Machine learning accelerates fuel cell life testing

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

REVIEW 5 major objections 5 minor 5 references

Machine learning accelerates fuel cell life testing

T0 review · 5 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims that fuel cell life testing can be accelerated 30-fold by predicting end-of-test aging indicators from performance differences measured within the first 1,000 stress cycles, and that full electrochemical characterization…

desk verdict A promising adaptation of battery early-prediction to fuel cells, but the headline 30x acceleration rests on an 11-cell split with no uncertainty quantification; the evidence is thinner than the abstract implies. read the letter →

arxiv 2504.18835 v2 pith:B5537FEM submitted 2025-04-26 stat.AP

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

show 1 more author

Xiaoping Ouyang

This is my paper · ORCID

classification stat.AP

keywords PEMFClifetestingacceleratedpredictionperformancecharacterizationdataelectrochemicalimpedancespectroscopydifferencecurvesSISSOdescriptorrandomforest

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

This paper claims that fuel cell durability testing can be shortened dramatically without giving up predictive accuracy. Its LP-ALT method uses the difference between performance curves measured at the 0th and 1,000th accelerated test cycle to predict aging indicators at the 30,000th cycle, reaching a minimum $R^2$ of 0.89 and a 30-fold acceleration. Its companion PCDP method predicts full electrochemical characterization curves from just four impedance values at two frequencies, so that slow offline measurements could be replaced by a quick impedance reading. If these claims hold, life-testing campaigns for fuel cells and possibly other devices could run in days rather than months. The paper also reports generalization to a water electrolysis cell and to capacitors.

What carries the argument

The central mechanism is the early-stage difference curve: standardized polarization curves, cyclic voltammetry curves, and impedance spectra from two early test stages are subtracted to produce curves that capture how the cell has already begun to change. From these difference curves, the method constructs candidate two-point features (absolute differences between every pair of points) and uses a compressed-sensing descriptor search called SISSO to select the few features most correlated with each target aging indicator at the 30,000th cycle. Random forest regressors then map those selected features, along with impedance-derived features, to the final aging-indicator values. For the PCDP method, the reconstruction mechanism is a clustering-selected pair of frequencies whose four real and imaginary impedance values feed random forest models that output the full characterization curves.

What would settle it

A fresh fuel cell cohort with catalyst compositions and accelerated stress protocols outside the training range, run to 30,000 cycles, would settle the claim: if the early-difference features and the same random forest protocol yield test-set $R^2$ below 0.89 for limiting current, total mass transport resistance, or electrochemically active surface area, the 30-fold acceleration claim fails; the four-impedance reconstruction would also fail if crossover-current estimates from predicted linear sweep voltammetry curves degrade materially on cells where membrane degradation dominates.

Watch

Extended reading notes

Core claim

On an open-source life-test data set of 42 polymer electrolyte membrane fuel cells, the LP-ALT method predicts limiting current, total mass transport resistance, and electrochemically active surface area at the 30,000th cycle with $R^2$ values of 0.94, 0.93, and 0.89, using only data from the 0th and 1,000th cycles. The PCDP method predicts electrochemical impedance spectroscopy, current-voltage curves, cyclic voltammetry curves, and linear sweep voltammetry curves from four impedances at two clustered frequencies; aging indicators derived from the predicted curves lose only 0.01 to 0.06 in $R^2$ compared with using measured curves. The authors interpret these results as showing that very early degradation differences carry enough information to forecast the end-of-test state, and that the same recipe transfers to other devices such as capacitors.

Load-bearing premise

The method assumes that what changes in the first 1,000 stress cycles is enough to determine the cell's state at the 30,000th cycle, and that a model trained on 28 cells will transfer to unseen cells with different catalyst-layer designs and test conditions.

Editorial extensions

If this is right

A 30,000-cycle fuel cell life test could be stopped after 1,000 cycles, cutting test time from months to days and reducing cost and labor.
Routine offline characterization could be replaced by a four-impedance reading, with the resulting aging-indicator predictions losing at most 0.06 in $R^2$ relative to measured curves.
Combining PCDP and LP-ALT enables diagnosis and prognosis of fuel cells and their components from early-life data alone.
The same LP-ALT recipe appears to transfer across device types: capacitor remaining capacitance at 5,105 hours was predicted from 125 hours of data with an acceleration ratio above 40 and a mean absolute percentage error of 2.45 percent.
The 30-fold acceleration ratio is limited by the available data, so denser early sampling could push the ratio higher.

Reading between the lines

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

Because LP-ALT needs only two early time points, the natural boundary of the method is how early the second measurement can be while still carrying predictive signal; testing cycle 100 instead of cycle 1,000 would reveal whether the 30-fold ratio is data-limited or signal-limited.
The four-impedance PCDP result suggests a sparse-sensing view of impedance spectroscopy: if two frequencies suffice for full characterization, the information dimension of the aging state is low, and frequency selection could be adapted online to each cell rather than fixed once by clustering.
If the selected two-point difference features correspond to specific physical losses, such as ohmic resistance or mass-transport resistance, the method could be used not only to forecast an aging indicator but also to identify which degradation mechanism dominates.
The capacitor transfer result implies that LP-ALT may work whenever a scalar health indicator is governed by a single dominant degradation process, making lithium-ion batteries and electrolyzers natural next testbeds.

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

Zhao et al. adapt Severson's early-prediction trick to PEM fuel cells and add a data-prediction shortcut: four impedances predict full EIS, I-V, CV, and LSV curves via random forests. That combination is new for fuel cells. The held-out test cells are genuinely separate, and the cross-domain checks (PEMWE for PCDP, capacitors for LP-ALT) show good instincts. If the method holds up, it would meaningfully cut the cost of fuel cell life testing.

Where it gets shaky is the size of the evidence. The headline 30x acceleration is computed from a single split of 11 test cells. No confidence intervals, no permutation tests, no repeated random splits. With n=11, one or two cells can move R^2 a lot, so the claim that min R^2 is at least 0.89 is not robust. The TP-SISSO feature selection runs on roughly 1,275 candidate EIS features and ~4,950 CV features with only 28 training cells, k=6, and no nested cross-validation; that is a real overfitting risk. The authors need to show stability via bootstrap or leave-one-out and report prediction intervals on the test set.

There is also an internal inconsistency worth probing: CV curve prediction R^2 is 0.50, yet ECSA predicted from those CVs has R^2 0.92. Either the CV errors occur in regions that don't matter for ECSA, or the RF exploits something else; the paper doesn't explain. Similarly, LSV curve R^2 is 0.43 but crossover current R^2 is 0.88. These need a mechanistic or at least a clear empirical explanation.

The capacitor and PEMWE generalization results are nice but small; the capacitor R^2 is 0.74, which is not as strong. Code and data are promised 'after publication,' which is fine but means the results aren't independently checkable now.

Who is this for? Researchers in fuel cell durability testing who want a quick screening tool. It is a serious engineering paper, not a methodological breakthrough. The central idea is plausible and the data splits are mostly handled carefully. But the evidence for the specific 30x number is thinner than the abstract suggests. I'd send it to peer review, but with the recommendation that the authors add uncertainty quantification, nested feature selection, and a comparison to simpler baselines (e.g., raw early-cycle values). If those hold up, it's a solid contribution.

Referee Report

5 major / 5 minor

Summary. The paper proposes two machine learning methods to accelerate polymer electrolyte membrane fuel cell (PEMFC) life testing. The first method, PCDP, predicts full performance characterization data (EIS, I-V, CV, LSV curves) from only four impedance values at two preset frequencies. The second method, LP-ALT, predicts aging indicators (limiting current, total mass transport resistance, ECSA, crossover current) at the 30,000th accelerated test cycle using PCD measured at the 0th and 1,000th cycles, yielding an acceleration ratio of 30. The methods are validated on open-source datasets: 39 PEMFCs with complete EIS data, a PEMWE cell, and 24 capacitors. Reported test-set R^2 values for LP-ALT are 0.94, 0.93, and 0.89 for the three PEMFC aging indicators, and the PCDP results show small R^2 losses (0.01-0.06) for aging indicators derived from predicted versus measured PCD.

Significance. If the claims hold, the methods could substantially reduce the time and cost of PEMFC durability testing, and the transfer to capacitors suggests broad applicability. Strengths of the manuscript include the use of public datasets, a train/test split that is genuinely held out, feature selection performed on the training set only, and the demonstration of cross-domain generalization to a PEMWE cell and electrolytic capacitors. The paper also contains detailed method descriptions and extensive supplementary material. However, the empirical evidence is not yet statistically robust: the headline R^2 values are point estimates from small test sets with no confidence intervals, and the central 30x acceleration ratio depends on a single hand-picked early test stage. The statistical strength of the evidence is currently disproportionate to the strength of the claims.

major comments (5)

[LP-ALT method, Step 4; Results (Fig. 7)] The central claim that the minimum R^2 for aging indicators is 'not less than 0.89' rests on only 11 held-out PEMFCs and no confidence intervals. With candidate feature spaces of 1,275 two-point EIS differences and 4,950 two-point CV differences, SISSO selection on 28 training cells followed by RF hyperparameter tuning is highly susceptible to overfitting; one or two test cells can move the point estimates substantially. The paper should provide bootstrap or leave-one-out confidence intervals for R^2, report sensitivity to the specific test split, and ideally use nested cross-validation for the full feature-selection-plus-regression pipeline.
[LP-ALT method, Step 1; Results] The acceleration ratio of 30 is the ratio 30,000/1,000 ATC and is therefore determined entirely by the hand-picked T2 stage of 1,000 cycles. The manuscript offers no analysis of how prediction accuracy varies with T2 (e.g., T2=500, 2,000, 5,000 cycles), so the 30x figure is not established as a robust property of the method. A systematic T2 sweep with reported R^2 and uncertainty would be needed to support the acceleration claim.
[PCDP method, Step 3; Figure 4] Direct prediction of CV curves has R^2=0.50 and LSV curves have R^2=0.43, yet the downstream aging indicators ECSA and I_cross lose only 0.05 and 0.06 in R^2 relative to measured PCD. This discrepancy is unexplained and suggests that the aging-indicator evaluation is not very sensitive to curve-level errors, likely because the indicator models are trained and tested on the same two stages (0 and 30k cycles) with a narrow target range. The authors should report the target distributions, show scatter plots of predicted versus measured indicators, and analyze how errors propagate from predicted PCD to indicator estimates.
[PCDP method, Step 3; Supplementary Note 4] The aging indicator models for R_O2,total, I_lim, ECSA, and I_cross are trained using only the 0 and 30k cycle stages per cell. The evaluation of predicted versus measured PCD is thus confined to the same two stages used in training, which does not constitute an independent test of the PCDP method for intermediate life stages. The manuscript should clarify the scope of the claim and, where PCD at intermediate stages exists (e.g., CV curves at 10, 100, 1k, 3k, 5k, 10k, 20k cycles), use those stages for a more stringent evaluation.
[Results, Step 4; Abstract] The main text states that the minimum R^2 is 'not less than 0.9', while the reported ECSA result is R^2=0.89 and the abstract states 0.89. This inconsistency must be resolved. More importantly, the claim 'not less than' requires a statistical lower bound, which is absent. The authors should either report a proper confidence interval for the minimum R^2 or rephrase the claim to reflect the point estimate.

minor comments (5)

[Throughout] The term 'MEA' is used for 'mean absolute error' in multiple places (e.g., PCDP Step 2, Step 3); this should be 'MAE'.
[Results, Datasets] The test set list for Dataset 3 contains 'ES10C3' twice; one of the entries is likely a different capacitor (e.g., 'ES10C4').
[Results, Datasets; LP-ALT Step 4] There are citation errors: 'please refer to [35]' for Dataset 3 details should be [36], and 'the RF regression model [36]' should be [37] (Breiman).
[PCDP method, Step 2] The sentence 'The MEA, RMSE, MAPE, and R^2 of the EIS prediction results the on the test set' contains a grammatical error ('results the on').
[Data availability, Code availability] Source data and code are stated to be available only after publication; providing them as supplementary material during review would substantially strengthen reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

The ML predictions rest on clean holdout splits, but the headline 30x acceleration ratio is the arithmetic ratio of the self-chosen T3 and T2 stages, so that specific number is definitional rather than derived.

self definitional [LP-ALT method, Step 1 ('PCD collection') and Step 4 ('Life test prediction'); also 'Generalization of LP-ALT method' for Dataset 3.]
"In Dataset 1, T1 stage is the test time corresponding to the 0th ATC (i.e. 0 s), and T2 stage is the test time corresponding to the 1000th ATC (i.e. 6500 s). ... Compared with the complete life test of 30,000 ATCs, the LP-ALT method only requires 1,000 ATCs, with an acceleration ratio of 30 times."

The acceleration ratio is not a model output or an empirical prediction: with T1=0, T2=1,000 and T3=30,000 ATCs, the ratio (T3-T1)/(T2-T1) equals 30 by the definitions chosen in Step 1. Any choice of T2 would fix the ratio as T3/T2, so the headline '30 times' is a restatement of the stage selection, not a result derived from the data. The empirical content (R2 of the aging-indicator predictions) is separate and remains a genuine holdout evaluation; only the ratio itself is circular.

full rationale

The paper's predictive pipeline is otherwise self-contained. The PCDP and LP-ALT models are trained on the 28-cell training split with TP-SISSO feature selection performed on training data only, and the reported R2 values are evaluated on the 11 held-out cells, so the aging-indicator predictions do not reduce to fitted inputs. The aging-indicator evaluation via models trained on measured PCD and applied to predicted PCD is a standard surrogate assessment, not a circular definition. The self-citation to [38] for TP-SISSO is not load-bearing because the feature-extraction procedure is fully specified in Supplementary Note 7 and independently implemented with the external SISSO algorithm [39]. The only constructed quantity is the acceleration ratio, which equals the quotient of the authors' chosen T3 and T2 stages (30,000/1,000 = 30; 5,105.5/125 > 40 on Dataset 3). That number is an arithmetic label rather than an emergent prediction, so the overall circularity is partial but the central predictive claim has independent content.

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

The paper's central claims rest on standard ML hyperparameters and on the choice of early test stages, not on new physical entities. The key unpaid-for inputs are the reliability of the public datasets' labels and the representativeness of the 28-cell training set.

free parameters (4)

T2 stage (early prediction time) = 1,000 ATCs (6500 s)
Determines the reported 30x acceleration ratio; chosen because Dataset 1 provides PCD at the 0th and 1000th ATC. Not optimized; the paper speculates higher ratios are possible.
Preset EIS frequencies f1 and f2 = 7.9433 Hz and 7943.3 Hz for Dataset 1
Selected by k-means clustering on the training set Re/f curves; these two frequencies are the only measurement inputs of the PCDP method.
Random forest hyperparameters = Grid-searched over 1080 combinations; final values not reported
All prediction models use random forests with hyperparameters tuned on the training set. The exact best combination is not disclosed.
SISSO hyperparameters = k=6, operators=[+, -], n_expansion=2
Feature selection for TP-SISSO features in the LP-ALT method; these choices affect which two-point differences are used.

assumptions (4)

domain assumption The aging indicator target values in Dataset 1 (R_O2_total, I_lim, ECSA, I_cross) are correct and reliable ground truth.
The paper uses these values as training targets without independent verification or uncertainty estimates.
domain assumption The open-source datasets used (PEMFC, PEMWE, capacitor) are free of major measurement or labeling errors.
All results inherit the quality of the public datasets; no data quality checks are reported.
domain assumption A mapping from early difference curves (cycle 0 to 1000) to the cycle-30,000 aging state can be learned from 28 training fuel cells and will generalize to unseen cells.
The 11-cell test set is small, so the generalization claim rests on the representativeness of the training set.
ad hoc to paper Acceleration ratio is appropriately measured by the number of ATC cycles, excluding the time for offline PCD measurements at T1 and T2.
The 30x claim compares 30,000 cycles to 1,000 cycles; the time spent acquiring the early performance data is not included in the denominator.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Machine learning accelerates fuel cell life testing." pith.science (2026). https://pith.science/paper/B5537FEM

@misc{pith2026250418835,
  author       = {Pith},
  title        = {Pith review of: Machine learning accelerates fuel cell life testing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B5537FEM}},
  note         = {Machine review of arXiv:2504.18835}
}

read the original abstract

Accelerated life testing (ALT) can significantly reduce the economic, time, and labor costs of life testing in the process of equipment, device, and material research and development (R&D), and improve R&D efficiency. This paper proposes a performance characterization data prediction (PCDP) method and a life prediction-driven ALT (LP-ALT) method to accelerate the life test of polymer electrolyte membrane fuel cells (PEMFCs). The PCDP method can accurately predict different PCD using only four impedances (real and imaginary) corresponding to a high frequency and a medium frequency, greatly shortening the measurement time of offline PCD and reducing the difficulty of life testing. The test results on an open source life test dataset containing 42 PEMFCs show that compared with the determination coefficient (R^2) results of predicted aging indicators, including limiting current, total mass transport resistance, electrochemically active surface area, and crossover current, obtained based on the measured PCD, the R^2 results of predicted aging indicators based on the predicted PCD is only reduced by 0.04, 0.01, 0.05, and 0.06, respectively. The LP-ALT method can shorten the life test time through early life prediction. Test results on the same open-source life test dataset of PEMFCs show that the acceleration ratio of the LP-ALT method can reach 30 times under the premise of ensuring that the minimum R^2 of the prediction results of different aging indicators, including limiting current, total mass transport resistance, and electrochemically active surface area, is not less than 0.89. Combining the different performance characterization data predicted by the PCDP method and the life prediction of the LP-ALT method, the diagnosis and prognosis of PEMFCs and their components can be achieved.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 3 canonical work pages

[1]

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
[2]

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

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

2001
[3]

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

work page
[4]

Ghiringhelli

Runhai Ouyang, Stefano Curtarolo, Emre Ahmetcik, Matthias Scheffler & Luca M. Ghiringhelli. SISSO: A compressed- sensing method for identifying the best low-dimensional descriptor in an immensity of offered candidates, Physical Review Materials, 2, 083802 (2018)

work page 2018
[5]

https://github.com/PaulsonLab/TorchSISSO

work page

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

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

[2] [2]

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

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

2001

[3] [3]

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

work page

[4] [4]

Ghiringhelli

Runhai Ouyang, Stefano Curtarolo, Emre Ahmetcik, Matthias Scheffler & Luca M. Ghiringhelli. SISSO: A compressed- sensing method for identifying the best low-dimensional descriptor in an immensity of offered candidates, Physical Review Materials, 2, 083802 (2018)

work page 2018

[5] [5]

https://github.com/PaulsonLab/TorchSISSO

work page