acceptodds
Under review as a conference paper at ICLR 2027

From Coarse Training to Fine Inference in Neural Operators via Resolution Adaptation

Abstract

We address this cross-resolution generalization problem using Polyphase Decomposition. Rather than adapting a pretrained neural operator to a finer discretization, we decompose the high-resolution problem into subproblems at the resolution seen during pretraining, apply the pretrained operator to each subproblem, and recombine their predictions at the target resolution. This yields a zero-shot method, Polyphase Inference, that requires no high-resolution training data. When high-resolution examples are available, the decomposition strategy can be learned and complemented with fine-tuning modules, giving Polyphase Fine-Tuning. Across several PDE benchmarks and neural architectures, Polyphase Inference consistently improves cross-resolution performance over direct evaluation on the target resolution. In the low-data setting, Polyphase Fine-Tuning substantially outperforms standard parameter-efficient adaptation methods while using the same or fewer additional parameters. These results suggest that matching the pretrained input resolution can be an effective component of cross-resolution transfer.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, transfer learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Zro8OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Decomposed Inference (specifically Polyphase Inference) as a way to transfer a neural PDE solver pretrained at low resolution rr to a higher resolution R>rR>r by decomposing the fine-grid input into multiple coarse-grid components, applying the pretrained model independently to each, and recomposing the outputs, thereby avoiding both the aliasing errors of direct high-resolution inference and the information loss of compressed inference. The authors provide a theoretical characterization showing that direct FNO inference suffers from cross-resolution discretization mismatch, that compressed inference has an irreducible error floor, and that component-wise marginal invariance is the key condition for risk preservation under decomposition. Experiments across multiple PDEs (Burgers, Darcy flow, Reaction-Diffusion, Compressible Navier-Stokes), architectures (FNO, UNet, Transolver, LaMO), and large pretrained physics models (DPOT up to 1.03B parameters) demonstrate that Polyphase Inference consistently improves zero-shot transfer, supports parameter-efficient fine-tuning, and offers favorable time-memory trade-offs. The method is limited to regular grids with integer resolution gaps, with future work targeting irregular geometries, point clouds, and other modalities such as vision, audio, and time series.

Critical strengths
  • Polyphase Inference is simple, requires no retraining, and works across architectures (FNO, UNet, Transolver, LaMO) and scales to large pretrained models (DPOT up to 1.03B parameters). It reopens a problem the field had largely abandoned and shifts the research question from "how do we train resolution-robust models?" to "how do we present high-resolution data to a fixed low-resolution model?"
  • The paper does not offer decorative theory. Proposition 1 explains why direct inference fails (layer-wise aliasing propagation), Proposition 2 explains why compression cannot work in principle (irreducible error), and Propositions 3–4 explain when the proposed method works and what remains to be fixed. The error decomposition directly motivates the design of Polyphase Fine-Tuning, with in-component error justifying PEFT and cross-component error justifying the lightweight depthwise convolutions. The honesty about limitations (regular grids, integer resolution gaps, approximate marginal invariance) strengthens rather than weakens credibility.

I personally think this is a strong paper that makes a genuine conceptual contribution to an important problem. The method is simple, theoretically grounded, empirically validated across a wide range of settings, and practically deployable.

Critical weaknesses
  • The paper's theoretical foundation rests on component-wise marginal invariance (Definition 1, Proposition 3). The assumption that each decomposed polyphase component's input-output distribution matches the pretraining distribution. The paper itself acknowledges this does not always hold, noting that "when exact translation invariance does not hold, spatial Sobolev regularity controls differences between phase distributions; comparison with μr\mu_r additionally contains the mismatch of a reference phase with the pretraining distribution." The paper's own Appendix F.5 formalizes this as an approximate condition governed by an unmeasurable term W2(ν0,μr)W_2(\nu_0, \mu_r), described as "separately measures the reference-phase mismatch with pretraining." This limits the method's practical utility and leaves the theoretical guarantee as a post hoc explanation rather than an actionable criterion.
  • Proposition 4 decomposes the excess risk into in-component and cross-component errors. The paper correctly identifies that cross-component error requires "exploiting interactions across components" and proposes lightweight depthwise convolutions before decomposition and after recomposition, motivated by Equation 15 showing that fine-grid convolution induces structured polyphase interactions. However, the paper provides no theoretical analysis of whether these convolutions actually reduce the cross-component error, nor any ablation isolating their contribution quantitatively.
Questions for authors
  • The paper positions Polyphase Inference as an alternative to the multi-resolution training advocated by Sakarvadia et al. (2026), but never directly compares the two under matched data budgets. Could the authors clarify whether, in the 1% and 10% few-shot settings, a multi-resolution training baseline was considered? This matters because the paper's practical recommendation ("use Polyphase instead of retraining") is only justified if Polyphase matches or exceeds the alternative at comparable cost. If multi-resolution training with the same 1% high-resolution data outperforms Polyphase + PEFT, the paper's central practical claim would need qualification.
  • The polyphase decomposition produces q=mdq = m^d components. For 3D problems with m=4m=4, this is 64 forward passes. Could the authors clarify whether the method was tested in any 3D setting, and whether the wall-clock or memory advantage over direct high-resolution inference persists as dd and mm grow? The efficiency claims in Section 5.5 are based on 2D data; extrapolation to 3D is not obvious.
  • The paper positions Polyphase Inference as an alternative to the multi-resolution training advocated by Sakarvadia et al. (2026), but never directly compares the two under matched data budgets.

I recommend acceptance, with the expectation that the authors will clarify the questions raised above in the camera-ready version.

5 comments
AuthorsOpenReview ↗

Part 1/2 - Response to Reviewer Zro8

First of all, thank you very much for the time and care you have invested in reviewing our paper. We sincerely appreciate it.

W1: The approximate condition given by W2(ν0,μr)W_2(\nu_0, \mu_r) is unmeasurable and limits the practical utility.

We agree that the condition involving W2(ν0,μr)W_2(\nu_0,\mu_r) has limited utility as a practical criterion: it cannot generally be verified from low-resolution training data alone, and estimating it reliably from high-dimensional samples can be difficult even when high-resolution data are available.

The purpose of this condition is to clarify the assumptions needed to transfer low-resolution accuracy to high-resolution inference. It shows how distributional mismatch affects the theoretical guarantee, but does not provide a practical way to check whether the method will work. Even though both resolutions describe the same physical system, their distributions are not necessarily close.

Our experiments show that Polyphase Inference works well across several PDE benchmarks, but they do not prove that the Wasserstein condition holds. We therefore see the theory as an explanation of when the method can be expected to work, rather than a guarantee that it will work in every setting

We have added a brief clarification on this point in Appendix F.5 of the revised manuscript.

W2: the paper provides no theoretical analysis of whether these convolutions actually reduce the cross-component error, nor any ablation isolating their contribution quantitatively.

Sec. 4.5 provides a theoretical motivation for the proposed convolutions: it shows that convolution on the high-resolution grid induces interactions between polyphase components, allowing each component to incorporate information from the others.

For fixed polyphase decomposition and recomposition operators, Proposition 4 shows that the cross-component error term is independent of the pretrained model. Hence, adapting the pretrained model alone cannot reduce this term. The proposed convolutions introduce interactions between polyphase components and make the decomposition and recomposition operators trainable, thereby providing a means to modify this error term. However, this does not guarantee that optimization will reduce it.

To isolate the contribution of the learned convolutions to learning decomposition and recompositions, we conducted an additional ablation study comparing PPFT + Adapter with Frozen PPI + Adapter, in which the polyphase decomposition and recomposition remain fixed and only the adapters are trained. Both methods use the same pretrained checkpoints and fine-tuning data.

The complete results are reported in Tab. 38-41 of App. L.2, the following table reports the relative L2L_2 errors on Darcy flow:

Pretrained resolutionPPI (zero-shot)Adapter 1%Frozen PPI + Adapter 1%PPFT + Adapter 1%Adapter 10%Frozen PPI + Adapter 10%PPFT + Adapter 10%
16×160.091900.115000.087210.066070.115600.087250.05497
32×320.052800.096990.049260.046830.095960.045800.04200
64×640.046100.072160.041330.037140.066330.040590.03573

First, Frozen PPI + Adapter consistently improves over both zero-shot PPI and adapters alone, showing that fine-tuning the pretrained model while retaining the fixed polyphase decomposition is already beneficial. Compared with zero-shot PPI, Frozen PPI + Adapter reduces the error by 5.1–10.3% with 1% of fine-tuning data and 5.1–13.3% with 10%.

Learning the decomposition and recomposition operators provides further improvements. On Darcy flow, PPFT + Adapter consistently outperforms Frozen PPI + Adapter across all pretrained resolutions and both fine-tuning budgets. The relative error is reduced by 24.2%, 4.9%, and 10.1% with 1% of fine-tuning data, and by 37.0%, 8.3%, and 12.0% with 10%, for pretrained resolutions of 16×16, 32×32, and 64×64, respectively.

On the other three datasets, the difference between Frozen PPI + Adapter and PPFT + Adapter is small. This suggests that fixed polyphase decomposition already captures most of the available benefit in these settings, leaving limited room for further improvements from learning the decomposition and recomposition operators.

AuthorsOpenReview ↗

Part 2/2 - Response to Reviewer Zro8

Q1, Q3: Comparison with multi-resolution training advocated by Sakarvadia et al. (2026).

We did not include this baseline in the original submission because our setting assumes an existing coarse-resolution pretrained model. However, we agree that multi-resolution training is an important alternative when additional high-resolution training data is available.

We have conducted additional experiments comparing multi-resolution FNO training with PPFT under the same high-resolution data budgets (1% and 10%) and under a favorable budget for multi-resolution training (25% of data for each resolution). The results are reported in the table below, where we compare multi-resolution training against the best PPFT configuration, with or without adapters.

With 1% high-resolution data, PPFT outperforms multi-resolution training in all 6 configurations across Darcy, Burgers, and Reaction–diffusion. With 10% data, PPFT achieves lower errors in 3 out of 6 configurations, while multi-resolution training remains competitive, particularly on Reaction–diffusion.

Overall, PPFT achieves better results in 9 out of 12 configurations. We report a table for comparison here, and we added more detailed results in Tab.32-33.

Dataset / splitBest epochMulti-resolution test relative L2 by resolutionBest PPFT at same % of high res. data
Darcy128² / 64² / 32² / 16²128²
25/25/25/251460.0481 / 0.0423 / 0.0430 / 0.0562—
90/0/0/101460.0563 / 0.0553 / 0.0540 / 0.05200.0534
99/0/0/11450.0927 / 0.0757 / 0.0727 / 0.05090.0631
0/0/99/11460.0753 / 0.0399 / 0.0428 / 0.07920.0364
0/0/90/101410.0539 / 0.0410 / 0.0428 / 0.07910.0355
Burgers1024 / 512 / 256 / 1281024
25/25/25/251450.0496 / 0.0469 / 0.0482 / 0.0554—
90/0/0/101460.0553 / 0.0509 / 0.0519 / 0.05500.0588
99/0/0/11480.0725 / 0.0531 / 0.0537 / 0.05620.0567
0/0/99/11450.0660 / 0.0451 / 0.0469 / 0.06120.0460
0/0/90/101460.0519 / 0.0459 / 0.0475 / 0.06020.0457
Reaction–diffusion128² / 64² / 32² / 16²128²
25/25/25/251420.2428 / 0.2397 / 0.2415 / 0.3099—
90/0/0/101310.3232 / 0.3118 / 0.3090 / 0.32340.3320
99/0/0/11490.4338 / 0.4093 / 0.4083 / 0.33150.3320
0/0/99/11460.2247 / 0.1977 / 0.2056 / 0.37560.2210
0/0/90/101350.2097 / 0.1976 / 0.2059 / 0.37570.2210

Q2. Efficiency and evaluation on 3D data.

We added two 3D datasets (Swift-Hohenberg and Kuramoto-Sivashinsky) generated with the solvers provided in APEBench (Koehler et al., 2024) at resolutions 323,643,128332^3, 64^3, 128^3 and 2563256^3. We added the results in Tab.7-8 and we included the results in Fig.3 and Fig.8 with latency and memory peak measurements. At these resolutions, polyphase inference is, roughly, as fast as the direct evaluation of FNO.

The efficiency benefits are significant at larger scale. To demonstrate this, we added experiments scaling the resolution in 1D, 2D, 3D, and 4D, using randomly generated inputs, since the analysis concerns computational efficiency rather than predictive accuracy, until reaching out-of-memory (OOM) on an NVIDIA A100 GPU with 80 GB of memory. At the largest resolutions, parallel PPI achieves speedups over direct FNO of 3.59× in 1D (50,331,648 points), 3.36× in 2D (716827168^2), 2.99× in 3D (3683368^3), and 3.04× in 4D (80480^4), with comparable peak memory consumption. We added these results as Tab.34-37 in App.L.1 and a visualization of the scaling trend in Fig.7.

Reviewer_Zro8OpenReview ↗

Re: Response to Reviewer Zro8

Thank you very much for the authors' careful and conscientious response, as well as the extensive additional experiments. I truly appreciate the effort and seriousness the authors have brought to this revision.

I personally find the core idea appealing. That said, the concerns raised by the other reviewers have also caught my attention, and I would very much like to understand how the authors view and plan to address those points. I would also kindly encourage the authors to directly and fully respond to Reviewer b6zW's second-round comments.

At this stage, and with genuine regret, I am afraid I must maintain my current assessment unchanged.

Thank you again for the authors' hard work and thoughtful response.

Reviewer_b6zWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper focuses on super-resolution PDE simulation, with the goal of minimizing discretization error. It argues that direct generalization retains discretization error, while compressed inference introduces a new source of error. It therefore proposes decomposed inference, which invertibly decomposes the high-resolution domain into multiple low-resolution domains that are simulated separately, and proves that this can be lossless under a strong joint-distribution assumption. Within the scope of the paper's experimental design, the proposed method is the only one that keeps the error nearly unchanged across resolutions in most settings.

Critical strengths
  1. Decomposing the high-resolution physical domain into multiple simulations at the training resolution is a very interesting and feasible approach, and I appreciate the novelty and insight of the paper.
  2. The chain of argument is fairly complete and the presentation is clear, which makes the paper easy to follow.
Critical weaknesses
  1. The experimental setup favors the proposed method and cannot test the application scenario the paper claims. The paper reuses the code of Sakarvadia et al., whose data loaders generate low-resolution data directly from the high-resolution simulation data by subsampling with x[::n]. This is a common, low-cost practice that avoids regenerating low-resolution training data and is usually reasonable; here, however, it happens to make the training data satisfy the very strong assumption of Prop. 3 (at least for Burgers and CNS, which have periodic boundaries). In the super-resolution application scenario, as the paper itself acknowledges, cheap low-resolution data are abundant while high-resolution data are scarce, and the strong joint assumption between low- and high-resolution data can hardly be guaranteed. The DPOT experiment, the only one in which the low-resolution training data and the high-resolution test data are generated by different pipelines, lacks both the native error at resolution 128 and compressed-inference baselines as references. Prop. 3 is itself very strong, and together with an experimental setup that heavily favors decomposed inference, this greatly undermines my confidence in the practical applicability of the method.

  2. The effectiveness of Polyphase Fine-Tuning (PPFT) lacks experimental support. PPFT is initialized from zero-shot Polyphase Inference, but Figure 4 uses direct zero-shot FNO as the reference, thereby counting the gain of zero-shot Polyphase Inference itself as a gain of fine-tuning. The claim in the abstract that PPFT substantially outperforms standard PEFT likewise comes from a comparison with PEFT applied directly to the FNO at high resolution. With zero-shot Polyphase Inference as the reference, Tables 7–10 show that PPFT brings a clear improvement only on Darcy, and no improvement or even degradation on the other three datasets. I therefore think the effectiveness of the proposed fine-tuning method currently lacks supporting evidence.

Questions for authors
  1. Please clarify the following experimental details:

    • (a) For each dataset, how were the low-resolution training and test data obtained from the high-resolution data, and do the test sets at different resolutions consist of the same trajectories? Are the "Native train res.", "Zero-shot" and "Polyphase inference" rows computed with the same checkpoint and the same evaluation pipeline, including the rollout length and the way errors are computed?
    • (b) Are the MSE, L2L^2 and relative L2L^2 errors computed in normalized or physical space, and over which of the samples, time steps, channels and grid points are they summed or averaged? Is the relative L2L^2 error an average of per-sample ratios, or a single ratio over the whole test set?
    • (c) Are the FNO results in Tables 11–14 obtained from the same models and the same evaluation as the corresponding settings in Tables 3–6? If not, what differs? From which pretrained checkpoint is each fine-tuning run in Tables 7–10 initialized, and how is the reported checkpoint selected?
  2. I do not find convincing the authors' reason for not comparing with CROP, namely that they "found the same negative results reported in Sakarvadia et al.". The negative result of Sakarvadia et al. is that CROP has large errors on unseen high frequencies, not that CROP fails to address discretization error, and the two are not necessarily related. Beyond citing Sakarvadia et al., how do the authors show that CROP does not solve the discretization-consistency problem the paper aims to address, or that CROP introduces larger errors than decomposed inference?

  3. Could the authors use the PDEBench solver already employed in Appendix G.4 to generate low-resolution training data directly on the coarse grid, and compare decomposed inference, Low+Pad, and the most direct compressed baseline (x[::n] subsampling followed by interpolation) on independently generated high-resolution trajectories? In addition, Low+Pad downsamples by Fourier truncation, so its input distribution differs from that of the training data obtained with x[::n]; is this comparison fair? Ideally, this should be done in a setting where resolution actually matters (i.e., the native error decreases clearly with resolution), rather than on datasets such as CNS whose native error barely changes with resolution.

7 comments
AuthorsOpenReview ↗

Part 1/2 - Response to Reviewer b6zW

First of all, thank you very much for the time and care you have invested in reviewing our paper. We sincerely appreciate it.

Q1: Please clarify the following experimental details.

For the original experiments on PDEBench, data at lower resolutions were obtained by subsampling, as in Sakarvadia et al. i.e. using a fixed stride, except for the experiments with DPOT, for which we generated the data using the PDEBench solver, since the CNS dataset was already present in the pretraining dataset and we needed to evaluate the model at a resolutions higher than the ones included in the pretraining. Train and Test sets at lower resolution are obtained by subsampling the high resolution train and test splits respectively (low-resolution train sets are obtained from the high-resolution train set and low-resolution test sets are obtained from the high resolution test set). For a given experimental set-up (e.g., dataset, pre-training resolution, and target resolution at inference), the experiments with different inference schemes (e.g., denoted as "Zero-shot" and "Polyphase inference") and the evaluation at the training resolution ("Native train res.") are computed with the same checkpoint and the same evaluation pipeline.

All metrics (MSE, L2L_2, rel. L2L_2) are computed in physical space. Data are normalised before being fed into the models and denormalised before evaluation. The metrics are computed over the test set and averaged over the trajectories of the test set, over the time-steps of each trajectory and over the channels. For each sample, Relative L2L_2 is computed as the L2L_2 distance between prediction and target divided by the norm of the target and is averaged over the samples to provide the relative L_2 that we report.

The FNO results in the original Tables 11–14 (Tables 13-16 in the revised version) are obtained using the same models and evaluation as in Tables 3–6 for each corresponding setting. We identified that some results in the original Tables 11–14 used different pretrained checkpoints. We have corrected these entries to ensure consistency with the checkpoints used in Tables 3–6. The checkpoints used for pretraining are the same as those reported for the zero-shot experiments, and the fine-tuning modules are initialised such that, at initialisation, the model we fine-tune performs exactly as the respective zero-shot variant. The fine-tuning checkpoint is selected based on a validation set, composed of data not used for fine-tuning and for testing.

W2, Q3: Could the authors use the PDEBench solver to generate low-resolution training data and compare with the baselines x[::n], inference with the pretrained model, followed by interpolation?

We agree that this is an important test of practical applicability and thank you very much for this remark. To address it, we added experiments using datasets generated directly at their native resolutions with the official PDEBench solvers. We also added two 3D datasets (Swift-Hohenberg and Kuramoto-Sivashinsky) generated with the solvers provided in APEBench (Koehler et al., 2024). We have added this visualization in Fig.8 and report the complete results in Tab.7-12. The trend is similar to that observed in experiments on the downsampled data, namely that Polyphase inference performs best across most datasets and resolution gaps. In addition, the baseline "x[::n] followed by interpolation", that we call "Stride+interpolation" in the revised paper, generally underperforms Polyphase inference and often Low+Pad, with particularly poor performance on the two 3D datasets. We have added these results in Tab.7-12 and Fig.8.

AuthorsOpenReview ↗

Part 2/2 - Response to Reviewer b6zW

Q2: Comparison with CROP

We agree that CROP is a very relevant work we should have included. We have now added CROP+FNO, which is the CROP model used in Sakarvadia et al. and presented in Gao et al., Section 5.1. We note that the main CROP model used in Gao et al. is CROP-CRNO, and we have added it as well. We tested both CROP models on downsampled and solver-generated data.

Unlike CROP, which targets resolution consistency through model design, Polyphase Inference is an adaptation strategy that can be applied to existing pretrained models without modifying their architectures. We therefore evaluate CROP on high-resolution data both directly and with Polyphase Inference. Across our experiments, Polyphase Inference generally achieves comparable performance when CROP generalises well to higher resolutions, while significantly improving its performance when direct evaluation degrades. On downsampled data, PPI improves direct evaluation in 11/12 configurations for CROP+FNO and 3/5 configurations for CROP-CRNO. We report below the comparison on solver-generated data, where PPI improves direct evaluation in 10/12 configurations for CROP+FNO and 5/5 configurations for CROP-CRNO. The complete results are reported in Tables 25–31 of Appendix I.

CROP+FNO — Test relative L2L^2 error (solver-generated data).

DatasetTraining gridDirect+ Polyphase
Darcy16×160.574460.38026
Darcy32×320.489750.29823
Darcy64×640.351350.29289
Reaction–diffusion16×1611.131314.28633
Reaction–diffusion32×327.642132.25830
Reaction–diffusion64×644.637702.47882
Compressible Navier–Stokes16×160.204820.08329
Compressible Navier–Stokes32×320.101990.08099
Compressible Navier–Stokes64×640.084170.08816
Burgers1280.952830.94565
Burgers2560.935810.93555
Burgers5120.950800.95132

CROP-CRNO — Test relative L2L^2 error (solver-generated data).

DatasetTraining gridDirect+ Polyphase
Darcy16×160.274170.22226
Darcy32×320.121370.10068
Darcy64×640.058420.05246
Reaction–diffusion64×643.660163.21456
Compressible Navier–Stokes64×640.071200.07079

Thank you for your detailed response. Below is my reassessment of my remaining main concerns.

W1. I thank the authors for the additional experiments. However, my original core concern appears to still hold. In the main experiments, which use the subsampling construction, the proposed method attains accuracy close to the native coarse-resolution accuracy, but this mainly relies on the subsampling construction making the strong assumption of Prop. 3 (approximately) hold: in the added setting where coarse- and fine-resolution data are generated separately, the error can reach many times the native error (Tables 9–12).

Q2. I thank the authors for the additional experiments and clarifications. However, the added experiments do not properly control the variables. The authors mainly compare direct CROP evaluation with CROP + Polyphase, i.e., they stack the two methods. To compare the accuracy of CROP and Polyphase, one should instead compare them with the same backbone. Among the provided results, the only same-backbone comparison is FNO + Polyphase versus CROP+FNO, and the key issue is that the CROP+FNO results look abnormal: even when trained directly at the target resolution, CROP+FNO is 2–15× worse than a plain FNO (Tables 25–28 vs. Tables 3–6). The authors' implementation may have a problem. Therefore, there is currently no valid same-backbone result for comparing the accuracy of CROP and Polyphase. On the other hand, since the authors do not report CROP's results at its native resolution, it also cannot be verified that CROP does not solve the problem the paper aims to address (i.e, making the model's accuracy at the fine resolution consistent with that at the coarse resolution), which the original CROP paper did at least demonstrate in its own experimental setting.

In summary, I still cannot recommend acceptance. The main reason is that this version of the paper oversells the current method: based on a possibly problematic implementation, it concludes that CROP "does not always generalize well to higher resolutions", thereby understating the existing contribution of Gao et al.; it describes the results too strongly; and it does not state reservations about the strong assumption prominently enough. If the authors honestly acknowledge CROP's contribution to this problem and accurately position the strength of their contribution, for example by removing the statements about achieving coarse-resolution accuracy and by replacing the misleading main experiments in the main text with the results under the separately generated setting, I would be inclined to consider acceptance. Given the current results, a statement along the lines of "Polyphase improves the accuracy of super-resolution simulation" may be more appropriate.

Reviewer_7KfDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors identify the problem of training neural operators in low resolution regimes, but having to use them on high resolution spaces at test time. This can be due to lack of data or computational constraints, especially in the field of physics problems defined by PDEs, with computationally heavy simulations. The authors identify and formalize problems with the existing neural operator methods, including FNO and downsampling/upsampling operations which they call ‘compressed inference’. A polyphase decomposition-based method is proposed to replace the downsampling/upsampling pipeline, which can better preserve informative, high-resolution details. The approach is evaluated on four datasets, with and without fine-tuning, and outperforms baselines on multiple architectures. The proposed model is also faster and more memory efficient than a zero-shot baseline.

Critical strengths

The problem is well described and motivated: Identifying and formalizing the resolution mismatch problem in neural operators, especially in the context of limited data; and formalizing the resolution mismatch of the existing FNO method as well as the limitations of the ‘compressed’ upsampling methods are interesting contributions. The writing is clear and easy to follow.

Critical weaknesses

Novelty: The novelty of the method needs to be clarified and compared with related work outside of the specific domain of neural operators for solving physics PDEs. Polyphase decomposition followed by interleaving the outputs is a relatively simple idea, while image and video superresolution through combining information under pixel shifts have been studied in detail before, for example in a variety of multi-image super resolution (MISR) methods. Burst super resolution is a research area directly focused on this problem, often employing polyphase decomposition (e.g. Cilia M, Valsesia D, Fracastoro G, Magli E. Multi-Level Fusion for Burst Super-Resolution with Deep Permutation-Invariant Conditioning. ICASSP 2023 (pp. 1-5)). Similarly, multiple existing super resolution methods employ polyphase decomposition tricks, for example, for improving computational efficiency (Sun B, Zhang Y, Jiang S, Fu Y. Hybrid pixel-unshuffled network for lightweight image super-resolution. AAAI 2023, Vol. 37:2375-2383), for enabling real-time latency (Shi W, Caballero J, Huszár F, Totz J, Aitken AP, Bishop R, Rueckert D, Wang Z. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. CVPR 2016 (pp. 1874-1883)), for remote sensing (Lin-Yang H, Jing-Hong L, Gang L. Super resolution of aerial image by means of polyphase components reconstruction. Acta Physica Sinica. 2015:64(11)), etc. The references are given simply as examples, so I wouldn’t expect the authors to necessarily cite them, but I would expect a discussion of the novelty of the paper’s main contributions in the context of the broader computer vision/super-resolution literature.

The polyphase fine-tuning setting: Is it basically translation augmentation? In other words, would the authors say that the main difference between the model with and without fine-tuning is the addition of subpixel-level translations?

Reproducibility: I found the experimental implementation details, especially for the polyphase fine-tuning experiments, very confusing. Even after looking at Appendix L, the fine-tuning setting was unclear to me. Will all the code be made publicly available? The ‘Reproducibility Statement’ does not mention it. I think sharing the experimental code after publication would greatly strengthen the paper and ease reproducibility concerns. For example in Fig. 3, it is unclear to me what kind of a network or architecture is used as the polyphase neural operator in the experiments. Is it similar to the baseline Low+Pad (Colagrande et al., 2026a)?

Baseline comparisons: Would it be possible to use one of the transformer-based methods (Li et al., 2023; Xiao et al., 2024; Colagrande et al., 2025) as baseline for showing the (in)effectiveness of resolution independent parametrization as discussed in Appendix B? Similarly, for the speed/efficiency comparisons in Fig. 6, only one baseline (DPOT) is used. Is it possible to add the latency and peak memory numbers for the other experiments (e.g. those in Fig. 5) as well?

Results: In Fig. 3, why are the red lines (direct zero-shot) not monotonically decreasing with resolution gap m, especially for Burgers and Reaction-diffusion datasets? Similarly, in Fig. 6, why is the L2 error not decreasing monotonically with model size for the ‘Zero-shot’ case?

Qualitative results: Some qualitative analysis/examples comparing different methods to the ground truth would be helpful (even if in the Appendix). Currently, the examples shown in Fig. 1 and Fig. 2 do not really show how the proposed method better conserves the fine, high-frequency details compared to compressed inference methods.

Minor points:

The abstract seems to be missing the introduction sentence(s).

The selection of the datasets are not justified/motivated.

Questions for authors
  • Can the authors discuss the paper’s main contributions in the context of the broader computer vision/super-resolution literature in which polyphase decomposition is used often?
  • Is the polyphase fine-tuning simply translation augmentation? Does it bring improvements on the identical model without fine-tuning?
  • Can the experimental implementation details, especially for the polyphase fine-tuning experiments, be clarified? Will all the experimental code be made publicly available?
  • In Fig. 3, what architecture is used as the polyphase neural operator?
  • Would it be possible to use one of the transformer-based methods (Li et al., 2023; Xiao et al., 2024; Colagrande et al., 2025) as baseline for showing the (in)effectiveness of resolution independent parametrization as discussed in Appendix B?
  • For the speed/efficiency comparisons in Fig. 6, is it possible to add the latency and peak memory numbers for the other experiments (e.g. those in Fig. 5) as well?
  • In Fig. 3, why are the red lines (direct zero-shot) not monotonically decreasing with resolution gap m, especially for Burgers and Reaction-diffusion datasets? Similarly, in Fig. 6, why is the L2 error not decreasing monotonically with model size for the ‘Zero-shot’ case?
  • Is it possible to add some qualitative analysis/examples comparing different methods to the ground (maybe in the Appendix)?
5 comments
AuthorsOpenReview ↗

Part 1/3 - Response to Reviewer 7KfD

First of all, thank you very much for the time and care you have invested in reviewing our paper. We sincerely appreciate it.

W1, Q1: Can the authors discuss the paper’s main contributions in the context of the broader computer vision/super-resolution literature in which polyphase decomposition is used often?

We agree that the connection to prior work that uses polyphase rearrangement should be discussed more explicitly. We acknowledge that polyphase decompositions and closely related space-to-depth/pixel-unshuffle operations are classical techniques in previous signal processing and computer vision works, and our novelty claim is not the polyphase rearrangement itself. However, conventional (or burst) super-resolution generally starts from one (or several) low-resolution observations and learns to reconstruct missing high-frequency content, usually using high-resolution supervision, whereas our setting starts from a high-resolution physical field and a model pretrained on a coarser discretization. In our work, polyphase inference exactly rearranges all fine-grid samples into native-resolution components, applies the existing operator to each component, and recomposes the low-resolution predictions into the high-resolution output: it does not synthesize missing samples. We added this distinction in the related work Sec.2 and discuss in depth the relevant sub-pixel/super-resolution literature in the Extended related work section in App.B.

W2, Q2: Is the polyphase fine-tuning simply translation augmentation? Does it bring improvements on the identical model without fine-tuning?

Polyphase decomposition can be viewed as the extraction of shifted coarse-grid sublattices, but polyphase fine-tuning (PPFT) goes beyond standard translation augmentation. In particular, a key aspect of our fine-tuning approach is that the decomposition and recomposition operators are learned rather than fixed. Specifically, they are initialized as polyphase decomposition and recomposition, respectively, and parameterized through convolutions in the high-resolution space. This allows the model to learn how to decompose and recombine the high-resolution information, going beyond simple translation augmentation.

To measure the benefit of fine-tuning, we compare PPFT with zero-shot polyphase inference (PPI), which corresponds to its initialization. Importantly, PPI and PPFT use exactly the same pretrained FNO checkpoints, with the FNO weights kept frozen in both cases. For PPFT, only the weights of the added convolutions are updated. So, the only difference between PPFT and PPI is that PPFT learns the decomposition and recomposition operators using a small amount of high-resolution data.

On Darcy flow, PPFT consistently improves over PPI, as shown by the relative L2L^2 errors below:

Resolution transferPPI (no fine-tuning)PPFT (1% data)PPFT (10% data)
162→128216^2 \to 128^20.09190.074940.06466
322→128232^2 \to 128^20.05280.049870.04668
642→128264^2 \to 128^20.04610.036350.03627

These results correspond to approximately 6–21% lower error using 1% of the high-resolution data and 12–30% using 10%, compared to PPI.

On the other three datasets, fine-tuning provides little additional improvement over zero-shot PPI. We note that on compressible Navier–Stokes, PPI already achieves errors comparable to those of a model trained directly at high resolution, while on Burgers and reaction–diffusion it approaches this performance for smaller resolution gaps. Overall, these experiments show that fine-tuning the decomposition and recomposition operators with very lightweight modules (a single convolution each) can improve performance even when the pretrained FNO is left unchanged, although the benefit depends on the dataset and resolution gap.

W3, Q3: Can the experimental implementation details be clarified? Will all the experimental code be made publicly available?

We have clarified the implementation details in the revised manuscript (Appendices N and O), and the full experimental code will be made publicly available after publication.

Q4: In Fig. 3, what architecture is used as the polyphase neural operator?

Fig. 3 adopts the Fourier Neural Operator pretrained on different datasets, at different resolutions, and evaluate it using different inference methods (direct evaluation, low+pad and Polyphase), we added that to the caption. Thanks for pointing it out.

AuthorsOpenReview ↗

Part 2/3 - Response to Reviewer 7KfD

Q5: Would it be possible to use one of the transformer-based methods (Li et al., 2023; Xiao et al., 2024; Colagrande et al., 2025) as baseline for showing the (in)effectiveness of resolution independent parametrization as discussed in Appendix B?

We have added experiments with ONO (Xiao et al., 2024), following the same setup as in Sec. 5.3. ONO was originally proposed with two attention variants: Nyström and Linear. We evaluated both variants and found that ONO with Nyström attention exhibits substantially higher errors at higher resolutions on two of the four datasets, while ONO with linear attention remains robust across all four datasets at the tested resolution gap m=4m=4. When direct high-resolution inference is already robust, PPI achieves comparable performance. Conversely, when ONO suffers from performance degradation at higher resolutions, PPI significantly reduces the error. We report these additional results in the following table.

Each entry reports test relative L2L^2 error / latency (ms) / peak memory (MiB). For Polyphase inference, latency and peak memory correspond to the parallel implementation. Lower is better.

DatasetONO Nyström: DirectONO Nyström: PolyphaseONO Linear: DirectONO Linear: Polyphase
Burgers1.4430 / 23.48 / 2169.61.4298 / 37.09 / 2861.11.6568 / 5.52 / 924.31.6548 / 5.55 / 945.1
Darcy0.5250 / 3.19 / 7209.40.0772 / 3.91 / 7436.10.0707 / 0.93 / 2455.90.0710 / 0.89 / 2472.1
Reaction–diffusion3.1201 / 289.15 / 19274.01.1164 / 353.92 / 19900.00.9152 / 83.39 / 10826.10.9160 / 157.72 / 11002.1
Compressible Navier–Stokes0.1568 / 35.86 / 8252.90.1353 / 43.90 / 8653.80.2711 / 10.78 / 3419.90.2711 / 18.54 / 3740.0

We updated Fig. 5 to include these results and added a detailed results table in Tab. 21.

Q6: For the speed/efficiency comparisons in Fig. 6, is it possible to add the latency and peak memory numbers for the other experiments (e.g. those in Fig. 5) as well?

We added latency and peak memory in the experiments of Fig.5 in Tab.17-21 and we added in Fig.9 a visualization similar to Fig.6. These results show that the sequential implementation of PPI saves significant memory (with the exception of side effects of memory allocation of cuDNN, visible for example in two UNet runs where the low-resolution inference and sequential PPI have higher memory peak than high-resolution inference) at the cost of longer runtime. The parallel implementation, for three of the four models, doesn't improve wall-clock time, but we note that in these cases PPI has roughly the same latency as the low-resolution inference, which we consider a lower bound. In the case of transolver, high-resolution inference is substantially slower than low-resolution inference, and in this case parallel PPI reduces the latency significantly.

The efficiency benefits are significant at larger scales. To demonstrate this, we report in Fig. 7 experiments scaling the resolution in 1D, 2D, 3D, and 4D until reaching out-of-memory (OOM) on an NVIDIA A100 GPU with 80 GB of memory for FNO models. At the largest resolutions, parallel PPI achieves speedups over direct FNO of 3.59× in 1D (50,331,648 points), 3.36× in 2D (716827168^2), 2.99× in 3D (3683368^3), and 3.04× in 4D (80480^4), with comparable peak memory consumption. We added these results as Tab.34-37 in App.L.1 and a visualization of the scaling in Fig.7.

AuthorsOpenReview ↗

Part 3/3 - Response to Reviewer 7KfD

Q7: In Fig. 3, why are the red lines (direct zero-shot) not monotonically decreasing with resolution gap m, especially for Burgers and Reaction-diffusion datasets? Similarly, in Fig. 6, why is the L2 error not decreasing monotonically with model size for the ‘Zero-shot’ case?

Although a higher pretraining resolution generally improves native-resolution accuracy, it does not guarantee better zero-shot generalization. Proposition 1 shows that the cross-resolution discrepancy of an FNO depends on aliasing in the input and intermediate representations, rather than solely on the resolution gap. This provides a possible explanation for the observed non-monotonicity. Since FNOs pretrained at different resolutions learn different representations, a model pretrained at a lower resolution can sometimes generalize better to the target resolution.

Q8: Is it possible to add some qualitative analysis/examples comparing different methods to the ground (maybe in the Appendix)?

We added visualizations of the predictions in Fig.10-13 and visualizations of the radially averaged power spectral density (RAPSD) in Fig.14. The RAPSD is particularly interesting as it shows that Low+Pad fails to capture high frequencies (which are set to zero by construction), while PPI predicts high-frequency content with a spectral profile that is, on average, closer to the ground truth. This is particularly visible at the highest frequencies, where direct zero-shot inference exhibits excessively high spectral power.

Minor point 1: The abstract seems to be missing the introduction sentence(s).

Thanks for pointing this out. We have added back the sentences that were cut by mistake.

Minor point 2: The selection of the datasets are not justified/motivated.

We selected the Burgers, Darcy and Compressible Navier-Stokes datasets to allow comparison with previous work, adopting those used in (Sakarvadia et al., 2026). We additionally included Reaction–Diffusion because its solutions contain high-frequency spatial structures, making the challenge of cross-resolution adaptation particularly relevant. In addition, following the suggestion of Reviewer Zro8, we have also added two 3D datasets (Swift–Hohenberg and Kuramoto–Sivashinsky) to evaluate our method on 3D data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.