acceptodds
Under review as a conference paper at ICLR 2027

Finite-Step Law Error and Learned Score Risk in Cahn–Hilliard–Cook Diffusion

Abstract

Physical simulators are increasingly used to supply training data for score-based generative models. A simulator can be stable and exactly conservative while still sampling the wrong distribution, and a score model trained on its output inherits that error. We give a quantitative account of this chain for Cahn–Hilliard–Cook (CHC) dynamics, from integrator to sampled law to learned score. For linearized CHC, the discrete Lyapunov equation gives the per-mode bias of an energy-stable IMEX step in closed form (up to 4.6× excess high-frequency variance), and a discrete fluctuation–dissipation covariance removes it at no cost. Under a shared Gaussian corruption, the excess target risk of learning the simulated law is a covariance-weighted relative Fisher divergence; we derive its exact fixed-time form and a risk-elasticity identity that predicts sub-quadratic growth at practical step sizes and a larger penalty for under- than over-dispersion. On a nonlinear 15 × 15 CHC Gibbs field at fixed physical time, coarser steps increase endpoint error and held-out target-DSM excess for a CNN and an FNO in all ten paired seeds, following the predicted pattern: law error ∝ h1.0 and risk exponents falling from about 1.9 to 1.5. Generation and inverse problem tests show where these gains carry over and where the sampler decides the outcome

Primary area:
generative models
Keywords:
score-based generative modeling, stochastic numerical integration, invariant-measure bias, denoising score matching, scientific machine learning, Cahn-Hilliard-Cook dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_v63sOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Diffusion models can be trained on data from physical simulators rather than real physical systems. This makes it easier to generate data, but one obvious downside is that simulation errors are inherited by (e.g.) score learning. Relatedly, constraints may or may not be exactly respected by the simulated data or downstream learned objects. The authors claim to quantitatively characterize this issue in the context of Cahn-Hilliard-Cook (CHC) dynamics.

Critical strengths

None

Critical weaknesses

This paper's major problem is that it is almost completely incoherent. It reads as almost completely AI-generated. There are a large number of mostly irrelevant comments, unimportant or conventional analyses presented with unnecessarily high detail, and few references. (There are many references in Sec. 2, but these are what some have called "drive-by" references: they are vaguely related but not meaningfully engaged with. This is different than deeply and specifically engaging with some aspect of any one reference.)

It is extremely difficult to make any sense of the organization of this paper. It is frustrating to read, and even with sincere effort, it is not clear to me how the story promised in the abstract maps onto what is actually in the main text of the paper. I strongly believe the authors, if they are invested in this work, should rewrite it from scratch.

Relatedly, there are many results, but some of them are clearly trivial (e.g., analyses of OU processes) and some of them may not be. Which results matter is not clearly indicated by the authors. The significance of their results is not clear either. Is the outcome a concrete recommendation about how to maintain constraints and/or minimize error? Is it a quantitative assessment of the statement that errors propagate? Why should someone care beyond this specific analysis of CHC dynamics?

Questions for authors
  1. Which results are most important and why?
  2. What are the practical consequences of this work?
Reviewer_8gZTOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper discusses score-based models trained on data samples produced by a physical simulator. In particular, the simulator for Cahn-Hilliard-Cook (CHC) dynamics is considered. A certain risk of diffusion score matching loss appears to be analyzed in terms of a step size. Experiments on simulated data of a periodic 15 ×\times 15 field appear to be conducted to verify the theoretical results.

Critical strengths

The topic of investigating score-based models based on physics is an interesting and important research direction.

Critical weaknesses

The overall writing quality might require significant improvement for soundness and readability. Several examples are listed below.

  • The background section lacks sufficient explanation to contextualize the research. Both explanations of score-based models and CHC dynamics appear insufficient for the main discussion. In addition, some important equations such as Eq. (1) and λ0=σ0=0\lambda_0 = \sigma_0 = 0 in Line 207 are used without derivation or references.
  • Some notations are introduced without prior definition, which makes it difficult to follow the mathematical formulation. For example:
    • R^ref\hat{\mathcal{R}}_{\mathrm{ref}} in Line 193.
    • The physical quantities in Line 201, such as kk, ahigha_\mathrm{high}, κ\kappa and σk\sigma_k.
  • The motivation of this paper may lack sufficient explanation. This paper appears to consider a situation where a physical simulator generates training data samples following a certain distribution and a score-based generative model is trained on those samples. However, the motivation and advantages of considering such a situation are not clearly explained. While the Related Work section presents some works that modify the forward and backward processes of typical score-based models inspired by physics and dynamical systems, the problem setup of this paper, according to Section 4.1, appears to be that training data X are drawn from simulated distributions π\pi or π~h,T\tilde{\pi}_{h,T} and the score-based model utilizes the typical variance-preserving process. Thus, there appears to be a mismatch between the problem setup and the related work, and the motivation of this paper is not well supported by the related work.
  • While several concepts are presented, the logical connections between them appear unclear, especially in Sections 5 and 6, leaving the overall explanation of the proposed method fragmented and disjointed.
  • Figure 1 and Figure 31 appear to be identical, suggesting that the manuscript has not been sufficiently proofread.
Questions for authors

To clarify the discussion, could further explanation of the motivation and purpose of this paper, the key takeaways of Proposition 1, and what is examined in the experiments be provided?

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

The paper examines how error propagates when physical simulators are used to supply training data for score-based generative models, focusing on Cahn-Hilliard-Cook dynamics, deriving theoretical bounds with experimental confirmation.

Critical strengths

Applying diffusion models to physical problems is a valuable research directions, and this paper considers the simulation part of the pipeline.

Critical weaknesses

It's unclear what the takeaways are from considering this highly specific model, with a highly specific discretization scheme. The paper is concerned with technical, over-specific calculations for the model, without giving a clear interpretation of what it means for training from physical simulators more generally. Each part of the error analysis could be different for a simulation of a different system, so the value of this end-to-end result for this model is unclear.

The exposition does not sufficiently set up and motivate the problem, and seems overly technical. The main Proposition 1 is difficult to understand and poorly interpreted, involving many quantities, while other results are not clearly demarcated in the text. It is unclear what is standard numerical analysis, and what are the novel contributions. The takeaways from the collection of experiments are also unclear.

Questions for authors

What lessons can be learned from this example more generally for training diffusion models with physical simulators, and why would they be justified from this one specific example?

What concrete takeaways for practitioners, for example, in allocating simulation and training resources?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.