acceptodds
Under review as a conference paper at ICLR 2027

XENO: Interface-Enriched Neural Operators for Stable Rollouts on Sharp Moving Interfaces

Abstract

One-step accuracy does not predict autoregressive rollout stability for neural operators on sharp moving interfaces. We study this failure in Stefan phase-change problems and find that FNO and Transolver, despite one-step interface errors below px, reach 20-step errors of and px, respectively, in the sharp R4 regime, while remaining stable in the smooth R3 regime. We introduce XENO, which augments a spectral neural operator with an Operator Enrichment Block (OEB): interface-conditioned basis functions modulated by a learned signed-distance-like field and gated by ReZero. On R4, XENO reduces the median 20-step Hausdorff error from px for FNO to px under held-out validation-rollout checkpoint selection. Matched ablations show that this improvement is not explained by rollout-noise training, parameter count, gating, phase access, or alternative geometric conditioning. The improvement is regime-specific: XENO is worse than the strongest baseline on R3 and Burgers, and of R4 trajectories still exceed px. We also find that spectral-energy collapse accompanies rollout failure, but do not establish it as a cause.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Operator Learning, Scientific Machine Learning, Partial Differential Equations, Moving Interfaces, Phase-Change Problems, Autoregressive Rollouts, Interface-Aware 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_YqJZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies how to improve the rollout stability of neural operators on Stefan phase-change problems with sharp moving interfaces. The authors propose XENO, using U-Net to predict a interface field and OEB to add corrections to neural operator. Experiments show substantially reduced rollout interface errors in the sharp R4 regime compared with baselines including FNO and Transolver.

Critical strengths
  1. The problem is clearly motivated and the method is effective in the target regime. The paper provides detailed analysis about the phenomenon and experiments about the method.
  2. The evaluation covers diverse baselines and includes informative ablations and scope analyses. Results on additional regimes, together with failure rates and computational costs, clarify the method’s benefits and limitations.
Critical weaknesses
  1. The mechanism is not fully explained. The spectral measurements and ablations support the effectiveness of interface-conditioned enrichment, but do not establish what drives error amplification during rollout or how OEB suppresses it.
  2. The evidence for broader applicability is limited. The largest gains are concentrated in the R4 Stefan regime, while improvements are small or absent in other equations. It would be helpful to know whether the improvement extends to other interface problems, or how to identify problems where XENO is beneficial.
Questions for authors
  1. The paper describes the R4 failure as heterogeneity-driven, but R3 and R4 differ in several factors at once, including material heterogeneity, interface sharpness and boundary conditions. The Allen–Cahn sweep is also confounded by its restoring reaction term. Could you vary one factor at a time, to identify which property of R4 causes the baselines to diverge?
  2. What evidence explains how OEB changes error propagation during rollout in R4? Could you track the evolution of a small, physically consistent interface perturbation under FNO and XENO in both R3 and R4?
  3. Beyond the regimes already evaluated, do the results suggest measurable characteristics that could identify other problems likely to benefit from XENO? How might the findings guide its use in a new physical setting?
3 comments

This comment addresses: Weakness W1 and Question Q2.

Thank you for the thoughtful review and the constructive questions. Your suggestions to track a perturbation and to vary one factor at a time were concrete enough that we could run both during the discussion period. The results are below.

Where each point is answered:

  • W1 + Q2 (mechanism; perturbation experiment): comment 1/3
  • Q1 (one factor at a time): comment 2/3
  • W2 + Q3 (broader applicability; how to identify suitable problems): comment 3/3

Regarding W1: "... do not establish what drives error amplification during rollout or how OEB suppresses it." and regarding Q2: "Could you track the evolution of a small, physically consistent interface perturbation under FNO and XENO in both R3 and R4?"

We ran this experiment. We grew the liquid region of the initial state by 1 px. In the new liquid cells we set T to the melting temperature and H from the trajectory's enthalpy relation, so the state stays physically consistent. Then we rolled out the original and perturbed states with the same model (50 test trajectories per regime). Median distance between the two predicted interfaces (px):

regimemodelstep 151020
R4FNO1.411.414.1426.75
R4XENO1.411.441.461.44
R3FNO0.660.921.462.36
R3XENO1.521.111.472.28

On R4, FNO grows a 1-px perturbation about 19× in 20 steps. XENO keeps it at its starting size. On R3, neither model shows the large growth seen for FNO on R4. The final separation is 2.36 px for FNO and 2.28 px for XENO, compared with 26.75 px for FNO on R4. This shows that the perturbation behaves differently under FNO and XENO. It does not isolate which part of OEB causes the difference, or explain why R4 produces the growth (see Q1).

A possible explanation, consistent with but not proven by these numbers: in FNO, a local error at the interface spreads into all Fourier modes and across the domain. On R4, where the temperature field has much sharper features (99th-percentile |∇T| 15× larger than on R3), this spread error grows step by step. XENO's enrichment acts near the predicted interface, so it can correct the error locally. This agrees with the spectral result in Appendix A: FNO's high-frequency error share rises during rollout on R4, but not on R3 and not for XENO.

Planned revision: a new subsection in Section 4 with this table and a figure. The spectral results (Section 4.3, Appendix A) are given as supporting evidence.

Resolution: on R4, a small, physically consistent interface perturbation grows 19× in 20 steps under FNO and does not grow under XENO. On R3 both end near 2.3 px, with no large growth. What drives the growth within R4 is in Q1.

This comment addresses: Question Q1.

Regarding Q1: "R3 and R4 differ in several factors at once ... Could you vary one factor at a time?"

We varied one factor at a time on R4.

Material information (full data). Giving FNO the exact material layout drops its 20-step error from 30.66 to 2.00 px. Giving XENO the same layout also helps (1.00 → 0.21 px). XENO stays ahead on moving fronts (0.79 vs 12.90 px) and at 100 steps (19.84 vs 30.85 px). Details in Reviewer: mXQd-Q1.

One-factor R4 variants. We regenerated R4 changing one factor at a time and kept everything else the same: (a) uniform material (same diffusivity everywhere); (b) mushy band twice as wide (a wider melting range); (c) constant ambient temperature (no time-varying boundary forcing). These are separate, lower-data experiments. Their original-R4 numbers should not be compared with the full-data results above. FNO and XENO were trained from scratch for each, with the same protocol, on 100 train, 50 validation and 50 test trajectories. This is the size that fit the discussion period. Median px on the 50 test trajectories:

conditionFNO 20 / 100 stepsXENO 20 / 100 steps
original R423.68 / 149.5522.16 / 24.42
uniform material24.32 / 28.5018.39 / 45.18
mushy band ×226.73 / 119.8624.27 / 27.27
constant ambient59.42 / 150.2622.06 / 31.83

With 1/16 of the training data, neither model gets below 5 px in any condition, including the original. Within that limit:

  1. No single change makes FNO stable at 20 steps.
  2. XENO has lower 20-step error than FNO in all four conditions (per-case: 41/50 original, 48/50 uniform).
  3. At 100 steps FNO blows up (about 120–150 px) on original, mushy and constant-ambient. XENO stays at 24–32 px. With uniform material FNO does not blow up (28.50 px), while XENO reaches 45.18 px.

Limits: (b) also changes the latent-heat release, so it is not a pure sharpness change. Run-to-run variation at this data size is large. In (c) only 13 of 50 test trajectories had time-varying forcing, yet FNO's error also changed (23.44 → 73.27 px) on the 37 where the change does nothing. So these small runs say more about which model is better within a condition than about differences between conditions.

Conclusion: the full-data material-input experiment shows that giving FNO the material layout greatly reduces its 20-step error. This suggests that missing material information contributes substantially to its failure. Our smaller experiments do not establish how much each factor contributes on its own.

Planned revision: Section 4.2 names all three factors and reports both controls. Section 6 states that a full-data one-factor study is not done.

Resolution: we tested material layout, mushy-band width and boundary forcing separately. No single factor explains the results conclusively. The full-data material-input experiment shows that material information matters. XENO has lower 20-step error than FNO in all four small-data conditions, and also when both get the material layout.

This comment addresses: Weakness W2 and Question Q3.

Regarding W2: "... how to identify problems where XENO is beneficial." and regarding Q3 "Do the results suggest measurable characteristics that could identify other problems likely to benefit from XENO?"

Our observations suggest two checks that can be run before training XENO. They are based on four regimes and are not validated general rules:

  1. One-step vs 20-step error of a plain FNO. On R4 it jumps (below 0.2 px at one step, 30.66 px at 20). On R3 and Burgers it does not.
  2. The perturbation test (Q2). If a plain FNO grows a 1-px interface perturbation, enrichment is likely to help. FNO grows it 19× on R4 (to 26.75 px). In the perturbation test, FNO ends at 2.36 px on R3, with no large growth. Separately, in the R3 accuracy results, XENO is only slightly better than FNO (1.6 vs 1.8 px), while Transolver reaches 1.0 px (Table 1).

Both need only one trained FNO.

Separately, after training, the learned gate of XENO can be inspected. It is not a pre-training test. The gate |g_E| (Appendix N) is 0.001 on R1 (1D Stefan with exact solution), 0.00 on Burgers, 0.68 on R3, 0.38 on Allen–Cahn and 1.94 on R4. It was highest where XENO helped most.

In our four regimes, XENO helped most when a plain FNO's error starts at a sharp moving interface and grows during rollout. Where the interface is wide or the plain model is already stable, the gain was small, or XENO was worse (Burgers: 3.0 vs 0.8 px for FNO).

Planned revision: Section 6 gives these checks as a guide for new problems.

Resolution: two pre-training checks (a plain FNO's one-step vs 20-step error gap, and its growth of a 1-px interface perturbation) and one post-training signal (XENO's gate). Both pre-training checks indicate potential benefit on R4, where the gain is large, and the post-training gate is also largest on R4. On R3, Burgers and Allen–Cahn the signals are weaker and the gain is small or negative. These are preliminary observations from four regimes.

We hope these answers address your questions on the mechanism, the R4 factors and applicability. Thank you again for the constructive suggestions. We are happy to clarify anything further during the discussion.

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

Enabling Neural Operator based models to handle sharp interfaces, spatial heterogeneity and strong discontinuities in the computational domain is an important problem. Recent works by Wang et al (2025), Guida & Roberts (2025), etc have proposed their solutions to this issue. The authors of this submission point out that in such conditions, accuracy at predicting a single step forward may be misleading and NO based models can suffer greatly under autoregressive rollouts. The authors claim that this instability is caused by the sharp interface. In this light, they propose a new NNO based model augmented by an Operator Enrichment Block. Tested on the Stefan benchmark, Allen Cahn and 1D Burgers, their model offer better accuracy over very short rollouts.

Critical strengths

The paper deals with an important problem and one that deserves focus. Neural Operator based models suffer at interfaces, spatial heterogeneity and discontinuities in the domain.

The XENO model is tested at a few different benchmark cases such as the 1D Burgers, the Stefan Benchmark, the Allen Cahn problem. At each it is able to show some improvement over other Neural Operator based models, including the Transolver.

Critical weaknesses
  1. The Lightweight CNN for Interface tracking: The downstream model depends on this CNN to apply the jump conditions or functions at the correct locations. However, using a CNN for such a critical task while computationally inexpensive, is dubious. If the interface is complex, the CNN may predict disjointed boundaries. Interface augmented methods require very accurate domain partitioning. A fragmented boundary causes the downstream model to apply the wrong physics to the wrong spatial regions. We dont even need a complex, disjointed interface for this. Such problems can arise if the data used to train the model is from a CFD simulation where the numerical dissipation is substantial, leading to a diffuse interface.

  2. Rollout Lengths: The authors test on rollout lengths of 20 steps to show that XENO leads to stable predictions. However, this is too short in my opinion. In any engineering problem, rollouts of lengths at least in the 100's, if not in the thousands are the norm and not the exception. In my opinion, a 20 step rollout does not show that XENO gives stability for NO based models on sharp moving interfaces.

Questions for authors

In light of the weaknesses that I have pointed out, my suggestions for the authors:

  1. Could the authors show results for problems with disjoint interfaces?

  2. Could the authors show results for problems with diffuse interfaces?

  3. Could the authors show comparative results for longer rollouts of at least a few hundred steps?

5 comments

This comment addresses: Weakness W1 and Question Q2.

Thank you for your review and for seeing the importance of this problem. Your questions on diffuse interfaces, disjoint interfaces and longer rollouts point to real limits of our evaluation, and we address each below.

Where each point is answered:

  • W1 (CNN, fragmented boundaries, domain partitioning): comment 1/3
  • W1 diffuse part + Q2 (diffuse interfaces): comment 1/3
  • Q1 (disjoint interfaces): comment 2/3
  • W2 + Q3 (longer rollouts): comment 3/3

Regarding: "The downstream model depends on this CNN to apply the jump conditions ... A fragmented boundary causes the downstream model to apply the wrong physics to the wrong spatial regions."

XENO does not split the domain into regions and applies no jump conditions. The CNN outputs a continuous field φ. The enrichment uses smooth sigmoids of φ, and the result is added to the spectral path through a learned gate (Section 3.1). A wrong φ changes the size of an added correction. It does not switch the physics in any region.

Two results agree with this:

  1. Replacing the learned φ with the exact signed distance makes results worse (A1, 19.98 vs 1.00 px).
  2. A2, with no interface field at all, reaches 1.12 px.

We also measured φ directly (500 test frames per regime). Within 3 px of the interface, sign(φ) matches the phase field on 99.8% of pixels on R4 (99.9% on R3). The zero level of φ is within 0.8 px of the f = 0.5 contour (R3: 0.5 px).

XFEM-type methods need an exact partition because they use different equations on each side. XENO keeps XFEM's idea of adding extra functions near the interface, but uses a soft weighting instead of a hard partition.

Planned revision: Section 3.2 states that φ is not a domain partition, no jump condition is applied, and reports the φ accuracy above.

Resolution: XENO does not partition the domain or apply jump conditions, so a fragmented φ cannot apply the wrong physics to a region. It can only change the size of a gated correction. We also measured φ near the interface: ≥99.8% sign agreement and ≤0.8 px contour distance.

Regarding W1 (diffuse part) and Q2: "Could the authors show results for problems with diffuse interfaces?"

We measured interface width as the number of pixels with 0.05 < f < 0.95 per interface pixel (median, 50 test trajectories × 10 frames). R3 is 1.2 px, R4 is 2.0 px, so both Stefan sets are sharp. "Smooth" in the paper refers to the temperature field of R3 (99th-percentile |∇T| 0.17 vs 2.6 on R4), not to a wide interface.

Our diffuse-interface data is Allen–Cahn:

  • Main Allen–Cahn set (width 8.4 px): XENO is best on every metric, by a modest margin: 7.29 vs 8.62 px (FNO), fail rate 64% vs 73–84% (Table 1, Appendix M).
  • Width sweep (100 test trajectories, Appendix A): XENO vs FNO at 20 steps: 9.71 vs 8.27 px (width 11.1 px), 9.85 vs 11.01 (12.5 px), 13.96 vs 14.22 (≈24 px).

So on wide interfaces, XENO and FNO are close, with small differences in either direction. The learned gate is also smaller on Allen–Cahn (mean |g_E| 0.38) than on R4 (1.94) (Appendix N). This could possibly be true because a wide, smooth interface is easier for a few Fourier modes, so there is less local error to correct.

Limits: Allen–Cahn has a restoring reaction term that Stefan problems do not (Appendix A), and we have no Stefan data with a numerically smeared interface.

Planned revision: Section 4.1 defines "smooth" via |∇T| and reports widths. Section 6 states that XENO targets sharp interfaces and diffuse Stefan data is untested.

Resolution: Allen–Cahn results at four widths (8–24 px) are given. On diffuse interfaces XENO is not worse than FNO overall, but the gain becomes small.

This comment addresses: Question Q1 (and the disjoint-boundary part of W1).

Regarding Q1: "Could the authors show results for problems with disjoint interfaces?"

Yes, two cases.

  • Allen–Cahn has many disjoint interfaces (median 8–12 blobs at t0, first 50 test trajectories). XENO is best on every reported metric, by a modest margin. Error grows with blob count for both models. When the blob count changes by step 20 (38 of 50 trajectories), XENO and FNO are close (8.90 vs 9.57 px).
  • R4. All test trajectories start with one interface. In 18 of them, new one-cell liquid pockets appear near the inner-box corners while the main front stays connected. We checked each model at the step where the pockets first appear. XENO, U-Net and A2 predict liquid at a new pocket in 0 of 144, 0 of 36 and 1 of 54 model–case pairs. FNO does in 12 of 54, but mostly while also producing extra wrong components elsewhere. On this group every model is near persistence (XENO 27.51, U-Net 26.82, A2 26.81, persistence 26.81 px). Under a max-distance metric, one missed cell costs about 27 px even if the main front is tracked well (figure in Appendix I).

Models without an interface field miss these pockets too, so the miss is not caused by a fragmented φ. Pockets are rare (18 of 200 test trajectories, 2 or 4 cells each). A likely reason all models miss them is that the training data has few such events.

Planned revision: Section 6 lists nucleation of new interfaces as a limitation of all tested models. Appendix I adds the pocket check and figure.

Resolution: with many disjoint interfaces (Allen–Cahn), XENO is best of the models tested. New interfaces that appear during the rollout (R4 pockets) are missed by every model, including those without a CNN, so this is not caused by the CNN.

This comment addresses: Weakness W2 and Question Q3.

Regarding W2: "... a 20 step rollout does not show that XENO gives stability ..." and regarding Q3: "Could the authors show comparative results for longer rollouts of at least a few hundred steps?"

Our R4 trajectories have 101 frames, so 100 steps is the longest horizon with ground truth. Median px, 200 test trajectories:

model2050100moving front, 100
persistence8.4024.8126.3827.34
FNO30.6624.2221.4219.53
Transolver25.84148.85148.85150.26
DeepONet23.97148.85150.26150.26
U-Net0.8620.7021.2115.31
A21.1222.0924.9015.75
XENO1.0020.3822.4511.67
  • Transolver and DeepONet blow up (about 150 px, the size of the domain). FNO is above the 5 px threshold at 20 steps (30.66 px), but its error does not keep growing (21.42 px at 100).
  • All models get worse between steps 20 and 50, when fronts move fastest (persistence goes from 8.4 to 24.8 px).
  • At 100 steps, XENO, U-Net and FNO are all at 21–23 px over all trajectories. On moving fronts XENO is lowest (11.67 px).

Beyond 100 steps we have no ground truth. We still rolled XENO, FNO, U-Net and A2 out to 300 steps on all 200 test trajectories. All predictions stay finite. (The phase field is clipped to [0, 1] in every model, so its range is not evidence.) This only checks that nothing becomes infinite or NaN. It does not measure accuracy, and the temperature drifts (U-Net max 40 at step 1, 328 at step 100, 587 at step 300; XENO 39 → 79).

A few-hundred-step accuracy comparison needs new, longer simulations (the R4 generator saves 101 frames).

Planned revision: the title and abstract drop the unqualified "stable rollouts". Section 4.2 limits the stability claim to 100 steps. Appendix E adds the per-step table and the 300-step check with its limits.

Resolution: results now cover 20, 50 and 100 steps, the full ground-truth length, plus a 300-step finiteness check. Transolver and DeepONet reach errors close to the domain size. FNO exceeds the 5 px threshold at 20 steps. XENO has the lowest moving-front error in this table at 100 steps. We limit our stability claim to the 100-step horizon.

We hope these answers address your concerns about the CNN, interface types and rollout length. Thank you again for your review. We are happy to clarify anything further during the discussion.

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

This paper introduces XENO, a neural operator for long-time prediction of moving solid-liquid interfaces in phase-change (Stefan) problems. XENO adds a branch to each FNO layer that uses a learned field marking the distance to the interface, so that the model can correct its output near the interface. On the hardest dataset (R4), it reduces the median interface error after 20 steps from 32.1 px for FNO to 1.09 px, but it is worse than the best baseline on R3 and Burgers.

Critical strengths

The motivation is clear and well supported: a model that is accurate for one step can still fail when it is applied repeatedly, and this is shown for eight architectures.

Critical weaknesses

The performance gains mostly come from cases where the interface does not move. On the 100 test cases with a moving interface, the median error is 12.4 px for XENO and 25.0 px for FNO, which is about a 2x improvement and still above the paper’s own 5 px failure threshold (Table 5).

It is unclear why the method works. Replacing the learned field with the exact distance to the interface does not work (27.8 px), while a version with no interface information reaches 2.8 px and was not retrained to check this result.

The comparison with baselines is not fully fair. XENO selects its checkpoint from the whole training run using rollout error on validation data, while most baselines have only two saved checkpoints and a single random seed.

The evidence is narrow and the novelty is limited. The method clearly helps on only one dataset created by the authors, and using an interface field as an input to neural operators has been studied before, for example in XI-DeepONet, φ-DeepONet and IANO.

Questions for authors

Is the material layout given to the model on R4? If all models receive it, does FNO still fail?

What is the result of the A2 model, which has no interface information, when it is retrained from scratch under the same setting as XENO?

What are the errors of all baselines, including U-Net and pushforward-FNO, on moving and non-moving interfaces separately?

How does a model that combines U-Net and FNO, with a similar number of parameters and the same training, perform on R4?

6 comments

This comment addresses: Weakness W1.

Thank you for the detailed review. Your points on fairness and on the moving-interface cases were right to raise, and checking them changed several of our numbers. We answer each point below.

Where each point is answered:

  • W1 (gains mostly on non-moving interfaces): comment 1/4
  • W2 + Q2 (why the method works; A2 retrained): comment 2/4
  • W3 (fairness of the comparison): comment 3/4
  • W4 (narrow evidence, novelty vs XI-DeepONet, φ-DeepONet, IANO): comment 3/4
  • Q1 (material layout): comment 4/4
  • Q3 (all baselines on moving vs non-moving): comment 4/4
  • Q4 (U-Net + FNO model): a separate comment

Regarding: "The performance gains mostly come from cases where the interface does not move ... 12.4 px for XENO and 25.0 px for FNO."

The 100 "moving" cases in Table 5 contain two different groups:

  • 18 cases: pocket nucleation. The main front does not move. Instead, 2 or 4 one-cell liquid pockets appear near the inner-box corners (steps 7–20). Hausdorff distance takes the farthest point, so one missed cell gives about 27 px for any model. Persistence: 26.81 px, XENO: 27.51.
  • 82 cases: the front itself moves. Median px:
model20 steps100 steps
persistence27.2727.34
FNO20.3819.53
No-OEB18.2424.03
A22.0915.75
U-Net1.1615.31
XENO5.9111.67

Although U-Net and A2 are better at 20 steps and XENO is still above failure threshold, but on moving fronts, XENO is 3.4× better than FNO at 20 steps and is also the lowest in this table at 100 steps.

On non-moving fronts, the model only has to keep the interface where it is. Persistence gets 0 px. Yet FNO, WNO, CNO and Transolver reach 23–52 px there, so this is where their rollout divergence shows most clearly. XENO gets 0.55 px.

Measured: from step 20 to 100 on moving fronts, XENO's error grows from 5.91 to 11.67 px, U-Net from 1.16 to 15.31, A2 from 2.09 to 15.75. This is one training run per model.

Planned revision: Table 5 (Section 4.2) reports non-moving, moving and pocket cases separately at 20 and 100 steps.

Resolution: On non-moving fronts XENO removes the divergence of global operators (0.55 vs 39.07 px for FNO). On moving fronts XENO beats FNO at 20 steps and is lowest in the table at 100 steps. The 12.4 vs 25.0 px figure included the 18 pocket cases, where every model is near 27 px.

This comment addresses: Weakness W2 and Question Q2.

Regarding: "It is unclear why the method works ..." and "What is the result of the A2 model ... when it is retrained from scratch under the same setting as XENO?"

A2 was trained from scratch with the same trainer, data, epochs and noise as XENO. Only the enrichment branch differs. The 2.8 px in the paper came from evaluating A2 with CCP on. CCP is never on in training, for any model. We will remove the footnote calling this a "train/inference mismatch".

With the common rule (median px):

20 steps100 stepsmoving, 20moving, 100
No-OEB20.7227.1118.2424.03
A21.1224.902.0915.75
XENO1.0022.455.9111.67

From this, we show that A2 is not "XENO without OEB". It keeps the OEB structure: a per-pixel branch added to each spectral layer through the same zero-initialised gate. It only drops the interface field. No-OEB removes the branch completely and diverges. So the gated local branch is important for 20-step accuracy, and the interface field is not needed for it. In this comparison, XENO has lower 100-step error than A2 on moving fronts (11.67 vs 15.75 px). This result comes from one run per model.

Two measurements support this. In the perturbation test (reference: reviewer YqJZ-Q2), a 1-px interface error grows to 26.75 px under FNO but stays at 1.44 px under XENO. FNO's high-frequency error share also rises during rollout on R4 (0.05 → 0.43), while XENO's stays flat (Appendix A). Both point to errors that start at the interface and are not corrected by the spectral layers alone.

About A1 (exact distance instead of the learned field, 19.98 px under the common rule): one possible reason is that the exact distance is computed from the model's own predicted phase field, so it carries that prediction's errors into the enrichment. The learned field is trained through the rollout loss and does not have to be a distance.

Planned revision: Section 4.3 and Appendix L state that the gated local branch gives stability and the interface field is not needed for 20-step stability. The contribution list describes OEB this way. The footnote is removed.

Resolution: A2 retrained from scratch gives 1.12 px, not 2.8 px. The comparison with No-OEB (20.72 vs 1.12 px) shows that the gated local branch is important for 20-step accuracy. In our runs, XENO has lower 100-step error than A2 on moving fronts (11.67 vs 15.75 px).

This comment addresses: Weaknesses W3 and W4.

Regarding W3: "The comparison with baselines is not fully fair ..."

We agree that checkpoint selection was not consistent, and we have fixed it. Every model, XENO included, now chooses between the same two checkpoints (best and last) with the same rule: the lower median 20-step rollout error on validation data.

This changed two baselines. U-Net improved from 9.4 to 0.86 px, and F-FNO from 13.5 to 10.23 px. XENO is at 1.00 px under the same rule. The main result did not change: FNO, WNO, CNO, Transolver and DeepONet still exceed 5 px on 84–100% of test trajectories at 20 steps. Adding OEB to FNO, with the same training, lowers the median 20-step error from 20.72 px (No-OEB) to 1.00 px (XENO).

Planned revision: Table 2 and Appendix Q (protocol).

Resolution: all models now use the same checkpoint rule, so the comparison is more consistent. The main comparison holds under it.

Regarding W4: "The evidence is narrow and the novelty is limited ..."

Novelty. XI-DeepONet, φ-DeepONet and IANO all receive the interface from outside the model:

  • XI-DeepONet takes a given level set and learns steady elliptic solution maps.
  • φ-DeepONet takes a given one-hot domain decomposition.
  • IANO takes an interface field I(t) from the solver or from measurements.

XENO does not receive the interface. It infers its interface field from the current state (T, f) at every rollout step, so it needs no interface data at test time. This matters for rollouts: after the first step, there is no external interface to give the model. XENO also differs in its goal (stable autoregressive rollouts, not one-shot solution maps) and in its form (a gated add-on to an existing spectral model, not a new branch–trunk architecture).

Scope: We agree that the clear gain is on one dataset, R4, which is where the baselines diverge. On R3 and Burgers the baselines do not diverge, and XENO gives little or no gain: on R3 it is 1.6 vs 1.8 px for FNO, with Transolver better at 1.0; on Burgers it is worse than FNO, 3.0 vs 0.8 px. On Allen–Cahn, a second PDE family, XENO is best by a small margin (7.3 vs 8.6 px). We will state this scope in Section 6: XENO helps where a spectral baseline diverges during rollout, and does not give a general gain.

On the datasets being our own: all four are generated by us, and the data-generation scripts are released with the code (Reproducibility statement, Appendix on data generation), so every dataset can be regenerated and checked.

Section 5 already cites all three methods and notes that IANO does not predict its own field. Planned revision: the same paragraph also states that XI-DeepONet and φ-DeepONet take the interface as a given input, and Section 6 states the scope above.

Resolution: none of the three cited methods infers its interface during the rollout; XENO does. On evidence, we agree the clear gain is on R4, and we now state this scope plainly instead of claiming a broad improvement.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.