acceptodds
Under review as a conference paper at ICLR 2027

Numerically Converged, Physically Wrong: Reference Validity for Samplers of Diffusion Physics Emulators

Abstract

Diffusion-based emulators of physical systems integrate a learned reverse process for every frame of a rollout, and cheaper samplers are judged by how closely they reproduce a fine-step reference. We ask when this yardstick can be trusted. First, we make it precise: a local error budget is the largest reverse step, at each noise level, whose endpoint stays within a tolerance of the reference. Measured with a censoring-aware protocol on pretrained autoregressive conditional diffusion models, budgets predict held-out step failures, transfer to rollout states, and follow an accuracy-limited rather than CFL-like law. Second, in a frozen evaluation on held-out physical conditions with 200-step rollouts scored against ground truth, fidelity to the reference ranks 20 coarse schedules like physical accuracy on transonic cylinder flow (Spearman and for two checkpoints) but in reverse on isotropic turbulence ( and ). The inversion appears only after about 50 frames, is absent over the upstream 100-frame evaluation horizon, and has a specific cause: the deterministic reference is numerically converged yet retains less than half of the true kinetic energy. On the same time grids, stochastic updates lower the physical error from to , match the model's native sampler, and restore the agreement. On an accurate incompressible-wake emulator, all samplers sit within 0.014 of the reference and the choice is moot. Budget-based schedule selection, frozen before testing, ties but does not beat direct pilot evaluation at a matched calibration budget. Before fidelity to a reference guides deployment, the reference itself must be validated on physical observables over the deployment horizon.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific machine learning, diffusion models, numerical sampling, error estimation, timestep selection, turbulence

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_RpTsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper asks whether a diffusion-based physics emulator can be trusted, where the criterion is whether numerical convergence implies physically accurate generation. Through experiments on three datasets, the authors show that numerical fidelity and physical accuracy can be poorly correlated depending on which dataset they are testing. They further propose a local error-budget framework for sampler selection.

Critical strengths
  • The question of whether numerical convergence implies physical correctness is important. This gives suggestions of whether a diffusion-based simulation model can be trusted.
  • The comparison between deterministic and stochastic sampling provides useful empirical evidence.
Critical weaknesses
  • The experiments cover only three datasets and a small number of checkpoints from a single model family. It remains a question whether the results generalize to other models and physical systems.
  • The practical benefit of the proposed error-budget framework is unclear.
Questions for authors
  • What practical benefit does the proposed error-budget framework offer?
  • How robust are the findings across different model architectures, training procedures, and physical systems?
Reviewer_FJ4jOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper questions a common practice in diffusion-based emulators: selecting cheaper samplers by their closeness to a fine-step reference sampler. Using local error budgets to quantify this closeness, it evaluates pretrained ACDM emulators on three flow datasets and finds:

  • Closeness tracks physical accuracy on transonic flow but inverts on isotropic turbulence over long rollouts.
  • When the reference sampler produces physically sound simulations, closeness to it is consistent with physical accuracy.
  • Budget-based sampler selection does not beat direct pilot evaluation.
Critical strengths
  • Carefully controlled and transparent evaluation. The paper fixes its data splits, candidate samplers, metrics, and selection rules before testing, and clearly labels analyses added afterward. It also reports negative results, including that its own budget-based selection does not beat a simple baseline.
Critical weaknesses
  • Expected conclusion without a practical alternative. It is expected that a reference sampler with poor physical accuracy is a poor proxy for physical accuracy. The more useful question is what proxy to use instead, and the paper does not offer one. Its suggested fix is to check the reference against ground truth. But if ground truth is available, candidate samplers can simply be compared against it directly, which the paper itself finds equally effective and cheaper.
  • Limited analysis beyond observation. Most of the paper runs many samplers and interprets which ones do better, without explaining why. For example, the paper does not analyze why stochastic sampling preserves energy better than deterministic sampling. Its main explanation, that stochastic sampling corrects errors of an imperfect denoiser, has already been explored for image diffusion models (e.g., Karras et al., 2022; Xu et al., 2023).
Questions for authors
  1. Validating the reference requires ground truth, but with ground truth, candidate samplers can be compared against it directly. In what practical setting is reference fidelity still useful?
  2. When ground truth is not available, which is the more practical setting, how should a sampler be chosen?
  3. Why is the deterministic reference physically inaccurate on Iso but not on Tra? Switching to stochastic sampling improves Iso, but it does not explain what property of Iso causes the failure.
Reviewer_ni9ZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper argues that often a deterministic sampler for diffusion models, like DDIM, is used as a baseline to find a cheaper sampler that requires fewer steps. The deterministic sampler is numerically converged, i.e., results do not change much when chosing a smaller step size. However, that does not mean that the numerically converged sampler is physically accurate. The authors demonstrate that this is not the case. For example for isotropic turbulence, the sampler is numerically converged, but samples lose kinetic energy over time. Therefore, samplers for diffusion-emulators should not be trusted based on convergence alone, but samples must be evaluated based on physical observables.

Critical strengths
  • The findings of the manuscript that numerical convergence does not imply physical validity are intuitive and seem correct.
  • The analysis and experiments are proper and support the claims.
Critical weaknesses
  • The manuscript is very difficult to follow. The language is very unnatural and the terminology is not standard. The manuscript reads like a LLM-generated technical report. The writing and presentation need to be improved significantly.
  • The overall premise seems wrong. The paper argues in the first sentence of the abstract that "cheaper samplers are judged by how closely they reproduce a fine-step reference." The fine-step reference here is a deterministic diffusion sampler. The paper then argues that the fine-step reference might be numerically converged, but physically inaccurate. I agree to that and I think most of the scientific ML community does that too. So the central argument of the paper is something that for me is not surprising and already well known. But I don't think that cheaper samplers are typically judged by how closely they reproduce the converged diffusion sampler as claimed in the abstract. Can you give some examples of that? For me, this significantly impacts the novelty and relevance of the paper.
  • The finding that fine-step samplers like the deterministic DDIM can drift over time and stochastic samplers are superior in many settings are well established already.
  • Experiments are conducted on specialized datasets, don't cover many newer variants of samplers/diffusion models and neural network architectures. Experiments are overall limited in scope.
Questions for authors
  • Can you give some examples where cheaper samplers are judged by how closely they reproduce a fine-step reference?
  • Running the fine-step reference accounts for the majority of the audit's calibration cost, and that the audit mostly ties with or underperforms simple direct pilots. Given this high overhead, what is the practical incentive for practitioners to use budget-based auditing instead of just running direct pilot evaluations from the start?
  • Findings are based on ACDM only. Do you expect these specific energy-loss and inversion phenomena to replicate in other physics emulator paradigms, such as continuous-time neural operators or diffusion models trained with explicit physical loss penalties?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.