acceptodds
Under review as a conference paper at ICLR 2027

ONCE: Optimized Neural Domain Decomposition for Scalable PDE Solving

Abstract

Neural surrogates enable efficient PDE solving for scientific discovery and engineering design, yet generalizing to larger domains and unseen geometries remains challenging due to the limited domain and geometric distributions observed during training. Domain decomposition methods (DDMs) provide a promising route to overcome this bottleneck by iteratively coupling the locally trained solver across subdomains. However, the local solver is typically trained under reference states and repeatedly evaluated on states generated by its own predictions. This training-inference mismatch can compromise global accuracy even when the iteration converges. To address these challenges, we present ONCE, a novel neural DDM framework for scalable PDE solving that unites optimized Schwarz methods (OSMs) with Trajectory-aligned Tuning (TAT). Specifically, TAT collects intermediate interface states from the pre-trained model rollouts on the original training cases and fine-tunes the local solver against the corresponding reference subdomain solutions, while retaining reference-state supervision to preserve consistency. This procedure requires neither additional numerical PDE solves nor backpropagation through the iterations. Our theoretical analysis characterizes fixed-point convergence of neural DDMs and establishes interface error bounds that account for trajectory supervision and distribution shift. Experimentally, ONCE consistently outperforms existing methods across benchmarks spanning varying domain scales and unseen geometries, demonstrating its superior accuracy and scalability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural PDE solver, Domain decomposition methods, Scalable PDE solving

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_G5LFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses neural domain decomposition methods, where a learned local solver is reused across subdomains inside a Schwarz iteration to reach larger domains and unseen geometries. It points out a training-inference mismatch: local solvers are trained on reference interface states but evaluated on interface states produced by their own earlier predictions.

ONCE combines: (i) a Transolver local operator pre-trained on subdomain restrictions of existing global solutions. (ii) restricted additive Schwarz with Robin transmission conditions (optimized Schwarz). (iii) Trajectory-aligned Tuning: the pre-trained model is rolled out on the training cases, intermediate interface states are collected and paired with the reference subdomain solutions, and the model is fine-tuned on them.

Critical strengths
  1. The approach is simple and cheap, and the results look promising. The observation that neural DDMs converge to a biased fixed point because of exposure bias is clearly articulated. TAT needs no extra numerical solves and no back-propagation through iterations. The controlled comparisons support the proposed approach, and the 19.7%/15.9% large-domain improvements from TAT are meaningful.

  2. The evaluation is comprehensive. HeatSink-3D and UrbanRANS-2D go beyond the usual elliptic problems. Using Transolver as a mesh-agnostic local operator with METIS partitions of irregular domains makes the framework general.

Critical weaknesses

W1. The scalability claim is not substantiated in absolute terms. On the 4× Large domains, ONCE still has 26–41% relative L2 error. Moreover, the largest domains are only 128×128, ONCE needs up to 200 Schwarz iterations, and no benchmark compares time or accuracy with a numerical solver, which solves such problems exactly in milliseconds.

W2. The novelty is limited. It appears that TAT is essentially a single round of DAgger-style relabeling applied to the Schwarz iteration.

Questions for authors

Q1. For HeatSink-3D and UrbanRANS-2D, what are the wall-clock time and accuracy of the reference numerical solver (FV/FEM direct solve, simpleFoam) on the same hardware class? In what regime of domain size or geometry complexity does ONCE become cheaper at comparable accuracy?

Q2. Can the error floor on Large domains be reduced by repeating TAT (DAgger rounds using the tuned model's own trajectories), or by adding a coarse-space correction as in two-level Schwarz?

4 comments
AuthorsOpenReview ↗

Response to Reviewer G5LF (Part 1)

We thank the reviewer for focusing on absolute accuracy, computational cost, and the relationship to prior learning strategies. We address these points below.

W1: The scalability claim is not substantiated in absolute terms. On the 4× Large domains, ONCE still has 26–41% relative L2 error. Moreover, the largest domains are only 128×128, ONCE needs up to 200 Schwarz iterations, and no benchmark compares time or accuracy with a numerical solver, which solves such problems exactly in milliseconds.

(1) The claim of scalability.
Thank you for this thoughtful critique. Our primary goal in the scaling experiments is to compare ONCE against global baseline under the exact same limited scale training data, whereas ONCE substantially reduces error across larger domains. At the same time, we acknowledge that noticeable error still accumulates at large scales. This is primarily due to the fundemental limitation within the one-level Schwarz method (Mathew, 2008), where the information exchanges cannot efficiently propagate low-frequency global modes as the number of subdomains grows. As detailed in our response to Q2, incorporating a coarse-space correction in two-level DDM may solve this problem. This can indeed serve as an important extension based on ONCE. We refer the reviewer to the below response in Q2 for details, and we have also acknowledged this in the limitation section.

Reference: Tarek Mathew. Domain Decomposition Methods for the Numerical Solution of Partial Differential Equations, volume 61 of Lecture Notes in Computational Science and Engineering. Springer Berlin Heidelberg, Berlin, Heidelberg, 2008. ISBN 978-3-540-77205-7 978-3-540-77209-5. doi: 10.1007/978-3-540-77209-5.

(2) Run time comparison with numerical solvers.
Following your suggestion, we added per-case timings for Poisson, HeatSink-3D, and UrbanRANS-2D in Appendix E.1, Table 5. The first row gives the mean time to generate a numerical reference solution. The second gives the ONCE speedup, computed as the reference generation time divided by the ONCE inference time.

MethodPoisson In. / Med. / LargeHeatSink-3D In. / OODUrbanRANS-2D In. / Med. / Large
Numerical solver (ms)9.95 / 43.38 / 164.8628.83 / 437.5819,916 / 139,897 / 294,409
ONCE (speedup)0.032× / 0.009× / 0.010×0.011× / 0.019×45.9× / 21.1× / 4.58×

The sparse direct solver is faster than ONCE across all Poisson and HeatSink-3D test sets—and notably, in these cases it is even faster than a single-pass global NO on Poisson and comparable in speed on HeatSink-3D. However, on UrbanRANS-2D, generating a steady-state RANS reference with OpenFOAM's simpleFoam takes approximately 20–294 s per case. ONCE achieves speedups of 45.9×, 21.1×, and 4.58× on the in-scale, medium, and large meshes, respectively. This also motivates ONCE to further extend to larger three-dimensional complex physical simulations.


W2: The novelty is limited. It appears that TAT is essentially a single round of DAgger-style relabeling applied to the Schwarz iteration.

We agree with the reviewer that TAT is related to DAgger-style supervision on learner-visited states. However, we respectfully point out that the motivation of our study comes from the core limitation we obsereved from neural DDMs. Also, the standard DAgger requires querying an interactive expert oracle at every visited state which is equivalent to involving an numerical solver on the fly for every perturbed interface condition, which is mostly expensive. On contrast, TAT requires zero additional PDE solves, directly reusing pre-computed reference solutions already in the training set to pull rollout trajectories back toward the true fixed-point.

AuthorsOpenReview ↗

Response to Reviewer G5LF (Part 2)

Q1: For HeatSink-3D and UrbanRANS-2D, what are the wall-clock time and accuracy of the reference numerical solver (FV/FEM direct solve, simpleFoam) on the same hardware class? In what regime of domain size or geometry complexity does ONCE become cheaper at comparable accuracy?

Please see details above in our reply in W1.(2).


Q2: Can the error floor on Large domains be reduced by repeating TAT (DAgger rounds using the tuned model's own trajectories), or by adding a coarse-space correction as in two-level Schwarz?

Thank you for this very insightful suggestions. Following your suggestions, we tested both extensions on Large Poisson.

ComparisonMean relative L2L^2 errorError reductionContraction factor ρ^\hat\rhoMean Schwarz sweeps
One TAT round → two TAT rounds0.2619 → 0.25233.7%0.940 → 0.937142.8 → 141.3
ONCE → ONCE with coarse correction0.2619 → 0.075971.0%0.940 → 0.706142.8 → 22.8

(1) Repeated TAT. We collected new trajectories with the one-round TAT checkpoint and fine-tuned it for a second round. The error shows slightly reduction reflects the combined effect of the new trajectories and additional training.

(2) Coarse-space correction. We applied a coarse-space correction after each learned Schwarz iteration. On large Poisson domains, the correction reduced the overall relative L2L^2 error by 71%. This result demonstrate an important extension of ONCE.


We appreciate your insightful questions and suggestions, which have helped us more carefully assess ONCE’s practical computational value and investigate coarse-space correction as an extension for improving large-domain performance.

Thank you for the detailed response and additional experiments. Some of my concerns have been addressed, while some remain. So I'll maintain my original score (weak acceptance).

For scalability/cost, the added timings demonstrate speedups on UrbanRANS, but whether this is established upon comparable accuracy is still unclear. For Q2, the coarse-space correction results are encouraging and address my suggestion. Repeated tat also offers a modest improvement. For novelty, reusing existing reference solutions without additional PDE solves is a useful design choice, but I feel like there is still conceptual similarity to DAgger-style training.

Thank you very much for your timely follow-up and for acknowledging the additional experiments. We are encouraged by your positive feedback on our work.

On the scalability, there is indeed a practical trade-off for neural methods: while classical numerical solvers already excel in both speed and accuracy on simple problems, for complex simulations (e.g., intricate wind flow across dense building layouts), neural methods offer promising speedups yet SOTA global methods still fall short at extrapolation to large scales, which is the ONCE trying to tackle. In our experiments, we show that the error is reduced by ONCE, though it is still hard to achieve comparable accuracy with numerical solvers. We acknowlege this and will work more to further improve on its speed and accuracy as mentioned in our limitation section.

Thank you again for reviewing our work and we are truly grateful for your support.

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

The paper proposes ONCE, a neural domain decomposition method for scaling PDE surrogates to larger domains and new geometries. A local neural operator (Transolver) is trained on subdomain pieces of global reference solutions and reused at inference inside a Schwarz iteration with Robin interface conditions. The main claimed contribution is Trajectory-aligned Tuning (TAT). The pre-trained local solver is run on the training cases, the intermediate interface states it visits are recorded, and the solver is fine-tuned on those states with the reference subdomain solutions as targets, plus the original loss as a regularizer. The authors argue this fixes the mismatch between the clean interface data seen in training and the imperfect interface data seen during iterative inference. Theory gives a fixed-point bound with a bias term. Experiments on Poisson, Darcy, a 3D heat sink benchmark and a 2D urban wind benchmark report lower errors than global neural operators and an adapted SNI baseline, especially on larger domains.

Critical strengths

Exposure bias in iterative neural solvers is a real problem, and the paper is right that a neural DDM can converge reliably and still converge to the wrong answer. TAT is simple, cheap and works with any architecture. Training the local solver on pieces of global solutions avoids the ill-posed synthetic local problems that SNI needs, which matters for problems like RANS. The benchmarks are varied, the ablations are broad, and the runtime and limitations sections are well established.

Critical weaknesses
  1. It is not shown that TAT is the source of the improvement, or that there is a meaningful improvement at all. The paper is framed around TAT, but the headline gains over SNI come from a bundle of changes: a different local architecture, RAS instead of additive Schwarz, Robin instead of Dirichlet, and different training data. The paper’s own Table 2 shows that most of the in-scale gain over SNI is already there before TAT is applied. The effect of TAT itself is a few percent on in-scale and medium domains and roughly 15 to 20 percent on the largest domain. No seeds or error bars are reported, and the error bars that do appear in Figure 4 overlap, so it is unclear whether even the largest gain is real. The one control that would settle this, continuing to train on reference data for the same number of epochs, is shown only as one curve on one case in Figure 6a and never in a table. Meanwhile, plain data augmentation in the appendix gives bigger gains than TAT. The paper needs a clean ablation ladder from SNI to ONCE with multiple seeds and a reference-only tuning control on every benchmark. Without that, the reader cannot tell whether this is a paper about TAT or a paper about a better-engineered neural DDM.

  2. Even if TAT helps, the paper does not explain why, and its own story does not hold together. The motivation is distribution shift: the solver sees imperfect interface data at inference, so it should be trained on such data. But TAT does not teach the solver the correct local answer for imperfect interface data. It teaches the solver to output the reference solution regardless of what interface data it receives. That is a learned correction, not a fix for distribution shift, and the two are different things with different failure modes. The theory does not resolve this. The bias bound in Theorem 1 depends on the reference-state error and the contraction rate. TAT does not reduce the first and the paper never measures the second, so the theory does not predict that TAT should help. The theorem also assumes the iteration contracts, while the appendix says the large-domain runs stall at an error floor and hit the iteration cap.

  3. The large-domain results remain poor, and the likely reasons are ones TAT cannot fix. On the largest domains the error is still 26 to 41 percent, which is hard to align with the scalability claims. The method has no coarse-level correction, so the error floor grows with the number of subdomains, exactly as the appendix reports. Pre-training also uses small domains cut into a handful of pieces, so nearly every training subdomain touches the physical boundary, whereas large domains are mostly interior subdomains the solver has never seen. TAT trajectories are collected only on small domains, so it cannot help here either. The paper treats a symptom while the structural problem is left in place.

  4. The comparisons and the practical case are weak. SNI is a modified re-implementation trained on different data from the original. The global operators are not meaningful large-domain baselines, since they are asked to extrapolate far outside their training coordinates. Several tuning and DDM hyperparameters are chosen on the medium and large test sets, with no validation split at those scales. ONCE is about two hundred times slower than a global neural operator on large domains, and there is no timing comparison with the classical solvers that generated the data, which for these problems are likely both faster and exact. The urban wind benchmark transmits only velocity across interfaces, so the local problems are not well posed and the DDM framing is heuristic there. The external AIJ case is introduced but never reported.

Questions for authors
  1. What is the error of ONCE without TAT, and of reference-only tuning with matched epochs, on every benchmark and scale, with seeds?

  2. How much of the gap to SNI comes from each of the architecture, assembly, transmission and data changes, versus TAT?

  3. How does TAT compare to tuning on the true local solutions for the same visited interface states?

  4. What are the contraction rates before and after TAT, and do the large-domain runs actually converge?

  5. How does the error split between interior and boundary-touching subdomains on large domains, and have you tried a coarse correction?

  6. How do the accuracy and runtime compare with the numerical solvers used to generate the data?

7 comments
AuthorsOpenReview ↗

Response to Reviewer rpHn (Part 1)

We would like to thank the reviewer for the detailed technical critique and insightful questions, and we really learn from your comments and rethink a lot to improve our manuscript during this rebuttal process. We try our best to address all of your concerns including the attribution of gains, the interpretation of TAT, and the evaluation protocol below.

W1: It is not shown that TAT is the source of the improvement, or that there is a meaningful improvement at all. The paper is framed around TAT, but the headline gains over SNI come from a bundle of changes: a different local architecture, RAS instead of additive Schwarz, Robin instead of Dirichlet, and different training data. The paper’s own Table 2 shows that most of the in-scale gain over SNI is already there before TAT is applied. The effect of TAT itself is a few percent on in-scale and medium domains and roughly 15 to 20 percent on the largest domain. No seeds or error bars are reported, and the error bars that do appear in Figure 4 overlap, so it is unclear whether even the largest gain is real. The one control that would settle this, continuing to train on reference data for the same number of epochs, is shown only as one curve on one case in Figure 6a and never in a table. Meanwhile, plain data augmentation in the appendix gives bigger gains than TAT. The paper needs a clean ablation ladder from SNI to ONCE with multiple seeds and a reference-only tuning control on every benchmark. Without that, the reader cannot tell whether this is a paper about TAT or a paper about a better-engineered neural DDM.

(1) Scope and design of ONCE, and the direct comparsion with SNI.
We agree that the overall gains of ONCE and the incremental contribution of TAT should be evaluated separately. ONCE is designed for reusing existing global simulations by extracting subdomain training pairs, with TAT serving as a tuning component towards the nature of neural DDM as a fixed-point iteration. In contrast, the original SNI assumes that local simulations can be independently generated under synthetic boundary conditions. Our setting avoids solving potentially ill-posed or numerically unstable synthetic local problems(such as in UrbanRANS-2D). We therefore train the adapted SNI on the same extracted data. Because the underlying problem assumption and data sampling pipeline are fundamentally different, in our original submission we only investigated the individual components of ONCE in the Appendix to quantify their respective performance gains, including the design of TAT, assembly method, local operator, and data augmentation.

We now acknowledge that this could lead to confusion. To directly address your concern, we reports the following cumulative comparison on Poisson starting from GNOT with additive Schwarz (AS) and Dirichlet transmission:

ConfigurationIn-scaleMediumLarge
SNI (AS, k=200k=200)0.0170.1530.802
SNI (AS, plateau)0.0170.0490.333
+ RAS (k=200k=200)0.0160.0490.333
+ Transolver0.0100.0380.335
+ Robin transmission0.0110.0260.326
+ TAT (ONCE)0.0090.0240.262

Overall, ONCE achieves both substantially faster convergence and lower prediction errors across all domain scales compared with SNI. Specifically, under the same iteration budget SNI with classical AS is not yet converge with high errors, whereas RAS reaches comparable plateau accuracy within fewer iterations. Along this cumulative sequence, replacing GNOT with Transolver reduces the in-scale and medium-domain errors, with a small increase in the large-domain error. Switching to Robin transmission then reduces the medium- and large-domain errors, while the in-scale error increases from 0.010 to 0.011. Finally, TAT consistently improves accuracy across scales, including an substaintial reduction on the large domain. We have also updated relevant presentation in our revised manuscript to improve its clarity.

AuthorsOpenReview ↗

Response to Reviewer rpHn (Part 2)

Continued reply on W1
(2) Multi-seed evaluation of TAT across benchmarks.
We compare the pre-trained model (Pre-train), reference-only fine-tuning (Ref-only), and TAT. Ref-only continues training on reference data for the same tuning costs as TAT. Each entry gives the mean ± standard deviation across three runs. For each run, we first average the relative L2L^2 error over the test cases at the indicated scale. We didn't run the experiments on UrbanRANS-2D due to the limited time available during the rebuttal period. We plan to extend the same comparison to UrbanRANS-2D in a subsequent revision.

MethodPoisson In. / Med. / LargeDarcy In. / Med. / LargeHeatSink-3D In. / OOD
Pre-train0.0109 ± 0.0001 / 0.02455 ± 0.0033 / 0.3073 ± 0.02390.0262 ± 0.0001 / 0.1423 ± 0.0003 / 0.4618 ± 0.00970.039 ± 0.002 / 0.056 ± 0.004
Ref-only0.0101 ± 0.0001 / 0.02418 ± 0.0035 / 0.3045 ± 0.02370.0262 ± 0.0001 / 0.1405 ± 0.0003 / 0.4613 ± 0.00980.038 ± 0.002 / 0.055 ± 0.004
TAT0.0090 ± 0.0001 / 0.0209 ± 0.0026 / 0.2511 ± 0.01780.0210 ± 0.0002 / 0.1326 ± 0.0007 / 0.4077 ± 0.00940.036 ± 0.001 / 0.050 ± 0.004
Error reduction vs. Pre-train17.4% / 14.8% / 18.3%19.9% / 6.8% / 11.7%7.7% / 10.2%

Across all three benchmarks, matched Ref-only tuning mostly yields minimal change over Pre-train, confirming that TAT's gains come from trajectory supervision rather than extra training epochs.

(3) On the data augmentation.
Thank you for mention this. As shown in Appendix E.4, Fig. 9 (also Appendix E.4 in the original submission), we want to point out that D4D_4 augmentation is applied on top of ONCE (which already includes TAT), providing complementary gains on Poisson and Darcy. However, we excluded it from default configurations for two reasons: (1) Fair comparison with the baselines: We report all main results in Table 1 without augmentation, maintaining the same setting for ONCE and other baselines. (2) Task-dependent effectiveness: As discussed in Appendix E.4, the same D₄ augmentation improves performance on Poisson and Darcy but provides no benefit on UrbanRANS-2D under the due to the significant different problem setup.


W2: Even if TAT helps, the paper does not explain why, and its own story does not hold together. The motivation is distribution shift: the solver sees imperfect interface data at inference, so it should be trained on such data. But TAT does not teach the solver the correct local answer for imperfect interface data. It teaches the solver to output the reference solution regardless of what interface data it receives. That is a learned correction, not a fix for distribution shift, and the two are different things with different failure modes. The theory does not resolve this. The bias bound in Theorem 1 depends on the reference-state error and the contraction rate. TAT does not reduce the first and the paper never measures the second, so the theory does not predict that TAT should help. The theorem also assumes the iteration contracts, while the appendix says the large-domain runs stall at an error floor and hit the iteration cap.

Thank you for the careful reading and insightful comments. We particularly appreciate your close examination of the relationship between TAT’s motivation, its supervision targets, and our theoretical analysis. Your comments highlight a distinction that is central to our formulation and that we should have explained more clearly. We clarify these connections and the scope of our theoretical claims below.

(1) TAT’s supervision target and its motivation.
We acknowledge that our claim of “distribution shift” was too broad and could lead to misunderstanding. The reviewer is correct that TAT does not teach the local solver to correctly solve the local BVP that haven't appeared in its training distribution (i.e., train a universally accurate local solver). Indeed, TAT is motivated by the nature of training-inference mismatch: the local solver is trained on reference interface states (true local BC and true solution) but receives model-generated intermediate states during iterative inference. Under the fixed-point iteration of neural DDMs, it requires an update operator that steers intermediate states toward the reference state λ∗\lambda^\ast while also preserving consistency at λ∗\lambda^\ast. You are also correct that pairing λ^i(k)\widehat{\lambda}_i^{(k)} with ui∗u_i^\ast therefore acts as a learned correction along the iteration trajectory. Compared with using the exact solution as step for supervision, using ui∗u_i^\ast could potentially requires zero additional PDE solves, and remains applicable when standalone local BVPs are not well posed.

AuthorsOpenReview ↗

Response to Reviewer rpHn (Part 3)

Continued reply on W2
(2) Our theoretical analysis.
In our analysis, Theorem 1 serves a diagnostic role for neural DDMs: it explains why a contractive learned iteration trained only on reference states can converge reliably yet still settle at a biased fixed point. We agree that the theorem alone does not by itself establish an improvement from TAT. Instead, its additional objective targets the errors of updates from intermediate interface states. Under the stated transmission assumptions, Proposition 1 and Eq. (15) connect the local trajectory loss to an upper bound on these next-step global interface errors. Reducing LTAT\mathcal L_{\mathrm{TAT}} tightens the bound in Eq. (15); the bound along the tuned trajectory also becomes tighter when the reduction in the trajectory-loss term is not offset by an increase in BθB_\theta. Together, these provide a theoretical rationale for TAT’s design, and we note that Propositions 1 and 2 do not assume contractivity.

In the revised manuscript, we have clarified TAT’s role as a learned correction within the coupled iteration and stated more precisely the scope of the training–inference mismatch it aims to mitigate.


W3: The large-domain results remain poor, and the likely reasons are ones TAT cannot fix. On the largest domains the error is still 26 to 41 percent, which is hard to align with the scalability claims. The method has no coarse-level correction, so the error floor grows with the number of subdomains, exactly as the appendix reports. Pre-training also uses small domains cut into a handful of pieces, so nearly every training subdomain touches the physical boundary, whereas large domains are mostly interior subdomains the solver has never seen. TAT trajectories are collected only on small domains, so it cannot help here either. The paper treats a symptom while the structural problem is left in place.

(1) On the large-domain accuracy and the scalability claims.
Regarding the scalability, as TAT only aims to improve the accuracy of the learned iterative updates (as we have discussed in W2.(2)), ONCE’s ability to perform better on larger domains is fully enabled by the DDMs itself, which is one-level DDMs as detailed in our methodology. Based on our experiments, we found that employing one-level DDMs does provide better scalability compared with global baselines under same training data. However, it does not establish accuracy independent of domain size (larger number of subdomains), which is exactly inline with the theory of classical DDMs (Gander and Vanzan, 2020), as this can be further addressed by coarse correction (i.e, two-level DDMs). We highly agree with you on this point.

Reference: Gander, M. J., & Vanzan, T. (2020). Multilevel optimized Schwarz methods. SIAM Journal on Scientific Computing, 42(5), A3170–A3201.

(2) ONCE with coarse correction.
To further enhance the scalablity of ONCE, we have added experiment to incoraporate coarse correction within ONCE. Specifically, we add bilinear coarse correction (H=32hH=32h, where hh is the size of the original grid) on the large Poisson case. We show that the error is significanly reduced. This demonstrates that ONCE can incorporate coarse-level information to substantially improve large-domain accuracy.

Poisson LargeRelative L2L^2 errorError reduction
ONCE0.2619—
ONCE + coarse correction0.075971.0%
Reviewer_6WPPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduced ONCE, a method dedicated to address complex geometries with neural solvers. By combining existing NO architecture with a Scwharz method, and a fine tuning stage on the interface trajectories, ONCE allows the method to handle complex geometries and to reduce the gap between training and inference samples. The paper provides a theoretical analysis and experiment to support the claim. The appendices contains proofs, experiment protocol and additional experiments and ablations.

Critical strengths
  • To the best of my knowledge the method is a new combination of a mathematical tool with current neural methods, and a lot of elements are introduced in the paper.
  • The appendices are complete and provide additional experiments (noise injection, ablations on several components such the model used or the model size, data augmentation...), proofs and details.
  • The datasets used by the authors seem new and introduce new challenges for current models.
  • I appreciated the honest discussion regarding the limitations of the method and proposing future directions to overcome them.
Critical weaknesses
  • One strong weakness for me is a missing part / discussion of crucial references of the literature, and a comparison of the 2 paradigms existing to handle complex geometries (natively in the model or with an external method as proposed here). For instance with a reference to Gino [1] and other mesh agnostic neural operators. Other methods using INRs exists (CORAL [2]), attention mechanisms (AROMA [3]) or continuous convolutions (CALM-PDE [4]). I think a clear and broader discussion on the large literature would help positioning the paper.

  • Moreover, a comparison with more recent neural operator methods would strengthen the paper and highlight the key properties of the method wrt existing mesh independent method. The underlying question behind my comment here is why one should chose ONCE wrt to another one ? Additionnaly, I believe more recent methods handles irregular geometries natively, and could strengthen the experimental section. For instance, why comparing with Transolver instead of the more recent version Transolver ++ (that is mentioned in the related work section) ?

  • I found the paper a bit hard to follow, but as I am not an expert in Schwarz method, I may have missed some points. As a result, I also could not check the theory. I reflect this with a low confidence on my review.

Minor

  • I think more references to the appendix would help. Especially, I almost missed some strong experiments validating a claim regarding the applicability of the method (Appendix E5 for instance).

References

[1] Geometry-Informed Neural Operator for Large-Scale 3D PDEs, Zongyi Li, Nikola Borislavov Kovachki, Chris Choy, Boyi Li, Jean Kossaifi, Shourya Prakash Otta, Mohammad Amin Nabian, Maximilian Stadler, Christian Hundt, Kamyar Azizzadenesheli, Anima Anandkumar. 2023.

[2] Operator Learning with Neural Fields: Tackling PDEs on General Geometries, Louis Serrano, Lise Le Boudec, Armand Kassaï Koupaï, Thomas X Wang, Yuan Yin, Jean-Noël Vittaut, Patrick Gallinari, 2023

[3] AROMA: Preserving Spatial Structure for Latent PDE Modeling with Local Neural Fields, Louis Serrano, Thomas X Wang, Etienne Le Naour, Jean-Noël Vittaut, Patrick Gallinari, 2024

[4] CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-dependent PDEs, Jan Hagnberger, Daniel Musekamp, Mathias Niepert. 2025

Questions for authors
  • How would methods behave without TAT ? The idea here is to isolate the improvement of TAT. In table 1, in my understanding the Transolver line corresponds to the comparison of the method with the TAT method ? I believe an ablation to isolate the contribution of each component would help ?
  • What is the cost of the subdomain decomposition before training ? and the cost of creating the transmission parameters ? Doesn't it induce an over computation wrt to other method to compute in parallel on overlapping subdomains ? (lines 210-213) Can you also study training time of the proposed method wrt to others ?
  • line 138-139, in practice, is the hypothesis often met ? (uniqueness solvability of the local BVP)

Comments on the paper

6 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all the reviewers for their detailed review and feedback of our study. We appreciate the reviewers’ recognition of ONCE as new combination of tools (Reviewer 6WPP), a simple, cheap and flexible method (Reviewers rpHn and G5LF), our evaluation is comprehensive (Reviewers 6WPP and G5LF), and the results look promising (Reviewer G5LF).

Based on the concerns, we briefly summarized the our rebuttal below:

  • Clarity, comprehensiveness, and baseline comparison (Reviewers 6WPP and rpHn): We expanded the discussion of geometry-aware neural operators, added Transolver++ as a baseline for comparison, and reported more quantitative results.
  • Theoretical justification and effectiveness of TAT (Reviewers 6WPP and rpHn): We revisited and clarified the scope and assumptions behind our theory. We also evaluated TAT through a step-by-step ablation from SNI to ONCE and multi-seed comparisons with matched-epoch reference-only tuning.
  • Computational costs (Reviewers 6WPP, rpHn, and G5LF): We added comparison between inference cost of neural methods and numerical solvers.
  • Potential integration with numerical solver for enhancement ONCE (Reviwers rpHn and G5LF): We investigated the potential of further incorporating numerical tools within ONCE, especially the coarse correction as an important extension beyond ONCE.

All changes in the revised manuscript are marked in blue, and we have provided point-by-point responses to all of your comments from each reviewer. Thanks again for your time and dedication, and we are happy to receive your feedback and further discussion with us.

AuthorsOpenReview ↗

Response to Reviewer 6WPP (Part 2)

Q1: How would methods behave without TAT ? The idea here is to isolate the improvement of TAT. In table 1, in my understanding the Transolver line corresponds to the comparison of the method with the TAT method ? I believe an ablation to isolate the contribution of each component would help ?

Thank you for raising this question. In our original submission, Table 1 presents the whole framework of ONCE including TAT. Section 4.3, Table 3 of the revised manuscript (Table 2 in the original submission) reports this ablation. The two settings below differ only in TAT, and we show that TAT improves the performance of ONCE.

SettingPoissonUrbanRANS-2D
ONCE w/o TAT0.0108 / 0.0262 / 0.32620.0833 / 0.2947 / 0.4713
ONCE0.0090 / 0.0238 / 0.26190.0782 / 0.2511 / 0.3964
Error reduction by TAT16.7% / 9.2% / 19.7%6.1% / 14.8% / 15.9%

We also report the multiple-seeds ablation in Appendix E.9, Table 10 in our revised paper, including the reference-only fine-tuning control to better confirm the performance of TAT.


Q2: What is the cost of the subdomain decomposition before training ? and the cost of creating the transmission parameters ? Doesn't it induce an over computation wrt to other method to compute in parallel on overlapping subdomains ? (lines 210-213) Can you also study training time of the proposed method wrt to others ?

(1) On the cost of domain decomposition methods.
Thank you for this careful observation. The decomposition and the transmission parameters do add little cost, and the overlap reduces the total runtime. We decompose each case once before the iteration, which takes 39 ms, 168 ms, and 715 ms per case on in-scale, medium, and large Poisson domains (approximately 12.5%, 3.5%, and 5% of the ONCE runtime). Training uses the same routine, so the decomposition before training costs 39 ms per in-scale case. Creating the transmission parameters itself has nearly no cost, because the Robin parameter γ\gamma is a fixed constant, the only related cost is reading the interface data from the current iterate, which takes 12 ms per iteration on large domains. The overlap does add computation per iteration: with 3 overlap layers, the subdomains together contain 1.42, 1.61, and 1.72 times the cells of the global mesh on in-scale, medium, and large domains. However, we'd like to note that using larger overlap could also reduces the iteration count, consistent with classical Schwarz theory. For example, on Large Poisson (Appendix E.6, Table 8(a)), increasing the overlap from 1 to 3 layers reduces the iterations from 180.5 to 68.0 and the runtime from 15.39 s to 7.02 s. In total, Fig. 7 in Appendix E.1 reports the DDM overhead (decomposition, interface extraction, transfers, and merging) as 20%, 24%, and 28% of the ONCE runtime on in-scale, medium, and large domains. We thank the reviewer again for raising this point.

(2) On the training cost.
We compare the training cost of global neural operator (Transolver) and ONCE (Transolver as local solver) on the Poisson benchmark:

MethodPre-trainingTrajectory collectionTATTotal
Transolver31.9 min––31.9 min
ONCE109.5 min2.0 min21.5 min133.0 min

The results shows that under the training configurations used in our experiments, ONCE requires approximately 4.2 times of the global method. This primarily attributes to the increase of training samples from global to local subdomains.


Q3: line 138-139, in practice, is the hypothesis often met ? (uniqueness solvability of the local BVP)

Indeed, the validity of this assumption depends on the PDE and the local boundary specifications. For example, the assumption is satisfied by the elliptic benchmark problems considered in this study under their prescribed boundary conditions. However, it is not hold for broader indutrail-scale simulations such as the nonlinear, coupled RANS system with turbulence closure. Accordingly, our experiment on UrbanRANS-2D should therefore be understood as an empirical extension beyond the established local well-posedness setting. We will clarify these boundaries and their implications for our theoretical scope in the revised version of our paper.


We thank you again for your constructive feedback, which has helped us better position ONCE within the literature, improve the clarity of its presentation, and strengthen the comparisons with existing methods.

AuthorsOpenReview ↗

We look forward to your response for further discussion

Dear Reviewers,

As the discussion period is approach its end, we first want to thank Reviewer G5LF for maintaining the positive feedback of our work. Meanwhile, we warmly invite Reviewers 6WPP and rpHn to share any further thoughts on our responses and additional experiments. We truly value your feedback and would be very glad to clarify any remaining concerns or questions.

Many thanks,
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.