pith. sign in

arxiv: 2604.18059 · v1 · submitted 2026-04-20 · ⚛️ physics.flu-dyn

Information decomposition for disentangled and interpretable manifold learning of fluid flows via variational autoencoders

Zhiyuan Wang , Iacopo Tirelli , Stefano Discetti , Andrea Ianiro This is my paper

Pith reviewed 2026-05-10 03:58 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn
keywords variational autoencodersKL divergence decompositiondisentangled representationsmanifold learningfluid dynamicsinterpretable machine learninginformation theoryflow field data
0
0 comments X

The pith

A three-term decomposition of the KL divergence in VAEs enables separate control of compression, disentanglement, and regularization to produce interpretable manifolds from fluid flow data.

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

The paper introduces an information-theoretic framework for variational autoencoders that decomposes the KL divergence into index-code mutual information, total correlation, and dimension-wise KL divergence. These terms separately regulate data compression, latent disentanglement, and geometric regularization. This targeted control yields compact manifolds that separate distinct physical effects in flow data while avoiding the capacity loss common in heavily regularized VAEs. The method is demonstrated on two synthetic unsteady flow datasets: cylinder flows with variable position, diameter, and Reynolds number, and NACA 0012 airfoil flows with varying angles of attack and vortex gust parameters. Comparisons against PCA, Isometric Feature Mapping, and beta-VAE confirm improved disentanglement and physical interpretability, with robustness to changes in loss weights.

Core claim

Decomposing the variational KL divergence into the index-code mutual information, the total correlation, and the dimension-wise KL divergence provides independent regulation of compression, disentanglement, and regularization. Applied to high-dimensional flow-field data, this yields latent coordinates that isolate individual physical parameters such as cylinder position and diameter or airfoil angle of attack and gust intensity, while preserving reconstruction accuracy.

What carries the argument

The three-term decomposition of the KL divergence in the VAE variational objective, where index-code mutual information controls compression, total correlation controls disentanglement, and dimension-wise KL divergence controls geometric regularization.

If this is right

  • Latent coordinates isolate distinct physical effects such as cylinder position, diameter, Reynolds number, airfoil angle of attack, and gust parameters.
  • Interpretability improves without the information-capacity reduction typical of heavily regularized VAEs.
  • The approach outperforms PCA, Isometric Feature Mapping, and beta-VAE in disentanglement and physical separation on both tested flow configurations.
  • Robustness to loss-weighting variations holds despite the larger number of tunable parameters.

Where Pith is reading between the lines

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

  • The explicit separation of effects could support reduced-order models that retain clear physical meaning for flow prediction or control tasks.
  • The framework might extend to other high-dimensional scientific datasets where isolating mechanisms matters, such as turbulence statistics or image-based flow sensing.
  • Combining the decomposition with additional physics-based loss terms could further constrain the latent space to obey conservation laws.

Load-bearing premise

That the three-term decomposition of KL divergence provides independent control over compression, disentanglement, and regularization without introducing new unintended trade-offs in the learned manifold for fluid data.

What would settle it

On a new unsteady flow dataset, increasing the total-correlation weight fails to improve separation of physical parameters like Reynolds number and gust length scale, or it reduces reconstruction fidelity more than the same weight change in a standard beta-VAE.

read the original abstract

We introduce an information-theoretic framework that uses variational autoencoders (VAEs) to extract compact, physically interpretable manifolds from high-dimensional flow-field data. To this end, the Kullback--Leibler (KL) divergence in the variational objective is decomposed into three complementary information-theoretic terms: the index-code mutual information, the total correlation, and the dimension-wise KL divergence. These terms explicitly regulate data compression, latent disentanglement, and geometric regularization. This establishes a principled basis for targeted latent-space design, allowing enhanced interpretability without sacrificing information capacity, a common drawback of heavily regularized VAE variants. The approach is evaluated on two synthetic unsteady flow datasets. First, we consider a flow around a cylinder in a channel with variable cylinder position, diameter, and Reynolds number. Later, we also consider the flow around a NACA 0012 airfoil at varying angles of attack and subjected to strong vortex gusts with variable intensity, position, and length scale. Comparisons with Principal Component Analysis, Isometric Feature Mapping, and $\beta$-VAE demonstrate clear advantages in disentanglement and physical interpretability. The learned latent coordinates successfully separate distinct physical effects. Moreover, the proposed method demonstrates strong robustness to variations in the loss-weighting parameters, despite involving a larger number of such parameters.

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

2 major / 0 minor

Summary. The paper introduces an information-theoretic decomposition of the KL divergence term in the VAE objective into three additive components—index-code mutual information, total correlation, and dimension-wise KL divergence—to separately regulate data compression, latent disentanglement, and geometric regularization. This is claimed to enable targeted latent-space design for physically interpretable manifolds from high-dimensional fluid flow data without the capacity loss typical of heavily regularized VAEs. The method is evaluated on two synthetic unsteady flow datasets (cylinder flow with variable position/diameter/Re and NACA 0012 airfoil with variable gust parameters), showing advantages over PCA, Isomap, and β-VAE in disentanglement and physical interpretability of the learned coordinates, along with robustness to the larger set of loss-weighting parameters.

Significance. If the three-term decomposition delivers the claimed independent control over compression, disentanglement, and regularization without new optimization-induced couplings, the work would provide a useful principled tool for interpretable reduced-order modeling in fluid dynamics. The use of synthetic datasets with explicitly varied physical parameters (Re, angle of attack, gust intensity) is a strength that supports direct validation of whether latent coordinates align with known physics. The approach rests on standard information-theoretic identities rather than ad-hoc inventions, which strengthens its foundation.

major comments (2)
  1. [Abstract] Abstract: the central claim that the decomposition 'establishes a principled basis for targeted latent-space design, allowing enhanced interpretability without sacrificing information capacity' is not supported by any reported quantitative metrics (e.g., reconstruction MSE, disentanglement scores such as MIG or SAP, or latent-space geometry measures), error bars, or statistical tests comparing the proposed method to β-VAE. Without these, it is impossible to verify that varying one loss weight leaves the other two effects (fidelity and capacity) statistically unchanged, especially given the coupled physical effects in the datasets (Re coupled to vortex shedding, gust parameters coupled to airfoil loading).
  2. [Abstract (and implied evaluation)] The manuscript does not demonstrate that the additive decomposition yields independent control in practice for fluid manifolds. Although the terms are mathematically additive, the shared encoder/decoder parameters and the structured correlations in unsteady flow data can induce indirect dependencies during joint optimization; no ablation is described that isolates the effect of each weight on reconstruction fidelity and latent geometry while holding the others fixed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and the positive assessment of the work's significance for interpretable reduced-order modeling. We respond point by point to the major comments, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the decomposition 'establishes a principled basis for targeted latent-space design, allowing enhanced interpretability without sacrificing information capacity' is not supported by any reported quantitative metrics (e.g., reconstruction MSE, disentanglement scores such as MIG or SAP, or latent-space geometry measures), error bars, or statistical tests comparing the proposed method to β-VAE. Without these, it is impossible to verify that varying one loss weight leaves the other two effects (fidelity and capacity) statistically unchanged, especially given the coupled physical effects in the datasets (Re coupled to vortex shedding, gust parameters coupled to airfoil loading).

    Authors: We acknowledge that the abstract claim would be more robust with explicit quantitative support. The manuscript demonstrates the advantages through qualitative evidence: the learned latent coordinates align with distinct physical parameters (e.g., separating Reynolds number effects from geometric variations in the cylinder case and gust parameters in the airfoil case), and comparisons with β-VAE show improved disentanglement without apparent loss of reconstruction quality. However, we agree that metrics such as reconstruction MSE (with error bars across runs), standard disentanglement scores (MIG, SAP), and statistical comparisons would allow direct verification of independent effects. These will be added to the revised manuscript, including analysis of how individual weight variations affect fidelity and capacity. revision: yes

  2. Referee: [Abstract (and implied evaluation)] The manuscript does not demonstrate that the additive decomposition yields independent control in practice for fluid manifolds. Although the terms are mathematically additive, the shared encoder/decoder parameters and the structured correlations in unsteady flow data can induce indirect dependencies during joint optimization; no ablation is described that isolates the effect of each weight on reconstruction fidelity and latent geometry while holding the others fixed.

    Authors: The referee is correct that mathematical additivity alone does not guarantee practical independence under joint optimization with shared parameters and correlated flow data. The current results show robustness to the larger set of loss weights and successful separation of physical effects, which provides indirect support. To directly address this, the revised manuscript will include a dedicated ablation study: each loss weight will be varied individually while holding the others fixed, with reported effects on reconstruction fidelity (MSE) and latent geometry (e.g., inter-dimension correlations). This will quantify the degree of independent control achieved for the fluid datasets. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation uses standard information-theoretic identities

full rationale

The paper's central step is decomposing the KL term in the VAE ELBO into index-code mutual information, total correlation, and dimension-wise KL divergence. This is a direct algebraic identity from information theory that holds for any joint distribution and does not depend on the fluid-flow data, the claimed interpretability gains, or any fitted parameters. The abstract and description present the decomposition as enabling 'targeted latent-space design' and 'enhanced interpretability without sacrificing capacity,' but these are interpretive claims supported by subsequent empirical comparisons (PCA, Isomap, β-VAE) rather than by construction. No equations reduce the performance assertions to self-referential definitions, no predictions are statistically forced by prior fits, and no load-bearing uniqueness theorems or ansatzes are imported via self-citation. The derivation chain is therefore self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on standard variational inference and information theory identities; the only notable free parameters are the multiple loss weights whose robustness is claimed but not derived.

free parameters (1)
  • loss-weighting parameters
    Multiple scalar weights control the relative strength of the three decomposed KL terms; the paper notes a larger number than typical VAEs yet reports robustness.
axioms (1)
  • standard math KL divergence decomposes additively into index-code mutual information, total correlation, and dimension-wise KL divergence.
    This identity is invoked to enable separate regulation of the three information terms in the variational objective.

pith-pipeline@v0.9.0 · 5545 in / 1366 out tokens · 42623 ms · 2026-05-10T03:58:46.431415+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

44 extracted references · 44 canonical work pages · 2 internal anchors

  1. [1]

    Annual Review of Fluid Mechanics49(Volume 49, 2017), 387–417 (2017)

    Rowley, C.W., Dawson, S.T.M.: Model reduction for flow analysis and control. Annual Review of Fluid Mechanics49(Volume 49, 2017), 387–417 (2017)

    work page 2017

  2. [2]

    The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science6(2), 559 (1901)

    Pearson, K.: Principal components analysis. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science6(2), 559 (1901)

    work page 1901

  3. [3]

    Psychometrika1(1), 27–35 (1936)

    Hotelling, H.: Simplified calculation of principal components. Psychometrika1(1), 27–35 (1936)

    work page 1936

  4. [4]

    Mathematis- che Annalen64(2), 161–174 (1907)

    Schmidt, E.: Zur theorie der linearen und nicht linearen integralgleichungen zweite abhandlung: Aufl¨ osung der allgemeinen linearen integralgleichung. Mathematis- che Annalen64(2), 161–174 (1907)

    work page 1907

  5. [5]

    Journal of Applied Meteorology and Climatology6(5), 791–802 (1967)

    Kutzbach, J.E.: Empirical eigenvectors of sea-level pressure, surface temperature and precipitation complexes over north america. Journal of Applied Meteorology and Climatology6(5), 791–802 (1967)

    work page 1967

  6. [6]

    Journal of computational Chemistry23(13), 1236–1243 (1967)

    Lumley, J., Yaglom, A., Tatarski, V.: Atmospheric turbulence and radio wave propagation. Journal of computational Chemistry23(13), 1236–1243 (1967)

    work page 1967

  7. [7]

    Annual Review of Fluid Mechanics25(Volume 25, 1993), 539–575 (1993)

    Berkooz, G., Holmes, P., Lumley, J.L.: The proper orthogonal decomposition in the analysis of turbulent flows. Annual Review of Fluid Mechanics25(Volume 25, 1993), 539–575 (1993)

    work page 1993

  8. [8]

    In: International Conference on Artificial Neural Networks, pp

    Sch¨ olkopf, B., Smola, A., M¨ uller, K.-R.: Kernel principal component analysis. In: International Conference on Artificial Neural Networks, pp. 583–588 (1997). Springer

    work page 1997

  9. [9]

    Philosophical transactions of the royal society A: Mathematical, Physical and Engineering Sciences374(2065), 20150202 (2016)

    Jolliffe, I.T., Cadima, J.: Principal component analysis: a review and recent developments. Philosophical transactions of the royal society A: Mathematical, Physical and Engineering Sciences374(2065), 20150202 (2016)

    work page 2065

  10. [10]

    Measurement Science and Technology34(4), 042001 (2023)

    Mendez, M.A.: Linear and nonlinear dimensionality reduction from fluid mechan- ics to machine learning. Measurement Science and Technology34(4), 042001 (2023)

    work page 2023

  11. [11]

    In: Proceedings A, vol

    Tirelli, I., Mendez, M.A., Ianiro, A., Discetti, S.: A meshless method to compute the proper orthogonal decomposition and its variants from scattered data. In: Proceedings A, vol. 481, p. 20240526 (2025). The Royal Society

    work page 2025

  12. [12]

    Neural networks2(1), 53–58 (1989)

    Baldi, P., Hornik, K.: Neural networks and principal component analysis: Learning from examples without local minima. Neural networks2(1), 53–58 (1989)

    work page 1989

  13. [13]

    Experimental Thermal and Fluid Science78, 354–369 (2016) 28

    Raiola, M., Ianiro, A., Discetti, S.: Wake of tandem cylinders near a wall. Experimental Thermal and Fluid Science78, 354–369 (2016) 28

    work page 2016

  14. [14]

    science290(5500), 2319–2323 (2000)

    Tenenbaum, J.B., Silva, V.d., Langford, J.C.: A global geometric framework for nonlinear dimensionality reduction. science290(5500), 2319–2323 (2000)

    work page 2000

  15. [15]

    Journal of Fluid Mechanics955, 34 (2023)

    Farzamnik, E., Ianiro, A., Discetti, S., Deng, N., Oberleithner, K., Noack, B.R., Guerrero, V.: From snapshots to manifolds – a tale of shear flows. Journal of Fluid Mechanics955, 34 (2023)

    work page 2023

  16. [16]

    Journal of Fluid Mechanics996, 26 (2024)

    Marra, L., Cornejo Maceda, G.Y., Meil´ an-Vila, A., Guerrero, V., Rashwan, S., Noack, B.R., Discetti, S., Ianiro, A.: Actuation manifold from snapshot data. Journal of Fluid Mechanics996, 26 (2024)

    work page 2024

  17. [17]

    Annual Review of Fluid Mechanics52(Volume 52, 2020), 477–508 (2020)

    Brunton, S.L., Noack, B.R., Koumoutsakos, P.: Machine learning for fluid mechanics. Annual Review of Fluid Mechanics52(Volume 52, 2020), 477–508 (2020)

    work page 2020

  18. [18]

    Progress in Aerospace Sciences157, 101130 (2025)

    Zhang, W., Xia, M., Kou, J.: A scientometric investigation of artificial intelligence for fluid mechanics: Emerging topics and active groups. Progress in Aerospace Sciences157, 101130 (2025)

    work page 2025

  19. [19]

    Journal of Computational Physics182(1), 1–26 (2002)

    Milano, M., Koumoutsakos, P.: Neural network modeling for near wall turbulent flow. Journal of Computational Physics182(1), 1–26 (2002)

    work page 2002

  20. [20]

    Journal of Computational Physics 404, 108973 (2020)

    Lee, K., Carlberg, K.T.: Model reduction of dynamical systems on nonlinear man- ifolds using deep convolutional autoencoders. Journal of Computational Physics 404, 108973 (2020)

    work page 2020

  21. [21]

    Archives of computational methods in engineering, 1–20 (2025)

    Mienye, I.D., Swart, T.G.: Deep autoencoder neural networks: A comprehensive review and new perspectives. Archives of computational methods in engineering, 1–20 (2025)

    work page 2025

  22. [22]

    Nature Communications14(1), 6480 (2023)

    Fukami, K., Taira, K.: Grasping extreme aerodynamics on a low-dimensional manifold. Nature Communications14(1), 6480 (2023)

    work page 2023

  23. [23]

    Journal of Fluid Mechanics1010, 4 (2025)

    Fukami, K., Taira, K.: Observable-augmented manifold learning for multi-source turbulent flow data. Journal of Fluid Mechanics1010, 4 (2025)

    work page 2025

  24. [24]

    Journal of Fluid Mechanics1021, 39 (2025)

    Fukami, K., Iwatani, Y., Maejima, S., Asada, H., Kawai, S.: Compact representa- tion of transonic airfoil buffet flows with observable-augmented machine learning. Journal of Fluid Mechanics1021, 39 (2025)

    work page 2025

  25. [25]

    Auto-Encoding Variational Bayes

    Kingma, D.P., Welling, M.: Auto-encoding variational bayes. In: International Conference on Learning Representations (ICLR) (2014). arXiv:1312.6114

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  26. [26]

    Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M., Mohamed, S., Lerchner, A.:β-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework, (2017)

    work page 2017

  27. [27]

    Nature Communications15(1), 1361 (2024)

    Solera-Rico, A., Sanmiguel Vila, C., G´ omez-L´ opez, M., Wang, Y., Almashjary, 29 A., Dawson, S.T., Vinuesa, R.:β-variational autoencoders and transformers for reduced-order modelling of fluid flows. Nature Communications15(1), 1361 (2024)

    work page 2024

  28. [28]

    Physics of Fluids34(7) (2022)

    Kang, Y.-E., Yang, S., Yee, K.: Physics-aware reduced-order modeling of transonic flow viaβ-variational autoencoder. Physics of Fluids34(7) (2022)

    work page 2022

  29. [29]

    Physics of Fluids33(8) (2021)

    Wang, J., He, C., Li, R., Chen, H., Zhai, C., Zhang, M.: Flow field prediction of supercritical airfoils via variational autoencoder based deep learning framework. Physics of Fluids33(8) (2021)

    work page 2021

  30. [30]

    AIAA Journal60(11), 6168–6182 (2022)

    Li, R., Zhang, Y., Chen, H.: Physically interpretable feature learning of supercrit- ical airfoils based on variational autoencoders. AIAA Journal60(11), 6168–6182 (2022)

    work page 2022

  31. [31]

    Journal of Fluid Mechanics1013, 41 (2025)

    Mousavi, H., Eldredge, J.D.: Low-order flow reconstruction and uncertainty quan- tification in disturbed aerodynamics using sparse pressure measurements. Journal of Fluid Mechanics1013, 41 (2025)

    work page 2025

  32. [32]

    Lucas, J., Tucker, G., Grosse, R., Norouzi, M.: Don’t blame the elbo! a linear VAE perspective on posterior collapse (2019)

    work page 2019

  33. [33]

    In: Dy, J., Krause, A

    Kim, H., Mnih, A.: Disentangling by factorising. In: Dy, J., Krause, A. (eds.) Pro- ceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 80, pp. 2649–2658. PMLR, ??? (2018)

    work page 2018

  34. [34]

    Chen, R.T.Q., Li, X., Grosse, R., Duvenaud, D.: Isolating sources of disentangle- ment in variational autoencoders31(2018)

    work page 2018

  35. [35]

    Sønderby, C.K., Raiko, T., Maaløe, L., Sønderby, S.K., Winther, O.: Ladder variational autoencoders29(2016)

    work page 2016

  36. [36]

    In: International Conference on Algorithmic Learning Theory, pp

    Gretton, A., Bousquet, O., Smola, A., Sch¨ olkopf, B.: Measuring statistical depen- dence with hilbert-schmidt norms. In: International Conference on Algorithmic Learning Theory, pp. 63–77 (2005). Springer

    work page 2005

  37. [37]

    Gaussian Error Linear Units (GELUs)

    Hendrycks, D.: Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415 (2016)

    work page internal anchor Pith review arXiv 2016

  38. [38]

    Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., Battaglia, P.W.: Learning mesh- based simulation with graph networks (2021)

    work page 2021

  39. [39]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(7), 3614–3633 (2022) 30

    Vandenhende, S., Georgoulis, S., Van Gansbeke, W., Proesmans, M., Dai, D., Van Gool, L.: Multi-task learning for dense prediction tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence44(7), 3614–3633 (2022) 30

    work page 2022

  40. [40]

    Machine Vision and Applications36, 104 (2025)

    Mogultay Ozcan, H., Kalkan, S., Yarman-Vural, F.T.: L-vae: variational auto- encoder with learnable beta for disentangled representation. Machine Vision and Applications36, 104 (2025)

    work page 2025

  41. [41]

    Journal of Fluid Mechanics1000, 95 (2024)

    Arranz, G., Lozano-Dur´ an, A.: Informative and non-informative decomposition of turbulent flow fields. Journal of Fluid Mechanics1000, 95 (2024)

    work page 2024

  42. [42]

    Physics of Fluids36(11), 117139 (2024)

    Franc´ es-Belda, V., Solera-Rico, A., Nieto-Centenero, J., Andr´ es, E., San- miguel Vila, C., Castellanos, R.: Toward aerodynamic surrogate modeling based onβ-variational autoencoders. Physics of Fluids36(11), 117139 (2024)

    work page 2024

  43. [43]

    Chinese Journal of Aeronautics37(3), 1–6 (2024)

    Zhang, W., Xuhao, P., Jiaqing, K., Xu, W.: Heterogeneous data-driven aero- dynamic modeling based on physical feature embedding. Chinese Journal of Aeronautics37(3), 1–6 (2024)

    work page 2024

  44. [44]

    AIAA Journal (2025) 31

    Wang, X., Huang, R., Kou, J., Tang, H., Zhang, W.: Goal-oriented feature extrac- tion: A novel approach to enhance data-driven surrogate models. AIAA Journal (2025) 31

    work page 2025