acceptodds
Under review as a conference paper at ICLR 2027

SimDFKs: Divergence-Free Kernels as Spatial Representations for Flow Simulations

Abstract

Fluid dynamics has posed enduring challenges in computational physics and computer graphics, where resolving rich, multiscale flow structures demands spatial representations that are both expressive and computationally efficient. Recent studies have shown that kernel-based spatial representations, which approximate flow fields as superpositions of localized basis functions, can outperform implicit neural representations (INRs) in both flow reconstruction and simulation while retaining a compact continuous parameterization. In this work, we extend divergence-free kernels (DFKs) from flow reconstruction to forward fluid simulation, introducing a key reformulation that makes evolving the continuous representation under PDEs practical and stable. Building on this representation, our optimization-based framework evolves the kernel parameters under advection, external forces, and complex boundary conditions while preserving incompressibility by construction. Our reformulation retains the compactness and expressiveness of the original DFK formulation while substantially improving the efficiency and numerical accuracy of derivative-sensitive PDE optimization. Without requiring training on precomputed data, our method produces high-quality forward simulations of a variety of 2D and 3D fluid phenomena from given initial conditions. The same framework also enables simulation recovery, i.e., reconstructing a fluid-dynamics-consistent flow field sequence from sparse vorticity observations of results produced by an existing solver. This capability is not directly available to conventional forward numerical solvers.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Fluid Simulation, Incompressible Flows, Divergence-Free Kernels

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_E16bOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SimDFKs, a fluid simulation method based on divergence-free kernels (DFKs). By optimizing the parameters of continuous kernels, the method simulates fluid motion while guaranteeing incompressibility by construction. It also improves vortex preservation, simulation accuracy, and representation efficiency.

Critical strengths
  1. The paper is well-written, clear, and easy to follow.

  2. The paper addresses a meaningful and interesting research problem.

  3. The proposed Single-Curl formulation improves upon the previous Double-Curl design and extends DFKs from flow reconstruction to forward simulation. The authors combine it with existing vorticity transport and optimization techniques, and demonstrate its effectiveness through extensive experiments.

Critical weaknesses
  1. High computational cost. While SimDFKs shows good accuracy and incompressibility, I am concerned about its practical value. In the Taylor–Green experiment, it is about 15,000 times slower than the Eulerian solver. Could a higher-resolution grid-based solver achieve similar or better accuracy with the same computational budget? I suggest adding a fair accuracy–runtime comparison to better demonstrate the advantages of SimDFKs.

  2. Limited evaluation of physical consistency. Although the paper evaluates incompressibility and smoke mass preservation, a more comprehensive analysis of physical consistency, particularly long-term energy and circulation conservation, would further strengthen the work.

  3. Concerns about sparse recovery. In Appendix A.6, the 10% observation experiment still uses full reference data for interpolation and normalization. Can the method achieve similar results using only the sparse observations? I suggest adding an experiment without this extra information.

Questions for authors

See Weaknesses.

1 comment

We sincerely thank the reviewer for the positive assessment of the paper's clarity, the single-curl reformulation, and the extension of DFKs to forward simulation. Following the reviewer's suggestions, we added several experiments to better characterize the accuracy–cost trade-off and sparse-recovery setting; these changes are summarized in the Revision Cover Letter above.

  1. Accuracy–runtime trade-off. Using the taylor vortex example, we add a comparison against the Eulerian method with a 1281×12811281\times 1281 MAC-grid, which takes up more computational time than our method. The measured last-frame vorticity MSE (Table 8) and visualized vorticity sequences (Figure 8) show that our method achieves higher accuracy than Eulerian method under same time budget.

  2. Energy and circulation. We do not claim exact energy or circulation conservation, since the optimization and reinitialization steps introduce approximation errors. Our 3D leapfrogging experiment instead provides qualitative evidence of long-term preservation of coherent vortex dynamics through repeated interactions. We have clarified the scope of this claim in the revision.

  3. Sparse recovery. We repeat the experiment without full-reference interpolation or normalization (Table 2). Filtering and normalization use retained observations only, while complete reference fields are reserved for post-recovery evaluation. The revised results still show substantial improvement over independent fitting. After 30 rounds, the four reported errors are 7.786×10−37.786\times 10^{-3}, 7.477×10−37.477\times 10^{-3}, 1.715×10−41.715\times 10^{-4}, and 9.756×10−49.756\times 10^{-4}, compared with independent-fit errors of 2.137×10−12.137\times 10^{-1}, 1.589×10−11.589\times 10^{-1}, 8.151×10−38.151\times 10^{-3}, and 4.637×10−24.637\times 10^{-2}. The stricter protocol slightly increases the errors, but the substantial improvement over independent fitting remains, confirming our original conclusion under a more rigorous and transparent evaluation.

We greatly appreciate the reviewer's suggestions, particularly the request for a more direct accuracy–runtime comparison and a stricter sparse-recovery protocol. We hope these additional experiments and clarifications address the reviewer's concerns and provide a more complete and balanced evaluation of SimDFKs. We would be happy to further discuss any remaining questions.

Reviewer_RnE4OpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SimDFKs, a kernel-based representation for incompressible fluid simulation. The proposed single-curl formulation guarantees divergence-free velocity fields while reducing the derivative order required for vorticity optimization. The method advances fluid dynamics through semi-Lagrangian vorticity transport and framewise kernel optimization, with additional boundary constraints and periodic reinitialization. It also supports flow recovery from incomplete vorticity observations. Experiments demonstrate promising results in vortex preservation, boundary handling, smoke simulation, and flow recovery.

Critical strengths

S1. Physically consistent representation. The single-curl formulation guarantees divergence-free velocity fields without additional divergence penalties or pressure projection.

S2. Efficient reformulation. Reducing the derivative order improves optimization efficiency and boundary fitting, as supported by the ablation results in Tables 3 and 4.

S3. Broad applicability. The framework supports both 2D and 3D simulations, including vortex dynamics, moving boundaries, external forces, and flow recovery.

S4. Promising experimental results. The method achieves competitive accuracy with compact representations. The piston experiment particularly demonstrates its advantage in preserving incompressible flow.

Critical weaknesses

W1. Limited validation of sparse recovery. As acknowledged in Section 5.3 and Appendix A.6, spatial recovery uses unobserved reference values for interpolation and full reference statistics for normalization. Thus, Figure 2 and Table 2 do not establish recovery solely from the retained observations. This concern does not extend to the separate temporal recovery experiment.

W2. Different boundary treatments. SimDFKs uses analytic exterior continuation, whereas the Eulerian baseline uses open boundaries (Appendices A.3 and B.3). This difference may affect the accuracy comparison in Table 1.

W3. Missing comparison and discussion of DDFKs. DDFKs (Xing et al., 2026) already applies divergence-free kernels to forward simulation. The paper should clarify its distinct contributions. Tables 3–4 isolate the kernel formulation but do not compare the complete methods. A direct comparison would be useful if official code is available; otherwise, a detailed technical discussion is appropriate. (DDFKs: Fluid Simulation with Dynamic Divergence-Free Kernels)

W4. Inconsistent visual comparisons. Figure 10 compares NMC at t=7.6t=7.6 with other methods at t=10t=10, using different color scales. These choices weaken direct visual comparison.

W5. Limited quantitative evaluation in 3D. Figures 9 and 13 mainly provide visual results. Additional leapfrogging events alone do not establish greater accuracy. Quantitative measures of vortex dynamics and smoke preservation would strengthen the evaluation.

W6. High computational cost. Table 1 reports about 23 seconds per frame for SimDFKs versus 0.00155 seconds for the Eulerian baseline. Hardware and accuracy differ, but the cost remains substantial. Runtime and peak memory measurements for 3D simulations are also needed.

Questions for authors

Q1. Can Figure 2 and Table 2 be repeated using only observed values for interpolation and normalization? How much improvement over independent fitting remains?

Q2. Can Table 1 use consistent boundary treatments, or quantify the effect of the current differences?

Q3. What distinguishes SimDFKs from DDFKs in representation and simulation procedure? Please provide a direct comparison if official code is available. If no code is available, a technical comparison is sufficient; a complete reimplementation is not required.

Q4. Can Figure 10 use common color scales and matching physical times? If NMC stopped earlier because of instability, please clarify.

Q5. Can the authors report quantitative 3D results, runtime per frame including reinitialization, and peak GPU memory? Relevant metrics include vortex trajectories, circulation and energy drift for vortex rings, and smoke mass and boundary flux for smoke.

3 comments

We thank the reviewer for the detailed questions and constructive suggestions. Several of these comments directly motivated revised experiments and a more transparent presentation, as summarized in the Revision Cover Letter.

  1. Sparse recovery. We repeat the experiment without full-reference interpolation or normalization (Updated in Table 2). Filtering and normalization use retained observations only, while complete reference fields are reserved for post-recovery evaluation. The revised results still show substantial improvement over independent fitting. After 30 rounds, the four reported errors are 7.786×10−37.786\times 10^{-3}, 7.477×10−37.477\times 10^{-3}, 1.715×10−41.715\times 10^{-4}, and 9.756×10−49.756\times 10^{-4}, compared with independent-fit errors of 2.137×10−12.137\times 10^{-1}, 1.589×10−11.589\times 10^{-1}, 8.151×10−38.151\times 10^{-3}, and 4.637×10−24.637\times 10^{-2}.

  2. Taylor–Green boundaries. We disabled reinitialization and exterior extrapolation for our method, use the analytic solution only for initialization, and use closed free-slip boundaries for the Eulerian baseline. This removes the boundary-treatment and extrapolation differences raised by the reviewer. We updated the corresponding results in Table 1. The revised Eulerian baseline improves substantially under the corrected setup, while our main conclusion remains unchanged: SimDFKs still achieves the lowest error among the compared methods.

  3. DDFKs. DDFKs is currently a non-peer-reviewed arXiv preprint. We therefore respectfully ask that DDFKs not be treated as an independent peer-reviewed prior work that diminishes the contribution of this submission. We nevertheless agree that clarifying the technical relationship is useful. DDFKs uses the original double-curl velocity formulation, whereas SimDFKs uses a single-curl formulation that reduces the derivative order required for vorticity fitting. Tables 3–4 provide controlled formulation-level comparisons under matched simulation settings.

  4. Kármán visualization. NMC is shown at frame 152 because it becomes numerically unstable afterward; we now clarify this explicitly in Appendix C.2. The method-specific color ranges are used to make the vortical structures of each method visible, and we clarify that this figure is intended as a qualitative comparison rather than a common-scale quantitative evaluation.

  5. 3D evaluation and resources. We now report end-to-end running times of 718 and 797 seconds per frame for the 3D leapfrogging-vortex-ring and smoky-sphere experiments, respectively, including scheduled reinitialization. We do not claim exact energy or circulation conservation, since the optimization and reinitialization steps introduce approximation errors. We use the number of successful leapfrogging interactions as qualitative evidence of longer preservation of coherent vortex dynamics, rather than as a standalone quantitative measure of accuracy.

We appreciate that these comments prompted us to make the experimental protocols and the scope of our claims considerably more explicit. We hope the revised experiments and clarifications address the reviewer's main concerns, and we would be happy to discuss any remaining questions.

I thank the authors for their response and revisions.

However, the computational cost of approximately 12–13 minutes per frame for the 3D experiments associated with a 1283128^3 grid remains substantial, raising concerns about the practical applicability of the proposed method.

Furthermore, DDFKs is indeed a closely related work. Although it has not yet undergone peer review, I believe it deserves a technical discussion and comparison.

Considering the authors' efforts to address my concerns, I am willing to raise my rating from 2 to 3.

We thank the reviewer for the thoughtful follow-up and continued consideration of our work. We greatly appreciate the constructive comments and are encouraged by the updated assessment.

We agree that the reported 3D runtime of 718–797 seconds per frame remains a significant practical limitation. Reducing the cost of repeated optimization is an important direction for future work. At the same time, we believe the results support continuous divergence-free representations as a promising alternative direction for fluid simulation, with computational efficiency remaining an important open challenge. In the next revision, we will make this limitation more explicit in Section 6 and further clarify the end-to-end timing discussion alongside the reported 3D results.

We also appreciate the reviewer's suggestion to discuss DDFKs more thoroughly. In the next revision, we will expand the technical comparison between DDFKs and SimDFKs, covering both the representation and simulation procedure, and more clearly delineate the scope of our contributions.

Thank you again for the thoughtful feedback and for helping us further improve the manuscript.

Reviewer_Vt41OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper extends divergence-free kernels to forward fluid simulation using a single-curl velocity representation and per-step optimization to fit transported vorticity and boundary conditions. The method is evaluated on 2D and 3D flows and on sequence recovery from incomplete vorticity observations, with additional reference information used in the spatially sparse experiment.

Critical strengths
  1. The paper demonstrates that DFKs can serve as a continuous spatial representation for forward fluid simulation. The single-curl formulation is well motivated for vorticity fitting, and the ablation shows reduced evaluation cost and improved fitting compared with the double-curl formulation. I consider this a useful extension of kernel-based flow representations beyond reconstruction, even though its advantages over conventional numerical methods require further justification.

  2. The validation covers several relevant fluid phenomena, including 3D vortex stretching and moving solid boundaries. These results support the value of enforcing incompressibility within the representation and make the approach worth sharing with the community.

Despite my concern about the computational justification, I believe the extension to forward simulation and the demonstrated results constitute a useful contribution. Therefore, I support acceptance, while encouraging the authors to clarify when their optimization-based recovery should be preferred over conventional vortex methods.

Critical weaknesses

My main concern is that the computational benefit of using DFKs for forward simulation is not sufficiently demonstrated. The framework is conceptually close to conventional vortex methods: it transports vorticity and then recovers a velocity field consistent with that vorticity and the boundary conditions. The main distinction lies in the recovery stage, where the authors optimize DFK parameters instead of using a conventional Poisson-based or Biot–Savart-based formulation. Constructing velocity as a curl guarantees analytical incompressibility, but this mechanism is also available in traditional vector-potential formulations. That property alone does not establish why the proposed recovery method should be preferred.

The kernel optimization therefore takes on the velocity-recovery work that a conventional solver performs through a numerical inversion, at the expense of thousands of Adam iterations per step. The single-curl ablation demonstrates an improvement within the DFK family, but does not establish the benefit of this replacement over mature vortex solvers. Likewise, the qualitative comparisons with vortex-based methods do not isolate the recovery stage or establish its accuracy–cost trade-off. Compact storage is valuable, yet the current results leave it unclear when that benefit outweighs the cost of recovering velocity through nonlinear fitting.

I find the motivation for DFKs more convincing in prior flow-reconstruction work, where a divergence-free function space directly helps fit physically consistent velocities to incomplete or indirect observations. For forward simulation, an established vorticity-to-velocity solver already provides a natural alternative. This makes me less enthusiastic about the present extension, although I still regard the demonstrated simulation capability as a useful contribution.

Questions for authors
  1. Could the authors clarify what practical advantage DFK-based velocity recovery provides over a conventional Poisson-based or Biot–Savart-based recovery under comparable initial conditions and boundary conditions? A controlled comparison on a representative existing example, keeping the transport scheme as similar as possible and measuring velocity error, boundary error, and runtime, would help determine whether the proposed representation offers an accuracy–cost benefit. Evidence of such a benefit would strengthen my assessment.

  2. How much of the per-step runtime is spent constructing transport targets versus optimizing the kernel parameters, and how does reducing the optimization budget affect the resulting velocity and boundary errors? This would help clarify whether the reported cost is essential to the demonstrated accuracy or whether a substantially cheaper operating point is available.

1 comment

We greatly appreciate the reviewer’s questions. As summarized in the Revision Cover Letter, we added several experiments and analyses in response to the reviews. In particular, the suggestions to compare against conventional velocity-recovery procedures and to characterize the optimization–cost trade-off helped us more clearly present both the strengths and current limitations of continuous DFK representations.

  1. Advantage over Poisson- or Biot–Savart-based recovery. A limitation of conventional vorticity-to-velocity recovery is that vorticity alone does not always uniquely determine the velocity field. In a multiply connected domain, the harmonic component can vary without changing either vorticity or divergence, and incorrect recovery of this component can lead to erroneous flow dynamics. We added a vortex-passage experiment (Lines 412–422 and Fig. 15) demonstrating this issue: the Biot–Savart vortex-particle baseline fails to maintain the required harmonic component, whereas SimDFKs directly evolves the velocity representation and avoids this artifact.

  2. Runtime breakdown and optimization budget. We added a per-step optimization-budget study on the Taylor-vortex example (Line 487 and Table 5). Transport-target construction accounts for 19.7% of the total advance time on average, with the remaining cost dominated by kernel optimization. The results in Sec. 5.4.2 further show that reducing the iteration budget can substantially lower computation with moderate accuracy degradation. For the main forward-simulation experiments, however, we use a common budget of 5,000 iterations per step to ensure sufficient convergence and avoid scene-specific tuning.

We believe these additional experiments and analyses provide a more balanced characterization of when continuous DFK representations are useful and where their current computational limitations remain. We hope they address the reviewer’s concerns, and we would be happy to further discuss any remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Cover Letter

We sincerely thank all three reviewers for their careful reading, constructive criticism, and recognition of the potential of divergence-free kernels as a continuous representation for fluid simulation. We also appreciate their positive assessment of our extension from flow reconstruction to forward simulation, the motivation for the single-curl reformulation, and the range of fluid phenomena demonstrated by our framework. Their feedback helped us further improve the experimental protocol and transparency of the manuscript, while more clearly articulating the scope and practical role of SimDFKs. We have submitted a revised manuscript (PDF) with all changes highlighted in blue and an updated supplementary webpage (HTML). The main revisions are summarized below.

  1. Accuracy–runtime comparison with high-resolution Eulerian solver. (Vt41, E16b)

    We added an Eulerian baseline using a 1281×12811281\times 1281 MAC grid to the Taylor-vortex experiment, whose running time is comparable to that of SimDFKs (Line 337 and Table 8). We enabled CFL-controlled time stepping and used closed free-slip boundaries for both Eulerian configurations. Under a comparable computational budget, SimDFKs still achieves lower final-frame vorticity error than the high-resolution Eulerian baseline.

  2. Optimization-budget study. (Vt41)

    We added a controlled study on the Taylor vortex by varying the number of Adam iterations per time step from 1,000 to 5,000 (Line 487 and Table 5 in the revised manuscript). We report the mean and final vorticity MSE, mean absolute boundary-normal velocity error, and total advance time. This experiment quantifies the accuracy–cost trade-off and shows that reducing the optimization budget can substantially reduce computation while retaining reasonable simulation accuracy.

  3. Comparison with conventional vortex-based recovery. (Vt41)

    We added a vortex-passage experiment comparing SimDFKs with a Biot–Savart vortex-particle baseline (Line 412-422 and Fig. 15 in the revised manuscript). The experiment highlights a complementary property of the continuous DFK representation: SimDFKs maintains the harmonic component required by the boundary-conditioned velocity field, whereas the Biot–Savart baseline does not.

  4. Updated Taylor–Green comparison. (RnE4)

    We revised the Taylor–Green comparison so that the evaluated methods use consistent closed free-slip boundaries (updated in Table 1). For SimDFKs, we additionally disabled reinitialization and exterior extrapolation in this experiment. We also clarified the evaluation locations, solver settings, hardware, and timing scope to make the comparison more transparent. The overall conclusions remain unchanged.

  5. Stricter spatially sparse recovery protocol. (RnE4, E16b)

    We repeated the spatially sparse recovery experiment under a stricter observation-only protocol. The observation mask and stored observation array now use the same nearest-mode discrete lookup, and only retained observations can enter an observation-conditioned target. The amplitude scale used for normalization is computed from retained observations only, while complete PFM fields are reserved for post-recovery evaluation. Under this revised protocol, we observe only a minor increase in recovery error, while retaining the substantial improvement over independent per-frame fitting (updated in Table 2).

  6. 3D computational cost. (RnE4)

    We added the average running times of both 3D experiments, including all per-frame computation and scheduled reinitialization. The 3D leapfrogging-vortex-ring and smoky-sphere experiments require 718 and 797 seconds per frame, respectively.

  7. Clarified scope and positioning (Vt41, RnE4, E16b)

    We revised the discussion to more clearly position SimDFKs as an exploration of continuous representations for fluid simulation, rather than a replacement for mature discrete solvers. Continuous representations offer complementary advantages in compactness, differentiability, spatial adaptivity, and inverse optimization, while currently remaining substantially more expensive in forward simulation. They also exhibit different numerical trade-offs: for example, SimDFKs does not inherit the same grid-resolution-dependent CFL stability restriction as many explicit grid-based schemes, allowing substantially larger timesteps in practice. The reviewers' comments, particularly on computational cost and comparisons with established solvers, helped us present these advantages, limitations, and remaining challenges more fairly and comprehensively. We believe that exploring this alternative representation paradigm, together with a transparent characterization of its trade-offs, provides a valuable direction at the intersection of fluid simulation and machine learning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.