acceptodds
Under review as a conference paper at ICLR 2027

Conditioning Spatial Representations in Fourier Neural Operators

Abstract

Geometric and operating parameters can change both the magnitude and spatial organization of a physical field. For Fourier neural operators (FNOs), this raises a design question: when should parameters participate in constructing spatial features, in addition to entering the input or modulating Fourier layers? We study this question with a parameter-conditioned sinusoidal coordinate network that supplies features to Fourier processing. Controlled heat-transport problems with translating and expanding material rings distinguish the benefit of learning spatial features from the additional benefit of conditioning them. A shared spatial representation is effective in the expanding-ring setting; conditioning provides a larger gain in the translating-ring setting, improving whole-field accuracy, localized thermal structure, and responses to geometry changes. In radio-frequency hollow cathode discharge, an industrial plasma problem in semiconductor manufacturing, the model predicts four coupled fields across operating conditions in a fixed cathode geometry and improves on the evaluated baselines. These results highlight spatial feature construction as a consequential choice in conditioning physical-field predictors.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, plasma physics, 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_g7GyOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a targeted architectural question for Fourier Neural Operators (FNOs): whether low-dimensional system parameters should also participate in constructing spatial features, not just in input channels or by modulating Fourier layers. The authors propose a parameter-conditioned sinusoidal coordinate network (a modulated SIREN) that produces configuration-dependent spatial feature maps, which are then processed by FiLM-conditioned Fourier blocks. Through controlled heat-transport problems (translating vs expanding material rings) and an industrial radio-frequency hollow cathode discharge (HCD) application, they show that learning spatial features is beneficial over standard FNOs, and conditioning these features yields additional gains, especially when geometry translates. On HCD (fixed geometry, varying operating parameters), their shared-embedding variant achieves the lowest errors across four coupled physical fields compared with FNO, Transolver, and DeepONet.

Critical strengths

The paper uses a principled conditioning point: parameters modulate the coordinate-based spatial representation before Fourier processing, complementing standard FiLM conditioning in FNO layers. The architectural decomposition is clear between “spatial feature construction” and “nonlocal mixing” (Fourier blocks), enabling controlled ablations via a static-spatial variant. The method also represents a simple and general mechanism (amplitude and phase modulation of SIREN layers) that can be plugged into operator-learning backbones. The motivation (“where to condition?”) is well articulated, and the architecture and training setup are described with sufficient detail.

Critical weaknesses
  1. The technical novelty is very incremental in that conditional neural representations and FiLM conditioning are established, and several geometry-aware operator methods exist. The core innovation is the careful placement of conditioning rather than a fundamentally new operator class.
  2. No theoretical analysis or principled criteria for when spatial conditioning helps, beyond empirical differentiation between translation and expansion.
  3. Comparisons miss some relevant geometry-/representation-aware variants (e.g., Geo-FNO/DAFNO/GINO/ACT, hpFNO/HyperFNO); while not all are strictly necessary for the fixed-grid setting, such baselines could contextualize the gains more broadly.
  4. Extrapolation and OOD robustness (e.g., beyond the trained geometry/parameter ranges) are not evaluated.
  5. Computational/runtime overhead of the added SIREN branch and modulation is not reported. the scaling implications are also unclear.
Questions for authors
  1. How does the approach perform under parameter/geometry extrapolation (e.g., translation or expansion beyond the training interval) and under sparser training coverage? Any signs of overfitting or failure modes?
  2. Can you provide runtime and memory overhead of the conditioned SIREN branch relative to FNO/FiLM FNO, and scaling considerations for higher resolutions or 3D?
  3. What is the sensitivity to SIREN hyperparameters (depth/width/frequency multiplier) and to the choice of positional encoding (pure coordinates vs Fourier features)? Are phase vs amplitude modulations equally important?
  4. In the heat-transport setup, what is the contribution of Fourier-layer FiLM vs spatial SIREN conditioning? An ablation that removes operator FiLM (keeping spatial conditioning) and vice versa would help quantify their roles.
  5. How would your method compare to coordinate-adaptation (e.g., ACT) or learned-deformation FNOs (Geo-FNO) and DAFNO on the same translating/expanding ring tasks? Do you expect complementary gains if combined?
  6. On HCD, did you try training a single shared model across multiple geometries (e.g., altering bore radius or electrode segments) to test whether spatial conditioning scales to modest geometry changes?
  7. Are the localized metrics (transition/hotspot) robust to threshold choices (e.g., different occupancy bands or top-k% regions)? A sensitivity analysis would strengthen conclusions.
  8. Could you share per-field loss contributions and qualitative failure cases on HCD (e.g., near electrode edges) to better understand where improvements are concentrated?
2 comments

Dear reviewer g7Gy,

Thank you for the detailed questions. We have added ablations, generalization tests, and cost analyses, and included ACT—which we agree is interesting concurrent work—in both our empirical comparisons and related-work discussion. We also agree that HCD as a important and practical problem deserves more detailed discussion, we agree to expand a varying geometry setup for HCD if the paper is accepted --- we need nontrivial time to generate the data, due to limited time and access it is not possible in the current rebuttal period.

Part ONE

The contribution: a spatial-conditioning principle

Our goal is to demonstrate spatial conditioning as a design principle, rather than introduce a new neural-operator family: system parameters can help construct the spatial representation before information is mixed across locations. FNO is our main example. To test the same principle with a different mixing mechanism, we also add static and conditioned spatial branches to Transolver, leaving its Physics-Attention blocks unchanged.

Unless stated otherwise, results are mean ± sample SD over three training seeds on fixed splits; percentage reductions are averaged over matched seed pairs. ACT uses two seeds and HCD uses five.

For Transolver at 2,000 updates, global scaled MSE (× 10⁻³) is:

TaskOriginal TransolverStatic spatial branchConditioned spatial branch
Translation3.475 ± 0.2591.065 ± 0.2320.824 ± 0.081
Matched two-source expansion2.595 ± 0.9320.610 ± 0.1410.488 ± 0.094

Conditioning improves all four field-error metrics over the static branch in every run, with 2,064 additional parameters. This supports the principle in both Fourier and attention-based backbones under the tested settings. The separate budget permits a within-Transolver comparison, not a ranking against the 32,000-update FNO results (Appendix I.2, Table 24).

A theory predicting when conditioning helps remains open. The results also go beyond a simple translation/expansion distinction: retaining translation's two-source forcing and diffusivity during expansion reduces Static-to-Joint MSE from 4.336 ± 0.774 to 3.193 ± 0.816 (× 10⁻⁵), with consistent global, physical, transition and response gains. Original single-source expansion is less consistent, indicating that forcing and material properties also matter (Appendix F.3, Tables 12–13).

1. Extrapolation and sparse coverage

We train and validate on ring positions s ∈ [0.1, 0.9], then test 128 in-range positions and 32 on each side. Models share cases, training-only normalization, 32,000 updates and validation selection. Global scaled MSE is:

ModelIn-range (× 10⁻⁵)Combined OOD (× 10⁻²)
FiLM FNO8.969 ± 1.8652.865 ± 0.354
Static Spatial8.389 ± 1.9452.351 ± 0.223
Ours4.338 ± 0.9590.942 ± 0.347

Our OOD MSE is 59.24 ± 16.85% lower than Static, with both tail means improving in every run. This test retains the original Static/Ours Fourier conditioning on propagation and coupling parameters, separately from the all-nine-parameter control below.

Our OOD MSE remains 157–250 times our own in-range MSE. In one run, Static wins on 10/32 positions just above s = 0.9 despite our lower upper-tail mean; our pooled peak-value error is also higher, at 7.57% versus 5.10% (Appendix G.1, Tables 14–16, Figure 12).

With 16 training geometries, physical relative L² (× 10⁻³) is 1.791 ± 0.614 for Ours, versus 3.156 ± 0.365 for FiLM FNO and 3.217 ± 0.455 for Static. All five field/response metrics improve over both controls in every run with either 16 or 32 training geometries. Removing s ∈ (0.4, 0.6) from training and validation gives mixed results: physical/response gains persist, but Static wins global, transition and hotspot MSE in one run. This interpolation gap reveals sensitivity to missing coverage without establishing overfitting as its cause (Appendix G.2, Tables 17–18).

2. Runtime, memory and scaling

At 256×256, batch 16 and 32,000 updates on DGX Spark, spatial conditioning adds 6.34 ± 0.67% fitting time and 7.05% peak memory over the matched all-parameter Static model. In the separate extrapolation experiment, the complete Ours model adds 27.73 ± 0.35% time and 25.49% memory over FiLM FNO. Time includes validation, checkpoint saving and replay; memory measures peak PyTorch allocation (Appendix J, Table 25).

Higher-resolution and 3D scaling remain unmeasured. The Fourier-mode study changes bandwidth on the same grid; at 24×24 modes, Joint improves all five field/response metrics in every run (Appendix G.3, Table 19).

Part TWO

3. SIREN settings and modulation

Depth, width, frequency and encoding sensitivity studies, and amplitude-only/phase-only ablations, remain incomplete. The ring branch uses three width-16 sine layers, frequency multiplier one and periodic coordinates; HCD uses three width-128 layers, frequency multiplier 30, normalized coordinates with Gaussian Fourier features, and a shared 128-dimensional condition embedding.

The capacity-matched conditional SIREN control, without Fourier mixing, is less accurate on both ring tasks. It tests mixing, rather than the relative importance of amplitude and phase or the optimal SIREN settings (Appendices A.1, A.2 and F.2).

4. Spatial conditioning versus Fourier FiLM

The 2×2 translation experiment gives every model identical direct inputs, including all nine parameters, and matches data and training. Global scaled MSE (× 10⁻⁵) is:

Spatial featuresNo Fourier FiLMAll-nine-parameter Fourier FiLM
Static11.581 ± 1.5029.811 ± 1.269
Conditioned3.664 ± 0.7634.435 ± 0.882

Spatial conditioning reduces MSE by 68.57 ± 2.63% without Fourier FiLM and 54.49 ± 9.30% with it, improving every run. The latter pair has 117,058 real parameters each; the no-FiLM pair has 114,994 for Static and 117,058 for Conditioned.

Adding Fourier FiLM to Conditioned instead increases MSE by 22.71 ± 24.61% in aggregate and worsens every run. Conditioned without Fourier FiLM has the lowest mean error. Because the combined model shares its encoder and modulation heads between routes, the cause of this adverse interaction remains unresolved (Section 5.3, Table 2; Appendix F.1, Tables 9–10).

5. ACT and other geometry-aware baselines

We add the author's ACT blocks after each FiLM FNO block, without a SIREN branch. All models use all-nine-parameter Fourier conditioning and matched data, preprocessing, optimization, 32,000 updates and validation selection. Translation reuses the factorial Static/Joint controls; expansion uses the original single-source task. Global scaled MSE from two validated seeds is:

TaskFiLM FNOStaticJointACT+FiLM
Translation MSE, × 10⁻⁵10.709 ± 4.26010.476 ± 0.7494.386 ± 1.2413.846 ± 0.044
Expansion MSE, × 10⁻⁶4.277 ± 1.5004.491 ± 2.3754.007 ± 2.2890.809 ± 0.163

ACT and Joint each win one translation run. On expansion, ACT improves all four main field metrics over Joint on every test case in both runs. Joint improves all five field/response metrics over Static on both tasks.

These gains do not extend to every metric. Joint sometimes has larger peak-position errors than Static, and in one expansion run it has higher global, physical and transition errors than FiLM FNO. ACT also does not improve peak-value or peak-position errors on every expansion case.

ACT has 131,040 real parameters versus Joint's 117,058 and uses about 17.4× the fitting time and 5.45× the peak memory in this author-block adaptation with deterministic sampling. Updates are matched, not computing time. The third seed remains unvalidated (Appendix I.1, Tables 22–23, Figure 17; Appendix J, Table 26).

Joint constructs input features; ACT resamples intermediate features. Their combination remains untested, as do direct Geo-FNO, DAFNO and GINO comparisons. Our spatial-shift control adapts the official hpFNOx blockwise hyper-shift to two dimensions, rather than benchmarking the full hpFNO/HyperFNO family.

Joint consistently outperforms this shift control on translation. Expansion is mixed: the shift control has lower global, physical, transition and response errors in one run, while Joint consistently has lower hotspot and peak errors.

6. HCD across device geometries

HCD uses one fixed device geometry with varying pressure, RF voltage and SEEC. A shared model across bore radii will be added if the paper is accepted as agreed before.

7. Threshold sensitivity

We rescored saved predictions using occupancy bands (0.05,0.95), (0.1,0.9), (0.2,0.8), (0.3,0.7) and reference hotspot fractions 1%, 2.5%, 5%, 10%. The matched all-parameter Conditioned model improves on Static for all eight choices in every run: 40.65–65.20% lower transition MSE and 48.31–68.52% lower hotspot MSE. Masks are evaluation-only; thresholds are not independent training repetitions (Appendix H.2, Table 21).

8. HCD field losses and failures

Table 4 reports normalized-space test relative L², not contributions to training loss. Over five seeds, Ours gives 3.246 ± 0.370, 2.313 ± 0.190, 2.092 ± 0.160 and 2.491 ± 0.100 (× 10⁻³) for potential, electrons, Ar⁺ and Ar*, respectively.

The field loss averages four standardized-space relative L² terms and adds 0.1 times joint relative Frobenius error, with 10⁻⁸ in denominators. Training also adds a 0.25-weighted radial adjacent-grid penalty on the predicted metastable field; validation excludes it.

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

This paper adds a spatial feature network before the existing Fourier neural operator. It simultaneously receives position and system parameters, generates spatial features that vary with conditions, and then feeds them into the subsequent network to predict the physical field.

Critical strengths
  1. The paper sets up a coordinate-only Static Spatial control, which retains the same direct input path and Fourier stage conditional information as the proposed method, and compares them under similar parameter counts, training budgets, and validation set modulus selection methods. This helps me distinguish the benefits of "introducing a learnable spatial representation" from "making the spatial representation vary with system parameters".

  2. The article introduces parameter conditionalization into the spatial representation construction stage, enabling the representation of the same spatial location to vary with geometry or operating conditions.

  3. The paper also evaluates the overall prediction error, the material transition region error, the high-temperature region error, and the field change error caused by geometric changes.

Critical weaknesses

1.Existing experiments do not provide a complete answer to the question of "at which stage should the parameters take effect." If subsequent Fourier modules are also allowed to adjust directly based on geometric changes, will spatial feature networks still have the same additional advantages?

  1. Appendix A.2 states: "Every model uses normalization statistics fitted over all 125 cases before splitting." However, the authors mention that the number of training, validation, and test samples were 101, 12, and 12, respectively. Information from the test data was used in the preprocessing. Typically, these statistics should only be calculated using the training set, and the same transformation should then be applied to the validation and test sets.
Questions for authors
  1. Appendix A.2 indicates that the standardized statistics are calculated from all 125 samples. Could it be changed to calculate the input and output statistics using only 101 training samples, and report whether the results in Table 3 are preserved?

  2. In the FiLM FNO heat conduction experiment, which parameters are used to adjust the Fourier module? Do geometric parameters participate in this process? If the Fourier module of Static Spatial also varies with the geometric parameters, will the new method still offer additional benefits?

  3. In the plasma model, what would be the result if we retained other structures such as output mixing and second-stage correction, but changed the spatial characteristics to be shared by all operating conditions?

1 comment

Thank you for pointing out these issues. We have added the geometry-aware Fourier control and clarified what the current HCD experiments can and cannot establish.

Normalization in the HCD experiment

You are right that normalization statistics should be fitted on the training cases only. Our HCD experiment computed them from all 125 cases before the 101/12/12 split, so information from the held-out cases entered preprocessing.

We have not yet completed a rerun using only the 101 training cases to fit these statistics. The HCD results in the revised manuscript, now Table 4, therefore remain results from the original preprocessing protocol. We cannot confirm that the reported rankings would remain unchanged after correcting normalization. This limitation is stated alongside the HCD results and in Appendix A.2, and the abstract and conclusion have been qualified accordingly.

The ring experiments use training-only normalization and provide separate evidence for the benefit of spatial conditioning.

Giving the Fourier blocks access to geometry

In the original ring experiments, FiLM FNO used all nine system parameters to modulate its Fourier blocks. Static Spatial and Ours used only the propagation and coupling parameters, which remain fixed during each geometry sweep. All models also received all nine parameters through their direct input.

To address your question, we added a translation experiment that combines static or conditioned spatial features with either no Fourier FiLM or Fourier FiLM using all nine parameters. The models share data, training-only normalization, shared-tensor initialization, 32,000 updates and validation-based checkpoint selection.

Global scaled MSE is shown below as mean ± sample SD over three training seeds:

Spatial featuresNo Fourier FiLMFourier FiLM using all nine parameters
Static11.581 ± 1.5029.811 ± 1.269
Conditioned3.664 ± 0.7634.435 ± 0.882

All values are × 10⁻⁵.

When both models' Fourier blocks receive geometry, spatial conditioning still reduces global MSE by 54.49 ± 9.30%, with improvements in global, physical, transition, hotspot and response errors in every run. This comparison has exactly 117,058 real parameters in each model. The percentage reduction is calculated within each seed before averaging.

The experiment also shows that combining the two conditioning routes does not necessarily help. Adding Fourier FiLM to the conditioned spatial model increases MSE by 22.71 ± 24.61%, worsening every run. Among these four variants, conditioned spatial features without Fourier FiLM give the lowest mean error. The combined model shares its encoder and modulation heads between the two routes, so this experiment does not identify the cause of that interaction. The no-FiLM pair has different parameter counts: 114,994 for Static and 117,058 for Conditioned.

Section 5.3, Table 2 and Appendix F.1, Tables 9–10, report the complete comparison.

Isolating spatial conditioning in HCD

The ablation you suggest—replacing the conditioned spatial features with static features while retaining field mixing and second-stage correction—has not been evaluated due to limited access to the original HCD dataset. Our HCD model combines these components, and its comparison with FNO changes several of them together. The observed improvement therefore cannot be attributed to spatial conditioning alone. We make this explicit in Appendix A.2. The matched ring controls isolate the effect in the ring setting, but the corresponding HCD ablation remains an open question. And since generating them is hard due to the short time given, we agree to test on extra setup of experiments on HCD if the paper is accepted --- we agree with you that HCD as an important practical problem deserves an more cautious treatment.

Reviewer_UeJ9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a Fourier Neural Operator augmented with a parameter-conditioned SIREN network, which learns spatial feature fields that change with the system’s geometry and physical parameters. These features are added to the FNO’s input to improve predictions. The paper presents experiments on heat transport simulations with moving or expanding rings of lower conductivity, and on fixed-geometry plasma discharge simulations, where the discharge conditions are changed. Both experiments show improved accuracy over baseline tests with DeepONet, Transolver and FNO. The paper is well written and seems to bring novelty for Neural Operators which could improve generalisation, accuracy, and potentially extrapolation.

Critical strengths

The paper provides clear and well-motivated augmentation of the Fourier Neural Operator with parameter-conditioned SIREN features, allowing the learned spatial representation to adapt to changes in geometry and physical conditions. The method is evaluated on two different use cases with very different underlying physics, although both 2D in nature. The method consistently improves prediction accuracy over the evaluated FNO, DeepONet and Transolver baselines.

Overall, the work is well presented and brings a meaningful and potentially useful contribution to neural-operator modelling, with promising implications for generalisation across varying physical configurations.

Critical weaknesses

There are no critical weaknesses, but rather some suggestive improvements below.

Questions for authors

It would be interesting to plot some of the learned spatial feature channels across a few cases, for example for the heat transport rings. Presumably at least some or all of the features contain information about the underlying parameters (eg. the location of the ring). Showing the SIREN output features that are fed into the FNO would help the reader visualise the method, for example these could be plotted in the Appendix for the two left plots of Figure-1 (with the ring at two different locations).

It would be interesting to know whether the conditioned spatial features also help for extrapolation beyond the training parameter range. Have the authors explored whether their advantage over the baselines persists for extrapolated input parameters (for example, if the ring position or radius is taken to be outside the training set interval)?

1 comment

Thank you for these suggestions. We agree with them and have added visualizations of the learned spatial features and an experiment testing extrapolation beyond the training range.

Visualizing the learned spatial features

Figures 13–14 show all 16 SIREN output channels and their projected features at three ring positions, s = 0.25, 0.50 and 0.75. Here, s controls the ring's translation while its radius remains fixed. The projected features are added to the FNO's input features before the Fourier blocks.

The conditioned features change in both magnitude and spatial pattern as the ring moves, whereas the Static Spatial features remain unchanged. Each channel uses the same color scale across the three positions, making these changes directly visible. The figures show one training run; Appendix H.1 and Table 20 report the numerical feature audit across three runs.

To check whether this adaptation helps prediction, we also compare static and conditioned spatial features while giving both models' Fourier blocks all nine system parameters. With identical parameter counts, spatial conditioning reduces global MSE by 54.49 ± 9.30% across three training seeds (Section 5.3, Table 2). Thus, its benefit persists when the Fourier blocks already receive geometry information.

Extrapolating to unseen ring positions

The average advantage also persists outside the training range. We train and validate on s ∈ [0.1, 0.9], then test on 128 positions within this interval and 32 positions on each side of it. All models use the same cases, training-only normalization and 32,000 updates. This experiment retains the original Static/Ours models, whose Fourier modulation uses propagation and coupling parameters; it is separate from the all-nine-parameter control above.

The table reports global scaled MSE as mean ± sample SD over three training seeds:

ModelIn-range (× 10⁻⁵)Below training range (× 10⁻²)Above training range (× 10⁻²)Combined OOD (× 10⁻²)
FiLM FNO8.969 ± 1.8654.458 ± 0.7451.273 ± 0.2122.865 ± 0.354
Static Spatial8.389 ± 1.9453.541 ± 0.3091.160 ± 0.3722.351 ± 0.223
Ours4.338 ± 0.9591.291 ± 0.6020.593 ± 0.1260.942 ± 0.347

Combining both outside ranges, our MSE is 59.24 ± 16.85% lower than Static Spatial, with the mean error on each side improving in every run. Percentage reductions are calculated within each seed before averaging.

Extrapolation errors nevertheless remain much larger than errors within the training range: our own OOD MSE is 157–250 times our in-range MSE. The advantage also varies across individual cases and metrics. In one run, Static performs better on 10 of the 32 positions just above the training boundary at s = 0.9, although our average MSE across all 32 positions is still 6.70% lower. In that same run, our average peak-value relative error across both outside ranges is higher: 7.57%, compared with 5.10% for Static.

These results show lower average extrapolation error, with exceptions that remain visible in the case-level results. The experiment tests ring-position extrapolation; radius extrapolation and more general geometry changes remain open. Appendix G.1, Tables 14–16 and Figure 12 provide the full results.

Comments on the paper

2 comments
AuthorsOpenReview ↗

A summary of the updated draft and a thank-you note to reviewers

Dear Reviewers,

We have uploaded the new draft in response to your concerns and suggestions for improvement. We thank you for all your carefully drafted reviews and summarize the main points below:

The revised manuscript makes our original focus clearer: how system parameters shape the spatial features used for prediction. FNO is our main example of this idea. The added Transolver experiments apply the same spatial-conditioning approach while leaving its attention blocks unchanged, and show improvements over static spatial features on both translation and expansion.

The new controls clarify what this conditioning contributes. Even when the Fourier layers receive all geometry parameters, conditioning the spatial features reduces global MSE by 54.49 ± 9.30% across three seeds, with identical parameter counts. The benefit also appears without Fourier FiLM, although combining the two modulation routes can hurt performance. Matching the forcing and material properties reveals gains on expansion as well, making the original translation–expansion comparison easier to interpret.

The additional generalization and diagnostic experiments give a fuller picture of these gains. Average position-extrapolation MSE falls by 59.24 ± 16.85% relative to Static, and improvements persist with sparse training geometries and increased Fourier bandwidth. Physical and response errors also improve across an unobserved interior interval. The feature plots show how the representation changes with geometry, while the threshold study shows that the regional improvements persist across different definitions of transition and hotspot regions. Case-level reversals remain in the reported results.

The alternative baselines place these findings in context. Joint improves on conditional SIREN on both ring tasks and on the adapted hpFNOx shift control for translation, with mixed expansion results. We include ACT as interesting concurrent work in both the comparisons and related work. In the two-seed comparison, ACT and Joint each win one translation run; ACT is stronger on expansion but substantially more expensive. Spatial conditioning itself adds 6.34 ± 0.67% fitting time over the matched Static model.

These additions appear in the new Section 5.3 and Appendices E–J, covering protocols, conditioning controls, generalization, feature diagnostics, alternative backbones and computational costs. The HCD discussion retains the lower mean errors across all four fields while making its preprocessing and attribution limitations explicit. The manuscript also retains our commitment to release the code upon acceptance. On top of that, we also added a new commitment to test the new setup of HCD dataset.

Best, Authors

AuthorsOpenReview ↗

Added results for GEO-FNO/DAFNO/GINO on our proposed translation/expansion heat transport problems

Dear reviewers,

Thank you again for the time and care you have put into reviewing our work. Following your suggestions, we have completed additional three-seed comparisons to further evaluate our spatial-conditioning design principle.

Joint achieves lower mean error than eDAFNO and GINO on both heat-transport tasks, and than Geo-FNO on translation.

Our two ring tasks test complementary changes in a material interface: translation moves a fixed-radius ring, while expansion changes its radius around a fixed center. Both operate within a fixed physical domain. Evaluating both tasks reveals differences in model performance that a single task would miss.

Global MSE, reported as mean ± sample SD over three seeds; lower is better:

ModelTranslation, ×10⁻⁵Expansion, ×10⁻⁶
Joint4.43 ± 0.883.94 ± 1.62
Joint + SDF6.15 ± 1.203.96 ± 1.49
Geo-FNO7.84 ± 0.321.85 ± 0.09
eDAFNO9.08 ± 4.597.02 ± 4.03
GINO (256×256)17.06 ± 8.584.52 ± 1.10

All models use 256×256 physical input/output grids. GINO and Joint + SDF receive the same signed-distance input.

Joint performs best on translation among these methods, while Geo-FNO performs best on expansion. The different rankings highlight the value of the task pair and help identify where our spatial-conditioning approach is particularly effective.

Your suggestions have helped us make the evaluation more complete and the conclusions better supported. Thank you again for your constructive feedback.

Best regards, 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.