acceptodds
Withdrawn from ICLR 2027

Spectral Stationarity Predicts Neural-Operator Transfer Across Resolutions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural operators may accept a new grid without preserving the physical response learned at the training resolution. We show that spectral stationarity predicts this transfer before target-resolution labels are available. At each physical frequency, it separates the stability of the learned response from its representability on the target grid. This yields a label-free estimator from band-limited probes evaluated on the training grid and one additional grid. In controlled one-dimensional studies, the estimator reduces false acceptance of resolution-anchored models from 68.6% to 2.8%. In a fixed-architecture two-dimensional study, eight probes achieve 0.962 concordance in paired model selection, compared with 0.464 for global training error at identical 0.259 coverage. The analysis also distinguishes failures caused by target-grid sampling, discretization-anchored representations, and uninformative probes. Spectral stationarity therefore identifies unsafe grids and usable frequency bands before labels are generated at the deployment resolution.

Primary area:
learning on time series and dynamical systems
Keywords:
neural operators, resolution transfer, spectral stationarity, label-free model selection, operator learning, PDE surrogate modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1kBrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on the cross-resolution transfer of neural operators (NOs), specifically on whether the performance of NOs on grids other than the training resolution can be predicted. If so, the cost of labeled validation at the target resolution can be saved. The main insight of the paper is to make this prediction by probing the frequency-domain response of NOs at low cost, and the paper formalizes a series of specific parameters, thresholds, and operations to make this prediction operational and accurate. The main result is that, when the probe does not abstain, the paper accurately predicts the cross-resolution generalization ability of NOs, and it uses a second grid to identify models anchored to the training grid.

Critical strengths
  1. Predicting in advance whether NOs generalize across resolutions is a meaningful problem, because it can save part of the cost. I therefore fully acknowledge the motivation of the paper.

  2. The paper validates its own claims from multiple angles and with good completeness: sensitivity to regimes, models, parameters, etc. is all examined. I therefore acknowledge the completeness and transparency of the paper's experiments.

Critical weaknesses
  1. The core idea of the paper is not novel, and the additional formalization is overly empirical and, in practice, not particularly accurate. Although few works directly target predicting, before deployment, whether NOs generalize after a change of resolution, probing the frequency-domain response of NOs in advance and using it to explain their generalization at higher resolutions has already been discussed or used in prior literature [1,2]. Compared with these existing works, the only additional contribution of the paper appears to be turning the qualitative inference of "judging generalization from band-wise responses" into operational quantitative thresholds and procedures. Therefore, I do not think the paper offers me any new insight. What is novel in the paper is the formalization and operationalization of the above idea. However, many parameters and thresholds are chosen rather empirically, which makes their generalizability in practice questionable: the safety threshold τ=0.25\tau = 0.25 defines both the ground-truth labels and the decisions, and neither the stationarity tolerance δ=0.10\delta = 0.10 nor the abstention margin of 0.020.02 (which the protocol states was selected on development seeds) is derived, among other choices. Moreover, concordances such as 0.99 are computed only on the ~37% of model pairs where the probe does not abstain; counting abstentions as 0.5, it drops to about 0.68. Therefore, the paper's formalization both lacks justification and is not accurate.

  2. More seriously, the so-called newly designed probe is equivalent in meaning to validation on held-out samples from the deployment distribution at the training resolution. In the supplementary code, the probes and the deployment evaluation data are generated by the same random-field generator with the same set of spectral parameters, and are labeled by the same reference solver; the only differences are the random seed and the fact that kmax⁡k_{\max} is capped at Ntrain/2−1N_{\text{train}}/2 - 1. Therefore, g(b)g(b) is simply the band-wise error of a small set of held-out deployment-distribution samples on the training grid, while the prediction target is the band-wise error of the same distribution on other grids; for stationary operators, the two are approximately equal by definition. Furthermore, I computed the simplest baseline using the authors' own records: selecting directly by the band-wise error of held-out data on the training grid performs on par with or better than the probe. Therefore, the paper's selection results cannot show that the probe provides any information beyond held-out validation at the training resolution, which directly undermines the core contribution of the paper.

[1] Wang et al. Breaking Scale Anchoring: Frequency Representation Learning for Accurate High-Resolution Inference from Low-Resolution Training. ICLR 2026.

[2] Sakarvadia et al. The False Promise of Zero-Shot Super-Resolution in Machine-Learned Operators. ICLR 2026.

Questions for authors
  1. Compared with the approach in [1,2] of qualitatively judging generalization from band-wise responses, what does the paper's quantitative procedure reveal that qualitative analysis cannot?

  2. Could the authors explain what essential difference exists between the probes and held-out samples from the deployment distribution, other than the random seed? And if they do differ, why is the accuracy of the former comparable to that of the latter?

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

This paper studies whether a trained neural operator can be expected to preserve its learned physical response when evaluated on a grid resolution different from the training resolution. The central proposal is to separate two factors at each physical frequency band: (i) the fidelity of the learned response, estimated using a small set of synthetic probes and reference solutions, and (ii) whether the target grid can represent that band. Their coefficient-free danger score is 1−repr⁡(b;Neval)gNtrain(b)1-\operatorname{repr}(b;N_{\mathrm{eval}})g_{N_{\mathrm{train}}}(b), where gg denotes probe-estimated band fidelity and repr⁡\operatorname{repr} denotes target-grid representability.

The paper argues that using the training-grid response at another resolution additionally requires spectral stationarity: the fidelity associated with a fixed physical band should remain approximately unchanged when the same trained model is evaluated on a different grid. The authors test this condition using matched probes at the training grid and one check grid. The method returns bandwise response, a stationarity decision, and target-grid representability rather than a single architecture-level resolution-invariance label.

The empirical study covers three periodic 1D PDEs and two periodic 2D PDEs, several neural-operator families, within-architecture model comparisons, a fixed-architecture FNO-2D spectral-loss tradeoff, and probe-count/cost analyses. The main reported findings are that a second-grid check reduces false acceptance of resolution-anchored models from 68.6% to 2.8%, and that eight probes obtain 0.962 paired concordance versus 0.464 for global training error at the same 0.259 coverage in the matched 2D comparison.

Critical strengths
  1. The paper provides a useful mechanism-based decomposition of resolution-transfer failure.

Separating target-grid representability from learned frequency-band fidelity clarifies why a finer grid cannot recover frequencies that the model did not learn, while a well-trained model cannot overcome frequencies that a coarse grid cannot represent. Adding a second-grid stationarity check addresses the important failure case in which a model is accurate only at its training discretization. The reduction in anchored-model false acceptance from 68.6% to 2.8% suggests that this check contains information not available from training-grid error alone.

  1. The empirical evaluation attempts to isolate band-specific model selection rather than architecture identity or global model quality.

The within-family comparisons and the fixed-architecture FNO-2D experiment are particularly valuable because they compare models whose architecture and most training conditions are held fixed. Reporting concordance together with coverage, using matched-coverage training-error baselines, and aggregating uncertainty at the PDE--seed level make the reported comparisons more informative than a pooled collection of resolution--band cells would be.

Critical weaknesses
  1. The practical advantage over direct labeled evaluation is not yet fully established.

The method is described as operating before target-resolution labels are available, but it still requires trusted reference solves for synthetic probes at the training grid and at an additional check grid. The reported cost depends strongly on the PDE (for example, the paper reports a much larger reference cost for Navier--Stokes than for linear advection--diffusion). I did not find a same-cost comparison against using a small number of labeled deployment cases or target-grid validation cases. Such a baseline addresses a different question, as the paper notes, but it is important for determining when the proposed procedure is operationally preferable. The requirement that probes match the deployment spectral envelope also appears to assume useful prior knowledge about deployment inputs.

  1. The evidence supports a local, band- and setting-specific diagnostic more clearly than a general prediction of cross-resolution transfer.

A single additional grid can expose some anchored responses, but agreement at two grids does not guarantee stationarity at an untested grid or throughout an interval; the appendix explicitly provides this limitation, and the results show that the direction and location of the check grid matter. In addition, repr⁡\operatorname{repr} counts represented modes rather than represented reference energy. The appendix explains that this is not a universal error bound when spectral energy is nonuniform. Most experiments use periodic domains and Fourier bands, so it is unclear how much of the result extends to nonperiodic boundaries, irregular meshes, anisotropic phenomena, or task-specific quantities of interest. These qualifications should be made more prominent in the main claims and abstract.

Questions for authors
  1. At matched total reference-solver cost, how does the proposed two-grid synthetic-probe procedure compare with directly evaluating a small number of randomly sampled or actively selected deployment cases at the target grid? Is there a regime of solver cost, number of candidate target grids, and deployment-set size in which the proposed method has a clear break-even advantage?

  2. How should a practitioner choose the check-grid resolution and direction when the final deployment resolution is not known in advance or when several target resolutions are possible? Can the authors provide a practical rule or an empirical coverage map showing which target-grid interval is supported by a check at N/2N/2, N/4N/4, or 2N2N?

Reviewer_Akh7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a spectral stationarity criterion to predict neural operator reliability across resolutions using bandwise errors measured on two grids. Experiments on 1D and 2D PDEs evaluate its effectiveness in detecting transfer failures and selecting models.

Critical strengths

The paper addresses the useful problem of assessing neural operator reliability across resolutions without requiring reference solutions at the target resolution.

Critical weaknesses
  1. Limited novelty. The method mainly uses training-resolution validation errors to predict performance at new resolutions, with an additional grid to check error stability. The ablation results show that simple two-grid validation already achieves most of the reported improvement. Moreover, the evaluation does not sufficiently cover different training strategies or establish an advantage over simple resolution-scaling baselines. These limitations weaken the claimed novelty and generalizability.
  2. The predictive value of the proposed stationarity criterion remains insufficiently established. The prformance in false acceptance is reported at a single operating point τ=0.25,δ=0.10\tau=0.25,\delta=0.10, without systematically evaluating the full two-grid criterion across different safety thresholds and delta.
Questions for authors

no further question.

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

Related papers

Discussion (0)

Sign in to comment.