acceptodds
Under review as a conference paper at ICLR 2027

MeshOp: Operator Learning for Adaptive Mesh Refinement in Parametric PDEs

23 views

Abstract

Adaptive finite-element methods (AFEM) construct problem-dependent meshes through repeated solve–estimate–mark–refine cycles, but this sequential mesh-discovery process must be repeated across PDE instances with different input functions or physical parameters. We propose MeshOp, an operator-learning framework for adaptive mesh refinement across parametric PDE families. MeshOp predicts AFEM-derived refinement fields directly from PDE inputs, thereby reducing the cost of sequential mesh discovery. The predicted field is converted into a conforming hierarchical mesh by a dedicated score-to-mesh realization procedure, while residual-based AFEM can continue from this learned initialization when higher accuracy is required. Across the three benchmarks, the mean exact-level disagreement between quantized predictions and AFEM targets ranges from to , yet the realized meshes attain finite-element errors comparable to those of fixed-cycle AFEM. This empirical downstream tolerance suggests that effective discretization does not require exact pointwise recovery of the AFEM refinement field. MeshOp reduces online time by and relative to reference-error-matched AFEM on Burgers and disk convection–diffusion, respectively, and by relative to comparable-accuracy fixed-cycle AFEM on three-dimensional reaction–diffusion.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Operator Learning, Adaptive Mesh Refinement, Adaptive Finite Element Methods, Parametric PDEs

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

The paper introduces MeshOp a Neural Operator approach to adaptive mesh refinement for finite element methods. MeshOp precits an adaptive finite-element refinement field which means effectively instead of having to go through multiple rounds of FEM solve and subsequent a-posteriori error estimation, the MeshOp is able to provide a one-shot estimate of regions in the domain that require fine meshes. The method is evaluated on three PDE test problems including a dynamic problem, a stationary case and a 3D example.

Critical strengths

First of all, I would like to highlight that the presentation of this paper is of significantly higher quality that most submissions I have reviewed for ML conference proceedings in recent months. The story is coherently presented and within the first few pages of the manuscript the problem statement and contributions are clearly presented. Thank you for not submitting AI slob. Beyond the presentation, I believe the scientific contribution of the manuscript to be of high value: mesh generation and adaptive meshing are of paramount importance to finite element methods which still constitute the most widely used computational technique across the sciences and engineering. The chosen numerical examples are also very appropriate for the requirement and the favourable performance of the method is clearly established.

Critical weaknesses

The manuscript contains some minor weaknesses which I would recommend the authors to consider for a camera-ready version (or revised resubmission as applicable):

  1. Baselines in the numerical experiments: Although the MeshOp framework is thoroughly evaluated, I believe a stronger case could be made for this approach by providing more extensive comparisons against standard baselines in adaptive mesh refinement. In particular, there are reinforcement learning approaches to adaptive mesh refinement (e.g. Freymuth et al., 2025, as cited) at least one of which could serve as a useful baseline here.

  2. To place the work more thoroughly in the appropriate literature I would encourage the authors to cite the following works:

(a) Bohn, J., and Feischl, M. (2021). Recurrent neural networks as optimal mesh refinement strategies. Computers & Mathematics with Applications.

(b) Rowbottom, J., Maierhofer, G., Deveney, T., et al. (2025). G-Adaptivity: optimised graph-based mesh relocation for finite element methods. Proceedings of the 42nd International Conference on Machine Learning (ICML 2025).

  1. Schematic of input/output of MeshOp: Although sections 3 and 4 do a good job in describing the MeshOp scheme, it would be nice if the authors could consider including an architectural schematic earlier on in the paper, that would showcase more clearly the input (solution field, domain representation etc) and the output (refinement field) that their model uses.
Questions for authors
  1. I wonder if it might be possible to derive heuristics about mesh refinement directly from the solution field. For example would it be possible to construct a "classical" refinement field from a coarse solve solution curvature field? I suspect this would be slightly more efficient than MeshOp because of negligible cost, but it might lead to worse refinement performance and error reduction?
6 comments
AuthorsOpenReview ↗

Response to Reviewer LJDV [Part 1]

Thank you for your careful review and encouraging assessment of our work.

1. Learning-based mesh-adaptation baselines.

Response: We agree that comparisons with learned mesh-adaptation methods are important. We explored several learning-based mesh-adaptation methods as potential baselines. For Adaptive Swarm Mesh Refinement (ASMR), a reinforcement-learning approach, we did not obtain a converged policy after adapting it to our Burgers space–time FEM setting. We therefore selected the data-free mesh mover (DMM) of Hu, Wang, and Ma [1] and added a comparison using its solution-conditioned mesh relocation followed by conventional FEM. DMM is an rr-adaptive mesh method that moves existing nodes while preserving their number and connectivity. In our implementation, an FNO first predicts the Burgers solution. DMM uses this prediction to move the nodes of an initially uniform space–time mesh, and the Burgers equation is then solved on the relocated mesh using periodic P1 FEM. The FNO is trained on 3000 instances; DMM is trained on the corresponding numerical solution fields with a physics-based loss, without optimal-mesh labels. At test time, mesh relocation uses only the FNO-predicted solution.

The following results are averaged over the same 100 test instances, with no additional AFEM correction and τ=0.5\tau=0.5 for MeshOp (FNO):

MethodMesh generatorST rel. L2L^2FT rel. L2L^2Free DOFTime (s)
MeshOp (FNO)NVB1.243e-31.858e-3148690.458
DMM (FNO)DMM4.694e-35.888e-3150000.665

At similar DOF counts, MeshOp has lower errors and online cost in this comparison. DMM can reuse solution datasets without requiring AFEM-derived mesh labels. These results are reported in the appendix subsection "Comparison with a learned mesh mover."

[1] Peiyan Hu, Yue Wang, and Zhi-Ming Ma. Better Neural PDE Solvers Through Data-Free Mesh Movers. ICLR, 2024.

2. Related work.

Response: Thank you for highlighting these important works on learning-based mesh adaptation, which we overlooked in the original manuscript. Both references are now cited and discussed in Section 2. These additions strengthen the related-work discussion and help clarify the positioning of MeshOp within the existing literature.

3. Input–output architecture schematic.

Response: We have added an architecture schematic as Figure 1. It shows the functional PDE inputs, the alternative FNO, DeepONet, and CNO backbones, and the predicted refinement field. The inputs are problem data such as forcing, initial conditions, or boundary data, as appropriate to each benchmark. Figure 2 focuses on the complete offline and online workflow, including mesh realization and FEM solution.

AuthorsOpenReview ↗

Response to Reviewer LJDV [Part 2]

4. Mesh generation guided by an approximate solution.

Response: Thank you for this suggestion. We conducted a preliminary Burgers experiment to test whether an approximate solution can guide mesh generation through a classical gradient-based monitor.

An FNO trained on 3000 instances predicts the Burgers solution u^\hat u, using the same architecture and training settings as the direct-solution baseline. We adapt the gradient-based monitor used by Hu et al. [1] to the normalized space–time coordinates (ξ,t)=((x+1)/2,t)(\xi,t)=((x+1)/2,t):

m=1+\frac{g}{0.01\langle g\rangle},\qquad \rho=\frac{m}{\langle m\rangle}.$$ where $\langle\cdot\rangle$ denotes the domain average. The mesh-size field is $h=\operatorname{clip}(C/\sqrt{\rho},1/2048,1/32)$, where $C$ controls the overall mesh size. Derivatives are evaluated by second-order finite differences, with periodic differences in $x$; near-constant predictions use a uniform monitor. Gmsh generates the mesh from this field, followed by the same periodic P1 FEM solver and error evaluator used in the backbone/mesh-generation comparison. Test reference solutions are used only for evaluation. Results below are means over the same 100 test instances, with no additional AFEM correction and $\tau=0.5$ for NVB: | Method | Mesh generator | $C$ | ST rel. $L^2$ | FT rel. $L^2$ | Free DOF | Time (s) | |---|---|---|---|---|---|---| | MeshOp (FNO) | NVB | — | 1.243e-3 | 1.858e-3 | 14869 | 0.458 | | MeshOp (FNO) | Gmsh | — | 8.267e-4 | 1.280e-3 | 15634 | 0.912 | | FNO + gradient | Gmsh | 0.009 | 1.323e-3 | 1.640e-3 | 14270 | 0.903 | | FNO + gradient | Gmsh | 0.011 | 1.943e-3 | 2.353e-3 | 9581 | 0.598 | | FNO + gradient | Gmsh | 0.0156 | 3.985e-3 | 4.717e-3 | 4794 | 0.320 | At $C=0.009$, the solution-guided baseline achieves comparable FEM accuracy at a similar DOF count: its space–time error is slightly higher and its final-time error lower than MeshOp (FNO) with NVB. Its online time is 0.903 s, compared with 0.458 s for the NVB pipeline. When both methods use Gmsh, direct prediction of the AFEM-derived mesh-size field gives lower errors with a somewhat larger mesh and a similar online time of 0.912 s. Increasing $C$ reduces DOFs and online time at the expense of accuracy. Constructing the gradient monitor and size field takes only about 0.001 s, consistent with the expectation that this step is inexpensive. Gmsh mesh generation is the main bottleneck in the current implementation. These preliminary results show that approximate solutions can provide useful mesh guidance. We would like to investigate more efficient ways to realize these solution-derived size fields in future work. [1] Peiyan Hu, Yue Wang, and Zhi-Ming Ma. *Better Neural PDE Solvers Through Data-Free Mesh Movers.* ICLR, 2024.

I am grateful to the authors for their careful consideration of my concerns and for addressing them so thoroughly. I can confirm that points 2-4 were addressed completely and that I am also grateful to see the additional baseline experiment performed by the authors and included in Appendix B of the revised manuscript. This experiment is particularly important and valuable as it helps place the advances of the current contribution within the literature.

I have therefore updated my informativeness score to 3 and I am content to keep my overall rating at 3 (weak acceptance). My main remaining concern is that the DMM comparison arguably shows only a limited advantage of the MeshOp framework (relative error from ∼4×10−3\sim 4\times10^{-3} to ∼10−3\sim10^{-3} and time from $\sim$0.6sec to $\sim$0.4sec) which on a single example might not be fully sufficient to warrant the novel methodology. To convincingly compare against baselines, in my opinion the authors would have to also include baseline comparisons in the main experiments shown in the main body of the paper.

In summary, my overall assessment is positive and elevated by the rebuttal, but the remaining concern is for the authors to demonstrate a convincing advantage of their method over baselines across a larger class of problems rather than in one specific instance.

Reviewer_8jdXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an operator-learning framework for accelerating adaptive mesh refinement in parametric PDE problems. Instead of directly predicting PDE solutions, MeshOp learns a mapping from problem-dependent inputs to refinement-generation fields derived from conventional adaptive finite-element methods.

The predicted refinement field is converted into a conforming hierarchical mesh through a score-to-mesh realization procedure based on recursive newest-vertex bisection and conformity closure. A conventional finite-element solver then computes the PDE solution on the resulting mesh. When higher accuracy is required, standard residual-based AFEM cycles can continue from the learned initialization.

The method is evaluated on three PDE families: space–time Burgers, two-dimensional convection–diffusion on a disk, and three-dimensional reaction–diffusion. Although predicted refinement levels differ from AFEM targets, the resulting meshes achieve comparable finite-element solution accuracy. The paper reports online speedups of approximately 2.10–3.06× against the corresponding AFEM baselines, with additional experiments on AFEM correction, alternative neural-operator backbones, and out-of-distribution Burgers inputs.

Critical strengths
  1. integration of operator learning with conventional adaptive finite-element solvers.
  2. downstream evaluation of mesh quality and computational efficiency.
Critical weaknesses
  1. The methodological novelty and source of improvement over existing learned mesh-adaptation approaches are not sufficiently isolated.
  2. The efficiency advantages depend on substantial offline training costs and a restricted family of problems.
Questions for authors
  1. Separate the contributions of refinement-field prediction and hierarchical mesh realization through controlled comparisons using a common mesh-generation procedure
  2. Given the estimates of approximately 2,500–42,000 queries, how sensitive is the overall computational advantage to the number of AFEM training trajectories?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 8jdX [Part 1]

Thank you for your thoughtful review and for recognizing the integration with conventional FEM and the downstream assessment of mesh quality and efficiency. Your questions motivated additional controlled comparisons and a more detailed analysis of the training-data and computational costs.

1. Methodological novelty and separation of prediction from mesh realization.

1.1. Contributions and relationship to existing methods.

Response. MeshOp introduces a function-to-field formulation of fixed-budget adaptive refinement: it learns the cumulative refinement demand of a prescribed residual-based AFEM procedure directly from functional PDE inputs. This formulation focuses on learning where mesh resolution is needed while retaining conventional FEM to compute the solution. Our experiments support this choice: direct solution prediction yields larger errors than FEM on MeshOp-generated meshes (Appendix B.1, Table 10), whereas imperfect refinement predictions can still produce effective FEM discretizations (Tables 2–3).

The predicted refinement-generation field is realized as a conforming hierarchical mesh through threshold-controlled NVB and conformity closure, allowing continued residual-based AFEM correction when higher accuracy is required. Specifically, our three contributions are:

  1. Refinement-operator formulation. We formulate fixed-budget adaptive refinement as a function-to-field operator, mapping functional PDE inputs to AFEM-derived refinement-generation fields on a common query grid.
  2. Hierarchical score-to-mesh realization. We introduce a threshold-controlled score-to-mesh procedure using recursive NVB refinement and conformity closure, producing conforming hierarchical meshes for conventional FEM and continued AFEM correction.
  3. Empirical downstream tolerance. Across three PDE families, we demonstrate that predicted refinement fields yield FEM accuracy comparable to fixed-cycle AFEM despite pointwise score errors, while reducing online cost relative to the reported AFEM baselines.

These contributions build on established FNO and NVB components; they are summarized in the Introduction and discussed in relation to prior work in Section 2.

1.2. Separating refinement-field prediction and mesh realization.

We address the requested separation through complementary controls:

  • Same realization, target versus predicted field. The Target-field reference in Table 3 uses AFEM-derived scores with the same realization procedure and FEM solver as MeshOp. For Burgers, the mean ST relative L2L^2 errors are 9.810e-4 for Target-field and 1.243e-3 for MeshOp (FNO). This measures the downstream effect of replacing target scores with learned predictions, including the resulting change in mesh size.
  • Same realization, different predictors. Appendix E.1, Table 16 compares FNO, CNO, DeepONet, and POD-DeepONet using the same NVB procedure, threshold, and FEM solver. These results establish compatibility with multiple backbones and show that prediction quality still influences FEM accuracy.
  • Crossed backbone and mesh-generator comparison. The new Appendix B.2, Table 11 compares FNO and U-Net with both NVB and Gmsh. All configurations use the same 3000 training instances, 1000 epochs, 100 test instances, periodic P1 FEM solver, reference solutions, and timing convention. FNO and U-Net have 48,190 and 48,064 trainable parameters, respectively. NVB uses τ=0.1\tau=0.1 in this comparison, and no configuration uses AFEM correction.

ST and FT denote space-time and final-time errors.

BackboneMesh generatorMean ST rel. L2L^2Mean FT rel. L2L^2Free DOFsOnline time (s)
FNONVB9.783e-41.485e-319,2910.549
U-NetNVB9.154e-41.384e-320,3290.565
FNOGmsh8.267e-41.280e-315,6340.912
U-NetGmsh8.764e-41.269e-316,1231.107

With either generator held fixed, the backbones give comparable FEM accuracy. The lower online cost of the NVB pipeline is observed with both backbones, while Gmsh achieves lower errors with fewer DOFs at these settings.

Each model is trained for its generator's control field: refinement-generation scores for NVB and mesh-size fields for Gmsh, both derived from the same AFEM meshes. Consequently, the cross-generator experiment compares complete pipelines; it does not isolate the generator alone under an identical predicted field or a matched DOF budget. The fixed-realization comparisons above provide the more direct control for the prediction component.

Section 5.5, Table 5 separately identifies the source of the improvement over AFEM: avoiding repeated assembly, estimation, marking, and mesh-management work. For example, 3D reaction-diffusion mesh-related time decreases from 6.091 s for fixed-10-cycle AFEM to 1.202 s for MeshOp, with total time decreasing from 9.259 s to 3.027 s.

AuthorsOpenReview ↗

Response to Reviewer 8jdX [Part 2]

2. Offline training costs, sensitivity to the number of trajectories, and the intended workload.

Response. We evaluate Burgers models trained on 25-3000 instances with the same FNO architecture, 1000 training epochs, and the same 100 test instances (Appendix A.1). For repeated queries within a prescribed PDE family and geometry, Appendix D.3 estimates the break-even query count, including target generation and training, as

NBE=Ttarget+TtrainTAFEM−TMeshOp.N_{\mathrm{BE}}=\frac{T_{\mathrm{target}}+T_{\mathrm{train}}}{T_{\mathrm{AFEM}}-T_{\mathrm{MeshOp}}}.

The following table reproduces the Burgers base-stage costs and estimated BE counts from Table 15.

Training instancesTarget generation (h)Training (h)Total offline (h)Base-stage BE
250.0070.0180.0254.86×1024.86\times10^{2}
500.0130.0290.0424.33×1024.33\times10^{2}
1250.0330.0490.0826.44×1026.44\times10^{2}
2500.0670.0820.1491.08×1031.08\times10^{3}
5000.1330.1480.2811.89×1031.89\times10^{3}
10000.2670.4300.6964.47×1034.47\times10^{3}
20000.5330.6041.1377.18×1037.18\times10^{3}
30000.8000.9101.7101.00×1041.00\times10^{4}

Across this sweep, total offline cost increases from 0.025 h to 1.710 h, while mean MeshOp online time remains between 0.3915 s and 0.4580 s (Table 7). The estimated break-even counts reflect both the offline investment and the per-query savings relative to reference-error-matched AFEM.

To relate these cost estimates to downstream accuracy, a useful operating point is 500 trajectories:

Quantity500 training instances3000 training instances
Mean ST rel. L2L^21.233e-31.243e-3
Mean FT rel. L2L^21.896e-31.858e-3
Maximum FT rel. L2L^26.682e-35.413e-3
Online time per instance0.4438 s0.4580 s

These results show that a sixfold smaller training set retains similar mean FEM accuracy and substantially lowers the estimated amortization threshold, although the maximum final-time test error increases. Further reducing the training set to 25 or 50 instances increases both mean FEM errors and maximum test errors (Table 7), indicating that overly small training sets can compromise downstream accuracy.

AuthorsOpenReview ↗

Official Comment by the Authors

Thank you again for your careful review. As the discussion period draws to a close, we would appreciate any feedback on whether our responses have addressed your main concerns. Please let us know if any concerns remain unresolved or if further clarification would be helpful. Thank you for your time and consideration.

AuthorsOpenReview ↗

Official Comment by the Authors (Additional Clarifications)

Thank you again for your careful review. As the discussion period draws to a close, we hope that our responses and the additional evidence have addressed your two main concerns:

  • Prediction and mesh realization: common-realization controls and backbone–mesh-generator comparisons, detailed in Part 1.
  • Offline cost: training-size sensitivity and break-even analysis, detailed in Part 2.

We have also added a learned mesh-relocation baseline: an FNO-predicted solution guides the data-free mesh mover (DMM), followed by conventional periodic P1 FEM.

MethodFree DOFsST rel. L2Online time (s)
MeshOp (FNO)14,8691.243e-30.458
DMM (FNO)10,0006.110e-30.433
DMM (FNO)15,0004.694e-30.665
DMM (FNO)25,0002.372e-31.210
This comparison is now in the main results (Table 3; details in Appendix B.3).

Please let us know if any main concerns remain unresolved or if further clarification would be helpful. Thank you for your time and consideration.

Reviewer_4FZiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

MeshOp is an operator-learning framework designed to bypass the costly, sequential solve-estimate-mark-refine cycles of Adaptive Finite Element Methods (AFEM) across parametric PDE families. It achieves this by mapping functional PDE inputs directly to an AFEM-derived target refinement-demand field on a common query grid, which is subsequently converted into a conforming hierarchical mesh using a dedicated score-to-mesh realization algorithm with Newest-Vertex Bisection (NVB). The paper's primary contributions include formulating fixed-budget adaptive refinement as a function-to-field operator learning task, introducing threshold-controlled hierarchical mesh realization, and revealing an empirical downstream tolerance where imperfect score fields still yield highly effective discretizations. Validated across space-time Burgers, 2D disk convection-diffusion, and 3D reaction-diffusion benchmarks, MeshOp attains finite-element solution accuracy competitive with standard fixed-cycle AFEM while delivering 2.10×2.10\times to 3.06×3.06\times online speedups. Additionally, the predicted meshes prove versatile by functioning either as direct terminal discretizations or as adaptive initializations from which standard residual-based AFEM can seamlessly continue when higher solution accuracy is required.

Critical strengths
  1. Novel Function-to-Field Operator Formulation Bypassing Sequential Discovery Cycles

The paper successfully reformulates adaptive mesh refinement as a supervised function-to-field operator task, mapping functional PDE inputs directly to cumulative AFEM refinement-demand fields on a fixed query grid. Combined with a dedicated score-to-mesh algorithm using Newest-Vertex Bisection (NVB) and local conformity closure, it allows one-shot generation of conforming hierarchical meshes without relying on continuous remeshing or iterative graph/RL predictions.

  1. Empirical Downstream Tolerance and Practical AFEM Warm-Start Flexibility

The authors identify and substantiate an "empirical downstream tolerance": exact pointwise recovery of target refinement scores is unnecessary because spatial aggregation, quantization, and conformity closure naturally absorb minor score errors. Furthermore, the generated mesh can function either as a direct terminal discretization or as an adaptive initial mesh for continued residual-based AFEM correction cycles.

Critical weaknesses
  1. MeshOp relies on fixed initial mesh templates T0\mathcal{T}_0 (e.g., a fixed 32×3232\times32 Cartesian partition or a fixed Delaunay disk mesh) and uniform query grids defined for a single specific geometry. The network cannot generalize across changing physical domain shapes, varying boundary topographies, or different PDE families without generating thousands of expensive AFEM offline trajectories and retraining the operator from scratch.

  2. Key terminology and experimental baselines are presented in a confusing manner. The acronym NVB (Newest-Vertex Bisection) appears for the first time on line 100 without being defined until much later in the text (line 193). Furthermore, in main tables such as Table 3 and Table 4, the proposed method is labeled generically as "FNO" (the neural operator backbone used) rather than "MeshOp", while "Target-field" is listed alongside standard baselines without clear initial context.

Questions for authors

N.A.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 4FZi

Thank you for your careful review and constructive suggestions.

1. Generalization scope and offline training cost.

Response: We agree that extending learned mesh refinement across geometries, boundary topologies, and PDE families is challenging. The current experiments use a separate model and a fixed initial mesh for each PDE family; cross-geometry and cross-PDE transfer is not established.

Offline cost and model reuse. The added Burgers study evaluates 25–3000 training instances on the same 100 test cases, with τ=0.5\tau=0.5 and no additional AFEM correction (Appendix A.1, Tables 6–7). Between 500 and 3000 instances, mean FEM errors change little, although refinement-field prediction continues to improve:

Training instancesMean STMean FTMaximum FTEstimated break-even queries
5001.233e-31.896e-36.682e-31.89×1031.89\times10^3
30001.243e-31.858e-35.413e-31.00×1041.00\times10^4

Thus, 500 training instances retain comparable mean downstream FEM accuracy in this Burgers setting, with a higher maximum final-time error. The estimated total offline cost decreases from 1.710 to 0.281 h, and the estimated break-even count from 1.00×1041.00\times10^4 to 1.89×1031.89\times10^3 queries (Appendix D.3, Table 15). In addition, one potential application is large-scale PDE dataset generation, where FEM solutions computed on MeshOp-generated meshes can serve as training data for subsequent models. Such workloads require solving many related PDE instances and therefore provide a natural setting for amortizing the offline cost.

The trained Burgers model is also evaluated without retraining on 100 OOD instances with an independent second harmonic added to the initial conditions (Appendix B.4, Table 13). Direct realization and realization followed by two AFEM correction cycles retain 2.19×2.19\times and 1.50×1.50\times speedups over the corresponding reference-error-matched AFEM baselines. These results support model reuse under the tested within-family input-distribution shift.

Geometric scope and extensions. The disk convection–diffusion benchmark uses a nonrectangular domain and an unstructured Delaunay initial mesh, demonstrating applicability beyond rectangular domains and Cartesian initial meshes. The fixed initial mesh gives refinement generations a common reference, while the final mesh varies with the PDE input. We could include geometry and initial-mesh information, such as local element sizes, as additional inputs and train the model on different initial meshes, although this would require additional AFEM data and a consistent definition of refinement labels across these meshes.

Appendix B.2 also examines an alternative realization route: predicting a mesh-size field for Gmsh instead of refinement generations for NVB. Gmsh generates a mesh from the prescribed geometry and size field without retaining the NVB hierarchy. At comparable FEM accuracy, total online time in the tested FNO configurations is 0.912 s for Gmsh versus 0.549 s for NVB (Table 11), with Gmsh attaining lower errors and using fewer DOFs at these settings.

For broader transfer, GINO [1] demonstrates solution-operator learning across varying geometries, while PDEformer-2 [2] supports pretraining and adaptation across multiple PDE families through PDE-graph conditioning. These approaches suggest possible directions for extending MeshOp across geometries and PDE families. Exploring these extensions will be a major focus of our future work.

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

[2] Zhanhong Ye, Zining Liu, Bingyang Wu, Hongjie Jiang, Leheng Chen, Minyan Zhang, Xiang Huang, Qinghe Meng, Jingyuan Zou, Hongsheng Liu, and Bin Dong. PDEformer-2: A Versatile Foundation Model for Two-Dimensional Partial Differential Equations. arXiv:2507.15409, 2026.

2. Terminology and baseline labels.

Response: These presentation issues have been addressed:

  • NVB: The acronym is expanded as “newest-vertex bisection” at its first occurrence in Section 2.
  • MeshOp versus FNO: The relevant FEM results are now labeled MeshOp (FNO), identifying the complete prediction–mesh-realization–FEM pipeline and specifying FNO as its backbone. This notation is introduced in Section 5.1 and used in the relevant tables and figures.
  • Target-field reference: Section 5.1 explicitly defines this reference: AFEM-derived target refinement scores replace the predicted scores, followed by the same mesh-realization and FEM procedures. It provides a diagnostic comparison of the effect of refinement-field prediction on downstream FEM accuracy.
AuthorsOpenReview ↗

Official Comment by the Authors

Thank you again for your careful review. As the discussion period draws to a close, we would appreciate any feedback on whether our responses have addressed your main concerns. Please let us know if any concerns remain unresolved or if further clarification would be helpful. Thank you for your time and consideration.

Thank the author for the clarification. My main concern is clarified and the presentation is improved after revision. Therefore, I raise my score to 3.

AuthorsOpenReview ↗

Official Comment by the Authors

Thank you for your careful reassessment and constructive feedback. We are glad that our clarifications have addressed your main concern. Your suggestions have helped us improve the clarity and presentation of the manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to AC and reviewers

Dear Area Chair and Reviewers,

Thank you for your careful reviews and constructive feedback. The manuscript has been revised to address the concerns raised, and the updated version has been uploaded to OpenReview with all changes highlighted in blue. The main updates are summarized below.

  • Training-size sensitivity and offline costs. Extended the Burgers experiments to training sets ranging from 25 to 3000 instances, reporting refinement-field prediction, downstream FEM accuracy including maximum test errors, and the corresponding offline costs and estimated break-even counts.

  • Backbone and mesh-generation comparisons. Added controlled Burgers comparisons of FNO and U-Net with both NVB and Gmsh. Each model is trained separately for the control field required by its mesh generator, helping distinguish the contributions of the backbone and mesh realization.

  • Additional mesh-adaptation baseline. Added a solution-guided baseline based on the data-free mesh mover (DMM), which uses FNO-predicted solutions to relocate a mesh before the downstream FEM solve.

  • Presentation and related work. Added an input–output architecture schematic, clarified terminology and baseline labels, and expanded the discussion of related mesh-adaptation methods.

Detailed responses to the individual concerns are provided in the corresponding reviewer threads. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.