Pith. sign in

REVIEW 2 major objections 29 references

Conditional normalizing flows learn a probabilistic mapping from low-fidelity to high-fidelity ROM coefficients to close the model with uncertainty quantification.

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 ↗

Conditional normalizing flows learn a probabilistic mapping from low-fidelity to high-fidelity ROM coefficients for closure correction in 2D Navier-Stokes vortex merging, with direct and residual strategies providing uncertainty quantification.

T0 review reviewed 2026-06-29 challenge →

load-bearing objection Conditional normalizing flows for multi-fidelity ROM closure work on the vortex merging case with residual learning ahead, but the generalization claim rests on thin evidence. the 2 major comments →

arxiv 2606.09857 v1 pith:JHTTDQ2L submitted 2026-05-27 cs.LG physics.comp-ph

Uncertainty-aware Multi-fidelity Closure via Conditional Normalizing Flows

Jice Zeng , Shady E. Ahmed , David Barajas-Solano , Panos Stinis This is my paper
classification cs.LG physics.comp-ph
keywords reduced order modelsmulti-fidelitynormalizing flowsclosure modelinguncertainty quantificationNavier-Stokesvortex mergingmachine learning
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 reading

The paper formulates the closure problem in reduced-order models as a multi-fidelity learning task. It uses conditional normalizing flows to learn a probabilistic mapping from low-fidelity ROM coefficients to high-fidelity ones. This improves accuracy on the vortex merging problem while providing uncertainty estimates. Residual learning of the discrepancy between fidelities outperforms direct learning of high-fidelity coefficients. The resulting framework supports more reliable application of ROMs by quantifying prediction confidence.

Core claim

The proposed uncertainty-aware multi-fidelity framework based on conditional normalizing flow learns a probabilistic mapping from low-fidelity ROM coefficients to high-fidelity coefficients, improving predictive fidelity while quantifying uncertainty; residual learning achieves consistently better performance than direct learning on the vortex merging problem.

What carries the argument

Conditional normalizing flow modeling the conditional probability distribution of high-fidelity reduced-order model coefficients given low-fidelity inputs.

Load-bearing premise

Sufficient paired low-fidelity and high-fidelity data must exist to train the conditional normalizing flow so that the mapping generalizes beyond training cases without the flow overfitting to irrelevant scale interactions.

What would settle it

Running the trained model on a new vortex merging case with parameters outside the training set and checking if the high-fidelity coefficient predictions match independent high-fidelity simulations within the reported uncertainty bounds.

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

Share X Bluesky LinkedIn Reddit HN

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

2 major / 0 minor

Summary. The paper formulates ROM closure modeling as a multi-fidelity learning problem and proposes an uncertainty-aware framework based on conditional normalizing flows to learn a probabilistic mapping from low-fidelity ROM coefficients to high-fidelity coefficients. Two strategies are investigated—direct learning of HF coefficients and residual learning of the LF-HF discrepancy—with the framework demonstrated on a 2D Navier-Stokes vortex merging problem. The abstract claims that both strategies improve accuracy over the uncorrected ROM, with residual learning performing better, while also providing UQ for the corrected coefficients.

Significance. If the central claims hold with proper validation, the work offers a generative modeling approach to data-driven closure that naturally incorporates uncertainty quantification, which could be valuable for reliable use of ROMs in multiscale applications. The explicit comparison of direct versus residual correction strategies is a positive aspect, as is the focus on probabilistic mappings via normalizing flows rather than deterministic corrections.

major comments (2)
  1. [Results section] Results section: the demonstration is restricted to the single vortex-merging trajectory governed by 2D Navier-Stokes; no out-of-distribution test cases, cross-validation across different initial conditions or Reynolds numbers, or ablation on training-set size are reported. This is load-bearing for the claim that the learned probabilistic mapping 'improves predictive fidelity' in a manner that supports practical ROM use, as the skeptic concern about memorization of training statistics rather than learning a transferable closure is not addressed.
  2. [Abstract] Abstract and results: the central performance claim states that 'both correction strategies improve ROM accuracy over uncorrected ROM, with residual learning achieving consistently better performance,' yet no quantitative metrics, error bars, baseline comparisons, dataset sizes, or validation protocol details are supplied. Without these, the soundness of the reported residual-learning advantage cannot be evaluated.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for their constructive feedback on our manuscript. We address each major comment below and outline the revisions we will make to improve the clarity and rigor of the presentation.

read point-by-point responses

  1. Referee: [Results section] Results section: the demonstration is restricted to the single vortex-merging trajectory governed by 2D Navier-Stokes; no out-of-distribution test cases, cross-validation across different initial conditions or Reynolds numbers, or ablation on training-set size are reported. This is load-bearing for the claim that the learned probabilistic mapping 'improves predictive fidelity' in a manner that supports practical ROM use, as the skeptic concern about memorization of training statistics rather than learning a transferable closure is not addressed.

    Authors: We agree that extending the validation to out-of-distribution cases would strengthen the claims regarding the generalizability of the learned closure. The current study focuses on demonstrating the framework on a canonical vortex merging problem, which is a standard benchmark in the ROM literature for 2D Navier-Stokes. To partially address the concern, we will add an ablation study on the effect of training set size in the revised results section. However, generating new HF simulations for different Reynolds numbers or initial conditions is computationally intensive and beyond the scope of this work, which prioritizes introducing the conditional normalizing flow approach for MF closure. revision: partial

  2. Referee: [Abstract] Abstract and results: the central performance claim states that 'both correction strategies improve ROM accuracy over uncorrected ROM, with residual learning achieving consistently better performance,' yet no quantitative metrics, error bars, baseline comparisons, dataset sizes, or validation protocol details are supplied. Without these, the soundness of the reported residual-learning advantage cannot be evaluated.

    Authors: We acknowledge the lack of specific quantitative information in the abstract and results. In the revised manuscript, we will update the abstract to include key performance metrics, such as the relative errors for the ROM coefficients under both strategies, and mention the dataset details (e.g., number of training snapshots from the HF simulation). The results section will be expanded to include error bars derived from sampling the normalizing flow, baseline comparisons with the uncorrected ROM, and a clear description of the validation protocol used to evaluate the probabilistic mappings. revision: yes

standing simulated objections not resolved
  • Complete out-of-distribution testing across varying Reynolds numbers and initial conditions, which would necessitate additional high-fidelity simulations not included in the current study.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper presents an empirical machine-learning framework that trains conditional normalizing flows on paired LF/HF ROM coefficient data to learn a probabilistic correction mapping. The claimed accuracy gains (residual learning outperforming direct learning) and UQ capability are demonstrated via numerical experiments on the 2D Navier-Stokes vortex-merging case rather than derived from any closed-form equations that reduce to the training inputs by construction. No self-definitional steps, fitted-input-as-prediction reductions, or load-bearing self-citations appear in the provided abstract or description; the central result remains an externally falsifiable empirical outcome on held-out snapshots within the same problem.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the normalizing flow successfully learning a generalizable probabilistic mapping from paired LF/HF data; this rests on standard assumptions about flow invertibility and data representativeness rather than new axioms or entities.

free parameters (1)
  • Conditional normalizing flow parameters
    Trainable neural network weights and flow parameters fitted during learning of the LF-to-HF mapping.
axioms (1)
  • domain assumption Conditional normalizing flows can represent the conditional distribution of HF coefficients given LF inputs
    Core modeling assumption invoked when formulating the probabilistic mapping.

reviewed 2026-06-29 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-aware Multi-fidelity Closure via Conditional Normalizing Flows." pith.science (2026). https://pith.science/paper/JHTTDQ2L

@misc{pith2026260609857,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-aware Multi-fidelity Closure via Conditional Normalizing Flows},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JHTTDQ2L}},
  note         = {Machine review of arXiv:2606.09857}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reduced-order models (ROMs) provide an efficient surrogate for complex multiscale systems, but their predictive accuracy is often compromised by truncation errors and the inadequate representation of interactions between resolved and unresolved scales. The missing effect of truncated (unresolved) scales on ROM (resolved) scales is often denoted as the closure problem. In this work, we formulate ROM closure modeling as a multi-fidelity (MF) learning problem and propose an uncertainty-aware MF framework based on conditional normalizing flow to enhance ROM predictive accuracy. The proposed approach learns a probabilistic mapping from low-fidelity (LF) ROM coefficients to high-fidelity (HF) coefficients, thereby improving predictive fidelity while quantifying the uncertainty associated with the learned closure. Two correction strategies are investigated: direct learning, in which HF coefficients are predicted directly from LF inputs, and residual learning, which learns the discrepancy between LF and HF coefficients and uses it to recover the corrected HF solution. The framework is demonstrated on a vortex merging problem governed by the two-dimensional Navier Stokes equations. Results show that both correction strategies improve ROM accuracy over uncorrected ROM, with residual learning achieving consistently better performance than direct learning. Moreover, the two proposed deep generative model-based strategies provide uncertainty quantification for the corrected ROM coefficients, which is critical for assessing prediction confidence and supporting the reliable use of ROMs in practical applications.

Figures

Figures reproduced from arXiv: 2606.09857 by David Barajas-Solano, Jice Zeng, Panos Stinis, Shady E. Ahmed.

Figure 1
Figure 1. Figure 1: A schematic diagram of the proposed probabilistic multi-fidelity framework. [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of conditional normalizing flow [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Probabilistic MF closure correction by the DL method for [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Predicted MF probability distributions for the POD coefficients [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Probabilistic MF closure correction by the RL method for [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Predicted MF probability distributions for the POD coefficients [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: MF predictions obtained by the RL method with 5% and 10% Gaussian noise [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: MF predictions obtained by the RL method with 15% and 20% Gaussian noise [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Quantitative comparison of the DL and RL methods in terms of the RL2E. A [PITH_FULL_IMAGE:figures/full_fig_p033_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Quantitative comparison of the DL and RL methods in terms of the LPP. [PITH_FULL_IMAGE:figures/full_fig_p033_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Quantitative comparison of the DL and RL methods in terms of the empirical [PITH_FULL_IMAGE:figures/full_fig_p034_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Calibration curves for the DL and RL methods under different noise levels. [PITH_FULL_IMAGE:figures/full_fig_p035_12.png] view at source ↗

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

29 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    D. J. Lucia, P. S. Beran, W. A. Silva, Reduced-order modeling: new approaches for computational physics, Progress in aerospace sciences 40 (2004) 51–117

    2004

  2. [2]

    Padula, M

    G. Padula, M. Girfoglio, G. Rozza, A brief review of reduced order models using intrusive and non-intrusive techniques, PAMM 24 (2024) e202400210

    2024

  3. [3]

    Liang, H

    Y. Liang, H. Lee, S. Lim, W. Lin, K. Lee, C. Wu, Proper orthogonal decomposition and its applications—part i: Theory, Journal of Sound and vibration 252 (2002) 527–544

    2002

  4. [4]

    Ullmann, M

    S. Ullmann, M. Rotkvic, J. Lang, Pod-galerkin reduced-order model- ing with adaptive finite element snapshots, Journal of Computational Physics 325 (2016) 244–258

    2016

  5. [5]

    S. E. Ahmed, S. Pawar, O. San, A. Rasheed, T. Iliescu, B. R. Noack, On closures for reduced order models—a spectrum of first-principle to machine-learned avenues, Physics of Fluids 33 (2021). 37

    2021

  6. [6]

    Sanderse, P

    B. Sanderse, P. Stinis, R. Maulik, S. E. Ahmed, Scientific machine learning for closure models in multiscale problems: A review, arXiv preprint arXiv:2403.02913 (2024)

    work page arXiv 2024

  7. [7]

    S. E. Ahmed, O. San, K. Kara, R. Younis, A. Rasheed, Multifidelity computing for coupling full and reduced order models, Plos one 16 (2021) e0246092

    2021

  8. [8]

    L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via deeponet based on the universal approximation theorem of operators, Nature machine intelligence 3 (2021) 218–229

    2021

  9. [9]

    S. E. Ahmed, P. Stinis, A multifidelity deep operator network approach to closure for multiscale systems, Computer Methods in Applied Me- chanics and Engineering 414 (2023) 116161

    2023

  10. [10]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stu- art, A. Anandkumar, Fourier neural operator for parametric partial differential equations, arXiv preprint arXiv:2010.08895 (2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  11. [11]

    Y. Lyu, X. Zhao, Z. Gong, X. Kang, W. Yao, Multi-fidelity prediction of fluid flow based on transfer learning using fourier neural operator, Physics of Fluids 35 (2023)

    2023

  12. [12]

    L. Lu, X. Meng, S. Cai, Z. Mao, S. Goswami, Z. Zhang, G. E. Karni- adakis, A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data, Computer Methods in Applied Mechanics and Engineering 393 (2022) 114778

    2022

  13. [13]

    Bülte, P

    C. Bülte, P. Scholl, G. Kutyniok, Probabilistic predictions with fourier neural operators, in: NeurIPS 2024 Workshop on Bayesian Decision- making and Uncertainty, 2024

    2024

  14. [14]

    J.L.Callaham, J.-C.Loiseau, S.L.Brunton, Multiscalemodelreduction for incompressible flows, Journal of Fluid Mechanics 973 (2023) A3

    2023

  15. [15]

    W.Snyder, C.Mou, H.Liu, O.San, R.DeVita, T.Iliescu, Reducedorder model closures: a brief tutorial, in: Recent Advances in Mechanics and Fluid-Structure Interaction with Applications: The Bong Jae Chung Memorial Volume, Springer, 2022, pp. 167–193. 38

    2022

  16. [16]

    S. E. Ahmed, O. San, A. Rasheed, T. Iliescu, A. Veneziani, Physics guided machine learning for variational multiscale reduced order mod- eling, SIAM Journal on Scientific Computing 45 (2023) B283–B313

    2023

  17. [17]

    Bhola, K

    S. Bhola, K. Duraisamy, Flow matching operators for residual- augmented probabilistic learning of partial differential equations, arXiv preprint arXiv:2512.12749 (2025)

    work page arXiv 2025

  18. [18]

    G.Papamakarios, E.Nalisnick, D.J.Rezende, S.Mohamed, B.Lakshmi- narayanan, Normalizing flows for probabilistic modeling and inference, Journal of Machine Learning Research 22 (2021) 1–64

    2021

  19. [19]

    C.Winkler, D.Worrall, E.Hoogeboom, M.Welling, Learninglikelihoods with conditional normalizing flows, arXiv preprint arXiv:1912.00042 (2019)

    work page arXiv 1912

  20. [20]

    J. Zeng, Y. Wang, A. M. Tartakovsky, D. A. Barajas-Solano, Solving high-dimensional inverse problems using amortized likelihood-free infer- ence with noisy and incomplete data, Computer Methods in Applied Mechanics and Engineering 443 (2025) 118064

    2025

  21. [21]

    Serrin, D

    J. Serrin, D. E. Varberg, A general chain rule for derivatives and the change of variables formula for the lebesgue integral, The American Mathematical Monthly 76 (1969) 514–520

    1969

  22. [22]

    Papamakarios, T

    G. Papamakarios, T. Pavlakou, I. Murray, Masked autoregressive flow for density estimation, Advances in neural information processing sys- tems 30 (2017)

    2017

  23. [23]

    Germain, K

    M. Germain, K. Gregor, I. Murray, H. Larochelle, Made: Masked au- toencoder for distribution estimation, in: International conference on machine learning, PMLR, 2015, pp. 881–889

    2015

  24. [24]

    Meunier, S

    P. Meunier, S. Le Dizès, T. Leweke, Physics of vortex merging, Comptes Rendus Physique 6 (2005) 431–450

    2005

  25. [25]

    Pawar, O

    S. Pawar, O. San, A. Nair, A. Rasheed, T. Kvamsdal, Model fusion with physics-guided machine learning: Projection-based reduced-order modeling, Physics of Fluids 33 (2021). 39

    2021

  26. [26]

    Arakawa, Computational design for long-term numerical integration of the equations of fluid motion: Two-dimensional incompressible flow

    A. Arakawa, Computational design for long-term numerical integration of the equations of fluid motion: Two-dimensional incompressible flow. part i, Journal of Computational Physics 1 (1966) 119–143

    1966

  27. [27]

    Gottlieb, C.-W

    S. Gottlieb, C.-W. Shu, Total variation diminishing Runge-Kutta schemes, Mathematics of Computation 67 (1998) 73–85

    1998

  28. [28]

    Y. Zong, D. Barajas-Solano, A. M. Tartakovsky, Randomized physics- informed neural networks for bayesian data assimilation, Computer Methods in Applied Mechanics and Engineering 436 (2025) 117670

    2025

  29. [29]

    S. I. Allec, M. Ziatdinov, Active and transfer learning with partially bayesian neural networks for materials and chemicals, Digital Discovery 4 (2025) 1284–1297. 40

    2025

This paper was first reviewed by grok-4.3 on June 29, 2026.