Machine learning bridging battery field data and laboratory data

Hao Liu; Haoyi Jiang; Lei Zhang; Tong Li; Xiaoping Ouyang; Xingkai Wang; Yanbin Zhao; Zhenfei Ling; Zhihua Deng

REVIEW 4 major objections 7 minor 4 references

Machine learning bridging battery field data and laboratory data

T0 review · 4 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Two impedance readings can rebuild laboratory battery test curves

desk verdict The two-point impedance reconstruction works on held-out lab cells, but the paper calls lab-discharge EIS 'field data' and never tests real field conditions. read the letter →

arxiv 2505.05364 v2 pith:XF2B2ZS6 submitted 2025-05-08 stat.AP

Yanbin Zhao , Hao Liu , Zhihua Deng , Tong Li , Haoyi Jiang , Zhenfei Ling , Xingkai Wang , Lei Zhang

show 1 more author

Xiaoping Ouyang

This is my paper · ORCID

classification stat.AP

keywords lithium-ionbatteryfielddatalaboratoryelectrochemicalimpedancespectroscopymachinelearningdiagnosisprognosisstateofhealth

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes a machine-learning bridge so that laboratory-trained battery diagnostics can be used on batteries in real vehicles and storage systems. Its claim is that just two real-impedance readings from a field battery—one at a medium frequency and one at a high frequency—are enough to predict the laboratory real-impedance curve, the laboratory charge and discharge Q/V curves, and the laboratory voltage-relaxation curve. With those laboratory curves reconstructed, existing laboratory data-driven methods can directly diagnose remaining capacity and prognose remaining life in the field. On two open datasets covering 249 nickel-manganese-cobalt (NMC) cells, the paper reports test-set MAPEs of 0.85% for the laboratory real-impedance curve, 4.72% for the charge curve, and 2.69% for the discharge curve; remaining-capacity prediction from predicted laboratory data gives MAPEs of 1.89% and 3.08% on the two datasets. If the claim holds, field battery health management would no longer need to collect massive, privacy-sensitive historical field data.

What carries the argument

The load-bearing object is the two-frequency real-impedance measurement, treated as a low-dimensional signature that encodes enough aging information to regenerate laboratory data. The two frequencies are chosen from the training set's laboratory Re/f curves by k-means clustering, one from the medium range (1–100 Hz) and one from the high range (100 Hz–1 kHz), giving $f_1=10$ Hz, $f_2=312.5$ Hz for Dataset 1 and $f_1=5.53$ Hz, $f_2=193.03$ Hz for Dataset 2. Random forest regression models carry every mapping step, and the argument rests on cited evidence that real impedance transfers across SOC and temperature at fixed frequency and aging state, that steady-state EIS can be measured online during dynamic operation, and that EIS, charge/discharge Q/V curves, and relaxation curves share overlapping aging-mode information. The final diagnostic and prognostic step uses best two-point features extracted from the predicted laboratory curves.

What would settle it

Take the trained bridge and apply it to genuine in-vehicle field EIS measurements collected under dynamic current, fluctuating temperature, and unknown exact SOC, then compare the predicted laboratory curves against controlled laboratory measurements of the same cells; the transfer claim is falsified if the error grows well beyond the reported MAPEs of 0.85%, 4.72%, and 2.69%.

Watch

Extended reading notes

Core claim

The central claim is that the real part of impedance measured at two carefully chosen frequencies is a compact bridge between field and laboratory battery data. The pipeline first converts the two field real impedances $R_{e1}^F$ and $R_{e2}^F$, together with field SOC and temperature, into laboratory real impedances $R_{e1}^L$ and $R_{e2}^L$ at a specified SOC and temperature using random forest models. A second random forest reconstructs the mid-high-frequency laboratory real-impedance curve from those two laboratory values, and further models convert that curve into the laboratory charge Q/V curve, discharge Q/V curve, and relaxation V/t curve. Best two-point features (BTPFs) are then extracted from the predicted laboratory curves and fed into models that predict remaining capacity and remaining cycles or days. The paper reports that on a 76-cell test set from the first dataset the reconstructed laboratory real-impedance curve has 0.85% MAPE, the charge Q/V curve has 4.72%, and the discharge Q/V curve has 2.69%, while downstream diagnosis and prognosis based on predicted laboratory data remain close to those based on measured laboratory data.

Load-bearing premise

The load-bearing premise is that 'field data' as the paper defines it—steady-state EIS measured in the laboratory during discharge at controlled temperatures and SOCs—represents real field batteries; if true field EIS measured under dynamic load, noise, and unknown state boundaries has a different distribution, the trained mappings may not transfer.

Editorial extensions

If this is right

Laboratory-trained diagnostic and prognostic models become directly applicable to field batteries, since the predicted laboratory curves match the input format those models were built on.
Field battery health monitoring can drop its dependence on massive historical field operation data, reducing development cost and avoiding user-privacy data such as driving and charging histories.
The two impedance readings are active, user-controlled measurements that take milliseconds, so diagnosis and prognosis can be performed on demand in the field.
The same bridge can predict regular offline performance characterization data such as EIS and Q/V curves during battery life testing, lowering the cost of life testing.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural stress test beyond the paper is to apply the trained bridge to true field EIS recorded under dynamic load, noise, and uncertain SOC and temperature; the paper's 'field' data are steady-state EIS measured in the laboratory during discharge, so this transfer is not yet demonstrated.
If the bridge generalizes across cell chemistries, formats, and pack designs, the two-frequency real impedance could become a standardized health signature for field batteries, an extension the paper does not itself test.
The predicted Q/V curves could feed physics-based aging-mode identification, such as loss of lithium inventory and loss of active material, since the paper's cited evidence indicates Q/V and EIS carry the same aging modes.
The same mapping idea could translate data between different laboratory protocols, making legacy laboratory datasets interoperable without new experiments.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the paper demonstrates a genuinely useful lab-to-lab mapping — two real impedance values predict full lab curves on held-out cells — but it does not test the claimed field bridge on actual field data. The word 'field' is doing too much work.

What's new: the specific pipeline of using two impedance points (one mid, one high frequency) to reconstruct the lab impedance curve, charge/discharge Q/V, and relaxation curves is not in the prior literature. Prior work predicted EIS from Q/V or vice versa; this two-point compression plus reconstruction is a new configuration. The evaluation is real out-of-sample: 76 test cells in Dataset 1, 10 in Dataset 2, with the impedance MAPE around 0.85% and discharge Q/V MAPE in the low single digits. That's solid empirical evidence that the mapping works within the distribution it was trained on. The DRT comparison is a nice sanity check.

The soft spots are real but not fatal. Most importantly, the 'field' data is steady-state EIS measured in the lab during discharge at controlled SOC and temperature — not field-collected impedance under dynamic load with noise and unknown state boundaries. The paper cites online EIS methods (Zhu et al., Sihvo et al.) but does not implement them, so there is zero evidence about transfer under true field conditions. Random forests are nonparametric; if real field EIS has a different distribution, the mapping can degrade. The prognosis results are also weak: remaining-life MAPE of 17–31% on Dataset 1 and no error bars anywhere. Removal of incomplete RPT samples is mentioned but not analyzed. And code/data are promised only after publication, which limits reproducibility checks.

I disagree with the stress-test on one point: it says the central claim is untested on true field EIS. That's fair, but the core mapping itself is not circular — held-out cell predictions are genuine. The frequency selection and bin boundaries are tuned on training data but evaluated on held-out cells, so the claimed accuracy numbers stand for the proxy distribution.

Bottom line: this is a competent engineering paper with a clear contribution to battery PHM, but the title and abstract overstate the field bridge. A serious referee should send it for review and request a clear separation of the lab-to-lab translation result from the unvalidated field-transfer claim, plus real field validation or at least a distribution-shift analysis. I'd cite it for the two-point impedance reconstruction idea, and bring it to reading group as an example of a well-executed held-out evaluation with an overbroad external validity claim.

Referee Report

4 major / 7 minor

Summary. The paper proposes a machine-learning pipeline that, from two measured real-part impedances (at a medium and a high frequency) of a battery in the field, together with the field SOC and temperature, predicts the laboratory real-impedance spectrum, charge/discharge capacity-voltage curves, and relaxation voltage-time curve. These predicted laboratory data are then used with the authors' 'best two-point features' method for remaining-capacity diagnosis and remaining-life prognosis. The method is evaluated on two open-source aging datasets (249 NMC cells) with held-out test cells, reporting out-of-sample MAPEs of 0.85% for the lab Re/f curve, 4.72% for charge and 2.69% for discharge curve predictions on Dataset 1, and diagnostic/prognostic MAPEs based on predicted lab data.

Significance. If the claimed transfer between field and laboratory data were established, the approach would be significant: it would allow the large body of laboratory-data-driven battery diagnostics to be applied with only two impedance measurements, avoiding the need for massive field datasets and addressing privacy concerns. The paper's strengths include genuine held-out test-cell evaluation, the use of two public datasets, and remarkable reconstruction accuracy for the laboratory impedance curve (0.85% MAPE). The central limitation, however, is that the 'field' data are themselves laboratory discharge EIS; the claimed applicability to real field batteries is thus not verified by the experiments as designed. The prognosis results also show large errors that are not contextualized against any baseline. With appropriate scope narrowing and additional analysis, the method's core idea is worth pursuing.

major comments (4)

[Results, Datasets / Abstract] The paper defines 'field data' in Dataset 1 as steady-state EIS measured during discharging at all SOC and operating temperatures, and in Dataset 2 as steady-state EIS at SOC<90% and T=25°C during discharging. These are laboratory measurements, not real field data. The cited online EIS techniques (refs [56,57]) are mentioned but neither implemented nor used in the experiments. Because the random forest mappings are trained and tested on this lab-discharge proxy, the reported MAPEs (0.85%, 4.72%, 2.69%) only quantify performance on the proxy. Real field EIS is collected under dynamic load, noise, uncertain SOC/temperature, and variable rest states; the paper provides no evidence that the proxy distribution matches the real field distribution. This is a load-bearing gap for the headline claim that the method 'makes all laboratory data-driven methods applicable to field battery diagnosis and prognosis.' The authors should either validate on a genuine field dataset or explicitly restrict the claims to laboratory discharge EIS.
[Step 5: Diagnosis and prognosis / Figures 11-12] The remaining-life prognosis results based on predicted laboratory data have MAPEs of 30.61% (Dataset 1) and 16.91% (Dataset 2), and the diagnostic MAPE for Dataset 2 is 3.08%. These are relatively large errors, and the paper does not compare against any existing field-data-driven method on the same data, despite the Discussion claiming 'higher accuracy' than such methods. The prognosis, which is a central component of the claimed contribution, is therefore not convincingly established. The authors should provide a baseline comparison or substantially temper the accuracy claims for prognosis.
[Abstract / Step 2: Laboratory Res prediction] The abstract states that 'only two field real impedances' are needed, but the prediction models in Step 2 take as inputs Re1_F (or Re2_F), SOC_F, and T_F. Thus the method requires not only the two impedances but also accurate field SOC and temperature, which are themselves difficult to obtain in real field use. The paper does not analyze the sensitivity of the predictions to errors in SOC_F or T_F, nor does it describe how these inputs would be obtained in practice. This oversimplification of the input requirements should be corrected and the robustness to input uncertainty discussed.
[Datasets / Supplementary Note 2] The paper states that 'the RPT data of some cells are incomplete, and we have removed these incomplete data' but does not report the number of removed samples or cells, nor does it analyze whether removal is balanced across test conditions. If incomplete RPT samples are non-randomly distributed (e.g., early failures or conditions with more frequent RPTs), the held-out evaluation could be biased. The authors should quantify the removals and demonstrate that the remaining data are representative.

minor comments (7)

[Datasets / throughout] The abbreviation 'PRT' is used for 'reference performance test' in the Datasets section, while 'RPT' is used everywhere else; please standardize the terminology.
[Figure 5 and other figure captions] The text uses 'MEA' in several places where 'MAE' is intended (e.g., Figure 5 caption and the corresponding main-text sentences); please correct this typo.
[Figure 11 and Figure 12 captions] The subpanels in these captions skip from (c) to (e), likely omitting (d); please fix the labeling.
[Step 5] The text refers to 'remaining days' for Dataset 1 and 'remaining cycles' for Dataset 2, but the figures and surrounding text use 'remaining cycles' interchangeably; the target variable should be defined consistently for each dataset.
[Supplementary Note 6] The hyperparameter grids for the random forest models are not specified; the reader is only told that grid search was used. Please provide the exact parameter ranges or the final hyperparameters for each model to make the results reproducible.
[Supplementary Note 5] The BTPF method is described only by reference to the authors' unpublished arXiv preprint [59]; to make the paper self-contained, the two-point feature extraction should be summarized in the main text or supplementary material, or the reference should be updated to a peer-reviewed version if available.
[Discussions] The claim that the proposed method 'can achieve higher accuracy than the field data-driven method when using the same amount of data' is not supported by any experiment or citation; please rephrase it as a conjecture or support it with a direct comparison.

Circularity Check

1 steps flagged · score 3.0 of 10

Mostly independent out-of-sample regression, but the Step 3 Re/f-curve metric partly predicts its own two input impedances by construction; no full circularity.

fitted input called prediction [Bridging Method, Step 3; Supplementary Note 3]
"The input of the ML model is 𝑅𝑒1𝐿 and 𝑅𝑒2𝐿, and the output is the laboratory Re/f curve in the mid-high frequency range. ... the inputs of the ML model are 𝑅𝑒1𝐿 and 𝑅𝑒2𝐿, and the outputs are laboratory Re/f curve in the frequency range of [2.08 Hz, 1000 Hz] (with 16 Re/f data)."

The Step 3 target is the 16-point Re/f curve over 2.08-1000 Hz, and f1=10 Hz and f2=312.5 Hz (Dataset 1) are two of those points. Hence Re(f1) and Re(f2) in the output are exactly the inputs Re1_L and Re2_L. Fitting a regressor to this target teaches it to copy its own inputs at those two output coordinates, so the reported curve MAPE (0.85%) and DRT verification include two 'predicted' points that equal the model's inputs by construction. The remaining 14 points and Step 4 remain genuinely predicted; the circularity is partial.

full rationale

Steps 2, 4, and 5 are evaluated on held-out cells, so the core field-to-laboratory mapping is not circular. Step 1 (frequency selection) and Step 2 (SOC/Re binning) use training data only; the field-to-laboratory Re regression is a true out-of-sample prediction. Step 4's charge/discharge Q/V and relaxation predictions are nontrivial functions of the Re/f curve, and Step 5's diagnostic/prognostic models are trained on measured laboratory features and applied to predicted features without fitting to the target. The only load-bearing self-citation is [59] for BTPF extraction, but the Supplementary Information re-derives the BTPF selection from the training set, so the citation is not load-bearing. The concern that 'field data' in both datasets is defined as laboratory steady-state discharge EIS rather than true dynamic field EIS is an external-validity limitation, not a circularity. Overall, the derivation is mostly self-contained; the Step 3 input-output overlap partially inflates the headline impedance-curve accuracy, warranting a modest score.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central empirical mapping is genuine held-out prediction, but the method is heavily parameterized: preset frequencies, bin boundaries, many per-bin RF models, and selected BTPF points all come from training data. The most consequential unstated assumption is that lab discharge EIS represents real field data.

free parameters (6)

Preset frequencies f1, f2 = 10 Hz and 312.5 Hz (Dataset 1); 5.53 Hz and 193.03 Hz (Dataset 2)
Selected by k-means clustering on training Re/f curves; not derived from physics.
SOC_F bin boundaries = 10 intervals of 10 percentage points
Chosen by hand in Supplementary Note 2 to improve per-bin model accuracy.
Re_F bin boundaries = 4 intervals per Re in Dataset 1; 6 intervals in Dataset 2
Chosen by hand from training set ranges; each bin gets its own random forest model.
RF hyperparameters = Tuned by grid search, exact values not reported in main text
Supplementary Note 6 describes grid search over n_estimators, max_depth, min_samples_leaf, max_features, and subsample.
BTPF two-point features = Specific voltage or frequency pairs selected per curve type
Selected from training set based on maximum Pearson correlation with target; details in Supplementary Figures 6-8.
Difference baseline (first RPT) = Subtract first RPT from each RPT
Choice of difference data affects BTPF computation; no sensitivity analysis.

assumptions (4)

domain assumption Re at any SOC and temperature can be predicted from Re at any other SOC and temperature at the same frequency and aging state
Invoked in Introduction and Step 2, based on Wang et al. [45] and Barcellona et al. [46].
domain assumption EIS, charge/discharge Q/V, and relaxation V/t curves share the same aging-mode information, so each can be predicted from the others
Invoked in Introduction; partly supported by Guo et al. [50] for Q/V and EIS, but relaxation prediction is labeled 'boldly speculate'.
ad hoc to paper Laboratory discharge EIS at controlled conditions is representative of field impedance data
Used to construct 'field' datasets in the Datasets section; not validated against real field data.
domain assumption Random forest models trained on training cells generalize to held-out cells from the same test conditions
Standard ML assumption; training and test cells share conditions, so extrapolation to unseen conditions is untested.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine learning bridging battery field data and laboratory data." pith.science (2026). https://pith.science/paper/XF2B2ZS6

@misc{pith2026250505364,
  author       = {Pith},
  title        = {Pith review of: Machine learning bridging battery field data and laboratory data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XF2B2ZS6}},
  note         = {Machine review of arXiv:2505.05364}
}

read the original abstract

Aiming at the dilemma that most laboratory data-driven diagnostic and prognostic methods cannot be applied to field batteries in passenger cars and energy storage systems, this paper proposes a method to bridge field data and laboratory data using machine learning. Only two field real impedances corresponding to a medium frequency and a high frequency are needed to predict laboratory real impedance curve, laboratory charge/discharge curve, and laboratory relaxation curve. Based on the predicted laboratory data, laboratory data-driven methods can be used for field battery diagnosis and prognosis. Compared with the field data-driven methods based on massive historical field data, the proposed method has the advantages of higher accuracy, lower cost, faster speed, readily available, and no use of private data. The proposed method is tested using two open-source datasets containing 249 NMC cells. For a test set containing 76 cells, the mean absolute percentage errors of laboratory real impedance curve, charge curve, and discharge curve prediction results are 0.85%, 4.72%, and 2.69%, respectively. This work fills the gap between laboratory data-driven diagnostic and prognostic methods and field battery applications, making all laboratory data-driven methods applicable to field battery diagnosis and prognosis. Furthermore, this work overturns the fixed path of developing field battery diagnostic and prognostic methods based on massive field historical data, opening up new research and breakthrough directions for field battery diagnosis and prognosis.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

4 extracted references · 3 canonical work pages

[1]

Two points are enough

Hao Liu, Yanbin Zhao, Huarong Zheng, Xiulin Fan, Zhihua Deng, Mengchi Chen, Xingkai Wang, Zhiyang Liu, Jianguo Lu & Jian Chen. Two points are enough, arXiv, 2408.11872 (2024)

work page Pith review arXiv 2024
[2]

Grid search, random search, genetic algorithm: a big comparison for NAS, arXiv, 1912.06059 (2019)

Petro Liashchynskyi & Pavlo Liashchynskyi. Grid search, random search, genetic algorithm: a big comparison for NAS, arXiv, 1912.06059 (2019)

arXiv 2019
[3]

Random forests, Machine learning, 45, 5-32 (2001)

Leo Breiman. Random forests, Machine learning, 45, 5-32 (2001)

work page 2001
[4]

RandomForestClassifier -scikit-learn 1.5.2 documentation, https://scikit- learn.org/1.5/modules/generated/sklearn.ensemble.RandomForestClassifier.html

work page

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Two points are enough

Hao Liu, Yanbin Zhao, Huarong Zheng, Xiulin Fan, Zhihua Deng, Mengchi Chen, Xingkai Wang, Zhiyang Liu, Jianguo Lu & Jian Chen. Two points are enough, arXiv, 2408.11872 (2024)

work page Pith review arXiv 2024

[2] [2]

Grid search, random search, genetic algorithm: a big comparison for NAS, arXiv, 1912.06059 (2019)

Petro Liashchynskyi & Pavlo Liashchynskyi. Grid search, random search, genetic algorithm: a big comparison for NAS, arXiv, 1912.06059 (2019)

arXiv 2019

[3] [3]

Random forests, Machine learning, 45, 5-32 (2001)

Leo Breiman. Random forests, Machine learning, 45, 5-32 (2001)

work page 2001

[4] [4]

RandomForestClassifier -scikit-learn 1.5.2 documentation, https://scikit- learn.org/1.5/modules/generated/sklearn.ensemble.RandomForestClassifier.html

work page