acceptodds
Withdrawn from ICLR 2027

Energy-Guided Functional Flow Matching for Soliton Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We study energy-guided Functional Flow Matching for generating solutions of the Korteweg–de Vries equation. A Fourier Neural Operator parameterizes a velocity field that transports a Gaussian reference measure toward a distribution of soliton solutions. We consider both complete space–time trajectories and joint initial–final state representations. Under explicit measure-theoretic assumptions, we derive an exact guidance identity in function space and a tractable covariance-based approximation for sampling. On the reported two- and three-soliton forecasting benchmarks, FFM with initial-condition guidance achieves lower median trajectory errors and PDE residuals than our DiffusionPDE and FunDPS implementations. Additional invariant guidance reduces momentum and energy drift, while its effect on prediction error depends on the soliton regime. These results distinguish conservation from dynamical accuracy and support guided functional generation as a promising approach to dispersive PDEs. The experiments assess performance on a common grid.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Functional Flow Matching, Generative Models, Korteweg-de Vries, Solitons

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_c7QTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This research studies energy-guided Functional Flow Matching (FFM) for KdV soliton generation, using a Fourier Neural Operator(FNO) to learn an unconditional velocity field from either full space–time trajectories or initial–terminal state pairs. The theoretical contributions include an exact function-space reweighting formulation under explicit assumptions, computing covariance-based guidance approximation via Jacobian–vector products, and an error bound for general FFM. The experiments show that initial-condition-guided FFM outperforms the implemented DiffusionPDE and FunDPS baselines. The joint representation also demonstrates low dynamical inconsistency and agreement across finite grids.

Critical strengths

The mathematical part is solid. The function-space guidance has an explicit, verifiable mathematical foundation in the report, which provides a technical foundation for using the existing guidance methods in function space.

The experiments are clear and comprehensive, showing the need for the individual components. The experiments cover the main components and clearly expose the tradeoff between conservation and prediction accuracy, making each design choice easy to understand.

Critical weaknesses

The method is incremental. The main additions are Hilbert-space derivations and adaptation to KdV solitons. But the overall method is a combination and extension of existing generation and guidance methods, without a sufficiently substantial new algorithmic idea.

A practical advantage over traditional numerical solvers for KdV has not been established. The paper studies one-dimensional soliton systems with known governing equations, for which full-initial-condition forecasting can be solved directly by mature numerical methods. The generative methods instead require additional data-generation, training, and iterative guided-sampling costs, and seems without quantifying corresponding gains in accuracy, efficiency, or conditional inference.

Questions for authors

See weaknesses above.

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

The paper extends energy-guided flow matching from finite-dimensional spaces to function spaces (Hilbert setting) and applies it to KdV soliton generation (final state generation). It derives an exact guidance identity (Theorem 1) and a covariance-based approximation (Eq. 15–17) in the infinite-dimensional setting, building on Functional Flow Matching (Kerrigan et al., 2024) and the finite-dimensional guidance framework of Feng et al. (2025).

Critical strengths

Generalizing finite dimensional flow-matching methods to infinite dimensional one and applying it to the generation of PDE solutions.

Critical weaknesses

The “discretization-invariance” claim is not empirically supported. The paper’s central selling point is that functional flow matching operates in function spaces and is therefore discretization-invariant. However, all experiments are conducted on a fixed 64×64 grid (or fixed 256-point spatial grid). Figure 4 claims to show “spatial refinement” consistency, but the model is trained on one discretization without querying on different grids at test time. This does not constitute a test of discretization invariance. To substantiate the claim, the authors would need to train a model on mixed resolutions and evaluate on a held-out resolution (e.g., 512×512) that was never seen during training. Without this, the functional-space framework is a theoretical claim rather than an empirically validated advantage.

The theoretical contribution is essentially a restatement of Feng et al. (2025) in Hilbert-space language, the “discretization-invariance” claim is not tested in a meaningful way, and the inconsistent effects of invariant guidance are left unexplained.

Questions for authors

See Critical Weaknesses.

Reviewer_AYt6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies energy-guided Functional Flow Matching (FFM) for Korteweg–de Vries soliton generation. It uses an FNO velocity field and considers two representations: complete space–time trajectories and joint initial–terminal states. The methodological component extends an energy-based guidance formulation to a Hilbert-space setting and develops a covariance-based approximation. Experiments examine unconditional generation, initial-condition forecasting, invariant guidance, and spatial-discretization consistency, with comparisons against implementations of DiffusionPDE and FunDPS.

Critical strengths

The evaluation distinguishes forecasting accuracy, PDE residuals, and conservation drift. This distinction is useful: the results demonstrate that improving invariant preservation does not necessarily improve the predicted dynamics. Appendix A.2 also provides an explicit illustration of this issue.

The joint initial–terminal representation produces physically consistent endpoint pairs in the reported experiments. Table 3 reports median consistency errors below 0.9%, providing evidence that the model learns a meaningful relationship between the two endpoints within the studied soliton distributions.

Critical weaknesses
  1. The baseline comparisons are substantially confounded and do not establish a reliable methodological advantage.

Appendix H explicitly states that trajectory FFM is trained for 100 epochs on 20,000 examples, whereas DiffusionPDE receives only 96,000 example presentations over 6,000 updates. This corresponds to approximately 2,000,000 versus 96,000 training-example presentations, a difference of more than twentyfold. Equal training exposure is not the only possible fairness criterion, but the paper provides no controlled comparison that separates the effects of training budget, architecture, and model family.

The FunDPS protocol raises additional concerns. Appendix I describes a target training budget “if the run is completed,” rather than clearly documenting the completed budget associated with the reported results. It also explicitly states that the observation-guidance coefficient is a starting value that has not yet been selected by validation. Consequently, the reported FunDPS results do not establish the performance of a sufficiently tuned baseline.

These issues are consequential because the comparisons in Table 2 provide much of the empirical support for the proposed approach. The authors acknowledge part of the budget mismatch, which is helpful, but acknowledgment does not resolve the evidential limitation. Matched-budget comparisons, documented completed runs, and consistently validated guidance parameters are needed before attributing the observed improvements to the proposed methodology.

  1. The incremental contribution and the practical value of the proposed guidance are insufficiently established.

The method combines established FFM, an FNO parameterization, and an energy-guidance construction. Section 2.4 explicitly follows Feng et al. (2025), and Equation (17) is described as the infinite-dimensional analogue of their covariance guidance. A Hilbert-space extension can be meaningful, but the paper needs to explain more precisely which new theoretical obstacle is resolved and why that extension materially improves the implemented solver.

The connection between the continuum theory and the practical algorithm remains incomplete. Section 2.5 explicitly excludes a convergence claim for the discretized guided sampler, while the covariance approximation is introduced without an error bound controlling its posterior-weighted Taylor remainders. Thus, the exact guidance identity does not by itself establish the accuracy of the implemented approximation.

The experimental evidence for the title’s central guidance component is also mixed. In Table 2, adding invariant guidance increases the two-soliton trajectory error from 0.27838 to 0.29378 and the final-state error from 0.43526 to 0.45491. For three solitons, the trajectory error decreases only from 0.24227 to 0.23985. Although momentum and energy defects improve, these results do not establish a consistent forecasting benefit. The reported sample medians also do not quantify variability across independently trained models.

Finally, validation is restricted to separately trained, low-order soliton datasets for a single integrable PDE. The introduction devotes substantial space to general generative-model background, but does not sharply articulate the methodological gap relative to the closest FFM and guidance approaches. Taken together, the theoretical positioning and empirical evidence do not sufficiently establish a contribution of the significance expected for acceptance.

Questions for authors
  1. Can you provide comparisons under matched training exposure or matched computational budgets, with completed training budgets and validation procedures documented for every method? In particular, was the reported FunDPS checkpoint fully trained, and how sensitive are its results to the currently unvalidated guidance coefficient?

  2. What specific theoretical result distinguishes the proposed guidance from a Hilbert-space extension of Feng et al. (2025)? Which assumptions required by that result are satisfied by the actual observation and invariant functionals used in the experiments?

  3. Does invariant guidance provide a reproducible forecasting improvement across independently trained models? Please quantify the variability of the Table 2 results and explain the deterioration in the two-soliton case.

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

Related papers

Discussion (0)

Sign in to comment.