acceptodds
Under review as a conference paper at ICLR 2027

GEMove: Unsupervised Geometric Mesh Movement Beyond Density

Abstract

Mesh movement adaptively relocates mesh elements under a fixed computational budget to achieve an accuracy-efficiency tradeoff in scientific simulations. Existing methods largely rely on costly theoretical supervision, whereas unsupervised approaches offer a promising alternative with zero-shot generalizability across diverse scenarios. However, under limited unsupervised constraints, local density equidistribution is prone to physically inconsistent movements and poorly coordinated over-allocation. To address these issues, we propose GEMove, which enforces global geometric equidistribution in a unified mapped space, thereby deriving vectorized movement indicators for each mesh element. Furthermore, GEMove models the mesh topology as a spring system, enabling global coordination through Tutte equilibrium. Extensive experiments across four PDE families demonstrate that GEMove achieves an average error reduction of 24.62%, with strong zero-shot generalization to unseen equations and domain geometries. Beyond simulation benchmarks, GEMove effectively adapts to real-world ocean systems, driving mesh movements in faithful response to multiscale spatial structures and long-term temporal variability. The unsupervised model learned from ocean systems further generalizes across the global Earth system, demonstrating broad scientific applicability across 54 regions and 16 variables. Code is available at https://anonymous.4open.science/r/GEMove/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Mesh Adaptation, Mesh Movement, Scientific Machine Learning, Partial Differential Equations, AI for Science

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_VgFwOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper studies unsupervised mesh movement under a fixed mesh budget. Scalar density equidistribution constrains element size but leaves shape and orientation underdetermined, which can lead to poorly coordinated mesh movement. GEMove addresses this limitation by combining Hessian-derived density with learned orientation and anisotropy to construct a metric tensor at each vertex. It then uses a weighted Tutte equilibrium system to determine the positions of interior vertices jointly, with an inversion-free guarantee under specific geometric conditions. The model is trained by encouraging uniform element size and shape in the metric space. Experiment results suggest that combining directional geometric information with global vertex coordination can improve unsupervised mesh adaptation.

Critical strengths

The paper identifies a genuine limitation of scalar-density-based mesh movement. Scalar density alone does not determine element shape or orientation, while the proposed anisotropic metric provides additional information about directional adaptation. The use of Tutte equilibrium also couples vertex positions through a global system rather than predicting them independently. Under the stated geometric conditions, this construction prevents element inversion. Although anisotropic metrics and Tutte embeddings are established concepts, combining them in an unsupervised mesh movement framework is a useful contribution.

The experimental evaluation covers several PDE families, unseen domain geometries, and real-world scientific fields. GEMove achieves consistent improvements over the tested learning-based baselines, with particularly strong results on unseen PDEs and geometries. The experiments also examine the effects of the geometric loss components and the equilibrium decoder. Additional results on real-world ocean data and Earth-system fields provide evidence of transfer across different physical settings. Overall, the experiments are relatively comprehensive and support the effectiveness of the complete method.

Critical weaknesses

The main concern is that the experiments do not clearly establish the contribution of the learned anisotropic metric. Although the metric tensor is central to the proposed method, its scalar density is computed directly from the Hessian magnitude, and only the orientation and anisotropic strength are learned. The existing Only-Size and Only-Direction ablations change the loss function but do not isolate the benefits of learning these parameters. The comparison with UGM2N also does not fully separate the effects of the density formulation, learned anisotropy, and global coordination. Classical Hessian-based methods can construct anisotropic metrics without training, but the paper does not compare against such alternatives. It is therefore difficult to determine what the learned directional representation contributes beyond existing geometric approaches.

The role of Tutte equilibrium also needs further examination. The learned metric specifies the desired local geometry, but the final vertex positions are determined by a separate equilibrium system. The paper does not directly establish how well the resulting mesh follows the predicted orientation and anisotropy.

The evidence for broader applicability is also limited. Although the method transfers to unseen synthetic PDEs and geometries, the real-world ocean experiments require retraining, leaving its cross-domain generalization less clear. The global transfer experiments use approximately 3,000 nodes per region, and scalability to substantially larger meshes has not been tested.

Questions for authors
  1. The paper does not fully explain why the scalar density is computed directly from the Hessian while orientation and anisotropy are learned. Appendix B.1 defines ρi\rho_i using the Hessian magnitude, whereas θi\theta_i and sis_i are predicted by the network. What motivates this separation? Does fixing the density improve training stability or generalization? Would learning the density jointly with the other parameters change the results?

  2. The existing ablations do not isolate the contribution of the learned anisotropic indicators. The Only-Size and Only-Direction variants remove individual loss terms but retain the learned metric representation. Can the authors compare the full model with an isotropic metric Mi=ρiIM_i=\rho_i I while keeping the same density, geometric objective, and Tutte decoder? In addition, are the density formulations and input features sufficiently matched between GEMove and UGM2N to attribute the reported improvements to the proposed directional representation rather than other differences?

  3. The need to learn orientation and anisotropy is not established relative to classical Hessian-based metric construction. Since the Hessian eigenvectors and eigenvalues can be used to construct an anisotropic metric directly, can the authors compare this analytical construction with the learned metric under the same mesh budget and boundary conditions? In particular, how does GEMove compare with a method that obtains all three metric parameters from Hessian information rather than learning θi\theta_i and sis_i?

  4. The learned metric is used in edge-weight prediction and the geometric loss, but the final vertex positions are determined by Tutte equilibrium. Although the paper explains this connection, it does not directly measure how well the resulting mesh follows the intended directional adaptation. Could the authors quantify the alignment between the learned metric and the final mesh geometry, and examine how this relates to reconstruction accuracy? This would help determine whether the equilibrium solution preserves the anisotropy specified by the metric.

  5. The need to retrain on ocean data raises questions about the claimed generalization across physical systems. The geometric objective does not explicitly depend on a particular PDE, and the method transfers to unseen equations and geometries in the synthetic experiments. Why is retraining required for the ocean experiments? Can a model trained only on rectangular synthetic PDE domains be applied directly to ocean fields without further training? This comparison would help establish the extent of cross-physics generalization.

  6. Iterative Mesh Adaptation and Inference Efficiency. UGM2N performs multiple iterative node updates within a single r-adaptation procedure. Does GEMove similarly require multiple iterations to obtain the final adapted mesh, or can it complete the adaptation with a single network forward pass followed by one Tutte equilibrium solve?

4 comments

Thank you for recognizing the contributions of directional learning and global coordination. We answer W1 together with Q1–Q3 because all concern metric design and its controlled evaluation: why density is prescribed, whether anisotropy helps, and whether its components should be learned. Reply 2 addresses Tutte coordination (W2/Q4), transfer and scaling (W3/Q5), and inference (Q6). References below are to the revised manuscript.

W1 + Q1–Q3. Metric design and controlled comparisons

Q1. Why prescribe density but learn orientation and strength? We prescribe the construction of density, not a constant density across inputs: ρi\rho_i is recomputed from each field's graph-normalized, bounded Hessian magnitude (Appendix B.1, Eq. (30)). Our motivation is to anchor resolution allocation to field variation while reducing freedom in jointly learning the metric and relocated mesh. At each node, det⁡(Mi)=ρi2\det(M_i)=\rho_i^2; learned θi\theta_i and sis_i change orientation and stretching at this fixed determinant scale (Section 4.2, Eq. (4)). Keeping density tied to the input field gives directional learning a stable reference: the network learns how to arrange the required resolution across neighboring elements, while the amount of resolution remains grounded in the field's variation.

W1 / Q2. Isotropic control and matched inputs. Following your request, the new Isotropic variant fixes Mi=ρiIM_i=\rho_i I but retains the full geometric objective and Tutte decoder, with matched training data, inputs, and optimization (Section 5.3; Appendix D.1). The complete Table 2 ablation is summarized below (mean error reduction, %):

VariantPois.Helm.SwirlBurg.Avg.
Isotropic20.0119.959.7225.3618.76
Only-Size20.3520.037.8226.0418.56
Only-Direction19.5118.9413.3924.7119.14
w/o Coordination21.8222.1615.70*31.3922.77
GEMove23.6623.8517.5533.4024.62

An asterisk marks element inversions on Swirl; independent-run standard deviations are reported in Appendix E.4, Table 7. With density, geometric objective, and equilibrium decoder matched, learned anisotropy raises average error reduction by 5.86 percentage points (18.76% to 24.62%), improving every PDE family. Only-Size/Only-Direction separately test the objective's two terms; w/o Coordination tests replacing the equilibrium decoder with direct coordinate prediction, as further discussed in Reply 2.

For the UGM2N comparison, our implementations use the same scalar-density formulation, physical/mesh input features, initial meshes, node budgets, and boundary constraints (Appendices B.1 and D). These matched settings ensure that GEMove's gains do not rely on additional field information or a different density formulation. The isotropic control identifies the contribution of anisotropy, while the decoder ablation examines the benefit of global coordination.

W1 / Q3. Learned versus fully Hessian-derived metric. We separately trained variants in which orientation θ\theta, strength ss, or both are prescribed from the Hessian while keeping the same ρ\rho, geometric objective, Tutte solver, mesh budget, and boundary treatment (Section 5.3, Table 3). Prescribing both gives the requested fully Hessian-derived metric:

MetricPois.Helm.SwirlBurg.Avg.
Hθ,HsH_\theta,H_s12.1416.479.1014.9213.16
Hθ,LsH_\theta,L_s15.4719.3310.3125.4617.64
Lθ,HsL_\theta,H_s18.2519.6214.1227.3519.84
Lθ,LsL_\theta,L_s (GEMove)23.6623.8517.5533.4024.62

H/LH/L denotes Hessian-derived/learned; Table 3 reports per-family means and standard deviations. Learning either component improves on the fully Hessian-derived metric, and learning both performs best on every PDE family. On unseen Swirl, error reduction increases from 9.10% to 17.55%, showing that the learned directional representation also benefits adaptation beyond the training equations. Together, the isotropic and Hessian controls establish both the value of directional geometry and the additional benefit of learning it from field and neighborhood context.

Reply 1 addresses metric design (W1/Q1–Q3). Here we address Tutte coordination (W2/Q4), transfer and scaling (W3/Q5), and inference (Q6). References are to the revised manuscript.

W2 + Q4. Tutte equilibrium, geometric alignment, and accuracy

Thank you for encouraging us to examine how the final mesh realizes the learned geometry. Our original examples already illustrated this alignment in metric space; we have now added a systematic comparison across all four PDE families.

Why global coordination matters. Achieving equal-area, equilateral geometry in metric space requires coordinating a triangle's three vertices, whose movement also affects neighboring triangles. Local geometric demands must therefore be coordinated across the mesh and its boundary. GEMove predicts positive symmetric edge weights using metric-dependent lengths and contextual features, then solves all interior coordinates jointly through Tutte equilibrium (Section 4.3, Eqs. (5)–(8)). The geometric loss is evaluated on the final mesh, and gradients through the solve train the metric and edge predictor together (Appendix C.1).

Alignment with the learned metric. The original cases in Appendix E.5, Figures 9–10 show how movement brings elements toward equal areas and equilateral shapes in metric space. In Helmholtz/Poisson, mapped-area CV decreases from 0.47/0.44 to 0.22/0.20, and mean edge aspect ratio from 1.39/1.35 to 1.13/1.09, alongside reconstruction-error reductions of 16.9%/31.9%.

Following your suggestion, the new Table 8 extends this analysis to all four PDE families, comparing final-mesh alignment with and without Tutte:

VariantPoissonHelmholtzSwirlBurgers
w/o Coord.0.22 / 1.110.25 / 1.090.27 / 1.140.21 / 1.16
GEMove0.24 / 1.130.25 / 1.110.28 / 1.170.21 / 1.17

Entries are mapped-area CV / mean edge aspect ratio, with ideals 0 / 1. The variant without Tutte retains the metric construction and full objective but predicts coordinates directly (Appendix D.1).

Both variants achieve comparable alignment in metric space. Without Tutte, local statistics can be slightly better, yet the mesh exhibits excessive clustering and Swirl inversions (Figure 4). Tutte coordinates the local demands into a valid mesh while preserving comparable geometric alignment: average error reduction rises from 22.77% to 24.62%, and Swirl from 15.70% to 17.55%, without the observed inversions (Table 2). This explains its practical contribution: globally compatible vertex positions improve reconstruction even when local shape statistics are similar.

W3 + Q5. Direct transfer and larger meshes

PDE-to-real-world transfer. Following your suggestion, we added PDE-trained GEMove and UGM2N to the Kuroshio, ENSO, and 54-region evaluations (Sections 5.4–5.5; Table 4; Appendix G.4, Table 17). Both models are trained only on rectangular Poisson and Helmholtz domains and applied without parameter updates.

BenchmarkPDE-trained UGM2NPDE-trained GEMoveOcean-trained UGM2NOcean-trained GEMove
Kuroshio2.88.34.114.4
ENSO 2°12.725.327.045.3
ENSO 3°21.137.129.257.5
ENSO 4°25.842.532.268.1
54-region average1.243.531.804.09

Values are error reductions (%); ENSO uses Niño 3.4 index RMSE. PDE-trained GEMove improves 52/54 regions without ocean training, demonstrating transfer across physical systems. Ocean training provides further gains, with GEMove improving all 54 regions.

Global SST scale and cost. We also added a global NOAA OISST experiment using existing checkpoints without further training (Section 5.4; Appendix G.3, Tables 15–16):

NodesError reduction (%)Time (ms)GEMove GPU (MiB)
2,7197.29 / 3.05247.70 / 240.78946.17
4,8393.50 / 1.18275.32 / 274.131,664.68
10,9065.22 / 0.74337.46 / 343.783,755.97
43,3149.87 / 3.59756.98 / 747.8515,058.00

Paired entries are GEMove/UGM2N. Error reductions average six months; timings use three selected months on an RTX 4090 and AMD EPYC 7542, with CUDA synchronization. They cover the full adaptation pipeline, excluding reconstruction; memory is peak GPU allocation.

A 15.93-fold increase in nodes requires 3.06-fold more adaptation time over this range, with GPU memory growing approximately proportionally to node count. GEMove retains higher accuracy at every resolution and adapts the largest mesh in under a second, at nearly the same runtime as UGM2N.

Q6. One pass or iterative updates? GEMove obtains the adapted mesh through one network forward pass and one Tutte equilibrium solve (Section 4.3, Eq. (8)). It predicts the metric and edge weights once, then jointly computes interior coordinates with fixed boundaries. This global solve coordinates mesh movement in a single adaptation step.

Reviewer_VgFwOpenReview ↗

Response to author

Thanks for the detailed response. Overall, the responses have addressed my main concerns. I'm keeping my weak accept and have updated the soundness and confidence scores.

Thank you very much for your positive assessment of our response. We are delighted that our clarifications and additional evidence have addressed your main concerns, and we are particularly encouraged by your updated soundness score and increased confidence in your overall positive assessment of the paper. Your continued support is greatly appreciated and means a lot to us. We have also incorporated all the additional experiments and analyses into the latest revised manuscript to further strengthen the paper.

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

This paper proposes GEMove, an unsupervised mesh movement method that constructs local anisotropic metrics from mesh geometry and physical field features. It encourages uniform element sizes and equilateral shapes in metric space, while using a spring system under Tutte equilibrium to coordinate mesh relocation globally. Experiments on PDE benchmarks and Earth-system datasets demonstrate strong performance and generalization across equations, geometries, and physical fields.

Critical strengths

The paper presents a coherent unsupervised mesh movement framework that integrates learnable anisotropic metrics with global coordination through Tutte equilibrium and empirically demonstrates strong generalization.

Critical weaknesses
  1. The contribution of the learned anisotropic metric is not sufficiently isolated. Setting Mi=ρiIM_i=\rho_iI would still allow the full geometric objective to be used, but the paper does not report this controlled comparison. The existing loss ablations therefore do not isolate the additional benefit of learning metric orientation and anisotropy.
  2. Experimental evaluation. The accuracy–cost trade-off against MA remains unclear. The paper reports neither training time nor MA mesh-generation costs, and lacks a direct MA accuracy baseline. Although Section 5.4 explains why supervised methods are omitted from the real-world experiments, the restriction to UGM2N leaves the broader practical advantage insufficiently established. Besides, the 3∘3^\circ GEMove curve in Figure 14 appears to exceed 100% improvement. The authors should clarify the metric used, since relative reduction in a nonnegative error cannot exceed 100%.
Questions for authors
  1. Could the authors clarify what enables GEMove to generalize across datasets with different geometric and physical-field structures, and the respective roles of input normalization, the prescribed density ρ\rho, and the learned metric? Which properties of the chosen training datasets support this transfer, and would the results persist with individual datasets or alternative combinations under comparable training budgets?

  2. Prior work [1] already combines metric-based equidistribution and alignment into a unified mesh objective. Could the authors clarify how GEMove geometric principle and spectral loss relate to this formulation, and what additional benefits arise from learning the metric and using the Tutte decoder?

[1] A. Kolasinski and W. Huang. A New Functional for Variational Mesh Generation and Adaptation Based on Equidistribution and Alignment Conditions. Computers & Mathematics with Applications, 75:2044–2058, 2018.

4 comments

Thank you for your careful review. We address W1 and Q1 together because both concern the roles of GEMove's learned components and the evidence supporting their contribution to adaptation and generalization. W2 and Q2 are addressed in our companion reply. All references are to the revised manuscript.

W1. Contribution of the learned anisotropic metric

We added the requested isotropic control, Mi=ρiIM_i=\rho_i I, preserving the density, full geometric objective, Tutte decoder, input features, and matched training protocol (Section 5.3, Table 2; Appendix D.1). Mean error reductions (%) are:

VariantPois.Helm.SwirlBurg.Avg.
Isotropic20.0119.959.7225.3618.76
Only-Size20.3520.037.8226.0418.56
Only-Direction19.5118.9413.3924.7119.14
w/o Coordination21.8222.1615.70*31.3922.77
GEMove23.6623.8517.5533.4024.62
  • denotes element-inversion failures; run-to-run standard deviations appear in Appendix E.4, Table 7. GEMove improves on isotropic adaptation in all four families, raising the average from 18.76% to 24.62% (+5.86 percentage points), and Swirl from 9.72% to 17.55%. With density, objective, and coordination held fixed, this isolates the benefit of directional geometry. We also test learned versus Hessian-derived anisotropy under Q1.

Q1. Generalization, component roles, and training diversity

Generalization. Physical fields can share fronts and elongated high-gradient structures even across different governing equations. GEMove encodes their resolution demands through density, orientation, and anisotropic strength under one geometric objective. Trained only on Poisson and Helmholtz, it improves reconstruction without further training in 52/54 Earth-system regions, averaging 3.53% error reduction versus 1.24% for PDE-trained UGM2N (Appendix G.4, Table 17).

Component roles. Per-channel, per-snapshot normalization reduces sensitivity to variable scale (Appendix F.5, Eq. (69)). The prescribed ρ\rho, obtained from normalized, bounded Hessian magnitude, anchors resolution density to the input (Appendix B.1). Learned θ\theta and ss use neighboring features to control direction and anisotropy. As det⁡(Mi)=ρi2\det(M_i)=\rho_i^2, ρ\rho fixes the determinant scale while θ,s\theta,s control the metric's orientation and shape.

To quantify the benefit of learning beyond a Hessian-derived metric, we prescribed θ\theta, ss, or both directly from the Hessian and trained each configuration separately under the same density, objective, and Tutte solver (Section 5.3, Table 3):

MetricPois.Helm.SwirlBurg.
Hθ,HsH_\theta,H_s12.1416.479.1014.92
Hθ,LsH_\theta,L_s15.4719.3310.3125.46
Lθ,HsL_\theta,H_s18.2519.6214.1227.35
Lθ,LsL_\theta,L_s (GEMove)23.6623.8517.5533.40

H/LH/L denotes Hessian-derived/learned components; values are mean error reductions (%). Learning both raises the average from 13.16% (fully Hessian-derived) to 24.62%, outperforming every fixed-component variant on all four families.

Training diversity. The default Poisson–Helmholtz training solutions contain 1–5 anisotropic Gaussians on multiresolution unstructured meshes (Appendix D). Following your suggestion, we added single-family and alternative-pair training with the same total sample count, training budget, and held-out test set (Appendix E.6, Table 9).

Training dataPois.Helm.SwirlBurg.Avg.
Poisson25.4321.4213.6331.2122.92
Helmholtz21.3524.1215.0329.3122.45
Swirl16.2514.7316.0125.5618.14
Burgers20.3819.1312.6634.1721.59
Poisson + Helmholtz (default)23.6623.8517.5533.4024.62
Poisson + Swirl24.1421.2716.9330.0223.09
Helmholtz + Burgers24.5925.0317.1133.1024.96

Values are mean error reductions (%). Every configuration improves all four test families, including unseen ones, and each pair outperforms its constituent single-family configurations on average. Helmholtz–Burgers achieves the highest average (24.96%), while the default pair gives the strongest Swirl result (17.55%) without training on Swirl. These results show that transfer extends beyond the original training pair and support the benefit of geometric diversity under a fixed training budget.

We address W2 and Q2 below. All references are to the revised manuscript.

W2. MA comparisons, accuracy–cost trade-off, and Figure 14

Synthetic PDEs and unseen geometries. We added direct MA comparisons (Section 5.2, Table 1; Appendices E.2–E.3, Tables 5–6).

BenchmarkMAGEMove
Poisson19.3223.66
Helmholtz17.5423.85
Swirl5.8917.55
Burgers28.1333.40
PDE average17.7224.62
Unseen polygons (avg.)3.4810.33

Values are mean error reductions (%). GEMove outperforms MA on all PDEs and unseen polygons, while avoiding MA's element inversions on Swirl and polygonal domains.

Real-world baselines and mesh validity. We extended comparisons to MA and PDE-trained UGM2N/GEMove on ocean data and 54 Earth-system regions (Sections 5.4–5.5; Table 4; Appendix G.4, Table 17). For Kuroshio (182 dates) and ENSO (60 months), which feature complex coastal and island boundaries, MA produces invalid meshes at all four resolutions despite matched initial nodes, connectivity, and fixed boundary vertices; inverted elements prevent reconstruction. Appendix F.4, Table 12 documents the solver settings and failures. Even a quasi-Newton test converges in residual but retains 101 inverted elements, confirming that convergence alone does not ensure mesh validity. GEMove successfully adapts these meshes. MA does succeed in some simpler open-ocean and atmospheric cases: North/South Pacific wind gains are 7.29%/5.56%, versus 12.79%/6.94% for GEMove.

Across 54 regions, GEMove improves every reconstruction (4.09% average). Without ocean training, PDE-trained GEMove improves 52/54 regions (3.53% average), versus 1.24% for PDE-trained UGM2N. The four field-group comparisons are:

Field groupMethodTime (ms)Error reduction (%)
OceanographicMA3266.94-0.33
UGM2N53.560.97
GEMove49.493.18
AtmosphericMA3271.892.39
UGM2N54.593.98
GEMove52.706.01
BiogeochemicalMA3553.40-0.32
UGM2N95.73-0.17
GEMove86.512.53
EnvironmentalMA2925.383.12
UGM2N122.241.05
GEMove91.623.61

Error reductions are regional averages (Appendix G.4, Table 17). MA averages exclude invalid outputs and include only successful cases for partial failures; learned methods cover all regions. Timings follow Appendix E.7, Table 11. GEMove leads in accuracy and adaptation time in all four groups.

Accuracy and cost. We now report synchronized PDE inference (including the Tutte solve), peak memory, and offline training cost (Appendix E.7, Tables 10–11).

MethodTime (ms)Peak GPU (MiB)
MA (CPU)5441.95—
UGM2N (GPU)24.97281.09
GEMove (GPU)25.11290.34

GEMove's inference cost is close to UGM2N; MA timings are not hardware-matched. A 1,000-epoch PDE run takes 7.45 h on an RTX 3090 (8.97 h including validation, checkpoints, and logging; 15.68 GiB peak allocated GPU memory); ocean training takes 8.73 h over 475 epochs. MA requires no training, whereas GEMove needs no target meshes and reuses its checkpoint. On 43,314-node global SST, GEMove achieves 9.87% error reduction versus 3.59% for UGM2N and 0.13% for kriging, with latency comparable to UGM2N (756.98 vs. 747.85 ms; Appendix G.3, Tables 15–16).

Figure 14 correction. We mistakenly used adapted-mesh rather than original-mesh error as the denominator. The corrected definition is

ErrorRed=100Eorig−EmethodEorig%.\mathrm{ErrorRed}=100\frac{E_{\mathrm{orig}}-E_{\mathrm{method}}}{E_{\mathrm{orig}}}\%.

We corrected Figure 14, checked other results, and specified the aggregate ENSO metric in Appendix F.3, Eq. (63). We apologize for the error.

Q2. Relation to Kolasinski and Huang (2018)

We added this reference to Section 2. Their formulation enforces equidistribution and alignment for a prescribed metric; GEMove learns directional metrics and a reusable mesh relocation rule without target meshes. Our spectral loss decomposes as

=2\left(\log\frac{a_K^M}{\alpha}\right)^2 +\frac12\left(\log\frac{\lambda_{K,1}}{\lambda_{K,2}}\right)^2,$$ where $\lambda_{K,r}$ are eigenvalues of $\bar G_K=G_K/\alpha$, $a_K^M=\sqrt{\det G_K}$, and $\alpha$ is the geometric mean of metric element areas. The terms penalize area variation and unequal metric-space stretching. Unlike their metric-volume-weighted inverse-Jacobian functional, our log penalty uses geometric-mean normalization and trains the metric and relocation jointly on the final mesh. GEMove predicts positive edge weights and determines vertex positions with one differentiable Tutte solve rather than a moving-mesh PDE. Tutte raises average error reduction from 22.77% to 24.62% and avoids Swirl inversions (Table 2; Figure 4).

Thank you for your clarification. I have updated the score accordingly.

Thank you very much for your positive assessment of our response and for reconsidering your evaluation. We are delighted that our clarifications and additional evidence have addressed your concerns and encouraged you to raise your score. Your recognition of these improvements and constructive feedback are greatly appreciated and mean a lot to us. We have also incorporated all the additional experiments and analyses into the latest revised manuscript to further strengthen the paper.

Reviewer_2MWKOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

GEMove is an unsupervised technique of moving vertices in a two-dimensional mesh for solving PDEs on this mesh. It is able to learn directional geometric information and uses a global Tutte equilibrium to coordinate vertex positions and avoid mesh singularities. The paper reports an average error reduction of around 25% across four PDE families and outperforms the closest competitor UGM2N together with the other baselines. The generalisability is tested on unseen equations different domain shapes and close-to-real Earth-system fields.

Critical strengths
  1. The method accounts for density and direction of the mesh triangles and also for their regular placement using Tutte equilibrium
  2. Unsupervised training without target meshes
  3. Rather good generalisability by experiments
Critical weaknesses
  1. There should be an ablation for directional metric
  2. It would be nice to see a coarse-to-fine interpolation baseline.
  3. The sizes of meshes around 40 000 elements are still inpractical for most of the real problems
Questions for authors

Q1. How will GEMove behave in comparison with an isotropic metric or a fixed Hessian-based metric using the same solver? Q2. How does GEMove compare with direct interpolation from the same coarse field? Q3. Can you report synchronized runtime, peak memory use and scaling with mesh size? Q4. How the real computational costs are compared for MA and GEMove considering different CPU/GPU hardware?

4 comments

Thank you for your constructive review and for recognizing GEMove's directional adaptation and generalization. In response, we added controlled metric comparisons, clarified the unchanged-mesh interpolation baseline, and expanded the global SST, runtime, and memory evaluations. We address your comments below; all references are to the revised manuscript.

W1. Isolating the contribution of the learned directional metric

We added an isotropic control, Mi=ρiIM_i=\rho_i I, in Section 5.3, Table 2, retaining the same density, full geometric objective, and Tutte decoder to isolate the effect of metric anisotropy.

ModelPoissonHelmholtzSwirlBurgersAverage
Isotropic20.0119.959.7225.3618.76
Only-Size20.3520.037.8226.0418.56
Only-Direction19.5118.9413.3924.7119.14
w/o Coordination21.8222.1615.70*31.3922.77
GEMove23.6623.8517.5533.4024.62

Values are mean error reductions (%); * indicates element-inversion failures. Standard deviations across independent runs are in Appendix E.4, Table 7.

GEMove improves average error reduction from 18.76% (isotropic) to 24.62%, outperforming the isotropic control on all four PDEs. This isolates the benefit of learned directional structure under the same objective and decoder. We also added comparisons with Hessian-derived orientation and anisotropic strength (Section 5.3, Table 3), which further support learning these components rather than prescribing them analytically; see our response to Q1.

W2. Coarse-to-fine interpolation in the global reconstruction benchmark

Direct interpolation from the same coarse field was already the baseline in our real-world experiments, but this was insufficiently explicit in the main text. We clarified in Section 5.1 (protocol in Appendix F.5) that the Original baseline keeps the coarse mesh unchanged and interpolates to fine-grid locations. GEMove uses the same coarse observations and node budget; relocated-node values are interpolated from the same coarse field, and the fine-resolution reference is used only for evaluation. Thus, the reported gains measure node redistribution. In the 54-region global transfer benchmark, GEMove achieves 4.09% average error reduction versus 1.80% for UGM2N relative to this baseline (Section 5.5).

We additionally introduced a global SST reconstruction benchmark using NOAA OISST v2.1 and comparing with kriging interpolation (Section 5.4; Appendix G.3, Table 15). It covers six monthly SST-anomaly fields (October 2015–March 2016), four resolutions (2,719–43,314 ocean nodes), and the native 0.25∘0.25^\circ evaluation reference. Models are tested without retraining; Regular denotes conventional interpolation on the unchanged mesh.

Method1∘1^\circ (43,314 nodes)2∘2^\circ (10,906 nodes)3∘3^\circ (4,839 nodes)4∘4^\circ (2,719 nodes)
Kriging interpolation0.130.170.150.21
UGM2N3.59±1.223.59\pm1.220.74±1.510.74\pm1.511.18±1.421.18\pm1.423.05±1.943.05\pm1.94
GEMove9.87±0.569.87\pm0.565.22±0.845.22\pm0.843.50±1.023.50\pm1.027.29±1.817.29\pm1.81

Values are percentage error reductions relative to Regular; reported standard deviations are across six months.

GEMove outperforms both UGM2N and kriging at every resolution. At 43,314 nodes, it achieves 9.87% error reduction versus 3.59% for UGM2N and 0.13% for kriging, confirming that adaptive node placement offers benefits beyond changing the interpolation method.

Q1. Isotropic and fixed Hessian-derived metrics with the same equilibrium solver

Thank you for this helpful suggestion. The isotropic comparison under W1 shows the benefit of directional information. To examine whether learning this information improves on deriving it directly from the Hessian, we have added a controlled metric-component comparison in Section 5.3.

Since ρi\rho_i is already Hessian-derived, we replace the orientation θi\theta_i, anisotropic strength sis_i, or both with Hessian-derived values, retaining the same geometric objective and Tutte solver. Each variant is trained separately, with its designated Hessian components prescribed from the input field and its remaining parameters trained for that configuration. Replacing both θi\theta_i and sis_i gives the requested fully Hessian-derived metric.

MetricPois.Helm.SwirlBurg.
Hθ,HsH_\theta,H_s(Hessian-derived)12.1416.479.1014.92
Hθ,LsH_\theta,L_s15.4719.3310.3125.46
Lθ,HsL_\theta,H_s18.2519.6214.1227.35
Lθ,LsL_\theta,L_s (GEMove)23.6623.8517.5533.40

HH and LL denote Hessian-derived and learned components, respectively. Values are mean error reductions; standard deviations are provided in Table 3.

Both learned components contribute. Starting from the fully Hessian-derived metric, learning strength raises average error reduction from 13.16% to 17.64%, while learning orientation raises it to 19.84%. Learning both achieves 24.62% and performs best on every PDE family.

These results support our design: Hessian provides a geometric starting point, while learning uses field and neighborhood context to refine directional allocation. Gains under the same solver justify both components.

Q2. Direct interpolation from the same coarse field

Thank you for emphasizing this comparison. As detailed under W2, the original benchmark already measures improvement over unchanged-mesh interpolation from the same coarse field (Appendix F.5), now clarified in Section 5.1. The new global SST experiment further shows that GEMove outperforms conventional interpolation and kriging at all four resolutions (Appendix G.3, Table 15).

Q3. Synchronized runtime, peak memory, and scaling with mesh size

Thank you for this suggestion. Inference times were already reported in the appendix. We now discuss efficiency in Section 5.2 and expand Appendix E.7 (Tables 10–11) with runtime and peak GPU/CPU memory on PDE and real-world benchmarks. We also add synchronized timing and memory measurements across four global SST mesh sizes (Section 5.4; Appendix G.3, Table 16), clarifying how costs scale.

On the PDE benchmarks, GEMove's complete forward procedure, including the Tutte equilibrium solve, takes 25.11 ms, compared with 24.97 ms for UGM2N (Table 10). Peak GPU allocations are 290.34 and 281.09 MiB, respectively, and peak CPU process memory is 1734.72 and 1678.41 MiB. Thus, the learned directional metric and equilibrium decoder provide the accuracy gains with only a small measured difference in inference cost. On the four real-world Earth-system groups, GEMove's mean adaptation times range from 49.49 to 91.62 ms and are lower than UGM2N's in every group; Table 11 reports the corresponding runtime and memory statistics.

The new global experiment further tests whether this remains practical beyond the regional meshes. GEMove's measurements in Table 16 are:

NodesAdaptation time (ms)Peak GPU (MiB)Peak CPU (MiB)
2,719247.70±2.77247.70\pm2.77946.171548.09
4,839275.32±5.32275.32\pm5.321664.681550.16
10,906337.46±7.60337.46\pm7.603755.971552.16
43,314756.98±18.67756.98\pm18.6715058.001567.11

These measurements use an RTX 4090 and AMD EPYC 7542 CPU, with CUDA synchronization before and after timing. They include preprocessing, data transfers, internal solves, and boundary processing for mesh adaptation. Peak GPU memory is allocated PyTorch memory, and peak CPU memory is process RSS. Table 16 gives the repeated-measurement protocol and full standard deviations.

Over this range, a 15.93-fold increase in node count requires a 3.06-fold increase in runtime; GPU memory grows approximately proportionally to the node count. At 43,314 nodes, GEMove takes 756.98 ms versus 747.85 ms for UGM2N, with nearly identical peak GPU allocations. At that same size, GEMove achieves 9.87% error reduction versus 3.59% for UGM2N. Together, these results show that the improvement in reconstruction is retained at global scale without a substantial increase in measured inference cost relative to the closest baseline.

Reviewer_2MWKOpenReview ↗

I race my grade to clear accept however some questions remain

Firstly I would like to show my appreciation for a great amount of additional experiments you made.

After some doubts I decided to increase my grade of your work although there are still some concerns related to practical use of your results and to some of your comparisons.

First, the 40 000 triangles is still a very small mesh when talking about real practical cases (even in 2D is like 200x200). So I would recommend you to think how the algorithm may be optimised in order to extend at list to 1M cells.

Second, some of your computational cost comparisons are made on different hardware (CPU vs GPU) and you better somehow normalise them.

Third, Table 9 in E6 shows that there is a strong influence of training PDEs on the results on the same or different families. That gives motivation to compare your results to the approaches you called "mesh adaptation into PDE-solving pipelines". Probably I would also recommend for you to add to consideration a paper "Ryabov, Alexander, et al. "Learnable stability-aware computational grid coarsening for accelerating physics simulations." Machine Learning: Science and Technology 7.1 (2026): 015019" as it also uses some kind of stability loss that can lead to the same effect as your Tutte equilibrium restrictions. And also this table might be extended to the model trained an all the PDE families, its performance will be particularly interesting.

Thank you very much for your thoughtful reassessment and for recognizing the extensive additional experiments we have conducted. We are genuinely delighted and greatly encouraged by your decision to raise your score. We especially appreciate the care and attention you have devoted to reviewing our work, as well as your constructive suggestions for further improving its practical relevance. Your recognition and support mean a great deal to us and provide valuable encouragement for our continued efforts to strengthen the paper.

Regarding scalability, thank you for highlighting the importance of evaluating much larger meshes. Due to time and computational resource constraints, our experiment with 40,000 triangles represents only a preliminary exploration of large-scale mesh adaptation. We fully agree that extending the framework toward one million cells is an important direction for practical applications. We will further investigate computational and memory optimizations to improve scalability and clarify this limitation and future direction in the manuscript.

Regarding computational cost, thank you for pointing out the potential limitations of comparisons across different hardware. We agree that CPU and GPU timings are not directly comparable, and your suggestion motivates us to reconsider how computational efficiency should be evaluated and presented more fairly. We will carefully revisit these comparisons, clarify the hardware configurations and measurement settings, and refine the corresponding discussion to ensure that our efficiency claims are appropriately supported.

Your observation about Table 9 is particularly valuable. We agree that the training PDE distribution influences cross-family generalization, and that this dependence deserves explicit discussion. Training a unified model on all four PDE families is an especially interesting suggestion, as it could help distinguish limitations of individual-family training from the potential of joint learning. We will investigate this setting and further clarify the scope of our generalization claims.

We also sincerely appreciate your recommendation of Ryabov et al. (2026). Their self-supervised, stability-aware grid coarsening framework is highly relevant to our discussion. We find your observation about the potential connection between their stability loss and our Tutte equilibrium particularly insightful. While both approaches may help prevent undesirable mesh configurations, they consider this issue from complementary perspectives: their approach incorporates numerical stability and simulation fidelity into learnable grid coarsening, whereas GEMove focuses on globally coordinated geometric adaptation under a fixed vertex budget. We will incorporate this reference and further discuss these connections and distinctions, as well as potential opportunities for integrating geometric adaptation with PDE-solving pipelines.

Thank you again for your careful reading, constructive feedback, and increased confidence in our work. Your suggestions have helped us identify important directions for strengthening both the current manuscript and its practical applicability. We have also incorporated the additional experiments and analyses from our rebuttal into the latest revised manuscript.

Comments on the paper

1 comment

General Response by Authors

Dear Program Chairs, (Senior) Area Chairs and Reviewers,

We thank all reviewers for their feedback. We highlight shared strengths and address two common concerns below; individual replies cover remaining points. All references are to the revised manuscript.

Reviewer Acknowledgment

We appreciate the reviewers' recognition of GEMove in four aspects:

1. Motivation: Beyond scalar-density adaptation.

  • R1: "The method controls element shape and orientation in addition to density."
  • R2: "It encourages uniform element sizes and equilateral shapes in metric space"
  • R3: "The paper identifies a genuine limitation of scalar-density-based mesh movement."

2. Method: Unsupervised directional learning and global coordination.

  • R1: "Training does not require target adapted meshes."
  • R2: "The paper presents a coherent unsupervised mesh movement framework that integrates learnable anisotropic metrics with global coordination through Tutte equilibrium"
  • R3: "combining them in an unsupervised mesh movement framework is a useful contribution."

3. Results: Consistent gains and broad evaluation.

  • R1: "The paper reports an average error reduction of around 25% across four PDE families and outperforms the closest competitor UGM2N."
  • R2: "Experiments on PDE benchmarks and Earth-system datasets demonstrate strong performance and generalization across equations, geometries, and physical fields."
  • R3: "GEMove achieves consistent improvements over the tested learning-based baselines, with particularly strong results on unseen PDEs and geometries."

4. Generalization: Transfer across equations, geometries, and scientific fields.

  • R1: "The experiments show generalisability to unseen equations and domain shapes."
  • R2: "empirically demonstrates strong generalization."
  • R3: "Additional results on real-world ocean data and Earth-system fields provide evidence of transfer across different physical settings."

Common Questions

Q1. What does the learned directional metric contribute? (R1, R2, R3)

We added an isotropic control (Mi=ρiIM_i=\rho_i I), retaining density, objective, Tutte decoder, and training protocol (Section 5.3, Table 2; Appendix D.1).

ModelPois.Helm.SwirlBurg.Avg.
Isotropic (new)20.0119.959.7225.3618.76
Only-Size20.3520.037.8226.0418.56
Only-Direction19.5118.9413.3924.7119.14
w/o Coordination21.8222.1615.70*31.3922.77
GEMove23.6623.8517.5533.4024.62

Mean error reductions (%); * flags element inversion. Std.: Appendix E.4, Table 7.

We also compare separately trained variants with Hessian-derived (H) or learned (L) orientation and strength, with the same density, objective, and Tutte solver (Section 5.3, Table 3).

MetricPois.Helm.SwirlBurg.
Hθ,HsH_\theta,H_s12.1416.479.1014.92
Hθ,LsH_\theta,L_s15.4719.3310.3125.46
Lθ,HsL_\theta,H_s18.2519.6214.1227.35
Lθ,LsL_\theta,L_s (GEMove)23.6623.8517.5533.40

Mean error reduction (%).

GEMove improves the average from 18.76% (isotropic) and 13.16% (fully Hessian-derived) to 24.62%, outperforming both across all four PDE families. The decoder ablation also confirms better accuracy and avoids the Swirl inversions seen without coordination (Figure 4).

Q2. Computational cost and scalability (R1, R2, R3)

We now report training cost, synchronized inference, memory, and mesh-size scaling (Appendices E.7 and G.3). PDE inference includes the network and Tutte solve (Appendix E.7, Table 10).

MethodTime (ms)Peak GPU (MiB)
MA (CPU)5441.95—
UGM2N (GPU)24.97281.09
GEMove (GPU)25.11290.34

GEMove has inference cost comparable to UGM2N. Training for 1,000 epochs takes 7.45 h (8.97 h including validation and logging). On a new 43,314-node SST case, GEMove achieves 9.87% versus 3.59% error reduction for UGM2N, with similar adaptation times (756.98 vs. 747.85 ms; Appendix G.3, Tables 15–16).

We include direct MA and interpolation baselines, training-data composition tests, and PDE-to-ocean transfer. Without retraining, PDE-trained GEMove improves 52/54 Earth-system regions over the unchanged mesh (Appendix G.4, Table 17). See individual replies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.