acceptodds
Under review as a conference paper at ICLR 2027

Electrochemistry Degradation Representation for Causal SOH Prediction in Batteries

Abstract

Battery state-of-health (SOH) forecasting is essential for reliable battery management, lifetime assessment, and risk-aware operation in electric vehicles and energy storage systems. However, accurate SOH prediction remains challenging because battery aging is governed by latent electrochemical degradation processes that are only indirectly reflected by heterogeneous cycle-level measurements, while historical SOH information may be incomplete or unreliable during deployment. To address these challenges, we propose EDR-BAT, an electrochemistry degradation representation framework for sequential causal SOH prediction in batteries. Guided by physically motivated degradation priors, EDR-BAT first transforms raw cycling measurements into multi-modal electrochemical embedding-based health features and uses series-level identity as a shared prior to query recent electrochemical history. This yields cycle-specific degradation representations that capture both shared aging regularities across similar batteries and cycle-specific variations within individual cells. To reduce drift under online prediction, EDR-BAT further introduces an SOH memory mechanism with memory-anchored training, allowing the model to operate with measured, incomplete, or autoregressively predicted SOH histories and reducing training-inference mismatch. Experiments on five battery aging datasets show that EDR-BAT consistently improves SOH prediction accuracy over representative baselines, reducing MAE/RMSE by 12.9%/11.4% on average compared with the best-performing deep learning baseline on each dataset.

Primary area:
learning on time series and dynamical systems
Keywords:
battery state-of-health prediction, electrochemistry degradation representation, causal SOH prediction, historical memory, cross-cell generalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pfvSOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This article presents a method named EDR-BAT to predict battery health (SOH) causal relationships. This method employs electrochemical feature embedding, series-specific learnable priors and cross-attention on historical observation. Moreover, low-rank bilinear interactions are also used. Besides, it utilizes the autoregressive memory on SOH and training based on memory to alleviate the deviation between training and inference. The effectiveness is validated through experiments of five battery datasets. It reduces average MAE by 12.9% and RMSE by 11.4% compared with the best deep learning baseline in each dataset.

Critical strengths

This paper discusses an important issue in predicting the SOH of battery: it's a real challenge for the lack of consistency between the training history of SOH and the autoregressive inference process. They conduct experiments on five datasets, many baselines and ablation studies are tested. The improvements are presented for the various battery degradation situations.

Critical weaknesses
  1. The proposed method is basically a combination of existing methods such as handcrafted electrochemical features, cross attention mechanism, low-rank bilinear fusion and scheduled sampling. Especially the memory-based training method seems very similar to scheduled sampling. There are also recently proposed works about battery SOH prediction that exploit electrochemical features and attention-based historical information modeling. But, this work mainly compares with some general baseline methods and there is no real comparison against these closely related recent works.

  2. The features used in this model are based on current cycle capacity, and labels (SOH) are also based on current cycle capacity. Even though the authors say that these features are different from true capacity, it's not well clarified how exactly those features are constructed and whether they are available in (so-called) online case or not. So, there is still possibility of information leakage issue.

  3. The experiments are only conducted under a scenario where unseen batteries in given dataset are tested. As the proposed model needs to learn some priors for a given series (learnable), it's still not clear how well this method would work for unseen battery series without training/tuning, unless it can work in other than experimental environment.

Questions for authors

Same as the weaknesses.

4 comments
AuthorsOpenReview ↗

Novelty and Closely Related Comparison

We thank the reviewer for the constructive assessment. We address the three concerns in the following comments.

1. Novelty and closely related comparison

EDR-BAT does not claim the individual attention, fusion, or sampling operators as new. Its contribution is their battery-specific organization: a shared query set is adapted through electrochemical history into a cycle-specific degradation representation, which then interacts separately with current-cycle evidence and SOH memory. Memory-anchored training is related to scheduled sampling, which we now cite explicitly; here it operates on an explicit SOH memory containing measured, missing, or stop-gradient predicted entries and works jointly with electrochemical-history querying.

Direct related-method comparison.

LSTM-LAM encodes multi-cycle electrochemical features using an LSTM and local attention. We implemented a controlled version preserving this core structure while using the same dataset-specific inputs, causal history lengths, target-cell splits, and training protocol as EDR-BAT; no target-cell SOH labels are used for fine-tuning.

Dataset-level macro-average results from the complete leave-one-cell-out evaluation. Each entry reports MAE/RMSE in percentage points. The final row gives equal weight to the five datasets. Best results are marked in bold.

DatasetLSTM-LAMEDR-BAT
CALCE0.65 / 0.960.45 / 0.61
MIT0.39 / 0.680.34 / 0.37
NASA1.20 / 1.501.12 / 1.35
Oxford0.44 / 0.610.22 / 0.29
In-house1.18 / 1.590.75 / 0.89
Avg.0.77 / 1.070.58 / 0.70

EDR-BAT achieves lower dataset-level macro-average MAE/RMSE on all five datasets and reduces the equal-dataset average by 25.2%/34.5%. This indicates that local-attention reweighting of a unified sequence does not fully replace the explicit separation of current-cycle evidence, electrochemical history, and autoregressive SOH states. Complete per-cell results are reported in revised Appendix C.

AuthorsOpenReview ↗

Condition-matched Mechanism Controls

Condition-matched mechanism controls.

We further evaluated Frozen query, History pooling, a Parameter-matched MLP, and a Dynamic MLP under identical inputs, SOH memory, training strategy, target cells, and splits. These variants respectively test the learnable query set, query-based historical adaptation, low-rank interaction, and a more flexible history-conditioned prediction mapping. Dynamic MLP was an early design explored during method development.

Controlled comparison of the learnable prior and interaction structure under identical inputs, training strategies, target cells, and data splits. MAE/RMSE are reported in percentage points, with small subscripts denoting standard deviations over five random seeds. Avg. denotes the equal-weight average across the five target cells. Best and second-best results are marked in bold and underlined.

Target cellFullFrozen queryHistory poolingParameter-matched MLPDynamic MLP
CS2_360.400.03/0.530.040.40_{0.03}/0.53_{0.04}0.460.02/0.580.050.46_{0.02}/0.58_{0.05}0.440.05/0.560.060.44_{0.05}/0.56_{0.06}0.450.05/0.590.050.45_{0.05}/0.59_{0.05}0.450.04/0.580.050.45_{0.04}/0.58_{0.05}
channel240.470.03/0.500.030.47_{0.03}/0.50_{0.03}0.510.04/0.540.040.51_{0.04}/0.54_{0.04}0.460.04/0.500.040.46_{0.04}/0.50_{0.04}0.520.06/0.560.060.52_{0.06}/0.56_{0.06}0.510.03/0.540.040.51_{0.03}/0.54_{0.04}
B00180.670.13/0.840.140.67_{0.13}/0.84_{0.14}0.900.12/1.060.090.90_{0.12}/1.06_{0.09}0.920.10/1.090.100.92_{0.10}/1.09_{0.10}0.960.12/1.120.110.96_{0.12}/1.12_{0.11}1.000.11/1.150.111.00_{0.11}/1.15_{0.11}
Cell40.180.02/0.220.020.18_{0.02}/0.22_{0.02}0.200.04/0.230.030.20_{0.04}/0.23_{0.03}0.180.04/0.220.040.18_{0.04}/0.22_{0.04}0.200.04/0.240.050.20_{0.04}/0.24_{0.05}0.190.03/0.240.030.19_{0.03}/0.24_{0.03}
In-house-30.820.07/0.960.080.82_{0.07}/0.96_{0.08}1.320.11/1.500.121.32_{0.11}/1.50_{0.12}1.210.14/1.460.171.21_{0.14}/1.46_{0.17}1.190.07/1.350.091.19_{0.07}/1.35_{0.09}1.160.10/1.390.131.16_{0.10}/1.39_{0.13}
Avg.0.51 / 0.610.68 / 0.780.64 / 0.770.66 / 0.770.66 / 0.78

The full model reduces average MAE/RMSE by 25.1%/22.0% over Frozen query, 20.9%/20.4% over History pooling, and 23.5%/21.0% over the Parameter-matched MLP, while also outperforming Dynamic MLP. Because all controls share identical information and training conditions, they isolate the learnable prior, history-conditioned representation, and low-rank interaction beyond historical access or generic fusion. Together with w/o Hist. and p=0p=0, these results distinguish the contributions of historical information, prediction-based SOH-memory training, and degradation-state construction. The complete controls are reported in revised Appendix E.

AuthorsOpenReview ↗

Capacity-related Inputs and Generalization Scope

2. Capacity-related inputs and information leakage

The SOH label, si(n)=Qidis(n)/Qirateds_i(n)=Q_i^{\mathrm{dis}}(n)/Q_i^{\mathrm{rated}}, is computed from complete-discharge capacity and used only for supervision. All model inputs are instead extracted from the current-cycle charging record before this label becomes available, without future-cycle information. Capacity-related inputs include charged throughput Qchg(t)=∫t0tI(τ),dτQ^{\mathrm{chg}}(t)=\int_{t_0}^{t}I(\tau)\\,\mathrm{d}\tau and IC-curve features derived from g(V)=dQchg/dVg(V)=\mathrm{d}Q^{\mathrm{chg}}/\mathrm{d}V; the remaining inputs describe charging duration, voltage rise, signal change, and distributional statistics.

The CALCE table below provides a representative example. Complete tables for CALCE, MIT, NASA, Oxford, and In-house, including every feature, extraction interval, definition, and earliest availability, are provided in the revised manuscript, Appendix A, “Input Feature Definitions and Availability.” Here, rj=(Vj+1−Vj)/(tj+1−tj)r_j=(V_{j+1}-V_j)/(t_{j+1}-t_j), h(V)=[g(V)−b(V)]+h(V)=[g(V)-b(V)]_{+}, and A[a,b]=∫abh(V),dVA[a,b]=\int_a^b h(V)\\,\mathrm{d}V.

CALCE input features. All quantities are extracted from the current-cycle charging record.

FeatureExtraction interval and definitionAvailable after
CC voltage-rise timet(4.00)−t(3.80)t(4.00)-t(3.80) over 3.803.80--4.00,V4.00\\,\mathrm{V}At 4.00,V4.00\\,\mathrm{V}
CC charging durationt(4.20)−tCC,startt(4.20)-t_{\mathrm{CC,start}}After CC
CC current--time integral∫CCI(t),dt\int_{\mathrm{CC}} I(t)\\,\mathrm{d}t; this is charged throughput during CC, not the discharge-capacity SOH labelAfter CC
Minimum CC voltage-change ratemin⁡j∈CCrj\min_{j\in\mathrm{CC}} r_jAfter CC
CC voltage-rate standard deviationStd⁡rj:Vj∈[3.80,4.00]\operatorname{Std}\\{r_j:V_j\in[3.80,4.00]\\}At 4.00,V4.00\\,\mathrm{V}
CC voltage skewnessSkew⁡Vj:Vj∈[3.80,4.00]\operatorname{Skew}\\{V_j:V_j\in[3.80,4.00]\\}At 4.00,V4.00\\,\mathrm{V}
IC main-peak voltageVp=arg⁡max⁡V∈[3.80,4.20]h(V)V_p=\arg\max_{V\in[3.80,4.20]}h(V)After CC
IC main-peak heightH=h(Vp)H=h(V_p), with VpV_p searched over 3.803.80--4.20,V4.20\\,\mathrm{V}After CC
IC main-peak areaA[Vp−0.15,Vp+0.15]A[V_p-0.15,V_p+0.15], clipped to the available IC domainAfter CC
IC main-peak FWHMW0.5=VR0.5−VL0.5W_{0.5}=V_R^{0.5}-V_L^{0.5}After CC

The complete input is assembled after the latest required CC, CV, or IC interval. Thus, charging-cycle-level online prediction uses only available current-cycle charging observations and previous history, without subsequent discharge capacity or future cycles. In the in-house dataset, the full CC current--time integral is the most direct capacity proxy; we mask this standardized feature while preserving the input dimension, architecture, training protocol, and split. Results use five random seeds.

Sensitivity to the direct charged-capacity proxy on the in-house dataset. Small subscripts denote standard deviations over five random seeds. MAE and RMSE are reported in percentage points. Avg. denotes the equal-weight average across the three target cells. Better results are marked in bold.

Target cellFull MAEFull RMSECapacity proxy masked MAECapacity proxy masked RMSE
In-house-10.580.060.58_{0.06}0.760.020.76_{0.02}0.670.080.67_{0.08}0.770.090.77_{0.09}
In-house-20.850.140.85_{0.14}0.940.140.94_{0.14}0.940.190.94_{0.19}1.020.181.02_{0.18}
In-house-30.820.070.82_{0.07}0.960.080.96_{0.08}1.130.091.13_{0.09}1.300.111.30_{0.11}
Avg.0.750.890.911.03

Masking increases average MAE/RMSE from 0.75/0.890.75/0.89 to 0.91/1.030.91/1.03 (21.321.3%/15.7%). The proxy is informative, but the remaining nine features still support prediction, showing that EDR-BAT is not driven solely by it. The revised manuscript clarifies the label--input distinction, feature availability, masking analysis, and prediction time.

3. Generalization beyond observed battery series

In the current protocol, each dataset is an independent battery-series task: the shared prior is learned from source cells and applied to an unseen cell from the same series without target-cell labels or fine-tuning. Because the five datasets are trained separately, these experiments do not establish direct transfer to an entirely unseen series. Such transfer also requires physically aligned inputs, whereas the current datasets use different charging protocols and feature definitions. Future work could construct common online charging features or aligned dataset-specific encoders, jointly learn a cross-series prior, and evaluate it through leave-one-series-out or leave-one-dataset-out protocols. The revised Discussion states this scope and the need for validation under varying temperatures, dynamic loads, partial charging, and fleet data.

Thank you again for your detailed feedback. We have provided point-by-point responses and added the requested clarifications and experiments to the revised manuscript. As the discussion period is approaching its end, if these additions have addressed your main concerns, we would greatly appreciate your reconsideration of the paper in light of the new evidence. If any major concern remains, we would be happy to clarify it further.

Reviewer_f79FOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper suggests EDR-BAT, a deep learning framework that combines electrochemical measurements with historical SOH history to estimate the battery's current SOH on a cycle-by-cycle basis. It is designed to reflect both the shared characteristics of batteries within the same class and the degradation patterns of individual cells, using the model's predictions as a history during traning to adapt to real operating conditions. First temporal, differential, integral,and statistical characteristics are extracted from observed values related to voltage, current temperature, and capacity to construct cycle-specific electrochemical embeddings. Then, the shareable prior for each battery series is queiried for it's recent history using attention, generating degradation related latent representations tailored to the specific cell and cycle. The results support the sequential SOH estimation performance of the overall framework, which combines characteristics, degradation reprresentation reflecting sequence information, and prediction-based SOH history.

It contributes, multi-modal embedding construction via electrochemical properties, proposal of a degradation representaiton and prediction architecture combining battery class infromation with individual cell observations and proposal of a memory based learning and self regression inference strategy to mitigate learning inference discrepancies.

It validates full cycle SOH prediction via 5 datasets (CALCE, MIT, NASA< Oxford, In-house) by comparing SVR, MLP, CNN, LSTM< Transformer, LSTM-W, Trans-W, PINNASOH.

Critical strengths

This paper provies evidence of predictive effectiveness on unseen cells. Appendix C reports complete leave-one-cell-out results for all 22 selected cells across the datasets, with means and standard deviations over five random seeds for the neural models. Averaging these reported results over either cells or datasets favors EDR-BAT in both MAE and RMSE. Although the gains vary across individual cells, this evaluation provieds meaningful evidence for the practical usefulness of the overall framework within the intended cross cell prediction setting.

The treatment of SOH history addresees a practical training inference mismatch and is supported by targeted ablations. EDR-BAT progressively constructs its SOH memory from previous predictions during inference, while traning exposes the model to prediction based histories. The (p=0) ablation supports the usefulness of thie design in Table 2.

Critical weaknesses

Lack of proof for the proposed series-aware degradation representation. According to the reported training protocol, the series prior is a learned query shared by all cells in each training problem. The ablation study shows the contributions of historical branches and prediction based SOH memory training, but they do not establish whether the proposed representation and bilinear interactions improve upon a simpler model given the same electrochemical history, autoregressive SOH context, and memory-training procedure. In my thought PCA analysis does not resolve this attribution gap: interpreting overlap between dataset-specific priors as shared degradation patterns requires establishing a common or aligned latent space. Since the paper identifies structured degradation representation as its main methodological contribution, the current evidence leaves its distinctive benefit insufficiently established.

Questions for authors
  1. Clarification of the electrochemical inputs and availability. Senction 3.1 describes four signal groups and four types of health indicators, while Table 3 reports ten elcetrochemical features. Please specify it, the curve segments required to compute them, and whether the same features are used across all datasets? In particular, how are the curve derived capacity features defined, and at what stage of the current cycle are the required inputs available for SOH prediction? This would clarify the measurement requirements and practical scope of the proposed online prediction setting.

  2. Clarification of what the ablation results establish. In table 2, which historical inputs and representation components are retained in the w/o E-Q and w/o S-Q variants? Do theses variants change the query interactions while preserving access to the same historical information, or do they also remove that information from the predictor? Please clarify which comparisons distinguish the benefit of the proposed structured degradation representation from the benefit of using historical observations and prediction based SOH memory training.

5 comments
AuthorsOpenReview ↗

Mechanism and Distinctive Value

We thank the reviewer for the constructive assessment and for recognizing the practical value of our cross-cell evaluation and SOH-memory training strategy. We address the three concerns in the following comments.

1. Mechanism and distinctive value

To close the attribution gap in the original ablations, we evaluated four condition-matched variants using identical current-cycle inputs, historical observations, SOH memory, training strategy, target cells, and data splits. Frozen query freezes the initial shared query set; History pooling replaces query-based adaptation with masked mean pooling and a parameter-matched encoder; Parameter-matched MLP replaces the low-rank interaction with concatenation-based MLP fusion; and Dynamic MLP, an early design explored during method development, converts the conditioned query set into the parameters of a three-layer prediction MLP.

Condition-matched comparison of the learnable prior and interaction mechanisms on the five target cells used in the main comparison. Each entry reports MAE/RMSE in percentage points, with small subscripts denoting standard deviations over five random seeds. Best and second-best results are marked in bold and underlined, respectively.

Target cellFullFrozen queryHistory poolingParameter-matched MLPDynamic MLP
CS2_360.400.03/0.530.040.40_{0.03}/0.53_{0.04}0.460.02/0.580.050.46_{0.02}/0.58_{0.05}0.440.05/0.560.060.44_{0.05}/0.56_{0.06}0.450.05/0.590.050.45_{0.05}/0.59_{0.05}0.450.04/0.580.050.45_{0.04}/0.58_{0.05}
channel240.470.03/0.500.030.47_{0.03}/0.50_{0.03}0.510.04/0.540.040.51_{0.04}/0.54_{0.04}0.460.04/0.500.040.46_{0.04}/0.50_{0.04}0.520.06/0.560.060.52_{0.06}/0.56_{0.06}0.510.03/0.540.040.51_{0.03}/0.54_{0.04}
B00180.670.13/0.840.140.67_{0.13}/0.84_{0.14}0.900.12/1.060.090.90_{0.12}/1.06_{0.09}0.920.10/1.090.100.92_{0.10}/1.09_{0.10}0.960.12/1.120.110.96_{0.12}/1.12_{0.11}1.000.11/1.150.111.00_{0.11}/1.15_{0.11}
Cell40.180.02/0.220.020.18_{0.02}/0.22_{0.02}0.200.04/0.230.030.20_{0.04}/0.23_{0.03}0.180.04/0.220.040.18_{0.04}/0.22_{0.04}0.200.04/0.240.050.20_{0.04}/0.24_{0.05}0.190.03/0.240.030.19_{0.03}/0.24_{0.03}
In-house-30.820.07/0.960.080.82_{0.07}/0.96_{0.08}1.320.11/1.500.121.32_{0.11}/1.50_{0.12}1.210.14/1.460.171.21_{0.14}/1.46_{0.17}1.190.07/1.350.091.19_{0.07}/1.35_{0.09}1.160.10/1.390.131.16_{0.10}/1.39_{0.13}
Avg.0.51 / 0.610.68 / 0.780.64 / 0.770.66 / 0.770.66 / 0.78

The full model achieves the best or tied-best result in nine of ten target-cell--metric comparisons. It reduces average MAE/RMSE by 25.1%/22.0% over Frozen query, 20.9%/20.4% over History pooling, and 23.5%/21.0% over the Parameter-matched MLP, while also outperforming Dynamic MLP. Since all variants use identical information and training conditions, these results isolate the benefits of the learnable prior, history-conditioned representation, and low-rank interaction beyond historical access or generic fusion. The complete controls are reported in Appendix E of the revised manuscript.

Figure 3 interpretation.

Figure 3 provides qualitative rather than performance evidence. The non-collapsed queries within each dataset show that the model learns a structured query set rather than duplicated initializations. Because the five query sets come from separately trained models without cross-dataset latent-space alignment, their overlap and separation are not interpreted as physically calibrated distances. The performance benefit is instead established by the condition-matched comparison above.

AuthorsOpenReview ↗

Electrochemical Inputs, Online Availability, and Ablation Attribution

2. Electrochemical inputs and online availability

Each dataset uses ten cycle-level charging features, with definitions and extraction intervals adapted to its chemistry and charging protocol. All inputs are extracted from current-cycle CC, CV, or IC curves; neither the subsequent discharge process nor future cycles are used. The CALCE table below provides a representative example. Complete tables for CALCE, MIT, NASA, Oxford, and In-house, including every feature, definition, extraction interval, and earliest availability, are provided in the revised manuscript, Appendix A, “Input Feature Definitions and Availability.”

CALCE input features. All quantities are extracted from the current-cycle charging record.

FeatureExtraction interval and definitionAvailable after
CC voltage-rise timet(4.00)−t(3.80)t(4.00)-t(3.80) over 3.803.80--4.00,V4.00\\,\mathrm{V}At 4.00,V4.00\\,\mathrm{V}
CC charging durationt(4.20)−tCC,startt(4.20)-t_{\mathrm{CC,start}}After CC
CC current--time integral∫CCI(t),dt\int_{\mathrm{CC}} I(t)\\,\mathrm{d}t; this is charged throughput during CC, not the discharge-capacity SOH labelAfter CC
Minimum CC voltage-change ratemin⁡j∈CCrj\min_{j\in\mathrm{CC}} r_jAfter CC
CC voltage-rate standard deviationStd⁡rj:Vj∈[3.80,4.00]\operatorname{Std}\\{r_j:V_j\in[3.80,4.00]\\}At 4.00,V4.00\\,\mathrm{V}
CC voltage skewnessSkew⁡Vj:Vj∈[3.80,4.00]\operatorname{Skew}\\{V_j:V_j\in[3.80,4.00]\\}At 4.00,V4.00\\,\mathrm{V}
IC main-peak voltageVp=arg⁡max⁡V∈[3.80,4.20]h(V)V_p=\arg\max_{V\in[3.80,4.20]}h(V)After CC
IC main-peak heightH=h(Vp)H=h(V_p), with VpV_p searched over 3.803.80--4.20,V4.20\\,\mathrm{V}After CC
IC main-peak areaA[Vp−0.15,Vp+0.15]A[V_p-0.15,V_p+0.15], clipped to the available IC domainAfter CC
IC main-peak FWHMW0.5=VR0.5−VL0.5W_{0.5}=V_R^{0.5}-V_L^{0.5}After CC

Capacity-related inputs.

The SOH label is si(n)=Qidis(n)/Qirateds_i(n)=Q_i^{\mathrm{dis}}(n)/Q_i^{\mathrm{rated}} and is used only for supervision. Capacity-related model inputs instead comprise the CC charged throughput QCC=∫CCI(t),dtQ^{\mathrm{CC}}=\int_{\mathrm{CC}}I(t)\\,\mathrm{d}t and IC-curve features such as peak position, height, width, and area, all computed from the current-cycle charging record. They are therefore distinct from the subsequent discharge-capacity label.

Online availability.

Each feature becomes available after its listed charging interval, and the complete input is assembled after the latest required interval. Thus, EDR-BAT performs charging-cycle-level online prediction: it estimates current-cycle SOH after the required charging observations, without using the subsequent discharge capacity or future-cycle information. The revised manuscript now states this prediction time explicitly.

3. Ablation definitions and attribution

The w/o E-Q variant removes only the low-rank interaction between the history-conditioned degradation representation and current-cycle electrochemical evidence; it retains the electrochemical-history window, learnable prior, history-querying procedure, SOH memory, S-Q interaction, and memory-anchored training. The w/o S-Q variant removes only the interaction with historical SOH memory while retaining the electrochemical-history pathway, learnable prior, cycle-specific representation, and E-Q interaction. Thus, neither variant removes the underlying historical electrochemical observations.

Accordingly, Full versus w/o Hist. evaluates historical information, Full versus p=0p=0 evaluates prediction-based SOH-memory training, and Full versus w/o E-Q or w/o S-Q evaluates the two interaction paths. The original ablations do not isolate the learnable prior or history-querying representation; these are evaluated by Frozen query and History pooling in the mechanism-control table reported in our first comment, while the two MLP variants evaluate the low-rank interaction. The revised manuscript now states explicitly what each variant retains and removes.

Thank you again for your valuable feedback. As the discussion period is approaching its end, we would be grateful if you could let us know whether our responses and revised manuscript have addressed your remaining concerns. We would be happy to provide any further clarification.

Reviewer_rsHKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies online battery state-of-health (SOH) prediction when a target cell's historical SOH measurements are incomplete or unavailable. The authors argue that accurate cross-cell prediction requires distinguishing current-cycle electrochemical evidence, historical degradation patterns, and evolving SOH trajectories, rather than treating all observations as an undifferentiated time series. They propose EDR-BAT, a structured neural framework with three main components: (1) cycle-level embeddings of temporal, differential, integral, and statistical electrochemical features; (2) a battery-series learnable prior that queries historical embeddings through attention to construct cycle-specific degradation representations, coupled to current observations through a low-rank bilinear interaction; and (3) an autoregressive SOH-memory branch trained with a mixture of ground-truth and previously predicted SOH values to mitigate training–inference mismatch. Here, causal refers to the chronological use of current and past observations rather than causal identification in the intervention-based sense.

The authors evaluate the method against SVR, CNN, LSTM, Transformer, sliding-window variants, and PINN4SOH on five battery aging datasets. The five representative target-cell results in Table 1 favor EDR-BAT, and Appendix C additionally reports complete leave-one-cell-out experiments on 22 cells across these datasets. Table 2 ablates the historical electrochemical and SOH-memory interactions, while other appendices examine memory lengths and sparse online calibration. Overall, the results demonstrate that structured history conditioning and prediction-based SOH memory can improve accuracy in several settings. However, the method mainly combines established feature engineering, learned queries, attention, low-rank fusion, and scheduled-sampling-like training; the experiments do not adequately isolate the distinctive contribution of the proposed physical degradation representation. Moreover, the full cross-cell results are more heterogeneous than the selected main-table results suggest.

Strengths. The work addresses a meaningful deployment constraint—online prediction without reliable ground-truth SOH history—and organizes different forms of battery evidence in an understandable architecture. The evaluation spans five datasets and includes full leave-one-cell-out results, component ablations, and additional sensitivity studies. Several improvements in SOH accuracy and the benefit of mixing predicted SOH into the training history are supported empirically.

Weaknesses. The learned representation is not convincingly linked to identifiable electrochemical degradation mechanisms, and key architectural choices lack matched alternatives. The cross-cell results include substantial failures against existing baselines, which limit the claim of consistent superiority. Finally, the exact feature construction, rollout protocol, and correspondence between main and ablation results need clarification, particularly given the risk that current-cycle capacity-derived inputs may act as close proxies for the SOH target.

Critical strengths

1. The deployment-aware formulation addresses a practically important problem.

The paper explicitly considers the absence of reliable measured SOH histories at inference time and avoids depending on future observations from the test cell. Section 3.3 maintains an evolving SOH memory that can be populated with past predictions and updated when measured values become available. This is a useful departure from protocols that assume a complete ground-truth SOH window for every prediction. The explicit division between current-cycle features, historical electrochemical observations, and historical SOH also provides a clear inductive bias for online estimation, even if its building blocks are not individually novel.

2. The experimental evaluation is broader than the five representative examples in the main table.

Table 1 compares EDR-BAT against a range of statistical, neural, and physics-informed baselines on CALCE, MIT, NASA, Oxford, and an in-house dataset. The full model achieves the lowest reported MAE and RMSE on the five selected target cells, including NASA B0018 (MAE 0.67, compared with 0.89 for LSTM) and the selected MIT cell (0.47, compared with 0.52 for CNN). Importantly, Appendix C, Tables 4–8, additionally evaluate all 22 cells under leave-one-cell-out splits rather than reporting only one target per dataset. The use of repeated seeds, training-cell normalization, and a stated cycle-by-cycle inference protocol improves the credibility and interpretability of the study.

3. The ablations provide evidence that historical information and memory-aware training are useful.

Table 2 evaluates the removal of electrochemical-history queries, SOH-memory queries, and both historical branches, as well as the case where training uses only ground-truth SOH history (p=0). For example, on the in-house target, the full model obtains 0.82 MAE, while the p=0, without-electrochemical-query, and without-history variants obtain 1.34, 1.66, and 1.89, respectively. The differences are particularly informative because they relate directly to the paper's motivating problem of autoregressive deployment. The additional memory-length analysis in Appendix F and the sparse online calibration examples in Appendix E provide useful boundary information rather than solely adding more benchmark scores.

Critical weaknesses

1. The claimed electrochemical degradation representation is not sufficiently distinguished from standard neural history conditioning.

The central prior \theta_c is a learned series- or dataset-specific query refined by attention over recent feature embeddings, and the final prediction uses low-rank bilinear fusion and a separately trained SOH-memory branch. These are reasonable architectural choices, but the paper does not isolate their necessity. Table 2 removes historical interactions, yet does not compare the learnable prior against a shared generic query, a fixed query, or a capacity-matched history encoder without the proposed prior. Nor does it replace the low-rank bilinear head with matched concatenation, additive fusion, or an MLP. The contributions of the temporal, differential, integral, and statistical feature groups are also not individually evaluated.

The representation analysis is primarily qualitative. Figures 2 and 3 show latent trajectories and PCA structure but do not establish correspondence with physically measurable degradation processes such as impedance growth or loss of lithium inventory. In particular, if the five dataset-specific priors are obtained from separately trained models, their raw latent coordinates are not guaranteed to inhabit a comparable aligned space; geometric proximity in a joint PCA plot would then have limited physical meaning. Thus the evidence supports a useful structured predictor more strongly than it supports the proposed interpretation as a general electrochemical degradation state representation.

2. The complete cross-cell evaluation reveals important failures that are obscured by the selected main-table results.

Table 1 reports the method as best on one target cell per dataset, but Appendix C shows substantial variability. For example, on NASA B0006, EDR-BAT has MAE 2.01, while Trans-W reaches 1.10 and Transformer 1.25 (Table 6). On CALCE CS2_35, EDR-BAT obtains 0.65 versus 0.48 for Trans-W (Table 4); on MIT channel14, it obtains 0.32 versus 0.23 for LSTM-W (Table 5). The method also loses to simpler baselines on multiple in-house cells (Table 8). In NASA, averaging the four reported cell-wise MAEs gives approximately 1.12 for EDR-BAT versus 1.05 for Transformer and 1.06 for Trans-W. This matters because Section 4.2 suggests consistently lower error averaged over held-out cells; the paper should specify exactly which aggregation supports that statement and discuss where the model loses.

The ablation evidence itself has nonuniform boundaries: on Oxford, removing the electrochemical-history query achieves 0.18 MAE versus 0.21 for the full model, and removing both history branches obtains 0.19 (Table 2). The main claims should therefore distinguish average benefits from uniform improvements, and analyze whether failure cases track chemistry, cycling protocols, trajectory shape, or limited training-cell diversity.

3. Input construction and training–inference details are not specified tightly enough to support the strongest causal-online interpretation.

SOH is defined from a cell's available capacity, while the current-cycle input also includes capacity-related indicators (Section 3.1). The authors state that these are curve-derived features rather than the full-cycle capacity used for the target label, but do not provide a complete list and formula for all ten input features or establish precisely which portions of the current cycle must have been observed. This distinction is essential for ruling out close label proxies and for determining whether the claimed setting corresponds to realistic partial-cycle online estimation or to full-cycle post-hoc estimation.

Moreover, Eq. (15) describes sampling from ground-truth SOH and detached prediction memory, but the practical schedule for p, handling of initialization and chronological state updates, and implementation with minibatches should be more explicit; Appendix B reports p_{max}=1 without resolving all these details. The authors show aggregate MAE improvements over p=0 but do not directly measure how error grows with autoregressive rollout length. Finally, the discrepancy between the main and ablation “Full” results for CALCE and Oxford should be traced to explicit target-cell identifiers and experimental protocols rather than left for readers to infer from Appendix C.

Questions for authors

1. Regarding the mechanism and distinctive value of the degradation representation.

Can the authors compare the series-level learnable prior against a fixed or globally shared query, and the low-rank bilinear interaction against a parameter-matched MLP or additive-fusion head, under identical inputs and splits? What is the individual contribution of the four electrochemical feature groups? Is there quantitative evidence that the learned representation tracks independent indicators of physical aging, rather than simply encoding cycle progress and predictive correlations? For Figure 3, are the priors obtained from a shared, aligned latent space or from separately trained models, and what makes their inter-dataset distances interpretable?

2. Regarding the scope of the cross-cell generalization claim.

How do the authors reconcile the consistent wins in Table 1 with the losses reported in Appendix C, particularly NASA B0006, CALCE CS2_35, and MIT channel14? What exact averaging procedure underlies the statement that MAE and RMSE are consistently lower across held-out cells? Can the authors provide dataset-wise leave-one-cell-out summaries with uncertainty and an analysis of the conditions under which EDR-BAT is weaker than LSTM-W, Trans-W, or Transformer? Why does the Oxford ablation improve after removing electrochemical-history interaction?

3. Regarding feature availability, online inference, and experiment reproducibility.

Could the authors enumerate the ten input features with precise extraction formulas and indicate whether each is available from a partial or only a completed charge–discharge cycle? How are the capacity-derived inputs separated from the capacity measurement used in the SOH label, and how much accuracy remains if these inputs are removed? What is the exact p schedule and memory-update order during training? Do the CALCE and Oxford “Full” results in Tables 1 and 2 correspond to different held-out cells, and, if so, why were different target cells used for the main comparison and ablation? A rollout-length error analysis would also help substantiate the claimed reduction in long-horizon drift.

6 comments
AuthorsOpenReview ↗

Mechanism and Distinctiveness of the Degradation Representation

We thank the reviewer for the detailed and constructive assessment. We address the three groups of concerns in the following comments.

1. Mechanism and distinctiveness of the degradation representation.

Fair controls for the learnable prior and low-rank interaction.

The series-level query set is shared by all source cells within each training task and is adapted to the current cycle by querying the available electrochemical history. We added four condition-matched controls: Frozen query retains the query dimensions and attention structure but freezes the initial shared query set; History pooling replaces query-based attention with masked mean pooling and a parameter-matched encoder; Parameter-matched MLP retains the complete history-query process but replaces the low-rank interaction with concatenation-based MLP fusion; and Dynamic MLP, an early design explored during method development, converts the cycle-specific query set into the parameters of a three-layer MLP. All controls use identical inputs, SOH memory, training strategy, target cells, and data splits.

Controlled comparison of the learnable prior and interaction structure under identical inputs, training strategies, target cells, and data splits. Each entry reports MAE/RMSE in percentage points, with subscripts denoting standard deviations over five random seeds. Avg. denotes the equal-weight average across the five target cells.

Target cellFullFrozen queryHistory poolingParameter-matched MLPDynamic MLP
CS2_360.400.03/0.530.040.40_{0.03}/0.53_{0.04}0.460.02/0.580.050.46_{0.02}/0.58_{0.05}0.440.05/0.560.060.44_{0.05}/0.56_{0.06}0.450.05/0.590.050.45_{0.05}/0.59_{0.05}0.450.04/0.580.050.45_{0.04}/0.58_{0.05}
channel240.470.03/0.500.030.47_{0.03}/0.50_{0.03}0.510.04/0.540.040.51_{0.04}/0.54_{0.04}0.460.04/0.500.040.46_{0.04}/0.50_{0.04}0.520.06/0.560.060.52_{0.06}/0.56_{0.06}0.510.03/0.540.040.51_{0.03}/0.54_{0.04}
B00180.670.13/0.840.140.67_{0.13}/0.84_{0.14}0.900.12/1.060.090.90_{0.12}/1.06_{0.09}0.920.10/1.090.100.92_{0.10}/1.09_{0.10}0.960.12/1.120.110.96_{0.12}/1.12_{0.11}1.000.11/1.150.111.00_{0.11}/1.15_{0.11}
Cell40.180.02/0.220.020.18_{0.02}/0.22_{0.02}0.200.04/0.230.030.20_{0.04}/0.23_{0.03}0.180.04/0.220.040.18_{0.04}/0.22_{0.04}0.200.04/0.240.050.20_{0.04}/0.24_{0.05}0.190.03/0.240.030.19_{0.03}/0.24_{0.03}
In-house-30.820.07/0.960.080.82_{0.07}/0.96_{0.08}1.320.11/1.500.121.32_{0.11}/1.50_{0.12}1.210.14/1.460.171.21_{0.14}/1.46_{0.17}1.190.07/1.350.091.19_{0.07}/1.35_{0.09}1.160.10/1.390.131.16_{0.10}/1.39_{0.13}
Avg.0.51 / 0.610.68 / 0.780.64 / 0.770.66 / 0.770.66 / 0.78

The full model achieves the best or tied-best result in nine of ten target-cell--metric comparisons. Using the unrounded results, it reduces macro-average MAE/RMSE by 25.1%/22.0% over Frozen query, 20.9%/20.4% over History pooling, and 23.5%/21.0% over the Parameter-matched MLP, while also outperforming Dynamic MLP. Because all controls share identical information and training conditions, these results isolate the contributions of the learnable prior, history-querying representation, and low-rank interaction beyond historical access or generic fusion. Full results are included in Appendix E of the revised manuscript.

AuthorsOpenReview ↗

Feature Groups and Figure 3

Feature groups and physical interpretation.

Figure 1(a) functionally groups the dataset-specific charging features into time-domain, differential, integral, and statistical categories. These are feature groups rather than independent model branches or independently validated aging mechanisms. Because the concrete features differ across datasets and the four groups are physically correlated, removing one complete group would simultaneously change the dataset-specific input definition and would not isolate an independent physical mechanism.

The revised manuscript reports all ten features for each dataset, including their extraction intervals, definitions, and earliest availability, in Appendix A, “Input Feature Definitions and Availability.” The available datasets do not provide independent diagnostic labels for impedance growth, loss of lithium inventory, or loss of active material; we therefore do not claim a one-to-one correspondence between the learned representation and a specific aging mode. Establishing such correspondence is an important direction for future work.

Interpretation of Figure 3.

The five query sets are obtained from models trained separately on the five datasets and visualized through a common PCA projection. Figure 3 provides qualitative evidence that each model learns a non-collapsed, dataset-related multi-query organization that can subsequently adapt to battery history through self- and cross-attention. Because no cross-dataset latent alignment is imposed, we do not interpret absolute inter-dataset distances as physically calibrated distances between chemistries, capacities, or charging protocols. This interpretation boundary is now explicit in the revised discussion.

AuthorsOpenReview ↗

Scope of Cross-cell Generalization

2. Scope of cross-cell generalization.

The original phrase “consistently lower across held-out cells” was imprecise: EDR-BAT has the lowest overall average error but does not outperform every baseline on every held-out cell. Table 1 evaluates one specified target cell per dataset, whereas Appendix C reports complete leave-one-cell-out results over all 22 cells. For each target cell, we average five random-seed results; dataset-level values equally average the corresponding held-out-cell means, and the final row equally weights all 22 cells. Subscripts below denote variation across held-out-cell means. The revised manuscript replaces the original claim with: “Averaged equally over all 22 held-out cells, EDR-BAT achieves the lowest average MAE and RMSE,” and adds an Avg. row to every Appendix C table.

Dataset-wise summary of the complete leave-one-cell-out evaluation. Each entry reports MAE/RMSE in percentage points, with subscripts denoting standard deviations across held-out-cell means. The final row gives equal weight to all 22 held-out cells. Best results are marked in bold.

DatasetTransformerLSTM-WTrans-WEDR-BAT
CALCE0.470.06/0.620.110.47_{0.06}/0.62_{0.11}0.590.04/1.050.060.59_{0.04}/1.05_{0.06}0.460.07/0.630.110.46_{0.07}/0.63_{0.11}0.450.14/0.610.160.45_{0.14}/0.61_{0.16}
MIT0.540.17/0.570.190.54_{0.17}/0.57_{0.19}0.420.23/0.770.150.42_{0.23}/0.77_{0.15}0.470.15/0.550.160.47_{0.15}/0.55_{0.16}0.340.12/0.370.120.34_{0.12}/0.37_{0.12}
NASA1.050.15/1.280.231.05_{0.15}/1.28_{0.23}1.240.44/1.540.541.24_{0.44}/1.54_{0.54}1.060.11/1.330.151.06_{0.11}/1.33_{0.15}1.120.64/1.350.751.12_{0.64}/1.35_{0.75}
Oxford0.300.08/0.370.100.30_{0.08}/0.37_{0.10}0.480.13/0.610.160.48_{0.13}/0.61_{0.16}0.380.07/0.470.070.38_{0.07}/0.47_{0.07}0.220.05/0.290.090.22_{0.05}/0.29_{0.09}
In-house0.900.60/1.020.590.90_{0.60}/1.02_{0.59}1.780.97/2.401.301.78_{0.97}/2.40_{1.30}0.780.38/1.020.530.78_{0.38}/1.02_{0.53}0.750.15/0.890.110.75_{0.15}/0.89_{0.11}
Avg.0.580.36/0.700.410.58_{0.36}/0.70_{0.41}0.810.61/1.130.770.81_{0.61}/1.13_{0.77}0.590.30/0.740.390.59_{0.30}/0.74_{0.39}0.520.43/0.630.500.52_{0.43}/0.63_{0.50}

EDR-BAT achieves the lowest dataset-wise average MAE/RMSE on CALCE, MIT, Oxford, and In-house, but not NASA, where Transformer obtains 1.05/1.28 versus 1.12/1.35 for EDR-BAT. Across all 22 equally weighted cells, EDR-BAT nevertheless achieves the lowest overall MAE/RMSE of 0.52/0.63. The reviewer-identified exceptions are also confirmed: EDR-BAT is weaker than Trans-W on CALCE CS2_35 and NASA B0006, while MIT channel14 exhibits a metric-dependent trade-off (0.32/0.36 for EDR-BAT versus 0.23/0.63 for LSTM-W). The claim is therefore limited to overall average performance rather than superiority on every dataset or target cell. The complete per-cell results and failure-case discussion are provided in revised Appendix C.

Oxford ablation.

Although removing E-Q improves Cell8 from 0.21/0.26 to 0.18/0.23, the absolute difference is small. One plausible explanation is that the relatively smooth Cell8 trajectory can already be represented effectively by the current-cycle features, making the additional electrochemical-history interaction partly redundant for this target cell. However, the matched-target experiment on Cell4 shows the opposite trend: Full obtains 0.18/0.22, compared with 0.19/0.25 for w/o E-Q and 0.24/0.30 for w/o Hist. The Cell8 improvement is therefore target-cell-specific rather than evidence that the E-Q interaction is generally ineffective on Oxford.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revised Manuscript Uploaded

We have uploaded a revised manuscript incorporating the clarifications and additional experiments described in our responses. All new and revised content is highlighted in blue for ease of review.

The main revisions include:

  • detailed definitions, extraction intervals, computation procedures, and availability of all ten charging features for each of the five datasets, together with a direct capacity-proxy sensitivity analysis (Appendix A);
  • clarified initialization, probability scheduling, memory updates, and mini-batch implementation of memory-anchored training (Appendix B);
  • complete leave-one-cell-out evaluation over all 22 cells, including the closely related LSTM-LAM baseline, dataset-level averages, and failure-case analysis (Appendix C);
  • four condition-matched structural controls on the five main target cells, additional matched-target ablations on CALCE and Oxford, and an autoregressive rollout-horizon analysis (Appendix E);
  • refined discussion of the learned-prior visualization, the scope of cross-cell generalization, and the conclusions supported by the ablation results.

Detailed quantitative results and point-by-point explanations are provided in our reviewer-specific responses. We sincerely thank the reviewers for their constructive feedback and the area chairs for their consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.