acceptodds
Withdrawn from ICLR 2027

EntroMesh: Mechanics-Conditioned Structural Entropy for Multi-Solid Stamping

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learned simulation of multi-solid stamping requires long-range reasoning over irregular finite-element meshes, where load, contact, material transitions, and localized plasticity govern inter-element couplings. Hierarchical mesh simulators reduce graph diameters, yet their coarse-level meshes are commonly derived from topology, geometry, algebraic connectivity, or standalone learned scores. These static coarsening schemes ignore instantaneous, state-dependent mechanical couplings and introduce substantial prediction error for evolving physical states. We introduce EntroMesh, a mechanics-conditioned structural-entropy hierarchy for multi-solid mesh prediction. EntroMesh converts each finite-element state into a dimensionless weighted graph, strengthens edges that carry stiffness, load, boundary, and strain-energy evidence, weakens edges across large stress jumps, and constructs a fixed-budget coding tree for a coarse-to-fine V-cycle graph simulator. Our hierarchy is physically interpretable, state-causal, and enables fair comparison under identical node budgets. Evaluated on the two large-scale long-time Unisoma stamping benchmarks against the state-of-the-art Unisoma model, EntroMesh reduces the sum of six-field relative- prediction error from 65.08 to 62.62 on BilateralStamping and from 64.19 to 53.65 on UnilateralStamping. Experiments demonstrate that mechanics-aware structural entropy is an effective hierarchical prior for large-scale multi-solid learned stamping simulation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
learned simulation, graph neural networks, structural entropy, solid mechanics, finite element method

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

The paper introduces an alternative method to construct a hierarchical graph for graph-based neural simulators. Here, the authors use mechanical information from the dynamics, such as stiffness, loads, and stress, to reweight the mesh edges. They then merge neighboring node groups based on changes in structural entropy. The resulting tree then defines the hierarchy used by the learned simulator.

Critical strengths
  • The author showed that the structural-entropy approach performs somewhat better than the random coarsening and the position attention on the BilateralStamping task.
Critical weaknesses
  • As stated in the discussion section, the evaluation is still limited to two solid-mechanics stamping tasks. Whether the approach can be extended to other dynamics remains unclear. In the paper's own experiment, on the larger UnilateralStamping task, it does not show a clear gain over the baselines.

  • More importantly, the paper does not explain how the coefficients in the mechanical weighting were chosen. This leads to a critical concern: how to select the "mechanical information" and its corresponding coefficients for different dynamics, e.g., fluid dynamics, which would severely limit the method’s broader applicability.

  • The hierarchy comparison only includes random coarsening and position attention. I would like to see comparisons with M4GN [1], which uses modal-decomposition features for grouping, and AMG-based approaches such as ROBIN [2]. Without these comparisons, it is unclear whether structural entropy is more useful than existing approaches or simply performs similarly.

  • The figures need substantial revision. For example, in Figure 5, the colors do not clearly highlight the groups, and it is unclear what the authors want to show or how the figure supports their approach.

[1] B. Lei, et al. M4GN: Mesh-based Multi-segment Hierarchical Graph Network for Dynamic Simulations. TMLR, 2025.

[2] T. Würth, et al. Diffusion-Based Hierarchical Graph Neural Networks for Simulating Nonlinear Solid Mechanics. NeurIPS, 2025.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer GjqM (Part 1 of 2)

W.1 Generalization Beyond Stamping

We have added CylinderFlow experiments on autoregressive velocity and pressure prediction over unstructured meshes. Results are included in Table 1, and Figure 10 in Appendix C.2 illustrates the corresponding physical inputs and hierarchy.

CylinderFlow errors, reported as mean ± standard deviation. Lower is better.

MethodRMSE 1 (×10⁻⁵)RMSE 50 (×10⁻⁴)RMSE all (×10⁻⁴)
GCN675 ± 28382 ± 691702 ± 310
g-U-Net401 ± 3.7179 ± 32758 ± 88
MGN246 ± 1362.8 ± 2.8412 ± 48
BSMS-GNN181 ± 22252 ± 8.21218 ± 83
EAGLE456 ± 2469.6 ± 3525 ± 24
M4GN290 ± 2565.2 ± 8340 ± 31
ROBIN700 ± 60180 ± 20900 ± 100
EntroMesh170 ± 1255 ± 3300 ± 18

EntroMesh has the lowest mean error at all three horizons, reducing complete trajectory error by 11.76% relative to M4GN. This provides evidence of applicability beyond solid stamping.

W.2 Effectiveness on UnilateralStamping

We have added M4GN and ROBIN comparisons in Table 1 and analyzed the field results in Section 4.2. The main comparisons are summarized below; Table 1 contains all baselines.

UnilateralStamping relative L₂ errors in ×10⁻² units. M denotes metal; R denotes rubber. Lower is better.

MethodGeo MGeo RStress MStress RPEEQ MPEEQ R
Unisoma0.710.638.077.7325.9121.14
M4GN0.770.667.877.5223.1919.62
ROBIN4.395.1914.2820.0573.5380.37
EntroMesh0.570.517.566.8020.7517.47

EntroMesh achieves state of the art performance among all methods in Table 1, with the lowest error across all six fields. Aggregate error decreases by 16.42% relative to Unisoma, covering improvements in geometry, stress, and plastic response.

W.3 Weight Selection and Cue Ablation

We have clarified cue selection in Section 3.2 and Figure 3 and added weighting ablations in Table 5 of Appendix F.1. Loads and boundaries describe operating conditions, stiffness represents material coupling, energy identifies concentrated response, and stress jumps identify candidate module boundaries. Available cues are normalized into dimensionless graph weights.

Both stamping tasks use stiffness, load, and energy coefficients of 1.0, boundary coefficient 0.75, and stress consistency coefficient 0.5. These empirical settings provide reference amplification, moderate boundary amplification, and smooth attenuation across stress jumps. Figure 6 reports local sensitivity.

BilateralStamping weighting ablations. Relative L₂ errors are in ×10⁻² units. M denotes metal; R denotes rubber. Lower is better.

ConfigurationGeo MGeo RStress MStress RPEEQ MPEEQ R
Full model0.500.512.5710.1826.1422.71
Without load term0.520.522.5810.7027.1824.11
Without boundary term0.520.522.5810.6527.1524.28
Geometry weights only0.520.542.5910.8527.1124.42

The full model achieves the lowest error across all fields. Removing load and boundary weighting increases rubber PEEQ error by 6.16% and 6.91%; geometry weights alone increase it by 7.53%. Rubber stress also worsens. These comparisons establish the individual contributions of load and boundary weighting and the combined benefit over geometry alone.

AuthorsOpenReview ↗

Response to Reviewer GjqM (Part 2 of 2)

W.4 Comparison with M4GN and AMG

We have added M4GN, which uses modal features for grouping, and ROBIN, which uses AMG, to Table 1. Results on both stamping tasks follow.

BilateralStamping relative L₂ errors in ×10⁻² units. M denotes metal; R denotes rubber. Lower is better.

MethodGeo MGeo RStress MStress RPEEQ MPEEQ R
Geo FNO1.251.173.9016.1242.6947.31
GINO2.992.435.4326.2155.2364.09
GNO2.852.814.6325.2458.3059.09
LSM5.986.155.5636.1670.0173.47
Galerkin5.995.945.6334.4269.6273.35
FactFormer2.971.054.5913.7862.6642.18
OFormer3.342.684.9531.7855.5151.57
ONO3.012.754.6432.9151.7652.31
Transolver0.830.813.4614.2535.9232.76
GraphSAGE2.272.294.4021.9649.3644.95
MGN1.951.974.2120.5946.4641.41
Unisoma0.570.522.7810.3928.1722.65
M4GN0.640.632.7011.2727.4424.82
ROBIN1.132.793.5927.8738.5756.59
EntroMesh0.500.512.5710.1826.1422.71

UnilateralStamping relative L₂ errors in ×10⁻² units. Lower is better.

MethodGeo MGeo RStress MStress RPEEQ MPEEQ R
Geo FNO1.091.5411.1311.2437.7135.74
GINO1.882.1912.6214.3846.5148.85
GNO4.264.2413.5118.1464.1368.59
LSM3.673.5713.2713.3558.8960.26
Galerkin10.0710.1017.2527.9479.8685.55
FactFormer1.781.6610.9510.3538.8630.22
OFormer1.190.9811.039.8132.6926.75
ONO4.334.3713.7713.1151.4047.48
Transolver0.790.758.398.6229.3826.93
GraphSAGE2.752.6010.5911.8954.4151.06
MGN3.913.8312.5214.1758.0556.94
Unisoma0.710.638.077.7325.9121.14
M4GN0.770.667.877.5223.1919.62
ROBIN4.395.1914.2820.0573.5380.37
EntroMesh0.570.517.566.8020.7517.47

EntroMesh outperforms M4GN and ROBIN across all six fields on both tasks. Table 1 compares complete models, while Table 2 compares hierarchy rules under the same architecture and node budgets.

W.5 Figure Clarity

We have redesigned the original Figure 5, now Figure 7, and added analysis in Section 4.3. It follows one real BilateralStamping input through physical inputs, edge weighting, and coarse modules.

AuthorsOpenReview ↗

Thank you for your review

We have uploaded the revised paper and the corresponding responses. The revised parts of the paper are marked in red. We hope our responses can address some of your concerns. We are also very grateful for your valuable feedback on our paper. If our response can satisfy you, we would be extremely honored that you could revise your previous opinion on our paper.

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

This paper addresses the learned simulation of multi-solid stamping. Traditional hierarchical graph networks rely on static geometry or topology to coarsen meshes, which easily disrupts load paths and blurs localized stress and plastic deformation. To resolve this issue, the paper proposes the EntroMesh framework. EntroMesh first builds a dimensionless mechanics-weighted graph using current loads, boundaries, stiffness, strain-energy proxies, and stress jumps. Next, it constructs a bottom-up coding tree by minimizing two-dimensional structural entropy. Finally, it takes multiscale snapshots under fixed node budgets to define the hierarchy, which guides a V-cycle graph neural network with skip fusion for long-range communication and localized feature prediction. The primary contribution is decoupling mesh hierarchy construction from neural network parameters, with theoretical proofs for the graph's causality and scale invariance. Evaluated on two large-scale Unisoma industrial stamping benchmarks against 12 baselines, EntroMesh reduces the sum of six-field relative-L2 prediction errors (geometry, stress, and plastic strain) by 3.78% and 16.42%, respectively, demonstrating significant performance gains.

Critical strengths
  1. The model innovatively converts current-state information—such as stiffness, loads, boundary constraints, and strain energy—into dimensionless edge weights. By introducing structural entropy from information theory to guide graph compression, it successfully translates complex mechanical requirements into an information-theoretic optimization goal of minimizing structural entropy. Furthermore, the authors provide rigorous theoretical proofs, which lends high credibility to the proposed method.
  2. It improves upon conventional hierarchical graph neural networks, which rely on static Euclidean distances, topological connectivity, or black-box learned scores for mesh coarsening. By incorporating physical cues like stiffness and loads, it shifts from static representations to real-time physical states and uses them as the core criterion for node coarsening. This allows the resulting graph network to naturally align with the actual load paths within the workpiece. The experiments demonstrate that without increasing neural network parameters and under identical coarse-node budgets, merely changing the hierarchical organization rule achieves substantial error reductions on both unilateral and bilateral stamping benchmarks.
Critical weaknesses
  1. When constructing the mechanics-weighted graph, the authors introduce multiple auxiliary physical quantities, including external loads, boundary constraints, material stiffness, and strain energy. However, the selection of these physical quantities appears to be grounded primarily in empirical intuition, lacking both explicit feature selection criteria and rigorous theoretical derivations or sources. More critically, the paper completely lacks a step-by-step feature ablation study for these physical priors. Consequently, it is impossible to assess the independent contribution of each physical cue to the final prediction accuracy, to verify whether all these quantities are genuinely effective, or to rule out potential feature redundancies and negative interference.
  2. The main framework overview diagram (Figure 2) carries noticeable presentation flaws that can cause confusion. In the top banner of Figure 2, the first three baseline panels explicitly illustrate the mesh coarsening process from fine to coarse scales along with their specific failure modes. In contrast, the subfigure for "Ours" displays virtually identical meshes on both sides of the arrow, breaking the comparative logic established by the baselines; clear legend explanations (e.g., for the red crosses and colored edges) are also missing. Furthermore, the relationship between the multiple tree structures depicted in step 3 (Structural-Entropy Coding Tree) of the middle panel and the unified coding tree proposed in the paper is insufficiently explained. Finally, the main text completely lacks an in-text call-out for Figure 2, and the brief caption fails to resolve the ambiguities arising from these visual representations.
Questions for authors
  1. When constructing the mechanics-weighted graph, the model incorporates multiple auxiliary physical quantities, including external loads, boundary constraints, material stiffness, strain energy, and stress jumps. To confirm the independent contribution of each physical prior to the final prediction accuracy, could the authors provide a step-by-step feature ablation study (for example, reporting the prediction error changes across all physical fields when removing loads, boundary indicators, strain energy, or stress jumps individually)? If a complete ablation study cannot be provided, could the authors elaborate on the theoretical foundations or sources for selecting this specific set of physical quantities, and explain whether this combination underwent any screening or feature selection process?
  2. As the core framework diagram, Figure 2 currently lacks an explicit in-text cross-reference, and several of its visual designs can easily mislead readers. Could the authors formally add an in-text citation to Figure 2 in the main manuscript and provide clear explanations of the relevant details in either the main text or the figure caption? Specifically, the authors need to explain what is indicated by the two identical graphs on either side of the arrow in the subfigure for "Ours" in the top banner, provide physical definitions for symbols such as the red lines and red crosses, and clarify the projection and mapping relationship between the multiple tree structures shown in step 3 of the middle panel and the unified coding tree proposed in the paper.
2 comments
AuthorsOpenReview ↗

Response to Reviewer Vkuz

W.1 Physical Prior Justification and Ablation

We have clarified cue selection in Section 3.2 and Figure 3 and added weighting ablations in Table 5 of Appendix F.1. Loads and boundaries describe operating conditions, stiffness represents material coupling, energy identifies concentrated response, and stress jumps identify candidate module boundaries. Available cues are normalized into dimensionless graph weights.

Both stamping tasks use stiffness, load, and energy coefficients of 1.0, boundary coefficient 0.75, and stress consistency coefficient 0.5. These empirical settings provide reference amplification, moderate boundary amplification, and smooth attenuation across stress jumps. Figure 6 reports local sensitivity.

BilateralStamping weighting ablations. Relative L₂ errors are in ×10⁻² units. M denotes metal; R denotes rubber. Lower is better.

ConfigurationGeo MGeo RStress MStress RPEEQ MPEEQ R
Full model0.500.512.5710.1826.1422.71
Without load term0.520.522.5810.7027.1824.11
Without boundary term0.520.522.5810.6527.1524.28
Geometry weights only0.520.542.5910.8527.1124.42

The full model achieves the lowest error across all fields. Removing load and boundary weighting increases rubber PEEQ error by 6.16% and 6.91%; geometry weights alone increase it by 7.53%. Rubber stress also worsens. These comparisons establish the individual contributions of load and boundary weighting and the combined benefit over geometry alone.

W.2 Figure 2 Interpretation and Referencing

We have revised Figure 2, its legend and caption, and explicitly referenced it in Section 3. Figures 3 and 4 explain weighting and tree construction, and Figure 7 provides a real case.

The top comparison shows seven fine nodes becoming six coarse nodes for every method. Our example retains the strongly coupled pair within one module. Colors identify modules, thick black edges indicate strong coupling, and red markers identify lost or spurious links. The displayed resolutions come from one coding tree, with parent maps defining pooling and lifting. Shields in the detailed diagrams represent importance and contribute a soft merge penalty.

AuthorsOpenReview ↗

Thank you for your review

We have uploaded the revised paper and the corresponding responses. The revised parts of the paper are marked in red. We hope our responses can address some of your concerns. We are also very grateful for your valuable feedback on our paper.

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

This paper introduces EntroMesh, a mechanics-conditioned structural-entropy hierarchy for multi-solid mesh prediction. It constructs a dimensionless weighted graph to represent each finite-element state, reinforces edges encoding stiffness, load and strain-energy evidence, attenuates edges having large stress jumps, and then builds a fixed-budget coding tree, and integrates the resulting hierarchy into a coarse-to-fine V-cycle graph simulator. On the two large-scale long-time Unisoma stamping benchmarks, EntroMesh reduces the aggregate six-field relative-L2 prediction error from 65.08 to 62.62 on Bilateral Stamping and from 64.19 to 53.65 on Unilateral Stamping, showing good performance against several representative neural-operator and graph-based baselines.

Critical strengths

S1: This paper addresses an important and practical challenge. In multi-solid mesh prediction, phenomena like long-range load transfer, contact, and localized plasticity are very common, and it’s essential to adjust the hierarchy based on current mechanical state rather than geometric information.

S2: The design of hierarchy is targeted. This paper combines geometric adjacency with mechanical information, and compares rules at different levels by fixed budgets of nodes, which can reduce the impact of hierarchy strategies and provide useful evidence on benchmark results.

S3: The reported results on Unilateral Stamping are promising. The results show that this method outperforms several other baselines on the two datasets and includes comprehensive results like hierarchy ablation and efficiency description.

Critical weaknesses

W1: This method is only limited to two datasets: Bilateral Stamping and Unilateral Stamping. Although these datasets are the largest Unisoma long-time stamping containing deformable solids, rigid solids and contacts, this paper doesn’t show whether the method can be applicable to other solid-state problems, different mesh generation methods or different contact and load conditions.

W2: The gains vary across different tasks. On Bilateral Stamping, the improvement is really minor, and rubber PEEQ is slightly worse than Unisoma, but on Unilateral Stamping, all the results are improved. So it’s important to clarify which task or which physical condition this method is more competitive.

W3: Physical weights can be explained more clearly. Although parameter sensitivity shows that the default operates in a locally stable regime, this paper can describe how default parameters are selected, and how this method works when deleting load, boundary or strain-energy evidence.

W4: The description of figure should be more detailed. The description of figure 2 is really short and doesn’t clearly explain how the framework works and what’s the meaning of the nodes of different colors or the shields. It’s better to use a real case as a demo.

Questions for authors

Q1: Are these two datasets splits dependently based on the geometry or design-case level, and what do the results show if the method uses different random seed or data split.

Q2: Parameter sensitivity shows that the remaining mechanics coefficients exhibit comparatively small variations, so what are the default coefficients and how they are selected? What do the results show if the method removes a cue like load, boundary, stress, or strain-energy evidence? And which physical signals above are essential in these tasks?

Q3: What is the runtime and memory cost of the whole framework like graph weighting, fixed-budget coding tree and V-cycle graph simulator? Can the authors provide a runtime and memory comparison against Unisoma on the same conditions like hardware and other cases and clarify how the cost scales with mesh size?

6 comments
AuthorsOpenReview ↗

Response to Reviewer ZeLG (Part 1 of 3)

W.1 Generalization

We have added CylinderFlow experiments on autoregressive velocity and pressure prediction over unstructured meshes. Results at three prediction horizons are included in Table 1 of the revised manuscript.

CylinderFlow errors, reported as mean ± standard deviation. Lower is better.

MethodRMSE 1 (×10⁻⁵)RMSE 50 (×10⁻⁴)RMSE all (×10⁻⁴)
GCN675 ± 28382 ± 691702 ± 310
g-U-Net401 ± 3.7179 ± 32758 ± 88
MGN246 ± 1362.8 ± 2.8412 ± 48
BSMS-GNN181 ± 22252 ± 8.21218 ± 83
EAGLE456 ± 2469.6 ± 3525 ± 24
M4GN290 ± 2565.2 ± 8340 ± 31
ROBIN700 ± 60180 ± 20900 ± 100
EntroMesh170 ± 1255 ± 3300 ± 18

EntroMesh achieves the lowest mean error at all three horizons and reduces complete trajectory error by 11.76% relative to M4GN. This supports applicability beyond solid stamping.

W.2 Performance Gains and Applicable Conditions

We have clarified aggregate and local plasticity results in Section 4.2 and added spatial diagnostics in Figure 16 of Appendix E.

Stamping errors in ×10⁻² units. Lower is better.

DatasetUnisoma aggregateEntroMesh aggregateReductionUnisoma rubber PEEQEntroMesh rubber PEEQ
BilateralStamping65.0862.623.78%22.6522.71
UnilateralStamping64.1953.6516.42%21.1417.47

On BilateralStamping, aggregate error improves by 3.78%, while rubber PEEQ is slightly higher. A possible explanation is that our hierarchy preserves load transmission and stress coupling without a dedicated plastic history preservation term. Figure 16 shows retained overall deformation but an underestimated local PEEQ maximum. Our strongest gains occur on UnilateralStamping, with more rigid bodies and contact interactions, where all six fields improve. The observed advantage is strongest in the evaluated task with more complex mechanical interactions.

W.3 Weight Selection and Cue Ablation; Q.2 Default Coefficients and Cue Importance

We have added coefficient settings and physical weighting ablations in Appendix F.1. Both stamping tasks use stiffness, load, and strain energy coefficients of 1.0, a boundary coefficient of 0.75, and a stress consistency coefficient of 0.5. After robust normalization, these empirical settings provide reference amplification, moderate boundary amplification, and smooth attenuation across stress jumps, respectively. We use the same settings without separate task tuning. Figure 6 reports sensitivity, and Table 5 reports the ablations.

BilateralStamping weighting ablations. Relative L₂ errors are in ×10⁻² units. M denotes metal; R denotes rubber. Lower is better.

ConfigurationGeo MGeo RStress MStress RPEEQ MPEEQ R
Full model0.500.512.5710.1826.1422.71
Without load term0.520.522.5810.7027.1824.11
Without boundary term0.520.522.5810.6527.1524.28
Geometry weights only0.520.542.5910.8527.1124.42

The full model gives the lowest error across all six fields. Removing load and boundary weighting increases rubber PEEQ error by 6.16% and 6.91%; geometry weights alone increase it by 7.53%. Rubber stress also worsens. These results establish the contributions of load and boundary weighting. Figure 6 supports local stability around the default coefficients.

AuthorsOpenReview ↗

Response to Reviewer ZeLG (Part 2 of 3)

W.4 Framework Illustration

We have revised Figure 2 and explicitly referenced it in Section 3. Figures 3 and 4 detail weighting and merging, and Figure 7 provides a real stamping example. Further cases appear in Figures 9 and 10 of Appendix C.2.

Figure 2 now shows seven fine nodes becoming six coarse nodes for every method, with our example retaining the strongly coupled pair in one module. Thick black edges indicate strong coupling, red markers indicate lost or spurious links, and node colors indicate modules. The graph resolutions are selected views of one coding tree across arbitrary levels. Parent maps define pooling and lifting. Shields in the detailed diagrams represent importance and contribute a soft merge penalty.

Q.1 Data Splits and Robustness

We have clarified the official partitions and added alternative splits in Tables 6 and 7 of Appendix F.2. Official training, validation, and test proportions are approximately 83.33%, 8.33%, and 8.33%. Seed 42 controls training only and does not alter the official split. Seeds 17 and 23 control both the additional random partitions and their corresponding training runs.

BilateralStamping results across splits and seeds. Relative L₂ errors are in ×10⁻² units. M denotes metal; R denotes rubber. Lower is better.

SplitSplit seedTraining seedGeo MGeo RStress MStress RPEEQ MPEEQ R
Officialn/a420.500.512.5710.1826.1422.71
Random17170.490.502.5510.1626.0322.64
Random23230.480.522.5510.0926.1222.68

UnilateralStamping results across splits and seeds. Relative L₂ errors are in ×10⁻² units. Lower is better.

SplitSplit seedTraining seedGeo MGeo RStress MStress RPEEQ MPEEQ R
Officialn/a420.570.517.566.8020.7517.47
Random17170.520.527.556.6121.0717.71
Random23230.560.517.646.8422.0917.34

BilateralStamping results remain close, with metal PEEQ ranging from 26.03 to 26.14 and rubber PEEQ from 22.64 to 22.71. UnilateralStamping geometry and stress errors also remain close, while metal PEEQ varies more, from 20.75 to 22.09. These runs support stability across the evaluated combinations of partitioning and training randomness.

Q.3 Efficiency and Scalability

We have added complete framework profiling against Unisoma on RTX 3090 in Table 8 of Appendix F.3.

Complete framework costs on RTX 3090. Memory is in MB.

DatasetMethodTraining ms/stepInference ms/caseMemory training / inferenceParams MFLOPs G
BilateralStampingUnisoma108.2743.45284.1 / 75.62.8246.987
BilateralStampingEntroMesh1187.661130.771180.4 / 163.72.27735.664
UnilateralStampingUnisoma186.2977.30528.5 / 140.75.66716.282
UnilateralStampingEntroMesh4150.103980.823745.8 / 489.13.623126.994

EntroMesh uses 19.37% and 36.07% fewer parameters on the two tasks. Its inference latency is 26.02 and 51.50 times that of Unisoma, with inference memory of 163.7 MB and 489.1 MB. Higher prediction accuracy therefore comes with greater runtime, memory, and computation despite fewer parameters.

AuthorsOpenReview ↗

Response to Reviewer ZeLG (Part 3 of 3)

Q.3 Efficiency and Scalability (continued)

We have also added complete framework measurements at five mesh sizes per task in Table 9 of Appendix F.4.

Costs across mesh sizes on RTX 3090. Time is in ms, memory in MB, parameters in millions, and FLOPs in billions.

DatasetMean nodesMethodTraining / stepInference / caseMemory training / inferenceParamsFLOPs
BilateralStamping1,724Unisoma102.0537.2887.2 / 23.72.8241.010
BilateralStamping1,724EntroMesh234.38171.04208.1 / 33.62.2776.032
BilateralStamping3,447Unisoma110.8440.36108.7 / 28.12.8241.856
BilateralStamping3,447EntroMesh357.21294.99330.3 / 49.42.27710.567
BilateralStamping6,893Unisoma116.2238.15153.3 / 37.22.8243.550
BilateralStamping6,893EntroMesh570.25605.69576.9 / 79.42.27719.697
BilateralStamping10,338Unisoma112.5937.00196.8 / 46.42.8245.243
BilateralStamping10,338EntroMesh902.81934.61799.9 / 108.02.27728.293
BilateralStamping13,785Unisoma115.0237.28240.2 / 55.82.8246.937
BilateralStamping13,785EntroMesh1293.091118.59998.9 / 133.02.27735.470
UnilateralStamping6,115Unisoma184.4471.38169.2 / 42.85.6672.383
UnilateralStamping6,115EntroMesh606.26421.07675.4 / 80.83.62323.357
UnilateralStamping12,231Unisoma184.6875.32216.9 / 55.15.6674.368
UnilateralStamping12,231EntroMesh1193.60885.521092.1 / 133.73.62339.275
UnilateralStamping24,461Unisoma181.5577.93315.0 / 79.95.6678.337
UnilateralStamping24,461EntroMesh1825.031788.781913.6 / 237.83.62370.630
UnilateralStamping36,680Unisoma197.3981.65413.8 / 107.05.66712.304
UnilateralStamping36,680EntroMesh2987.632840.542705.0 / 335.03.623100.605
UnilateralStamping48,911Unisoma194.2985.13509.5 / 130.95.66716.274
UnilateralStamping48,911EntroMesh4086.494011.523391.4 / 424.83.623126.607

For approximately eight times as many nodes, EntroMesh inference time increases by factors of 6.54 and 9.53 on BilateralStamping and UnilateralStamping, respectively. Inference memory increases by factors of 3.96 and 5.26. Parameter counts remain unchanged within each task. These results characterize mesh size scaling within each task separately from differences between the two workloads.

Comments on the paper

1 comment
Area_Chair_u1W6OpenReview ↗

Discussion encouraged during this period!

Hi all reviewers,

We’re about halfway through the discussion period. This is a friendly reminder to check the authors' replies. If the authors have answered questions or responded to critiques in the reviews, please leave a comment for the authors!

Thank you so much!

Best, Your AC

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

Related papers

Discussion (0)

Sign in to comment.