acceptodds
Under review as a conference paper at ICLR 2027

PACE: A Physics-Aware Mesh Coarsening Framework based on Error Equidistribution

Abstract

Accurate mesh-based simulation relies on sufficient mesh resolution. However, increasing the global mesh resolution directly raises computational cost, making the balance between physical fidelity and efficiency a fundamental bottleneck in numerical simulation. While careful artificial mesh layout or adaptive mesh coarsening can mitigate this burden by eliminating redundant degrees of freedom, existing strategies often rely on costly iterative optimization. We introduce PACE, a physics-aware mesh coarsening framework based on error equidistribution. PACE designs a graph neural network for efficient edge-collapse selection that couples a coarse physical field with mesh geometric features to predict an edge-wise coarsening score map. PACE then realizes the predicted score map through topology-constrained parallel edge collapse to produce the final coarse mesh. Without assuming a uniquely defined target coarse mesh, PACE formulates error equidistribution as a self-supervised objective for learning how limited mesh resolution should be allocated according to local physical importance. We evaluate on multiple cases covering steady and transient regimes, and observe consistent gains in accuracy and efficiency compared to traditional mesh coarsening baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Network, Mesh coarsening, Deep Learning, Self-supervised learning

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

The paper presents PACE, a physics-aware mesh coarsening framework. A gnn predicts edge-wise coarsening scores and a topology-constrained parallel edge-collapse procedure applies the predicted coarsening. The training objective is motivated by error equidistribution. The paper evaluates the method on steady and transient PDE problems, including flow cases, and reports solution errors, rollout RMSE, coarsening ratios, and runtime comparisons against learned and QEM-based baselines.

Critical strengths
  • The paper addresses a practical problem: reducing mesh size while preserving the accuracy of downstream PDE simulations. Incorporating physical-field information into edge selection is a reasonable direction.
  • The evaluation covers multiple PDE settings and reports both numerical accuracy and computational efficiency. The comparisons with learned coarsening baselines and QEM variants provide some evidence that the proposed approach can be useful.
Critical weaknesses
  • The form is motivated by a general interpolation-error estimate, but the paper does not show that this proxy tracks actual post-coarsening error or report its correlation with PDE solution error. The value and selection procedure for gamma are also unclear.

  • The paper reports 60 training samples across four PDEs, but does not provide the sample count for each PDE or explain how samples are split by equation, initial condition, and mesh instance.

  • QEM baselines are evaluated only on steady-state problems. In the large-scale CFD experiments, MeshGraphNet does not reach the target coarsening ratio, so the main comparison is against uniform coarsening. It is therefore unclear whether PACE remains better than other learned methods at the same mesh size.

  • The runtime results are ambiguously reported and strongly scale-dependent. Table 2 reports only “10k-level” and “1k-level” meshes, without exact vertex or element counts. One row labels both meshes “10k-level,” despite a nearly twofold difference in solve time. The total speedup also ranges from 1.07× at the 10k scale to 18.75× at the 100k scale. Exact mesh sizes and per-seed results are needed to assess these claims.

  • Text and labels in several figures are too small to read, which makes the results difficult to interpret, please improve the figure layouts and increase the font sizes to make all text and labels legible

Questions for authors
  • Could the authors clarify the theoretical connection between the error proxy and the actual post-coarsening error, and provide sensitivity analyses for key hyperparameters,?
  • Could the authors report the sample counts and data splits for each PDE, as well as the exact numbers of vertices and elements before and after coarsening in Table 2?
3 comments
AuthorsOpenReview ↗

Author response to reviewer pNtm

We thank the reviewer for the thoughtful feedback. Below, we provide our responses to the weaknesses (W1 to W3) and questions (Q1) raised in the review.

Response to W1 and Q1

Thank you for your comment. We agree that the interpolation-based motivation alone does not establish whether the proxy accurately reflects the error after coarsening. To clarify this connection, we have added Appendix B.3 to explain the connection using a linear coercive elliptic problem. For conforming piecewise-linear finite elements, under standard solution regularity and mesh regularity assumptions, Cea's lemma and interpolation estimates bound the coarse-mesh solution error through the interpolation accuracy of the coarse mesh. The discrepancy from the fine-mesh reference further satisfies the triangle inequality:

\leq \lVert u-u_h\rVert_{H^1(\Omega)} + \lVert u-u_H\rVert_{H^1(\Omega)}.$$ Here, $u$ is the exact solution on the computational domain $\Omega$, while $u_h$ and $u_H$ denote the conforming piecewise-linear finite element solutions on the fine and coarsened meshes, respectively. The notation $\lVert\cdot\rVert_{H^1(\Omega)}$ denotes the standard $H^1$ norm. The local gradient interpolation error depends on the element diameter and the local second-order variation of the solution. PACE defines the target edge length as $\hat l_e=s_e\bar l_e$, where $\bar l_e$ is the local reference length and $s_e$ is the predicted size factor for edge $e$. If the realized element diameter $h_K$ is bounded by a constant multiple of the target length, the local interpolation estimate gives $$\lvert u-I_Hu\rvert_{H^1(K)} \leq C_{\mathrm{int}}c_l \left( \bar l_e \lvert u\rvert_{H^2(K)} \right)s_e.$$ Here, $I_Hu$ is the piecewise-linear interpolant on the coarse mesh, $K$ denotes a coarse element, and $\lvert u\rvert_{H^2(K)}$ measures the local second-order variation. The constants $C_{\mathrm{int}}$ and $c_l$ are independent of the element size and predicted size factors, respectively, under the stated assumptions. This bound relates the local approximation error to the solution variation and the size factor, providing an interpolation-based motivation for the proxy $\widetilde E_e=\eta_e s_e^{\gamma}$. Here, $\eta_e$ is the normalized local physical indicator combining first- and second-order variations, and $\gamma$ controls the influence of the predicted size factor. # Response to W2 Thank you for your insightful comment. The 60 samples constitute the complete dataset used for model development, with 36 samples for training, 12 for validation, and 12 for internal evaluation. We have added the per-PDE breakdown in Appendix D.1, as shown below. | Governing equation | Total | Training | Validation | Internal evaluation | |:---|---:|---:|---:|---:| | Burgers | 12 | 8 | 2 | 2 | | Pressure diffusion | 12 | 8 | 2 | 2 | | Scalar advection-diffusion | 18 | 10 | 4 | 4 | | Vector wave | 18 | 10 | 4 | 4 | | Total | 60 | 36 | 12 | 12 | For each PDE, we split the samples by physical case. All temporal states and mesh instances associated with the same initial condition are assigned to the same subset. # Response to W3 Thanks for your comment. Classical QEM evaluates geometric distortion at a given mesh configuration through quadric error metrics. Extending this formulation to transient problems requires additional rules for aggregating physical information across time and incorporating it into the collapse cost. Therefore, We only evaluate the QEM baselines on steady-state problems, where the spatial physical field is used to guide coarsening. The large-scale evaluation also includes a comparison with MeshGraphNet in the FX63137 airfoil experiment at an angle of attack of 15 degrees, as shown in Figure 4(b). The pressure-error curves show that PACE maintains lower relative $L_2$ errors than MeshGraphNet and uniform coarsening across the saved steady-solver iterations after initialization. At the final saved iteration, the errors are approximately 12% for PACE, 83% for MeshGraphNet, and 69% for uniform coarsening, as estimated from the plotted curves. The mesh visualization further shows that PACE preserves finer resolution near the airfoil while applying stronger coarsening in the far field. The timing results in Figure 4(b) show that PACE requires 20 coarsening rounds, compared with approximately 53 for MeshGraphNet, and achieves a lower recorded coarsening time.
AuthorsOpenReview ↗

Author response to reviewer pNtm

We thank the reviewer for the thoughtful feedback. Below, we provide our responses to the weaknesses (W4 to W5) and questions (Q2) raised in the review.

Response to W4

Thanks for your comment. We have revised Table 2 to report the exact numbers of triangular elements before and after coarsening. The updated results are reproduced below.

Dense elems.Coarse elems.Coarsening (s)Dense solve (s)Coarse solve (s)Solve speedupTotal speedup
14,5168,71511.5627.6814.261.94×1.07×
25,88615,56718.1555.5128.631.94×1.19×
173,20810,315102.833,250.28147.5222.03×12.98×
248,9245,446142.404,241.9283.8450.60×18.75×

We have clarified the timing definitions in Appendix E.3. The total speedup is calculated as the dense-mesh solve time divided by the sum of the coarsening time and coarse-mesh solve time. The pre-solve time is reported separately, as the calculation assumes that the coarse input trajectory is available.

Response to W5

Thanks for your comment. We have revised the layout of all figures and enlarged the font size of text and labels to ensure readability. Please refer to the revised manuscript.

Response to Q2

Thank you for this suggestion. The per-PDE sample counts and the 36/12/12 split are added in Appendix D.1 and detailed in Response to W2. The case-level grouping preserves temporal frames, paired solutions and mesh instances within each subset.

The current Table 2 (detailed in Response to W4) lists exact triangular element numbers before and after coarsening for all four configurations, along with uniform coarse-mesh solve time and PACE timing results. We have clarified the statistical aggregation of timing results in the table caption.

Thank you for the response. However, I believe the paper still has the following issues:

  1. You use an interpolation-based proxy to drive the training objective. However, the rebuttal does not empirically demonstrate that this proxy correlates with the actual post-coarsening PDE solution error. Furthermore, the requested sensitivity analysis for the hyperparameter is missing.
  2. In the large-scale airfoil experiment, MeshGraphNet fails to reach the same target coarsening ratio as PACE. Comparing errors at different mesh sizes is unfair.
  3. Due to the significant differences in coarsening ratios across configurations, the runtime speedups vary drastically. Although the authors updated the exact element counts in their response, a comparison at a fixed coarsening ratio is still missing.

Although the authors added an appendix using Cea's lemma to justify the theoretical connection, it only covers linear coercive elliptic problems with conforming P1 elements and does not fully cover the experimental cases in this paper. The lack of empirical correlation between the proxy and actual error remains a critical flaw.

Given that the theoretical justification and the mentioned experimental issues remain unresolved, I maintain my rating of weak reject.

Reviewer_GfUKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PACE learns an edge-wise size field for coarsening a mesh before numerical PDE simulation. A GNN combines mesh geometry with an interpolated coarse physical trajectory, and a constrained parallel edge-collapse procedure realizes the predicted size field. Training balances a gradient/Laplacian-based equidistribution objective, a teacher term computed from dense solutions, and a coarsening penalty. Experiments cover several canonical PDEs, unseen domains, and airfoil and cylinder-flow cases. The reported benefits are lower solution error at comparable mesh reduction and reduced coarsening and solve times.

Critical strengths
  1. The method produces an explicit mesh for a conventional solver, and the evaluation measures the resulting numerical solution. This is a useful design choice: the accuracy gains are not confounded by a learned simulator. The vector reaction-diffusion results in Table 1 are particularly encouraging. For the second case, rollout RMSE falls from 45.73 to 19.39 in the table's units relative to MeshCNN, at essentially the same solving speedup.

  2. The implementation addresses mesh validity rather than treating coarsening as unconstrained graph pooling. Appendix F.2 reports no invalid meshes with the full constraints and includes a serial-versus-parallel comparison using identical edge priorities. Those experiments help separate the benefit of the collapse implementation from the predictor and make the proposed workflow more credible for numerical use.

Critical weaknesses
  1. The collapse rule as written does not explain how the reported reductions are obtained. Section 3.3 says that the decoder ends in a sigmoid, while Eq. (2) sets the target length to s_e times the local reference length. Thus the target is smaller than the reference. On a mesh with equal edge lengths, the candidate test in Algorithm 1 admits no edges; Eq. (36), with the additional factor 0.8, is stricter still. Is there an omitted rescaling or a target-ratio-dependent multiplier? Algorithm 1 also uses an OR in its stopping condition, which can continue with no candidates if the target ratio is unmet. These are central details, especially given the reductions from roughly 100k elements to 1k. Without the actual mapping and termination logic, I don't fully understand the inference procedure.

  2. The experiments do not cleanly establish how much of the end-to-end advantage comes from learning. The inputs include a coarse physical sequence, and Appendix E.1 normalizes over the complete valid sequence, but the total-time expression includes only coarsening and the final coarse solve. It is unclear whether producing the input trajectory is included in coarsening time or omitted. Separately, the paper already derives an analytic size factor from physical importance in Eq. (48), but evaluates it through QEM rather than through PACE's parallel collapse operator. Comparing learned and analytic size fields using the same operator, input field, and final element count would directly test the learned contribution. The current QEM timings mix predictor quality with implementation differences, and Table 2's largest speedups are not accompanied by numerical errors in that table.

  3. The qualitative evidence for the transient simulations is also limited to static figures; no videos or other animated results are included in the supplied material. For a method evaluated on evolving physical fields, side-by-side animations of the dense reference, PACE, and baselines would help assess whether moving structures are preserved throughout the rollout, and whether localized artifacts or phase errors are obscured by selected snapshots and aggregate error curves. This is an important gap in the evidence for temporal fidelity, particularly for the cylinder-flow example.

Questions for authors
  1. Clarification: What is the exact range of s_e after all transformations, how is the requested reduction ratio enforced, and what are the actual candidate and termination tests? Please explain the equal-edge-length case above.

  2. Clarification: Which time frames of the coarse solution are available at inference, is one mesh constructed for the whole trajectory, and where is the cost of obtaining that trajectory included? Please give the full timing breakdown and associated solution errors for Table 2.

  3. Assessment-changing control: How does the analytic field in Eq. (48) perform when passed through the same parallel collapse code at matched final element counts? This would distinguish the value of the GNN from that of the coarsening implementation.

  4. Supporting evidence: Can you provide synchronized animations of representative transient simulations, especially cylinder flow, comparing the dense reference, PACE, and baselines at matched coarsening ratios, with consistent color scales and error fields? These would help assess temporal fidelity beyond the static examples.

3 comments
AuthorsOpenReview ↗

Author response to reviewer GfUK

We thank the reviewer for the thoughtful feedback. Below, we provide our responses to the weaknesses (W1 to W3) and questions (Q1 to Q4) raised in the review.

Response to W1 and Q1

Thank you for pointing out the missing output transformation. The sigmoid output is rescaled by an affine transformation before it is used as the size factor. Let zez_e denote the output after the sigmoid. The mapping added in Section 3.3 is:

se=0.5+2.5ze,0<ze<1.s_e = 0.5 + 2.5z_e, \qquad 0 < z_e < 1.

The predicted size factor therefore lies in (0.5,3)(0.5,3). It can enlarge the target length sufficiently to admit collapse candidates. On an equal-edge-length mesh, the stricter test with the factor 0.8 admits an edge when its size factor exceeds 1.25, before the target sizes are averaged onto nodes.

Algorithm 1 first forms candidates using the predicted target lengths. Appendix C describes the updated test using 0.8 times the current target length after target-size transfer. Candidate edges also undergo one-ring conflict filtering and the boundary, topology, and mesh quality checks.

We have made the target ratio explicit through the element count in Appendix C and Algorithm 1. Let FF and F0F_0 denote the current and original element counts, respectively. The target is:

Ftarget=round⁡((1−Rtar)F0).F_{\mathrm{target}} = \operatorname{round}((1-R_{\mathrm{tar}})F_0).

After each collapse batch, the element count is updated. The loop continues while the candidate set is nonempty and the current count exceeds the target. We have corrected OR to AND in this condition. The requested ratio is specified through this count criterion; the affine output mapping is fixed across target ratios.

Response to W2, Q2 and Q3

Thank you for your insightful comment. The physical input consists of 16 frames uniformly sampled over the complete pre-solved coarse trajectory. These sampled frames are available during mesh construction and are used for sequence normalization. We now describe the input construction in Appendix D.1, with the test data following the same procedure in Appendix D.2. One coarse mesh is constructed for each case and target budget and remains fixed during the final solve, as stated in Appendix E.1.

The reported runtime covers mesh coarsening and the final coarse solve. GNN inference is part of the coarsening stage. Preprocessing, input-trajectory generation, and offline GNN training are excluded from the reported total. The pre-solve cost is listed separately in Table 2. The revised table also includes the corresponding relative L2L_2 errors. The stage timings and errors are reproduced below.

Stage timings and solution errors from Table 2. All times are in seconds.

Dense elems.Coarse elems.Pre-solveCoarseningDense solveCoarse solveRelative L2 error
14,5168,71516.32911.5627.6814.264.85 × 10⁻²
25,88615,56730.49818.1555.5128.639.03 × 10⁻⁴
173,20810,31583.334102.833,250.28147.521.29
248,9245,44641.508142.404,241.9283.840.55

To separate the learned size field from the collapse implementation, we have added the control in Appendix F.2.1 and Figure 10. The analytical field and the learned field use the same coarse physical input, parallel collapse operator, and matched element budgets. The control includes six steady cases. As can be seen from the results, PACE has lower relative L2L_2 error than the analytical field in all six cases. The additional cost of the learning module limits the coarsening-time advantage in most cases.

Response to W3 and Q4

Thank you for this suggestion. We have prepared supplementary videos for cylinder flow and the FX63137 airfoil. For details, please refer to the supplementary materials.

Thank you for the revision. The explicit size-factor mapping in Eq. (16) and corrected stopping condition address my original concern about how collapse candidates are admitted. Figure 10 also supplies the requested same-operator control: PACE improves over the analytical field in all six cases, although LDT has lower error in two, so the advantage of learning is not uniform.

My remaining concern is the practical accuracy/runtime trade-off. Appendix E.3 now explicitly excludes the pre-solve from the reported total speedup. Including it gives approximately 0.66x, 0.72x, 9.74x, and 15.84x for Table 2, before other excluded preprocessing. These totals should accompany the conditional speedups. Table 2 also reports relative L2 errors of 1.29 and 0.55 for the larger cases, without a percentage scale; under the definition in Appendix E.3, these imply 129% and 55%. Please confirm the units and explain the utility of these operating points, or show the speedups at a lower error tolerance. The implementation is clearer, but this trade-off remains important to my assessment.

AuthorsOpenReview ↗

Author response to reviewer GfUK

We thank the reviewer for the positive assessment of the revised inference procedure and the added control experiment. We apologize for not explaining the timing scope and error scale clearly in our previous response.

We have updated Table 2 and Appendix E.3 to include the cost of generating the input coarse trajectory in the total speedup. The revised definition is:

Stotal=TsolvedTpre+Tcoarse+Tsolves.S_{\mathrm{total}}=\frac{T_{\mathrm{solve}}^{d}} {T_{\mathrm{pre}}+T_{\mathrm{coarse}}+T_{\mathrm{solve}}^{s}}.

Here, TsolvedT_{\mathrm{solve}}^{d} and TsolvesT_{\mathrm{solve}}^{s} denote the dense and coarse mesh solving times, respectively, TpreT_{\mathrm{pre}} is the time required to generate the input coarse trajectory, and TcoarseT_{\mathrm{coarse}} is the mesh coarsening time. The resulting total speedups are 0.66x, 0.72x, 9.74x, and 15.84x, respectively. We have updated Section 4.2 accordingly.

We also clarify the error scale in Table 2. The previous entries were reported as dimensionless relative-error ratios, and we now explicitly report them as percentages. Using the values before rounding, the two larger-case errors are 128.60% and 54.92%.

We have additionally included the relative L2L_2 errors of uniform coarse meshes for comparison. The updated results are shown below.

Dense to coarse elementsTotal speedup (including pre-solve)PACE relative L2L_2 error (%)Uniform relative L2L_2 error (%)
14,516 → 8,7150.66x4.854.97
25,886 → 15,5670.72x0.090.12
173,208 → 10,3159.74x128.60145.27
248,924 → 5,44615.84x54.9281.38

We acknowledge that the absolute errors in the two larger cases remain high, and these configurations represent aggressive coarsening regimes rather than high-accuracy operating points. However, at comparable mesh sizes, PACE achieves lower relative L2L_2 errors than uniform coarsening in all four cases. In particular, the error reductions in the two larger cases are 16.67 and 26.46 percentage points, respectively. These results indicate that physics-aware resolution allocation can reduce the accuracy loss associated with aggressive coarsening while retaining substantial computational speedups. For applications with stricter accuracy requirements, less aggressive coarsening levels would be more appropriate.

We have added this discussion to Section 4.2 to make the accuracy-runtime trade-off explicit.

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

The paper presents a physics-aware mesh coarsening framework using a graph neural network and loss function based on error equidistribution, which guides the model to learn an edge-wise coarsening score map that reflects the local high gradient region of the mesh. The framework is applied to several PDE, both steady and unsteady simulations. The results show improvement in coarsening speed and solution accuracy compared to the tested baselines.

Critical strengths

The paper proposes an alternative for computationally expensive fine-mesh simulations. Both physical information and geometry were used to guide coarsening. The experiments cover both steady and transient problems, and evaluate both computational cost and solution error.

Critical weaknesses

The results don’t highlight how well the method generalizes to physical cases that differ from those used during training. For transient simulations, it is not clear enough how the method handles flow structures that develop or move as the simulation progresses. For instance, if a vortex or sharp gradient develops in a region that was heavily coarsened, the mesh may no longer resolve it adequately.The reported speedups are interesting, but it is important to understand the total computational cost, including the GNN and any preparation or training needed.

Questions for authors

For transient simulations, is the coarse mesh generated once or updated as the physical solution evolves? If the mesh is generated once, how does the method handle new flow structures that develop in regions where the mesh has been heavily coarsened? How does PACE perform on physical cases that are different from its training data, presenting high gradients in new unidentified regions? Does the reported runtime include the cost of training the GNN and preparing its input features, or only the coarsening stage? How does the method compare with the strongest physics-aware coarsening baselines when accuracy and total runtime are considered together? How is the physical solution represented on the coarse mesh, and how is the resulting error measured against the fine-mesh solution? How sensitive are the results to the chosen coarsening level? How does the proposed error-equidistribution loss compare with established edge-based interpolation error estimators using the length distribution tensor, particularly for Navier–Stokes flows with strong directional features such as boundary layers and vortices?

2 comments
AuthorsOpenReview ↗

Author response to reviewer Pnej

We thank the reviewer for the thoughtful feedback. Below, we provide our responses to the weaknesses (W1 to W3) and questions (Q1 to Q3) raised in the review.

Response to W2 and Q1

W2: For transient simulations, it is not clear enough how the method handles flow structures that develop or move as the simulation progresses. For instance, if a vortex or sharp gradient develops in a region that was heavily coarsened, the mesh may no longer resolve it adequately.

Q1: For transient simulations, is the coarse mesh generated once or updated as the physical solution evolves? If the mesh is generated once, how does the method handle new flow structures that develop in regions where the mesh has been heavily coarsened?

PACE constructs the coarse mesh once for each case and target element budget, and the resulting mesh remains fixed throughout the final simulation. To account for evolving physical structures, we uniformly sample 16 frames from the complete pre-solved coarse trajectory. These sampled frames generally capture the dominant flow structures in the evaluated cases. The temporal encoder then aggregates information across these frames to guide mesh resolution in regions affected by the evolving physical structures.

Response to W1 and Q2

W1: The results don’t highlight how well the method generalizes to physical cases that differ from those used during training.

Q2: How does PACE perform on physical cases that are different from its training data, presenting high gradients in new unidentified regions?

We evaluate PACE on physical cases that differ from those used during training. The training data consist of relatively simple PDE cases, while the test cases include more complex flow conditions, such as the cylinder and airfoil simulations in Figure 4 and the additional airfoil configurations in Figure 9 of Appendix F.1. In these cases, PACE generally identifies regions with strong physical variations and preserves higher mesh resolution near the dominant flow structures. These results demonstrate the ability of PACE to transfer the learned coarsening strategy from the training PDEs to more complex physical conditions.

Response to W3 and Q3

W3: The reported speedups are interesting, but it is important to understand the total computational cost, including the GNN and any preparation or training needed.

Q3: Does the reported runtime include the cost of training the GNN and preparing its input features, or only the coarsening stage?

The reported runtime focuses on the online computational cost of mesh coarsening and PDE solving. The GNN is trained offline and can be reused across different simulation cases, allowing the one-time training cost to be amortized over multiple evaluations.

The recorded coarsening time includes GNN inference, input preparation within the prediction stage, and mesh construction. Preprocessing outside the recorded timing stages is excluded. The coarse physical trajectory is generated through a separate pre-solve stage, and its computational cost is explicitly reported in Table 2.

Accordingly, the reported total speedup measures the computational benefit of mesh coarsening and subsequent PDE solving when the coarse input trajectory is available. The timing definitions and the computational stages included in the reported speedups are clarified in Appendix E.3.

AuthorsOpenReview ↗

Author response to reviewer Pnej

We thank the reviewer for the thoughtful feedback. Below, we provide our responses to the questions (Q4 to Q7) raised in the review.

Response to Q4 and Q7

Q4: How does the method compare with the strongest physics-aware coarsening baselines when accuracy and total runtime are considered together?

Q7: How does the proposed error-equidistribution loss compare with established edge-based interpolation error estimators using the length distribution tensor, particularly for Navier–Stokes flows with strong directional features such as boundary layers and vortices?

We compare PACE with both learning-based and conventional mesh coarsening methods. Figure 2 and Table 1 show that PACE achieves lower solution errors than the evaluated learning-based baselines on the transient benchmarks, while Figure 3 compares numerical accuracy and runtime against QEM-based methods on steady-state problems. These results demonstrate the favorable performance of PACE across the evaluated settings.

We have also added a controlled comparison in Appendix F.2.1 and Figure 10 using the analytical size field and a field based on the Length Distribution Tensor (LDT). All three variants use the same coarse physical input and parallel collapse operator, with matched element budgets. The experiments cover three Poisson and three Helmholtz cases.

PACE achieves lower relative L2L_2 errors than the analytical field across all six cases, while its advantage over LDT is smaller. This smaller gap may be attributed to the use of local interpolation-error information in LDT to guide mesh resolution. The additional computational cost of the learning module also limits the coarsening-time advantage of PACE in most cases. These results provide a more direct comparison of the contribution of the learned size field under the same coarsening implementation.

Response to Q5

Q5: How is the physical solution represented on the coarse mesh, and how is the resulting error measured against the fine-mesh solution?

The physical solution is represented by values defined at mesh nodes. For network input, these values are converted into edge-associated physical features, since each node of the input graph corresponds to an edge of the original mesh. This conversion is described in Section 3.3. For error evaluation, the coarse-mesh solution is interpolated onto the fine reference mesh and compared with the fine-mesh solution at the corresponding time steps. Appendix E.3 provides the detailed definitions of the relative L2L_2 error and rollout RMSE.

Response to Q6

Q6: How sensitive are the results to the chosen coarsening level?

As shown in Figure 2, the numerical error generally increases as the element reduction ratio increases. However, PACE consistently achieves lower errors than the evaluated baselines and uniform coarsening across the reported reduction ratios. These results show that PACE maintains better numerical accuracy under different mesh resolution budgets.

Comments on the paper

1 comment
Area_Chair_HexrOpenReview ↗

Requesting reviewer feedback

Dear Reviewers,

Thank you for your reviewing service. The authors have provided detailed rebuttals to the comments. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations. Thank you for your continued engagement.

Thank you,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.