acceptodds
Withdrawn from ICLR 2027

DRESO: An Adaptive Complementary Spectral Neural Operator for Robust Autoregressive Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural operators learn mappings between function spaces and provide fast surrogates for partial differential equations (PDEs), but accurate one-step prediction does not guarantee stable autoregressive rollouts. DRIFT-Net combines a spectral branch for global low-frequency structure with an image branch for local details and nonstationary patterns. The two branches are fused through lightweight bandwise weighting (Li & Salim, 2026). However, DRIFT-Net’s hard spectral partition does not support direct gradient-based optimization of the cutoff, while its bandwise energy summary provides limited information about the spectral structure within each region. To address these limitations, we propose DRESO, a dual-region evidence-guided spectral operator, which consists of a learnable complementary filter (LCF) and an adaptive gain controller (AGC). First, LCF designs a smooth Gaussian low-pass filter with a learnable cutoff and its complementary high-pass filter. This design enables each spectral layer to adaptively learn its low–high frequency allocation during training. Then, based on the low- and high-frequency regions provided by LCF, AGC extracts ten spectral descriptors from each region, converts their normalized comparisons into sample–channel-specific gains, and uses these gains to recombine the independently transformed regional spectra before the inverse FFT. These descriptors provide richer regional spectral information for adaptive routing than raw features or low-dimensional spectral summaries. On the six CamLab-ETHZ tasks, DRESO-L achieves the lowest error on all six one-step problems. DRESO-B ranks second in each comparison and outperforms DRIFT-Net with 32.9% fewer parameters. DRESO-B also uses 44.0% fewer parameters than NESTOR-Small, the strongest competing baseline. Code is available at anonymous code repository.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, partial differential equations, spectral neural networks, frequency-domain modeling, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pzCyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DRESO, extending the DRIFT-Net backbone with a learnable complementary filter (LCF) and an adaptive gain controller (AGC). LCF forms low- and high-frequency regions using a smooth Gaussian response and its complement, with separate channel transformations. AGC uses ten regional spectral descriptors to generate two gains for each sample and channel, controlling their recombination. Evaluation covers one-step and autoregressive prediction on CamLab-ETHZ and The Well, alongside component ablations, gate comparisons, and longer rollouts.

Critical strengths
  1. The prediction results are appealing. Among the comparisons in the main tables, DRESO-L achieves the lowest one-step error on all six CamLab tasks and four The Well tasks, while the 11.48M-parameter DRESO-B improves over the 17.10M-parameter DRIFT-Net on all twelve CamLab one-step and rollout metrics, showing a useful accuracy–parameter trade-off.
Critical weaknesses
  1. I understand the motivation to reduce long-horizon error by better handling non-dominant spectral components, but I am not yet clear why neglecting these components is a key bottleneck on the tasks studied here. The introduction mainly relies on intuition and prior work, while the appendix’s low/high-frequency comparisons do not directly distinguish dominant from non-dominant components, leaving the connection between the method and the forecasting problem insufficiently explained.

  2. I am still unsure how much the ten descriptors and state-dependent gating each contribute to AGC’s gains. Table 5 supports the benefit of adding a gate, but Tables 6 and 7 only mask features at inference, so I would appreciate comparisons with input-independent learned gains and controllers retrained with fewer descriptors to better understand these design choices.

Questions for authors
  1. Are spectral diagnostics available for the baselines that connect errors in non-dominant components to subsequent rollout error growth? If so, could you point to these results and explain how they support the motivation in the introduction?

  2. Have input-independent learned regional gains or controllers retrained with fewer descriptors been evaluated? If not, I view these as suggestions for additional validation and would appreciate clarification of which specific AGC design choices the existing experiments establish.

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

The paper proposes DRESO, a neural operator for predicting how PDE systems evolve. It builds on DRIFT-Net and introduces a learnable filter that separates low and high frequencies while keeping the full spectrum. A controller uses ten spectral statistics to decide how strongly each region contributes to the prediction. The method is tested on CamLab-ETHZ and The Well, with additional studies on longer rollouts, transfer, and efficiency. The main results are strong, although the advantage is less consistent on longer rollouts.

Critical strengths
  1. The improvements are substantial across several tasks. DRESO-L wins all six CamLab-ETHZ one-step comparisons and five of six rollout comparisons in the main setting. It also wins the four main one-step tasks from The Well. DRESO-B beats DRIFT-Net on all twelve CamLab-ETHZ metrics with about 33% fewer parameters. This makes the results useful beyond a small gain on one benchmark.

  2. The paper makes a good effort to check which parts of the design help. The component study in Table 5, fixed-filter comparison in Table 8, and shared-mixer study in Table 16 support the proposed choices. Table 17 also shows that much of the advantage remains without the extra high-frequency loss. These checks make the main results more convincing.

Critical weaknesses
  1. I am not yet convinced that all ten controller inputs are needed. Removing inputs at test time shows that the trained model relies on them, but a simpler controller might learn to work just as well. It would help to show how much the gates actually change with the input (see also Weakness 5).

  2. The rollout advantage has clear limits and comes at a large speed cost. DRESO-L wins only two of four rollout tasks from The Well. On the 199-step Shear Flow test, its error is 204.3 versus 94.5 for NESTOR-Small, and it is also worse than DRIFT-Net at 194.3 (Table 12). Its latency is about 158 ms, compared with 39 ms for DRIFT-Net and 15 ms for NESTOR-Small (Table 13). These limits deserve more attention when discussing robustness and practical efficiency.

  3. The headline results depend on scaled-down baselines. The main tables use smaller variants of NESTOR and PDE-ReFiner, and Table 18 shows that the full-sized versions do better. Full-sized NESTOR has lower rollout error than DRESO-L on all six CamLab-ETHZ tasks. It also has lower one-step error on NS-Gauss, CE-RP, CE-KH, and NS-Sines, and the full PDE-ReFiner is lower on CE-CRP (14.747 versus 17.498). DRESO-L stays best only on CE-Gauss one-step. Full-sized NESTOR is larger (80.8M parameters) but still much faster than DRESO-L (13 ms versus 158 ms, Table 13). The claim of the lowest error on all six one-step tasks should therefore be stated as relative to the scaled-down baselines.

  4. The spectral layer is simpler than its framing suggests. The mixers in Equation 7 are shared across frequencies and MH=1−MLM_H = 1 - M_L, so Equations 7 and 12 reduce to

y=F−1[(gHWH+(gLWL−gHWH) ML)X].y = \mathcal{F}^{-1}\big[\big(g_H W_H + (g_L W_L - g_H W_H)\, M_L\big) X\big].

For the real parts of the mixers, this is a pointwise channel mixer plus a channel mixer applied after a Gaussian blur. If ρ=1\rho = 1 at the Nyquist frequency, the learned cutoffs of about 0.24–0.29 (Appendix C.4) give a blur with a spatial standard deviation of only about one grid cell. At the finer levels of the encoder–decoder, the layer is therefore close to a local operator rather than a global spectral kernel. The imaginary parts of the complex mixers act differently: applied to the half-spectrum of a real FFT, they produce a Hilbert transform along one grid axis. That term is non-local but direction-dependent, and it comes from the FFT convention rather than from the radial design. The paper should state this reduced form, explain the role of the imaginary parts, and compare with the same layer built from a spatial Gaussian convolution. This would also show whether the FFT is worth its cost, given the 158 ms latency.

  1. The normalization in Equation 9 limits what the controller and the masking study can show. With only two regions, it gives

zL=−zH=dd2+10−5,z_L = -z_H = \frac{d}{\sqrt{d^2 + 10^{-5}}},

where dd is half the low–high difference. Unless the two values are almost equal, each input is close to ±1\pm 1, so the controller mainly sees the sign of each comparison, and almost no gradient passes back through it. This has three consequences. First, Appendix C.2 calls zero the training-distribution reference value. It is the average of each low–high pair, but it almost never occurs as an input value, so the masking in Tables 6 and 7 is an out-of-distribution intervention. Second, some comparisons are largely fixed by the masks themselves. Because MH/MLM_H / M_L increases with radius, the high region should have the higher spectral centroid in essentially every sample (exactly so if the small ϵ\epsilon in Appendix C.1 is ignored). That input is then nearly constant, and masking it mostly removes a learned bias rather than input-dependent information. Third, if both soft regions are defined over the whole grid, log⁡P\log P and qq order the two regions in the same way and therefore give the same normalized value. The large masking effects in Section 5.5 may therefore not reflect adaptive use of the descriptors. Reporting how often each sign flips across samples and time steps would resolve this weakness of mine.

Questions for authors
  1. How much variation across inputs and time survives the normalization in Equation 9? Statistics for a fixed block and channel would help establish how adaptive the gates are.

  2. Could you retrain a simpler controller using only energy statistics, on a representative subset of tasks, and report a few independent runs? This would help establish the benefit of the additional inputs.

  3. How were the different learning rates for DRESO and DRIFT-Net/Poseidon-T selected? What exactly is the Table 5 baseline, and were the other training settings held fixed across its rows?

  4. Were the full-sized models in Table 18 trained with the same data budget and input history as the main comparisons, or were original pretrained checkpoints used? If they were trained under the same protocol, how do you position DRESO against full-sized NESTOR, which is both more accurate and faster? A comparison at matched parameter count or matched latency would help here.

  5. How is the normalized radius ρ\rho defined, and are the mixers WL,WHW_L, W_H used with nonzero imaginary parts after training? How does the spectral layer compare with the same design implemented as a spatial Gaussian convolution, in both accuracy and latency?

  6. For each AGC descriptor, how often does the sign of the low–high comparison change across samples and rollout steps? Are both regions defined over the full Fourier grid when computing PP and qq?

I would raise my score if the authors reframe the claims and show where DRESO is preferable to larger, faster baselines.

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

This paper proposes DRESO, a neural operator for PDE prediction that combines a learnable complementary spectral filter (LCF) and an adaptive gain controller (AGC). LCF employs a smooth Gaussian low-pass filter that enables direct gradient-based optimization of the cutoff frequency, addressing a limitation of the existing DRIFT-Net baseline. AGC uses regional spectral descriptors to adjust the contribution of different frequency bands. The method is evaluated on six CamLab-ETHZ and four The Well tasks in both one-step prediction and autoregressive rollout settings, with extensive ablations. DRESO demonstrates competitive one-step prediction accuracy and promising autoregressive rollout performance.

Critical strengths
  1. DRESO-L achieves the best one-step RelL1 error among the evaluated baselines on all ten main tasks. The paper also provides broad experimental coverage, including multiple evaluation settings (e.g., Pretrain, Finetune-50 and Finetune-100), one-step and several rollout modes, and cross-resolution evaluations. These results show that DRESO is a promising approach to improving spectral neural operators, even if its advantages do not consistently extend to long-horizon rollouts.
  2. The proposed LCF and AGC mechanisms provide an intuitive way to separate and recombine low- and high-frequency information. The paper investigates several architectural choices through ablations, including filter types (ideal vs Gaussian filters), mixer sharing, and alternative gating mechanisms (FreqMoE and TF-SNO gates). This provides useful initial evidence about the method design, although stronger controlled experiments are needed to establish which components are responsible for the observed gains.
Critical weaknesses
  1. Limited novelty and insufficient evidence supporting the proposed components. While intuitive, the core ideas behind DRESO (learnable spectral filtering and adaptive gating) are closely related to established techniques. At the same time, the use of a Gaussian filter, rather than a learnable sigmoid or another smooth filter family, lacks a principled justification, and alternative smooth filter shapes are not evaluated. Similarly, while the authors provide intuitive explanations for the individual descriptors, it remains unclear why exactly 10 descriptors are necessary, rather than a smaller or larger set. Furthermore, the masking study (Tables 6 and 7) sets descriptors to 0, a value the gate may rarely see in training, so the large error increases may reflect distribution shift rather than feature importance. Thus, ablations with retraining DRESO without selected features are missing to establish their individual contributions.

  2. The baseline comparisons are not properly controlled to support the central performance claims. a. The main results focus on comparisons against reduced-size baselines, while the full-size NESTOR reported in the Table 18 outperforms DRESO-L in most of the one-step and rollout comparisons. A parameter-matched comparison is a legitimate framing, but it should be stated more explicitly rather than implied as a general win in the main claims. b. Also, long-horizon and The Well rollouts show mixed results (Tables 11 and 12), which weakens the “robust autoregressive prediction” claim. c. The Table 5 baseline is not properly defined in the text, yet it already beats DRIFT-Net by a wide margin (e.g., 3.118 vs 7.140 on NS-Gauss), which is more than the improvements provided by LCF and AGC in the same ablation. This suggests the improvement may result from the training procedure or implementation rather than the proposed modules. This needs clarification. d. Finally, the paper states that random seeds are fixed for each run, but it does not report multi-seed means or uncertainty estimates for the main comparisons, even though some margins (e.g., Tables 4 and 5) are quite small. Thus, without properly matched training protocols and multiple independent runs reported, it is difficult to attribute the gains reliably to DRESO and assess the robustness of these results.

Questions for authors
  1. Can the authors report means and standard deviations over multiple independent seeds for the main results, especially the ones used to support the central claims?
  2. What exact architecture and training configuration does the Table 5 baseline use? If it corresponds to DRIFT-Net, why does the baseline demonstrate better performance? Are the training protocol, loss, and evaluation settings identical? (see weakness 2)
  3. In Equation 9, each descriptor is standardized over only two values (the low- and high-frequency regions). Does it imply that for every sample b, channel c and descriptor j: z_L = -z_H, with |z|=1 whenever the regional difference is large relative to the stabilization constant, so that the normalized descriptors are effectively only +1 and -1? If so, doesn’t this normalization remove potentially useful information about the magnitude of the regional difference? And could setting a descriptor to 0 for masking in Table 6 consequently cause a distribution shift?
  4. Could the authors evaluate the contribution of descriptor features by retraining DRESO with each descriptor removed individually?

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

Related papers

Discussion (0)

Sign in to comment.