Pith. sign in

REVIEW 4 major objections 4 minor 49 references

The paper claims a physics-embedded residual network predicts hydrogen crossover with near-perfect accuracy and extrapolates to 200 bar, but the body reports a different model and different extrapolation numbers.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 23:24 UTC pith:CG2SOEE3

load-bearing objection The arXiv abstract promises a hard-constraint PR-Net that the body never defines; the actual contribution is a soft-constraint PINN with a circular extrapolation trick. the 4 major comments →

arxiv 2511.05879 v5 pith:CG2SOEE3 submitted 2025-11-08 cs.LG cs.AI

Hard-constraint physics-residual networks for hydrogen crossover prediction and high-pressure extrapolation in PEM water electrolysis

Yong-Woon Kim , Jihyeok Lee , Chulung Kang , Yung-Cheol Byun This is my paper
classification cs.LG cs.AI
keywords hydrogen crossoverPEM water electrolysisphysics-informed neural networkhard-constraint residual networkpressure extrapolationreal-time monitoringHenry's lawFick's law
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper aims to show that hydrogen crossover in PEM water electrolyzers—a safety and efficiency constraint near the 4 mol% explosive limit—can be predicted accurately and, more importantly, extrapolated to high pressures with a network that treats Henry's, Fick's, and Faraday's laws as a fixed backbone and learns only residual corrections. The headline claim is a hard-constraint 'PR-Net' with R² = 99.57 ± 0.16% across six membranes and R² = 94.02 ± 0.92% at 200 bar, 2.5× beyond the training range, with 9-fold lower variability than NN/PINN baselines. The body does not describe that hard-constraint model; it reports a soft-constraint PINN (R² = 99.84% ± 0.15%) whose extrapolation to 200 bar comes from a 50:50 inference-time average with a physics model (R² = 86.4%). A sympathetic reading is that the authors are trying to establish that embedding transport laws in a network preserves interpolation accuracy while enabling pressure extrapolation that pure data-driven models cannot deliver.

Core claim

The central claim, stated on the paper's own terms, is that a compact network embedding Henry's solubility, Fick's diffusion, and Faraday's production laws as deterministic constraints—learning only a residual correction—predicts hydrogen crossover with near-perfect accuracy and degrades gracefully when pressure is pushed to 200 bar, 2.5× the training maximum. The authors report PR-Net at R² = 99.57 ± 0.16% and 200-bar extrapolation R² = 94.02 ± 0.92%, p < 0.001 vs baselines. The body reports different figures: a soft-constraint PINN at R² = 99.84 ± 0.15% and a PINN + physics fusion extrapolating at R² = 86.4% (pure NN: 43.4%). Both versions rest on the same mechanism: the physics term const

What carries the argument

The load-bearing object is the physics-informed loss (or its hard-constraint variant): L_total = (1−β)L_data + βL_physics, where L_physics penalizes deviations from the physics-derived crossover concentration Φ_phys^H2 computed from Henry's law, Fick's law with effective diffusivity, and Faraday's law for oxygen production. In the body, extrapolation is carried by the inference-time fusion y_hat = 0.5·y_PINN + 0.5·y_physics (Eq. 8), with the physics model's empirical convection parameters α, β and Arrhenius diffusivity fitted to the training data (Supplementary A.6.4). The residual network's job is to correct interpolation error; the physics term's job is to keep predictions physical outside

Load-bearing premise

The claim that the model extrapolates to 200 bar rests on the fitted physics model (used both in the fusion and as the backbone) remaining valid at 120–200 bar even though its empirical convection and diffusivity parameters are optimized on data up to 80 bar; if the physics is wrong out there, the extrapolation is just a weighted average of two data-fitted curves.

What would settle it

Re-run the same training and fusion protocol with a held-out set of independent 120–200 bar measurements from a membrane other than Nafion 117 and at a temperature other than 25°C, comparing the fused predictions against the physics model alone. If the fusion's R² at 200 bar falls below ~86% (or is not significantly better than the physics model alone), the extrapolation claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim is correct, sub-millisecond inference (0.18 ms on desktop, ~4.5 ms on Raspberry Pi 4) would allow real-time crossover monitoring and adaptive control, keeping H₂ in O₂ below the 4 mol% explosion threshold.
  • The 50:50 physics fusion would let operators trust predictions at 120–200 bar without collecting new high-pressure experimental data.
  • The approach would provide a template for other electrochemical systems where data are scarce, such as battery degradation monitoring and fuel-cell optimization.
  • The claimed 15–25% membrane lifetime extension and $200k–1.5M annual savings per facility would follow from earlier detection of crossover-driven degradation.
  • The model would confirm a transport-regime transition near 0.23 A cm⁻² between diffusion-dominated and Faradaic-production-dominated crossover.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • My reading: the 200-bar extrapolation is only as trustworthy as the fitted physics model; since α, β, and diffusivity parameters are fit to 1–80 bar data, the fusion at 120–200 bar is essentially physics extrapolation dressed as a network result.
  • Given the abstract/body mismatch, the R² = 94.02 and R² = 86.4 numbers should not be conflated; any replication attempt should specify which model and which evaluation protocol is being tested.
  • A direct test would train on 1–80 bar and evaluate on independent published 120–200 bar data from a second membrane at a second temperature; the current evidence is 24 points from Nafion 117 at 25°C.
  • The claimed residual correction for high-pressure gas-phase non-ideality could be checked by comparing the learned residual to an equation-of-state correction; if they disagree, calling it a capture of non-ideality is coincidental.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper (as submitted under the arXiv title) claims to develop a hard-constraint physics-residual network (PR-Net) for hydrogen crossover prediction in PEM electrolyzers, reporting R²=99.57±0.16% on the augmented dataset and R²=94.02±0.92% at 200 bar pressure extrapolation, with 9-fold lower variability than NN and PINN. The body of the manuscript, however, describes a soft-constraint physics-informed neural network (Eqs. 1–2, β=0.3), an 8→128→128→1 feedforward architecture (Section 4.2), and inference-time 50:50 averaging with a physics model (Eq. 8). The body reports cross-validated R²=99.84±0.15% on 1,114 augmented points and R²=86.4% at 200 bar for the fusion approach. The central claim of the abstract is thus not the model or the numbers that appear in the rest of the manuscript.

Significance. If a hard-constraint PR-Net with R²=99.57% and 94% extrapolation at 200 bar were actually demonstrated, it would be a practically important monitoring tool for high-pressure PEM electrolysis. The paper also compiles a useful multi-study dataset (184 experimental points, eight sources, six membranes) and includes extensive cross-validation, ensemble UQ, hardware timing, and statistical significance testing. However, as written, the headline model and results are absent from the body; the physics constraints are applied as soft penalties and the physics model itself is calibrated to the same experimental data. The manuscript also evaluates performance largely on spline-interpolated points. These issues prevent the reported claims from being verified, let alone adopted.

major comments (4)
  1. [Abstract vs. Section 4.2, 4.3, 4.8, Table 6] The abstract claims a hard-constraint PR-Net with R²=99.57±0.16% and 200-bar extrapolation R²=94.02±0.92%. No such model appears in the Methods: Eq. (1) defines a soft-constraint loss with β=0.3, Eq. (2) is a penalty term, and Section 4.2 describes a plain feedforward network. Table 6 reports extrapolation R²=86.4% at 200 bar for PINN+physics fusion, and standalone PINN R²=51.0%. The 99.57% and 94.02% numbers are not reported anywhere in the body. The abstract's p<0.001 comparison is against the NN, not against any PR-Net. This is an internal inconsistency in the central claim, not a presentation issue.
  2. [Sections 2.5, 4.8, A.6.4, Eq. (8)] The claimed extrapolation advantage is circular in its current form. The 'physics' component in Eq. (8) is not an independent benchmark: Eq. (6) uses convection parameters α and β that are 'membrane-specific values optimized from experimental data' (A.6.4), and Eq. (9) is an Arrhenius fit. These parameters are calibrated to the same 1–80 bar training data, then Eq. (8) half-weights the PINN prediction with this data-fitted physics model at 120–200 bar. The paper does not report the physics-model-only extrapolation R² at 120/160/200 bar, nor how the fusion result changes with α and β. A concrete test would be to refit α and β on a pressure-restricted subset and show that the fusion result is stable, or to report a physics model that was not fitted to the test pressure range.
  3. [Section 4.1, Table 1, A.6.1] The accuracy claims are based on cross-validation over 1,114 points, of which 930 are cubic-spline interpolations between 184 experimental measurements. Table 1 shows that the same PINN drops from R²=99.84%±0.15% on the augmented set to R²=98.91%±0.92% on the original 184 points. Spline-interpolated points are not independent measurements; unless augmentation is performed inside each training fold without ever using test-fold experimental points to build splines, the CV error is optimistically biased. The manuscript does not describe a leakage-free augmentation protocol. The interpolation accuracy claim should be evaluated on held-out original experimental points only.
  4. [A.4.1, Table 6, Fig. 10] The extrapolation claim rests on n=24 test points from a single membrane (Nafion 117) at a single temperature (25°C). Table 6 reports R² values at 120/160/200 bar from these 24 points, and Fig. 10e shows overlapping bootstrap confidence intervals for PINN vs. NN MAE in the no-fusion comparison. This sample size and coverage cannot support a general claim that the method 'extrapolates to 200 bar' across membrane types and temperatures. At minimum, the extrapolation evaluation should include multiple membranes, multiple temperatures, and more test points with confidence intervals reported for each pressure.
minor comments (4)
  1. [Title and headings] The arXiv title names a 'hard-constraint physics-residual network (PR-Net)', but the manuscript title and body describe a 'physics-informed neural network (PINN)'. This naming inconsistency should be resolved.
  2. [A.7.2] The reporting convention states that the best single model uses β=0.1 while the cross-validation mean uses β=0.3, yet Table 1 and Section 2.1 report β=0.3 as the primary. Clarify which configuration underlies the abstract and headline claims.
  3. [Section 4.9 and references] Code and data are promised only 'upon publication' at a placeholder repository, preventing verification. Several references have incomplete placeholders, e.g., '[? ?]' after the ISO 26142 statement. These should be fixed before any resubmission.
  4. [Figure captions] Supplementary Figures 6 and 7 contain typos such as 'Datraset Analysis'.

Circularity Check

2 steps flagged

Headline accuracy and extrapolation claims rest on self-generated augmented data and a 50:50 fusion with a data-fitted physics model; abstract PR-Net numbers are absent from the body.

specific steps
  1. fitted input called prediction [Section 4.8, Eq. (8); Supplementary A.6.4]
    "ˆyfusion =α× ˆyPINN + (1−α)× ˆyphysics ... The physics model parameters were calibrated using the training data to ensure consistency. ... Note that empirical parameters α and β are membrane-specific values optimized from experimental data."

    The extrapolation headline (Table 6: PINN+Physics R²=86.4% at 200 bar) is a 50:50 average with a physics model whose empirical convection parameters α and β are optimized on experimental data (A.6.4). Because Eq. (2) trains the PINN to minimize deviation from this same fitted Φ_phys^H2, at out-of-range pressures the fusion output is essentially the fitted physics curve, not an independent first-principles extrapolation. The paper presents this as physics-based extrapolation, while the physics half is a data-fitted curve; the 2.5×-beyond-training prediction therefore reduces, in large part, to extrapolation of a fitted empirical model.

  2. self definitional [Section 4.1 (data augmentation); Section 2.1 (cross-validation metrics)]
    "Data augmentation via cubic spline interpolation expanded the dataset to 1,114 points ... (3) all interpolated points were validated against physics model predictions with a tolerance of ±5% ... Cross-validation on 1,114 augmented data points yielded R 2 = 99.84%±0.15%."

    The main interpolation accuracy is benchmarked on a test set largely composed of cubic-spline interpolated points generated under monotonicity constraints and validated against the same physics model later used as the soft constraint in Eq. (2). The interpolation generator and the physics regularizer are therefore the same input, so the reported R² on 1,114 augmented points is partly a self-consistency check rather than independent experimental validation. The 'exceptional accuracy (R² = 99.84%)' claim is thus circular in its validation.

full rationale

The two circular steps are in the validation/benchmark chain, not in the neural-network training itself. First, the extrapolation claim (Section 2.5, Table 6) is produced by Eq. (8), a 50:50 average with a physics model whose empirical convection parameters α and β are 'optimized from experimental data' (A.6.4). Since Eq. (2) trains the PINN to match that same physics model, at out-of-range pressures the PINN inherits the fitted curve and the fusion is dominated by it; the 'physics-based extrapolation' is therefore largely extrapolation of a fitted empirical model. Second, the main interpolation accuracy (R² = 99.84%) is reported on 1,114 points that include cubic-spline interpolations generated under monotonicity constraints and 'validated against physics model predictions' (Section 4.1), the same model used as the soft regularizer; the test set is not independent of the model family. Separately (not a circularity), the abstract claims a hard-constraint PR-Net with R² = 99.57 and extrapolation R² = 94.02 that does not appear anywhere in the body: the body defines a soft-constraint PINN (Eqs. 1–2) and reports fusion R² = 86.4 at 200 bar (Table 6). That mismatch is a reproducibility/integrity problem, but there is no derivation to reduce. No load-bearing self-citations were found; references [18]–[26] are external. Overall, because the headline interpolation benchmark and the extrapolation benchmark each reduce partly to self-generated/fitted inputs, the circularity score is 6.

Axiom & Free-Parameter Ledger

6 free parameters · 8 axioms · 0 invented entities

The central claims rest on a physics model (Eqs. 3-7) whose parameters are fitted to the same experimental data, on spline-augmented training points, and on an inference-time fusion weight chosen by hand. The body does not implement the hard-constraint PR-Net described in the abstract. No new physical entities are introduced.

free parameters (6)
  • Physics weight β = 0.3
    Selected by cross-validation over [0, 0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0] (Tables 3-4). Controls how strongly the model is pulled toward the fitted physics model.
  • Fusion weight α (Eq. 8) = 0.5
    Chosen by hand as 'equal contribution'; no sensitivity analysis is provided for the extrapolation result.
  • Convection parameter α in Eq. (6) = 5.52e-3 for Nafion 117
    Empirical, membrane-specific, stated as 'optimized from experimental data' in Supplementary A.6.4. Entered into the physics model that drives L_physics and fusion.
  • Convection exponent β in Eq. (6) = computed via log-linear model (a=0.440, b=0.0439)
    Fitted to experimental data; used in the physics model and therefore in the loss and in inference-time fusion.
  • Arrhenius diffusivity parameters (Eq. 9) = 7.734e-6 m2/s and 2225.4 K
    Stated as 'parameters fitted for the operating temperature range' (A.6.4).
  • Spline augmentation constraints = max 10 interpolated points between experimental measurements; monotonicity and 0-20% bounds
    Choice of augmentation density and constraints affects every cross-validation metric on the augmented dataset.
axioms (8)
  • domain assumption Henry's law with a temperature-dependent solubility coefficient describes dissolved H2 at the membrane interface at all pressures up to 200 bar
    Used in Eq. (3). Assumes no interfacial resistance or saturation effects at high pressure.
  • domain assumption Fick's law with effective diffusivity D_eff = (epsilon_w/tau) * D_H2,w captures membrane transport
    Used in Eq. (5). Assumes homogeneous, non-degraded membrane and steady-state behavior.
  • domain assumption Faraday's law gives the O2 production rate as N_O2 = i/(4F)
    Used in Eq. (7). Assumes 100% Faradaic efficiency for oxygen evolution.
  • ad hoc to paper The Omrani et al. empirical convection model with fitted α, β is valid across all operating conditions and extrapolates to 200 bar
    This model enters L_physics (Eq. 2) and the inference-time fusion (Eq. 8); its parameters are fitted to the training data (A.6.4).
  • ad hoc to paper Cubic spline interpolated points lie on the true response surface and can be used as training/validation data
    Data augmentation in Section 4.1; cross-validation on augmented data may overstate generalization to real experimental conditions.
  • ad hoc to paper A 50:50 weighted average of PINN and physics model predictions improves extrapolation
    Eq. (8), Section 4.8. No theoretical justification or sensitivity analysis is given for the fusion weight.
  • domain assumption Data extracted from published figures via WebPlotDigitizer are accurate, and missing parameters can be imputed (e.g., anode pressure assumed atmospheric)
    Section 4.1. Measurement uncertainty from figure extraction and imputation is not quantified.
  • standard math A two-layer tanh feedforward network with 128 hidden units can represent the crossover function with the available data
    Standard neural network approximation assumption; no error bounds are provided.

pith-pipeline@v1.3.0-alltime-deepseek · 5009 in / 5363 out tokens · 142594 ms · 2026-08-03T23:24:55.049283+00:00 · methodology

0 comments
read the original abstract

Hydrogen crossover is a critical safety and efficiency constraint in high-pressure polymer electrolyte membrane water electrolysis (PEMWE), but accurate prediction remains difficult because data are limited, transport physics are strongly coupled, and industrial operation requires reliable extrapolation beyond observed conditions. This study develops a hard-constraint physics-residual network (PR-Net) for hydrogen crossover prediction in PEMWE and compares it with a purely data-driven neural network (NN) and a soft-constraint physics-informed neural network (PINN). PR-Net embeds Henry's, Fick's, and Faraday's laws as a deterministic backbone and learns only a residual correction for unmodelled nonlinear effects. The benchmark includes 184 observations from eight peer-reviewed sources across six membrane types, covering 1-200 bar, $25-85{\deg}C$, and $0.05-5.0 A cm^{-2}$. PR-Net achieves $R^2 = 99.57 \pm 0.16%$, with 9-fold lower prediction variability than NN and PINN. In pressure-axis extrapolation, PR-Net attains $R^2 = 94.02 \pm 0.92%$ at 200 bar, 2.5 times beyond the training pressure range, compared with $68.06 \pm 5.52%$ for PINN and $58.00 \pm 8.60%$ for NN (p < 0.001). Residual analysis indicates that the learned correction captures part of the high-pressure gas-phase non-ideality and recovers a transport-regime transition near $0.23 A cm^{-2}$ between Fickian diffusion-dominated and Faradaic production-dominated transport. With a computation time of $1.08 \pm 0.34 ms$ on low-power embedded hardware, PR-Net provides a practical framework for real-time crossover monitoring, adaptive process control, and safer high-pressure green-hydrogen operation.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 8 canonical work pages

  1. [1]

    International Journal of Hydrogen Energy38, 4901–4934 (2013)

    Carmo, M., Fritz, D.L., Mergel, J., Stolten, D.: A comprehensive review on pem water electrolysis. International Journal of Hydrogen Energy38, 4901–4934 (2013)

    2013

  2. [2]

    International Journal of Hydrogen Energy45, 26036–26058 (2020)

    Grigoriev, S.A., Fateev, V.N., Bessarabov, D.G., Millet, P.: Current status, research trends, and challenges in water electrolysis science and technology. International Journal of Hydrogen Energy45, 26036–26058 (2020)

    2020

  3. [3]

    Materials Science for Energy Technologies2, 442–454 (2019)

    Shiva Kumar, S., Himabindu, V.: Hydrogen production by pem water electrolysis – a review. Materials Science for Energy Technologies2, 442–454 (2019)

    2019

  4. [4]

    International Journal of Hydro- gen Energy42(21), 14355–14366 (2017) https://doi.org/10.1016/j.ijhydene.2017

    Trinke, P., Bensmann, B., Hanke-Rauschenbach, R.: Current density effect on hydrogen permeation in pem water electrolyzers. International Journal of Hydro- gen Energy42(21), 14355–14366 (2017) https://doi.org/10.1016/j.ijhydene.2017. 03.231

    work page doi:10.1016/j.ijhydene.2017 2017

  5. [5]

    International Journal of Hydrogen Energy38(35), 14921–14933 (2013) https://doi.org/10.1016/j.ijhydene.2013.09

    Schalenbach, M., Carmo, M., Fritz, D.L., Mergel, J., Stolten, D.: Pressurized pem water electrolysis: Efficiency and gas crossover. International Journal of Hydrogen Energy38(35), 14921–14933 (2013) https://doi.org/10.1016/j.ijhydene.2013.09. 049

    work page doi:10.1016/j.ijhydene.2013.09 2013

  6. [6]

    Electrochimica Acta377, 138085 (2021)

    Omrani, R., Shabani, B.: Hydrogen crossover in proton exchange membrane elec- trolysers: The effect of current density, pressure, temperature, and compression. Electrochimica Acta377, 138085 (2021)

    2021

  7. [7]

    Journal of Power Sources366, 33–55 (2017)

    Feng, Q., Yuan, X.Z., Liu, G., Wei, B., Zhang, Z.: A review of proton exchange membrane water electrolysis on degradation mechanisms and mitigation strate- gies. Journal of Power Sources366, 33–55 (2017)

    2017

  8. [8]

    International Journal of Hydrogen Energy42, 30470–30492 (2017)

    Schmidt, O., Gambhir, A., Staffell, I., Hawkes, A., Nelson, J.: Future cost and per- formance of water electrolysis: An expert elicitation study. International Journal of Hydrogen Energy42, 30470–30492 (2017)

    2017

  9. [9]

    International Journal of Hydrogen Energy 42, 26203–26216 (2017)

    Aubras, F., Deseure, J., Kadjo, J.J., Dedigama, I., Majasan, J.: Two-dimensional model of low-pressure pem electrolyser: Two-phase flow regime, electrochemical 15 modelling and experimental validation. International Journal of Hydrogen Energy 42, 26203–26216 (2017)

    2017

  10. [10]

    International Journal of Hydrogen Energy40, 7006–7016 (2015)

    Han, B., Steen, S.M., Mo, J., Zhang, F.Y.: Electrochemical performance mod- eling of a proton exchange membrane electrolyzer cell for hydrogen production. International Journal of Hydrogen Energy40, 7006–7016 (2015)

    2015

  11. [11]

    Energy152, 237–246 (2018)

    Toghyani, S., Afshari, E., Baniasadi, E., Atyabi, S.A.: Thermal and electrochemi- cal performance assessment of a high temperature pem electrolyzer. Energy152, 237–246 (2018)

    2018

  12. [12]

    Energy Conversion and Management49, 2748–2756 (2008)

    Ni, M., Leung, M.K.H., Leung, D.Y.C.: Energy and exergy analysis of hydrogen production by a proton exchange membrane (pem) electrolyzer plant. Energy Conversion and Management49, 2748–2756 (2008)

    2008

  13. [13]

    International Journal of Hydrogen Energy 37, 1927–1938 (2012)

    Garc ´ ıa-Valverde, R., Espinosa, N., Urbina, A.: Simple pem water electrolyser model and experimental validation. International Journal of Hydrogen Energy 37, 1927–1938 (2012)

    1927

  14. [14]

    International Journal of Hydrogen Energy36, 14779–14786 (2011)

    Awasthi, A., Scott, K., Basu, S.: Dynamic modeling and simulation of a proton exchange membrane electrolyzer for hydrogen production. International Journal of Hydrogen Energy36, 14779–14786 (2011)

    2011

  15. [15]

    Energies15, 6657 (2022)

    Mohamed, A., Ibrahem, H., Yang, R., Kim, K.: Optimization of proton exchange membrane electrolyzer cell design using machine learning. Energies15, 6657 (2022)

    2022

  16. [16]

    Journal of Power Sources556, 232486 (2023)

    Ding, R.,et al.: Machine learning utilized for the development of proton exchange membrane electrolyzers. Journal of Power Sources556, 232486 (2023)

    2023

  17. [17]

    Applied Energy368, 123550 (2024)

    Chen, X.,et al.: Machine learning in proton exchange membrane water electrolysis — a knowledge-integrated framework. Applied Energy368, 123550 (2024)

    2024

  18. [18]

    Nature Reviews Physics3, 422–440 (2021)

    Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S.: Physics- informed machine learning. Nature Reviews Physics3, 422–440 (2021)

    2021

  19. [19]

    Journal of Computational Physics378, 686–707 (2019)

    Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics378, 686–707 (2019)

    2019

  20. [20]

    Nature Reviews Physics3, 422–440 (2021)

    Karniadakis, G.E.,et al.: Physics-informed machine learning. Nature Reviews Physics3, 422–440 (2021)

    2021

  21. [21]

    Computer Methods in Applied Mechanics and Engineering361, 112732 (2020) 16

    Sun, L., Gao, H., Pan, S., Wang, J.X.: Surrogate modeling for fluid flows based on physics-constrained deep learning without simulation data. Computer Methods in Applied Mechanics and Engineering361, 112732 (2020) 16

    2020

  22. [22]

    Journal of Computational Physics426, 109951 (2021)

    Jin, X., Cai, S., Li, H., Karniadakis, G.E.: Nsfnets (navier-stokes flow nets): Physics-informed neural networks for the incompressible navier-stokes equations. Journal of Computational Physics426, 109951 (2021)

    2021

  23. [23]

    Journal of Heat Transfer143, 060801 (2021)

    Cai, S., Wang, Z., Wang, S., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks for heat transfer problems. Journal of Heat Transfer143, 060801 (2021)

    2021

  24. [24]

    Energy AI20, 100474 (2025)

    Zerrougui, I., Li, Z., Hissel, D.: Physics-informed neural network for modeling and predicting temperature fluctuations in proton exchange membrane electrolysis. Energy AI20, 100474 (2025)

    2025

  25. [25]

    Applied Energy382, 125048 (2025)

    Ko, T., Kim, D., Park, J., Lee, S.H.: Physics-informed neural network for long- term prognostics of proton exchange membrane fuel cells. Applied Energy382, 125048 (2025)

    2025

  26. [26]

    arXiv preprint arXiv:2507.02887 (2025)

    Polo-Molina, A., et al.: Modeling membrane degradation in pem electrolyzers with physics-informed neural networks. arXiv preprint arXiv:2507.02887 (2025)

    Pith/arXiv arXiv 2025

  27. [27]

    International Journal of Hydrogen Energy36(3), 2721–2728 (2011) https://doi

    Grigoriev, S.A., Porembskiy, V.I., Korobtsev, S.V., Fateev, V.N., Auprˆ etre, F., Mill´ et, P.: High-pressure pem water electrolysis and corresponding safety issues. International Journal of Hydrogen Energy36(3), 2721–2728 (2011) https://doi. org/10.1016/j.ijhydene.2010.12.084

    work page doi:10.1016/j.ijhydene.2010.12.084 2011

  28. [28]

    International Journal of Hydrogen Energy45(7), 4008–4014 (2020) https://doi.org/10.1016/j.ijhydene.2019.12.029

    St¨ ahler, M., St¨ ahler, A., Scheepers, F., Carmo, M., Lehnert, W., Stolten, D.: Impact of porous transport layer compression on hydrogen permeation in pem water electrolysis. International Journal of Hydrogen Energy45(7), 4008–4014 (2020) https://doi.org/10.1016/j.ijhydene.2019.12.029

    work page doi:10.1016/j.ijhydene.2019.12.029 2020

  29. [29]

    Journal of The Electrochemical Society168(9), 094509 (2021) https://doi.org/10.1149/1945-7111/ac275b

    Martin, A., Abbas, D., Trinke, P., B¨ ohm, T., Bierling, M., Bensmann, B., Thiele, S., Hanke-Rauschenbach, R.: Communication—proving the importance of pt- interlayer position in pemwe membranes for the effective reduction of the anodic hydrogen content. Journal of The Electrochemical Society168(9), 094509 (2021) https://doi.org/10.1149/1945-7111/ac275b

    work page doi:10.1149/1945-7111/ac275b 2021

  30. [30]

    Journal of The Electrochemical Society166(13), 873 (2019) https: //doi.org/10.1149/2.0111913jes

    Garbe, S., Babic, U., Nilsson, E., Schmidt, T.J., Gubler, L.: Communication—pt- doped thin membranes for gas crossover suppression in polymer electrolyte water electrolysis. Journal of The Electrochemical Society166(13), 873 (2019) https: //doi.org/10.1149/2.0111913jes

    work page doi:10.1149/2.0111913jes 2019

  31. [31]

    Journal of The Electrochemical Society167(12), 124502 (2020) https: //doi.org/10.1149/1945-7111/abaa68

    Bernt, M., Schr¨ oter, J., M¨ ockl, M., Gasteiger, H.A.: Analysis of gas permeation phenomena in a pem water electrolyzer operated at high pressure and high current density. Journal of The Electrochemical Society167(12), 124502 (2020) https: //doi.org/10.1149/1945-7111/abaa68

    work page doi:10.1149/1945-7111/abaa68 2020

  32. [32]

    International Journal of Hydrogen Energy135, 499–506 (2025) https://doi.org/10.1016/j.ijhydene.2025.04.404

    Wu, J., Zhang, G., Dong, B., Chang, Y., Chen, Y.: Experimental and simulation 17 study of h2 crossover in pem water electrolysis for high-pressure hydrogen pro- duction up to 20 mpa. International Journal of Hydrogen Energy135, 499–506 (2025) https://doi.org/10.1016/j.ijhydene.2025.04.404

    work page doi:10.1016/j.ijhydene.2025.04.404 2025

  33. [33]

    Journal of The Electrochemical Society133, 88–92 (1986)

    Sakai, T., Takenaka, H., Torikai, E.: Gas diffusion in the dried and hydrated nafions. Journal of The Electrochemical Society133, 88–92 (1986)

    1986

  34. [34]

    International Journal of Hydrogen Energy36, 2462–2470 (2011)

    Buttner, W.J., Post, M.B., Burgess, R., Rivkin, C.: An overview of hydrogen safety sensors and requirements. International Journal of Hydrogen Energy36, 2462–2470 (2011)

    2011

  35. [35]

    IEA Publications, ??? (2019)

    International Energy Agency: The Future of Hydrogen: Seizing Today’s Oppor- tunities. IEA Publications, ??? (2019)

    2019

  36. [36]

    Brussels: European Commission (2020)

    European Commission: A hydrogen strategy for a climate-neutral Europe. Brussels: European Commission (2020)

    2020

  37. [37]

    Journal of Computing and Information Science in Engineering24, 040802 (2024)

    Faroughi, S.A.,et al.: Physics-guided, physics-informed, and physics-encoded neu- ral networks and operators in scientific computing: Fluid and solid mechanics. Journal of Computing and Information Science in Engineering24, 040802 (2024)

    2024

  38. [38]

    Computers and Fluids280, 106370 (2024)

    Sun, R., Jeong, H., Zhao, J., Gou, Y., Sauret, E., Li, Z., Gu, Y.: A physics- informed neural network framework for multi-physics coupling microfluidic problems. Computers and Fluids280, 106370 (2024)

    2024

  39. [39]

    Es’kin, V.A., Malkhanov, A.O., Smorkalov, M.E.: Are two hidden lay- ers still enough for the physics-informed neural networks? arXiv preprint arXiv:2404.10695 (2024)

    Pith/arXiv arXiv 2024

  40. [40]

    In: Proceedings of the 13th International Conference on Artificial Intelligence and Statistics, pp

    Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the 13th International Conference on Artificial Intelligence and Statistics, pp. 249–256 (2010)

    2010

  41. [41]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  42. [42]

    International Journal of Hydrogen Energy48, 22486–22501 (2023)

    Tao, L.,et al.: Present and future cost of alkaline and pem electrolyser stacks. International Journal of Hydrogen Energy48, 22486–22501 (2023)

    2023

  43. [43]

    Technical Report NREL/TP-6A20-72740, National Renew- able Energy Laboratory (2019)

    Ruth, M., et al.: Manufacturing cost analysis for proton exchange membrane water electrolyzers. Technical Report NREL/TP-6A20-72740, National Renew- able Energy Laboratory (2019)

    2019

  44. [44]

    Technical Report NREL/TP-6A20-72740, National Renewable Energy Laboratory (2019)

    Mayyas, A., Ruth, M., Pivovar, B., Bender, G., Wipke, K.: Manufacturing cost analysis for proton exchange membrane water electrolyzers. Technical Report NREL/TP-6A20-72740, National Renewable Energy Laboratory (2019)

    2019

  45. [45]

    Department of Energy: Clean hydrogen production cost pem electrolyzer

    U.S. Department of Energy: Clean hydrogen production cost pem electrolyzer. 18 DOE Hydrogen Program Record 24005, U.S. Department of Energy (2024)

    2024

  46. [46]

    Department of Energy: Manufacturing cost analysis for pem electrolyzers

    U.S. Department of Energy: Manufacturing cost analysis for pem electrolyzers. H2new: Hydrogen from next-generation electrolyzers of water, U.S. Department of Energy (2025). DOE Hydrogen Shot Initiative

    2025

  47. [47]

    Technical report, Center on Global Energy Policy, Columbia University (2023)

    Odenweller, A., et al.: Demystifying electrolyzer production costs. Technical report, Center on Global Energy Policy, Columbia University (2023)

    2023

  48. [48]

    Technical report, Pacific Northwest National Laboratory (2020)

    Mongird, K., et al.: Energy storage grand challenge cost and performance assessment 2020. Technical report, Pacific Northwest National Laboratory (2020)

    2020

  49. [49]

    Technical Report DOE/EERE-2019, U.S

    Peterson, D., et al.: Hydrogen production cost from pem electrolysis - 2019. Technical Report DOE/EERE-2019, U.S. Department of Energy (2019) 19 A Supplementary Information A.1 Supplementary Figures A.1.1 Dataset Analysis Figures Fig. 4 Dataset Analysis #1: Comprehensive dataset analysis of hydrogen crossover in PEM water electrolysis under varying operat...

    2019