arxiv: 2604.14424 · v1 · submitted 2026-04-15 · 💻 cs.LG · physics.flu-dyn

Recognition: unknown

Non-intrusive Learning of Physics-Informed Spatio-temporal Surrogate for Accelerating Design

Sudeepta Mondal , Soumalya Sarkar

Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:02 UTC · model grok-4.3

classification 💻 cs.LG physics.flu-dyn

keywords physics-informed surrogateKoopman autoencodersspatio-temporal modelingfluid flownon-intrusive learningcylinder wakegeneralization in dynamical systems

0 comments

The pith

A physics-informed surrogate learns fluid dynamics from limited data and predicts behavior under new conditions.

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

The paper introduces a framework to replace slow, high-fidelity simulations of nonlinear systems with fast machine-learning models that still respect the underlying physics. It does this by first using Koopman autoencoders to extract the system's evolution rules in a non-intrusive way from simulation snapshots, then training a separate surrogate to forecast how those rules change when operating conditions shift. The approach targets the common failure of purely data-driven surrogates to work outside their training set, using the cylinder flow problem as the test case. If successful, engineers could evaluate many more design options without repeating full simulations for every new parameter value.

Core claim

The PISTM framework learns the underlying spatio-temporal dynamics of nonlinear systems in a non-intrusive manner with Koopman autoencoders and couples this representation to a spatio-temporal surrogate model that predicts the behavior of the Koopman operator over a time window for previously unseen operating conditions, as demonstrated on two-dimensional incompressible flow around a cylinder.

What carries the argument

The PISTM framework, which embeds physics through Koopman autoencoders to extract dynamics and uses an additional surrogate to forecast operator evolution outside the training distribution.

If this is right

Designers can explore wider ranges of operating conditions for fluid systems without incurring the cost of new high-fidelity runs.
The same non-intrusive workflow can be applied to other nonlinear spatio-temporal problems once training snapshots are available.
Surrogate predictions become available orders of magnitude faster than the original simulations, enabling tighter integration into optimization loops.

Where Pith is reading between the lines

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

The method may scale to three-dimensional or multi-physics problems if the autoencoder can be trained on coarser grids or reduced-order snapshots.
Long-term stability of the predicted trajectories could be improved by adding explicit constraints on the Koopman eigenvalues during surrogate training.
Real-time monitoring or control applications become feasible if the surrogate inference time drops below the physical time step.

Load-bearing premise

That combining the Koopman embedding with the surrogate predictor will yield accurate forecasts for cylinder-flow conditions that lie outside the training data distribution.

What would settle it

Running the trained model on cylinder flow at Reynolds numbers far from the training range and measuring whether the predicted velocity fields deviate from high-fidelity simulation results by more than a few percent in norm.

Figures

Figures reproduced from arXiv: 2604.14424 by Soumalya Sarkar, Sudeepta Mondal.

**Figure 1.** Figure 1: Schematic overview of the proposed physics-informed spatio-temporal modeling (PISTM) framework of dynamical systems. challenging for practical engineering applications. To this end, researchers have looked into physics-informed approaches which are informed by the physics of dynamical systems in general, and not the underlying physical laws. Some of the notable works in this regard are the ones by Erichso… view at source ↗

**Figure 2.** Figure 2: Comparison of εE,εKE, εK on the 5 test conditions, along with a histogram of the 45 training conditions as a function of Re [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Qualitative comparison of the true data, Koopman predictions, emulated data, absolute Koopman emulation error, absolute Koopman error and absolute emulation error for the test case with Re = 172, at t = 1,5 and 9. case. Koopman autoencoders are learned for each of these training conditions, which predict the velocity fields for 10 timesteps in the future (t = T,T +1,...,T +k). For simplicity of notation, … view at source ↗

read the original abstract

Most practical engineering design problems involve nonlinear spatio-temporal dynamical systems. Multi-physics simulations are often performed to capture the fine spatio-temporal scales which govern the evolution of these systems. However, these simulations are often high-fidelity in nature, and can be computationally very expensive. Hence, generating data from these expensive simulations becomes a bottleneck in an end-to-end engineering design process. Spatio-temporal surrogate modeling of these dynamical systems has been a popular data-driven solution to tackle this computational bottleneck. This is because accurate machine learning models emulating the dynamical systems can be orders of magnitude faster than the actual simulations. However, one key limitation of purely data-driven approaches is their lack of generalizability to inputs outside the training distribution. In this paper, we propose a physics-informed spatio-temporal surrogate modeling (PISTM) framework constrained by the physics of the underlying dynamical system. The framework leverages state-of-the-art advancements in the field of Koopman autoencoders to learn the underlying spatio-temporal dynamics in a non-intrusive manner, coupled with a spatio-temporal surrogate model which predicts the behavior of the Koopman operator in a specified time window for unknown operating conditions. We evaluate our framework on a prototypical fluid flow problem of interest: two-dimensional incompressible flow around a cylinder.

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.

Desk Editor's Note private letter to a colleague

The paper combines a Koopman autoencoder with a separate surrogate for the operator to target generalization on unseen flow conditions, but supplies no numbers to show whether it actually works.

read the letter

The core idea is to train a Koopman autoencoder on simulation snapshots from known operating conditions so the latent dynamics become linear, then train a second spatio-temporal model that takes new parameters (Reynolds number, inlet speed) and outputs the operator or its action over a time window. This split lets the framework handle new conditions without retraining the encoder-decoder pair from scratch. The cylinder-flow test case is a sensible choice because it has clear vortex shedding and force quantities that matter for design work. The framing around computational cost in multi-physics engineering loops is straightforward and matches real bottlenecks. The non-intrusive angle is also practical since it works from existing simulation data rather than requiring intrusive adjoint or residual terms. The soft spot is exactly where the stress-test note lands: the abstract asserts accurate forecasts for unknown conditions yet gives zero quantitative support. No training versus test ranges, no prediction-window lengths, no L2 errors on velocity or integrated forces, and no ablation on whether the physics constraint actually improves extrapolation. Without those, it is impossible to tell if the predicted operator stays faithful or if the fixed encoder-decoder still reconstructs well. The paper is aimed at people building fast surrogates for fluid-design workflows who already know Koopman methods or physics-informed networks. A reader with that background can extract the architectural choice and see whether the full methods section fills in the missing evidence. If the results and ablations are solid, the work is worth referee time because the practical goal is clear even if the current write-up leaves the central claim unverified.

Referee Report

1 major / 1 minor

Summary. The manuscript proposes the Physics-Informed Spatio-Temporal Surrogate Modeling (PISTM) framework, which combines Koopman autoencoders to non-intrusively learn the latent linear dynamics of nonlinear spatio-temporal systems with a separate spatio-temporal surrogate that predicts the Koopman operator (or its action over a time window) for operating conditions outside the training distribution. The framework is evaluated on the two-dimensional incompressible flow around a cylinder as a prototypical fluid dynamics problem.

Significance. If the quantitative results demonstrate faithful extrapolation to unseen conditions while preserving reconstruction accuracy, the work would offer a practical advance in non-intrusive surrogate modeling for engineering design. The explicit separation of dynamics learning via Koopman operators from parameter-to-operator mapping addresses a recognized limitation of purely data-driven models and could reduce reliance on expensive high-fidelity simulations.

major comments (1)

[Results section] Results section (cylinder-flow experiments): the central claim that the framework produces accurate forecasts for unknown operating conditions is load-bearing yet unsupported by concrete evidence. No ranges are given for Reynolds numbers or inlet velocities separating training from test cases, no prediction-window lengths are stated, and no error metrics (relative L2 velocity error, integrated force coefficients, or comparison to baselines) are reported for held-out conditions. Without these, it is impossible to determine whether the predicted operator remains a faithful linearization or whether the encoder/decoder pair (trained only on the original distribution) continues to reconstruct accurately.

minor comments (1)

[Abstract] Abstract: the description of the evaluation omits any quantitative indicators (error magnitudes, parameter ranges, or window sizes), which would allow readers to gauge the strength of the generalization claim at first reading.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive and detailed review of our manuscript. We address the single major comment below and will revise the manuscript to incorporate the requested clarifications.

read point-by-point responses

Referee: [Results section] Results section (cylinder-flow experiments): the central claim that the framework produces accurate forecasts for unknown operating conditions is load-bearing yet unsupported by concrete evidence. No ranges are given for Reynolds numbers or inlet velocities separating training from test cases, no prediction-window lengths are stated, and no error metrics (relative L2 velocity error, integrated force coefficients, or comparison to baselines) are reported for held-out conditions. Without these, it is impossible to determine whether the predicted operator remains a faithful linearization or whether the encoder/decoder pair (trained only on the original distribution) continues to reconstruct accurately.

Authors: We agree that the results section requires additional quantitative detail to fully support the extrapolation claims. In the revised manuscript we will add explicit information on the Reynolds-number and inlet-velocity ranges used to define the training versus held-out sets, the exact lengths of the prediction time windows, and tabulated error metrics (relative L2 velocity errors, integrated drag/lift coefficient errors, and comparisons against baseline models) evaluated on the unseen operating conditions. These additions will enable readers to assess both the fidelity of the predicted Koopman operator and the reconstruction accuracy of the autoencoder outside the training distribution. revision: yes

Circularity Check

0 steps flagged

No circularity: framework is a novel combination of existing techniques with independent evaluation

full rationale

The paper presents PISTM as a combination of Koopman autoencoders (trained non-intrusively on known conditions) and a separate spatio-temporal surrogate that maps operating parameters to the learned operator for new conditions. No equation or claim reduces a prediction to a fitted input by construction, nor does any load-bearing step rely on self-citation for uniqueness or ansatz smuggling. The central claim of generalizability is supported by evaluation on the cylinder-flow benchmark rather than by re-deriving fitted quantities; the derivation chain remains self-contained against external data and prior non-author work on Koopman methods.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, so no explicit free parameters, axioms, or invented entities can be extracted; typical ML models of this type would involve latent dimension, regularization weights, and time-window length chosen during training.

pith-pipeline@v0.9.0 · 5527 in / 1032 out tokens · 24275 ms · 2026-05-10T13:02:41.544875+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages

[1]

org/abs/2003.02236

URL https://arxiv. org/abs/2003.02236. Chakraborty, S. Transfer learning based multi-fidelity physics informed deep neural network.CoRR, abs/2005.10614,

work page arXiv 2003
[2]

Erichson, N

URL https://arxiv.org/abs/ 2005.10614. Erichson, N. B., Muehlebach, M., and Mahoney, M. W. Physics-informed au- toencoders for lyapunov-stable fluid flow prediction,

work page arXiv 2005
[3]

Physics-informed Autoencoders for Lyapunov-stable Fluid Flow Prediction

URL https: //arxiv.org/abs/1905.10866. Fukami, K., Hasegawa, K., Nakamura, T., Morimoto, M., and Fukagata, K. Model order reduction with neural networks: Application to laminar and tur- bulent flows.SN Computer Science, 2(6):467, Sep

work page Pith review arXiv 1905
[4]

doi: 10.1007/s42979-021-00867-3

ISSN 2661-8907. doi: 10.1007/s42979-021-00867-3. URL https://doi.org/10.1007/ s42979-021-00867-3. Gonzalez, F. J. and Balajewicz, M. Deep convolutional recurrent autoencoders for learning low-dimensional feature dynamics of fluid systems,

work page doi:10.1007/s42979-021-00867-3
[5]

Greco, C

URL https://arxiv.org/abs/1808.01346. Greco, C. S., Paolillo, G., Astarita, T., and Cardone, G. The von kármán street behind a circular cylinder: flow control through synthetic jet placed at the rear stagnation point.Journal of Fluid Mechanics, 901:A39,

work page arXiv
[6]

Karniadakis, G., Kevrekidis, Y ., Lu, L., Perdikaris, P., Wang, S., and Yang, L

doi: 10.1017/jfm.2020.427. Karniadakis, G., Kevrekidis, Y ., Lu, L., Perdikaris, P., Wang, S., and Yang, L. Physics-informed machine learning. pp. 1–19, 05

work page doi:10.1017/jfm.2020.427 2020
[7]

Hamiltonian Systems and Transformation in Hilbert Space,

doi: 10.1073/pnas.17.5.315. URL https://www.pnas.org/doi/abs/10. 1073/pnas.17.5.315. Krüger, T., Kusumaatmaja, H., Kuzmin, A., Shardt, O., Silva, G., and Viggen, E. M. The Lattice Boltzmann Method - Principles and Practice. 10

work page doi:10.1073/pnas.17.5.315
[8]

doi: 10.1007/978-3-319-44649-3

ISBN 978-3- 319-44647-9. doi: 10.1007/978-3-319-44649-3. Kutz, J. N. Deep learning in fluid dynamics.Journal of Fluid Mechanics, 814:1–4,

work page doi:10.1007/978-3-319-44649-3
[9]

Maulik, R., Botsas, T., Ramachandra, N., Mason, L

doi: 10.1017/jfm.2016.803. Maulik, R., Botsas, T., Ramachandra, N., Mason, L. R., and Pan, I. Latent-space time evolution of non-intrusive reduced-order models using gaussian process emulation.Physica D: Nonlinear Phenomena, 416:132797, Feb

work page doi:10.1017/jfm.2016.803 2016
[10]

doi: 10.1016/j.physd.2020.132797

ISSN 0167-2789. doi: 10.1016/j.physd.2020.132797. URL http://dx.doi.org/ 10.1016/j.physd.2020.132797. Meng, X. and Karniadakis, G. E. A composite neural network that learns from multi- fidelity data: Application to function approximation and inverse pde problems. Journal of Computational Physics, 401:109020,

work page doi:10.1016/j.physd.2020.132797 2020
[11]

doi: https: //doi.org/10.1016/j.jcp.2019.109020

ISSN 0021-9991. doi: https: //doi.org/10.1016/j.jcp.2019.109020. URL https://www.sciencedirect. com/science/article/pii/S0021999119307260. Mohan, A. T. and Gaitonde, D. V . A deep learning based approach to reduced order modeling for turbulent flow control using lstm neural networks,

work page doi:10.1016/j.jcp.2019.109020 2019
[12]

Mondal, S

URL https://arxiv.org/abs/1804.09269. Mondal, S. and Sarkar, S. Multi-fidelity modeling of spatio-temporal fields. In AAAI 2022 Workshop on AI for Design and Manufacturing (ADAM),

work page arXiv 2022
[13]

Raissi, M., Perdikaris, P., and Karniadakis, G

doi: https: //doi.org/10.4271/2021-01-0550. Raissi, M., Perdikaris, P., and Karniadakis, G. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics, 378: 686–707,

work page doi:10.4271/2021-01-0550 2021
[14]

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

ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.2018.10.045. URL https://www.sciencedirect.com/science/article/pii/ S0021999118307125. Non-intrusive Learning of Physics-Informed Spatio-temporal Surrogate for Accelerating Design Rasmussen, C. and Williams, C.Gaussian Processes for Machine Learning (Adap- tive Computation and Machine Learning). The MIT P...

work page doi:10.1016/j.jcp.2018.10.045 2018
[15]

URL https://arxiv.org/abs/2010.00399. Tu, J. H., Rowley, C. W., Luchtenburg, D. M., Brunton, S. L., and Kutz, J. N. On dy- namic mode decomposition: Theory and applications.Journal of Computational Dynamics, 1(2):391–421, 2014

work page arXiv 2010