acceptodds
Under review as a conference paper at ICLR 2027

Learning Large-Step Fluid Dynamics with Navier–Stokes Structure

Abstract

While the Navier–Stokes equations provide a continuous-time description of fluid motion, learning discrete-time fluid dynamics directly from observations remains highly challenging. Over large time intervals, fluid evolution becomes strongly nonlinear, making it difficult for purely data-driven neural networks to recover the underlying dynamical mechanisms. Consequently, existing methods may fit observed trajectories well with low per-step errors, yet still suffer from rapidly accumulating errors during autoregressive rollout and often generalize poorly beyond the training distribution. We introduce NeNS, a neural surrogate for incompressible flow that learns large-step dynamics with Navier–Stokes structure. On rectangular domains with periodic or free-slip boundaries, NeNS represents the flow using analytical eigenfunctions of the Stokes operator, thereby satisfying incompressibility and boundary conditions by construction. In this modal basis, the viscous term is naturally diagonalized. Guided by the algebraic structure of the equations, we parameterize the remaining nonlinear dynamics with a low-rank skew-symmetric bilinear operator. This operator is learned directly from finite-time state pairs, without requiring physics-informed losses, while inherently preserving the kinetic-energy conservation property of the nonlinear term. On in-domain benchmarks, NeNS reduces held-out relative vorticity error by approximately , compared with the strongest evaluated baselines. Moreover, although trained only on randomized initial conditions, NeNS exhibits strong out-of-distribution generalization, achieving the lowest errors on independently generated leapfrogging-vortex and Taylor-vortex dynamics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
fluid simulation, reduced-order models, neural operators

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qfNKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Purely data-driven neural networks for fluid dynamics often suffer from rapidly accumulating errors during autoregressive rollout over large time intervals, motivating the introduction of NeNS^2, a neural surrogate designed to learn large-step incompressible flow dynamics with an explicit Navier-Stokes structure. By representing flow on rectangular domains using analytical eigenfunctions of the Stokes operator, the framework satisfies incompressibility and boundary conditions by construction while parameterizing nonlinear dynamics with a low-rank skew-symmetric bilinear operator to inherently preserve kinetic-energy conservation.

Critical strengths
  • The authors provide comparison with numerical solver baselines.

  • The model demonstrates outstanding empirical performance compared to strong baselines and achieving excellent out-of-distribution generalization on complex independent scenarios.

Critical weaknesses
  • The experiments are only conducted on rectangular domains with periodic or free-slip boundaries, making it unclear how the method generalizes to irregular meshes and other boundary conditions (particularly whether the Stokes eigenbasis can be calculated on an irregular mesh).

  • While the authors claim their method satisfies incompressibility and boundary conditions by construction, they should report the incompressibility error, boundary error, and boundary residual to justify this, and clarify exactly how uncommon conditions (such as integral boundary conditions) are mathematically satisfied.

  • The authors should discuss and compare their approach with existing neural PDE solvers that use eigenbases and spectral representations to decompose spatial or temporal dynamics into characteristic modes, which generalize beyond the Navier-Stokes equation.

  • It is unclear if all baselines also use RK4; if not, the comparison is unfair, as the baselines could potentially improve their performance by integrating RK4. The authors must also report wall-clock inference time, as the proposed model will be significantly slower if it is the only one using RK4.

  • More details are needed in Section 4.3 (Efficiency Against Classical Spectral Simulation), specifically regarding what ground truth data was used to calculate the error against the classical method and whether the evaluation follows the fair protocols established by [1, 2].

[1] Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations

[2] FD-Bench: A Modular and Fair Benchmark for Data-driven Fluid Simulation

Questions for authors
  • Which specific step in the methodology is restricted to the Navier-Stokes equation, and can this framework be extended to other PDEs?

  • How does the methodology actively support large-step predictions as highlighted in the title, and exactly how large is the experimental step compared to existing works?

  • Why did FNO and LNO diverge during testing? This divergence, particularly for FNO on periodic Navier-Stokes equations, directly contradicts existing literature and requires further analysis to validate the experimental results.

3 comments
AuthorsOpenReview ↗

(1/3) Response to Reviewer qfNK's comments

We sincerely appreciate the reviewer’s thoughtful feedback and constructive suggestions. We address each point in detail below:

  • Boundary Conditions and Geometry (W1)
  • Physical Constraints and Boundary Residuals (W2)
  • Related Spectral Methods and PDE Generalization (W3, Q1)
  • Baseline Fairness and Inference Efficiency (W4)
  • Classical Solver Comparison (W5)
  • Large-Step Prediction (Q2)
  • Analysis of Baseline Divergence (Q3)

Please refer to the revised PDF with all changes highlighted in blue. The updated supplementary web viewer at CLICKME.html further provides side-by-side video comparisons of the evaluated rollouts and baselines.


1. Boundary Conditions and Geometry (W1)

Our framework already supports irregular geometries and no-slip boundary conditions through the immersed-boundary method (IBM) [1]. As demonstrated in the single-cylinder and two-cylinder experiments in Section 4.4, NeNS² successfully reproduces vortex deflection and transport around unseen obstacles without retraining. IBM imposes no-slip constraints through boundary forces, allowing us to handle complex geometries while retaining the original analytical Stokes basis.

Our choice of IBM reflects a practical trade-off between boundary accuracy and computational efficiency. While IBM imposes boundary constraints approximately, a more rigorous alternative would be to compute geometry-adapted numerical Stokes eigenfunctions, which incorporate the prescribed boundary conditions directly into the modal basis. However, computing thousands of eigenmodes on irregular domains can incur substantial computational overhead. We therefore favor IBM for its efficiency and flexibility, while viewing numerical Stokes bases as a promising direction for improving boundary accuracy in future work.

We have expanded the discussion in the revised manuscript (see Section 4.4 and Appendix E.7) to clarify our treatment of complex geometries and no-slip boundaries, along with the associated computational trade-offs. The corresponding obstacle flow rollouts are available in the supplementary web viewer under Passive-dye rollouts.


2. Physical Constraints and Boundary Residuals (W2)

Our Stokes basis satisfies incompressibility and the prescribed periodic or free-slip boundary conditions by construction. To further verify these properties numerically, we evaluated a 121-frame free-slip rollout on a 128×128128\times128 grid, computing spatial derivatives analytically from the Stokes basis (see Appendix E.5 and Table 11 in the revised manuscript). We report the maximum normalized residual over the entire rollout, using the RMS velocity for normalization.

MetricMaximum Relative Residual ↓
Incompressibility (divergence)1.06×10−61.06\times10^{-6}
Normal velocity at boundaries1.98×10−71.98\times10^{-7}
Tangential stress at boundaries (excluding viscosity)1.64×10−51.64\times10^{-5}

All measured residuals are near numerical precision, consistent with the structural guarantees of our formulation. The boundary residuals correspond to the no-penetration and zero-tangential-stress conditions of the homogeneous free-slip setting.

For more general boundary conditions, including compatible integral constraints, the immersed-boundary formulation can incorporate additional constraint forces. Specifically, a discrete constraint Cwn+1=bCw^{n+1}=b can be enforced by correcting the predicted modal coefficients w∗w^{\ast} through a force term GλG\lambda, where λ\lambda is obtained by solving CGλ=b−Cw∗CG\lambda=b-Cw^*. This provides a possible extension to more general boundary constraints.


3. Related Spectral Methods and PDE Generalization (W3, Q1)

Following this suggestion, we have expanded the Related Work section to include Spectral Neural Operators [2] and the Laplace Neural Operator [3], which also exploit spectral representations for learning PDE dynamics.

Our approach differs in that the spectral representation is coupled with the specific algebraic structure of the Navier–Stokes equations. In particular, the divergence-free Stokes basis enforces incompressibility and boundary conditions, the viscous operator is diagonalized, and the skew-symmetric quadratic generator preserves the energy-conserving structure of nonlinear advection.

The general framework could also be extended to other PDEs by replacing these Navier–Stokes-specific components with appropriate spectral bases, linear operators, and nonlinear structures. The underlying idea of learning structured modal dynamics through differentiable temporal integration remains applicable.

AuthorsOpenReview ↗

(2/3) Response to Reviewer qfNK's comments

(Continue)


4. Baseline Fairness and Inference Efficiency (W4)

We appreciate the reviewer's concern regarding the fairness of numerical integration settings. In the original submission, the neural baselines used their native prediction schemes rather than RK4 integration. To address this concern, we have added a new comparison in which all models use RK4 with the same two internal substeps per prediction interval, ensuring consistent integration settings across methods (see Table 10 in the revised manuscript). All baseline checkpoints are reused without retraining.

We evaluate both prediction accuracy and wall-clock inference time over the complete 120-step leapfrogging OOD rollout:

ModelVorticity Error ↓Runtime ↓
NeNS²5.033%0.301 s
U-Net102.898%6.759 s
PDE-T SC-100 (scratch)48.539%30.524 s
PDE-T SC-B (scratch)47.956%30.877 s
PDE-T SC-B (pretrained)48.222%30.487 s

NeNS² achieves both the lowest prediction error and the fastest inference time under the same RK4 integration settings. We have also clarified the integration settings and wall-clock timing protocol in the revised manuscript.


5. Classical Solver Comparison (W5)

We are grateful for this constructive point regarding reference solutions and benchmarking fairness. We have clarified the ground-truth construction and evaluation protocols in Appendix E.8.

For the trajectory prediction experiments in Section 4.1, the reference solutions are generated using an independent float64 DOP853 integration of the resolved Galerkin system. For the efficiency comparison in Section 4.3, we use the same retained modes, initial conditions, prediction interval, and NVIDIA A100 GPU to evaluate the classical solver and NeNS².

All reported timings include the complete numerical integration process, including the classical solver's internal substeps and all generator evaluations required by NeNS². We also report the classical periodic solver at its native 1542154^2 resolution to make the effect of numerical configuration explicit.

We have further clarified the error definitions and measurement protocols in Appendix E.8, following the principles of controlled comparisons emphasized by FD-Bench [4] and McGreivy and Hakim [5].


6. Large-Step Prediction (Q2)

We would like to clarify that our use of “large-step” is relative to the time-step restrictions of conventional explicit modal integration. As discussed by Cui et al. [6], explicit integration can require very small time steps to maintain stability, with a reported maximum stable step size of approximately 10−610^{-6} at 1,000 modes. This motivated their use of implicit time integration.

To provide further evidence, we have added a comparison with the explicit integration scheme discussed by Cui et al. [6] (see Appendix E.6, Table 13, and Fig. 8 in the revised manuscript). We use an output interval of Δt=0.02\Delta t=0.02 and evaluate 120-step rollouts up to T=2.4T=2.4, varying the number of internal substeps.

MethodSubstepsMean Vorticity Error ↓Mean Velocity Error ↓
Cui et al. (explicit)1–16Diverged—
Cui et al. (explicit)324.716%1.461%
NeNS²25.033%1.356%

As shown above, the explicit scheme requires at least 32 substeps to remain stable in this experiment, whereas NeNS² achieves stable prediction with only two substeps, corresponding to a 16× larger internal time step.

Side-by-side rollout videos illustrating the divergence of explicit stepping across different substeps are available in the supplementary web viewer under Structural analysis →\rightarrow Large-step.

We hope this additional comparison helps clarify the intended meaning of “large-step” in our work and provides further empirical support for our choice of prediction interval.

AuthorsOpenReview ↗

(3/3) Response to Reviewer qfNK's comments

(Continue)


7. Analysis of Baseline Divergence (Q3)

The divergence of FNO and LNO over long rollouts stems primarily from the evaluation setup: all models are evaluated under the same single-frame autoregressive setting, where the current state is used to predict the next state. In contrast, previous FNO and LNO evaluations commonly use multiple historical frames (e.g., 10 input frames) to predict the next frame.

This single-frame setting is more challenging for these baselines, as they cannot leverage temporal information from previous observations. However, it ensures a consistent comparison with NeNS², which also predicts the next state from a single current state.

We believe this difference in evaluation settings helps explain the weaker performance and eventual divergence observed in some baselines during long-term autoregressive rollout.


Summary

We sincerely thank the reviewer for the careful and constructive evaluation. We hope the additional experiments, clarified evaluation protocols, and expanded discussion address the concerns regarding physical consistency, baseline fairness, and large-step prediction. We welcome any further discussion.


References

[1] K. Taira and T. Colonius. "The Immersed Boundary Method: A Projection Approach." Journal of Computational Physics, 225(2), 2118–2137, 2007. https://doi.org/10.1016/j.jcp.2007.03.005

[2] V. S. Fanaskov and I. V. Oseledets. "Spectral Neural Operators." Doklady Mathematics, 108(Suppl 2), S226–S232, 2023. https://doi.org/10.1134/S1064562423701107

[3] "Laplace Neural Operator." Nature Machine Intelligence, 2024. https://doi.org/10.1038/s42256-024-00844-4

[4] "FD-Bench: A Modular and Fair Benchmark for Data-driven Fluid Simulation." https://arxiv.org/abs/2505.20349

[5] N. McGreivy and A. Hakim. "Weak Baselines and Reporting Biases Lead to Overoptimism in Machine Learning for Fluid-Related Partial Differential Equations." https://arxiv.org/abs/2407.07218

[6] Q. Cui, P. Sen, and T. Kim. "Scalable Laplacian Eigenfluids." ACM Transactions on Graphics, 37(4), Article 87, 2018. https://doi.org/10.1145/3197517.3201352

Reviewer_SJpWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes NeNS², a structured surrogate for large-step prediction of two-dimensional incompressible fluid dynamics. The velocity field is represented in a Stokes eigenbasis, enforcing incompressibility and the prescribed boundary conditions by construction. Viscous dynamics are represented analytically in this basis, while nonlinear modal interactions are modeled through a low-rank skew-symmetric bilinear operator. The skew-symmetric construction ensures that the learned nonlinear term preserves kinetic energy in the inviscid limit.

Rather than directly fitting time derivatives or learning an unconstrained finite-step mapping, the model is trained on finite-time state pairs by differentiating through an RK4 integration of the structured vector field. The paper evaluates the method on periodic and free-slip two-dimensional flows, including held-out and out-of-distribution initial conditions. The reported experiments indicate improved accuracy and rollout stability relative to several grid-based neural operator and neural network baselines.

Critical strengths

Physics-informed formulation: The model incorporates incompressibility, boundary conditions, viscous dissipation, and energy conservation of the inviscid nonlinear term by construction

Efficient representation of nonlinear interactions: The low-rank skew-symmetric factorization avoids the cubic parameter cost of a dense quadratic interaction tensor while preserving a meaningful physical invariant.

Strong empirical performance within the evaluated scope: The proposed method performs consistently well for both periodic and free-slip settings and exhibits stable long-horizon rollouts for the limited use cases presented.

Interesting formulation. The paper provides an interesting bridge between Galerkin reduced-order modeling, operator inference, neural differential equations, and neural operators.

Critical weaknesses

Restricted geometry and boundary conditions: The method depends on a suitable Stokes eigenbasis and is demonstrated only on periodic or rectangular free-slip domains. It is unclear how the approach would extend to irregular geometries and no-slip boundaries, or complexities.

Missing comparisons with the closest methodological baselines. The main comparisons are against grid-based neural operators and neural networks. However, the closest alternatives appear to be operator inference, quadratic reduced-order models, structure-preserving ROMs, and learned modal dynamical systems. Direct comparisons with these methods would be important, especially the formulation in 'On the representation of energy-preserving quadratic operators with application to operator inference'.

Large practical parameter count: Although the factorization is efficient relative to a dense third-order tensor, the largest reported model remains large even for the simplified use cases presented. A systematic analysis of accuracy, rank, memory, and computational cost would strengthen the practical case.

Narrow empirical scope: The experiments are limited to 2D incompressible flows. How does the method scale to three-dimensional flows, turbulent regimes, varying Reynolds numbers, or different forcing conditions.

Missing benchmark studies: The readers would benefit from seeing the results against some benchmark problems such as the Kolmogorov flow (somewhat similar to the results here), Dynamic stall etc. used in other papers

Questions for authors

Some additional questions to consider if possible:

What approach would you take to extend the method to irregular geometries and no-slip boundaries?

How sensitive are the results to the number of retained modes M and factorization rank r? Please provide an accuracy-cost curve if possible.

How much performance is lost if the Stokes basis is retained but the structured generator is replaced by an unconstrained finite-time mapping?

Does the energy-preserving property of the continuous nonlinear vector field lead to any corresponding stability guarantee for the finite-step RK4 rollout?

How does the method behave when a test trajectory contains substantial energy outside the retained modal subspace?

5 comments
AuthorsOpenReview ↗

(1/3) Response to Reviewer SJpW's comments

We are grateful for the reviewer’s valuable comments and insightful suggestions. Our detailed responses to each concern are as follows:

  • Boundary Conditions and Geometry (W1, Q1)
  • Comparison with Structured ROMs (W2)
  • Model Complexity and Efficiency (W3, Q2)
  • Extension to Broader Flow Regimes (W4)
  • Additional Fluid Benchmarks (W5)
  • Effectiveness of Structured Dynamics (Q3)
  • Energy Conservation and Stability (Q4)
  • Impact of Modal Truncation (Q5)

Please refer to the revised PDF with all changes highlighted in blue, and to the updated supplementary web viewer at CLICKME.html, which provides side-by-side video comparisons across the newly evaluated settings.


1. Boundary Conditions and Geometry (W1, Q1)

Our framework already supports irregular geometries and no-slip boundary conditions through the immersed-boundary method (IBM) [1]. As demonstrated in the single-cylinder and two-cylinder experiments in Section 4.4, NeNS² successfully reproduces vortex deflection and transport around unseen obstacles without retraining. IBM imposes no-slip constraints through boundary forces, allowing us to handle complex geometries while retaining the original analytical Stokes basis.

Our choice of IBM reflects a practical trade-off between boundary accuracy and computational efficiency. While IBM imposes boundary constraints approximately, a more rigorous alternative would be to compute geometry-adapted numerical Stokes eigenfunctions, which incorporate the prescribed boundary conditions directly into the modal basis. However, computing thousands of eigenmodes on irregular domains can incur substantial computational overhead. We therefore favor IBM for its efficiency and flexibility, while viewing numerical Stokes bases as a promising direction for improving boundary accuracy in future work.

We have expanded the discussion in the revised manuscript (see Section 4.4 and Appendix E.7) to clarify our treatment of complex geometries and no-slip boundaries, along with the associated computational trade-offs. The corresponding obstacle flow rollouts are available in the supplementary web viewer under Passive-dye rollouts.


2. Comparison with Structured ROMs (W2)

We thank the reviewer for highlighting Gkimisis et al. [2] and the importance of comparisons with structure-preserving reduced-order models. We have expanded the Related Work section to clarify the connections and distinctions between these approaches and our method.

Gkimisis et al. [2] study energy-preserving quadratic operators and infer their coefficients through sequential least-squares optimization. However, directly applying this approach to thousands of modes and large training datasets would incur substantial computational and memory costs. To provide a practical comparison, we instead consider Goyal et al. [3], suggested by Reviewer DKP5, which also develops a structure-preserving quadratic ROM but employs Adam-based optimization, making it more feasible at our experimental scale.

Specifically, we have added a POD-based energy-preserving quadratic ROM baseline following the skew-structured representation of these works (see Sec. 4.2.2, Table 2, and Appendix E.3 in the revised manuscript). The baseline retains 586 POD modes extracted from the training data and uses approximately the same parameter budget (100M) and optimization procedure as NeNS², with identical finite-time supervision.

Mean relative error ↓NeNS²POD Structured ROM
ID velocity0.81%18.79%
ID vorticity2.76%38.04%
Leapfrogging OOD velocity1.36%18.26%
Leapfrogging OOD vorticity5.03%53.10%

NeNS² achieves the lowest errors across all evaluated metrics, highlighting the effectiveness of combining physically defined Stokes coordinates with low-rank, structure-preserving modal dynamics.

Side-by-side rollout videos comparing NeNS² with the POD-based structured ROM are available in the supplementary web viewer under Structural analysis →\rightarrow POD ROM.


3. Model Complexity and Efficiency (W3, Q2)

To systematically examine how model capacity and computational efficiency scale with the factorization rank, we have added a systematic rank-scaling study (see Fig. 6 in the revised manuscript).

Specifically, we vary the factorization rank rr from 512 to 8192 while fixing the number of Stokes modes at M=4096M=4096, allowing us to examine how model expressiveness changes with parameter count at a fixed spatial representation capacity. Increasing the rank consistently improves prediction accuracy on both in-distribution and OOD trajectories, accompanied by increases in parameter count, GPU memory usage, and inference time.

These results illustrate the accuracy–efficiency trade-off of our low-rank formulation and provide further evidence of its scalability to thousands of fluid modes.

AuthorsOpenReview ↗

(2/3) Response to Reviewer SJpW's comments

(Continue)


4. Extension to Broader Flow Regimes (W4)

We thank the reviewer for these insightful suggestions. We agree that extending the evaluation to more diverse flow regimes would further demonstrate the applicability of our framework.

  • Extension to 3D. Our structured formulation naturally extends to three-dimensional flows, as the quadratic, energy-preserving generator can also be defined over 3D vector Stokes modes. While preliminary experiments suggest the feasibility of this extension, resolving more complex 3D and turbulent dynamics requires substantially more modes and greater computational efficiency. We consider this an important direction for future work.
  • Varying Reynolds numbers. Our formulation separates learned nonlinear dynamics from analytical viscous dissipation, allowing the same trained model to accommodate different Reynolds numbers through the viscosity term e−νΛΔte^{-\nu\Lambda\Delta t}. Our existing experiments already cover a broad range of Reynolds numbers: Re=983Re=983–16031603 for free-slip ID flows (200 trajectories), Re=1596Re=1596–22432243 for periodic ID flows (200 trajectories), and Re=1092Re=1092 for the leapfrogging OOD scenario. These results demonstrate accurate predictions across varied flow conditions.
  • External forcing. Our framework also accommodates sustained external forcing by projecting the force into the Stokes basis and incorporating it through the forcing split. We have added a Kolmogorov flow experiment to evaluate this capability without retraining (see Table 3 in the revised manuscript). NeNS² achieves a mean vorticity error of 3.61%, compared with 9.02% for the strongest baseline.

5. Additional Fluid Benchmarks (W5)

Following the reviewer's suggestion, we have added a Kolmogorov flow benchmark to further evaluate the generalization capability of NeNS² (see Section 4.4, Table 3, Fig. 4, and Appendix E.2 in the revised manuscript).

The experiment considers continuously forced Kolmogorov flow on the unit periodic domain, with forcing mode nf=4n_{f}=4, viscosity ν=3×10−4\nu=3\times10^{-4}, and 120 prediction steps at Δt=0.02\Delta t=0.02. All models reuse their periodic checkpoints without retraining and incorporate the same external forcing.

ModelVorticity Error ↓Velocity Error ↓
NeNS²3.61%2.78%
PDE-T SC-100 (scratch)9.02%7.59%
PDE-T SC-B (pretrained)9.66%8.24%
PDE-T SC-B (scratch)16.19%14.33%
CNO37.45%36.14%

NeNS² achieves the lowest errors across all evaluated methods, demonstrating its ability to generalize from unforced training flows to a continuously forced benchmark.

Side-by-side rollout videos comparing all evaluated models on continuously forced Kolmogorov flow are available in the supplementary web viewer under Flow comparisons →\rightarrow Kolmogorov flow.


6. Effectiveness of Structured Dynamics (Q3)

To further examine the contribution of our structured dynamics, we have added a comparison with parameter-matched FNO1D and U-Net1D models operating on the same Stokes coefficients (see Table 9 in the revised manuscript).

ModelOOD Vorticity Error ↓OOD Velocity Error ↓Runtime ↓
NeNS²5.03%1.36%0.299 s
Coefficient FNO1D129.60%102.29%1.544 s
Coefficient U-Net1D109.92%67.70%14.064 s

NeNS² achieves substantially better prediction accuracy and efficiency, despite all models operating in the same modal space. These results suggest that the benefits of our approach arise not only from the Stokes basis, but also from the structured representation of the underlying dynamics.

AuthorsOpenReview ↗

(3/3) Response to Reviewer SJpW's comments

(Continue)


7. Energy Conservation and Stability (Q4)

Our skew-symmetric generator guarantees energy conservation for the continuous-time nonlinear dynamics, providing a structural foundation for stable evolution. With RK4 integration, the local energy error is O(h5)O(h^5) for smooth bounded trajectories, allowing accurate energy preservation at finite time steps.

Our added 120-step energy analysis (see Fig. 7 in the revised manuscript) further evaluates this property empirically. We retain the original viscous trajectories and subtract the energy loss from each discrete viscosity update. We measure DN=(EN−E0+∑n=0N−1Lnvisc)/E0D_N=(E_N-E_0+\sum_{n=0}^{N-1}L_n^{\mathrm{visc}})/E_0, where LnviscL_n^{\mathrm{visc}} is the energy loss caused by the actual discrete viscosity update at step nn. The maximum absolute corrected deviation over the full rollout, max⁡0≤N≤120∣DN∣\max_{0\leq N\leq120}|D_N|, is 9.02×10−69.02\times10^{-6} for NeNS², compared with 2.76×10−32.76\times10^{-3} for no-skew, 3.63×10−23.63\times10^{-2} for coefficient-space FNO1D, and 5.22×10−15.22\times10^{-1} for coefficient-space U-Net1D. The no-RK4 variant becomes non-finite at t=0.36t=0.36 and is not assigned a full-rollout score. This diagnostic measures non-viscous energy drift along the original trajectories, including numerical integration and floating-point effects.


8. Impact of Modal Truncation (Q5)

Energy outside the retained modal subspace introduces an irreducible projection error and may also influence the evolution of resolved modes. By orthogonality, the reconstruction error can be decomposed as

∥u−u^M∥2=∥PMu−u^M∥2+∥(I−PM)u∥2,\|u-\widehat{u}_M\|^2 = \|P_Mu-\widehat{u}_M\|^2 + \|(I-P_M)u\|^2,

where the second term represents the energy excluded by modal truncation.

Increasing the number of retained modes MM reduces this projection error, while incorporating a subgrid closure could further account for the influence of unresolved dynamics. Our low-rank formulation makes larger modal spaces computationally feasible, and our current benchmarks evaluate reference and predicted dynamics at the same modal resolution.


Summary

We sincerely thank the reviewer for the thoughtful and constructive feedback. We especially appreciate your recognition of our physics-informed design and strong empirical performance. Your suggestions have helped us further strengthen our experimental evaluation, and we welcome any further discussion.


References

[1] K. Taira and T. Colonius. "The Immersed Boundary Method: A Projection Approach." Journal of Computational Physics, 225(2), 2118–2137, 2007. https://doi.org/10.1016/j.jcp.2007.03.005

[2] L. Gkimisis, I. Pontes Duff, P. Goyal, and P. Benner. "On the Representation of Energy-Preserving Quadratic Operators with Application to Operator Inference." Applied Mathematics Letters, 173, 109761, 2026. https://doi.org/10.1016/j.aml.2025.109761

[3] P. Goyal, I. Pontes Duff, and P. Benner. "Guaranteed Stable Quadratic Models and Their Applications in SINDy and Operator Inference." Physica D: Nonlinear Phenomena, 483, 134893, 2025. https://doi.org/10.1016/j.physd.2025.134893

Reviewer_DKP5OpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces NeNS$^2$, a novel neural solver for the two-dimensional incompressible Navier–Stokes equations on rectangular domains with periodic or free-slip boundaries. The method respects the structure of the Navier–Stokes equations, enabling accurate and efficient large-step inference, even for out-of-distribution (OOD) fluid scenarios. Using a Stokes basis to represent the flow, it decomposes the overall physics into two parts: (1) viscous decay, governed by fixed physics, and (2) the remaining nonlinear advection dynamics, which are learned. The proposed skew-symmetric bilinear operator conserves kinetic energy by construction. Combined with RK4 time integration, the method achieves better stability than a simpler integration scheme such as Euler.

This paper could be an interesting methodological contribution to the current machine learning for fluid dynamics community, several clarification and improvement could further improve the quality of the work.

Critical strengths
  1. Physically consistent design: The method separates the overall physics into (a) learned nonlinear advection and (b) analytical diffusion through viscous decay. This design can enable the trained model to have the potential to adapt to different viscosities and flow condition. The energy-preserving bilinear operator ensures energy conservation for the advection dynamics.

  2. OOD inference capability: Section 4.4 provides a valuable discussion of generalization to different flow configurations and unseen geometries, suggesting promise for further two-dimensional applications. The dye transport experiments in Section 4 and the supplementary video are intuitive and visually convincing.

  3. RK4 versus Euler ablation: The ablation in Table 2 provides strong evidence that RK4 integration is a necessary component for maintaining the model’s temporal stability. This effectively supports the overall methodology.

Critical weaknesses

Major

  1. Large-step claim: Δt=0.02\Delta t = 0.02 is used consistently throughout training. Although this interval might be considered “large” for the cases presented, its significance is less clear without reference to the relevant physics. For example, flows with matching Reynolds numbers under same geometry and initial condition can exhibit similar dynamics on different timescales, so discussing the scale of the time step becomes less meaningful without considering other flow property such as velocity and characteristic length. The authors should consider reporting a more physically meaningful reference quantity, such as (but not limited to) the CFL number associated with the specified Δt\Delta t. Testing the proposed model and baseline models across several values of Δt\Delta t would also make the large-step claim more defensible.

  2. Novelty claim: Although, to the reviewer’s knowledge, this is a noval work to apply structure-preserving learning with Runge–Kutta integration to two-dimensional incompressible flow using a finite rank factorization, existing literature also enforces energy preservation through skew-symmetric quadratic operators combined with Runge–Kutta integration (see https://doi.org/10.1016/j.physd.2025.134893).

  3. Energy conservation claim: Although the energy-conserving construction of the bilinear operator is technically sound, RK4 integration can affect energy conservation slightly, as discussed in Appendix B. Additional ablations comparing the full model with the variant without skew symmetry and other baseline models would substantially strengthen the energy-conservation claim. Measuring relative energy drift over a fixed number of rollout steps would help readers assess the effectiveness of the proposed method in conserving the system’s energy.

Minor

Boundary-condition scope and practical applicability: The paper acknowledges that the analytic basis restricts the method to rectangular periodic and free-slip domains. However, many practical applications require no-slip boundary conditions. The paper could outline how the framework might extend to no-slip conditions. Identifying applications where free-slip boundaries are useful would also help clarify the method’s practical impact.

Summary

While the formulation is promising, I lean toward weak rejection because the distinction from closely related work and the empirical support for the central large-step claim remain insufficiently established. Clarifying the methodological novelty and providing comparisons across prediction intervals could change my assessment.

Questions for authors
  1. What are the main differences between the proposed method and existing approaches such as this related work (https://doi.org/10.1016/j.physd.2025.134893)? Could the authors clarify which contributions are methodologically new and which arise from adapting established techniques?

  2. Given the acknowledged restrictions on boundary conditions, what modifications would be necessary to support no-slip boundaries and irregularly shaped domains? Could the authors clarify whether the current limitation stems primarily from computational challenges or from the Stokes basis formulation or something else?

6 comments
AuthorsOpenReview ↗

(1/2) Response to Reviewer DKP5's comments

We appreciate the reviewer’s insightful comments and constructive feedback. Below, we provide detailed responses to each concern as follows:

  • Large-Step Prediction (W1)
  • Methodological Novelty (W2, Q1)
  • Energy Conservation (W3)
  • Boundary Conditions and Generalization (W4, Q2)

Please refer to the revised PDF with all changes highlighted in blue, as well as the updated supplementary web viewer at CLICKME.html, which hosts side-by-side video comparisons of the evaluated flow dynamics.


1. Large-Step Prediction (W1)

We clarify that our claim of “large-step” prediction (Δt=0.02\Delta t=0.02) is defined relative to the severe stability restrictions of conventional explicit modal integration. As discussed by Cui et al. [1], explicit integration can require very small time steps to maintain stability, with a reported maximum stable step size of approximately 10−610^{-6} at 1,000 modes, which historically motivated the use of implicit time integration.

To provide direct empirical evidence, we have added a systematic comparison with the explicit modal stepping of Cui et al. [1] (see Appendix E.6, Table 13, and Fig. 8 in the revised manuscript). We evaluate 120-step rollouts (T=2.4T=2.4, Δt=0.02\Delta t=0.02), varying the number of internal substeps:

MethodSubstepsMean Vorticity Error ↓Mean Velocity Error ↓
Cui et al. (explicit)1–16Diverged—
Cui et al. (explicit)324.716%1.461%
NeNS²25.033%1.356%

As shown above, the explicit scheme requires at least 32 substeps to remain stable in this experiment, whereas NeNS² achieves stable prediction with only two substeps, corresponding to a 16× larger internal time step.

The side-by-side rollout videos illustrating the divergence of explicit stepping across different substeps are available in the supplementary web viewer under Structural analysis →\rightarrow Large-step.

We hope this additional comparison helps clarify the intended meaning of “large-step” in our work and provides further empirical support for our choice of prediction interval.


2. Methodological Novelty (W2, Q1)

We agree that energy-preserving quadratic operators and RK4-based learning have been explored previously, including in Goyal et al. [2], and we have expanded the Related Work section accordingly.

Our key contribution is to make structured learning practical at thousands of physically defined fluid modes by combining the Stokes basis with a low-rank skew-symmetric parameterization.

Specifically, the Stokes basis enforces incompressibility and boundary conditions while diagonalizing viscous dynamics. Our low-rank parameterization preserves the energy-conserving structure of nonlinear advection while reducing storage and evaluation complexity from O(M3)O(M^3) to O(Mr)O(Mr). For example, our 4096-mode model uses 100.66M parameters, compared with 68.72B entries in a dense quadratic tensor before symmetry reductions.

To further validate this distinction, we have added an energy-preserving quadratic ROM baseline following Goyal et al. [2] using data-driven POD modes and finite-time RK4 supervision under a strictly matched $\sim$100M parameter budget (see Sec. 4.2.2, Table 2, and Appendix E.3 in the revised manuscript). Because dense skew-symmetric quadratic tensors scale as O(M3)O(M^3), a 100M parameter budget can only accommodate up to 586 POD modes, whereas our low-rank O(Mr)O(Mr) factorization enables scaling to 4,096 modes at the same parameter count:

Mean relative error ↓NeNS² (4096 modes, ~100M)POD quadratic ROM (586 modes, ~100M)
ID velocity0.81%18.79%
ID vorticity2.76%38.04%
Leapfrogging OOD velocity1.36%18.26%
Leapfrogging OOD vorticity5.03%53.10%

NeNS² reduces vorticity error by 92.7% on ID and 90.5% on OOD flows, demonstrating the effectiveness of combining physically defined modal coordinates with low-rank, structure-preserving dynamics.

Side-by-side rollout videos comparing NeNS² with the POD-based structured ROM on both ID and leapfrogging trajectories are available in the supplementary web viewer under Structural analysis →\rightarrow POD ROM.

AuthorsOpenReview ↗

(2/2) Response to Reviewer DKP5's comments

(Continue)


3. Energy Conservation (W3)

Following this suggestion, we evaluate energy preservation over the original 120-step viscous rollouts by subtracting the energy loss attributable to each discrete viscosity update. This retains the original trajectories while measuring the cumulative energy defect of the nonlinear updates. We compare NeNS² with the no-skew and no-RK4 (Euler) variants, as well as coefficient-space FNO1D and U-Net1D baselines (see Fig. 7 in the revised manuscript).

Specifically, we define the viscosity-corrected cumulative energy deviation as

DN=EN−E0+∑n=0N−1LnviscE0,D_N=\frac{E_N-E_0+\sum_{n=0}^{N-1}L_n^{\mathrm{visc}}}{E_0},

where LnviscL_n^{\mathrm{visc}} is the energy loss caused by the actual discrete viscosity update at step nn. We report max⁡0≤N≤120∣DN∣\max_{0\leq N\leq120}|D_N| for each complete rollout.

The maximum absolute viscosity-corrected energy deviation is 9.02×10−69.02\times10^{-6} for NeNS², compared with 2.76×10−32.76\times10^{-3} for no-skew, 3.63×10−23.63\times10^{-2} for coefficient-space FNO1D, and 5.22×10−15.22\times10^{-1} for coefficient-space U-Net1D. The no-RK4 (Euler) variant becomes non-finite at t=0.36t=0.36 and is not assigned a full-rollout score. These results demonstrate substantially smaller non-viscous energy drift for NeNS² under the evaluated rollout setting.

The learned nonlinear generator conserves advective kinetic energy exactly in continuous time, while RK4 introduces an O(h5)O(h^5) local energy defect along smooth bounded trajectories. The viscosity-corrected diagnostic measures the accumulated energy defect of the nonlinear updates along the original viscous trajectories, including numerical integration and floating-point effects. Although finite-step energy conservation is not exact, the results show that the proposed structure substantially reduces energy drift in practice.


4. Boundary Conditions and Generalization (W4, Q2)

Our framework already supports irregular geometries and no-slip boundary conditions through the immersed-boundary method (IBM) [3]. As demonstrated in the single-cylinder and two-cylinder experiments in Section 4.4, NeNS² reproduces vortex deflection and transport around unseen obstacles without retraining. IBM imposes no-slip constraints through boundary forces, allowing us to handle complex obstacle geometries while retaining the original analytical Stokes basis.

For stricter enforcement of these constraints, an alternative is to compute geometry-adapted numerical Stokes eigenfunctions. Such a basis would satisfy the prescribed boundary conditions by construction, but computing thousands of eigenmodes on irregular domains can incur substantial computational overhead. We therefore adopt IBM as a practical trade-off between boundary accuracy and computational efficiency, while identifying geometry-adapted Stokes bases as a promising direction for future work.

Furthermore, free-slip boundary conditions have established applications in ocean-circulation modeling [4] and thermal-convection research [5]. They allow the study of bulk circulation and transport without resolving near-wall shear layers.

We have expanded the discussion in the revised manuscript (see Section 4.4 and Appendix E.7) to clarify the boundary treatment, computational trade-offs, and potential extensions.


Summary

We sincerely thank the reviewer for the thoughtful and constructive feedback. We have carefully addressed the concerns raised and provided additional experimental evidence to strengthen our claims. We hope these clarifications and new results will help address your reservations. We remain open to further discussion.


References

[1] Q. Cui, P. Sen, and T. Kim. "Scalable Laplacian Eigenfluids." ACM Transactions on Graphics, 37(4), Article 87, 2018. https://doi.org/10.1145/3197517.3201352

[2] P. Goyal, I. Pontes Duff, and P. Benner. "Guaranteed Stable Quadratic Models and Their Applications in SINDy and Operator Inference." Physica D: Nonlinear Phenomena, 483, 134893, 2025. https://doi.org/10.1016/j.physd.2025.134893

[3] K. Taira and T. Colonius. "The Immersed Boundary Method: A Projection Approach." Journal of Computational Physics, 225(2), 2118–2137, 2007. https://doi.org/10.1016/j.jcp.2007.03.005

[4] B. Deremble et al. "On the Application of No-Slip Lateral Boundary Conditions to 'Coarsely' Resolved Ocean Models." Ocean Modelling, 39, 411–415, 2011. https://doi.org/10.1016/j.ocemod.2011.05.002

[5] R. B. Rodakoviski and N. L. Dias. "Direct Simulation of Two-dimensional Bénard Flow with Free-slip Boundary Conditions." Computers & Fluids, 228, 105040, 2021. https://doi.org/10.1016/j.compfluid.2021.105040

Thank you for the additional experiments. I have increased my informativeness score to acknowledge the insights provided. The substep comparison supports the large-step claim in the tested setting, although its broader applicability remains unclear without physical timescale normalization and evaluation across different prediction intervals. I therefore maintain my overall rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.