pith. sign in

arxiv: 2605.30112 · v1 · pith:SEINM7NBnew · submitted 2026-05-28 · 💻 cs.LG

Striding Across Reynolds Numbers: Representation Geometry in Neural PDE Generalisation

Jianing Shi This is my paper

Pith reviewed 2026-06-29 08:50 UTC · model grok-4.3

classification 💻 cs.LG
keywords cross-Reynolds generalisationrepresentation geometryneural PDE solversconvolutional autoencoderlatent space matchingNavier-Stokeszero-shot transferFourier Neural Operator
0
0 comments X

The pith

Matching states in a source-trained autoencoder latent space enables cross-Reynolds generalisation in PDE solvers using only source data.

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

The paper examines why neural PDE solvers struggle or succeed when Reynolds number changes by a factor of ten. On the forced 2D Navier-Stokes benchmark, a trained Fourier Neural Operator reaches 46.68 percent relative L2 error, yet simple retrieval from a source database already improves to 41-42 percent. The authors introduce ConvAE-Relay, which encodes states with a convolutional autoencoder trained only on the source regime, finds nearest matches in that latent space, and applies the corresponding source dynamics. This yields 38.34 percent error without any target-regime training or data. Ablations and oracle checks indicate that latent-space matching quality drives the gain and that source dynamics directions stay usable when matches remain on-manifold.

Core claim

ConvAE-Relay matches states in a source-trained convolutional autoencoder latent space and borrows dynamics from a source-regime database, achieving 38.34+/-0.07 percent relative L2 error on target-regime queries. This uses only the source-regime database with no target-regime fitting, labels, or entries. Oracle experiments show source dynamics directions remain transferable with cosine similarity around 0.84 when matches stay on-manifold, while autoregressive drift accounts for about 12 percentage points of error. A U-Net with multi-scale skip connections reaches 34.72+/-0.60 percent, supporting that local multi-scale representations organise the transfer.

What carries the argument

ConvAE-Relay: matching query states to source states inside the latent space of a convolutional autoencoder trained on the source regime, then retrieving and applying the associated source dynamics.

If this is right

  • Matching quality in the latent space dominates the choice of update rule for cross-regime performance.
  • Source-regime dynamics directions transfer with cosine similarity near 0.84 provided matches stay on-manifold.
  • Autoregressive rollout drift forms the primary remaining error source, contributing roughly 12 percentage points.
  • Multi-scale local representations, as used in U-Nets with skip connections, support improved cross-Reynolds transfer.

Where Pith is reading between the lines

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

  • The same latent-space retrieval could be tested on other parameter shifts if the on-manifold condition holds for those shifts.
  • Pre-training the autoencoder on the most accessible regime might allow zero-shot application to less accessible regimes without retraining.
  • Increasing the density of the source database could reduce the impact of autoregressive drift while still avoiding any target data.

Load-bearing premise

Target regime states remain close enough to the source manifold that nearest-neighbor matches in the autoencoder latent space select dynamics that remain useful after the tenfold Reynolds shift.

What would settle it

Target states under the 10x Reynolds shift fall sufficiently far off the source manifold that nearest matches in latent space produce dynamics whose error exceeds the 41-42 percent retrieval baseline.

Figures

Figures reproduced from arXiv: 2605.30112 by Jianing Shi.

Figure 1
Figure 1. Figure 1: Cross-Reynolds diagnostic summary. (a) Vorticity predictions and absolute error maps at Re=10,000 (rollout step 5, median￾error trajectory). FNO produces visibly smoother fields; ConvAE-Relay and U-Net preserve finer structure. (b) 2×2 representation ablation (seed 42): upgrading matching space from PCA to ConvAE latent improves the best result by 2.37 pp, while the update-rule effect is ≤1.1 pp; matching … view at source ↗
Figure 2
Figure 2. Figure 2: Per-step dynamics cosine similarity ct = ⟨∆dωt, ∆ωt⟩/(∥∆dωt∥2∥∆ωt∥2) between borrowed and actual dy￾namics. Under oracle matching, similarity remains stable at ∼0.84; under standard relay, it collapses by step 10. source-regime dynamics and actual target-regime dynamics remains stable at ∼0.84 across all 10 steps. Under standard relay, this similarity collapses from 0.84 to −0.07 by step 10 ( [PITH_FULL_I… view at source ↗
Figure 4
Figure 4. Figure 4: Wavenumber-resolved spectral relative error at rollout step 10. All methods show higher error at low wavenumbers (large￾scale dynamics); U-Net achieves the lowest error across most scales. Shaded bands denote 95% bootstrap confidence intervals. 4.5. Boundary Characterisation Re=100,000. At Re=100,000, U-Net remains the strongest tested method (58.85 ± 0.32%), outperforming FNO (63.08%) by 4 pp. However, Co… view at source ↗
Figure 5
Figure 5. Figure 5: Cross-regime comparison at Re=10,000 and Re=100,000. At 10× shift, ConvAE-Relay outperforms FNO; at 100× shift, this advantage reverses. U-Net retains its advantage at both regimes. Error bars denote 95% bootstrap confidence inter￾vals. ConvAE-Relay (38.34%) despite U-Net’s stronger predic￾tive encoder. A representation can be predictively sufficient but not metrically indexable: the latent geometry that s… view at source ↗
Figure 6
Figure 6. Figure 6: Relay scaling behaviour. (a) Database size ablation: approximately 75% of improvement is achieved by the first 500 source trajectories, with diminishing returns beyond. (b) Long-horizon degradation: by T=20, ConvAE-Relay converges toward persistence, consistent with drift accumulation eroding matching quality over extended rollouts. Error bars denote 95% bootstrap confidence intervals. 10 epoch 50 epoch 0 … view at source ↗
Figure 7
Figure 7. Figure 7: U-Net epoch saturation (seed 42). OOD performance at Re=10,000 saturates by 10 epochs, while in-distribution error continues improving to 50 epochs. Further source-regime training does not reduce OOD error. 0 5 10 15 20 25 30 Radial wavenumber k 0.0 0.2 0.4 0.6 0.8 1.0 Enstrophy ratio Zpred/Ztrue FNO PCA relay ConvAE-Relay U-Net [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Enstrophy spectrum ratio Zpred(k)/Ztrue(k) at rollout step 10. Values above 1 indicate spectral energy amplification; below 1 indicates damping. The dashed line marks perfect reconstruction (ratio=1). 12 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
read the original abstract

Cross-Reynolds generalisation in neural PDE solvers remains poorly characterised. On the canonical forced 2D Navier-Stokes benchmark, a trained Fourier Neural Operator reaches 46.68% relative L2 error under a 10x Reynolds-number shift, yet zero-forward-model retrieval baselines already improve to 41-42%. This suggests representation geometry as a major organising variable among the tested methods. We test this hypothesis through ConvAE-Relay, which matches states in a source-trained convolutional autoencoder latent space and borrows dynamics from a source-regime database, achieving 38.34+/-0.07% using only a source-regime database and no target-regime fitting, labels, or database entries. A 2x2 ablation isolates matching quality as dominant over the update rule. Oracle experiments confirm that source-regime dynamics directions remain transferable (cosine similarity ~0.84) when matching stays on-manifold; autoregressive drift is the primary bottleneck (~12 percentage points). From the learned-prediction side, a U-Net with multi-scale skip connections achieves 34.72+/-0.60%, consistent with the retrieval-side finding that local, multi-scale representations organise cross-Reynolds transfer among tested methods. All claims are scoped to this benchmark.

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 / 0 minor

Summary. The paper examines cross-Reynolds generalization for neural PDE solvers on the forced 2D Navier-Stokes benchmark. A Fourier Neural Operator reaches 46.68% relative L2 error under a 10x Re shift, while retrieval baselines improve to 41-42%. ConvAE-Relay matches states in a source-trained ConvAE latent space and borrows source dynamics, achieving 38.34+/-0.07% with no target fitting, labels, or data. A 2x2 ablation isolates matching quality as dominant; oracle cosine similarity is ~0.84 when on-manifold. Autoregressive drift costs ~12 points. A U-Net with multi-scale skips reaches 34.72+/-0.60%. All claims are scoped to this benchmark.

Significance. If the results hold, the work shows that representation geometry (local, multi-scale) organizes cross-Re transfer among tested methods and supplies a concrete no-target-data baseline (ConvAE-Relay) that beats the FNO while using only source data. The numerical claims are supported by reported means with standard deviations, a 2x2 ablation, and oracle measurements; these elements strengthen the empirical grounding.

major comments (1)
  1. [Abstract] Abstract: the reported 38.34% performance of ConvAE-Relay and its advantage over the 41-42% retrieval baselines rest on target-regime states remaining sufficiently on the source manifold in the ConvAE latent space under the 10x Re shift. No direct verification (reconstruction error, latent-space density, or distance statistics for target queries) is supplied; the oracle cosine-similarity result (~0.84) is explicitly conditional on on-manifold matching. This assumption is load-bearing for the central performance claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful review and for highlighting the importance of verifying the manifold assumption underlying the ConvAE-Relay results. We address the single major comment below and will strengthen the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the reported 38.34% performance of ConvAE-Relay and its advantage over the 41-42% retrieval baselines rest on target-regime states remaining sufficiently on the source manifold in the ConvAE latent space under the 10x Re shift. No direct verification (reconstruction error, latent-space density, or distance statistics for target queries) is supplied; the oracle cosine-similarity result (~0.84) is explicitly conditional on on-manifold matching. This assumption is load-bearing for the central performance claim.

    Authors: We agree that direct verification of how well target-regime states align with the source manifold would strengthen the central claim. The reported oracle cosine similarity (~0.84) is indeed conditional on on-manifold matches, and while the performance gap versus retrieval baselines provides indirect support, it does not substitute for explicit checks. In the revised manuscript we will add: (i) reconstruction error statistics for target queries passed through the source-trained ConvAE, (ii) latent-space distance histograms comparing target queries to the source database, and (iii) a brief density comparison (e.g., nearest-neighbor distances) between source and target latent points. These additions will be placed in the methods/results section and referenced from the abstract. revision: yes

Circularity Check

0 steps flagged

No circularity: all performance claims are direct empirical measurements on held-out target data

full rationale

The paper reports empirical error rates (e.g., ConvAE-Relay at 38.34+/-0.07%, U-Net at 34.72+/-0.60%) obtained by training on source-regime data and evaluating on held-out target-regime trajectories under a 10x Reynolds shift. No equations, fitted parameters, or self-citations are used to derive these numbers; the results are measured quantities. The on-manifold assumption is stated as a scope condition for the oracle cosine-similarity check but is not used to algebraically reduce any reported performance figure. No self-definitional loops, fitted-input predictions, or load-bearing self-citations appear in the derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on empirical performance comparisons on a single forced 2D Navier-Stokes benchmark; no new physical constants, particles, or mathematical entities are postulated.

axioms (1)
  • domain assumption The forced 2D Navier-Stokes equations define the benchmark dynamics under varying Reynolds number.
    The paper selects this PDE system as the canonical test case for cross-Re generalization.

pith-pipeline@v0.9.1-grok · 5746 in / 1472 out tokens · 36518 ms · 2026-06-29T08:50:53.326787+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

30 extracted references · 12 canonical work pages · 2 internal anchors

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    and Zawadzki, I

    Atencia, A. and Zawadzki, I. A comparison of two techniques for generating nowcasting ensembles. part II : Analogs selection and comparison of techniques. Monthly Weather Review, 143 0 (7): 0 2890--2908, 2015. doi:10.1175/MWR-D-14-00342.1. URL https://doi.org/10.1175/MWR-D-14-00342.1

    work page doi:10.1175/mwr-d-14-00342.1 2015

  3. [3]

    Representation learning: A review and new perspectives.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8): 1798–1828, 2013

    Bengio, Y., Courville, A., and Vincent, P. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35 0 (8): 0 1798--1828, 2013. doi:10.1109/TPAMI.2013.50. URL https://doi.org/10.1109/TPAMI.2013.50

    work page doi:10.1109/tpami.2013.50 2013

  4. [4]

    E., and Welling, M

    Brandstetter, J., Worrall, D. E., and Welling, M. Message passing neural PDE solvers. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=vSix3HPYKSU

    2022

  5. [5]

    L., Noack, B

    Brunton, S. L., Noack, B. R., and Koumoutsakos, P. Machine learning for fluid mechanics. Annual Review of Fluid Mechanics, 52: 0 477--508, 2020. doi:10.1146/annurev-fluid-010719-060214. URL https://doi.org/10.1146/annurev-fluid-010719-060214

    work page doi:10.1146/annurev-fluid-010719-060214 2020

  6. [6]

    Choose a transformer: Fourier or Galerkin

    Cao, S. Choose a transformer: Fourier or Galerkin . In Advances in Neural Information Processing Systems, volume 34, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/d0921d442ee91b896ad95059d13df618-Abstract.html

    2021

  7. [7]

    Turbulence modeling in the age of data

    Duraisamy, K., Iaccarino, G., and Xiao, H. Turbulence modeling in the age of data. Annual Review of Fluid Mechanics, 51: 0 357--377, 2019. doi:10.1146/annurev-fluid-010518-040547. URL https://doi.org/10.1146/annurev-fluid-010518-040547

    work page doi:10.1146/annurev-fluid-010518-040547 2019

  8. [8]

    Multiwavelet-based operator learning for differential equations

    Gupta, G., Xiao, X., and Bogdan, P. Multiwavelet-based operator learning for differential equations. In Advances in Neural Information Processing Systems, volume 34, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/c9e5c2b59d98488fe1070e744041ea0e-Abstract.html

    2021

  9. [9]

    GNOT : A general neural operator transformer for operator learning

    Hao, Z., Wang, Z., Su, H., Ying, C., Dong, Y., Liu, S., Cheng, Z., Song, J., and Zhu, J. GNOT : A general neural operator transformer for operator learning. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp.\ 12556--12569. PMLR, 2023. URL https://proceedings.mlr.press/v202/h...

    2023

  10. [10]

    Physics-informed machine learning,

    Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., and Yang, L. Physics-informed machine learning. Nature Reviews Physics, 3: 0 422--440, 2021. doi:10.1038/s42254-021-00314-5. URL https://doi.org/10.1038/s42254-021-00314-5

    work page doi:10.1038/s42254-021-00314-5 2021

  11. [11]

    Generalization through memorization: Nearest neighbor language models

    Khandelwal, U., Levy, O., Jurafsky, D., Zettlemoyer, L., and Lewis, M. Generalization through memorization: Nearest neighbor language models. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=HklBjCEKvH

    2020

  12. [12]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. URL https://arxiv.org/abs/1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  13. [13]

    B., Li, Z., Liu, B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A

    Kovachki, N. B., Li, Z., Liu, B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A. M., and Anandkumar, A. Neural operator: Learning maps between function spaces with applications to PDEs . Journal of Machine Learning Research, 24 0 (89): 0 1--97, 2023. URL https://jmlr.org/papers/v24/21-1524.html

    2023

  14. [14]

    Kraichnan, R. H. Inertial ranges in two-dimensional turbulence. Physics of Fluids, 10 0 (7): 0 1417--1423, 1967. doi:10.1063/1.1762301. URL https://doi.org/10.1063/1.1762301

    work page doi:10.1063/1.1762301 1967

  15. [15]

    Learning skillful medium-range global weather forecasting.Science, 382(6677): 1416–1421, 2023

    Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., and Battaglia, P. Learning skillful medium-range global weather forecasting. Science, 382 0 (6677): 0 1416--1421, 2023. doi:10.1126/scie...

    work page doi:10.1126/science.adi2336 2023

  16. [16]

    B., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A

    Li, Z., Kovachki, N. B., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A. M., and Anandkumar, A. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=c8P9NQVtmnO

    2021

  17. [17]

    Z., Liu, B., and Anandkumar, A

    Li, Z., Huang, D. Z., Liu, B., and Anandkumar, A. Fourier neural operator with learned deformations for PDEs on general geometries. Journal of Machine Learning Research, 24 0 (388): 0 1--26, 2023. URL https://jmlr.org/papers/v24/23-0064.html

    2023

  18. [18]

    Scale-consistent learning with neural operators

    Li, Z., Lanthaler, S., Deng, C., Wang, Y., Azizzadenesheli, K., and Anandkumar, A. Scale-consistent learning with neural operators. In NeurIPS 2024 Workshop on Foundation Models for Science: Progress, Opportunities, and Challenges, 2024. URL https://neurips.cc/virtual/2024/105933

    2024

  19. [19]

    E., and Brandstetter, J

    Lippe, P., Veeling, B., Perdikaris, P., Turner, R. E., and Brandstetter, J. PDE -refiner: Achieving accurate long rollouts with neural PDE solvers. In Advances in Neural Information Processing Systems, volume 36, 2023. URL https://papers.nips.cc/paper_files/paper/2023/hash/d529b943af3dba734f8a7d49efcb6d09-Abstract-Conference.html

    2023

  20. [20]

    Lorenz, E. N. Atmospheric predictability as revealed by naturally occurring analogues. Journal of the Atmospheric Sciences, 26 0 (4): 0 636--646, 1969. URL https://journals.ametsoc.org/view/journals/atsc/26/4/1520-0469_1969_26_636_aparbn_2_0_co_2.xml

    1969

  21. [21]

    Lu, L., Jin, P., Pang, G., Zhang, Z., and Karniadakis, G. E. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, 3: 0 218--229, 2021. doi:10.1038/s42256-021-00302-5. URL https://doi.org/10.1038/s42256-021-00302-5

    work page doi:10.1038/s42256-021-00302-5 2021

  22. [22]

    FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

    Pathak, J., Subramanian, S., Harrington, P., Raja, S., Chattopadhyay, A., Mardani, M., Kurth, T., Hall, D., Li, Z., Azizzadenesheli, K., Hassanzadeh, P., Kashinath, K., and Anandkumar, A. FourCastNet : A global data-driven high-resolution weather model using adaptive Fourier neural operators. arXiv preprint arXiv:2202.11214, 2022. URL https://arxiv.org/ab...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  23. [23]

    Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A., and Battaglia, P. W. Learning mesh-based simulation with graph networks. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=roNqYL0_XP

    2021

  24. [24]

    Pope, S. B. Turbulent Flows. Cambridge University Press, 2000. URL https://www.cambridge.org/core/books/turbulent-flows/B6051EBC741C237DCAF9C517ED3E0539

    2000

  25. [25]

    A., Ross, Z

    Rahman, M. A., Ross, Z. E., and Azizzadenesheli, K. U-NO : U -shaped neural operators. Transactions on Machine Learning Research, 2023. URL https://openreview.net/forum?id=j3oQF9coJd

    2023

  26. [26]

    Raissi, M., Perdikaris, P., and Karniadakis, G. E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378: 0 686--707, 2019. doi:10.1016/j.jcp.2018.10.045. URL https://doi.org/10.1016/j.jcp.2018.10.045

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

  27. [27]

    U-Net: Convolutional networks for biomedical image segmentation

    Ronneberger, O., Fischer, P., and Brox, T. U-Net : Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, volume 9351 of Lecture Notes in Computer Science, pp.\ 234--241. Springer, 2015. doi: https://doi.org/10.1007/978-3-319-24574-4_28

    work page doi:10.1007/978-3-319-24574-4_28 2015

  28. [28]

    Sanchez-Gonzalez, A., Godwin, J., Pfaff, T., Ying, R., Leskovec, J., and Battaglia, P. W. Learning to simulate complex physics with graph networks. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pp.\ 8459--8468. PMLR, 2020. URL https://proceedings.mlr.press/v119/sanchez-gonz...

    2020

  29. [29]

    W., and Gholami, A

    Subramanian, S., Harrington, P., Keutzer, K., Bhimji, W., Morozov, D., Mahoney, M. W., and Gholami, A. Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. In Advances in Neural Information Processing Systems, volume 36, 2023. URL https://papers.nips.cc/paper_files/paper/2023/hash/e15790966a4a9d85d688635...

    2023

  30. [30]

    Tran, A., Mathews, A., Xie, L., and Ong, C. S. Factorized Fourier neural operators. In International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=tmIiMPl4IPa

    2023