acceptodds
Under review as a conference paper at ICLR 2027

SS-REPA: Spectral-Spatial Representation Alignment for Onboard Sparse Physical Field Reconstruction

Abstract

Onboard sparse physical field reconstruction aims to recover dense physical states from a limited number of sensor measurements under tight memory and latency constraints. Compact convolutional encoder–decoders are attractive for deployment, but their capacity-constrained bottlenecks may struggle to capture the field-wide context required to infer unobserved regions. We propose Spectral-Spatial Representation Alignment (SS-REPA), a framework for improving bottleneck representations in compact convolutional encoder–decoders through alignment with a pretrained spectral neural operator. Specifically, SS-REPA introduces a dual-branch bottleneck with a mean-shift spatial branch and a spectral-context branch, providing separate pathways for local features reconstruction and global spectral interaction. Only the spectral-context branch features are mapped to the spectral neural operator feature space, where a Gaussian-weighted frequency-domain objective aligns their real and imaginary components while emphasizing global low-frequency structure. Extensive experiments on three sparse-sensor reconstruction benchmarks demonstrate that our method achieves competitive reconstruction accuracy with approximately 94% fewer parameters and 83.3–85.5% lower inference latency than the standard baseline on an edge computing device.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Onboard sparse physical field reconstruction, convolutional encoder–decoder network, spectral neural operator, bottleneck representation, representation alignment

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_2c2VOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The work proposes a spatial field reconstruction method for onboard deployment. The work uses an existing U-Net architecture with its bottleneck changed. A decoupled dual branch bottleneck is proposed with separate spatial and spectral branches, where the spectral branch is supervised by a frozen spectral neural operator. Experiments are shown on AirfRANS, Three-cylinder flow and Turbulent channel flow with parameters, FLOPS and Inference time is given for all 3 datasets strengthening edge device applicable robustness. Ablations are provided with various settings of the proposed model to support contribution of each component.

Critical strengths
  1. The edge device deployment problem is of high practical relevance and is also supported with evidence.
  2. The paper is clearly written and very easy to read and follow and architectural improvements are supported by ablations.
Critical weaknesses
  1. The work is a combination of existing papers or components from different papers rather than a novel contribution: Specifically U-REPA (NeurIPS 2025) and REPA-P (ICML 2026) are some of closely related papers, including applications to scientific domain.
  2. Claimed efficiency gain in abstract is not a fair comparison, as it largely reflects the choice of baseline: The reported 94% parameter reduction and around 84% lower inference latency than standard baseline are obtained by comparing SS-REPA (built on compact U-Net 0.86M) with Standard 13.39M parameter U-Net.

References:

  1. Tian, Y., Chen, H., Zheng, M., Liang, Y., Xu, C., & Wang, Y. (2026). U-repa: Aligning diffusion u-nets to vits. Advances in Neural Information Processing Systems, 38, 11003-11024.
  2. Jia, H., Yin, P., Chen, W., Liang, S., Wang, L., Tian, B., ... & Yue, Y. (2026). Learning to Think in Physics: Breaking Shortcut Learning in Scientific Diffusion via Representation Alignment. arXiv preprint arXiv:2605.20780.
Questions for authors
  1. Have the authors compared SS-REPA against a parameter-matched UNet (around 0.86M) with a standard convolutional bottleneck? Reported Standard U-Net small is (0.5 M parameters from Figure 2) which seems not a fair ablation.
  2. Could the authors compare SS-REPA against conventional feature-level and output-level knowledge distillation using the same RecFNO teacher and compact student backbone?
  3. The Gaussian bandwidth is fixed at σ = 0.5 throughout the experiments. How sensitive is the performance to this choice?
  4. Since the method aims onboard sparse-sensor reconstruction, how robust is SS-REPA to sensor noise, missing sensors, and changes in sensor placement?
  5. Could the authors report mean ± standard deviation over multiple seeds ?
  6. Given that SS-REPA targets sparse-sensor physical-field reconstruction, could the authors extend their comparisons to recent latent neural operators (LNOs) and implicit neural representation methods, particularly MMGN (ICLR 2024), SIREN, and WIRE (CVPR 2023)?

References:

  1. Wang, T., & Wang, C. (2024). Latent neural operator for solving forward and inverse pde problems. Advances in Neural Information Processing Systems, 37, 33085-33107.
  2. Luo, X., Xu, W., Nadiga, B. T., Ren, Y., & Yoo, S. (2024, May). Continuous field reconstruction from sparse observations with implicit neural networks. In International Conference on Learning Representations (Vol. 2024, pp. 3281-3305).
  3. Sitzmann, V., Martel, J., Bergman, A., Lindell, D., & Wetzstein, G. (2020). Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33, 7462-7473.
  4. Saragadam, V., LeJeune, D., Tan, J., Balakrishnan, G., Veeraraghavan, A., & Baraniuk, R. G. (2023, June). Wire: Wavelet implicit neural representations. In 2023 IEEE/CVF conference on computer vision and pattern recognition (CVPR) (pp. 18507-18516). IEEE.
4 comments

W1: Novel contribution.

We thank the reviewer for raising this point and for pointing us to these closely related works.

We would like to clarify the nature of our contribution. SS-REPA is designed for compact convolutional encoder–decoders under onboard deployment constraints, where memory footprint and inference latency are first-order design requirements. Multi-level spatial convolutions are well suited to capturing local detail in physical fields and are deployment-friendly on edge devices, but they are weak at capturing global structure. Spectral neural operators, by contrast, encode global information through spectral features and therefore provide a useful spectral prior. SS-REPA is built on this complementarity in a deployment-driven way: it keeps the compact convolutional encoder–decoder as the deployed model and injects the spectral prior during training through branch-level selective alignment.

U-REPA (diffusion U-Net → ViT) differs from SS-REPA in four fundamental respects. (i) Application domain: U-REPA targets generative diffusion modeling, where the goal is to improve sample generation quality; SS-REPA targets onboard sparse physical-field reconstruction, where the goal is to recover dense physical states from sparse sensor measurements under tight memory and latency budgets. (ii) Model: U-REPA aligns a diffusion U-Net student to a ViT teacher; SS-REPA aligns a compact convolutional encoder–decoder. (iii) Motivation: U-REPA's alignment is intended to improve generative training dynamics; SS-REPA's alignment is intended to compensate for the limited global aggregation of a capacity-constrained convolutional bottleneck. (iv) Architectural design: U-REPA aligns a full hidden block and has no notion of branch-level selectivity; SS-REPA introduces a decoupled dual-branch bottleneck and aligns only its spectral branch, leaving the spatial branch free to preserve local detail.

REPA-P (teacher-free, physics-prior-driven) differs from SS-REPA in its supervision signal, its output, and its task. REPA-P is teacher-free and supervises training with PDE/boundary-condition residuals, and it decodes features into physical states; it is a generative model for scientific diffusion. SS-REPA is data-driven: it aligns to features of a pretrained spectral operator, with no physics residual and no physical-state decoding, and it is a discriminative reconstructor for sparse-field reconstruction. The two methods therefore rely on fundamentally different sources of prior knowledge—physics residuals versus a learned spectral representation—and address fundamentally different tasks—generative modeling versus discriminative reconstruction. The resemblance is at the level of using representation alignment as a training mechanism; the substance of what is aligned, why, and for what deployed purpose is different.

W2 and Q1: parameter-matched UNet (around 0.86M).

We thank the reviewer. We have added a near-parameter-matched reference, the Standard UNet (0.92M), as described in lines 359–360.

Q2: compare SS-REPA against conventional feature-level and output-level knowledge distillation?

We thank the reviewer for this suggestion. This comparison is reported in the ablation study (lines 411–413), with results in Table 3.

Q3: How sensitive is the performance to σ = 0.5?

We thank the reviewer for this suggestion. The sensitivity of SS-REPA to the bandwidth σ is reported in the revised Section 4 ("Effect of bandwidth σ", lines 507–510).

Q4: How robust is SS-REPA to sensor noise, missing sensors, and changes in sensor placement?

We thank the reviewer for this question, which we address along three axes. Sensor noise. A noise-robustness analysis is added in Section 4.3 (“Noisy Robustness”, lines 512-518).

Missing sensors and changes in sensor placement. These are evaluated on the Turbulent channel flow benchmark, whose protocol and sensor encoding are described in Appendix A.1.1.

Q5: Report mean ± standard deviation over multiple seeds.

We thank the reviewer for this request. We repeated the main accuracy comparisons over three random seeds (42, 123, 456); Appendix Tables 6–8 and 10 report means ± standard deviations. Tables 3 and 5 report three-seed means for the ablation settings.

Q6: Comparison with latent neural operators.

We thank the reviewer for this constructive suggestion. We have added comparisons against LNO, MMGN, SIREN, and WIRE to the main experiments (Table 1).

Among these, SIREN, WIRE, and LNO were originally proposed for other purposes: SIREN as a general implicit neural representation with periodic activations, WIRE as a wavelet-based implicit representation for image and signal reconstruction, and LNO for forward and inverse PDE problems in a latent space. We therefore adapt them to the sparse-field reconstruction setting rather than applying them out of the box, as described below and in Appendix A.1.3 (lines 910–951).

Reviewer_2c2VOpenReview ↗

Response to Official Comment by Authors

I thank the authors for the detailed rebuttal and substantial revision, many of my concerns got resolved. Considering the rebuttal, i still have some concerns:

  1. Novelty is incremental, authors tried to explain the differences from existing closely related methods, but that does not rule out incremental novelty issue.
  2. The abstract and contributions still report the 94% parameter and 83.4–85.5% latency reductions against the 13.39M UNet.
  3. Improvement of model over baselines is marginal and within noise, Considering Tables 6, 7 and 8 (seed robustness tests), gains over the strongest baselines are small relative to seed variance on AirfRANS and partly on turbulent channel flow.

Therefore, i retain my score.

AuthorsOpenReview ↗

Novelty: the contribution is a deployment-driven design, not a combination of existing components.

We thank the reviewer very much for the careful re-reading and for the constructive follow-up. Our responses to the three remaining points are as follows.

1. Novelty: the contribution is a deployment-driven design, not a combination of existing components.

We respectfully clarify that SS-REPA is not proposed as a generic accuracy-oriented architecture, but as a deployment-driven method for onboard sparse physical-field reconstruction under tight memory and latency budgets. This framing is not a niche motivation: the escalating memory, latency, and energy costs of deploying deep models on edge devices are widely recognized as a central bottleneck for edge AI, and this problem has been explicitly highlighted in leading venues.

Nature Communications (2025) states:

“Artificial intelligence (AI) has made significant strides towards efficient online processing of sensory signals at the edge through the use of deep neural networks with ever-expanding size. However, this trend has brought with it escalating computational costs and energy consumption, which have become major obstacles to the deployment and further upscaling of these models.”

Science (2022) similarly notes:

“Advanced machine learning models are currently impossible to run on edge devices such as smart sensors and unmanned aerial vehicles owing to constraints on power, processing, and memory.”

These statements reflect a broadly shared concern across the community rather than isolated observations: the deployment constraints of memory, latency, and energy are first-order obstacles, not secondary considerations. This is precisely the setting that motivates SS-REPA. References

  • Exploiting neuro-inspired dynamic sparsity for energy-efficient intelligent perception. Nature communication 2025.

  • Delocalized photonic deep learning on the internet‘s edge. Science 2022.

In this setting, accuracy is not the sole—or even the first-order—design criterion; memory footprint and inference latency are first-order requirements. The relevant question is whether a compact deployed model can retain accuracy comparable to much larger or stronger baselines after aggressive compression.

Our specific contributions are:

  1. a decoupled dual-branch bottleneck for a deployment-friendly convolutional encoder–decoder,
  2. selective alignment of only the spectral-context branch to a frozen spectral neural operator teacher, with the teacher discarded at inference.

The decoupled dual-branch bottleneck is a general, backbone-agnostic architectural component: it can be inserted into arbitrary convolutional encoder–decoder backbones, and the selective spectral alignment is a training-time add-on that leaves inference unchanged. To our knowledge, this decoupled dual-branch design with branch-level selective spectral alignment has not been proposed in prior work. More importantly, the ablation shows that neither component alone explains the result. Using the data in Table 1 and the ablation study:

UNet + decoupled dual-branch already reduces error relative to the parameter-matched Standard UNet (0.92M) by 2.4% / 34.1% / 17.8% on AirfRANS (8 sensors), Three-cylinder (4/cyl), and Turbulent (100 sensors), respectively.

UNet + alignment (i.e., the U-REPA-style single-branch model with alignment but without the decoupled dual-branch design) does not improve over the parameter-matched UNet on AirfRANS (−9.4%, i.e. worse), and is only marginally better on Three-cylinder (+1.1%) and Turbulent (+6.7%). Relative to UNet + decoupled dual-branch, it is actually 12.1% / 50.0% / 13.6% worse.

Thus, simply transferring an alignment mechanism to a single-branch compact model is insufficient. These results show that the gain is not a simple additive combination of existing components: the decoupled dual-branch architecture and the selective spectral-branch alignment are mutually necessary. The ablation therefore supports a design contribution, not an incremental stacking of known modules.

VariantAirfRANS (8)Three-cyl. (4/cyl)Turbulent (100)
Standard UNet (0.92M)0.2970.0910.359
UNet + decoupled dual-branch0.2900.0600.295
UNet + alignment0.3250.0900.335
UNet + decoupled dual-branch (w/o mean-shift)0.2810.0750.304
UNet + decoupled dual-branch + alignment (Full)0.2510.0430.231
  • UNet + decoupled dual-branch + alignment vs. UNet + alignment: 22.8 / 52.2 / 31.0% lower error.
  • UNet + alignment vs. UNet + decoupled dual-branch: −12.1 / −50.0 / −13.6% (worse).

On effective receptive field and long-range dependence. Figure 5 shows that the architectural improvement effectively enlarges the effective receptive field and strengthens long-range input–output dependence, which existing single-branch compact methods do not provide.

AuthorsOpenReview ↗

Fair comparison and marginal gains.

2. Fair comparison.

We thank the reviewer for this observation. We have revised both the abstract and the contributions to report two complementary comparisons rather than the efficiency numbers alone. We agree that these efficiency numbers are obtained relative to the 13.39M standard UNet and that they reflect this choice of baseline. We retain this comparison as a primary one because the paper targets onboard sparse-field reconstruction, where the standard UNet is the baseline conventionally considered for deployment, and the efficiency numbers directly quantify the end-to-end deployment cost relative to that reference. At the same time, the revision adds a parameter-matched baseline comparison to show that the accuracy advantage is not merely a consequence of capacity: relative to the parameter-matched baseline, SS-REPA reduces relative ℓ₂ error by 15.5% on AirfRANS-8, 52.7% on Three-cylinder, and 35.7% on Turbulent-100. These two comparisons answer different questions and are intended to complement each other.

3. Marginal gains and seed variance. We thank the reviewer for this careful reading of Tables 6–8. We agree that on AirfRANS and parts of the turbulent channel flow benchmark, the absolute accuracy gap over the strongest baselines is small and close to seed variance, and we do not claim that this gap is statistically decisive on every configuration. However, as also emphasized in our response to the novelty concern, SS-REPA is deployment-driven: under limited onboard deployment resources, accuracy is not the only—or even the first-order—design criterion; memory footprint and inference latency are first-order requirements. This point is more clearly illustrated in Figure 2, which presents the accuracy-efficiency trade-off across the three benchmarks. Although the absolute accuracy margin is modest on some configurations, SS-REPA consistently lies on the favorable side of the trade-off, achieving competitive reconstruction accuracy at substantially lower parameter count and inference latency. Judging the contribution solely by the accuracy margin is therefore incomplete and does not reflect the deployment-oriented objective of the paper.

Against the strongest accuracy baseline on each benchmark, SS-REPA reduces parameters by about 91–92%, FLOPs by 59.1–99.4%, and inference latency by 93.5–98.6%, while remaining competitive in accuracy. This is the deployment-relevant trade-off the paper targets, and we have made this framing more explicit in the revision.

Efficiency comparison against the strongest accuracy baseline per benchmark (lowest average error across sensor settings, excluding SS-REPA variants). Each cell shows baseline → SS-REPA (reduction). Inference time is ms/sample.

MetricAirfRANS vs. SenseiverThree-cylinder vs. IRNOTurbulent vs. GNO
Params (M)10.11 → 0.86 (−91.5%)10.57 → 0.86 (−91.9%)9.59 → 0.86 (−91.0%)
FLOPs (G)31.38 → 1.66 (−94.7%)5.52 → 2.26 (−59.1%)54.55 → 0.31 (−99.4%)
Latency (ms)53.46 → 3.47 (−93.5%)126.73 → 4.69 (−96.3%)59.26 → 0.81 (−98.6%)

We hope this clarifies that the intended claim is not that SS-REPA dominates by a large accuracy margin, but that it achieves competitive accuracy at a small fraction of the deployed parameter and latency cost—the trade-off that matters for onboard deployment.

We sincerely thank the reviewer again for the detailed engagement.

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

This paper proposes SS-REPA for improving compact convolutional encoder–decoder networks for sparse physical-field reconstruction. A frozen, pretrained RecFNO provides intermediate feature targets during training. The student incorporates a dual-branch bottleneck consisting of a mean-subtracted spatial branch and a frequency-domain channel-mixing branch. The latter is aligned with the teacher through a Gaussian-weighted spectral loss. Experiments cover AirfRANS, three-cylinder flow, and turbulent channel flow, including inference measurements on an NVIDIA Orin Nano.

Critical strengths
  1. Important and well-motivated problem. Recovering dense physical fields from sparse measurements under strict computational constraints is relevant to onboard estimation and real-time scientific applications. The paper appropriately treats inference efficiency as a first-class objective rather than focusing exclusively on reconstruction accuracy.

  2. Reasonable architectural decomposition. Separating spatial-detail and spectral-context pathways is a sensible way to address the competing requirements of local fidelity and global structure. Aligning only the spectral-context branch also avoids imposing the same representation constraint on features serving different purposes.

  3. Promising empirical results. The method is evaluated on three benchmarks with different flow characteristics and multiple sensor configurations. The ablations in Table 3 indicate that removing the alignment objective, dual-branch structure, or mean-shift component can degrade performance, particularly on the three-cylinder benchmark.

Critical weaknesses
  1. The methodological novelty relative to existing representation alignment and teacher–student approaches needs to be established more clearly. The main component, dual-branch feature extraction, alignment to a pretrained teacher, frequency-domain losses, and progressive loss weighting, are individually plausible but not entirely new. The paper's contribution appears to lie in their particular combination for sparse physical field reconstruction.

  2. The frequency-selective alignment mechanism is not sufficiently validated. The paper argues that low-frequency components provide global physical structure, whereas high-frequency components are more effectively captured by spatial convolutions. This is a plausible hypothesis, but the current evidence primarily consists of downstream reconstruction metrics and component ablations.

  3. The efficiency claims require a more comprehensive accounting. The reported parameter and latency reductions are encouraging, but the comparison against a standard U-Net does not by itself establish that SS-REPA is the most effective accuracy–efficiency trade-off. The paper should distinguish the cost of the deployed student from the total training cost, including teacher pretraining, teacher inference during alignment, and additional optimization.

Questions for authors
  1. How much of the improvement remains when the dual-branch architecture is retained but the spectral teacher and alignment loss are removed?

  2. Does the proposed frequency-domain loss consistently improve low-frequency reconstruction, and what happens to high-frequency error?

  3. What is the total training-time overhead of SS-REPA relative to training the student alone?

  4. How does the method perform when evaluated on unseen geometries or sensor configurations rather than unseen samples from the same configurations?

1 comment

W1: The methodological novelty relative to existing representation alignment and teacher–student approaches needs to be established more clearly.

We thank the reviewer for this comment. Our contribution is not representation alignment itself, but a narrower, deployment-driven combination: branch-level selective alignment—only the spectral branch of the bottleneck is aligned to a frozen teacher, while the spatial branch stays free—with a spectral operator as the teacher, under strict onboard parameter and latency constraints. The teacher is discarded at inference.

Why the problem is hard. Onboard real-time field reconstruction is constrained by memory and latency. Under a strict budget, the compact convolutional encoder–decoder is the natural backbone, but reducing width and depth also restricts what its bottleneck can represent. The bottleneck must aggregate widely separated measurements and encode global context to infer large unobserved regions—exactly where a capacity-constrained convolutional model fails, since its bottleneck is dominated by local spatial features. Existing remedies either enlarge the deployed model or replace it with a spectral operator or Transformer, both defeating the onboard constraint. The question is therefore: how can the bottleneck of a compact encoder–decoder be improved without substantially enlarging the deployed model?

Why existing alignment approaches do not directly apply. These methods assume representational compatibility that does not hold here: the student's bottleneck is a spatial convolutional feature, while the teacher's feature is spectral, with different resolution, channel organization, and statistics. Aligning the entire bottleneck imposes conflicting demands—preserving local detail while assimilating global spectral structure.

What is new in our design. This motivates the decomposition: a decoupled dual-branch bottleneck, separating a mean-shift spatial branch from a spectral-context branch and fusing them before decoding; and selective alignment of only the spectral branch, leaving the spatial branch free to preserve local detail. The teacher is used only at training time and removed at inference. To our knowledge, this is the first framework using a pretrained spectral neural operator as a training-time prior to improve the bottleneck of a compact convolutional encoder–decoder for onboard sparse field reconstruction.

Relation to U-REPA and REPA-P. Both U-REPA (NeurIPS 2025) and REPA-P (ICML 2026) are generative models. U-REPA aligns a full U-Net hidden block to a ViT for generative diffusion modeling and has no branch-level selective alignment; SS-REPA is a discriminative reconstructor built on a compact convolutional encoder–decoder, aligning only one branch under the onboard accuracy–latency trade-off. REPA-P is a teacher-free, physics-prior-driven generative model that supervises with PDE/boundary-condition residuals and decodes features into physical states; SS-REPA is data-driven, aligning to features of a pretrained spectral operator with no physics residual and no physical-state decoding. The tasks also differ—scientific diffusion generation versus sparse-field reconstruction with a compact convolutional encoder–decoder.

Q1: How much of the improvement remains when the dual-branch architecture is retained but the spectral teacher and alignment loss are removed.

We thank the reviewer for this question. This comparison was already reported in the ablation study as SS-REPA w/o alignment. In this revision we have moved this row into the main comparison (Table 1).

Q2: Does the proposed frequency-domain loss consistently improve low-frequency reconstruction, and what happens to high-frequency error?

We thank the reviewer for this question. To address it directly, we have added an analysis to Appendix A.2 ("Training low-/high-frequency reconstruction losses") that decomposes the reconstruction error into low- and high-frequency components during training. This addresses the question: the frequency-domain loss consistently improves low-frequency reconstruction, and it does not do so at the expense of high-frequency error.

Q3: What is the total training-time overhead of SS-REPA relative to training the student alone?

We thank the reviewer for this question. It is discussed in Section 4.3 (“Training Overhead”).

Q4: How does the method perform when evaluated on unseen geometries or sensor configurations rather than unseen samples from the same configurations?

We thank the reviewer for this question. We have evaluated such settings on two benchmarks, and the revision states this more explicitly (lines 484–485, Appendix A.1.1). Unseen geometries and operating conditions (AirfRANS). Unseen sensor configurations (Turbulent channel flow).

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

The paper looks at reconstructing dense physical fields from sparse sensor readings on hardware with tight memory and latency budgets. The proposed method, SS-REPA, trains a small convolutional encoder-decoder (a 0.86M-parameter UNet) with the help of a frozen RecFNO teacher. Its bottleneck has two branches: one subtracts the spatial mean, and the other mixes channels in the Fourier domain. Only the Fourier branch is aligned to the teacher, by projecting it to the teacher's feature space and applying a Gaussian-weighted loss on the Fourier coefficients. The teacher and the projection head are dropped at inference. Experiments cover three flow datasets (AirfRANS, three-cylinder flow, turbulent channel flow) under several sensor counts. Besides the main comparison, there are ablations and some analysis of the bottleneck and the receptive field, and latency is measured on an NVIDIA Orin Nano.

Critical strengths

Accuracy at a small deployment cost, with latency measured on real hardware. The paper is motivated by onboard use, and it backs this up with Orin Nano timings instead of stopping at parameter and FLOP counts, which I appreciate. With 0.86M parameters, SS-REPA matches or beats every baseline in Table 1 at every sensor setting, and its latency (3.47/4.69/0.81 ms on the three benchmarks) is 83-85% lower than that of the 13.39M standard UNet. Part of that saving just comes from using a small backbone. The comparison I find more telling is MobileNet-UNet, the only baseline of similar size (0.71M): SS-REPA is faster on the device (3.47 vs. 5.56 ms on AirfRANS) and clearly more accurate (0.254 vs. 0.297 with 8 sensors). The teacher is only needed during training.

The ablations deal with the most obvious alternative explanation, that the small backbone alone is responsible. Table 3 compares the full model with a UNet-small of the same size, a dual-branch model trained without alignment, a version without the mean-shift step, and a version that replaces the frequency-domain loss with a plain spatial ℓ2\ell_2 loss. Adding alignment to the dual-branch model lowers the error from 0.292/0.061/0.296 to 0.254/0.045/0.257, and UNet-small is well behind at 0.320/0.091/0.369. Whatever the mechanism turns out to be (see Weakness 1), these numbers convince me that the alignment training adds something beyond picking a smaller network.

Critical weaknesses
  1. The "spectral" branch and the frequency-domain loss reduce to simple spatial operations. According to Eq. (4) and Appendix A.1.2, WrW_r and WiW_i are bias-free 1x1 convolutions over channels, so the same channel map is applied at every frequency. For a real-valued map zz, Re(Fz)=F((z+Rz)/2)\mathrm{Re}(\mathcal{F}z) = \mathcal{F}\big((z + Rz)/2\big) and i Im(Fz)=F((z−Rz)/2)i\,\mathrm{Im}(\mathcal{F}z) = \mathcal{F}\big((z - Rz)/2\big), where Rz(p)=z(−p mod (HS,WS))Rz(p) = z(-p \bmod (H_S, W_S)) reflects the map through the origin with periodic wrap-around. Plugging this into Eq. (4) gives zSg=12(Wr+Wi) zS+12(Wr−Wi) RzSz_S^g = \tfrac{1}{2}(W_r + W_i)\, z_S + \tfrac{1}{2}(W_r - W_i)\, R z_S: a 1x1 conv of the bottleneck plus a 1x1 conv of its mirrored copy. If "shared" in Appendix A.1.2 means Wr=WiW_r = W_i, the branch is just a 1x1 convolution. Either way, it can't learn spatial interactions the way an FNO layer with frequency-dependent weights does. At most it links each location to one fixed mirrored location. The loss has the same issue. By Parseval's theorem, Eq. (7) equals, up to a constant, a spatial ℓ2\ell_2 loss between low-pass-filtered features, so the "w/o frequency-domain loss" ablation really compares low-pass weighting with no weighting. There may be steps in the code that the paper doesn't show (see Q2), and those could change this picture. As written, though, the paper's central story, a spectral prior injected through a spectral branch, rests on operations with simple spatial equivalents. This is the main reason for my score.

  2. It isn't clear where the gains come from, or how stable they are. No experiment checks whether the teacher needs to be a spectral operator (see Q3). Without that check, plain regularization could explain the results just as well: the RecFNO baseline (same configuration as the teacher) is worse than the student in every setting of Table 1, and the biggest gains are on three-cylinder flow, where the model sees only 50 training samples. If a generic teacher works equally well, the paper offers a regularizer for small UNets, and that is a narrower contribution than the one claimed. The analyses in Section 4.3 don't settle this either. Table 4 compares SS-REPA's fused bottleneck zS∗z_S^{\ast} with UNet-small's zSz_S (Appendix A.3), but zS∗z_S^{\ast} is partly built from zS−GAP(zS)z_S - \mathrm{GAP}(z_S), so it has a global pathway by construction. If the ERF is computed on the same tensor, even a small uniform component can look wide on a log color scale. The LogDet entropy also depends on the overall scale of the features, and that is where Table 4 shows the largest gain. Finally, no run-to-run variance is reported, while some margins in Table 1 are very small (0.290 vs. 0.290 and 0.233 vs. 0.235 on AirfRANS). I can't tell how much of the improvement would survive a change of seed.

Questions for authors
  1. How are the benchmarks split at the level of individual fields or snapshots, for both the student and the teacher? For turbulent channel flow, Appendix A.1.1 says the model is trained on all available simulation data and then evaluated on the corresponding fields with 30, 50, and 100 sensors. If the evaluated fields were also dense training targets, this benchmark tests robustness to new sensor layouts, not reconstruction of unseen flow states. That's a reasonable setting, but it should be said explicitly, and results on held-out snapshots would change how much weight I give this benchmark. For three-cylinder flow, are the 1000 samples snapshots of one simulation, and how close in time is each test snapshot to the nearest training one? I also couldn't find how the sparse measurements are turned into a grid input for SS-REPA and for the baselines (Voronoi tessellation, a masked field, something else?).

  2. In Eq. (4), are WrW_r and WiW_i the same weights (is that what "shared" means?), and are they applied identically at every Fourier mode? Does the code do anything the equation doesn't show, such as frequency-dependent weights, mixing between the real and imaginary parts (as in FFC), a nonlinearity in the frequency domain, or mode truncation? Any of these could change the analysis in Weakness 1, so a few lines of pseudocode for this branch would settle it.

  3. Do the gains hold up if the teacher is not a spectral operator? Two simple checks: the same projection and loss with (a) the 13.39M standard UNet as teacher, which is more accurate than RecFNO on AirfRANS, and (b) a randomly initialized RecFNO. For the dual-branch model trained without alignment, could you also report the ERF, the distance-conditioned Jacobian, and the Table 4 metrics, computed on the same tensor (zS∗z_S^{\ast}) as for the full model? That comparison would separate what the architecture contributes from what the teacher contributes. And if Eq. (4) is implemented exactly as written, how do you see the role of this branch, given the spatial form in Weakness 1?

  4. Are the accuracy numbers from a single run? On the AirfRANS AoA and Reynolds splits, the gaps to the standard UNet are tiny (0.2596 vs. 0.2600 and 0.2700 vs. 0.2738 in Table 6), and I'd want the mean and spread over a few seeds before reading them as a real advantage. A small thing in the same table: U-FNO and U-FFNO have identical AoA and Reynolds errors (0.3211/0.3211 and 0.3164/0.3164). Is that a copy error?

1 comment

W1 and Q2: The "spectral" branch and the frequency-domain loss reduce to simple spatial operations.

We thank the reviewer for this careful analysis, which identifies the key point that our implementation must be described precisely enough to rule out reduction to a plain spatial operation. We clarify below, and the revision expands the relevant implementation details in Appendix A.1.2. We also distinguish the linear limit from the actual nonlinear implementation and describe the frequency-weighted loss without overstating its mathematical novelty.

Linear limit and implemented nonlinearity. W_r and W_i are independently parameterized, bias-free 1×1 channel maps applied to the real and imaginary parts, respectively. They are frequency-uniform in the sense that the same channel map is applied at every Fourier mode, but they are distinct operators. Crucially, a GeLU nonlinearity is applied in the frequency domain after each channel mixing, before the inverse transform (Appendix A.1.2; Algorithm 1). We agree with the reviewer’s algebraic observation that, for real input z and frequency-independent bias-free maps W_r and W_i, if the activation is replaced by the identity, the inverse transform has the form

((W_r + W_i)/2) z(x) + ((W_r − W_i)/2) z(−x),

with z(−x) denoting the periodic spatial reversal. Thus even W_r ≠ W_i does not prevent the linear branch from reducing to a spatial mixing operation. In the implemented branch, however, the nonlinear frequency-domain operation is not, in general, the same fixed linear combination of z(x) and z(−x). We do not claim absolute non-equivalence to all spatial-domain operators; the claimed advantage is supported empirically by the spatial-mixing control in Table 5 rather than by such an assertion.

Controlled branch comparison. Appendix A.2 and Table 5 compare five instantiations of the branch under the same dual-branch topology: frequency-dependent weighting, FFC-style real/imaginary mixing, mode truncation, pure spatial channel mixing, and our nonlinear frequency-domain variant. Pure spatial mixing is worse than ours on the representative AirfRANS/Three-cylinder/Turbulent settings at the same reported 0.86M parameter count. A frequency-dependent variant attains lower errors than ours on two datasets, but uses 2.93M rather than 0.86M parameters. These results support an accuracy–parameter trade-off for the chosen nonlinear branch and indicate that the frequency-domain operation, not merely the channel-mixing capacity, is what the branch contributes.

Parseval equivalence. We have revised lines 226–228 to clarify this point. We state this explicitly and do not claim that the alignment objective is irreducible to spatial filtering. The design choice is its frequency selectivity: Gaussian weighting smoothly attenuates higher-frequency teacher constraints. Table 3 compares this with conventional unweighted feature MSE under the same reported teacher and student configuration, supporting its empirical usefulness.

Q1: benchmark splits, temporal proximity, and sensor-to-grid encoding.

We appreciate the request for clarity. We have revised lines 762–781 to explicitly describe the benchmark splits and sensor configurations.

Q3: Do the gains hold up if the teacher is not a spectral operator?

We thank the reviewer for these concrete checks. The revised version now reports these experiments in the main experiments, the ablation study, and the sections on bottleneck information improvement and effective receptive field.

(a) Standard UNet as teacher. Table 1 includes SS-REPA (UNet), aligned to bottleneck features of the 13.39M Standard UNet, using the same projection and frequency-weighted loss. This improves over the dual-branch model without alignment in eight of nine sensor settings. The effect is weaker than with RecFNO and GNO teachers across the representative comparisons, even though the full Standard UNet itself performs better than RecFNO on AirfRANS. These observations suggest that transfer benefit cannot be predicted from teacher output accuracy alone; what matters is whether its representation carries the global spectral structure the compact student lacks.

(b) Randomly initialized RecFNO as teacher. The ablation “Dual-branch w/ randomly initialized alignment” uses features from a randomly initialized RecFNO, with everything else unchanged. It degrades clearly on all three benchmarks, indicating the gain requires a trained spectral representation and is not produced by any fixed random projection.

Q4: Are the accuracy numbers from a single run?

We thank the reviewer for identifying both issues. The U-FNO/U-FFNO duplicate was a copy error; corrected values are in Table 10. The revised main benchmark (Appendix Tables 6–8) and AirfRANS split comparisons (Table 10) report mean±std over three seeds (42, 123, 456).

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

The authors introduce SS-REPA, a training-time regularization method that aligns the spectral branch of a dual-branch latent space of a UNet architecture with features from a pretrained Fourier neural operator via a loss in the frequency domain. The method yields meaningful efficiency gains on edge hardware, though the absolute accuracy improvements over a standard UNet are modest on two of three benchmarks.

The method is evaluated on three 2d CFD benchmarks and achieves good accuracy in most settings, with a much lower number of parameters and correspondingly lower inference time compared to standard UNet. Detailed ablation studies confirm each component contributes as expected.

Critical strengths
  • Interesting idea to only regularize on the spectral part, combining the large model benefits from a trained neural operator and the edge-device setting.
  • Well-motivated practical setting with thorough edge-device evaluation.
  • Very detailed ablation studies and computational experiments
  • May be possible to translate this idea to other fields, regularizing a smaller network with a large neural operator is applicable more generally
Critical weaknesses
  • The method has only modest accuracy gains on two of three benchmarks (roughtly 1-5%). These margins are small enough that a standard UNet (which is simpler and does not require a pretrained operator during training) might be preferable in practice; the three-cylinder flow example with large gains ( approx 30%) appears to be the setting where the baseline UNet performs unusually poorly.

  • The method is framed as knowledge distillation (Section 2), but no standard knowledge- or feature-distillation baseline (e.g., FitNets, CRD) is included. Without this, it is unclear whether the gains come from the specific frequency-domain formulation or simply from having any structured teacher signal. Similarly, it is not clear how other teacher networks would impact performance.

  • It is strange that the main table with comparisons to other approaches does not include the approach without teacher. It is included in the ablation studies later; but the benefit of the key innovation of the paper (regularization with teacher) should also be clearly shown in the main table, not just in an ablation.

Questions for authors
  • The paper fixes sigma = 0.5 without reporting sensitivity. Given that the central argument rests on selective low-frequency transfer, how much does performance degrade for, say,0.3 or 0.8, and is there a principled way to set sigma per dataset (e.g., based on the spectral energy profile of the target fields)?

  • Necessity of the dual-branch design vs. a single spectral branch with a residual spatial skip: In the ablation (Table 3), "Dual-branch w/o mean-shift" still outperforms the single-branch variant. What would happen if the spatial branch were replaced by a simple skip/identity residual rather than the explicit mean-shift? This would clarify whether the mean-shift is doing something specific.

  • The standard UNet's poor performance on three-cylinder flow: The authors write that the standard UNet and its pruned variant "perform poorly" on this benchmark and attributes this to factors beyond capacity. Could the authors elaborate on what these factors might be, since this benchmark seems to be critical to show accuracy improvement?

  • Would the method still hold if the teacher were a different architecture class (e.g., a graph neural operator or a spectral transformer)? Since the claimed benefit is injecting global spectral structure, demonstrating this with at least one alternative operator class would strengthen the generality claim.

1 comment

W1: The method has only modest accuracy gains on two of three benchmarks.

We thank the reviewer and agree that on AirfRANS and Turbulent, the accuracy gain over the standard UNet is modest. We respectfully note, however, that this comparison treats accuracy as the sole criterion, which is not the operative frame for onboard reconstruction. Whether a model can run within the power, memory, and latency budget is a first-order constraint. Φ-Sat-1 illustrates this directly: the on-board CNN was selected not as the most accurate cloud-detection model, but because it ran under a 2 W power budget on the Myriad 2 VPU with a ≈5 MB memory footprint suitable for uplink updates. Aircraft state-assessment work follows the same logic—recent UAV fault-diagnosis work (KD-LiteTCN) reports 97.6% parameter reduction and 2.4 ms latency as core contributions, with accuracy framed as acceptable under extreme compression. In these settings, acceptable accuracy under a fixed resource budget is itself the contribution.

Our method targets exactly this bottleneck: a compact encoder–decoder whose capacity-limited bottleneck cannot aggregate field-wide context. The relevant comparison is therefore not "SS-REPA vs. standard UNet in accuracy," but "SS-REPA vs. what is actually deployable." Table 2, the standard UNet (13.39 M) has 6–7× higher latency on the Orin Nano than SS-REPA (0.86 M). MobileNet-UNet (0.71 M) is the other lightweight option, but is substantially less accurate on all three benchmarks and is actually slower on AirfRANS. SS-REPA thus occupies a previously unavailable working point: ≈6% of the UNet's parameters and ≈15% of its latency, while matching or exceeding its accuracy on all three benchmarks.

References:

  1. The Φ-Sat-1 Mission: The First On-Board Deep Neural Network Demonstrator for Satellite Earth Observation[J]. IEEE Transactions on Geoscience and Remote Sensing,2022,Vol.60(1): 1-14.

  2. An Ultra-Lightweight Knowledge-Distilled TCN Framework with Multi-Objective Output Distillation for Real-Time UAV Fault Diagnosis[C]//2026 2nd International Conference on Unmanned Systems and Technology. 2026.

W2. no standard knowledge- or feature-distillation baseline is included.

We thank the reviewer for this observation. The revised Section 4.3 adds a direct comparison against conventional distillation objectives.

W3. the main table with comparisons to other approaches does not include the approach without teacher.

We thank the reviewer for this observation. The revised main comparison (Table 1) now includes SS-REPA w/o alignment.

Q1: The paper fixes sigma = 0.5 without reporting sensitivity.

We thank the reviewer for this suggestion. The sensitivity of SS-REPA to the bandwidth σ is reported in the revised Section 4 ("Effect of bandwidth σ", lines 504–508).

On the suggested principled per-dataset rule: we agree it is a natural extension. The observed trend suggests σ could be set from the spectral energy profile of the target fields, and we now note this as future work, clarifying that σ = 0.5 is an empirical default validated across all three benchmarks rather than a dataset-specific tuned value.

Q2: Necessity of the dual-branch design with a residual spatial skip.

We thank the reviewer for this question, which is already addressed by an ablation in the paper; we have made its definition more explicit in the revision (line 416-417).

Q3: The standard UNet's poor performance on three-cylinder flow.

We thank the reviewer for pushing on this point. We have made the explanation more concrete in the revision (lines 345–346).

The factor we point to is the sensor placement configuration: sensors are placed predominantly on the cylinder surfaces, leaving large unobserved regions in the wake. Reconstructing the wake therefore requires propagating information from surface measurements across a spatially extended unobserved domain. One plausible explanation is that the convolutional UNet bottleneck aggregates this information less effectively than a spectral representation, which is consistent with its error concentrating in the unobserved wake (Figure 3) and with the observation that increasing UNet capacity does not close the gap. We do not claim this is the only factor, but it aligns with the failure mode motivating our design, and the benchmark is informative because this setting is where the difference in global aggregation is most visible.

Q4: Would the method still hold if the teacher were a different architecture class.

We thank the reviewer for raising this point.

We agree that the framework should not depend on one particular teacher implementation. We therefore tested GNO, a graph-based spectral–spatial neural operator, as an additional teacher (Appendix A.1.3) . Results are reported in the main experiments (Table 1, "SS-REPA (GNO)"). SS-REPA with a GNO teacher is competitive across all benchmarks and sensor settings, and is best or second-best in several.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the Area Chair and Reviewers

We thank the Area Chair and all reviewers for their constructive feedback and for the careful evaluation of our work. We have revised the manuscript accordingly and resubmitted it; all changes are highlighted in blue.

Core contribution.

  • Target setting and problem: onboard sparse physical-field reconstruction under tight memory and latency budgets, where a capacity-constrained convolutional encoder–decoder bottleneck struggles to aggregate field-wide context and infer unobserved regions.
  • Method: a deployment-driven, branch-level selective alignment framework built on a decoupled dual-branch bottleneck. A compact convolutional encoder–decoder is retained as the deployed model; only the spectral-context branch is aligned during training to a frozen spectral neural operator teacher, while the spatial-detail branch remains free to preserve local detail.
  • Result: on an NVIDIA Orin Nano edge platform, the model reduces parameters by approximately 94% and inference latency by approximately 83.4–85.5% relative to the standard convolutional baseline, while maintaining competitive reconstruction fidelity on AirfRANS, three-cylinder flow, and turbulent channel flow.

Main revisions in response to the reviewers.

  • Added a near-parameter-matched Standard UNet (0.92M) comparison.
  • Added conventional feature-level and output-level distillation baselines under the same teacher and student.
  • Included the no-alignment variant in the main comparison table.
  • Reported bandwidth sensitivity for σ.
  • Added teacher-architecture generalization results (GNO, Standard UNet, and randomly initialized teachers).
  • Added sensor-noise robustness analysis.
  • Reported three-seed means and standard deviations.
  • Added comparisons with LNO, MMGN, SIREN, and WIRE.
  • Added low-/high-frequency reconstruction error decomposition.
  • Reported training overhead.
  • Clarified Parseval equivalence and the implementation details of the spectral-context branch (frequency-domain nonlinearity and real-output convention), and added instance-level comparison experiments, etc.

We sincerely thank the Area Chair and all reviewers for their valuable comments, which have helped us strengthen the evidence and clarity of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.