acceptodds
Under review as a conference paper at ICLR 2027

SSF-Net: A Spatial–Spectral Fusion Network for Accurate PDE Operator Learning

13 views

Abstract

Machine learning methods for solving partial differential equations (PDEs) significantly improve computational efficiency over traditional numerical solvers such as finite element method (FEM). However, PDE dynamics inherently span both local and global dynamics: disturbances propagate and decay gradually through the domain, while global coupling effects can influence the whole system simultaneously. Capturing both types of dynamics with a low complexity model remains a key challenge. To address this, we propose the Spatial-Spectral Fusion Network (SSF-Net), built around a dual-branch Spatial–Spectral Fusion (SSF) block comprising a spatial branch and a spectral branch. The spatial branch employs a multi-directional Mamba, which mitigates its inherent causal, single-direction scanning bias and capture local interactions with reduced directional bias at linear complexity. The spectral branch leverages a Fourier transform with a blockdiagonal MLP to efficiently capture global spectral features with reduced parameter count and complexity. The SSF block is embedded in a multiresolution convolutional encoder-decoder, which extracts multi-scale features. We evaluate SSFNet on a comprehensive PDE benchmark spanning multiple task types and assess its long-horizon stability. Compared with attention-based and convolution-based baselines, SSF-Net achieves lower prediction error with fewer parameters while retaining near-linear asymptotic complexity in the number of spatial tokens.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
PDE, mamba, spectral, neural operator

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_KU6nOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents SSF-Net, a neural operator architecture for learning PDE solution operators. It uses a U-Net-style encoder-decoder with two branches: (1) a spatial branch that uses four-directional Mamba scans, and (2) a spectral branch that uses Fourier transforms and block-diagonal MLPs. The input features are split between the two branches and fused at the end. The authors evaluate their architecture on 12 PDE benchmark datasets from PDE GYM and APEBench, in the settings of one-step prediction and autoregressive rollout. They show that their method achieves better prediction accuracy than the considered baselines.

Critical strengths

The evaluation covers 12 PDE benchmarks.

Critical weaknesses
  1. Major reproducibility concerns: The baseline results reported in Table 1 differ substantially from those published in the original DriftNet paper (arXiv:2509.24868), even though the authors of SSF-Net use the same benchmark datasets and closely related training and evaluation protocols. For example, on NS-PwC, SSF-Net reports a final-step L1 error of 0.68 for FNO and 4.74 for DriftNet, whereas the original DriftNet paper reports 4.57 and 1.09, respectively. Likewise, on NS-Tracer-PwC, SSF-Net reports 1.85 for FNO and 8.78 for DriftNet, compared to 9.46 and 4.19 in the original paper. These discrepancies are large and change the relative performance of the baselines. The SSF-Net paper provides no explanation for them. Although Appendix E describes the general training setup, much important information is missing. Furthermore, no code or reproducibility instructions are provided, and there is no reproducibility statement. Given the magnitude of the reported improvements and the discrepancies with previously published results, I consider this a major weakness.
  2. Missing evidence for fair baseline comparisons: Although Appendix E.3 lists the training configurations, it is unclear how the hyperparameters and architectures of the different baselines were selected, i.e., according to which criteria. The authors use different learning rates, training objectives, and learning-rate schedules across models, but provide neither a tuning study nor an explanation of how they ensured comparable optimization quality. For example, SSF-Net is trained with a learning rate of 5e-4, whereas DriftNet uses 3e-4 and additionally uses its frequency-weighted loss. The models also differ substantially in architecture size and configuration. Moreover, using the same number of training epochs does not by itself establish a fair comparison, since many other moving parts, such as patch sizes, also play a role.
  3. Having worked with Wave-Gauss myself, I see a real problem with how the authors handle it: Table 3 in Appendix E.1 states that the model input for Wave-Gauss is (ut,cω)(u^t, c_\omega), with ut+1u^{t+1} as the prediction target. However, the wave equation is second order in time, so in general the state u(t)u(t) alone does not determine the future; one also needs ∂tu(t)\partial_t u(t), or a proxy thereof, as input. Otherwise, the operator that the network is supposed to learn is not well-defined on arbitrary wave-equation states.
Questions for authors
  1. Could the authors explain the large differences between the baseline results in Table 1 and those reported in Table 1 of the original DriftNet paper?
  2. How were the hyperparameters selected when tuning the baselines? Which hyperparameters were varied, over which ranges, and according to which criteria? Were they tuned on validation sets? How did the authors ensure that the baselines were adequately optimized, given their different architectures, training objectives, and learning-rate schedules? 3 ) Could the authors provide a comparison with DriftNet in which DriftNet’s spatial branch is replaced by the proposed Mamba branch, while keeping everything else fixed (remaining architecture, parameter budget, training protocol, etc.)?
  3. Regarding the Wave-Gauss input described under Weaknesses: does the implementation include the temporal derivative (as in the original Poseidon formulation) as a model input, or is the model trained only on the displacement field? If the latter, what restrictions on the dataset make this one-step prediction problem a well-defined forecasting problem?
  4. Could the authors provide a qualitative evaluation of the predicted PDE dynamics? I would be happy to see visualizations of the predictions and the corresponding error maps at different forecast horizons.
4 comments
AuthorsOpenReview ↗

The answer for Q1

Could the authors explain the large differences between the baseline results in Table 1 and those reported in Table 1 of the original DriftNet paper?

The model configurations of FNO, CNO, scOT follows the setting in Poseidon, and we use the same setting of Drift-Net and PDE-Transformer as their original paper.

Our FNO baseline was retrained using a different configuration from that reported in Drift-Net, including 20 rather than 16 Fourier modes and an initial learning rate of 5×10−45\times10^{-4} rather than 3×10−43\times10^{-4}. This setting is from the Poseidon.

We acknowledge that errors were introduced when porting the scOT and Drift-Net implementations into our training framework. We have corrected these errors and rerun the affected experiments. The revised results supersede the corresponding values in the original submission. With the corrected baselines, SSF-Net retains the lowest errors across all eight completed benchmarks. Relative to the stronger of scOT and Drift-Net on each task, it reduces the step-averaged, full-rollout, and final-step errors by 59.2%, 59.2%, and 58.2%, respectively, averaged equally over the seven time-dependent tasks. On Poisson-Gauss, the error reduction is 61.9%. These improvements are achieved with 11M parameters, compared with 20M for scOT and 17M for Drift-Net. Experiments on Kolm-Smooth, Kolm-Turb, KS-Strong, and KS-Weak are still running. We will report the updated results once these runs are complete.

We will explain why the results are still differenct from the Drift-Net paper.

DatasetscOTDrift-NetSSF-Net
Params20M17M11M
Poisson-Gauss0.430.420.16
NS-PwC0.34/0.34/0.600.50/0.50/0.850.10/0.10/0.17
NS-Tracer-PwC0.48/0.48/0.820.86/0.85/1.850.18/0.18/0.33
NS-SL0.66/0.67/1.250.86/0.88/1.670.23/0.24/0.48
FNS-KF0.31/0.31/0.520.57/0.57/1.030.10/0.10/0.16
CE-RP5.06/4.99/9.977.65/7.55/15.202.75/2.72/5.71
Wave-Gauss6.30/6.15/9.765.59/5.46/8.524.43/4.32/7.10
ACE0.11/0.12/0.210.12/0.13/0.250.02/0.02/0.03

There is a difference in how training examples are organized. The released Drift-Net dataset interface indexes individual trajectory–time pairs, allowing pairs from different temporal positions to be mixed during training. In our implementation, each batch contains complete trajectories. We then process their adjacent transitions sequentially, performing a separate optimizer update for each time index. Thus, samples within each update share the same temporal position. This introduces temporal homogeneity within each optimization batch, which may contribute to the substantially lower scOT and Drift-Net errors observed in our reruns compared with the original paper, as well as the stronger performance of scOT relative to Drift-Net on some tasks.

This training scheme allows us to systematically use all adjacent transitions from each sampled trajectory batch, with a fixed prediction interval matching the one-step rollout task. Processing the transitions in temporal order also ensures that each optimizer update contains examples at the same temporal position, rather than mixing states from different stages of evolution. This provides a consistent batching protocol across the compared architectures, although we do not claim that it is universally preferable to shuffled trajectory–time pairs.

Consequently, our experiments are retraining comparisons under our stated protocol, rather than direct reproductions of the published Drift-Net results.

We have updated the affected experimental results and the corresponding analyses in the revised manuscript. We have also clarified the architecture and configuration of each baseline in Appendix E.4 (Model Configurations).

AuthorsOpenReview ↗

Answers for Q2-Q4

How were the hyperparameters selected when tuning the baselines? Which hyperparameters were varied, over which ranges, and according to which criteria? Were they tuned on validation sets? How did the authors ensure that the baselines were adequately optimized, given their different architectures, training objectives, and learning-rate schedules? 3 ) Could the authors provide a comparison with DriftNet in which DriftNet’s spatial branch is replaced by the proposed Mamba branch, while keeping everything else fixed (remaining architecture, parameter budget, training protocol, etc.)?

Our baseline model hyperparameters were primarily adopted from published configurations rather than selected through a new systematic hyperparameter search. The model configurations of FNO, CNO, scOT follows the setting in Poseidon, and we use the same setting of Drift-Net and PDE-Transformer as their original paper.

We adapted these reference settings to our experimental training budgets and retained architecture-specific choices, including CNO’s batch size and learning-rate schedule and Drift-Net’s frequency-weighted training objective.

ACEFNS-KF
DriftNet with Mamba branch (33M)0.17/0.19/0.340.49/0.49/0.92
original DriftNet (17M)0.12/0.13/0.250.57/0.57/1.03

We replaced Drift-Net’s depthwise convolutional spatial mixer with the four-directional Mamba operator, retaining its spectral processing, radial gating, pointwise channel mixing, remaining backbone, and training protocol. To isolate the replacement within a fixed backbone, we preserved the stage widths and depths. This increases the parameter count from 17M to 33M.

On FNS-KF, the replacement reduces step-averaged/full-rollout/final-step errors from 0.57/0.57/1.03% to 0.49/0.49/0.92%, corresponding to reductions of approximately 14.0/14.0/10.7%. On ACE, however, the errors increase from 0.12/0.13/0.25% to 0.17/0.19/0.34%, despite the larger model.

The FNS-KF improvement is consistent with our motivation for combining scan-mediated interactions with spectral processing, whereas the ACE result shows that Mamba is not uniformly preferable to convolution.

Regarding the Wave-Gauss input described under Weaknesses: does the implementation include the temporal derivative (as in the original Poseidon formulation) as a model input, or is the model trained only on the displacement field? If the latter, what restrictions on the dataset make this one-step prediction problem a well-defined forecasting problem?

Our implementation uses the displacement field and the static wave-speed field, (ut,c)(u_t,c), without an explicit temporal-derivative channel. We checked the released Wave-Gauss dataloaders in the original code of both Poseidon and Drift-Net: both likewise construct two field channels, [u,c][u,c], and neither loads nor computes v=∂tuv=\partial_tu. Thus, the augmented state [u,v,c][u,v,c] in the Poseidon paper describes the first-order mathematical formulation, rather than the field-channel configuration of its released dataloader. Moreover, The publicly released Wave-Gauss in PDEGym dataset itself contains only the displacement trajectories (u) and the static wave-speed field (c). It does not provide v=∂tuv=\partial_tu as a separate variable.

Nevertheless, we agree that consistency with these implementations does not establish that displacement alone defines a closed one-step forecasting state. Our current model learns (ut,c)↦ut+1(u_t,c)\mapsto u_{t+1} from consecutive snapshots, and we have not established a dataset restriction that guarantees the missing temporal derivative is uniquely determined by these inputs.

Could the authors provide a qualitative evaluation of the predicted PDE dynamics? I would be happy to see visualizations of the predictions and the corresponding error maps at different forecast horizons.

We give FNS-KF qualitative visualization here, which shows GT, SSF-Net and scOT at 5 time steps in appendix I.

Dear Reviewer,

Thank you for your careful examination of our experiments. As the discussion period ends on November 18, we would like to follow up on our response and revised manuscript , in which all changes are marked in blue:

  • Baseline discrepancies: We corrected implementation errors affecting scOT and Drift-Net, reran the eight completed benchmarks, and updated the results and analyses. SSF-Net retains the lowest errors on these benchmarks. Appendix E.4 details the model configurations, and our response explains the training-protocol differences from the released implementations of Drift-Net, which is alse revised in Appendix E.3.
  • Hyperparameters and branch replacement: We clarified that hyperparameters were primarily adopted from published configurations rather than selected through a new systematic search in Appendix E.4. We also reported the requested Mamba replacement within Drift-Net: it improves FNS-KF but worsens ACE, while increasing parameters from 17M to 33M. We therefore do not present it as a parameter-matched comparison or evidence of a universal Mamba advantage.
  • Wave-Gauss and qualitative evaluation: We confirmed that our inputs are displacement and wave speed, without an explicit temporal-derivative channel.
  • Qualitaitve evaluation: Appendix I provides predicted fields and error maps at multiple forecast horizons on FNS-KF.

We hope these corrections and additional experiments help clarify the evidence supporting our conclusions. We would appreciate your feedback and, if your main concerns have been addressed, your consideration of an updated assessment.

Sincerely,
The Authors

Thank you for the additional experiments and clarifications. I appreciate the authors' efforts, but my main concerns remain unresolved. The baselines were not systematically retuned for the modified training procedure. With a different training procedure, it may be that the baselines need different HPs to perform well. I am also concerned about how the improvements are presented. For example, the reported 61.9% improvement on Poisson-Gauss is relative to Drift-Net, whereas the improvement over the strongest baseline, PDE-Transformer, is only 0.16 VS. 0.18 (see Table 1) is not reported in your text. I think, comparisons should be always against the strongest baseline.

I therefore maintain my score of 1.

Reviewer_FSpwOpenReview ↗
Rating
3
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SSF-Net, a spatial-spectral fusion network for PDE operator learning, which efficiently captures local interactions and global dependencies. It employs multi-directional Mamba scanning to model spatial interactions while reducing directional bias. In addition, a spectral branch incorporates Fourier transforms and block-diagonal MLPs to efficiently capture global dependencies. Experimental results on multiple 2D PDE benchmarks demonstrate that SSF-Net outperforms five neural operator baselines across all tasks while using fewer parameters. The authors also provide additional experimental results to discuss the experimental settings and the contributions of individual modules.

Critical strengths

1.SSF-Net achieves the lowest prediction errors among the compared methods on all 12 PDE benchmarks while using substantially fewer parameters (11M). The improvements span multiple PDE families, and the model maintains competitive stability over 100-step autoregressive rollouts. 2.The paper provides theoretical characterizations and Jacobian-based effective-kernel visualizations to examine the interaction patterns of its spatial and spectral branches. The observed local attenuation and global sensitivity patterns are consistent with the intended architectural design.

Critical weaknesses

1.The proposed SSF-Net primarily integrates existing techniques, including multi-directional Mamba, AFNO-based spectral processing, and a multiresolution encoder–decoder architecture. In particular, prior methods such as Drift-Net have already explored spatial–spectral fusion for PDE operator learning. The main architectural distinction appears to be the introduction of multi-directional Mamba for spatial feature extraction. However, the paper lacks controlled comparisons that isolate the benefits of this modification under otherwise identical architectural configurations. Consequently, it remains unclear whether the reported improvements arise from the proposed design or other architectural differences.

2.The ablation study is conducted on only two of the twelve PDE benchmarks (ACE and FNS-KF), without clearly explaining the criteria for selecting these tasks. Moreover, the spatial-only variant outperforms the full model on ACE and achieves the same step-averaged error on FNS-KF, although the variants differ in parameter count. While the authors provide plausible explanations based on PDE dynamics and model capacity, the results do not consistently demonstrate the benefits of combining both branches. Extending the ablation study to additional PDE families would help establish whether the proposed fusion provides systematic improvements.

Questions for authors
  1. What are the specific advantages of multi-directional Mamba over existing spatial modeling approaches in spatial–spectral fusion architectures? Given that prior methods such as Drift-Net have already explored spatial–spectral fusion for PDE operator learning, could the authors clarify the specific contribution of introducing multi-directional Mamba? In particular, can the authors provide controlled comparisons replacing the Mamba branch with a convolutional alternative while keeping the remaining architecture and computational budget comparable?
  2. Why were only ACE and FNS-KF selected for the ablation study, and how generalizable are the observed effects of spatial–spectral fusion? The current ablation results show that the spatial-only variant outperforms the full model on ACE and matches its step-averaged error on FNS-KF, although the variants have different parameter counts. Could the authors clarify the selection criteria for these two benchmarks and provide evidence from additional PDE families?
2 comments
AuthorsOpenReview ↗

Answers for all questions

What are the specific advantages of multi-directional Mamba over existing spatial modeling approaches in spatial–spectral fusion architectures? Given that prior methods such as Drift-Net have already explored spatial–spectral fusion for PDE operator learning, could the authors clarify the specific contribution of introducing multi-directional Mamba? In particular, can the authors provide controlled comparisons replacing the Mamba branch with a convolutional alternative while keeping the remaining architecture and computational budget comparable? Such evidence would help determine whether the reported improvements can be attributed to the proposed spatial modeling mechanism.

Our motivation emphasizes a different aspect of the interaction mechanism: Rather than primarily emphasizing the complementary modeling of low- and high-frequency features, our approach focuses on spatial interaction and information propagation. The Fig 1 and Fig 4 show our model can extract both directional response that rapidly weakens away from the query and interaction having broad spatial coverage.

In the Mamba pathway, information propagates through recurrent state updates, introducing an attenuation bias along scan paths. Multi-directional scanning mitigates the causal ordering and directional bias of a single scan. Our effective-kernel analysis provides empirical evidence of localized influence in the evaluated models. In contrast, the effective-kernel of CNO in Fig 1 shows plain convolution cannot model this attenuation.

To isolate the contribution of the spatial modeling mechanism, we replace the Mamba pathway with a ConvNeXt-style convolutional alternative while retaining the spectral pathway, fusion operation, embedding, and multiresolution backbone. The ablation study shows the improvements can be attributed to our multi-Mamba.

We have added these additional experiments and the corresponding discussion to Appendix F.2.

ACEFNS-KF
replace spatial branch with ConvNeXt (11M)0.022/0.024/0.0350.108/0.108/0.175
Full SSF (11M)0.020/0.021/0.0330.098/0.098/0.157

Why were only ACE and FNS-KF selected for the ablation study, and how generalizable are the observed effects of spatial–spectral fusion? The current ablation results show that the spatial-only variant outperforms the full model on ACE and matches its step-averaged error on FNS-KF, although the variants have different parameter counts. Could the authors clarify the selection criteria for these two benchmarks and provide evidence from additional PDE families? This would help assess whether spatial–spectral fusion offers consistent benefits across different physical dynamics and whether the observed differences arise from branch composition or model capacity.

The two tasks provide contrasting settings: ACE involves local reaction–diffusion dynamics, whereas FNS-KF combines advective transport with nonlocal pressure coupling. This contrast makes them informative case studies for examining the two branches.

Regarding parameter counts, our original ablations keep the backbone configuration fixed, including the number of resolution stages, depth per stage, stage-wise channel dimensions, patch size, and skip connections. The retained branch in each single-branch variant is widened to preserve the block width. Because the spatial and spectral branches have different parameterizations, this produces different parameter counts despite the shared backbone configuration. Thus, these experiments assess branch composition under a fixed backbone configuration, rather than a matched parameter budget.

Matching parameters would require adjusting widths or other capacity settings, providing a complementary comparison to our fixed-backbone ablation. We adjust the embedding dimension to implement larger spectral branch only model and smaller spatial branch only model.

VariantParamsACEFNS-KF
Spectral branch only10M0.022/0.024/0.0340.122/0.122/0.206
Spatial branch only8M0.022/0.023/0.0320.109/0.109/0.172
Spectral branch only (fixed backbone)9M0.021/0.023/0.0330.115/0.115/0.188
Spatial branch only (fixed backbone)16M0.012/0.014/0.0260.098/0.098/0.168
full model11M0.020/0.021/0.0330.098/0.098/0.157

We have added these additional experiments and the corresponding discussion to Appendix F.1, and clarified their implications in Section 4.3.

The ablation experiments on other datasets are still running. We will report the results once the additional runs are complete.

Dear Reviewer,

Thank you again for your helpful comments. As the discussion period ends on November 18, we would like to follow up on our response and revised manuscript , in which all changes are marked in blue.

To examine the contribution of the spatial modeling mechanism, we added a ConvNeXt replacement experiment in Appendix F.2. This comparison retains the spectral pathway, fusion operation, embedding, multiresolution backbone, and training protocol, with both models containing approximately 11M parameters. The Mamba-based variant achieves lower errors under all three metrics on ACE and FNS-KF.

We also added adjusted-width single-branch comparisons in Appendix F.1 and clarified their implications in Section 4.3. Our response explains why ACE and FNS-KF were selected as contrasting physical settings and distinguishes task-dependent benefits from a universal advantage of fusion. Ablations on additional PDE families remain in progress.

Sincerely,
The Authors

Reviewer_vAf3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an architecture for neural operators built on a U-Net style multiresolution encoder–decoder, wherein each block contains both a spatial (meant to capture local information) and fourier (meant to capture global information) branch. The proposal is evaluated on 12 2D benchmarks from PDE GYM and APEBench. This also includes ablations, 100-step autoregressive rollouts, and effective kernel analysis to illustrate the complementarity of the two branches.

Critical strengths

The general motivation for the paper is well justified, and the results appear encouraging; the effective kernel analysis in Figures 1, 4, and 5 and Appendix C offer useful evidence that the two branches have complementary responses and provides a better motivation than simply relying on accuracy.

The proposed architecture is shown to be best in every row on 12 examples, often by 50–80%, at the smallest parameter count (11M).

Critical weaknesses
  • The source of the gains is not established; compute is confounded with architecture; limitations are ommited.

Although the full model performs well against external baselines, the ablation results in Table 2 provide mixed evidence for the central architectural contribution. On ACE, the spatial-only variant achieves lower prediction errors than the full model (0.012 versus 0.020). On FNS-KF, the spatial-only variant matches the full model's step-averaged error (0.098), while the eight-directional variant performs slightly better (0.096). What is more, these cover only 2/12 benchmarks, and the single-branch variants have substantially different parameter counts.

The efficiency is reported only as parameter count, while the limitations section clearly says practical training and inference costs are higher than the baselines. Yet no FLOPs, throughput or memory figures are given. Near-linear complexity is shared by FNO, AFNO and Swin-type models, so it does not distinguish the method.

  • Unclear baseline credibility; prior work.

The Drift-net is used here as a baseline, but it already proposed a dual-branch design: a spectral branch for global low-frequency structure and an image branch for local detail. It fuses the branches additively without widening channels, and pitches the unit as a drop-in replacement for windowed self-attention blocks. The proposed SSF-Net differs mainly in three ways: a VMamba-style scan instead of a ConvNeXt local branch; AFNO instead of masked low-frequency mixing with radial gating; finer, overlapping patches. Stronger controlled comparisons are therefore necessary to establish the specific contribution of the proposed fusion and Mamba-based spatial branch to understand the actual gap.

What is worse, is that the numbers don’t reconcile with DRIFT-Net’s published evaluation, despite an apparently identical protocol (Drift-net paper has a comparison against scOT with Drift-Net better on all shared PDE GYM tasks, while here its the opposite). Similarly, scOT on 100-step Kolmogorov and FNO on Poisson-Gauss have completely different numbers.

There appear to be no SSM-based baselines. The SSM branch is the main architectural novelty, yet none is compared. MambaNO already combines cross-scan Mamba integration with local convolution integration in a U-shaped operator.

Questions for authors
  1. Why patch size 2 for SSF-Net but 4 for the patch-based baselines
  2. Which implementation and configuration did you use for each baseline -- there appear to be discrepancies. Provide details for implementations.
  3. Please report mean \pm std over at least 3 seeds for Table 2, and ideally for the top two methods in Table 1.
  4. The spectral-only variant already beats the ablated baselines, so which component explains the gap to scOT? Two ablations would answer this directly: overlapping -> non-overlapping patches, and conv down/up-sampling -> patch merging.
  5. Please compare against an SSM-based neural operator (MambaNO or LaMO) under the same budget.
  6. MambaNO already incorporates multi-directional Mamba scanning and convolutional processing within a U-shaped neural operator architecture. Given the similarity to the proposed spatial branch, can the authors clarify the architectural differences?
  7. Can the authors report comparable training time, inference latency or throughput, FLOPs, and peak GPU memory usage for SSF-Net and the strongest baselines?
  8. Can you ablate the patch sizes? and the channel size splits?
5 comments
AuthorsOpenReview ↗

Answer for Q1-Q3

Why patch size 2 for SSF-Net but 4 for the patch-based baselines

We used patch size 2 for SSF-Net to retain a finer spatial representation before multiresolution processing. The patch-based baselines used patch size 4 in our evaluated configurations. We agree that this difference affects both prediction accuracy and computational cost.

To quantify its effect, we added an SSF-Net ablation with patch size 4, keeping the remaining architectural and training settings fixed. Changing the patch size from 2 to 4 increases the error, supporting the benefit of finer tokenization in the evaluated tasks. For a 128×128128\times128 input, patch size 2 produces four times as many initial tokens as patch size 4. Therefore, we do not interpret the lower parameter count as a matched computational budget.

Importantly, SSF-Net with patch size 4 still outperforms baselines on the ablation tasks, indicating that the observed advantage cannot be explained solely by the finer token grid.

We have added these additional experiments and the corresponding discussion to Section 4.3.

VariantParamsACEFNS-KF
patch=412M0.047/0.052/0.0930.158/0.158/0.238
patch=211M0.020/0.021/0.0330.098/0.098/0.157

Which implementation and configuration did you use for each baseline -- there appear to be discrepancies. Provide details for implementations.

DatasetscOTDrift-NetSSF-Net
Params20M17M11M
Poisson-Gauss0.430.420.16
NS-PwC0.34/0.34/0.600.50/0.50/0.850.10/0.10/0.17
NS-Tracer-PwC0.48/0.48/0.820.86/0.85/1.850.18/0.18/0.33
NS-SL0.66/0.67/1.250.86/0.88/1.670.23/0.24/0.48
FNS-KF0.31/0.31/0.520.57/0.57/1.030.10/0.10/0.16
CE-RP5.06/4.99/9.977.65/7.55/15.202.75/2.72/5.71
Wave-Gauss6.30/6.15/9.765.59/5.46/8.524.43/4.32/7.10
ACE0.11/0.12/0.210.12/0.13/0.250.02/0.02/0.03

We acknowledge that errors were introduced when porting the scOT and Drift-Net implementations into our training framework. We have corrected these errors and rerun the affected experiments. The revised results supersede the corresponding values in the original submission. With the corrected baselines, SSF-Net retains the lowest errors across all eight completed benchmarks. Relative to the stronger of scOT and Drift-Net on each task, it reduces the step-averaged, full-rollout, and final-step errors by 59.2%, 59.2%, and 58.2%, respectively, averaged equally over the seven time-dependent tasks. On Poisson-Gauss, the error reduction is 61.9%. These improvements are achieved with 11M parameters, compared with 20M for scOT and 17M for Drift-Net.Experiments on Kolm-Smooth, Kolm-Turb, KS-Strong, and KS-Weak are still running. We will report the updated results once these runs are complete.

The model configurations of FNO, CNO, scOT follows the setting in Poseidon, and we use the same setting of Drift-Net and PDE-Transformer as their original paper.

There is a difference in how training examples are organized. The released Drift-Net dataset interface indexes individual trajectory–time pairs, allowing pairs from different temporal positions to be mixed during training. In our implementation, each batch contains complete trajectories. We then process their adjacent transitions sequentially, performing a separate optimizer update for each time index. Thus, samples within each update share the same temporal position. This introduces temporal homogeneity within each optimization batch, which may contribute to the substantially lower scOT and Drift-Net errors observed in our reruns compared with the original paper, as well as the stronger performance of scOT relative to Drift-Net on some tasks.

This training scheme allows us to systematically use all adjacent transitions from each sampled trajectory batch, with a fixed prediction interval matching the one-step rollout task. Processing the transitions in temporal order also ensures that each optimizer update contains examples at the same temporal position, rather than mixing states from different stages of evolution. This provides a consistent batching protocol across the compared architectures, although we do not claim that it is universally preferable to shuffled trajectory–time pairs.

Consequently, our experiments are retraining comparisons under our stated protocol, rather than direct reproductions of the published Drift-Net results.

We have updated the affected experimental results and the corresponding analyses in the revised manuscript. We have also clarified the architecture and configuration of each baseline in Appendix E.4 (Model Configurations).

Please report mean \pm std over at least 3 seeds for Table 2, and ideally for the top two methods in Table 1.

The multi-seed experiments is still running. We will report mean ± standard deviation once the additional runs are complete.

AuthorsOpenReview ↗

Answers for Q4-Q8

The spectral-only variant already beats the ablated baselines, so which component explains the gap to scOT? Two ablations would answer this directly: overlapping -> non-overlapping patches, and conv down/up-sampling -> patch merging.

VariantParamsACEFNS-KF
non-overlapping12M0.058/0.065/0.1010.189/0.189/0.322
patch merging11M0.025/0.027/0.0370.156/0.157/0.266
full model11M0.020/0.021/0.0330.098/0.098/0.157

We added both ablations while retaining the dual-branch configuration and the remaining backbone and training settings. Replacing the overlapping patch embedding/reconstruction with its non-overlapping counterpart increases final-step error from 0.033% to 0.101% on ACE and from 0.157% to 0.322% on FNS-KF. Replacing convolutional down/up-sampling with patch merging/expanding increases these errors to 0.037% and 0.266%, respectively.

These results indicate that patch embedding and multiresolution resampling also contribute to the full model’s performance.

We have added these additional experiments and the corresponding discussion to Section 4.3.

Please compare against an SSM-based neural operator (MambaNO or LaMO) under the same budget.

The LaMO experiments is still running. We will report the results once the additional runs are complete.

MambaNO already incorporates multi-directional Mamba scanning and convolutional processing within a U-shaped neural operator architecture. Given the similarity to the proposed spatial branch, can the authors clarify the architectural differences?

We agree that multi-directional Mamba scanning, convolutional processing, and a U-shaped multiresolution architecture are established components in MambaNO. We do not claim these components individually as novel. Beyond the four-directional configuration, we additionally evaluate an eight-directional variant by introducing forward and reverse scans along both diagonal families. This ablation examines whether additional scan directions improve prediction accuracy and whether the gains justify the increased parameter cost.

The main architectural distinction lies in how spatial interactions are parameterized within each block. MambaNO combines Mamba-based scanning with convolutional processing to capture global and local information. In SSF-Net, each resolution stage instead contains an explicit parallel spatial–spectral block: a multi-directional Mamba pathway processes one channel group, while a Fourier-domain pathway processes the other. Their outputs are summed and projected back to the block width. Thus, the distinction is the explicit coupling of scan-mediated, input-adaptive interactions with direct Fourier-domain interactions at each scale, rather than the introduction of multi-directional scanning itself.

We also distinguish the intended interaction biases from strict representational limitations: the Mamba pathway can capture long-range dependencies, but communication is mediated by recurrent scan paths, whereas the spectral pathway provides domain-wide interactions without sequential propagation. Our effective-kernel analysis is consistent with these complementary patterns.

Can the authors report comparable training time, inference latency or throughput, FLOPs, and peak GPU memory usage for SSF-Net and the strongest baselines?

We give the single sample inference latency and throughput on FNS-KF when batch size is 40 here.

We have added these additional experiments and the corresponding discussion to Appendix G.

modelinference latency(10−410^{-4}/s/sample)throughput(sample/s)
FNO2.93164.25
CNO9.2953.03
PDE-Transformer6.3077.94
scOT4.64105.02
Drift-Net4.8999.86
SSF-Net(patch=2)10.0549.13
SSF-Net(patch=4)6.7372.97

Can you ablate the patch sizes? and the channel size splits?

We have added the patch-size ablation before. Regarding channel allocation, our current SSF block uses element-wise addition to couple the two branch outputs. Since each branch preserves its channel width, this requires an equal spatial–spectral split. An unequal split would require additional projections or a different fusion mechanism, introducing architectural changes beyond channel allocation alone.

AuthorsOpenReview ↗

Answer for Q5

Please compare against an SSM-based neural operator (MambaNO or LaMO) under the same budget.

We have gotten the results of LaMO on some datasets. We follow the model configration of original LaMO paper, and transfer it to our training framework. All the error results are in percent.

DatasetLaMOSSF-Net
Params10M11M
Poisson-Gauss1.070.16
NS-PwC2.07/2.07/3.660.10/0.10/0.17
NS-Tracer-PwC1.61/1.61/3.210.18/0.18/0.33
NS-SL0.64/0.65/1.250.23/0.24/0.48
CE-RP4.13/4.08/8.072.75/2.72/5.71
Wave-Gauss5.43/5.28/9.024.43/4.32/7.10
ACE0.11/0.12/0.160.02/0.02/0.03

The dataset experiments are still running. We will report the results once the additional runs are complete.

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

The paper studies how to capture local and global interactions in PDE operator learning. The authors propose SSF-Net, which combines a spatial branch using multi-directional Mamba and a spectral branch using a Fourier transform within a multiresolution encoder-decoder. Experiments on twelve two-dimensional PDE benchmarks show lower prediction errors with fewer parameters than the compared baselines. The paper also evaluates prediction accuracy over 100 autoregressive steps.

Critical strengths
  1. The method and architecture are well designed. The motivation is clear and supported by Fig. 1, which shows that existing operators do not capture decaying local interactions. Within the SSF block, the multi-directional Mamba spatial branch and the AFNO-style spectral branch are both well motivated, and Appendix justifies their pairing by casting both branches as parameterizations of one kernel integral operator.
  2. The experiments covers various PDE benchmarks . Ablation studies and effective analysis help clarify the contributions and interaction patterns of the two branches.
Critical weaknesses

The ablation study covers only ACE and FNS-KF, and the spatial branch alone performs better on ACE with more parameters. Comparisons at similar parameter counts across more tasks would better establish the benefit of spatial–spectral fusion.

Questions for authors
  1. The paper explains the different ablation results on ACE and FNS-KF in terms of their local and global interactions. Can this explanation help guide the choice between a single branch and spatial–spectral fusion for a new PDE problem?
  2. How should we interpret the ablation results given the different parameter counts of the variants? Would the same conclusions hold when the models have similar parameter counts?
2 comments
AuthorsOpenReview ↗

Replies for all questions

The paper explains the different ablation results on ACE and FNS-KF in terms of their local and global interactions. Can this explanation help guide the choice between a single branch and spatial–spectral fusion for a new PDE problem?

The explanation provides a physics-informed starting point for architecture selection, although we view it as a qualitative guideline rather than a universal rule. When localized, state-dependent interactions dominate the target solution operator, a spatial-only model is a reasonable candidate. This is consistent with ACE, where the spatial-only variant achieves the lowest errors.

When local transport coexists with nonlocal coupling, spatial–spectral fusion is a natural candidate. For FNS-KF, which combines advection with nonlocal pressure coupling, the full model achieves the lowest final-step error, while matching the spatial-only model’s step-averaged error with fewer parameters.

For a new PDE problem, we therefore suggest using the PDE’s physical mechanisms to assess whether local interactions dominate or whether nonlocal coupling is also important. Our current results support this selection rationale, but do not establish a definitive criterion for unseen PDEs.

How should we interpret the ablation results given the different parameter counts of the variants? Would the same conclusions hold when the models have similar parameter counts?

Our ablations keep the backbone configuration fixed, including the number of stages, depth per stage, and encoder–decoder layout. In each single-branch variant, the retained branch is widened to preserve the block’s channel dimensions. Because the spatial and spectral branches have different parameterizations, this produces different parameter counts despite the shared backbone configuration. Thus, the current ablations compare branch choices within a fixed backbone, rather than under a matched parameter budget.

Matching parameters would require adjusting widths or other capacity settings, providing a complementary comparison to our fixed-backbone ablation. We adjust the embedding dimension to implement larger spectral branch only model and smaller spatial branch only model, and found our current fixed backbone configuration is better than those models.

We have added these additional experiments and the corresponding discussion to Appendix F.1, and clarified their implications in Section 4.3.

VariantParamsACEFNS-KF
Spectral branch only10M0.022/0.024/0.0340.122/0.122/0.206
Spatial branch only8M0.022/0.023/0.0320.109/0.109/0.172
Spectral branch only (fixed backbone)9M0.021/0.023/0.0330.115/0.115/0.188
Spatial branch only (fixed backbone)16M0.012/0.014/0.0260.098/0.098/0.168
full model11M0.020/0.021/0.0330.098/0.098/0.157

Dear Reviewer

Thank you for your feedback. As the discussion period ends on November 18, we would like to follow up on our response and revised manuscript , in which all changes are marked in blue. To address the influence of parameter counts, we added adjusted-width single-branch comparisons in Appendix F.1 and clarified their interpretation in Section 4.3. These experiments distinguish the original fixed-backbone comparison from comparisons at similar, though not identical, model sizes and show that both branch composition and capacity affect performance.

We also clarified that the local/global interaction explanation provides a qualitative starting point for architecture selection, rather than a validated criterion for unseen PDEs.

We would appreciate hearing whether these additions address your concerns. If so, we would be grateful if you could consider updating your assessment. We welcome any further questions.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.