acceptodds
Under review as a conference paper at ICLR 2027

Learning to decompose PDE domains with geometry-aware local operators and global communication

Abstract

Neural operators provide an efficient approach to repeated partial differential equation (PDE) solving, but practical use on complex geometries and large meshes remains constrained by the challenge of balancing local spatial resolution, global physical coupling, and computational cost. A Geometry-Aware Domain-Decomposition Neural Operator (GDDNO) is introduced to organize local computation and global communication through domain decomposition. Within each subdomain, local interactions are encoded by a shared graph operator. The global information is exchanged between the subdomain tokens through attention, and the resulting context is combined with retained cell features for prediction. Overlap domain assembly and flux consistency constraints coordinate predictions between subdomains. GDDNO is evaluated on five RealPDEBench systems, AirfRANS, and a three-dimensional reacting flow dataset. It obtains the lowest simulation-trained relative error on all five RealPDEBench systems. On full-field AirfRANS, the relative error is 0.0815, compared with 0.3511 for the strongest baseline, and the prediction of the aerodynamic coefficients is also the best. On cavity combustion flow, the relative errors for the predicted fields and combustion efficiency are 0.4148 and 0.2112, respectively, both the lowest among the compared models. These results indicate that domain decomposition can be used as an effective strategy to combine local geometric detail with long-range interactions in neural PDE operators.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operator, AI for PDE, Domain decomposition

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tjx4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GDDNO, a domain-decomposition neural operator for predicting PDE solutions on complex geometries and large meshes. The method partitions the computational domain into overlapping subdomains, employs a shared geometry-aware graph network to capture local interactions, and uses attention over compact subdomain tokens to exchange global information. Local features and global context are combined for field prediction, while overlap averaging and consistency losses coordinate predictions across subdomain boundaries. Experiments on AirfRANS, five RealPDEBench systems, and a three-dimensional reacting-flow dataset demonstrate improvements in prediction accuracy and selected physical quantities over existing neural operator baselines.

Critical strengths
  1. The integration of geometry-aware local graph operators, compact global token communication, and overlapping domain assembly provides a coherent approach to balancing local geometric detail and long-range interactions. Although the individual components are not entirely novel, their combination addresses a practically important challenge in neural PDE operators.
  2. The paper reports improvements across AirfRANS, five RealPDEBench systems, and a challenging 3D reacting-flow dataset, including substantial gains on full-field AirfRANS and improvements in engineering quantities. This breadth is particularly valuable because it suggests that the proposed architecture may generalize beyond a single PDE benchmark, provided the reported results can be verified.
Critical weaknesses
  1. Multiple numerical discrepancies appear across the paper. For example, Tables 6 and 7 report different R² and lift-coefficient errors for the same configuration (P=48, H=2), while Table 10 repeatedly labels the proposed model as "HPNO (ours)" rather than GDDNO. For the Foil task, the simulation-trained relative L2 error is reported as 0.0344 in the main text but 0.0360 in Table 3. These issues raise substantial concerns about experimental result management and reliability. Since the paper's main contribution is largely supported by empirical improvements, the authors should reconcile these discrepancies and provide verifiable experimental records.
  2. The proposed interface consistency loss penalizes differences between neighboring cell states without directly constraining numerical fluxes (Sections 3.2–3.5). Its connection to the claimed flux consistency is therefore unclear, and the penalty may suppress physically valid gradients. Furthermore, the ablation study (Section 4.2) does not isolate the contribution of global token communication or separately evaluate the overlap and interface losses. Consequently, the evidence does not adequately establish the necessity of the key architectural components or support the claimed physical consistency. More rigorous physical validation and targeted ablations are needed.
Questions for authors
  1. Can the authors clarify the inconsistencies in the reported experimental results? Tables 6 and 7 report different R2R^2 and Cl errors for the same configuration (P=48, H=2), while Table 10 repeatedly refers to the proposed model as "HPNO (ours)" rather than GDDNO. How can the authors ensure that the reported results correspond to the correct model and experimental configurations? Please provide corrected results and verifiable supporting evidence.
  2. How much does global token communication independently contribute to the performance gains? The current ablations do not evaluate the model without global attention, nor do they separately examine overlapping subdomains and interface consistency losses. Can the authors provide targeted ablations to establish which components are responsible for the improvements?
  3. What justifies interpreting the proposed interface consistency loss as a flux-consistency constraint? The loss penalizes differences between neighboring cell states rather than discrepancies in numerical fluxes. Since physically valid solutions can exhibit substantial gradients across adjacent cells, could this regularization introduce artificial smoothing? Can the authors justify its physical interpretation and evaluate its impact on conservation errors?
4 comments
AuthorsOpenReview ↗

Response to reviewer tjx4 (W1,Q1, W2, Q2)

W1 Multiple numerical discrepancies appear across the paper. The authors should reconcile these discrepancies and provide verifiable experimental records.

Q1 How can the authors ensure that the reported results correspond to the correct model and experimental configurations? Please provide corrected results and verifiable supporting evidence.

We sincerely apologize for these errors. The discrepancies in Tables 6 and 7 (now Table 7and 8) resulted from a keyboard shortcut failure when copying data from the experimental logs during manuscript preparation, while the inconsistency in the airfoil results was due to a manual writing error. We understand that these mistakes may have affected the reviewer's confidence in our results. To facilitate independent verification, we have uploaded our original experimental logs as supplementary materials. The original data for (P=48, H=2) reported in Tables 6 and 7 can be found in lines 6058–6070 of of-metrics.json, while the airfoil results reported in Table 3 can be verified at line 68 of eval.log.

For the name HPNO, during the early development stage, we named the model "Halo Patch Neural Operator (HPNO)" to reflect our design of introducing overlapping boundaries between subdomains. We renamed the model to GDDNO only one week before the paper submission. Unfortunately, one of our co-authors had not updated their code branch to the new name in time and therefore continued to use the earlier one. The eval.log file also provides evidence of this naming history. We once again apologize for this oversight and have corrected the corresponding errors in the revised manuscript.

Finally, we commit to publicly releasing the code within one month of paper acceptance. Due to the double-blind review requirements, we will provide the repository URL in the camera-ready version if the paper is accepted. The code included in the supplementary materials is already executable and provides the implementation needed to reproduce the reported experiments.

Q2: How much does global token communication independently contribute to the performance gains? The current ablations do not evaluate the model without global attention, nor do they separately examine overlapping subdomains and interface consistency losses. Can the authors provide targeted ablations to establish which components are responsible for the improvements?

W2: The proposed interface consistency loss penalizes differences between neighboring cell states without directly constraining numerical fluxes (Sections 3.2–3.5). Its connection to the claimed flux consistency is therefore unclear

We thank the reviewer for the constructive suggestion. To thoroughly evaluate the contribution of our proposed components, we have conducted comprehensive ablation studies. We compared GDDNO with two variants: (a) a model replacing the cross-subdomain attention with a simple global pooling baseline, and (b) a model without any inter-subdomain communication. As shown in Table below, GDDNO reduces the overall relative L2L_2 error from 0.0996 to 0.0816 and achieves the lowest CpC_p error among all settings. This demonstrates that our proposed communication mechanism is significantly more effective than uniform aggregation or isolated subdomains

communication mechanismRMSERel L2L_2R2R^2CpC_pClC_l
no cross-subdomain communication0.08260.09960.98820.05100.0742
simple global pooling0.08030.09090.99170.04990.0665
GDDNO0.07210.08160.99340.04800.0736

We evaluated the individual and joint impacts of the interface and overlap losses. Incorporating either loss independently improves the prediction accuracy, and their combination further boosts the overall performance. This fully proves the effectiveness and complementarity of both consistency constraints.

LossRMSERel L2L_2R2R_2CpC_pClC_l
LdataL_{\mathrm{data}}0.08070.09140.99170.07030.1106
Ldata+LoverlapL_{\mathrm{data}}+L_{\mathrm{overlap}}0.07500.08330.99060.04920.0932
Ldata+LinterfaceL_{\mathrm{data}}+L_{\mathrm{interface}}0.07830.08860.99340.06960.1236
GDDNO0.07210.08160.99340.04800.0736
AuthorsOpenReview ↗

Response to reviewer tjx4 (Q3, W2)

Q3: What justifies interpreting the proposed interface consistency loss as a flux-consistency constraint? The loss penalizes differences between neighboring cell states rather than discrepancies in numerical fluxes. Since physically valid solutions can exhibit substantial gradients across adjacent cells, could this regularization introduce artificial smoothing? Can the authors justify its physical interpretation and evaluate its impact on conservation errors?

W2: Furthermore, the ablation study (Section 4.2) does not isolate the contribution of global token communication or separately evaluate the overlap and interface losses.

We thank the reviewer for this insightful suggestion. In response, we have supplemented a more detailed continuity analysis of the predicted physical fields. Specifically, we evaluated four quantitative metrics,EnetE_{\mathrm{net}}, EdivE_{\mathrm{div}}, EpatchE_{\mathrm{patch}}, and Eϕ,ΓE_{\phi,\Gamma},whose physical meanings are summarized below:

Metricsexplanation
EnetE_{\mathrm{net}}Net flux imbalance across the entire computational domain
EdivE_{\mathrm{div}}Volume-weighted average of cell divergence
EpatchE_{\mathrm{patch}}Mean net-flow imbalance of subdomains
Eϕ,ΓE_{\phi,\Gamma}Relative error of subdomain interface flux compared to CFD ground truth

Due to space limitations, their exact mathematical definitions and formulations are detailed in the revised manuscript Appendix D.5. The statistical results of these metrics are presented in the table below. As shown, GDDNO’s consistency strategy significantly improves both global and local flow field continuity as well as physical conservation, demonstrating the effectiveness of our two proposed constraints within the domain decomposition framework.

FieldsEnetE_{\mathrm{net}}EdivE_{\mathrm{div}}EpatchE_{\mathrm{patch}}Eϕ,ΓE_{\phi,\Gamma}
CFD6.0463×10−56.0463\times10^{-5}1.3531×10−61.3531\times10^{-6}6.5693×10−66.5693\times10^{-6}0.00000.0000
Ldata\mathcal{L}_{\mathrm{data}}3.0459×10−33.0459\times10^{-3}3.1638×10−43.1638\times10^{-4}5.9878×10−45.9878\times10^{-4}4.7221×10−34.7221\times10^{-3}
Ldata+Loverlap\mathcal{L}_{\mathrm{data}}+\mathcal{L}_{\mathrm{overlap}}3.2234×10−33.2234\times10^{-3}2.6985×10−42.6985\times10^{-4}5.1084×10−45.1084\times10^{-4}4.6739×10−34.6739\times10^{-3}
\mathcal{L\}_{\mathrm{data}}+\mathcal{L}_{\mathrm{interface}}3.3476×10−33.3476\times10^{-3}3.1550×10−43.1550\times10^{-4}6.3973×10−46.3973\times10^{-4}3.8355×10−33.8355\times10^{-3}
Transolver1.5224×10−21.5224\times10^{-2}7.4794×10−37.4794\times10^{-3}1.0214×10−21.0214\times10^{-2}2.2915×10−12.2915\times10^{-1}
Mondrian1.0799×10−21.0799\times10^{-2}9.4544×10−49.4544\times10^{-4}1.3047×10−31.3047\times10^{-3}4.0638×10−24.0638\times10^{-2}
GDDNO1.9372×10−31.9372\times10^{-3}2.1747×10−42.1747\times10^{-4}4.8618×10−44.8618\times10^{-4}2.6714×10−32.6714\times10^{-3}

Furthermore, to provide a more intuitive illustration, around Page 10, Line 500 of the revised manuscript, we have visualized the predicted fields with zoomed-in local details under three ablation settings: removing the overlap loss, removing the interface loss, and removing both. Severe discontinuities (or named fragmentation) clearly emerge across subdomain boundaries when these losses are omitted, confirming that both consistency losses substantially enhance the continuity and conservation of the solution.

Similarly, eegarding the reviewer's concern about potential artificial smoothing, sharp discontinuities in physical problems, most notably shock waves in supersonic flows, induce steep gradient variations. In numerical simulations, handling strong discontinuities inherently involves a certain degree of numerical dissipation due to discretization schemes and stability considerations. GDDNO captures fine grained local flow structures through its geometry-aware local operators, which largely preserves local features, including steep physical gradients, and avoids introducing excessive smoothing.

In fact, our cavity combustion task serves as strong evidence for this concern. This task is a 3D supersonic combustion case naturally characterized by complex shock waves and turbulence–chemistry interactions. Because GDDNO extracts fine-scale local flow features prior to global communication, the two consistency losses do not eliminate the flow discontinuities, while achieving substantially higher accuracy compared to other baselines (please refer to Figure 3(c), Figure4, Table 6, and Table 14 in the revised manuscript).

Finally, we sincerely appreciate the reviewer's rigor. We acknowledge that the current version of GDDNO does not directly enforce constraints on the numerical fluxes of the predicted solutions, and we have revised the corresponding statements in the manuscript. But we all think incorporating explicit numerical flux conservation is indeed a very promising direction, which we plan to explore in our future work.

Thanks to the authors for their detailed response and additional experiments. The new ablations provide useful evidence for the contributions of global communication and the consistency losses, which address some of my earlier concerns. I also appreciate the authors' efforts to clarify the numerical discrepancies and provide experimental logs.

However, I still have some reservations. The explanations for the numerical inconsistencies are plausible, but the repeated reporting issues make it difficult to fully restore confidence in the empirical results. In addition, while the new experiments show improvements in several conservation-related metrics, the interface loss does not directly enforce flux conservation, and its individual effects are mixed across different metrics.

Overall, the rebuttal has improved my assessment of the paper, but I do not think it fully resolves my main concerns. I will maintain my current rating, although I would not strongly oppose acceptance if the other reviewers find the additional evidence convincing.

AuthorsOpenReview ↗

Response to reviewer tjx4

Thank you for taking the time to assess the additional experiments, and for noting that our additional ablations and logs address part of the earlier concerns. We fully understand why repeated reporting discrepancies make it difficult to immediately restore full confidence in the results, and we appreciate your fair and constructive stance.

Why indirect consistency constraints is beneficial for flux/conservation and why individual effects are mixed:

As noted by the reviewer, GDDNO does not impose an explicit numerical-flux loss. However, in our evaluation (Appendix D.5), face flux is computed by interpolating the predicted velocity field onto mesh faces, and cell divergence aggregates these face fluxes. Interface loss reducing artificial discontinuities across subdomain boundaries and overlap loss aligning overlapping predictions for shared cells, these two losses prevent mismatches that would distort interpolated fluxes and conservation residuals.(Refer to Figure 4 of the manuscript and Figure 2 in Supplementary-response.pdf) At the same time, our geometry-aware local graph encoder still captures genuine steep gradients, such as those in supersonic cavity combustion.

The effects of interface and overlap alone are mixed across metrics because each term's target is differen:

  • Adding interface loss alone penalizes state differences across core boundaries and directly reduces interface flux error from 4.7221e-3 to 3.8355e-3 and overall Rel L2 from 0.0914 to 0.0886.
ComparisonEϕ,ΓE_{\phi,\Gamma}
Data-only → Data + interface−18.8%
  • Adding overlap loss alone enforces agreement among local copies of shared cells, improving cell divergence residual (from 3.1638e-4 to 2.6985e-4), subdomain net-flow imbalance (from 5.9878e-4 to 5.1084e-4), and AirfRANS errors (pressure coefficient error, lift coefficient error, and relative L2 error), while having a smaller effect on cross-boundary interface flux error (4.6739e-3).
ComparisonEdivE_{\mathrm{div}}EpatchE_{\mathrm{patch}}
Data-only → Data + overlap−14.7%−14.7%
  • Combining both consistency losses with data loss coordinates shared-cell representations and cross-boundary transitions together, yielding the strongest results across all four conservation metrics as well as the main accuracy metrics. It also has a very large improvement compared to other baselines.
ComparisonEnetE_{\mathrm{net}}EdivE_{\mathrm{div}}EpatchE_{\mathrm{patch}}Eϕ,ΓE_{\phi,\Gamma}
CFD6.0463×10−56.0463\times10^{-5}1.3531×10−61.3531\times10^{-6}6.5693×10−66.5693\times10^{-6}0.00000.0000
Data only3.2234×10−33.2234\times10^{-3}2.6985×10−42.6985\times10^{-4}5.1084×10−45.1084\times10^{-4}4.6739×10−34.6739\times10^{-3}
Data+Interface+Overlap1.9372×10−31.9372\times10^{-3}2.1747×10−42.1747\times10^{-4}4.8618×10−44.8618\times10^{-4}2.6714×10−32.6714\times10^{-3}
Improv−39.9%−19.4%−4.8%−42.8%
ComparisonEnetE_{\mathrm{net}}EdivE_{\mathrm{div}}EpatchE_{\mathrm{patch}}Eϕ,ΓE_{\phi,\Gamma}
Transolver1.5224×10−21.5224\times10^{-2}7.4794×10−37.4794\times10^{-3}1.0214×10−21.0214\times10^{-2}2.2915×10−12.2915\times10^{-1}
Mondrian1.0799×10−21.0799\times10^{-2}9.4544×10−49.4544\times10^{-4}1.3047×10−31.3047\times10^{-3}4.0638×10−24.0638\times10^{-2}
Transolver → GDDNO−87.3%−97.1%−95.2%−98.8%
Mondrian → GDDNO−82.1%−77.0%−62.7%−93.4%

We again thank Reviewer tjx4 for the thorough engagement and candid feedback throughout the discussion.

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

This paper proposes a framework that partitions the computational mesh into overlapping subdomains and extracts local features using a shared geometry-aware graph neural network. Each subdomain is then compressed into a small number of tokens, which communicate globally through attention. The model combines retained cell-level features with global information to generate predictions, while averaging predictions at shared cells and introducing a consistency loss to coordinate outputs across subdomains. Experiments cover AirfRANS, five systems from RealPDEBench, and 3D combustion flows, including field prediction, engineering quantity estimation, simulation-to-real fine-tuning, and some out-of-distribution evaluations. The results demonstrate competitive predictive performance. However, the independent contributions of the proposed components have not been sufficiently validated, particularly regarding the physical interpretation of the interface constraints and the actual role of global communication.

Critical strengths

1. The evaluation covers diverse settings and practically relevant outputs.

The authors report not only field errors but also pressure coefficients, lift coefficients, and combustion efficiency. In Table 2, GDDNO reduces the full-field AirfRANS relative L2 error from Transolver’s 0.3511 to 0.0815, and the lift-coefficient error from 0.7171 to 0.0735, both substantial improvements. The simulation-trained and real-data fine-tuning results in Table 3, together with the operating-condition extrapolation tests in Appendix D.1, extend the evaluation beyond a single data distribution.

2. The combination of local modeling and boundary treatment has some empirical support.

In Table 5, replacing the local graph operator with node-only encoding increases the relative L2 error from 0.0816 to 0.0945. Jointly removing overlapping cells and the interface consistency loss increases the lift-coefficient error from 0.0736 to 0.1106. These results support the usefulness of local neighborhood information and boundary treatment, although they do not isolate the contribution of each component.

Critical weaknesses

1. The physical interpretation does not match the actual loss, and the ablations do not resolve this.

The abstract claims “flux consistency,” but Appendix A.3 defines the interface loss as the mean squared difference between predicted states in neighboring cells across core boundaries. This penalizes state differences rather than numerical flux imbalance. Since neighboring cells can have genuine physical gradients, I do not see how this loss alone establishes flux consistency. Table 5 removes overlapping cells and the interface loss together, changing both neighborhood information and the objective. I would like to see the two consistency losses ablated separately with the overlap structure fixed, alongside flux or conservation-error measurements. Otherwise, I would describe this term as a boundary-state smoothing regularizer.

2. I am not convinced that the added value over prior work is sufficiently established.

Mondrian [1] already combines local subdomain operators with cross-subdomain attention, while Transolver [2] and UPT [3] use compact representations for global interaction. GDDNO has distinct design choices, but their additional value remains unclear. Also, §3.2 states that the mesh is “partitioned in advance,” which makes “Learning to Decompose” in the title somewhat misleading. I could not find ablations disabling cross-subdomain communication or removing the local-feature skip connection, despite both being central contributions. Tables 6–7 vary the decomposition, but changing the subdomain count also changes the total token count. Direct comparisons under comparable budgets, together with measured runtime and memory, would make the case stronger. Finally, Table 5 describes removing “mesh topology,” while the text says edge attributes. Tables 6–7 also report different R² (0.9788/0.9905) and lift-coefficient errors (0.0871/0.0679) for P=48, H=2, despite identical other metrics. These need clarification. Without run-to-run variability, smaller gains such as 0.4237→0.4148 in Table 4 are also difficult to assess.

Reference:

[1] Feeney, A., Huang, K.-H., and Chandramowlishwaran, A. Mondrian: Transformer Operators via Domain Decomposition. arXiv:2506.08226, 2025.

[2] Wu, H., Luo, H., Wang, H., Wang, J., and Long, M. Transolver: A Fast Transformer Solver for PDEs on General Geometries. ICML, 2024. arXiv:2402.02366.

[3] Alkin, B., Fürst, A., Schmid, S., Gruber, L., Holzleitner, M., and Brandstetter, J. Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators. NeurIPS, 2024. arXiv:2402.12365.

Questions for authors
  1. Appendix A.3 penalizes differences between neighboring states. What justifies interpreting this as flux consistency? Do you have separate ablations of the two consistency losses with overlap fixed, or measurements of flux imbalance?

  2. How much do cross-subdomain communication and the local-feature skip connection each contribute? I could not find ablations isolating these two central components.

  3. What is the main advance over Mondrian, Transolver, and UPT? Comparisons under similar budgets and measured runtime and memory would help clarify the added value.

  4. Could you clarify whether Table 5 removes mesh connectivity or only edge attributes, and reconcile the conflicting results for P=48, H=2 in Tables 6–7? Were the reported gains consistent across random seeds?

4 comments
AuthorsOpenReview ↗

Response to reviewer TVCz (W1, Q1)

W1: I would like to see the two consistency losses ablated separately with the overlap structure fixed, alongside flux or conservation-error measurements.

Q1: Appendix A.3 penalizes differences between neighboring states. What justifies interpreting this as flux consistency? Do you have separate ablations of the two consistency losses with overlap fixed, or measurements of flux imbalance?

We thank the reviewer for this comment. We have revised the statements regarding the direct constraint of flux. Specifically, our designed loss functions substantially improve continuity across the computational domain and physical consistency between subdomains.

To illustrate how these improvements are reflected in the results, we first present the prediction accuracy of the model under different loss configurations. As shown in table below, adding either the overlap loss or the interface loss alone improves the prediction accuracy. Combining the two losses leads to further improvements, providing clear evidence of the effectiveness of both consistency losses.

LossRMSERel L2L_2R2R_2CpC_pClC_l
LdataL_{\mathrm{data}}0.08070.09140.99170.07030.1106
Ldata+LoverlapL_{\mathrm{data}}+L_{\mathrm{overlap}}0.07500.08330.99060.04920.0932
Ldata+LinterfaceL_{\mathrm{data}}+L_{\mathrm{interface}}0.07830.08860.99340.06960.1236
GDDNO0.07210.08160.99340.04800.0736

More importantly, we have added a detailed analysis of the physical consistency and continuity of the predicted flow fields. Specifically, we introduce four metrics:

  • EnetE_{\mathrm{net}}: Net-flow imbalance over the entire computational domain.
  • EdivE_{\mathrm{div}}: Volume-weighted average of cell divergence.
  • EpatchE_{\mathrm{patch}}: Mean net-flow imbalance of subdomains.
  • Eϕ,ΓE_{\phi,\Gamma}: Relative error of subdomain interface flux compared to CFD ground truth.

Due to space limitations, the detailed definitions and calculation methods for these metrics are provided in the revised manuscript. The corresponding results are presented in the table below. As shown, GDDNO's consistency constraints substantially improve both global and local flow-field continuity and physical conservation. These findings further demonstrate the value of incorporating the two proposed constraints within the domain-decomposition framework.

FieldsEnetE_{\mathrm{net}}EdivE_{\mathrm{div}}EpatchE_{\mathrm{patch}}Eϕ,ΓE_{\phi,\Gamma}
CFD6.0463×10−56.0463\times10^{-5}1.3531×10−61.3531\times10^{-6}6.5693×10−66.5693\times10^{-6}0.00000.0000
Ldata\mathcal{L}_{\mathrm{data}}3.0459×10−33.0459\times10^{-3}3.1638×10−43.1638\times10^{-4}5.9878×10−45.9878\times10^{-4}4.7221×10−34.7221\times10^{-3}
Ldata+Loverlap\mathcal{L}_{\mathrm{data}}+\mathcal{L}_{\mathrm{overlap}}3.2234×10−33.2234\times10^{-3}2.6985×10−42.6985\times10^{-4}5.1084×10−45.1084\times10^{-4}4.6739×10−34.6739\times10^{-3}
Ldata+Linterface\mathcal{L}_{\mathrm{data}}+\mathcal{L}_{\mathrm{interface}}3.3476×10−33.3476\times10^{-3}3.1550×10−43.1550\times10^{-4}6.3973×10−46.3973\times10^{-4}3.8355×10−33.8355\times10^{-3}
Transolver1.5224×10−21.5224\times10^{-2}7.4794×10−37.4794\times10^{-3}1.0214×10−21.0214\times10^{-2}2.2915×10−12.2915\times10^{-1}
Mondrian1.0799×10−21.0799\times10^{-2}9.4544×10−49.4544\times10^{-4}1.3047×10−31.3047\times10^{-3}4.0638×10−24.0638\times10^{-2}
GDDNO1.9372×10−31.9372\times10^{-3}2.1747×10−42.1747\times10^{-4}4.8618×10−44.8618\times10^{-4}2.6714×10−32.6714\times10^{-3}

In addition, we provide a more intuitive visualization for the reviewer. Around line 500 on page 10 of the revised manuscript, we plot the results obtained by removing the overlap loss, the interface loss, and both losses, respectively, with enlarged views of local regions. Clear discontinuities can be observed across subdomain boundaries when the overlap and interface losses are removed. These visual results provide further evidence of the contributions of the two consistency losses to solution continuity and physical conservation.

Finally, we appreciate the reviewer's careful observation. We acknowledge that the current GDDNO does not directly impose constraints on the numerical fluxes, and we have revised the statements in the manuscript. But explicit numerical flux constraints are also a promising direction, we will explore it in future work.

AuthorsOpenReview ↗

Response to reviewer TVCz (Q2, W2,Q3–cross-subdomain communication and the local-feature skip connection)

Q2: How much do cross-subdomain communication and the local-feature skip connection each contribute? I could not find ablations isolating these two central components.

W2: I could not find ablations disabling cross-subdomain communication or removing the local-feature skip connection, despite both being central contributions

W2: Mondrian [1] already combines local subdomain operators with cross-subdomain attention, while Transolver [2] and UPT [3] use compact representations for global interaction. GDDNO has distinct design choices, but their additional value remains unclear.

Q3: What is the main advance over Mondrian, Transolver, and UPT?

We thank the reviewer for these comments. These three comments largely cover the core motivations behind the design of the GDDNO, So please allow us to answer these comments together.

First, we have added ablation experiments on cross-subdomain communication and local feature skip connections. The results are presented in the table below. The simple global pooling configuration replaces attention among subdomain tokens with simple global pooling. Compared with the setting where no information is exchanged between subdomains, GDDNO's subdomain communication mechanism reduces the overall Rel L2L_2 error from 0.0996 to 0.0816 and achieves the lowest CpC_p error among the three configurations. This indicates that GDDNO is more effective in predicting both the overall flow field and the physical quantities near the airfoil surface.

communication mechanismRMSERel L2L_2R2R^2CpC_pClC_l
no cross-subdomain communication0.08260.09960.98820.05100.0742
simple global pooling0.08030.09090.99170.04990.0665
GDDNO0.07210.08160.99340.04800.0736

Second, to evaluate local-feature skip connection, we have added an ablation experiment that removes the local feature preservation branch. This increases the Rel L2L_2 error from 0.0816 to 0.1549, nearly doubling the error of the full model. Together, these two groups of ablation experiments demonstrate the important roles of local operator modeling and global communication in GDDNO.

local feature retentionRMSERel L2L_2R2R_2CpC_pClC_l
Remove hidden skip0.14930.15490.96360.09790.1674
GDDNO0.07210.08160.99340.04800.0736

Based on these results, we can further discuss the advantages of GDDNO over Transolver, UPT, and Mondrian. GDDNO offers three main advantages:

  • Preservation of the original discrete geometry. GDDNO operates directly on variable-size cell adjacency graphs, it does not require the unstructured computational domain to be converted into a regular grid beforehand.
  • An independent information pathway for local predictions. The graph-level hidden features of each cell can be directly passed to the decoder. This reduces the extent of local predictions depend exclusively on compact contextual representations and strengthens the model's ability to capture native details.
  • Explicit treatment of prediction consistency across subdomains. The overlap and interface constraints further improve physical continuity across the entire computational domain and between individual patches.

We have also added an analysis of model efficiency in Appendix D.6 of the revised manuscript. Compared with other baselines, GDDNO achieves a favorable balance between computational cost and accuracy. We invite the reviewer to review the Figure 7 of manuscript for a more detailed comparison.

AuthorsOpenReview ↗

Response to reviewer TVCz (W2,Q4)

W2: Tables 6–7 vary the decomposition, but changing the subdomain count also changes the total token count. Direct comparisons under comparable budgets, together with measured runtime and memory, would make the case stronger.

We agree with the reviewer on this point. When the number of tokens per subdomain (KK) is fixed, the total number of tokens (PKPK) varies with the number of subdomains (PP). This may indeed introduce additional effect between partition strategy and the size of the global representation. We have conducted additional ablation experiments, with the results presented in the table below. Overall, the results show a general trend between model accuracy and the number of subdomain tokens, with more tokens tending to improve prediction accuracy. But it also leads to higher costs. Considering the balance between cost and accuracy, we select NP=32N_P=32 and NK=16N_K=16 as the default configuration for the main results presented in the paper.

NKN_KRMSERel L2L_2R2R_2CpC_pClC_lTime<br>(s/epochs/epoch)Memory<br>(GBGB)
10.08130.08990.98270.06570.070819333.57
40.07810.08840.99220.04930.075120034.04
80.07700.08710.99240.05070.070720234.09
160.07210.08160.99340.04800.073620734.18
320.07350.08320.99310.05140.065421134.92
640.07170.08110.99340.04990.054421735.23
1280.07090.08030.99360.04690.059824438.1
2560.06960.07880.99380.04180.049935956.94

W2 Finally, Table 5 describes removing “mesh topology,” while the text says edge attributes.

Q4: Could you clarify whether Table 5 removes mesh connectivity or only edge attributes, and reconcile the conflicting results for P=48, H=2 in Tables 6–7? Were the reported gains consistent across random seeds?

In Table 5, "removing mesh topology" refers to removing geometric attributes such as inter-cell distances, shared-face areas, while preserving the connectivity between nodes. We have revised the corresponding description in Table 5 to avoid ambiguity.

Regarding the conflicting results for P=48P=48 and H=2H=2 in Tables 6–7 (now Tables 7–8), we sincerely apologize for this error. It occurred during manuscript preparation when a keyboard shortcut failed while we were copying data from the original experimental logs. We have carefully rechecked all reported results and confirm that the values in the revised submission are correct. We commit to publicly releasing our code within one month of acceptance. Given the double-blind review requirements, the code in the supplementary materials is functional. we will provide the repository URL in the camera-ready version if the paper is accepted.

The improvements reported in our manuscript are also reproducible across different random seeds. We supplemented the random seed experiment we had previously conducted. As shown in table below, GDDNO consistently maintains its superior prediction accuracy across different random seeds. The standard deviation of GDDNO is small under different random seeds. These results are added in Table12,Table13 of revised manuscript.

AirfRANS

seedRMSERel L2L_2R2R_2CpC_pClC_l
00.07220.08170.99330.04870.0659
100.07430.08210.99290.04770.0690
420.07210.08160.99340.04800.0736
xˉ±s\bar{x} \pm s0.0729$\pm$0.00120.0818$\pm$0.00030.9932$\pm$0.00030.0481$\pm$0.00050.0695$\pm$0.0039

Cavity combustion

seedRMSERel L2L_2R2R_2η\eta
00.34610.40340.66550.1884
100.33580.40960.65750.2065
420.34130.41480.66120.2112
xˉ±s\bar{x} \pm s0.3411$\pm$0.00050.4091$\pm$0.00060.6614$\pm$0.00040.2020$\pm$0.0120

W2: Also, §3.2 states that the mesh is “partitioned in advance,” which makes “Learning to Decompose” in the title somewhat misleading.

Thank you for this comment. In the title, “Learning to Decompose” refers to learning a domain-decomposed representation and computation for the PDE solution operator. Specifically, the geometry-aware operator learns local cell interactions, learned queries determine how cell features contribute to subdomain tokens, and attention learns how information is exchanged across subdomains (§3.3–3.4). A fixed partition therefore does not prescribe what each subdomain represents or how its information influences predictions elsewhere.

The phrase “with geometry-aware local operators and global communication” identifies these learned components. We believe the original title captures this contribution at the operator level and will clarify the distinction between prescribed geometric partitioning and learned representations and interactions in §3.2.

AuthorsOpenReview ↗

Response to reviewer TVCz (W2, Q4)

W2: Without run-to-run variability, smaller gains such as 0.4237→0.4148 in Table 4 are also difficult to assess.

Q4: Were the reported gains consistent across random seeds?

We can answer your question from two perspectives. First, we have provided supplementary experimental results using different random seeds. As you can see, the performance gains achieved by GDDNO are reproducible, and the standard deviation remains within a narrow range.

seedRMSERel L2L_2R2R_2η\eta
00.34610.40340.66550.1884
100.33580.40960.65750.2065
420.34130.41480.66120.2112
xˉ±s\bar{x} \pm s0.3411$\pm$0.00050.4091$\pm$0.00060.6614$\pm$0.00040.2020$\pm$0.0120

Furthermore, Table 4 reports the performance of GDDNO on the three-dimensional cavity combustion problem. The relatively challenging nature of this task is primarily due to three specific characteristics. First, each cavity combustion case contains approximately 600,000 mesh cells. Second, this is a supersonic combustion problem, as described in Section 4 and Appendix 4.1 of the manuscript. Supersonic combustion involves strong turbulence-chemistry interactions. In addition, the supersonic flow conditions make the chemical kinetics equations more numerically stiff. Third, cavity combustion exhibits strong multiphysics coupling. In the experiments reported in Table 4, we attempt to use GDDNO to simultaneously predict 13 variables in a single forward propagation. These include eight chemical species, namely H, H$_2$, H$_2$O, H$_2$O$_2$, HO$_2$, O, O$_2$, and OH, which are derived from a detailed hydrogen combustion mechanism, as well as three flow quantities, pp, TT, and U\mathbf{U}, where U=[Ux,Uy,Uz]\mathbf{U}=[U_x,U_y,U_z] is a three-dimensional velocity vector.

Therefore, compared with AirfRANS and RealPDEBench, cavity combustion task represents a considerably more challenging task. Despite these challenges, GDDNO still achieves the best overall performance.

We also provide another perspective on GDDNO's performance in cavity combustion. Specifically, we decouple the prediction targets and allow each baseline model and GDDNO to focus on predicting a single variable, similar to the settings in airfRANS and RealPDEBench, where the prediction targets consist of only a few variables, such as [Ux,Uy][U_x,U_y] and pp. The results are presented in the table below. As shown, prediction accuracy improves substantially under this setting. Due to space limitations, we report only the Rel. L2L_2 errors for three representative quantities, the chemical species OH and temperature TT, as well as velocity U\mathbf{U}. The complete results are provided in Appendix D.4 of the revised paper.

ModelOHOHTTU[Ux,Uy,Uz]U[U_x,U_y,U_z]
U-Net0.75180.16430.2234
GraphSAGE0.92370.29080.3129
pointNet0.35940.13470.1659
FNO0.39820.13600.1978
Transolver0.86930.28600.2511
GDDNO(ours)0.29330.11280.1528
Reviewer_MZpbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Geometry-Aware Domain-Decomposition Neural Operator (GDDNO). Drawing on domain decomposition from traditional geometry-processing methods, it uses graph networks to capture local geometric details, integrates these details with global coupling, and produces a complete PDE solution through overlapping-subdomain assembly and consistency constraints. The authors present a promising local–global neural-operator framework for complex meshes and provide experimental evaluations across multiple physical tasks. The method is evaluated on AirfRANS, RealPDEBench, and a three-dimensional cavity combustion dataset, with particularly large improvements in full-field AirfRANS predictions and related aerodynamic metrics. However, the current ablation studies do not sufficiently support the effectiveness of the geometry-based domain decomposition, the flux-consistency interpretation, the overall efficiency advantages, or the independent contributions of the core components.

Critical strengths

The paper uses classical domain decomposition to organize computation and communication in neural operators, aiming to enable low-cost information exchange between distant regions while preserving cell-level details on large-scale meshes.

Critical weaknesses

The ablation studies are insufficient to establish the independent contributions of the core components.

Questions for authors

I encourage the authors to add at least the following three sets of ablation studies to show the individual contributions of the proposed components. The key findings should also be verified on at least one additional type of task.

  1. Evaluate the contribution of cross-subdomain communication. Keeping the local encoder, domain partition, and decoder unchanged, compare three settings: no cross-subdomain communication, simple global pooling or averaging-based communication, and the full cross-subdomain communication mechanism. This comparison would help establish the effectiveness of the proposed attention-based communication

  2. Evaluate the two consistency losses separately. Using the same overlapping mesh and assembly procedure, evaluate the independent contributions of the overlap loss and the interface loss, as well as their interaction.

  3. Evaluate local feature retention and token compression. Verify the necessity and effectiveness of retaining local features and compressing subdomain information into tokens.

2 comments
AuthorsOpenReview ↗

Response to reviewer MZpb (Q1,Q2)

W: The ablation studies are insufficient to establish the independent contributions of the core components.+

Q: The key findings should also be verified on at least one additional type of task.

Response: We thank the reviewer for pointing out the limitations of our original manuscript. We have conducted a comprehensive set of ablation experiments covering three aspects: local operator modeling, global communication, and loss functions. All ablation experiments are conducted on both airfRANS and RealPDEBench-FSI task. The results provide strong evidence for the individual contributions of GDDNO's key components and their combined effectiveness.

Q1: Evaluate the contribution of cross-subdomain communication. Keeping the local encoder, domain partition, and decoder unchanged, compare three settings: no cross-subdomain communication, simple global pooling or averaging-based communication, and the full cross-subdomain communication mechanism. This comparison would help establish the effectiveness of the proposed attention-based communication

We have added two groups of experiments. Compared with the setting where no information is exchanged between subdomains, GDDNO's subdomain communication mechanism reduces the overall Rel L2L_2 error from 0.0996 to 0.0816 and achieves the lowest CpC_p error among the three configurations. These results demonstrate that GDDNO provides more effective predictions of both the overall flow field and the physical quantities on the wing surface than models using uniform aggregation or no communication mechanism.

communication mechanismAirfRANS RMSEAirfRANS Rel L2L_2AirfRANS R2R^2AirfRANS CpC_pAirfRANS ClC_lFSI(numerical) Rel L2L_2FSI(numerical) R2R^2FSI(real) Rel L2L_2FSI(real) R2R^2
No communication0.08260.09960.98820.05100.07420.15730.91860.05840.9751
simple global pooling0.08030.09090.99170.04990.06650.15580.91970.05800.9813
GDDNO0.07210.08160.99340.04800.07360.14130.93250.05760.9885

Q2: Evaluate the two consistency losses separately. Using the same overlapping mesh and assembly procedure, evaluate the independent contributions of the overlap loss and the interface loss, as well as their interaction.

We compare four loss configurations while keeping other strategy unchanged. Compared with the data-only setting, adding the overlap loss improves all four error metrics. Adding the interface loss alone improves the overall Rel L2L_2 error. When both losses are incorporated, all four metrics, namely RMSE, Rel L2L_2, CpC_p error, and ClC_l error, are better than those obtained with either individual loss. Relative to the data-only, the full loss formulation reduces the Rel L2L_2, CpC_p, and ClC_l errors by approximately 10.77%, 31.68%, and 33.45%, respectively. It also achieves a further 2.14% reduction in Rel L2L_2 compared with using the overlap loss alone. These results support the value of jointly incorporating the two constraints within the domain-decomposition framework.

LossAirfRANS RMSEAirfRANS Rel L2L_2AirfRANS R2R^2AirfRANS CpC_pAirfRANS ClC_lFSI(numerical) Rel L2L_2FSI(numerical) R2R^2FSI(real) Rel L2L_2FSI(real) R2R^2
Ldata\mathcal{L}_{\mathrm{data}}0.08070.09140.99170.07030.11060.15490.92020.05940.9835
Ldata+Loverlap\mathcal{L}_{\mathrm{data}}+\mathcal{L}_{\mathrm{overlap}}0.07500.08330.99060.04920.09320.15490.92050.05780.9819
Ldata+Linterface\mathcal{L}_{\mathrm{data}}+\mathcal{L}_{\mathrm{interface}}0.07830.08860.99340.06960.12360.15530.91990.05870.9826
GDDNO0.07210.08160.99340.04800.07360.14130.93250.05760.9885
AuthorsOpenReview ↗

Response to reviewer MZpb (Q3)

Q3: Evaluate local feature retention and token compression. Verify the necessity and effectiveness of retaining local features and compressing subdomain information into tokens.

We have added an ablation experiment that completely removes the local feature retention branch, retaining only the local graph encoding and global token pathways. Removing this branch increases the Rel L2L_2 error from 0.0816 to 0.1549, nearly doubling the error of the full model. The results of FSI also show the important role of local feature retention.

LossAirfRANS RMSEAirfRANS Rel L2L_2AirfRANS R2R^2AirfRANS CpC_pAirfRANS ClC_lFSI(numerical) Rel L2L_2FSI(numerical) R2R^2FSI(real) Rel L2L_2FSI(real) R2R^2
Remove hidden skip0.14930.15490.96360.09790.16740.15290.91910.12380.9418
GDDNO0.07210.08160.99340.04800.07360.14130.93250.05760.9885

Local feature preservation and token compression work together. Without token compression, attention need to be computed over the entire mesh. Each airfRANS case contains approximately 250,000–300,000 mesh cells. Taking N=3×105N=3\times10^5 cells, a hidden dimension of D=256D=256, and 16 attention heads as an example, the attention matrices would contain approximately 16N216N^2 elements in total. This would require approximately 5.76 TB of GPU memory in FP32 precision. We hope the reviewer understands that conducting such a expensive experiment is currently beyond our available resources.

In contrast, our current configuration with NP=32N_P=32 and NK=16N_K=16 uses only 512 tokens to represent both local features and the relative relationships between subdomains. This is one of the key contributions of our work. We have analyzed the computational cost and prediction accuracy under different values of KK. The results are presented below. As can be observed, weaker compression generally leads to higher accuracy, but also incurs greater computational costs. Considering the balance between accuracy and cost, we ultimately select NK=16N_K=16 as our default configuration.

NKN_KRMSERel L2L_2R2R_2CpC_pClC_lTime(s/epochs/epoch)Memory(GBGB)
10.08130.08990.98270.06570.070819333.57
40.07810.08840.99220.04930.075120034.04
80.07700.08710.99240.05070.070720234.09
160.07210.08160.99340.04800.073620734.18
320.07350.08320.99310.05140.065421134.92
640.07170.08110.99340.04990.054421735.23
1280.07090.08030.99360.04690.059824438.1
2560.06960.07880.99380.04180.049935956.94

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers and Area Chair

Dear Reviewers and Area Chair,

We hereby submit our revised manuscript along with our detailed responses to all reviewers' comments. Based on your constructive feedback, we have made the following major updates during this rebuttal period:

  1. Additional Ablation and Analysis: We have conducted further ablation studies covering the global communication mechanism, the overlap and interface consistency losses, and local feature connections. In addition, we have included multi-seed evaluations and model efficiency analyses, which make the validation of our work much more comprehensive.

  2. Continuity and Physical Conservation Analysis: From the perspective of our two consistency losses, we have systematically analyzed the continuity and physical conservation properties of the solutions predicted by GDDNO. The results confirm that both consistency losses contribute positively to addressing these challenges.

Due to the formatting limitations of OpenReview comments, we have also uploaded a supplementary response PDF in the supplementary material (Supplementary-response.pdf), which systematically addresses the common issues raised across the reviews. We kindly invite the Reviewers and Area Chair to review both the revised manuscript and the supplementary document.

We sincerely thank the reviewers for their thorough and insightful evaluation of our manuscript. These constructive comments that have helped us strengthen both the clarity and rigor of this paper. We have carefully addressed all concerns below, and we believe these revisions have substantially improved the manuscript. Looking forward to receiving your reply

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.