The reviewed record of science sign in
Pith

arxiv: 2605.26416 · v1 · pith:ZD3PZIU4 · submitted 2026-05-26 · physics.comp-ph · nlin.CD· physics.flu-dyn

A Differentiable Programming Framework for Accurate and Stable Reduced-Order Modeling of Chaotic Flows

Anant Kumar , Oliver Morales , Rohit Deshmukh This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-01 17:01 UTCgrok-4.3pith:ZD3PZIU4record.jsonopen to challenge →

classification physics.comp-ph nlin.CDphysics.flu-dyn
keywords reduced-order modelingPOD-Galerkindifferentiable programmingchaotic flowsmodel stabilizationhybrid losslid-driven cavity
0
0 comments X

The pith

Tuning POD-Galerkin tensors via differentiable programming stabilizes chaotic flow models with only 20 modes.

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

The paper shows how classical POD-Galerkin reduced-order models for chaotic flows demand many modes or extra stabilization terms to stay accurate over long times. The authors instead optimize the linear and quadratic tensors directly through differentiable programming, training on short trajectory segments. A hybrid loss that adds an energy-conservation penalty to the usual trajectory error proves essential for long-term behavior. Demonstrated on the chaotic lid-driven cavity at Re=30,000, the resulting model stays stable and accurate with far fewer modes than the untuned version. This produces a substantial drop in computational cost while preserving the projection structure of the original equations.

Core claim

Optimizing the linear and quadratic tensors of a POD-Galerkin reduced-order model by gradient descent on short-term trajectory data, using a hybrid loss that combines point-wise error with a conservation-of-energy constraint, produces long-term accurate and stable simulations of chaotic flows without additional closure models or higher mode counts, as verified on the lid-driven cavity at Reynolds number 30,000 where 20 modes suffice in place of 80.

What carries the argument

Differentiable programming applied to the linear and quadratic tensors of the POD-Galerkin system, optimized under a hybrid loss that merges trajectory matching with an energy-conservation term.

If this is right

  • The tuned model remains accurate and stable using only 20 modes for a flow where the classical POD-Galerkin method requires 80 modes.
  • Computational cost for long-time integration drops by roughly an order of magnitude relative to the untuned approach.
  • Short-term trajectory data alone, paired with the hybrid loss, suffices to produce long-term stability without extra closure terms.
  • The projection-based structure is preserved while the tensors are adjusted, avoiding the need to add new model terms.

Where Pith is reading between the lines

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

  • The same tensor-tuning procedure could be applied to other projection-based reduced-order models beyond standard POD-Galerkin.
  • Hybrid losses that embed conservation laws may address the closure problem more broadly in data-driven modeling of chaotic systems.
  • Online re-tuning of the tensors during a simulation could further extend accuracy when the flow regime slowly changes.

Load-bearing premise

That coefficients fitted to short trajectory segments with the hybrid loss will remain stable and accurate over much longer times in chaotic dynamics.

What would settle it

Integrate the tuned 20-mode model for times several times longer than the training window and check whether solution energy stays bounded while flow statistics continue to match those of the full-order simulation.

Figures

Figures reproduced from arXiv: 2605.26416 by Anant Kumar, Oliver Morales, Rohit Deshmukh.

Figure 1
Figure 1. Figure 1: Two-dimensional LDC Configuration measure if the learned systems are truly understanding underlying conservation laws and shifting towards a more stable system. 3.1.1 Galerkin Projection The uncalibrated Galerkin projection model (GP-ROM) acts as the baseline comparison in this study to evaluate the stabilizing efficacy of the proposed DPG-ROM framework. The baseline system tracks the evolution of the moda… view at source ↗
Figure 2
Figure 2. Figure 2: Calibration parameter sweep for GP-ROMs with [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Trajectory-based training and validation loss for DPG-ROMs with [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Short-term trajectory comparison of POD temporal coefficients for NeuralGP ROMs [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Long-term comparison of POD temporal coefficients for ROMs with [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Turbulent Kinetic Energy Evolution for DPG-ROMs with [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Hybrid training loss and trajectory based validation loss for DPG-ROMs with [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Normalized root mean squared error (NRMSE) in turbulent kinetic energy for different [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Short-term trajectory comparison of POD temporal coefficients for DPG-ROMs with [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Long-term comparison of POD temporal coefficients for ROMs with [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Turbulent Kinetic Energy Evolution for NeuralGP ROMs with [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Kinetic Energy PSD for DPG-ROMs with N = 20 and 30 modes (Hybrid loss). Eigenvalue Evolution [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Eigenvalue evolution in the complex plane for DPG-ROMs with [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Production and dissipation values for DPG-ROMs with [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
read the original abstract

Classical Proper Orthogonal Decomposition (POD)-based Galerkin projection models of chaotic flows typically require a large number of modes as well as stabilization or closure terms to achieve adequate accuracy and long-term stability. We present a novel differentiable programming framework that stabilizes low-rank POD-Galerkin models without increasing the number of modes or introducing additional closure terms, thereby delivering both high efficiency and high accuracy. Model stabilization is achieved by tuning the linear and quadratic tensors in the POD-Galerkin using differentiable programming, trained on short-term trajectory data. A key finding of this study is that a purely point-wise trajectory-based loss function yields poor long-term accuracy for chaotic systems. In contrast, a hybrid loss function that combines trajectory error with a physics-based conservation-of-energy term provides superior long-term performance. We demonstrate the approach on a chaotic lid-driven cavity flow at Re = 30,000. The stabilized ROM achieves an order-of-magnitude reduction in computational cost compared with the classical POD-Galerkin method: it remains accurate and stable with only 20 modes, whereas the classical ROM requires 80 POD modes.

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 / 2 minor

Summary. The manuscript introduces a differentiable programming framework to tune the linear and quadratic tensors of classical POD-Galerkin reduced-order models for chaotic flows. Stabilization is achieved solely by optimizing these tensors on short-term trajectory data using a hybrid loss that combines pointwise trajectory error with a physics-based energy-conservation term; no additional closure models or mode increases are introduced. The central demonstration is on a lid-driven cavity flow at Re=30,000, where the tuned 20-mode ROM is claimed to remain accurate and stable while the untuned classical ROM requires 80 modes, yielding an order-of-magnitude computational-cost reduction. A key reported finding is that a pure trajectory loss produces poor long-term behavior whereas the hybrid loss succeeds.

Significance. If the long-term attractor fidelity and stability claims hold under quantitative scrutiny, the work would provide a practical route to low-mode, closure-free ROMs for chaotic flows by leveraging differentiable optimization and a minimal physics-informed loss. The explicit contrast between pure trajectory and hybrid losses supplies a useful methodological insight for data-driven stabilization of Galerkin models. Reproducible code or parameter tables would further strengthen the contribution to computational physics.

major comments (3)
  1. [Abstract] Abstract and results section: the headline claim that the stabilized 20-mode ROM 'remains accurate and stable' with an order-of-magnitude cost reduction versus the classical 80-mode model is load-bearing, yet the abstract supplies no quantitative long-time error metrics (e.g., time-averaged kinetic-energy error, modal energy spectra, or Lyapunov exponents) or integration horizon beyond the training window; without these the robustness for chaotic dynamics cannot be assessed.
  2. [Method] Hybrid-loss formulation (method section): the energy-conservation term is presented as a global scalar constraint that replaces the need for closure; however, it is unclear whether this term preserves the correct modal energy cascade or merely enforces a single integral invariant, which is critical because the skeptic concern is that short-trajectory tuning may still permit slow drift from the true attractor once the solution leaves the training interval.
  3. [Results] Validation protocol (results section): the demonstration is confined to a single flow configuration (lid-driven cavity, Re=30,000) and a single training-trajectory length; the central claim that retuning the Galerkin tensors alone suffices for long-term fidelity would be strengthened by reporting sensitivity to initial-condition ensemble or training-window duration, as these directly test whether the optimized tensors generalize beyond the fitted short-time manifold.
minor comments (2)
  1. Provide explicit numerical values or ranges for the optimized linear and quadratic tensor entries (or at least their norms) so that the tuned model can be reproduced independently of the differentiable-programming pipeline.
  2. Clarify the relative weighting hyper-parameter between the trajectory and energy terms in the hybrid loss and report any ablation on its value.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed report. We address each major comment below with our responses and indicate planned revisions to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract and results section: the headline claim that the stabilized 20-mode ROM 'remains accurate and stable' with an order-of-magnitude cost reduction versus the classical 80-mode model is load-bearing, yet the abstract supplies no quantitative long-time error metrics (e.g., time-averaged kinetic-energy error, modal energy spectra, or Lyapunov exponents) or integration horizon beyond the training window; without these the robustness for chaotic dynamics cannot be assessed.

    Authors: We agree that the abstract would be improved by including quantitative long-time metrics. The results section already presents time-averaged kinetic-energy errors, modal energy spectra, and integration horizons beyond the training window. We will revise the abstract to explicitly reference these quantitative metrics and the validation horizon to better substantiate the claims regarding accuracy and stability for chaotic dynamics. revision: yes

  2. Referee: [Method] Hybrid-loss formulation (method section): the energy-conservation term is presented as a global scalar constraint that replaces the need for closure; however, it is unclear whether this term preserves the correct modal energy cascade or merely enforces a single integral invariant, which is critical because the skeptic concern is that short-trajectory tuning may still permit slow drift from the true attractor once the solution leaves the training interval.

    Authors: We will expand the method section with a derivation showing that the energy term arises from the inner product of the momentum equation with the velocity field. This global constraint acts on the sum of modal energies but, through the quadratic tensor couplings, influences the inter-modal transfers. We will add text explaining how this helps maintain the energy cascade and mitigates slow drift, referencing the long-term stable integrations shown in the results. revision: yes

  3. Referee: [Results] Validation protocol (results section): the demonstration is confined to a single flow configuration (lid-driven cavity, Re=30,000) and a single training-trajectory length; the central claim that retuning the Galerkin tensors alone suffices for long-term fidelity would be strengthened by reporting sensitivity to initial-condition ensemble or training-window duration, as these directly test whether the optimized tensors generalize beyond the fitted short-time manifold.

    Authors: We agree that sensitivity tests would strengthen the generalization claim. However, a full ensemble study over multiple initial conditions and training lengths would require new computations not present in the current work. We will revise the results section to discuss the rationale for the selected training trajectory and initial condition, and to note the method's performance on this challenging case as evidence of robustness, while indicating that broader sensitivity analysis is reserved for future extensions. revision: partial

Circularity Check

0 steps flagged

No significant circularity; data-driven tensor tuning validated empirically on external benchmark

full rationale

The paper optimizes the linear/quadratic Galerkin tensors via differentiable programming to minimize a hybrid loss (trajectory error + independent energy-conservation term) on short data, then reports long-term stability/accuracy on the Re=30,000 lid-driven cavity as a separate numerical outcome. This is a standard parameter-fitting procedure whose success is not guaranteed by construction and must be demonstrated against the full-order solver. No self-definitional equations, fitted quantities renamed as predictions, or load-bearing self-citations appear in the claims. The energy term is an external physics constraint, not derived from the fit. The 20-mode vs. 80-mode comparison is therefore an independent empirical result, not a tautology.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the validity of POD-Galerkin projection for the flow, the ability of short-trajectory data to inform long-term behavior when combined with an energy constraint, and the assumption that the learned tensors remain stable outside the training window. No new physical entities are introduced.

free parameters (1)
  • linear and quadratic tensors
    Coefficients in the reduced-order model that are adjusted via differentiable programming on trajectory data.
axioms (2)
  • domain assumption POD-Galerkin projection yields a valid reduced-order representation of the Navier-Stokes equations for the lid-driven cavity flow
    Standard modeling assumption invoked when the method is applied to the Re=30,000 case.
  • domain assumption A conservation-of-energy term can be meaningfully added to the loss without altering the underlying model structure
    Physics-based component of the hybrid loss described in the abstract.

pith-pipeline@v0.9.1-grok · 5733 in / 1576 out tokens · 60365 ms · 2026-07-01T17:01:30.191803+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

31 extracted references · 31 canonical work pages · 2 internal anchors

  1. [1]

    Balajewicz, Earl H

    Maciej J. Balajewicz, Earl H. Dowell, and Bernd R. Noack. Low-dimensional modelling of high-Reynolds-number shear flows incorporating constraints from the Navier–Stokes equa- tion.Journal of Fluid Mechanics, 729:285–308, 2013

    work page 2013

  2. [2]

    Benner, S

    P. Benner, S. Gugercin, and K. Willcox. A Survey of Projection-Based Model Reduction Methods for Parametric Dynamical Systems.SIAM Review, 57(4):483–531, 2015

    work page 2015

  3. [3]

    Bergmann.Optimisation A ´erodynamique par R ´eduction de Mod `ele POD et Contr ˆole Optimal: Application au Sillage Laminaire d’un Cylindre Circulaire

    M. Bergmann.Optimisation A ´erodynamique par R ´eduction de Mod `ele POD et Contr ˆole Optimal: Application au Sillage Laminaire d’un Cylindre Circulaire. PhD thesis, Institut National Polytechnique de Lorraine, 2004

    work page 2004

  4. [4]

    The proper orthogonal decomposition in the analysis of turbulent flows.Annual review of fluid mechanics, 25(1):539–575, 1993

    Gal Berkooz, Philip Holmes, and John L Lumley. The proper orthogonal decomposition in the analysis of turbulent flows.Annual review of fluid mechanics, 25(1):539–575, 1993

    work page 1993

  5. [5]

    Bourguet, M

    R. Bourguet, M. Braza, and A. Dervieux. Reduced-Order Modeling of Transonic Flows Around an Airfoil Submitted to Small Deformations.Journal of Computational Physics, 230(1):159–184, 2011

    work page 2011

  6. [6]

    J. L. Callaham, S. L. Brunton, and J.-C. Loiseau. On the Role of Nonlinear Correlations in Reduced-Order Modeling, 2021. arXiv preprint arXiv:2106.02409

    work page arXiv 2021

  7. [7]

    Mohan, Datta V

    Surya Chakrabarti, Arvind T. Mohan, Datta V . Gaitonde, and Daniel Livescu. Full trajectory optimizing operator inference for reduced-order modeling using differentiable programming, 2023

    work page 2023

  8. [8]

    Chinesta, A

    F. Chinesta, A. Huerta, G. Rozza, and K. Willcox. Model Order Reduction. InEncyclopedia of Computational Mechanics. Wiley, 2 edition, 2016

    work page 2016

  9. [9]

    Cordier, B

    L. Cordier, B. A. El Majd, and J. Favier. Calibration of POD Reduced-Order Models Using Tikhonov Regularization.International Journal for Numerical Methods in Fluids, 63(2):269– 296, 2010

    work page 2010

  10. [10]

    Couplet, C

    M. Couplet, C. Basdevant, and P. Sagaut. Calibrated Reduced-Order POD-Galerkin System for Fluid Flow Modelling.Journal of Computational Physics, 207(1):192–220, 2005

    work page 2005

  11. [11]

    Z. Deng, Y . Chen, Y . Liu, and K. C. Kim. Time-Resolved Turbulent Velocity Field Recon- struction Using a Long Short-Term Memory (LSTM)-Based Artificial Intelligence Frame- work.Physics of Fluids, 31(7):075108, 2019

    work page 2019

  12. [12]

    Deshmukh, J

    R. Deshmukh, J. J. McNamara, and J. Z. Kolter. Energy balance of reduced order models for unsteady flows using sparse coding.Nonlinear Dynamics, 94(1):197–210, 2018. 27

    work page 2018

  13. [13]

    McNamara, Zongxian Liang, J

    Rohit Deshmukh, Jack J. McNamara, Zongxian Liang, J. Zico Kolter, and Abhijit Gogulapati. Model order reduction using sparse coding exemplified for the lid-driven cavity.Journal of Fluid Mechanics, 808:189–223, 2016

    work page 2016

  14. [14]

    Gad-el Hak

    M. Gad-el Hak. Flow Control: The Future.Journal of Aircraft, 38(3):402–418, 2001

    work page 2001

  15. [15]

    Hochreiter and J

    S. Hochreiter and J. Schmidhuber. Long Short-Term Memory.Neural Computation, 9(8):1735–1780, 1997

    work page 1997

  16. [16]

    Holmes, J

    P. Holmes, J. L. Lumley, G. Berkooz, and C. W. Rowley.Turbulence, Coherent Structures, Dynamical Systems and Symmetry. Cambridge University Press, Cambridge, UK, 2nd edi- tion, 2012

    work page 2012

  17. [17]

    M. Ilak, S. Bagheri, L. Brandt, C. W. Rowley, and D. S. Henningson. Model reduction of the nonlinear complex ginzburg-landau equation.SIAM Journal on Applied Dynamical Systems, 9(4):1284–1302, 2010

    work page 2010

  18. [18]

    Kalashnikova, B

    I. Kalashnikova, B. van Bloemen Waanders, S. Arunajatesan, and M. Barone. Stabilization of projection-based reduced order models for linear time-invariant systems via optimization- based eigenvalue reassignment.Computer Methods in Applied Mechanics and Engineering, 272:251–270, 2014

    work page 2014

  19. [19]

    Lorenzi, A

    S. Lorenzi, A. Cammi, L. Luzzi, and G. Rozza. POD-Galerkin Method for Finite V olume Approximation of Navier–Stokes and RANS Equations.Computer Methods in Applied Me- chanics and Engineering, 311:151–179, 2016

    work page 2016

  20. [20]

    Physics-informed time- integrated deeponet: Temporal tangent space operator learning for high-accuracy inference

    Luis Mandl, Dibyajyoti Nayak, Tim Ricken, and Somdatta Goswami. Physics-informed time- integrated deeponet: Temporal tangent space operator learning for high-accuracy inference. Computer Methods in Applied Mechanics and Engineering, 455:118917, 2026

    work page 2026

  21. [21]

    Mittal, H

    R. Mittal, H. Dong, M. Bozkurttas, F. M. Najjar, A. Vargas, and A. von Loebbecke. A versatile sharp interface immersed boundary method for incompressible flows with complex boundaries.Journal of Computational Physics, 227(10):4825–4848, 2008

    work page 2008

  22. [22]

    Compressed Convolutional LSTM: An Efficient Deep Learning framework to Model High Fidelity 3D Turbulence

    A. Mohan, D. Daniel, M. Chertkov, and D. Livescu. Compressed Convolutional LSTM: An Efficient Deep Learning Framework to Model High Fidelity 3D Turbulence, 2019. arXiv preprint arXiv:1903.00033

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  23. [23]

    A. T. Mohan and D. V . Gaitonde. A Deep Learning Based Approach to Reduced Order Modeling for Turbulent Flow Control Using LSTM Neural Networks, 2018. arXiv preprint arXiv:1804.09269

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  24. [24]

    A. T. Mohan, K. Nagarajan, and D. Livescu. Learning Stable Galerkin Models of Turbulence with Differentiable Programming, 2021. arXiv preprint arXiv:2107.07559

    work page arXiv 2021

  25. [25]

    B. R. Noack, M. Morzynski, and G. Tadmor.Reduced-Order Modelling for Flow Control, volume 528. Springer Science & Business Media, 2011. 28

    work page 2011

  26. [26]

    Noack, P

    Bernd R. Noack, P. Papas, and Peter A. Monkewitz. The need for a pressure-term representa- tion in empirical Galerkin models of incompressible shear flows.Journal of Fluid Mechanics, 523:339–365, 2005

    work page 2005

  27. [27]

    D. Rempfer. On Low-Dimensional Galerkin Models for Fluid Flow.Theoretical and Com- putational Fluid Dynamics, 14(2):75–88, 2000

    work page 2000

  28. [28]

    C. W. Rowley. Model reduction for fluids, using balanced proper orthogonal decomposition. International Journal of Bifurcation and Chaos, 15(03):997–1013, 2005

    work page 2005

  29. [29]

    C. W. Rowley, T. Colonius, and R. M. Murray. Model Reduction for Compressible Flows Using POD and Galerkin Projection.Physica D: Nonlinear Phenomena, 189(1-2):115–129, 2004

    work page 2004

  30. [30]

    Nihar Sawant, Boris Kramer, and Benjamin Peherstorfer. Physics-informed regularization and structure preservation for learning stable reduced models from data with operator infer- ence.Computer Methods in Applied Mechanics and Engineering, 404:115836, 2023

    work page 2023

  31. [31]

    Turbulence and the dynamics of coherent structures

    Lawrence Sirovich. Turbulence and the dynamics of coherent structures. part i: Coherent structures.Quaterly of Applied Mathematics, XLV(3):561–571, 1987. A Supplementary Results 29

    work page 1987