [25]

    Berrut J P and Trefethen L N 2004SIAM Review46501–517

  26. [26]

    Trefethen L N 2000Spectral Methods in MATLAB(Philadelphia: SIAM)

  27. [27]

    Boyd J P 2001Chebyshev and Fourier Spectral Methods2nd ed (New York: Dover)

  28. [28]

    Light J C, Hamilton I P and Lill J V 1985Journal of Chemical Physics821400

  29. [29]

    Colbert D T and Miller W H 1992Journal of Chemical Physics961982–1991

    1991

  30. [30]

    Wang Z, Hao S, Zhang Y and Zhang L 2024SIAM Journal on Numerical Analysis621702–1720 (Preprint 2405.14099)

    arXiv

  31. [31]

    Kosloff D and Tal-Ezer H 1993Journal of Computational Physics104457–469

  32. [32]

    XIed Brezis H and Lions J L (Pitman Research Notes in Mathematics) pp 13–51

    Bernardi C, Maday Y and Patera A T 1994 A new nonconforming approach to domain decomposition: The mortar element methodNonlinear Partial Differential Equations and Their Applications, Collège de France Seminar, vol. XIed Brezis H and Lions J L (Pitman Research Notes in Mathematics) pp 13–51

    1994

  33. [33]

    Lacour C and Maday Y 1997BIT Numerical Mathematics37720–738

  34. [34]

    Chen W, Howard A A and Stinis P 2025Journal of Computational Physics(Preprint 2407.01613) URL https://www.sciencedirect.com/science/article/pii/S0021999125005091

    arXiv

  35. [35]

    Liu D C and Nocedal J 1989Mathematical Programming45503–528

  36. [36]

    Byrd R H, Lu P, Nocedal J and Zhu C 1995SIAM Journal on Scientific Computing161190–1208 24 Neural Spectral Element Methods for stiff multiphysics PDEs with electrochemical transport benchmarksA PREPRINT

  37. [37]

    Doumèche N, Biau G and Boyer C 2023 On the convergence of PINNs arXiv preprint arXiv:2305.01240 (Preprint 2305.01240)

    arXiv 2023

  38. [38]

    Newman J and Thomas-Alyea K E 2004Electrochemical Systems3rd ed (Wiley)

  39. [39]

    Bard A J and Faulkner L R 2001Electrochemical Methods: Fundamentals and Applications2nd ed (Wiley)

  40. [40]

    Bazant M Z, Thornton K and Ajdari A 2004Physical Review E70021506 (Preprintcond-mat/0401118)

    arXiv

  41. [41]

    Huang X, Wang F, Zhang B and Liu H 2025Mathematics and Computers in Simulation237231–246 (Preprint 2402.01768)

    arXiv

  42. [42]

    Wang S, Yu X and Perdikaris P 2022Journal of Computational Physics449110768 (Preprint2007.14527)

    arXiv

  43. [43]

    Macdonald J R 1953Physical Review924–17

  44. [44]

    Lasia A 2014Electrochemical Impedance Spectroscopy and Its Applications2nd ed (Springer)

  45. [45]

    Orazem M E and Tribollet B 2017Electrochemical Impedance Spectroscopy2nd ed (Wiley (ECS Series))

  46. [46]

    Doyle M, Fuller T F and Newman J 1993Journal of The Electrochemical Society1401526–1533 25