pith. machine review for the scientific record. sign in

arxiv: 2604.17156 · v1 · submitted 2026-04-18 · 💻 cs.LG · physics.comp-ph

Recognition: unknown

Uncertainty Quantification in PINNs for Turbulent Flows: Bayesian Inference and Repulsive Ensembles

George Em Karniadakis, Khemraj Shukla, Michael Triantafyllou, Theo Kaeufer, Zongren Zou

Pith reviewed 2026-05-10 06:16 UTC · model grok-4.3

classification 💻 cs.LG physics.comp-ph
keywords physics-informed neural networksuncertainty quantificationBayesian inferencerepulsive ensemblesturbulent flowsinverse problemsReynolds-averaged Navier-Stokes
0
0 comments X

The pith

Bayesian PINNs provide the most consistent uncertainty estimates for reconstructing turbulent flows from sparse data.

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

This paper develops and tests probabilistic extensions of physics-informed neural networks to quantify epistemic uncertainty in inverse problems for turbulent flows. It compares Bayesian PINNs that use Hamiltonian Monte Carlo sampling with a tempered multi-component likelihood, Monte Carlo dropout, and repulsive deep ensembles that enforce diversity in function space. The methods are evaluated on the Van der Pol oscillator plus cylinder flows at Reynolds numbers 3900 and 10000 using both simulation and experimental data. Bayesian PINNs deliver the most consistent uncertainty across inferred quantities, while repulsive ensembles achieve competitive accuracy on primary flow variables at lower computational cost. The work supplies practical guidance on the accuracy-cost-calibration trade-offs when using physics-informed learning for data-driven turbulence modeling.

Core claim

By incorporating Hamiltonian Monte Carlo sampling with a tempered multi-component likelihood in Bayesian PINNs and enforcing diversity in function space for repulsive ensembles, the framework shows that Bayesian PINNs deliver the most consistent uncertainty estimates across all inferred quantities while function-space repulsive ensembles provide a computationally efficient approximation with competitive accuracy for primary flow variables.

What carries the argument

The combination of Hamiltonian Monte Carlo sampling with a tempered multi-component likelihood in Bayesian PINNs together with repulsive deep ensembles that enforce diversity in function space.

If this is right

  • Bayesian PINNs should be preferred when reliable uncertainty quantification is required across all flow variables in data-driven RANS modeling.
  • Repulsive ensembles can serve as a practical substitute when computational resources limit the use of full Bayesian sampling but primary flow accuracy remains essential.
  • Likelihood tempering and function-space diversity are the mechanisms that drive improved calibration in these PDE-constrained settings.
  • The hierarchy of test cases supports extension of the same uncertainty methods to other sparse-data turbulent reconstruction tasks.

Where Pith is reading between the lines

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

  • The same probabilistic extensions could be tested on inverse problems governed by other PDEs outside fluid dynamics to check whether the consistency-efficiency trade-off persists.
  • Hybrid approaches that combine repulsive ensembles with partial Bayesian sampling might retain most of the calibration benefit at reduced cost.
  • Function-space diversity techniques may also reduce over in standard deep ensembles for non-physics machine-learning tasks.

Load-bearing premise

The tempered multi-component likelihood and enforced ensemble diversity improve uncertainty calibration for PDE-constrained inverse problems without introducing systematic biases that are not captured by the chosen test cases.

What would settle it

Empirical coverage rates of the uncertainty intervals falling significantly outside the nominal levels on a new turbulent flow dataset with different sparsity, noise, or Reynolds number would falsify the claim of consistent calibration.

Figures

Figures reproduced from arXiv: 2604.17156 by George Em Karniadakis, Khemraj Shukla, Michael Triantafyllou, Theo Kaeufer, Zongren Zou.

Figure 1
Figure 1. Figure 1: Van der Pol oscillator: comparison of uncertainty quantification obtained versus [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Van der Pol oscillator: comparison of Repulsive Deep Ensembles in parameter space [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Re = 3,900: Computational domain and data used for training and inference. The top-left panel shows the spatial distribution of sparse observation points used for Bayesian inference across the domain. The remaining panels in the top row illustrate the perturbed velocity measurements, where Gaussian noise is added to the underlying data. The first two panels in the bottom row show the Reynolds stress compon… view at source ↗
Figure 4
Figure 4. Figure 4: Re = 3,900 — Bayesian PINN velocity diagnostics. The panels show the reference solution, posterior mean prediction, absolute error, and predictive standard deviation, ratio for the velocity components (U, V ). The results demonstrate accurate reconstruction of the velocity field and good alignment between predicted uncertainty and pointwise error across the domain [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Re = 3,900 — Bayesian PINN Reynolds stress diagnostics. The panels present the reference fields, posterior mean predictions, absolute error, and predictive standard deviation for the Reynolds stress divergence components (fx, fy). The results indicate accurate recovery of the closure terms and well-calibrated uncertainty estimates throughout the computational domain. physics constraints via a cosine ramp, … view at source ↗
Figure 6
Figure 6. Figure 6: Re = 3,900 — Bayesian PINN: inferred pressure field. No pressure data are used during training, and the pressure is recovered solely through the governing equations. The posterior mean prediction closely matches the reference solution, while the predictive uncer￾tainty provides a meaningful estimate of the local error, which is particularly important in practical scenarios where reference pressure data are… view at source ↗
Figure 7
Figure 7. Figure 7: Re = 3,900 — RDE-PINN (function space): velocity diagnostics. The panels show the reference solution, ensemble mean prediction, absolute error, predictive standard deviation, and the error-to-uncertainty ratio for the velocity components (U, V ). The ensemble accurately reconstructs the velocity field, while the uncertainty estimates capture the spatial distribution of prediction error with reasonable cali… view at source ↗
Figure 8
Figure 8. Figure 8: Re = 3,900 — RDE-PINN (function space): Reynolds stress divergence diagnostics. The panels present the reference fields, ensemble mean predictions, absolute error, predictive standard deviation, and the error-to-uncertainty ratio for the closure terms (fx, fy). While the mean predictions capture the overall structure, the uncertainty estimates are less well calibrated, particularly in regions with large mo… view at source ↗
Figure 9
Figure 9. Figure 9: The corresponding uncertainty estimates are conservative, yielding [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: Re = 3,900 — RDE-PINN (function space): inferred pressure field. No pressure data are used during training, and the pressure is recovered solely through the governing equations. The ensemble mean captures the large-scale structure of the pressure field; however, the associated uncertainty is relatively conservative, reflecting increased model variability in the absence of direct supervision [PITH_FULL_IMA… view at source ↗
Figure 10
Figure 10. Figure 10: A limited number of measurement locations are distributed across [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Re = 10,000 — Bayesian PINN velocity diagnostics. The panels show the reference solution, posterior mean prediction, absolute error, and predictive standard deviation, for the velocity components (U, V ). The results demonstrate accurate reconstruction of the velocity field and good alignment between predicted uncertainty and pointwise error across the domain after recalibration [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 12
Figure 12. Figure 12: Re = 10,000 — Bayesian PINN Reynolds stress diagnostics. The panels present the reference fields, posterior mean predictions, absolute error, and predictive standard deviation for the closure terms (fx, fy). While the mean predictions capture the overall structure, the uncertainty estimates reflect increased modeling difficulty in the closure terms [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Re = 10,000 — Bayesian PINN: inferred mean pressure field (P). No pressure data are used during training, and the pressure is recovered solely through the governing equations. The posterior mean captures the large-scale structure of the pressure field, while the predictive uncertainty provides a measure of confidence in the absence of direct observations. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Re = 10,000 — RDE-PINN velocity diagnostics. The panels show the reference solution, posterior mean prediction, absolute error, and predictive standard deviation, for the velocity components (U, V ). The results demonstrate accurate reconstruction of the velocity field and good alignment between predicted uncertainty and pointwise error across the domain after recalibration [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 15
Figure 15. Figure 15: Re = 10,000 — RDE-PINN Reynolds stress diagnostics. The panels present the reference fields, posterior mean predictions, absolute error, and predictive standard deviation for the closure terms (fx, fy). While the mean predictions capture the overall structure, the uncertainty estimates reflect increased modeling difficulty in the closure terms. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
read the original abstract

Physics-informed neural networks (PINNs) have emerged as a promising framework for solving inverse problems governed by partial differential equations (PDEs), including the reconstruction of turbulent flow fields from sparse data. However, most existing PINN formulations are deterministic and do not provide reliable quantification of epistemic uncertainty, which is critical for ill-posed problems such as data-driven Reynolds-averaged Navier-Stokes (RANS) modeling. In this work, we develop and systematically evaluate a set of probabilistic extensions of PINNs for uncertainty quantification in turbulence modeling. The proposed framework combines (i) Bayesian PINNs with Hamiltonian Monte Carlo sampling and a tempered multi-component likelihood, (ii) Monte Carlo dropout, and (iii) repulsive deep ensembles that enforce diversity in function space. Particular emphasis is placed on the role of ensemble diversity and likelihood tempering in improving uncertainty calibration for PDE-constrained inverse problems. The methods are assessed on a hierarchy of test cases, including the Van der Pol oscillator and turbulent flow past a circular cylinder at Reynolds numbers Re=3,900 (direct numerical simulation data) and Re = 10,000 (experimental particle image velocimetry data). The results demonstrate that Bayesian PINNs provide the most consistent uncertainty estimates across all inferred quantities, while function-space repulsive ensembles offer a computationally efficient approximation with competitive accuracy for primary flow variables. These findings provide quantitative insight into the trade-offs between accuracy, computational cost, and uncertainty calibration in physics-informed learning, and offer practical guidance for uncertainty quantification in data-driven turbulence modeling.

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

1 major / 2 minor

Summary. The manuscript develops and evaluates probabilistic extensions of physics-informed neural networks (PINNs) for uncertainty quantification in inverse problems for turbulent flows. It introduces Bayesian PINNs using Hamiltonian Monte Carlo with a tempered multi-component likelihood, Monte Carlo dropout, and repulsive deep ensembles enforcing function-space diversity. These are assessed on a test hierarchy consisting of the Van der Pol oscillator and cylinder flow at Re=3900 (DNS data) and Re=10,000 (PIV data). The central claim is that Bayesian PINNs deliver the most consistent uncertainty estimates across inferred quantities, while repulsive ensembles provide a computationally efficient approximation with competitive accuracy on primary flow variables.

Significance. If the empirical ranking holds under broader validation, the work supplies practical guidance on accuracy-cost-calibration trade-offs for physics-informed learning in data-driven turbulence modeling, particularly for ill-posed RANS-type inverse problems where epistemic uncertainty matters.

major comments (1)
  1. [Test cases and results] The evaluation hierarchy (Van der Pol oscillator plus cylinder at Re=3900 DNS and Re=10,000 PIV) primarily exercises data sparsity and moderate nonlinearity. It does not include regimes in which PDE residual errors dominate or in which tempering materially shifts posterior mass on secondary quantities such as Reynolds stresses or pressure. Consequently, the claim that Bayesian PINNs yield the 'most consistent' uncertainty estimates (and that repulsive ensembles are competitive) rests on test cases that may not expose systematic mis-calibration introduced by the tempered likelihood or enforced diversity; this is load-bearing for the central ranking.
minor comments (2)
  1. [Abstract] The abstract states comparative results on standard test cases yet supplies no quantitative metrics, error bars, or ablation details on tempering/diversity parameters; the full manuscript should make these explicit in the results section.
  2. [Methods] Clarify the precise definition and implementation of the 'tempered multi-component likelihood' and the function-space repulsion term (including any hyperparameters) so that the reported calibration behavior can be reproduced.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the manuscript's significance. We address the major comment on the evaluation hierarchy below, providing clarification on our test case rationale while agreeing to strengthen the discussion of limitations.

read point-by-point responses

  1. Referee: [Test cases and results] The evaluation hierarchy (Van der Pol oscillator plus cylinder at Re=3900 DNS and Re=10,000 PIV) primarily exercises data sparsity and moderate nonlinearity. It does not include regimes in which PDE residual errors dominate or in which tempering materially shifts posterior mass on secondary quantities such as Reynolds stresses or pressure. Consequently, the claim that Bayesian PINNs yield the 'most consistent' uncertainty estimates (and that repulsive ensembles are competitive) rests on test cases that may not expose systematic mis-calibration introduced by the tempered likelihood or enforced diversity; this is load-bearing for the central ranking.

    Authors: We appreciate this observation on the scope of our test cases. The hierarchy was selected to span increasing levels of complexity relevant to turbulent flow inverse problems: from the nonlinear Van der Pol oscillator (exercising the tempered multi-component likelihood) to cylinder flows at Re=3900 (DNS) and Re=10,000 (PIV), where data sparsity, turbulence effects, and PDE-data balance are central. In these cases, the inverse problem inherently involves non-negligible PDE residuals due to the approximate RANS-like modeling and sparse observations, and our results demonstrate that tempering improves calibration for both primary variables and secondary quantities such as Reynolds stresses. Nevertheless, we agree that additional regimes with more dominant PDE errors or explicit sensitivity analysis of tempering on posterior mass could further substantiate the ranking. We will revise the manuscript to expand the discussion of these limitations, include a dedicated paragraph on potential mis-calibration risks from tempering and diversity enforcement, and add a forward-looking statement on future validation in more challenging regimes. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical comparisons with independent DNS and PIV benchmarks

full rationale

The paper proposes probabilistic extensions of PINNs (Bayesian HMC with tempered likelihood, MC dropout, repulsive ensembles) and evaluates them on a hierarchy of test problems including the Van der Pol oscillator and cylinder flow at Re=3900 (DNS) and Re=10000 (PIV). The central claims concern relative consistency of uncertainty estimates and computational trade-offs; these are supported by direct numerical comparisons against external data rather than by any derivation that reduces to a fitted parameter, self-citation chain, or redefinition of inputs. No load-bearing step equates a reported uncertainty or ranking to its own construction by definition. Minor self-citations of prior PINN work are present but do not carry the empirical ranking results.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review; the framework rests on the domain assumption that PINNs can be extended to probabilistic inference for RANS-type inverse problems without violating the underlying PDE constraints.

axioms (1)
  • domain assumption PINNs can be trained to solve ill-posed inverse problems for turbulent flows governed by RANS equations
    Invoked as the starting point for all probabilistic extensions

pith-pipeline@v0.9.0 · 5590 in / 1253 out tokens · 41014 ms · 2026-05-10T06:16:36.772042+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Posterior Concentration of Bayesian Physics-Informed Neural Networks for Elliptic PDEs

    math.ST 2026-05 unverdicted novelty 7.0

    Bayesian PINNs for elliptic PDEs have posteriors that contract around the true solution at near-optimal rates, with the prior adapting automatically to unknown smoothness.

  2. Functional-prior-based Bayesian PDE-constrained inversion using PINNs

    physics.geo-ph 2026-05 unverdicted novelty 6.0

    Presents fpBPINN framework using FPI-BPINN and fParVI-PINN to enable functional priors in Bayesian PINN-based PDE inversion, with random Fourier features aiding Gaussian prior representation.

Reference graph

Works this paper leans on

47 extracted references · 9 canonical work pages · cited by 2 Pith papers

  1. [1]

    D. C. Wilcox, Turbulence Modeling for CFD, DCW Industries, 2006

    2006

  2. [2]

    S. B. Pope, Turbulent Flows, Cambridge University Press, 2000

    2000

  3. [3]

    P. A. Durbin, Some recent developments in turbulence closure modeling, Annual Review of Fluid Mechanics 50 (2018) 77–103

    2018

  4. [4]

    P. R. Spalart, Detached-eddy simulation, Annual Review of Fluid Mechan- ics 41 (2009) 181–202

    2009

  5. [5]

    P. Moin, K. Mahesh, Direct numerical simulation: A tool in turbulence research, Annual Review of Fluid Mechanics 30 (1998) 539–578

    1998

  6. [6]

    P.Sagaut, LargeEddySimulationforIncompressibleFlows, Springer, 2006

    2006

  7. [7]

    Duraisamy, G

    K. Duraisamy, G. Iaccarino, H. Xiao, Turbulence modeling in the age of data, Annual Review of Fluid Mechanics 51 (2019) 357–377

    2019

  8. [8]

    J. Ling, A. Kurzawski, J. Templeton, Reynolds averaged turbulence mod- elling using deep neural networks, Journal of Fluid Mechanics 807 (2016) 155–166

    2016

  9. [9]

    Emory, J

    M. Emory, J. Larsson, G. Iaccarino, Modeling of structural uncertainties in Reynolds-averaged Navier–Stokes closures, Physics of Fluids 25 (11) (2013) 110822.doi:10.1063/1.4824659

    work page doi:10.1063/1.4824659 2013

  10. [10]

    Iaccarino, A

    G. Iaccarino, A. A. Mishra, S. Ghili, Eigenspace perturbations for un- certainty estimation of single-point turbulence closures, Physical Review Fluids 2 (2) (2017) 024605.doi:10.1103/PhysRevFluids.2.024605

    work page doi:10.1103/physrevfluids.2.024605 2017

  11. [11]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics 378 (2019) 686–707.doi:10.1016/j.jcp.2018.10.045

    work page doi:10.1016/j.jcp.2018.10.045 2019

  12. [12]

    Shukla, P

    K. Shukla, P. C. Di Leoni, J. Blackshire, D. Sparkman, G. E. Karniadakis, Physics-informed neural network for ultrasound nondestructive quantifica- tion of surface breaking cracks, Journal of Nondestructive Evaluation 39 (3) (2020) 61

    2020

  13. [13]

    Shukla, A

    K. Shukla, A. D. Jagtap, G. E. Karniadakis, Parallel physics-informed neu- ral networks via domain decomposition, Journal of Computational Physics 447 (2021) 110683

    2021

  14. [14]

    Patel, V

    Y. Patel, V. Mons, O. Marquet, G. Rigas, Turbulence model augmented physics-informed neural networks for mean-flow reconstruction, Physical Review Fluids 9 (3) (2024) 034605. 32

    2024

  15. [15]

    Bureš, Y

    H. Eivazi, M. Tahani, P. Schlatter, R. Vinuesa, Physics-informed neural networks for the Reynolds-averaged Navier–Stokes equations, Journal of Computational Physics 452 (2022) 110912.doi:10.1016/j.jcp.2021. 110912

    work page doi:10.1016/j.jcp.2021 2022

  16. [16]

    Barrett, K

    Z.Zhang, K.Shukla, Z.Wang, A.Morales, T.Käufer, S.Salauddin, N.Wal- ters, D. Barrett, K. Ahmed, M. Triantafyllou, et al., Turbulence closure in rans and flow inference around a cylinder using pinns and sparse experi- mental data, arXiv preprint arXiv:2510.06049 (2025)

    work page arXiv 2025

  17. [17]

    A. F. Psaros, X. Meng, Z. Zou, L. Guo, G. E. Karniadakis, Uncertainty quantification in scientific machine learning: Methods, metrics, and com- parisons, Journal of Computational Physics 477 (2023) 111902

    2023

  18. [18]

    L.Yang, X.Meng, G.E.Karniadakis, B-PINNs: Bayesianphysics-informed neural networks for forward and inverse pde problems with noisy data, Journal of Computational Physics 425 (2021) 109913

    2021

  19. [19]

    Z. Zou, X. Meng, G. E. Karniadakis, Uncertainty quantification for noisy inputs–outputs in physics-informed neural networks and neural operators, Computer Methods in Applied Mechanics and Engineering 433 (2025) 117479

    2025

  20. [20]

    H. Liu, Z. Wang, R. Deng, S. Wang, X. Meng, C. Xu, S. Cai, Flow recon- struction with uncertainty quantification from noisy measurements based on bayesian physics-informed neural networks, Physics of Fluids 36 (11) (2024)

    2024

  21. [21]

    X. Meng, L. Yang, Z. Mao, J. del Águila Ferrandis, G. E. Karniadakis, Learning functional priors and posteriors from data and physics, Journal of Computational Physics 457 (2022) 111073

    2022

  22. [22]

    Z. Zou, G. E. Karniadakis, Multi-head physics-informed neural networks for learning functional priors and uncertainty quantification, Journal of Computational Physics 531 (2025) 113947

    2025

  23. [23]

    X. Meng, H. Babaee, G. E. Karniadakis, Multi-fidelity Bayesian neural networks: Algorithms and applications, Journal of Computational Physics 438 (2021) 110361

    2021

  24. [24]

    Aikawa, N

    Y. Aikawa, N. Ueda, T. Tanaka, Improving the efficiency of training physics-informed neural networks using active learning, New Generation Computing 42 (4) (2024) 739–760

    2024

  25. [25]

    Z. Zou, T. Meng, P. Chen, J. Darbon, G. E. Karniadakis, Leveraging vis- cous Hamilton–Jacobi PDEs for uncertainty quantification in scientific ma- chine learning, SIAM/ASA Journal on Uncertainty Quantification 12 (4) (2024) 1165–1191. 33

    2024

  26. [26]

    Linka, A

    K. Linka, A. Schäfer, X. Meng, Z. Zou, G. E. Karniadakis, E. Kuhl, Bayesian physics informed neural networks for real-world nonlinear dynam- icalsystems, ComputerMethodsinAppliedMechanicsandEngineering402 (2022) 115346

    2022

  27. [27]

    R. M. Neal, Bayesian Learning for Neural Networks, Vol. 118 of Lecture Notes in Statistics, Springer, 1996.doi:10.1007/978-1-4612-0745-0

    work page doi:10.1007/978-1-4612-0745-0 1996

  28. [28]

    R. M. Neal, MCMC using Hamiltonian dynamics, in: Handbook of Markov Chain Monte Carlo, Chapman and Hall/CRC, 2011, pp. 113–162

    2011

  29. [29]

    M. D. Hoffman, A. Gelman, The No-U-Turn sampler: Adaptively setting path lengths in Hamiltonian Monte Carlo, Journal of Machine Learning Research 15 (47) (2014) 1593–1623

    2014

  30. [30]

    Y. Gal, Z. Ghahramani, Dropout as a Bayesian approximation: Repre- senting model uncertainty in deep learning, in: Proceedings of the 33rd International Conference on Machine Learning (ICML), 2016

    2016

  31. [31]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, C. Blundell, Simple and scalable predic- tive uncertainty estimation using deep ensembles, in: Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  32. [32]

    D’Angelo, V

    F. D’Angelo, V. Fortuin, Repulsive deep ensembles are Bayesian, in: Ad- vances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  33. [33]

    arXiv preprint arXiv:2505.17308

    P. Pilar, M. Heinonen, N. Wahlström, Repulsive Ensembles for Bayesian Inference in Physics-informed Neural Networks, arXiv preprint arXiv:2505.17308 (2025)

    work page arXiv 2025

  34. [34]

    Zhang, L

    D. Zhang, L. Lu, L. Guo, G. E. Karniadakis, Quantifying total uncer- tainty in physics-informed neural networks for solving forward and inverse stochastic problems, Journal of Computational Physics 397 (2019) 108850

    2019

  35. [35]

    Wenzel, K

    F. Wenzel, K. Roth, B. S. Veeling, J. Świątkowski, L. Tran, S. Mandt, J. Snoek, T. Salimans, R. Jenatton, S. Nowozin, How good is the Bayes posterior in deep neural networks really?, in: Proceedings of the 37th In- ternational Conference on Machine Learning (ICML), 2020

    2020

  36. [36]

    Aitchison, A statistical theory of cold posteriors in deep neural networks, in: International Conference on Learning Representations (ICLR), 2021

    L. Aitchison, A statistical theory of cold posteriors in deep neural networks, in: International Conference on Learning Representations (ICLR), 2021

    2021

  37. [37]

    Kapoor, A

    S. Kapoor, A. Narayanan, Leakage and the reproducibility crisis in machine-learning-based science, Patterns 4 (9) (2023) 100804.doi:10. 1016/j.patter.2023.100804

    work page arXiv 2023

  38. [38]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, in: International Conference on Learning Representations (ICLR), 2015. 34

    2015

  39. [39]

    D. C. Liu, J. Nocedal, On the limited memory BFGS method for large scale optimization, Mathematical Programming 45 (1–3) (1989) 503–528. doi:10.1007/BF01589116

    work page doi:10.1007/bf01589116 1989

  40. [40]

    Z. Zou, Z. Wang, G. Em Karniadakis, Learning and discovering multiple solutionsusingphysics-informedneuralnetworkswithrandominitialization and deep ensemble, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 481 (2325) (2025)

    2025

  41. [41]

    Z. Zou, X. Meng, G. E. Karniadakis, Correcting model misspecification in physics-informed neural networks (PINNs), Journal of Computational Physics 505 (2024) 112918

    2024

  42. [42]

    Kuleshov, N

    V. Kuleshov, N. Fenner, S. Ermon, Accurate uncertainties for deep learn- ing using calibrated regression, in: Proceedings of the 35th International Conference on Machine Learning (ICML), 2018

    2018

  43. [43]

    J. Yao, W. Pan, S. Ghosh, F. Doshi-Velez, Quality of uncertainty quantifi- cation for Bayesian neural network inference, in: Workshop on Uncertainty and Robustness in Deep Learning, ICML, 2019

    2019

  44. [44]

    Z. Zou, X. Meng, A. F. Psaros, G. E. Karniadakis, NeuralUQ: A compre- hensivelibraryforuncertaintyquantificationinneuraldifferentialequations and operators, SIAM Review 66 (1) (2024) 161–190

    2024

  45. [45]

    Fischer, S

    P. Fischer, S. Kerkemeier, M. Min, Y.-H. Lan, M. Phillips, T. Rathnayake, E. Merzari, A. Tomboulides, A. Karakus, N. Chalmers, et al., NekRS, a GPU-accelerated spectral element Navier-Stokes solver, Parallel Comput- ing 114 (2022) 102982

    2022

  46. [46]

    D. Fan, G. Jodin, T. Consi, L. Bonfiglio, Y. Ma, L. Keyes, G. E. Karni- adakis, M. S. Triantafyllou, A robotic intelligent towing tank for learning complex fluid-structure dynamics, Science Robotics 4 (36) (2019) eaay5063

    2019

  47. [47]

    Westerweel, F

    J. Westerweel, F. Scarano, Universal outlier detection for PIV data, Ex- periments in Fluids 39 (6) (2005) 1096–1100. 35

    2005