{"id":"b96aec6b-23c4-4fc3-92a6-91b39dfb4df1","arxiv_id":"2512.10055","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SOBER and BASQ, two previously published Bayesian algorithms, are adapted for battery simulators and demonstrated on six case studies, including impedance-based model selection.","lead":"A battery-science primer wraps two existing Bayesian algorithms (SOBER, BASQ) into a PyBOP toolbox for model parameterisation and evidence-based model selection. Six worked examples — battery sizing, voltage relaxation, knee-point detection, inverse surrogates, electrolyte identifiability, and impedance model choice — make the case that this accelerates data-informed battery model development.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Key model-selection result rests on an unexplained post hoc rescaling of DFN evidence by sqrt(2π)^3; without it the impedance rankings in Table 6 reverse, so the central claim of accurate automatic model sorting is not yet supported.","rationale":"Good-faith reading: the paper is a tutorial/showcase; its central contribution is to demonstrate that SOBER+BASQ enable fast, uncertainty-quantified parameterisation and reliable evidence-based model selection. The parameterisation examples (voltage relaxation, inverse surrogates) are plausible and the code is available. The load-bearing step for the 'sort through models' claim is the impedance model comparison in Example 6, because it is the only real-data head-to-head comparison of physics-based models and it is used in the conclusion. There the reported rankings are not robust: they depend on multiplying DFN evidence by sqrt(2π)^3, a constant not derived from the evidence integral. The reader's weakest assumption was fixed evaluation budgets, but even a perfectly converged BASQ would not fix this, because the reported quantity is no longer the evidence. I therefore identify the rescaling as the most load-bearing concern. It is internally inconsistent with the paper's own definition of evidence (Eq. 10-11). A single computational check—removing the factor or computing Laplace evidence with actual covariances—would settle whether the DFN preference is real. My recommendation is to keep the reader's CONDITIONAL verdict: the paper should not be accepted until the rescaling is justified or removed and the rankings recomputed.","tokens_in":23272,"tokens_out":11420,"duration_ms":118742,"concrete_test":"Recompute Table 6(c,d) using the raw BASQ log-evidence values (no sqrt(2π)^3 factor) and, separately, using a Laplace approximation with the actual posterior covariance matrices and prior densities from the SOBER runs. If the DFN/SPM ratios for 42µm and 80µm drop below 1, or if their one-sigma intervals overlap 1, the claimed preference for DFN on thick electrodes is an artifact of the rescaling. The code in PyBOP/SOBER should allow this directly from saved posterior samples.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Example 6 is the paper's real-data validation of the central claim that SOBER+BASQ automatically and accurately sort models. To compare SPM (3 parameters) with DFN (6 parameters), the authors multiply all DFN evidences by sqrt(2π)^3 ≈ 15.75, calling it a 'correction for the integral penalty for the extra 3 fit parameters' (Section 'Example 6: Impedance model selection', Table 6). This is not part of the evidence integral p(D|M)=∫p(D|θ,M)p(θ)dθ that BASQ is designed to compute; that integral already contains any Occam/complexity penalty via the prior and the integrated posterior. The multiplier is a constant favoring the higher-dimensional model, independent of data and prior ranges. It is decisive: in Table 6(c), the raw DFN/SPM mean-evidence ratios are roughly 0.03 (17µm), 0.05 (42µm), and 0.49 (80µm); after multiplication they become 0.46, 0.78, and 7.7 (table reports 51%, 117%, 792%). Thus the paper's headline conclusion—that DFN is preferred for thicker electrodes—is an artifact of this arbitrary rescaling. At minimum, the reported 'DFN (corr.)' values are not BASQ evidences; they are BASQ outputs altered by a hand-chosen constant. Without a derivation of this factor from the actual posterior covariances and prior volumes, the model-selection conclusions in the key validation are unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a tutorial-style application paper introducing the Bayesian optimisation algorithm SOBER and the Bayesian quadrature algorithm BASQ to the battery modelling community, with the software wrapped into the PyBOP package. A brief derivation of Bayesian inference and likelihood-free inference is given, followed by six worked examples: battery sizing, voltage relaxation parameterisation, knee-point model selection from degradation data, inverse modelling via a global inverse surrogate, electrolyte transport identifiability analysis, and impedance-based model selection between SPM and DFN models. The central claim is that SOBER and BASQ together allow users to 'automatically, accurately, and quickly sort through this plethora of models' (Conclusion), enabling uncertainty-quantified parameter estimation and evidence-based model comparison for battery simulators.","tokens_in":23737,"tokens_out":6152,"duration_ms":61437,"significance":"If the evidence calculations were fully valid, the paper would be a valuable methodological introduction for the battery community, with the practical contribution of open-source implementations and a wide range of demonstration cases. The first five examples are instructive and the predictive-posterior checks are a good practice. However, the key validation of the central model-selection claim in Example 6 is compromised: the DFN evidence is rescaled by an unexplained constant, and the reported evidence credible intervals are so wide that the model rankings are not robust. Because this example is the main evidence for the paper's headline claim, the manuscript needs substantial revision before the contribution can be considered established. The paper is more a primer and software showcase than a new theoretical contribution, but that is consistent with its stated purpose.","major_comments":[{"comment":"The DFN evidence is multiplied by (sqrt(2*pi))^3 and reported as 'DFN (corr.)'. This factor is not part of the marginal likelihood p(D|M)=int p(D|theta,M)p(theta)dtheta that BASQ is designed to compute; that integral already contains any complexity/Occam penalty through the prior and the integrated likelihood. The multiplier is a constant independent of the data, the prior, and the posterior, so it cannot be derived merely from the normalisation of a Gaussian density. The raw mean-evidence ratios from Table 6(a) are approximately 3%, 5%, and 49% for the 17, 42, and 80 um electrodes; after multiplication they become 51%, 117%, and 792%. Thus the conclusion that the DFN is preferred for thicker electrodes is an artifact of this rescaling. The text itself is contradictory: it says one must 'correct' for the integral penalty by multiplying by (sqrt(2*pi))^3, yet a sentence later says 'We pre","section":"Example 6, Table 6"},{"comment":"The evidence credible intervals are extremely wide, often spanning more than an order of magnitude (e.g., SPM at 17 um: [0.6, 95]; DFN at 80 um: [0.3, 17]), and the intervals for competing models overlap substantially at 17 and 42 um. The statement that ranking by mean evidence is valid 'as long as we ensure the variances are smaller than the means' is not established; for several entries the interval width implies a variance comparable to or larger than the mean. No per-example convergence diagnostics are shown beyond two predictive-posterior plots for DFN fits. The fixed evaluation budget of 48 initial samples and 19 SOBER iterations at 48 samples each for all impedance examples, independent of the number of fitted parameters, is a strong assumption. The paper should demonstrate, at least for the model-selection examples, that the BASQ evidence estimates are converged with respect to i","section":"Example 6, Result and Discussion, Table 6"},{"comment":"The statement 'We fix kappa_e,peak and kappa_e,spread at 1.0, as we found them to be virtually unidentifiable in a first test run' is circular with respect to the example's stated aim of assessing whether electrolyte parameters are uniquely identifiable from a constant-current pulse. Excluding two of the four transport-shape parameters based on a preliminary run means the subsequent identifiability analysis covers only the remaining parameters and cannot make claims about the fixed ones. If the intended message is that these parameters are unidentifiable, the paper should demonstrate this from the full model and report the preliminary evidence. If they are fixed for numerical conditioning, the clustering and correlation results should be explicitly described as conditional on that choice, and the generality of the identifiability conclusion should be qualified.","section":"Example 5, Method"}],"minor_comments":[{"comment":"The caption does not define 'DFN (corr.)'. State explicitly that this column is the BASQ evidence multiplied by (sqrt(2*pi))^3, and distinguish it from the unmodified evidence in Table 6(a).","section":"Example 6, Table 6 caption"},{"comment":"The sentence 'We prefer comparing the determinant of Sigma between different models directly, without the arbitrary dimensionality penalty' conflicts with the rescaling actually applied. Rewrite to explain the exact relationship between the correction factor and the posterior covariance determinant, or remove the correction.","section":"Example 6, text near Eq. (23)"},{"comment":"The Kronecker-structure multi-output GP notation in Eq. (21) is introduced without a definition of the kernel K or the fidelity kernel K_f. A reader unfamiliar with multi-fidelity GPs will not be able to reproduce this example.","section":"Example 4, Figure 7 and Eq. (21)"},{"comment":"The fixed number of integration nodes for BASQ is '3 to the power of the number of fit parameters', which grows quickly (e.g., 729 nodes for three parameters, 531441 for six parameters). Please state whether this is the number of nodes per dimension or the total number, and how the evidence variance is estimated.","section":"Example 6, 'For all impedance examples...'"},{"comment":"The paper uses the authors' own previously published algorithms (SOBER, ref. 12; BASQ, ref. 13; EP-BOLFI, ref. 4) and, in Example 6, the authors' own impedance data (ref. 53). This should be stated more explicitly in the main text so that readers understand that the demonstrations are not independent third-party validations.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is largely self-referential: the three main algorithms (SOBER, BASQ, EP-BOLFI) are all from the same research group, and the real-data validation in Example 6 uses the authors' own earlier dataset. This is not disqualifying, but the paper would be stronger with at least one independent dataset or a comparison against a standard MCMC baseline for one of the examples. The decisive issue is the unjustified rescaling of DFN evidence in Table 6; without it, the impedance model-selection rankings reverse, and the paper's central claim of automatic and accurate model sorting is not yet supported. I therefore cannot recommend acceptance in the current form, but the problem is local enough that a careful reanalysis and honest reporting of raw evidences could make the paper publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-written primer that will genuinely help battery researchers use SOBER and BASQ, but its key real-data validation — impedance-based model selection in Example 6 — is compromised by a post hoc rescaling of DFN evidence that flips the rankings. That needs to be fixed before the central claim is credible.\n\nWhat's actually new: the paper isn't claiming new algorithms; SOBER and BASQ are already published. The value is in the applications: impedance-based DFN/SPM/SPMe model selection on real NMC data, electrolyte identifiability via an actively trained global inverse surrogate, the DRT-distance cost function, and the PyBOP wrapper. The pedagogy is good — the covariance/identifiability explanations and the knee-point example are clear. Code and data links are provided.\n\nThe soft spots are real, though. In Example 6, to compare SPM (3 params) with DFN (6 params), the authors multiply DFN evidences by sqrt(2π)^3, calling it a 'correction' for dimensionality. But BASQ is supposed to compute the actual marginal likelihood; that integral already contains any Occam penalty. The multiplier is a data-independent constant that favors the larger model. It's decisive: raw DFN/SPM evidence ratios of roughly 0.03, 0.05, and 0.49 become 0.46, 0.78, and 7.7 after the correction. So the paper's own reported conclusion — DFN preferred for thicker electrodes — is an artifact of this choice. The authors even call the factor 'arbitrary' in the text, which doesn't help. They need to either derive the factor from actual posterior covariances and prior volumes, or drop the cross-model comparison.\n\nSecond, the evidence credibility intervals in Table 6 are extremely wide (e.g., SPM 17 µm [0.6, 95], scaled by 1000), and no per-example convergence diagnostics are shown beyond predictive-posterior plots. The fixed budgets (48 initial + 19×48 samples for all impedance examples) may or may not be enough; the paper doesn't demonstrate it.\n\nThird, in Example 5 they fix κe,peak and κe,spread to 1.0 after an initial test run. They disclose this, and the identifiability reasoning is plausible, but it means the electrolyte identifiability analysis actually uses only 4 of 6 parameters. The claim about electrolyte identifiability is correspondingly narrower.\n\nThe self-citation pattern (SOBER, BASQ, EP-BOLFI, ref 53) is not itself a problem — the applications are independent demonstrations, and the circularity burden is low.\n\nBottom line: this deserves peer review, but as major revision. The primer and most examples stand; the impedance model-selection results need a real justification for the rescaling and better convergence evidence.","headline":"Useful Bayesian primer for battery modelers, but Example 6's model-ranking claims rest on an arbitrary evidence rescaling the authors themselves flag; the 'automatically, accurately sort models' conclusion is not yet supported.","tokens_in":24157,"tokens_out":3093,"would_cite":false,"duration_ms":31053,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62L05","62P30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Two Bayesian algorithms, SOBER and BASQ, can parameterize and rank battery models from test data in a fraction of the simulator runs that Markov-chain sampling requires, and the paper demonstrates the pair on six battery problems.","keywords":["Bayesian inference","Model selection","Battery simulation","Likelihood-free inference","Bayesian optimisation","Bayesian quadrature","Gaussian processes","Electrochemical impedance spectroscopy"],"falsifier":"Run a battery test case through SOBER+BASQ under the paper's fixed evaluation budgets and compare the resulting posterior and evidence ranking against a gold-standard MCMC posterior and a precise evidence estimate (e.g., thermodynamic integration or annealed importance sampling). Disagreement beyond the stated credibility intervals on a dataset where the candidate models are known to be measurably different would directly contradict the central claim.","tokens_in":23217,"feed_emoji":"🔋","tokens_out":3961,"duration_ms":44853,"temperature":0.7,"pith_summary":"The paper argues that combining two Bayesian algorithms—SOBER for parameter estimation and BASQ for model evidence—turns battery model comparison into an automated, uncertainty-quantified procedure. The central claim is that this pairing works with only a few hundred simulator evaluations, making Bayesian inference practical for expensive physics-based battery simulators. The authors demonstrate the approach on battery sizing, voltage relaxation analysis, knee-point detection in degradation data, inverse surrogate training, electrolyte identifiability, and impedance model selection. If the claim holds, battery researchers can replace slow MCMC sampling and heuristic curve fitting with posterior distributions and evidence-based model ranking, which matters because new battery chemistries generate a growing range of candidate models.","feed_headline":"Two Bayesian algorithms make battery model choice automatic","feed_subtitle":"They replace slow sampling with fast, uncertainty-aware parameter fits and evidence scores that rank candidate models.","key_machinery":"SOBER is the parameter-estimation engine: it models the discrepancy between simulator output and data with a Gaussian process, uses an uncertainty-sampling acquisition function to choose batches of simulator runs, and returns an approximate posterior distribution over parameters. BASQ is the model-selection engine: it computes the marginal-likelihood integral (evidence) over that posterior, producing a scalar score for comparing candidate models. The pairing works because SOBER makes inference sample-efficient enough for expensive battery simulators, while BASQ turns each parameterisation into a model-comparison criterion that also serves as a convergence check.","core_discovery":"On its own terms, the paper establishes that a likelihood-free Bayesian inference pipeline—SOBER, which frames parameter fitting as Bayesian optimisation with an uncertainty-sampling acquisition function, and BASQ, which evaluates the model-evidence integral—can be applied to battery simulators with modest numbers of model evaluations. The key demonstration is that model evidence, not just point fits, can be computed accurately enough to rank models: the one-knee vs. two-knee degradation model, the single-particle vs. full-transport impedance model, and the ordering of double-layer vs. SEI timescales are all decided by comparing evidence values. The same machinery also trains a global invers","pith_inferences":["The evidence-based workflow could serve as a screening tool in high-throughput materials discovery, where the evidence value itself is a compact, uncertainty-aware descriptor of whether an experiment carries information about a proposed model.","A testable extension would apply the same pipeline to streaming degradation data, checking whether the evidence for a second knee point rises before the knee is visually apparent, enabling early warning in second-life battery applications.","The fixed evaluation budgets used throughout the paper suggest that the method may be less robust when the posterior is multimodal, a case the paper only briefly mentions; applying the approach to deliberately multimodal synthetic problems would test that boundary."],"forward_implications":["Battery researchers can rank candidate physics-based models by computed evidence, replacing manual inspection of fit curves or single-point RMSE comparisons.","Parameterisation outputs include posterior distributions and correlation matrices, so identifiability problems—parameters that cannot be separated from data—become visible directly from the fit.","The same evidence calculation can flag when data are too weak to support a model choice, since wide evidence intervals signal that ranking by mean values is unreliable.","Active learning with SOBER can build inverse surrogates that map measured signals back to operating conditions or material parameters with far fewer simulator runs than naive design-of-experiments.","The approach generalises naturally to automated laboratories: the algorithm can decide which experiment to run next based on where the surrogate or evidence is most uncertain."],"fun_headline_variants":["Faster Bayesian fits rank battery models by evidence","Bayesian short-cuts pick correct battery models","Model evidence via two fast algorithms ends battery guesswork","Bayesian algorithms SOBER and BASQ make model choice clear","Uncertainty-aware fits rank battery models fast"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The preset number of simulator evaluations is large enough for SOBER and BASQ to converge to the true posterior and to evidence values accurate enough for ranking; this is mostly assumed rather than verified per example, and some reported evidence intervals are wide enough that rankings by means may not be trustworthy.","fun_headline_variants_meta":{"raw":{"variants":["Faster Bayesian fits rank battery models by evidence","Bayesian short-cuts pick correct battery models","Model evidence via two fast algorithms ends battery guesswork","Bayesian algorithms SOBER and BASQ make model choice clear","Uncertainty-aware fits rank battery models fast"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2324,"prompt_tokens":618,"completion_tokens":1706,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":362,"completion_tokens_details":{"reasoning_tokens":1645}},"tokens_in":362,"tokens_out":1706,"duration_ms":11407,"temperature":1.0,"reasoning_tokens":1645,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T17:15:39.016541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a battery test case through SOBER+BASQ under the paper's fixed evaluation budgets and compare the resulting posterior and evidence ranking against a gold-standard MCMC posterior and a precise evidence estimate (e.g., thermodynamic integration or annealed importance sampling). Disagreement beyond the stated credibility intervals on a dataset where the candidate models are known to be measurably different would directly contradict the central claim.","supporting_citations":[],"review_version":1}