pith. machine review for the scientific record. sign in

arxiv: 2604.20735 · v1 · submitted 2026-04-22 · 💻 cs.LG · cs.SY· eess.SY· physics.comp-ph

Recognition: unknown

Fast Bayesian equipment condition monitoring via simulation based inference: applications to heat exchanger health

Peter Collett , Alexander Johannes Stasik , Simone Casolo , Signe Riemer-S{\o}rensen
Authors on Pith no claims yet

Pith reviewed 2026-05-10 00:41 UTC · model grok-4.3

classification 💻 cs.LG cs.SYeess.SYphysics.comp-ph
keywords simulation-based inferencecondition monitoringheat exchangersBayesian inferenceneural posterior estimationfault diagnosisdigital twins
0
0 comments X

The pith

Simulation-based inference matches MCMC accuracy for heat exchanger degradation but infers parameters 82 times faster.

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

The paper demonstrates that a neural network trained on simulated sensor data can map thermal-fluid observations directly to the full posterior distribution over latent degradation parameters such as fouling and leakage. Traditional Markov Chain Monte Carlo sampling delivers rigorous uncertainty quantification but is too slow for real-time use. By amortizing the inference through the trained network, the method produces comparable diagnostic accuracy and uncertainty estimates across synthetic failure scenarios while delivering near-instant results. If the mapping generalizes, it removes the computational barrier that has kept probabilistic condition monitoring out of live industrial control loops and digital-twin systems.

Core claim

Training neural density estimators on a simulated dataset of thermal-fluid observations yields a likelihood-free mapping from measurements to the posterior over degradation parameters; across synthetic fouling and leakage cases the resulting posteriors match the accuracy and calibration of MCMC while reducing inference time by a factor of 82.

What carries the argument

Amortized neural posterior estimation, which learns a direct mapping from observations to posterior distributions after a single training phase on simulations.

If this is right

  • Real-time probabilistic fault diagnosis becomes feasible for process-control loops that must act within seconds.
  • Digital-twin models can update their internal degradation state estimates continuously rather than in batch.
  • Low-probability sparse failure events can still be diagnosed with quantified uncertainty without prohibitive compute cost.
  • The same trained network can be reused across many units or operating conditions once the initial simulation training is complete.

Where Pith is reading between the lines

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

  • The approach could be applied to other sensor-rich equipment such as pumps or turbines if equivalent high-fidelity simulators exist.
  • Online retraining or domain-adaptation layers might allow the network to track gradual changes in system behavior without full re-simulation.
  • Integration with control systems could enable proactive maintenance scheduling driven by the full posterior rather than point estimates.

Load-bearing premise

The simulated thermal-fluid data and failure modes must be representative enough that the trained network produces accurate posteriors when applied to real sensor readings.

What would settle it

Apply the trained network to real heat-exchanger sensor streams whose true degradation levels are independently measured or known, then check whether the predicted posteriors are calibrated and contain the true parameter values at the reported credible intervals.

Figures

Figures reproduced from arXiv: 2604.20735 by Alexander Johannes Stasik, Peter Collett, Signe Riemer-S{\o}rensen, Simone Casolo.

Figure 1
Figure 1. Figure 1: FIG. 1: Schematic of a counterflow heat exchanger with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Fouling and Leakage evolution in time with the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: The applied SBI scheme. Blue: training data (priors) and process, orange: input data, green: inferred [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Prior probability densities for the changepoint [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Scatterplot comparing posterior median [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: The marginalized distributions of the posterior medians inferred via MCMC (orange) and SBI (blue) across [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: Continuous Ranked Probability Score (CRPS) [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9: Posterior predictive checks for a Batch Process [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10: Resource comparison at the minimal [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
read the original abstract

Accurate condition monitoring of industrial equipment requires inferring latent degradation parameters from indirect sensor measurements under uncertainty. While traditional Bayesian methods like Markov Chain Monte Carlo (MCMC) provide rigorous uncertainty quantification, their heavy computational bottlenecks render them impractical for real-time process control. To overcome this limitation, we propose an AI-driven framework utilizing Simulation-Based Inference (SBI) powered by amortized neural posterior estimation to diagnose complex failure modes in heat exchangers. By training neural density estimators on a simulated dataset, our approach learns a direct, likelihood-free mapping from thermal-fluid observations to the full posterior distribution of degradation parameters. We benchmark this framework against an MCMC baseline across various synthetic fouling and leakage scenarios, including challenging low-probability, sparse-event failures. The results show that SBI achieves comparable diagnostic accuracy and reliable uncertainty quantification, while accelerating inference time by a factor of82$\times$ compared to traditional sampling. The amortized nature of the neural network enables near-instantaneous inference, establishing SBI as a highly scalable, real-time alternative for probabilistic fault diagnosis and digital twin realization in complex engineering systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The manuscript proposes an amortized simulation-based inference (SBI) framework using neural posterior estimation to perform fast Bayesian inference of latent degradation parameters (e.g., fouling factors, leakage rates) in heat exchangers from indirect thermal-fluid sensor observations. It benchmarks the approach against MCMC on synthetic datasets simulating fouling and leakage scenarios, claiming comparable diagnostic accuracy and uncertainty quantification with an 82× inference-time speedup.

Significance. If the central claims hold, the work would demonstrate a practical route to real-time probabilistic condition monitoring via amortization, addressing the computational barrier that has limited Bayesian methods in industrial digital-twin applications. The use of synthetic data with known ground-truth parameters enables direct, falsifiable evaluation of posterior recovery and calibration, which is a methodological strength. The focus on low-probability sparse-event failures further strengthens the relevance for safety-critical equipment.

major comments (3)
  1. [§4] §4 (Neural Posterior Estimation): The specific neural density estimator (e.g., SNPE-C, MAF, or MDN), its architecture (layers, hidden units, activation), and training procedure (number of simulations, prior specification, loss function, optimizer schedule) are not described in sufficient detail to allow reproduction or to diagnose the source of the reported accuracy.
  2. [§5.2] §5.2 and Table 2: The quantitative metrics underlying the claim of 'comparable diagnostic accuracy' (posterior mean error, credible-interval coverage, or calibration error) are not defined or tabulated against the MCMC baseline, making it impossible to judge whether the posteriors are statistically equivalent or merely qualitatively similar.
  3. [§5.3] §5.3 (Timing Experiments): The 82× speedup is stated without reporting MCMC implementation details (number of chains, samples, thinning, convergence diagnostics such as R-hat), hardware platform, or whether the one-time neural-network training cost is excluded from the comparison; this undermines the load-bearing performance claim.
minor comments (3)
  1. [Abstract] Abstract: the inline LaTeX '82$×$' renders incorrectly in plain text; use consistent notation throughout.
  2. [§2] The manuscript should add a short related-work subsection contrasting the proposed amortized SBI approach with existing surrogate-model or reduced-order Bayesian methods already used in heat-exchanger monitoring.
  3. [Figures] Figure captions for posterior visualizations should explicitly state the number of held-out test cases and the exact coverage probability shown.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which have helped us strengthen the reproducibility and clarity of the manuscript. We address each major comment below and have revised the paper accordingly.

read point-by-point responses

  1. Referee: [§4] §4 (Neural Posterior Estimation): The specific neural density estimator (e.g., SNPE-C, MAF, or MDN), its architecture (layers, hidden units, activation), and training procedure (number of simulations, prior specification, loss function, optimizer schedule) are not described in sufficient detail to allow reproduction or to diagnose the source of the reported accuracy.

    Authors: We agree that the original manuscript did not provide sufficient implementation details for reproducibility. In the revised version, Section 4 has been expanded to fully specify the neural density estimator (SNPE-C with MAF), its architecture, the number of simulations used for training, the prior distributions, the loss function, and the optimizer schedule. These additions enable readers to reproduce the experiments and assess the source of the reported performance. revision: yes

  2. Referee: [§5.2] §5.2 and Table 2: The quantitative metrics underlying the claim of 'comparable diagnostic accuracy' (posterior mean error, credible-interval coverage, or calibration error) are not defined or tabulated against the MCMC baseline, making it impossible to judge whether the posteriors are statistically equivalent or merely qualitatively similar.

    Authors: We acknowledge that the metrics supporting the accuracy claim were not explicitly defined or compared quantitatively. The revised Section 5.2 now defines the evaluation metrics (posterior mean error, credible-interval coverage, and calibration error), and Table 2 has been updated to report these values for both SBI and MCMC across all scenarios. This allows a direct, quantitative judgment of whether the posteriors are statistically comparable. revision: yes

  3. Referee: [§5.3] §5.3 (Timing Experiments): The 82× speedup is stated without reporting MCMC implementation details (number of chains, samples, thinning, convergence diagnostics such as R-hat), hardware platform, or whether the one-time neural-network training cost is excluded from the comparison; this undermines the load-bearing performance claim.

    Authors: We agree that the timing comparison lacked necessary context. The revised Section 5.3 now reports the MCMC implementation details (number of chains, samples, thinning, and R-hat convergence diagnostics), the hardware platform used for both methods, and explicitly states that the 82× factor measures only amortized inference time, excluding the one-time training cost. These clarifications support the performance claim with full transparency. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper trains neural density estimators on synthetic data from a thermal-fluid simulator and benchmarks posterior recovery and inference speed directly against MCMC on held-out synthetic fouling/leakage cases. Ground-truth parameters are known by construction in this controlled setting, so the reported 82× speedup and comparable accuracy follow from amortization and explicit comparison rather than any self-referential fit, self-citation load-bearing premise, or renaming of inputs as predictions. No equations or claims reduce to their own inputs by definition; the validation is external to the fitted network within the synthetic benchmark.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework depends on the fidelity of the simulation model to real physics, which is a standard but unverified domain assumption for SBI methods.

axioms (1)
  • domain assumption Simulated thermal-fluid observations and degradation scenarios are sufficiently representative of real heat exchanger behavior
    The neural network is trained exclusively on simulated data to map to real posterior distributions.

pith-pipeline@v0.9.0 · 5509 in / 1025 out tokens · 21893 ms · 2026-05-10T00:41:21.515257+00:00 · methodology

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

59 extracted references · 11 canonical work pages

  1. [1]

    2: Fouling and Leakage evolution in time with the effect of the failure parameters (λ, βf , βl) over the trajectories at a fixed failure onset timeτ

    T ube F ouling.In this work, grounded in recent analyses of physical deposition mechanisms [40], fouling is modeled as a reduction in the overall conductanceU A via a non-negative, dimensionless fouling factorR(t): U A(t) = U Aclean 1 +R(t) .(10) FIG. 2: Fouling and Leakage evolution in time with the effect of the failure parameters (λ, βf , βl) over the ...

  2. [2]

    Leakage with Fluid Loss.Internal leakage is modeled as a diversion of fluid from the primary hot- stream flow path, representing a shell-side bypass or a loss of integrity in the tube-to-header seals. We define a time-dependent leak fractionL(t)∈[0,1)that reduces the effective mass flow rate participating in the heat ex- change: ˙mhot,out(t) = ˙mhot,in(1−...

    2000

  3. [3]

    Ahmed Murtaza, A

    A. Ahmed Murtaza, A. Saher, M. Hamza Zafar, S. Ku- mayl Raza Moosavi, M. Faisal Aftab, and F. Sanfilippo, Results in Engineering24, 102935 (2024). 14

    2024

  4. [4]

    Villa and C

    L. Villa and C. Zanini Brusamarello, The Canadian Jour- nal of Chemical Engineering103, 1786 (2025)

    2025

  5. [5]

    R.Qi, J.Zhang,andK.Spencer,Algorithms16,9(2022)

    2022

  6. [6]

    A. Sahu, S. K. Palei, and A. Mishra, Expert Systems41, 10.1111/exsy.13360 (2023)

    work page doi:10.1111/exsy.13360 2023

  7. [7]

    J. Zou, T. Hirokawa, J. An, L. Huang, and J. Camm, Frontiers in Energy Research11, 10.3389/fenrg.2023.1294531 (2023)

    work page doi:10.3389/fenrg.2023.1294531 2023

  8. [8]

    Sundar, M

    S. Sundar, M. C. Rajagopal, H. Zhao, G. Kuntumalla, Y. Meng, H. C. Chang, C. Shao, P. Ferreira, N. Miljkovic, S. Sinha, and S. Salapaka, Intl. J. of Heat and Mass Transf.159, 120112 (2020)

    2020

  9. [9]

    Z. Wu, B. Zhang, H. Yu, J. Ren, M. Pan, C. He, and Q. Chen, Chem. Eng. Sci.282, 119285 (2023)

    2023

  10. [10]

    Jradi, C

    R. Jradi, C. Marvillet, and M. R. Jeday, Sci. Rep12, 20437 (2022)

    2022

  11. [11]

    Markowski and P

    M. Markowski and P. Trzcinski, Sci. Rep15, 42078 (2025)

    2025

  12. [12]

    A. A. Bash, H. Karah, and A. A. Khail, Engineering Ap- plications of Artificial Intelligence151, 110723 (2025)

    2025

  13. [13]

    G. Hou, D. Zhang, Q. Yan, S. Wang, L. Ma, and M. Jiang, International Communications in Heat and Mass Transfer164, 108809 (2025)

    2025

  14. [14]

    G. Hou, Z. An, D. Zhang, X. Du, Y. Ding, and J. Fu, In- ternational Communications in Heat and Mass Transfer 175, 111129 (2026)

    2026

  15. [15]

    H. Xu, S. Yan, X. Qin, W. An, and J. Wang, Processes 13, 10.3390/pr13010219 (2025)

    work page doi:10.3390/pr13010219 2025

  16. [16]

    Hosseini, A

    S. Hosseini, A. Khandakar, M. E. Chowdhury, M. A. Ayari, T. Rahman, M. H. Chowdhury, and B. Vaferi, En- ergy Reports8, 8767 (2022)

    2022

  17. [17]

    Berce, M

    J. Berce, M. Bucci, M. Zupančič, M. Može, and I. Golo- bič, Applied Thermal Engineering277, 126954 (2025)

    2025

  18. [18]

    Aboul Khail and A

    A. Aboul Khail and A. Karah Bash, Thermal Science and Engineering Progress71, 104565 (2026)

    2026

  19. [19]

    Ben-Mansour, S

    R. Ben-Mansour, S. El-Ferik, M. Al-Naser, B. A. Qureshi, M. A. M. Eltoum, A. Abuelyamen, F. Al-Sunni, and R. Ben Mansour, Energies16, 10.3390/en16062812 (2023)

    work page doi:10.3390/en16062812 2023

  20. [20]

    Etminan, K

    A. Etminan, K. Pope, and K. Mashayekh, AI Thermal Fluids4, 100022 (2025)

    2025

  21. [21]

    Gelman, J

    A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin, Bayesian data analysis(Chapman and Hall/CRC, 1995)

    1995

  22. [22]

    Moradi, S

    R. Moradi, S. Cofre-Martel, E. Lopez Droguett, M. Modarres, and K. M. Groth, Reliability Engineering & System Safety222, 108433 (2022)

    2022

  23. [23]

    M. E. Cholette, H. Yu, P. Borghesani, L. Ma, and G. Kent, Reliability Engineering & System Safety183, 184 (2019)

    2019

  24. [24]

    Bardeeniz, C

    S. Bardeeniz, C. Panjapornpon, P. Chomchai, and M. A. Hussain, Reliability Engineering & System Safety262, 111250 (2025)

    2025

  25. [25]

    G. L. Jones and Q. Qin, Annual Review of Statistics and Its Application9, 557 (2021)

    2021

  26. [26]

    Xiao and W

    S. Xiao and W. Nowak, Reliability Engineering & System Safety266, 111718 (2026)

    2026

  27. [27]

    Zio and G

    E. Zio and G. Peloni, Reliability Engineering & System Safety96, 403 (2011)

    2011

  28. [28]

    Nazemzadeh, J

    N. Nazemzadeh, J. Loyola-Fuentes, G. Righetti, E. Diaz- Bejarano, S. Mancin, and F. Coletti, AI Thermal Fluids 5, 100030 (2026)

    2026

  29. [29]

    Mo and W.-J

    J. Mo and W.-J. Yan, Reliability Engineering & System Safety264, 111337 (2025)

    2025

  30. [30]

    Dasgupta and E

    A. Dasgupta and E. A. Johnson, Reliability Engineering & System Safety242, 109729 (2024)

    2024

  31. [31]

    Cranmer, J

    K. Cranmer, J. Brehmer, and G. Louppe, Proceedings of the National Academy of Sciences117, 30055 (2020)

    2020

  32. [32]

    Goncalves, and Jakob H

    M. Deistler, J. Boelts, P. Steinbach, G. Moss, T. Moreau, M. Gloeckler, P. L. C. Rodrigues, J. Linhart, J. K. Lap- palainen, B. K. Miller, P. J. Gonçalves, J.-M. Lueck- mann, C. Schröder, and J. H. Macke, Simulation-based inference: A practical guide (2025), arXiv:2508.12939 [stat.ML]

    work page arXiv 2025

  33. [33]

    Boelts, M

    Álvaro Tejero-Cantero, J. Boelts, M. Deistler, J.-M. Lueckmann, C. Durkan, P. J. Gonçalves, D. Green- berg, and J. H. Macke, arXiv 10.48550/arxiv.2007.09114 (2020)

    work page doi:10.48550/arxiv.2007.09114 2007

  34. [34]

    Advances in neural information processing systems , year=

    G. Papamakarios and I. Murray, Fastϵ-free inference of simulation models with bayesian conditional density es- timation (2018), arXiv:1605.06376 [stat.ML]

    work page arXiv 2018

  35. [35]

    Papamakarios, D

    G. Papamakarios, D. Sterratt, and I. Murray, inProceed- ings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, Proceedings of Ma- chine Learning Research, Vol. 89, edited by K. Chaudhuri and M. Sugiyama (PMLR, 2019) pp. 837–848

    2019

  36. [36]

    S. He, Y. Ye, M. Wang, J. Zhang, W. Tian, S. Qiu, and G. Su, Nuclear Engineering and Design432, 113759 (2025)

    2025

  37. [37]

    CoRR , volume =

    J. Degrave, M. Hermans, J. Dambre, and F. Wyffels, A differentiable physics engine for deep learning in robotics (2018), arXiv:1611.01652 [cs.NE]

    work page arXiv 2018

  38. [38]

    Majumdar, V

    R. Majumdar, V. Jadhav, A. Deodhar, S. Karande, L. Vig, and V. Runkana, Numerical Heat Transfer, Part B: Fundamentals86, 1910 (2025)

    1910

  39. [39]

    M. Shi, L. Sun, Z. Bi, and R. He, Computers & Chemical Engineering201, 109259 (2025)

    2025

  40. [40]

    T. L. Bergman, A. S. Lavine, F. P. Incropera, and D. P. DeWitt,Fundamentals of heat and mass transfer(John Wiley & Sons, 2011)

    2011

  41. [41]

    Zhang, J.-L

    J.-X. Zhang, J.-L. Zhang, Z.-X. Zhang, T.-M. Li, and X.-S. Si, Reliability Engineering & System Safety250, 110223 (2024)

    2024

  42. [42]

    P. Wang, Q. Liu, D. Xu, G. Jiang, W. Peng, F. Liu, and J. Hou, International Journal of Heat and Mass Transfer 256, 128183 (2026)

    2026

  43. [43]

    Tankov and R

    P. Tankov and R. Cont,Financial Modelling with Jump Processes(Chapman and Hall/CRC, Boca Raton, 2003)

    2003

  44. [44]

    Ebert and C

    W. Ebert and C. B. Panchal, inFouling mitigation in in- dustrial heat exchanger equipment, edited by C. B. Pan- chal, T. R. Bott, E. F. C. Somerscales, and S. Toyama (Begell House, 1997) pp. 451–460

    1997

  45. [45]

    van Ravenzwaaij, P

    D. van Ravenzwaaij, P. Cassey, and S. D. Brown, Psy- chonomic Bulletin & Review25, 143 (2018)

    2018

  46. [46]

    D. M. Blei, A. Kucukelbir, and J. D. McAuliffe, Jour- nal of the American Statistical Association112, 859–877 (2017)

    2017

  47. [47]

    M. D. Hoffman and A. Gelman, J. Mach. Learn. Res.15, 1593–1623 (2014)

    2014

  48. [48]

    Sunnåker, A

    M. Sunnåker, A. G. Busetto, E. Numminen, J. Corander, M. Foll, and C. Dessimoz, PLoS Computational Biology 9, e1002803 (2013)

    2013

  49. [49]

    D. S. Greenberg, M. Nonnenmacher, and J. H. Macke, in International Conference on Machine Learning(2019)

    2019

  50. [50]

    Boelts, M

    J. Boelts, M. Deistler, M. Gloeckler, Álvaro Tejero- Cantero, J.-M. Lueckmann, G. Moss, P. Steinbach, T. Moreau, F. Muratore, J. Linhart, C. Durkan, J. Vet- ter, B. K. Miller, M. Herold, A. Ziaeemehr, M. Pals, 15 T. Gruner, S. Bischoff, N. Krouglova, R. Gao, J. K. Lappalainen, B. Mucsányi, F. Pei, A. Schulz, Z. Ste- fanidi, P. Rodrigues, C. Schröder, F. A...

    2025

  51. [51]

    J. E. Matheson and R. L. Winkler, Management Science 22, 1087 (1976)

    1976

  52. [52]

    C. Yang, M. Ai, Z. Tang, and Y. Xie, Engineering Appli- cations of Artificial Intelligence162, 112236 (2025)

    2025

  53. [53]

    Chang and Y.-H

    L. Chang and Y.-H. Lin, Engineering Applications of Ar- tificial Intelligence142, 109980 (2025)

    2025

  54. [54]

    D. Phan, N. Pradhan, and M. Jankowiak, arXiv preprint arXiv:1912.11554 (2019)

    work page arXiv 1912

  55. [55]

    Bingham, J

    E. Bingham, J. P. Chen, M. Jankowiak, F. Obermeyer, N. Pradhan, T. Karaletsos, R. Singh, P. A. Szerlip, P. Horsfall, and N. D. Goodman, J. Mach. Learn. Res. 20, 28:1 (2019)

    2019

  56. [56]

    Durkan, A

    C. Durkan, A. Ivanov, A. Maddison, and G. Papamakar- ios, inAdvances in Neural Information Processing Sys- tems, Vol. 32 (2019) pp. 7511–7522

    2019

  57. [57]

    Delaunoy, B

    A. Delaunoy, B. K. Miller, P. Forré, C. Weniger, and G. Louppe, Preprint (2023), arXiv:2304.10978 [stat.ML]

    work page arXiv 2023

  58. [58]

    Falkiewicz, N

    M. Falkiewicz, N. Takeishi, I. Shekhzadeh, A. Wehenkel, A. Delaunoy, G. Louppe, and A. Kalousis, Preprint (2023), arXiv:2310.13402 [stat.ML]

    work page arXiv 2023

  59. [59]

    X. Xu, J. Zhou, X. Weng, Z. Zhang, H. He, F. Steyskal, and G. Brunauer, Reliability Engineering & System Safety250, 110250 (2024)

    2024