{"id":"4c59f0d7-7a2e-4de0-8d55-0b433c724c54","arxiv_id":"2507.00353","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid battery model that learns sparse corrections to a reduced-order physics model reduces unseen-cycle voltage MSE by up to 46%, with conformal intervals above 90% coverage.","lead":"This paper adds a data-driven correction to a physics-based battery model, cutting voltage prediction error by up to 46% on new driving cycles. It also packages the predictions with interval estimates that stayed around 97% accurate, useful for battery management and safety.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 45.96% MSER claim may be inflated: Eq. 8 uses the true previous voltage error e[k] as a regressor, so the hybrid model sees the measured voltage Vexp[k] at test time, unlike the ESPM baseline.","rationale":"The reader correctly identified Eq. 8 as the weakest spot, but focused on the representational adequacy of the sparse library rather than the more fundamental issue: the regressor e[k] is the measured voltage error, not a self-generated prediction. This makes the reported MSER an unfair comparison against the physics-only ESPM, which does not receive the previous measured voltage. The conformal prediction results are also overstated as 'theoretically guaranteed' when SPCI's guarantee is asymptotic and conditional, but that is secondary. The method itself has merit as a one-step-ahead residual correction for real-time voltage estimation, and the ensemble/conformal apparatus is reasonable. However, the central quantitative claim of up to 46% MSER on unseen data needs to be re-evaluated against a persistence baseline and in a recursive evaluation mode. If the recursive performance collapses, the paper's contribution should be repositioned as a filtering/correction method rather than a standalone predictive model. The verdict should remain conditional, but with an additional explicit condition that the evaluation protocol be corrected and compared against autoregressive baselines.","tokens_in":15250,"tokens_out":5344,"duration_ms":61847,"concrete_test":"Re-run the test-set evaluation with two additional baselines on the same SC03/JC/NEDC data: (i) persistence e[k+1] = e[k] (equivalently Vh[k+1] = VESPM[k+1] + e[k]), and (ii) the AESI I model in recursive/autonomous mode, where the predicted error ˆe[k+1] from Eq. 8 is fed back as the input e[k+1] instead of the measured error, so no Vexp is used after time 0. If persistence matches or beats AESI I's MSER, or if recursive AESI I's MSER drops below roughly 15%, the headline claim should be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline improvement (Table I: AESI I 45.96% MSER) is computed by comparing ESPM voltage predictions with hybrid predictions from Vh[k+1] = VESPM[k+1] + Θ(e[k], I[k], cs,p[k], …)ξ (Eqs. 7–8). In both training and test evaluation, e[k] is the actual voltage error Vexp[k] − VESPM[k] computed from the measured voltage, not a predicted or filtered value. Thus, at every test time step the hybrid model has access to the true battery voltage from the previous sample, whereas the ESPM baseline does not. The discovered library includes T1(e[k]) (Fig. 9, rank 8), i.e., an autoregressive term; since voltage errors are smooth on a 10 Hz grid, a predictor that simply carries the last error forward can already capture a large fraction of the residual. The paper never compares against a persistence or other autoregressive baseline, nor does it evaluate a closed-loop mode in which predicted errors are fed back recursively. Consequently, the 45.96% test-set MSER conflates genuine learned error dynamics with the trivial predictability of the previous error, and the generalization claim in the abstract is not established by the reported experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Adaptive Ensemble Sparse Identification (AESI) framework that augments an Extended Single Particle Model (ESPM) of a lithium-ion cell with a data-driven correction for voltage error dynamics. The correction is learned from moving-block-bootstrap ensembles of sparse regressors, with either bagging or stability selection, and the framework further wraps the point predictions in Sequential Predictive Conformal Inference (SPCI) intervals. The authors report mean squared error reductions up to 45.96% on unseen driving cycles (SC03, JC, NEDC) relative to the ESPM baseline, and empirical conformal coverage of 96.85% and 97.41% for the two ensemble variants. The evaluation uses a 10 Hz experimental dataset from a 30 Ah NMC cell and includes a Welch ANOVA on squared residuals.","tokens_in":15582,"tokens_out":4844,"duration_ms":61740,"significance":"If the accuracy claim is established, the paper is a useful step toward low-cost, interpretable correction of physics-based battery models with rigorous uncertainty quantification. The combination of ensemble sparse identification with conformal prediction is sensible, and the use of genuinely held-out driving cycles rather than a random split is a strength. The work also compares against a tuned SINDy-C baseline and gives a clear algorithmic description. However, the central evaluation currently conflates the value of the learned sparse dynamics with the trivial information contained in the current measured voltage error, and the uncertainty quantification is reported only for one-step-ahead, oracle-informed prediction. These issues must be resolved before the headline claims can be accepted.","major_comments":[{"comment":"The hybrid predictor in Eq. (7) uses e[k] = Vexp[k] - VESPM[k] as a regressor, so at every test step the model has access to the measured voltage from the previous sample, whereas the ESPM baseline does not. The discovered library contains T1(e[k]) (Fig. 9, rank 8), and voltage errors on a 10 Hz grid are smooth, so a large part of the reported 45.96% MSER may come from simply carrying the last error forward rather than from the learned sparse error dynamics. The paper does not compare against a persistence baseline (e_hat[k+1] = e[k]) or a linear autoregressive baseline, nor does it evaluate a closed-loop mode in which predicted errors are fed back recursively. Please add these baselines and report MSER for recursive prediction to substantiate the generalization claim in the abstract.","section":"III, Eq. (8) and V, Table I"},{"comment":"The SPCI intervals are updated online with the true residual at each test step (Algorithm 1, lines 5-6), so the reported coverage of 96.85% is a one-step-ahead coverage for a predictor that is corrected by the current measurement, not coverage for the hybrid model used as a standalone forward simulator. This limitation should be stated explicitly. If the intended deployment is recursive or multi-step prediction, the paper should provide coverage and interval-width results for that mode, since the current numbers do not quantify uncertainty for the simulated trajectory.","section":"IV, Algorithm 1 and Table II"},{"comment":"The central MSER figures are point estimates over one concatenated test set without confidence intervals, and the Welch ANOVA in Fig. 8 reports a small effect size (omega-squared = 0.04). Because the test set is a concatenation of three different cycles (SC03, JC, NEDC), the authors should report per-cycle MSER values and bootstrap or other uncertainty intervals for the 45.96% figure, to show that the improvement is not driven by a particular segment and to give the reader a sense of the precision of the headline number.","section":"V, Table I and Fig. 8"}],"minor_comments":[{"comment":"The terminal voltage equation uses Rc, but the surrounding text refers to R0; please use consistent notation for the ohmic resistance.","section":"II, Eq. (6)"},{"comment":"The expression for the expected number of unique points per MBB sample is stated without derivation and appears to ignore the block structure of the resampling; please clarify the approximation or provide a citation that supports it.","section":"III, Eq. (15)"},{"comment":"The reported confidence interval for omega-squared is shown as [0.04, 1.00], which is not plausible for an effect size of 0.04; this is likely a computational or reporting error and should be corrected.","section":"V, Fig. 8"},{"comment":"The SVD-based feature importance in Eqs. (34)-(38) is not standard and the coefficients u_i,k are not identified if U is rank-deficient; please provide a reference or a numerical validation of the importance scores.","section":"V.A.1"},{"comment":"The notation with \\hat Q_t(\\hat \\beta) and \\hat Q_t(1 - \\alpha + \\hat \\beta) is confusing; please state explicitly which quantile levels are used and how \\hat \\beta is calibrated.","section":"IV, Eq. (31)"},{"comment":"The phrase '5C charging pulse' should define the C-rate convention and clarify why 5C was chosen, as this is an unusually high charge rate for a 30 Ah cell.","section":"III.C"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for IEEE Transactions on Transportation Electrification, and the methodology is a reasonable extension of the authors' earlier SINDy-C work. The main obstacle is not the method itself but the evaluation protocol: the headline MSER and coverage numbers are obtained in an oracle-informed one-step-ahead setting, which is not clearly disclosed in the abstract. If the authors add persistence and autoregressive baselines, report closed-loop results, and provide uncertainty intervals for the MSER estimates, the paper could become acceptable. I do not see a novelty disclosure problem; prior work [22] is the authors' own and is cited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent engineering paper, but its headline number is probably better than its true merit. The 46% MSE reduction is an open-loop, one-step-ahead evaluation that conditions on the measured voltage at every previous step, and the model includes a T1(e[k]) term. Without a persistence baseline, you can't tell how much of that reduction is just 'the error at k+1 looks like the error at k.' The authors owe the reader that comparison.\n\nWhat's genuinely new: they've assembled sparse regression, moving block bootstrap ensembles, evolutionary library selection, and SPCI conformal prediction into one pipeline for correcting a reduced-order Li-ion battery model. The experimental design is sensible — training on UDDS/HWFET/US06/WLTC, validation on Artemis Urban, test on SC03/JC/NEDC. The Welch ANOVA on squared residuals is a reasonable check that the hybrid models are better than ESPM alone, even if the effect size is small.\n\nWhere it's soft:\n\n- The look-ahead issue. Eq. 8 feeds the true voltage error e[k] into the predictor at every time step. That's fine for real-time monitoring with sensor feedback, but it's not a 'prediction' in the sense of a forward simulation. The paper should either evaluate a closed-loop recursive mode, or at least compare against e[k] itself as a baseline. Given that T1(e[k]) ranks 8th in feature importance, the persistence baseline could well eat up a large chunk of the claimed reduction.\n\n- The conformal claim is overstated. 'Statistically valid prediction intervals' is too strong for an empirical coverage of 96.85% on one test set. SPCI provides asymptotic conditional coverage under assumptions; reporting empirical coverage with a nominal 90% means the intervals are wide (average width 3.4e-4 V) and the method is conservative. That's fine, but the language should be toned down.\n\n- MSER is a single point estimate with no uncertainty quantification. A bootstrap or a confidence interval would help.\n\n- No code or data. For a methods paper, that's a real limitation for independent verification.\n\nWho should read it: people working on physics-ML hybrid battery models, especially those interested in conformal prediction for time series. It's worth a serious referee — the authors have done careful experiments and the method is clearly described — but it needs revision to address the baselines and temper the claims. I'd send it to review with 'major revision' in mind.","headline":"A competent hybrid-modelling paper whose headline 46% MSE reduction is likely inflated by an autoregressive term that uses the true previous voltage error; the authors need a persistence baseline.","tokens_in":16078,"tokens_out":3225,"would_cite":true,"duration_ms":35310,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid physics-and-sparse-learning model cuts Li-ion voltage prediction error by up to 46% on unseen drive cycles, with conformal intervals that stay valid.","keywords":["lithium-ion battery","hybrid modeling","sparse identification","ensemble learning","conformal prediction","voltage prediction","reduced-order model","uncertainty quantification"],"falsifier":"Compute, on the held-out cycles, the projection residual of the true error $e[k+1]$ onto the selected library terms: if the best possible sparse linear fit from Eq. 8 leaves a large residual compared with the reported $45.96\\%$ MSE reduction, then the improvement is not a true learned dynamics but an artifact of the training set. Alternatively, run the trained AESI correction on a long, sustained high-C-rate pulse profile or a different cell chemistry: if the empirical coverage of the SPCI intervals falls well below the nominal $90\\%$ level, the reliability claim fails outside the tested distribution.","tokens_in":15054,"feed_emoji":"🔋","tokens_out":5983,"duration_ms":62491,"temperature":0.7,"pith_summary":"The paper tries to establish that a reduced-order electrochemical model of a lithium-ion cell can be brought much closer to experimental voltage data by adding a sparse, data-driven correction that is learned once and then runs alongside the physics. The correction is not a static offset: it models the time evolution of the voltage error as a function of the previous error, current, and estimated internal concentrations. On driving cycles the model never trained on, the hybrid predictor cuts mean squared voltage error by up to 46 percent relative to the physics model alone. The same framework supplies prediction intervals with stated coverage guarantees, reaching 96.85 percent empirical coverage in the bagged version.","feed_headline":"Sparse correction cuts battery voltage error by 46 percent","feed_subtitle":"A physics model plus an ensemble-learned correction predicts voltage better on unseen cycles, with 96.85% interval coverage.","key_machinery":"The load-bearing object is the adaptive ensemble sparse identification (AESI) correction, defined by the discrete error model $\\hat{e}[k+1] = \\Theta(e[k], I[k], c_{s,p}[k], c_{s,n}[k], c_{e,0}[k], c_{e,L}[k])\\,\\xi$, where $\\Theta$ is a library of Chebyshev polynomials of the state and inputs plus trigonometric and hyperbolic terms, and $\\xi$ is a sparse coefficient vector. An evolutionary algorithm selects a subset of the library and tunes the regularization and sparsification parameters; moving-block bootstrap resampling creates 100 training sets, and the top 10 percent of models by out-of-bag error are aggregated either by averaging coefficients (bagging) or by refitting on terms whose inclusion probability exceeds a threshold (stability selection). Uncertainty is handled by sequential predictive conformal inference (SPCI), which feeds the most recent 200 residuals into a quantile regression forest to produce adaptive prediction intervals.","core_discovery":"The central claim is that the residual voltage error of the Extended Single Particle Model obeys a low-dimensional sparse dynamics that can be identified from data and used for one-step-ahead correction. Specifically, the paper asserts that the error at time $k+1$ is well approximated by a sparse linear combination of Chebyshev and trigonometric basis functions of the current error, the current, and the electrolyte and solid-phase concentrations at time $k$, with coefficients found by thresholded ridge regression on bootstrap resamples. Two ensemble strategies are compared: bagging, which averages coefficients across selected sparse models, and stability selection, which refits on the terms most often selected. The paper reports that the bagged hybrid model achieves $45.96\\%$ mean squared error reduction on the held-out SC03, JC, and NEDC cycles, and that the conformal intervals built on these corrections cover the true voltage error $96.85\\%$ of the time at a nominal $90\\%$ level.","pith_inferences":["If the same AESI recipe is applied to other reduced-order models, the library must be re-derived for the new error dynamics; Chebyshev-and-trigonometric dictionaries are not universal.","The coverage guarantees hold for the tested drive-cycle distribution; a stress test with a long sustained high-C-rate pulse or a different cell chemistry would reveal whether the learned error model transfers beyond the training envelope.","The discovered sparse terms are plausible correlates of current and transport effects, but the paper's own caveat applies: the error correction is not uniquely identifiable, so the terms should not be read as a causal mechanism.","Extending the library with aging features could let the same framework track state-of-health, as the authors suggest for future work; one testable version is to re-estimate $\\xi$ periodically as capacity fades."],"forward_implications":["The hybrid voltage predictor can be deployed in battery management systems at a small computational overhead over the ESPM, since the correction is a fixed linear combination of precomputed basis functions.","Bagging generalizes better than stability selection on unseen cycles, suggesting that variance reduction matters more than sparsity when the training data are limited and temporally dependent.","Conformal intervals from SPCI give a practical certificate of reliability for voltage prediction, not just a point estimate.","The learned terms can be ranked by singular value decomposition to point at which physical variables drive the model error, such as current and electrolyte boundary concentrations."],"supporting_citations":[{"why":"Provides the ESPM equations, boundary conditions, and electrolyte dynamics that the hybrid model builds on.","marker":"[5]"},{"why":"Defines the sparse identification of nonlinear dynamics formulation that Eq. 8 extends to error dynamics.","marker":"[29]"},{"why":"The earlier hybrid sparse identification method that AESI improves with ensemble learning and conformal prediction.","marker":"[22]"},{"why":"Provides the Ensemble-SINDy strategy for bootstrap resampling, top-model selection, and bagging/stability aggregation.","marker":"[34]"},{"why":"Supplies stability selection, the inclusion-probability criterion used by the AESI II variant.","marker":"[36]"},{"why":"Defines sequential predictive conformal inference, the method used to build adaptive prediction intervals.","marker":"[38]"},{"why":"Supplies quantile regression forests, the model that estimates conditional residual quantiles inside SPCI.","marker":"[40]"},{"why":"Galerkin projection used to approximate the PDE concentrations in the ESPM, providing the latent states that enter the library.","marker":"[28]"}],"fun_headline_variants":["Sparse correction cuts battery voltage error by 46%","Battery model gains 46% accuracy with sparse ensemble","Conformal prediction validates battery voltage model boost","Hybrid model trims voltage error, hits 96.85% coverage","Sparse learning sharpens battery voltage forecasts by 46%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the voltage error at the next time step can be expressed as a sparse linear combination of the chosen Chebyshev and trigonometric functions of the current error, current, and internal concentrations, and that this expression learned on the training cycles still holds on the held-out SC03, JC, and NEDC cycles.","fun_headline_variants_meta":{"raw":{"variants":["Sparse correction cuts battery voltage error by 46%","Battery model gains 46% accuracy with sparse ensemble","Conformal prediction validates battery voltage model boost","Hybrid model trims voltage error, hits 96.85% coverage","Sparse learning sharpens battery voltage forecasts by 46%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1270,"prompt_tokens":964,"completion_tokens":306,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":223}},"tokens_in":580,"tokens_out":306,"duration_ms":4092,"temperature":1.0,"reasoning_tokens":223,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:18:33.878630+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, on the held-out cycles, the projection residual of the true error $e[k+1]$ onto the selected library terms: if the best possible sparse linear fit from Eq. 8 leaves a large residual compared with the reported $45.96\\%$ MSE reduction, then the improvement is not a true learned dynamics but an artifact of the training set. Alternatively, run the trained AESI correction on a long, sustained high-C-rate pulse profile or a different cell chemistry: if the empirical coverage of the SPCI intervals falls well below the nominal $90\\%$ level, the reliability claim fails outside the tested distribution.","supporting_citations":[{"cited_title":"Modeling of li-ion cells for fast simulation of high c-rate and low temperature operations,","cited_arxiv_id":null,"evidence_quote":"Provides the ESPM equations, boundary conditions, and electrolyte dynamics that the hybrid model builds on."},{"cited_title":"Improving Low-Fidelity Models of Li-ion Batteries via Hybrid Sparse Identification of Nonlinear Dynamics","cited_arxiv_id":"2411.12935","evidence_quote":"The earlier hybrid sparse identification method that AESI improves with ensemble learning and conformal prediction."},{"cited_title":"Ensemble- sindy: Robust sparse model discovery in the low-data, high-noise limit, with active learning and control,","cited_arxiv_id":null,"evidence_quote":"Provides the Ensemble-SINDy strategy for bootstrap resampling, top-model selection, and bagging/stability aggregation."},{"cited_title":"Stability selection,","cited_arxiv_id":null,"evidence_quote":"Supplies stability selection, the inclusion-probability criterion used by the AESI II variant."},{"cited_title":"Sequential predictive conformal inference for time series,","cited_arxiv_id":null,"evidence_quote":"Defines sequential predictive conformal inference, the method used to build adaptive prediction intervals."},{"cited_title":"Quantile regression forests","cited_arxiv_id":null,"evidence_quote":"Supplies quantile regression forests, the model that estimates conditional residual quantiles inside SPCI."},{"cited_title":"A reduced-order electrochemical model of li-ion batteries for control and estimation applications,","cited_arxiv_id":null,"evidence_quote":"Galerkin projection used to approximate the PDE concentrations in the ESPM, providing the latent states that enter the library."}],"review_version":1}