acceptodds
Under review as a conference paper at ICLR 2027

Whitening the Spectrum: Preconditioned Flow Matching for Turbulence Generation

Abstract

Generative models of turbulence transport a spectrally flat source onto fields whose energy follows a steep power law, and the flow matching objective inherits that spectrum through the regression target and the interpolant's signal-to-noise ratio, so energy-containing scales dominate the gradient and dissipative scales are learned last. Our method, WSFM, instead solves transport in a whitened Fourier representation, an invertible, mode-wise affine transform estimated from the training ensemble that equalizes the latent's second-order statistics across wavenumbers, preserves the relative Fourier phases carrying coherent vortical structure, and leaves the modeled distribution unchanged. It therefore improves the conditioning of the optimization rather than altering the target, and we prove the whitened transport problem is uniformly conditioned. At matched backbone and parameter count, WSFM lowers relative and high-wavenumber spectral error against raw Fourier flow matching and retains the small-scale structure the raw coordinate smooths away. Across benchmark flows the gains concentrate in the physics, with structure-function exponents, enstrophy and dissipation statistics reproduced more faithfully than by the raw coordinate or the published surrogates, while pointwise accuracy stays comparable.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, turbulence modeling, spectral methods, scientific machine learning

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_s7PFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Whitening the Spectrum for Flow Matching (WSFM), a turbulence generation framework that applies an invertible, mode-wise affine normalization to Fourier coefficients before flow matching. The main motivation is that the strongly nonuniform energy spectrum of turbulent flows leads to imbalanced regression targets and frequency-dependent signal-to-noise ratios. By performing transport in spectrally standardized coordinates, the method aims to improve optimization conditioning and preserve small-scale turbulence statistics. The approach is implemented using a frequency-coordinate transformer and evaluated against deterministic neural operators, generative baselines, and several coordinate-system variants.

Critical strengths
  1. The paper identifies a meaningful frequency imbalance in turbulence generation and offers a simple representation-level intervention rather than introducing substantial architectural complexity.

  2. Under a controlled comparison, the method improves relative L2 error and high-frequency spectral accuracy while preserving the overall model capacity.

  3. The inclusion of spectral, vorticity, enstrophy, dissipation, structure-function, and distributional metrics provides a broader assessment than pointwise prediction accuracy alone.

Critical weaknesses
  1. The central contribution is not convincingly supported by the ablation results.

The main distinction of WSFM is its use of exact mode-wise whitening rather than radial whitening. However, Table 2 shows that radial whitening outperforms the proposed approach on eight of the nine reported metrics, including relative L2, vorticity, structure-function, enstrophy, and dissipation errors. Some of these differences are statistically significant, as reported in Appendix A.9.

The improvement in high-wavenumber spectral error over radial whitening is relatively modest, and the anisotropic spectrum improvement is not statistically significant. Consequently, the experiments establish the benefit of spectral normalization over raw Fourier coordinates, but do not establish the superiority of the proposed mode-wise formulation.

  1. The theoretical conditioning claims require clarification.

The propositions demonstrate equalization of marginal Fourier-mode variances. However, diagonal variance normalization does not generally whiten the full covariance matrix, as cross-frequency correlations remain.

In particular, the claim that an arbitrary Gaussian field becomes an independent standard complex Gaussian after per-mode normalization requires additional assumptions about its covariance structure.

Moreover, uniform marginal variances do not directly imply uniform conditioning of the neural flow matching optimization objective. The paper should distinguish representation-level variance equalization from guarantees concerning optimization dynamics.

  1. There is an inconsistency in the variance formulation.

Equation (2) expresses the variance of the target velocity using the second moment of the data Fourier coefficient. For nonzero-mean coefficients, the correct expression involves the centered variance defined in Equation (1). This distinction is particularly relevant because the proposed affine transformation explicitly removes nonzero ensemble means.

Questions for authors

What empirical evidence supports choosing mode-wise whitening over radial whitening, given the current ablation results?

Can the authors clarify the mathematical definition of conditioning used in the propositions and distinguish it from full covariance whitening and neural optimization conditioning?

How does WSFM compare with raw Fourier flow matching using both a spectrally matched source and inverse-variance loss weighting?

Can the authors provide exact pseudocode and numerical verification for the Hermitian projection?

How robust are the results to the estimation of whitening statistics, near-zero-variance modes, increased resolution, and different training seeds?

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

This paper studies spectral scale imbalance in generative modeling of turbulence. The authors argue that differences in Fourier-mode energy produce imbalanced flow-matching targets and signal-to-noise ratios, making small-scale structures difficult to learn. WSFM estimates a mean and standard deviation for each velocity component and Fourier mode from the training ensemble, performs flow matching in the standardized spectral coordinates, and reconstructs physical fields through the inverse transform. Using SpectralDiT, the paper evaluates prediction errors, spectra, and physical statistics on cylinder wakes and additional flow datasets. Whitening substantially improves over raw Fourier flow matching, although the proposed per-mode formulation does not consistently outperform simpler radial whitening.

Critical strengths

The method is simple and motivated by a relevant physical issue: large scales dominate energy, while small scales are important for vorticity and dissipation. Fixed spectral standardization is a reasonable and accessible way to address this imbalance. It requires neither a separately trained encoder nor lossy compression and can be incorporated as a representation change in an existing generative model. The distinction between normalizing spectral amplitudes and learning non-Gaussian structure also provides useful intuition about the remaining modeling task.

Critical weaknesses
  1. The experiments establish the benefit of whitening over raw Fourier coordinates, but do not establish the practical value of the proposed per-mode formulation over radial whitening. In Table 2, radial whitening achieves lower relative L2L_2 error, 0.190 versus 0.215, and lower dissipation error, 0.153 versus 0.210. Appendix A.9 reports statistically significant paired differences, and the vorticity metrics also favor radial whitening. Conversely, the anisotropic-spectrum advantage of per-mode whitening is not statistically resolved. I appreciate the authors’ acknowledgment that the per-mode formulation adds exactness, but its practical necessity and significance remain unclear. Section 6 additionally discusses equivalence to a spectrally matched source combined with per-mode loss weighting; the incremental contribution beyond these components needs clearer delineation.

  2. The theoretical account conflates mode-wise variance equalization with stronger whitening and optimization-conditioning claims. Equation (4) normalizes marginal Fourier-mode variances but does not generally remove cross-mode correlations. Consequently, Proposition 3 requires additional covariance assumptions to conclude that a Gaussian field becomes CN(0,I)\mathrm{CN}(0,I). Likewise, a condition number of one for the diagonal scale matrix does not establish uniform conditioning of general regression or neural-network optimization. Equation (2) also conflates variance with the second moment when modal means are nonzero: under Equation (1), the target-velocity variance is Dk2+1D_k^2+1, rather than E∣qk∣2+1\mathbb E|q_k|^2+1. I appreciate the narrower empirical interpretation in Section 5.2, but the abstract, propositions, and derivations need consistent assumptions and scope.

Questions for authors
  1. Given that radial whitening achieves lower relative L2L_2, dissipation, and several vorticity errors, which results establish a practical advantage of the per-mode formulation? How should the statement that the methods do not separate on bulk metrics be reconciled with the significant paired differences reported in Appendix A.9?

  2. Please identify the precise matrix or operator described as uniformly conditioned and state the required assumptions. Could the authors also provide the raw-coordinate source and weighted objective corresponding to the L2L_2 flow-matching loss actually used? Appendix A.5 derives the L1L_1 case, whereas the L2L_2 correspondence involves weights proportional to Dk−2D_k^{-2}; the derivation should be aligned with the experimental objective.

  3. Appendix A.4 states that deterministic models receive 20 genuine historical frames while generative models use single-frame guidance, but subsequently describes every model as observing the same single frame. Please specify the historical frames, physical parameters, guidance, and frame-replacement rules available to each model on each dataset. For the unguided foil and combustion settings, how are generated samples associated with particular reference trajectories for pointwise evaluation?

Reviewer_TXMhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a method to improve generative turbulence models. Motivated by the universal scaling law of the energy spectrum, the authors transform the prior distribution of a generative model (DiT as an example) to match the marginal distribution of a fully-developed turbulence spectrum. The authors argue to do this by normalizing the per-mode magnitude in the Fourier space directly, called "whiten" in this paper. Experimental results show that this simple scheme helps learned turbulence models some ways.

Critical strengths

The proposed method is simple and principled. Guided by the belief that generative model learns better when two distributions match closer, normalizing the prior distribution is a direct way to improve generative models.

Critical weaknesses
  1. The result is mixed. Across 4 test cases, WSFM is not the best model on both error metric AND statistical metrics, while in most cases, a simple FNO is competitive enough. It makes me question the effectiveness of the method despite its seemingly true argument.
  2. (minor) the writing needs to be improved. The way authors introduce their work is abrupt to most people in the ML community. The paper also contains paragraphs with visible AI-generated texts that are hard to read
Questions for authors
  1. Can you plot the raw energy spectrum curve? The method is built on spectrum but no spectrum is shown.
  2. Fourier-mode normalization is the most straightforward way to match the marginal distribution to a -5/3 curve. However, given that the improvement is not consistent, have you tried other variants? Examples could be prescribing spatial correlations over the random field.

Comments on the paper

1 comment

We sincerely thank all reviewers for their careful evaluation and constructive feedback. The concerns regarding mathematical rigor, presentation clarity, and experimental evidence are well taken. In particular, the theoretical assumptions and claims require greater precision, the writing needs clearer organization, and the experimental results require more careful analysis and interpretation. We will take these comments seriously and carefully revisit and polish the manuscript to improve its rigor, clarity, and overall quality.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.