{"id":"4fecb29e-ef8c-43c2-ae00-b31e77bda23c","arxiv_id":"2506.17755","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"PIMOE uses a mixture-of-experts network with physics-derived features to forecast second-life battery capacity trajectories from a single cycle of partial charging and relaxation data, reporting 0.88% mean absolute percentage error over 50-cycle horizons.","lead":"A new neural network predicts how retired electric vehicle batteries will keep losing capacity over the next 50 to 150 cycles using only one partial charge and voltage-relaxation measurement, plus assumed future usage. It reports 0.88% average error across 207 lab batteries, but the practical value depends on whether future operating conditions can really be known in advance.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation split is not documented; sliding-window overlap may let the model see test horizon capacities during training, so the 0.88% MAPE may not reflect history-free generalization.","rationale":"The paper's central claim is an empirical generalization claim: from one field-accessible cycle and an assumed future load sequence, PIMOE computes future capacity trajectories of retired batteries not seen during training. The most load-bearing condition for that claim is that test trajectories are genuinely out-of-sample. The manuscript never documents the train/test split. The sliding-window construction (Supp. Note 1) makes this omission consequential: every cycle of each battery generates a training sample whose target is the next H capacities, so a random cycle-level split creates overlapping target windows between train and test. In that case the model could memorize or interpolate the exact future capacities it is evaluated on, and the 0.88% MAPE would not transfer to a new retired battery. Supplementary Note 10 references a 78-battery training set and a fixed test set but does not specify the partition by battery, condition, or cycle, nor any temporal exclusion. This is not an internal inconsistency, but it is a missing verification step for the central claim. The reader's designated weakest assumption (future operating conditions must be specified) is acknowledged by the authors in the Discussion and is part of the claim's conditional wording; it does not by itself invalidate the method as a conditional forecaster. The split issue, if it lands, invalidates the empirical evidence for the conditional forecaster. The proposed concrete check—publishing the exact partition and rerunning with a strict battery-level plus temporal-exclusion split—would settle it. If the leak-free MAPE remains near the reported values, the concern does not land and the conditional acceptance stands. Credit is due for the breadth of experiments, ablations, noise analysis, and public code and data links; the issue is one of evaluation protocol, not of fraudulent reporting.","tokens_in":29825,"tokens_out":6425,"duration_ms":66539,"concrete_test":"Publish the exact battery/condition/cycle partition used for training, validation, and testing. Then rerun the full PIMOE pipeline under a strict leak-free split: (1) all windows from each held-out battery are excluded from training; (2) for any test window starting at cycle t, no training window from the same battery may have a target horizon containing any cycle at or after t. Report MAPE per dataset (UL, TPSL-Random, TPSL-Fixed), per initial SOC, and for the 50- and 150-cycle horizons. If the leak-free MAPE remains close to the reported 0.88% (UL) and 2.8-3.0% (TPSL), the concern is resolved; if it rises substantially or approaches baseline Transformer errors, the headline accuracy was inflated by overlapping training and test windows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PIMOE's headline accuracy is presented as validation of history-free trajectory computation across 207 batteries, but the manuscript never specifies how training and test windows are separated. The sliding-window construction in Supplementary Note 1 creates, for each battery, a training sample at every cycle whose target is the next H capacity values (Eqs. 1-2 there). If windows are split randomly by cycle rather than held out as a contiguous block, training samples at cycles t-1, t-2, ... have target horizons that overlap the test sample at cycle t, so the model has directly been trained on the future capacities it is asked to predict. Supplementary Note 10 mentions a complete training set of 78 batteries and a fixed test set, but the exact battery/condition/cycle partition is never given, and no temporal exclusion is described. Without proof that test trajectories come from batteries or contiguous cycle blocks unseen in training and with non-overlapping target cycles, the reported MAPE does not establish generalization to a new retired battery with unknown history. This is the load-bearing issue because the central claim is an empirical generalization claim, and the supporting experiment is not shown to be leak-free.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes PIMOE, a mixture-of-experts architecture combined with a recurrent network (FORNN) to compute future capacity-degradation trajectories of retired lithium-ion batteries from a single partial charging curve and a relaxation-voltage measurement at a random initial state of charge, together with an assumed future load profile. The method is evaluated on the Uniform-Life dataset (130 cells) and the Two-Phase Second-Life dataset (77 cells) and is compared with PatchTST and Informer, with additional ablations, noise-perturbation studies, data-scarcity experiments, and horizon sweeps. The headline results are an average MAPE of 0.88% over a 50-cycle horizon and 1.50% over 150 cycles, with an inference time of 0.43 ms per battery.","tokens_in":29995,"tokens_out":7125,"duration_ms":68373,"significance":"If the reported accuracy is leak-free, PIMOE would be a practically valuable history-free trajectory forecaster for second-life battery management: it uses field-accessible partial-cycle signals, incorporates future load scenarios, and is computationally very light. The authors should be credited for the breadth of the evaluation (207 cells across multiple conditions), the ablations of the AMDP and FORNN modules, the feature-perturbation and data-scarcity analyses, and the public code/data link. The expert-weight specialization shown over the lifecycle is suggestive and could be a useful interpretability tool. However, the significance is conditional: the headline metric currently lacks a documented train/test split, a trivial baseline, a precise statement of the protocol underlying the 0.88% figure, and independent validation of the claimed degradation-mode labels. These issues must be resolved before the central empirical claim can be taken at face value.","major_comments":[{"comment":"The manuscript never specifies how training and test windows are separated. Supplementary Note 1 constructs sample pairs by sliding over every cycle t, with target horizon [t+1, t+H] (Eqs. (1)-(2)); if the split is random by cycle rather than by battery or by temporally non-overlapping blocks, a model trained on cycles t-1, t-2, ... has already seen target capacity values that overlap the test horizon at cycle t, so the reported 0.88% MAPE would not reflect generalization to a new retired battery. Supplementary Note 10 states only that the test set is \"unchanged\" during data reduction; the exact battery/condition/cycle partition is absent. Please provide the split protocol explicitly (battery-level or contiguous-block, with no target-cycle overlap) and report the per-condition test sizes. This is load-bearing for the history-free generalization claim.","section":"Results (Dataset; Model performance and generalization capability); Supplementary Notes 1 and 10"},{"comment":"The headline 0.88% average MAPE is presented without a trivial baseline and without a precise protocol for how the number is composed. No persistence or linear-extrapolation forecaster is reported, and the initial-SOC composition of the 0.88% figure is unspecified: Fig. 3f reports MAPEs of 1.44%, 2.81%, and 3.43% at 50% SOC for the three datasets, so the aggregate number depends on the mix of initial SOCs and horizons used. Because capacity fade over 50 cycles is gradual, a \"repeat last capacity\" baseline may already achieve low MAPE; without such a baseline, the improvement over PatchTST and Informer does not by itself establish that the trajectory shape is being predicted. Please add persistence and linear baselines and state exactly which initial-SOC conditions and horizons enter the 0.88% average.","section":"Results, Fig. 3c-3f; Abstract"},{"comment":"Supplementary Table 9 is identical to Supplementary Table 7 in every row. This duplication is a clear copy/paste error and invalidates the 80-cycle results as currently presented. The table must be regenerated from the actual experimental runs, and the claims about horizon sensitivity in the Results section and Fig. 5d should be updated if the corrected numbers differ.","section":"Supplementary Table 9 (80-cycle results) and Supplementary Table 7 (40% training data)"},{"comment":"The expert networks are labeled as SEI formation, SEI thickening, and lithium plating based on the temporal evolution of router weights, but no independent measurement (post-mortem analysis, incremental-capacity or dQ/dV analysis, electrochemical impedance spectroscopy) is used to validate that each expert corresponds to the claimed degradation mode. The Discussion itself states that the employed features are \"essentially statistical correlations, lacking systematic integration of deeper physicochemical mechanisms.\" Therefore, the interpretability claims in Fig. 4 should be reframed as a hypothesized correspondence, or supported with external mechanistic measurements; otherwise the physics-informed/mechanistic interpretation is not evidenced.","section":"Results (Rationalization of statistical model performance), Fig. 4a, 4c, 4d; Discussion"}],"minor_comments":[{"comment":"The number of batteries is reported as 207 in the Abstract and as 203 independent cells in the Discussion; these counts should be reconciled.","section":"Abstract; Discussion"},{"comment":"The MAPE values for TPSL-Fixed and TPSL-Random appear to be swapped between the main text (2.96% for TPSL-Fixed, 2.81% for TPSL-Random) and Supplementary Table 4 (2.957 for TPSL-Arbitrary and 2.806 for TPSL-Fixed); please correct the inconsistency.","section":"Results, Fig. 3d; Supplementary Table 4"},{"comment":"The placeholder \"see Supplementary Material X\" should be replaced with a real reference to the prediction-horizon selection note.","section":"Supplementary Note 1"},{"comment":"The statement that the method \"reduces computational time and MAPE by 50%, respectively\" is ambiguous: the inference-time reduction is about 50% relative to Informer, but the MAPE reduction relative to the baseline models is much larger; please state the exact comparison baseline for each quantity.","section":"Abstract; Results, Fig. 3e"},{"comment":"The captions of Supplementary Figs. 13 and 14 appear mismatched with their descriptions (one is captioned TPSL-Arbitrary but describes TPSL-Fixed, and the neighboring figure has the opposite inconsistency); the figure numbering and captions should be checked.","section":"Supplementary Figures 13 and 14"},{"comment":"The manuscript needs a careful language and copy-editing pass; numerous typographical errors remain (e.g., \"relavant\", \"larbor\", \"Desipte\", \"valuse\", \"statisitical\", \"pipleline\"), and several informal phrases should be revised for a journal submission.","section":"Whole manuscript"},{"comment":"The 0.43 ms inference-time measurement should be accompanied by a specification of the hardware and software environment on which it was obtained.","section":"Results, Fig. 3e"}],"recommendation":"major_revision","confidential_remarks":"The duplicated Supplementary Table 9/7 and the swapped TPSL labels suggest a quality-control lapse during manuscript assembly. I recommend that the editor request an internal data-integrity audit before the revision is sent back to the reviewers, and that the authors be asked to deposit the exact train/test split definitions with the code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious ML-for-batteries paper with a real new combination and a lot of experimental work behind it, but the headline generalization claim is not supported until the train/test split is documented. I would send it to peer review rather than desk-reject, with the split as the mandatory revision.\n\nWhat is actually new: the packaging of degradation-mode routing (MoE over physics-derived features from one partial charge and relaxation curve) feeding a future-load-conditioned LSTM, all aimed at computing trajectories for retired batteries without any historical cycling data. The ablations, noise perturbation, data-reduction, and horizon studies show the architecture does something internally coherent. If the evaluation is leak-free, this is a deployable tool for second-life sorting and grid storage planning, and the 0.43 ms inference is a real practical advantage.\n\nWhere the soft spots are: the biggest one is the evaluation protocol. The paper never states whether test windows come from batteries or contiguous cycle blocks that were absent from training. Supplementary Note 1 constructs sliding-window samples whose targets are the next H capacity values; if training windows from earlier cycles of the same battery include those targets, the 0.88% MAPE reflects leakage and tells you nothing about a new retired battery. The mention of a 78-battery training set suggests a battery-level split, but the paper does not say which batteries or cycles are held out, and no temporal exclusion is described. This must be clarified before the central claim is credible.\n\nAlso: the 0.88% headline conflicts with the per-scenario MAPEs in Fig. 3d unless a weighting is disclosed; the \"reduces MAPE by 50%\" wording is not what Fig. 3e shows; and Supplementary Tables 7 and 9 are identical, a copy-paste duplicate. These are medium and small issues, but they add up.\n\nThe \"physics-informed\" part is best read as feature engineering plus interpretability. The authors themselves say the connections are statistical correlations, so the claims should stay bounded.\n\nWho this is for: battery degradation researchers, second-life repurposing people, and anyone working on MoE with physical features. It deserves a serious referee: the idea is worth engaging with and the experimental breadth is real. In current form I would not rely on the 0.88% number in my own work until the split and weighting are pinned down.","headline":"Useful architecture and real experimental breadth, but the headline accuracy is not credible until the train/test split is disclosed.","tokens_in":30598,"tokens_out":4377,"would_cite":false,"duration_ms":41505,"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":"Without any historical cycling data, PIMOE computes a retired battery's future capacity trajectory from one partial charging curve plus relaxation voltage and an assumed future load schedule, reporting 0.88% average error over 50 cycles.","keywords":["second-life battery","degradation trajectory prediction","physics-informed learning","mixture of experts","state of health","capacity fade","partial charging curve","relaxation voltage"],"falsifier":"Deploy the trained PIMOE on retired cells with logged future loads but hidden capacity history, specifying the load sequence that is actually used, and compare predicted and measured capacity at cycles 50 and 150. If the reported 0.88% and 1.50% average MAPE reproduce only when the given load sequence matches the true one, the history-free claim holds; if accuracy collapses when load assumptions deviate even slightly, the effective constraint moves from data accessibility to forecastability of future operating conditions.","tokens_in":29585,"feed_emoji":"🔋","tokens_out":7929,"duration_ms":76623,"temperature":0.7,"pith_summary":"Retired electric-vehicle batteries are hard to repurpose because their degradation history is usually unavailable and their future loads are uncertain. The paper proposes a physics-informed mixture-of-experts network (PIMOE) that claims to compute a retired battery's capacity-fade trajectory from one field-accessible cycle alone: a partial charging curve starting at a random state of charge plus a 30-minute relaxation voltage measurement, together with an assumed schedule of future charge/discharge currents and temperatures. On 207 batteries across 77 use conditions and 67,902 cycles, it reports average errors of 0.88% over a 50-cycle horizon and 1.50% over a 150-cycle horizon, with a 0.43 ms inference time. The significance, if the claim holds, is that second-life batteries could be screened and routed for reuse without offline capacity tests or historical records, and the same expert weights can indicate whether a cell is fit for demanding storage duty or should be recycled.","feed_headline":"No history: one partial charge curve forecasts 150 cycles","feed_subtitle":"PIMOE reads only current-cycle field data, then computes capacity fade over 50-150 future cycles to about 1% error.","key_machinery":"The load-bearing object is the degradation router inside the Adaptive Multi-degradation Prediction (AMDP) module. It takes twelve physics-informed features, six from the partial charging curve and six from relaxation voltage, including charge acceptance in a 0.05 V window and voltage rise per 200 mAh, applies a noisy top-k softmax over five expert networks, and synthesizes a latent degradation-trend vector. Each expert maps the 50-point charging curve to a short-horizon trend, and the router's weights select which degradation mode dominates at that moment in life. The Future-Operation Recurrent Neural Network (FORNN) then concatenates that trend with per-cycle future load triplets $(I_{\\text{charge}}, I_{\\text{discharge}}, T)$ and feeds an LSTM that emits the capacity trajectory, which is what lets the model respond to changing second-life loads rather than assuming a constant profile.","core_discovery":"PIMOE establishes a history-free route from in-field electrical signals to future degradation: the input is a random-SOC partial charge curve (50 voltage bins) and a relaxation voltage trace; twelve statistics encode polarization and aging state; a noisy top-2-of-5 degradation router produces a latent trend embedding; and a recurrent network walks that embedding forward through the assumed load profile to output capacity per future cycle. The central claims are that this pipeline computes 50-cycle capacity trajectories at 0.88% average MAPE without historical data, extends to 150 cycles at 1.50% average MAPE, stays below 6.26% maximum MAPE, and works across three material systems and both fixed and randomly varying second-life loads. The paper further claims that the router's expert weights alone classify retired cells by state of health, with high-confidence separation between cells retired at 95% SOH ('excellent') and 75% SOH ('scrap'), so the interpretable routing doubles as a screening tool.","pith_inferences":["A natural extension is to treat the router's confidence as a distribution-shift alarm: when expert weights are uncertain or fall outside trained regions, the forecast should be flagged rather than trusted, since the training data do not include thermal-runaway or internal-short regimes.","The model's expert-to-mechanism mapping (SEI formation, SEI thickening, lithium plating) is inferred from trajectory slope and literature, not from direct cell teardown; an independent validation could check whether the expert weights align with differential-voltage signatures or post-mortem diagnostics.","Because the method requires an assumed future load sequence, real deployments should pair PIMOE with a load-forecasting module or scenario ensemble; otherwise the validation protocol implicitly bundles trajectory-model error with load-forecast error.","The input features are deliberately simple statistics, so replacing or augmenting them with incremental-capacity peaks or other mechanistic descriptors might sharpen loss-of-lithium versus loss-of-active-material separation, and this is a cheap, testable modification."],"forward_implications":["Battery recyclers could classify and route retired cells using one partial charge and relaxation measurement, replacing full-capacity calibration tests and historical data audits.","Because random initial state of charge is supported, field batteries need no pre-conditioning to a fixed SOC before measurement.","The same trained network can be applied across first-life and second-life regimes, since the results show it tracks capacity when loads switch abruptly.","Long-horizon planning is feasible: 150-cycle forecasts remain at 1.50% average MAPE, so battery-management decisions could be made against predicted future fade rather than current state of health alone.","Small training budgets suffice, with the model remaining usable on a pruned 5 MB training set, lowering the data barrier for deployment."],"supporting_citations":[{"why":"Defines the prior approach requiring at least one full cycle of capacity-voltage data for degradation prediction under uncertain future conditions, which PIMOE replaces with partial random-SOC data.","marker":"[35]"},{"why":"Physics-informed trajectory prediction from early data that still requires historical cycling data, the requirement PIMOE removes.","marker":"[34]"},{"why":"Shows random-SOC single-point SOH estimation for second-life batteries, motivating the extension from one capacity point to a full trajectory.","marker":"[12]"},{"why":"Forecasts degradation trajectories under domain shift using historical capacity data, providing a comparison point for data-requirement claims.","marker":"[33]"},{"why":"The patch-based transformer sequence model used as a baseline for accuracy and inference-time comparisons.","marker":"[44]"},{"why":"The sparse-attention transformer sequence model used as a second baseline for accuracy and inference-time comparisons.","marker":"[45]"},{"why":"Supplies the three degradation phases (SEI formation, SEI thickening, lithium plating) that the expert router is claimed to discover.","marker":"[8]"},{"why":"Supports the nonlinear aging-stage interpretation used to explain expert specialization across battery life.","marker":"[46]"}],"fun_headline_variants":["No history, one cycle: battery fade forecast to 1% error","One partial charge curve predicts 150 cycles of battery fade","PIMOE: battery fade forecast from a single cycle, no history","Interpretable AI reads one cycle to forecast battery fade"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For every claimed trajectory, the future charge current, discharge current, and temperature of each future cycle are assumed known and are fed into the network; if those future loads cannot be specified or forecast, the history-free trajectory computation cannot be run as validated.","fun_headline_variants_meta":{"raw":{"variants":["No history, one cycle: battery fade forecast to 1% error","One partial charge curve predicts 150 cycles of battery fade","PIMOE: battery fade forecast from a single cycle, no history","Interpretable AI reads one cycle to forecast battery fade"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00076,"raw_usage":{"total_tokens":3410,"prompt_tokens":1015,"completion_tokens":2395,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":2323}},"tokens_in":631,"tokens_out":2395,"duration_ms":18322,"temperature":1.0,"reasoning_tokens":2323,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:01:51.485921+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the trained PIMOE on retired cells with logged future loads but hidden capacity history, specifying the load sequence that is actually used, and compare predicted and measured capacity at cycles 50 and 150. If the reported 0.88% and 1.50% average MAPE reproduce only when the given load sequence matches the true one, the history-free claim holds; if accuracy collapses when load assumptions deviate even slightly, the effective constraint moves from data accessibility to forecastability of future operating conditions.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the prior approach requiring at least one full cycle of capacity-voltage data for degradation prediction under uncertain future conditions, which PIMOE replaces with partial random-SOC data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Physics-informed trajectory prediction from early data that still requires historical cycling data, the requirement PIMOE removes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The sparse-attention transformer sequence model used as a second baseline for accuracy and inference-time comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the three degradation phases (SEI formation, SEI thickening, lithium plating) that the expert router is claimed to discover."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the nonlinear aging-stage interpretation used to explain expert specialization across battery life."}],"review_version":2}