{"id":"87d2331a-63d5-45f6-94d0-dc77dc788e52","arxiv_id":"2412.13200","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural-network solver with a learned bypass term and an integral conservation constraint solves the full nonlinear P2D battery model and estimates battery lengths from data.","lead":"This paper introduces two fixes that let physics-informed neural networks solve the full nonlinear pseudo-two-dimensional model of lithium-ion batteries, including the stiff Butler-Volmer reaction term. The result is a forward solver with about one percent error and an inverse solver that estimates battery geometry from simulated data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All quantitative claims rest on an undocumented reference solver; without solver details, parameter values, and a grid-convergence study, the reported errors cannot be interpreted.","rationale":"The paper's central claim is that the proposed bypassing-plus-conservation-law PINN method successfully solves the fully nonlinear P2D model, with forward errors of order 1e-2 and inverse parameter errors below 9%. The evidence for this claim is entirely relative to a reference solution, yet the reference is described in a single sentence with no solver details, no parameter values, and no convergence or external validation. This is the most load-bearing weakness because every reported number inherits the reference's accuracy: if the reference is biased, the forward errors are meaningless and the inverse problems are fit to biased synthetic data. The reader's weakest_assumption identifies exactly this gap, and I agree. I also note the 'first successful' claim appears to conflict with the cited work of Hassanaly et al. [25], which handled the nonlinear BV equation, but that is secondary to the validation gap. The ablation study is internally coherent and supports the necessity of both ingredients, but it does not compensate for the missing reference benchmark. The appropriate disposition is CONDITIONAL, as the reader already concluded: the paper could become acceptable if the authors provide the missing solver and parameter details and demonstrate reference accuracy. Therefore I do not change the verdict.","tokens_in":21034,"tokens_out":6278,"duration_ms":61673,"concrete_test":"Release a complete parameter table and the reference solver, or reproduce the reference with PyBaMM/LIONSIMBA using the same parameters. Then run a grid-convergence study: compare the current reference (Nx=50, Delta t=1) against a refined computation (e.g., Nx=200, Delta t=0.1, with a higher-order time integrator). If the coarse and refined references differ by more than the reported PINN relative errors (e.g., >1e-2 for psi), the reported accuracy is not validated. Additionally, re-run the B+SC forward and inverse experiments against the refined reference; if the relative errors change materially, the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The forward relative L2 errors in Table 2 and A.6, and the inverse parameter estimates in Tables 3–5, are all measured against a reference solution described only as: 'For reference solution, we used Delta t = 1 for time-stepping up to the final time tau = 3500, and Nx = 50 for spatial discretization' (Section 6). No spatial or temporal discretization scheme, no physical parameter table, no open-circuit potential functions, and no radial mesh details are given. There is no comparison to established battery solvers (e.g., PyBaMM, LIONSIMBA) and no grid-convergence study. If the coarse reference carries discretization errors comparable to or larger than the claimed ~1e-2 relative errors, then the reported forward accuracy is not established, and the inverse problem is trained on biased synthetic data. The 'first successful PINN simulation' assertion is also not supported relative to prior work: Hassanaly et al. [25] explicitly addressed the nonlinear BV equation in the P2D model with an additional division step, so the claim needs a narrow definition (e.g., 'without any linearization or multi-stage initialization') before it can be evaluated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two modifications to physics-informed neural network (PINN) training for the pseudo-two-dimensional (P2D) lithium-ion battery model with the fully nonlinear Butler–Volmer (BV) equation. First, a bypassing neural network βθ approximates the scaled overpotential inside the sinh term of the BV flux, which is shown to reduce the Hessian condition number of the loss. Second, a secondary conservation law for the solid-phase potential is added to the loss to prevent convergence to incorrect solutions that yield small residuals. The authors report forward relative L2 errors of 4.8e-3 to 7.2e-2 for various state variables at 0.5C, 1C, and 2C discharge rates, and inverse estimates of battery length and section lengths with relative errors below 9% even with 5% noise. They claim this is the first successful PINN simulation of a Li-ion battery incorporating the fully nonlinear BV equation.","tokens_in":21270,"tokens_out":5310,"duration_ms":46273,"significance":"If the reported accuracy is trustworthy, the work addresses two genuine obstacles in applying PINNs to battery models: the ill-conditioning caused by the exponential sinh term and the loss-accuracy mismatch caused by small reaction-rate coefficients. The ablation study (Table 2) cleanly shows that neither the bypassing term nor the secondary conservation law is redundant, and the toy problems in Sections 4 and 5 provide useful insight into the failure modes. The inverse results are internally consistent and show robustness to added noise. However, the significance is currently limited by the lack of external validation: all quantitative claims are measured against an undocumented reference solver, and the inverse experiments use synthetic data from that same solver. The novelty claim also needs careful qualification relative to prior work by Hassanaly et al. [25]. With additional reproducibility details and independent validation, the contribution would be valuable to the computational physics and battery modeling communities.","major_comments":[{"comment":"The reference solution is described only as 'For reference solution, we used Δt = 1 for time-stepping up to the final time τ = 3500, and Nx = 50 for spatial discretization.' No discretization scheme, physical parameter table, open-circuit potential functions, radial mesh details, or convergence study are provided. Because Table 2 and Table A.6 report relative L2 errors against this reference, and Tables 3–5 use it to generate the inverse-problem data, an inaccurate or biased reference would invalidate all quantitative conclusions. The authors should supply the full parameter set, the numerical scheme, and a grid-convergence study (or a comparison with an established solver such as PyBaMM or LIONSIMBA) demonstrating that the reference is converged to an accuracy well below the reported errors.","section":"Section 6"},{"comment":"The inverse problems are tested exclusively with synthetic data generated by the same reference solver used to validate the forward problem ('using the reference solution as our data'). This demonstrates self-consistency but not the ability to estimate parameters from real or independent data. Since the reference solver is also undocumented, the inverse errors in Tables 3–5 may partly reflect the solver's own discretization bias. The authors should add at least one test using data from an independent solver or a finer-grid reference, or an experimental dataset, and discuss how model mismatch would affect the parameter estimates.","section":"Section 6.2"},{"comment":"The claim that this is 'the first successful PINN simulation of a Li-ion battery incorporating the fully nonlinear BV equation (1)' needs qualification. The manuscript itself discusses Hassanaly et al. [25], who addressed the nonlinear BV equation in the P2D model with an additional division step, and Remark 1 acknowledges similar work for the single-particle model. As written, the claim is too broad and potentially misleading. It should be qualified, for example, as 'first single-step end-to-end PINN training without linearization or multi-stage initialization,' and the differences in accuracy, robustness, and computational cost relative to [25] should be quantified.","section":"Section 6.1 and Introduction"},{"comment":"The Hessian conditioning analysis is performed on a smaller network than the one used for the full P2D model, as the paper acknowledges ('memory constraints required us to perform this analysis with a smaller neural network compared to the one used for the full P2D model in Section 6.1'). Without specifying the reduced architecture, the included loss terms, and the collocation data, the reported condition numbers (2.017×10^108 without vs. 4.599×10^9 with bypassing) cannot be directly taken as evidence for the full model. Since the ablation study in Table 2 already demonstrates the necessity of the bypassing term, this issue is not fatal, but the analysis should be clearly labeled as illustrative or repeated on the actual model.","section":"Section 5.1 and Figure 6"}],"minor_comments":[{"comment":"The loss weights λ_BC, λ_Inter, λ_Lβ, λ_SC, and λ_data are never given numerical values. For reproducibility, the authors should provide the values used in the reported experiments.","section":"Section 3.2 and Section 5"},{"comment":"The derivation of the secondary conservation laws is stated to follow from integrating the solid-potential equation, but the relevant equation is not numbered (the text refers to '(3.1)'). Adding an explicit derivation with the boundary conditions would make the sign conventions and normalization clear.","section":"Section 5.2, Eq. (12)"},{"comment":"The 2C results show relative errors up to 7.15e-2 for φ and 4.69e-2 for cs; the text says the method 'accurately' predicts behavior, but a 7% error may warrant more careful wording.","section":"Appendix A, Table A.6"},{"comment":"There are several typos and formatting issues: 'can appears' in Section 3, 'instable' in Section 5.1, 'V anilla' and 'O urs' in Figure 9 labels, and the sentence 'inverse problems that might not be infeasible' in the conclusion. These should be corrected.","section":"Throughout"},{"comment":"The implementation note is useful, but the claimed speedup (14 vs. 9 iterations per second) is modest and would benefit from a benchmark with wall-clock time and hardware details beyond what is stated.","section":"Section 6.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of physics.comp-ph and the core ideas are promising, but the missing reference solver details and parameter table are a serious reproducibility problem that must be fixed before the results can be trusted. I would also encourage the authors to make code available, as this would substantially increase the value of the work. The novelty claim relative to Hassanaly et al. [25] needs careful rephrasing, as the current wording is likely to draw criticism from informed readers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this paper presents a genuinely useful trick for making PINNs handle the P2D lithium-ion battery model with the fully nonlinear Butler-Volmer equation, and it demonstrates that trick with a clean ablation study. The bypassing network for the overpotential inside sinh, plus the secondary conservation law for the solid-phase potential, are real contributions. The Hessian conditioning argument explains why the bypass helps, and the toy examples in Sections 4-5 make the failure mode concrete.\n\nWhat the paper does not do is give you any way to trust the reported numbers. The reference solution is described only as 'Delta t = 1 ... Nx = 50' with no discretization scheme, no physical parameter table, no OCP functions, no radial mesh, and no code. On a stiff model like P2D, Nx=50 is coarse enough that the reference itself could easily have 1e-2 errors, which is exactly the accuracy the method claims. There is no grid-convergence study and no comparison against PyBaMM, LIONSIMBA, or any established solver. The inverse results are all synthetic, measured against that same reference, so they inherit whatever bias it has. These are not fatal to the method, but they are fatal to the current evidence.\n\nThe 'first successful PINN simulation' claim also needs a qualifier. Hassanaly et al. [25] already handled the nonlinear BV equation in P2D, albeit with a division step. The claim as written is too broad.\n\nThe strengths are real. The ablation shows that neither the bypass nor the conservation law alone is sufficient; both are needed. The secondary conservation law argument is a neat answer to the loss-vs-accuracy problem for the small-flux regime. The inverse experiments are a reasonable proof of concept, and even with 5% noise the relative errors stay below ~9%.\n\nWho is this for? Anyone doing PINNs for battery models, or for stiff systems with embedded transcendental nonlinearities. It deserves a serious referee because the technical idea is interesting and the ablation is convincing. But the revision must supply a reproducible reference: full parameters, solver scheme, a convergence study, and ideally a comparison with an open-source battery solver. Without that, the quantitative claims are not interpretable. I would send it to review with that as the required revision.","headline":"A genuinely useful PINN stabilization trick for the P2D battery model, but the quantitative claims need a reproducible reference solver before they can be trusted.","tokens_in":21801,"tokens_out":2661,"would_cite":false,"duration_ms":23512,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","65M99"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper reports the first physics-informed neural network for the lithium-ion P2D model that keeps the fully nonlinear Butler-Volmer term, reaching forward errors of 0.5–7 percent and inverse length estimates within 9 percent.","keywords":["physics-informed neural networks","Butler-Volmer equation","pseudo-two-dimensional model","lithium-ion battery","forward problem","inverse problem","Hessian conditioning","conservation law"],"falsifier":"Recompute the P2D reference solution with an independent, grid-converged finite-volume or finite-difference solver using the same physical parameters, and re-run both the forward comparisons and the inverse estimation of $L$, $L_p$, and $L_n$ against that new reference; if the relative $L^2$ errors leave the reported ranges, or if removing the secondary conservation law lets a constant-shifted solid potential $\\psi_p + C$ pass the PINN residual test, the paper's central claims would fail.","tokens_in":20825,"feed_emoji":"🔋","tokens_out":16183,"duration_ms":124144,"temperature":0.7,"pith_summary":"This paper claims the first physics-informed neural network (PINN) that solves the full pseudo-two-dimensional (P2D) model of a lithium-ion battery while keeping the fully nonlinear Butler-Volmer equation, the stiffest and most failure-prone part of the battery model. The difficulty is that the Butler-Volmer flux contains a hyperbolic sine of the overpotential with a coefficient near $19.5$, which severely ill-conditions the PINN loss—training can start at losses around $10^{75}$—and also lets visibly wrong solutions pass with tiny residuals. The authors fix both failure modes with two additions: a bypassing term, a dedicated network for the argument of the hyperbolic sine that cuts the Hessian condition number from about $10^{108}$ to about $10^9$, and a secondary conservation law that pins the total ionic flux and rules out those wrong solutions. If the method works as reported, it predicts the four main battery state variables with relative $L^2$ errors of roughly 0.5–7 percent across discharge rates, and it recovers the total battery length and the section-length ratio from data with errors below 9 percent even when the data carry 5 percent noise.","feed_headline":"Neural nets crack the battery model's stiffest equation","feed_subtitle":"A bypass network plus an ionic-flux constraint make the full nonlinear P2D battery model trainable.","key_machinery":"The central object is the Butler-Volmer reaction flux $j_i = 2k\\,(c_{s,\\max}-c_{s,\\mathrm{surf}})^{0.5}c_{s,\\mathrm{surf}}^{0.5}c_i^{0.5}\\sinh(19.5\\,\\eta_i)$, the algebraic coupling through which all the P2D equations exchange information. Two devices carry the argument. The first is the bypassing term: a dedicated neural network $\\beta_\\theta$ replaces the argument of the hyperbolic sine, with a penalty $L_\\beta = \\int(\\beta_\\theta - 19.5\\,\\eta)^2\\,dx$ tying it to the true overpotential, which removes the exponential sensitivity and collapses the Hessian spectrum. The second is the secondary conservation law: the integral identities $(L_p a_p F/I_{\\mathrm{app}})\\int_0^{L_p} j_p\\,dx = 1$ and $(L_n a_n F/I_{\\mathrm{app}})\\int_{L_p+L_s}^{1} j_n\\,dx = -1$, derived by integrating the solid-potential equation, fix the total ionic flux and rule out shifted-potential impostors that the residual alone accepts. Around these sit the scaffolding of non-dimensionalization, hard-constrained initial conditions, and domain decomposition with soft interface penalties.","core_discovery":"On its own terms, the paper's central claim is that the P2D model becomes trainable by PINNs without discarding the nonlinear Butler-Volmer kinetics, and that this is the first successful PINN simulation of a lithium-ion battery with the fully nonlinear Butler-Volmer equation. The two ingredients are a bypassing network $\\beta_\\theta$ that absorbs the argument $19.5\\,\\eta$ inside the hyperbolic sine, isolating the stiffest nonlinearity and dropping the Hessian condition number from about $2\\times10^{108}$ to about $5\\times10^9$, and secondary conservation laws that fix the integrated ionic flux in each electrode, eliminating the family of constant-shifted solid-potential solutions that the PDE residual alone cannot distinguish. With both ingredients, the trained network reproduces the reference solid concentration, liquid concentration, solid potential, and liquid potential at 1C, 0.5C, and 2C discharge, and the same representation supports inverse estimation of the total battery length and the electrode-to-separator length ratio, with relative errors below 9 percent even under 5 percent observational noise. The ablation study shows that each ingredient is necessary: the bypassing term alone yields a tiny loss but inaccurate solutions, while the conservation law alone does not even stabilize training.","pith_inferences":["The bypassing recipe is generic: any stiff or near-singular algebraic coupling inside a PDE system could be isolated with its own network plus a consistency penalty, so the method likely transfers to other electrochemical, combustion, or reaction-diffusion models.","Because the reference solution used for validation and inverse data is described only by its time step and grid size, with no solver specified and no grid-convergence or experimental check, the quantitative error claims are best read as upper bounds until an independent converged reference is used; the qualitative two-failure-mode diagnosis does not depend on that reference.","The stability of the estimates under 5 percent noise suggests the architecture could be pointed at real cell measurements—voltage and current traces—to track capacity fade or parameter drift over aging, an application the authors mention as future work but do not test.","The condition-number drop offers a practical selection rule for when bypassing is needed: if the Hessian spectrum of a PINN residual spans more than about nine orders of magnitude, isolating the stiffest term is a reasonable first move."],"forward_implications":["The full P2D model with nonlinear Butler-Volmer kinetics becomes solvable by PINNs in a single training pass, without the single-particle approximation, linearized kinetics, or two-step pretraining used in earlier work.","The reported error levels hold across discharge rates from 0.5C to 2C; the worst case is the liquid potential at 2C at about $7\\times10^{-2}$ relative $L^2$ error, and most variables sit near $10^{-3}$–$10^{-2}$.","The same framework estimates battery geometry—total length, and the ratio of electrode and separator lengths—with relative errors under 9 percent, and the estimates remain in that range when the observational data carry 5 percent Gaussian noise.","The Hessian analysis gives a quantitative explanation of why vanilla PINNs fail here: without bypassing, the condition number is on the order of $10^{108}$ and the initial loss reaches $10^{75}$, while with bypassing the condition number falls to about $10^9$.","The ablation study shows the two ingredients are complementary: the bypassing term restores trainability, the conservation law restores accuracy, and neither alone suffices."],"supporting_citations":[{"why":"Defines the pseudo-two-dimensional battery model whose governing equations the PINN must satisfy.","marker":"[10]"},{"why":"Finite-difference P2D solver whose device of treating overpotential and flux as independent unknowns directly inspires the bypassing term.","marker":"[14]"},{"why":"Introduces the physics-informed neural network framework that the two proposed strategies extend.","marker":"[19]"},{"why":"Earlier PINN attempt at the P2D model that needed a two-step training strategy and a linearized Butler-Volmer equation, the approach the paper aims to supersede.","marker":"[25]"},{"why":"Catalogues PINN failure modes that frame the observed loss-versus-accuracy discrepancy in the P2D setting.","marker":"[32]"},{"why":"Prior PINN study of the single-particle model with the nonlinear Butler-Volmer equation whose two-step remedy the end-to-end strategy is designed to replace.","marker":"[36]"},{"why":"Analyzes Hessian and gradient pathologies in PINN training; its conditioning analysis underlies the condition-number comparison that validates the bypassing term.","marker":"[37]"}],"fun_headline_variants":["Bypass term and flux law make battery PINNs trainable","Neural nets solve full nonlinear battery model with two tricks","Stable PINN design for lithium-ion battery simulations","Two fixes let neural nets handle stiff battery equations","Battery model's toughest equation now yields to neural nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything reported is measured against a reference solution that the paper computes with time step 1 and 50 spatial points but does not describe further—no solver details, no parameter table, no grid-convergence study, and no experimental comparison—so if that reference is biased or wrong, the forward errors and inverse parameter estimates do not establish that the method works.","fun_headline_variants_meta":{"raw":{"variants":["Bypass term and flux law make battery PINNs trainable","Neural nets solve full nonlinear battery model with two tricks","Stable PINN design for lithium-ion battery simulations","Two fixes let neural nets handle stiff battery equations","Battery model's toughest equation now yields to neural nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000392,"raw_usage":{"total_tokens":2078,"prompt_tokens":981,"completion_tokens":1097,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":1017}},"tokens_in":597,"tokens_out":1097,"duration_ms":10140,"temperature":1.0,"reasoning_tokens":1017,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:43:00.535314+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the P2D reference solution with an independent, grid-converged finite-volume or finite-difference solver using the same physical parameters, and re-run both the forward comparisons and the inverse estimation of $L$, $L_p$, and $L_n$ against that new reference; if the relative $L^2$ errors leave the reported ranges, or if removing the secondary conservation law lets a constant-shifted solid potential $\\psi_p + C$ pass the PINN residual test, the paper's central claims would fail.","supporting_citations":[{"cited_title":"A fast solver for the pseudo-two-dimensional model of lithium-ion batteries","cited_arxiv_id":"2111.09251","evidence_quote":"Finite-difference P2D solver whose device of treating overpotential and flux as independent unknowns directly inspires the bypassing term."},{"cited_title":"PINN surrogate of Li-ion battery models for parameter inference. Part II: Regularization and application of the pseudo-2D model","cited_arxiv_id":"2312.17336","evidence_quote":"Earlier PINN attempt at the P2D model that needed a two-step training strategy and a linearized Butler-Volmer equation, the approach the paper aims to supersede."},{"cited_title":"Krishnapriyan, A","cited_arxiv_id":null,"evidence_quote":"Catalogues PINN failure modes that frame the observed loss-versus-accuracy discrepancy in the P2D setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Analyzes Hessian and gradient pathologies in PINN training; its conditioning analysis underlies the condition-number comparison that validates the bypassing term."}],"review_version":1}