acceptodds
Under review as a conference paper at ICLR 2027

GeoSlice-MGN: Analysis-Guided Geometric Routing for Impact Response Prediction in Realistic Structures

Abstract

Localized impact on complex structures produces high-gradient responses near the impact region and long-range structural coupling. Predicting these responses requires effective representations of loading position, structural boundaries, and local mesh geometry. We examine these geometric dependencies through finite-plate mechanics. Modal response analysis shows the roles of impact-relative and structure-relative positions. Analysis of bending and stretching strains shows that local spatial variations remain closely related to neighborhood directions and distances under finite deflection. Based on these relationships, we propose GeoSlice-MGN, which uses impact-relative geometry and bounding-box-relative position to modulate content-driven node-to-slice assignment. This makes load–structure geometry directly available to global routing in addition to its role in node features. The model retains local MeshGraphNet propagation before and after global interaction and augments slice states with relative soft occupancy. We also publicly release AutoPanelImpact, an open benchmark of 1,500 parameterized nonlinear impact cases on three real automotive floor panels, with paired displacement and stress fields under varying impact conditions. GeoSlice-MGN achieves the lowest average error on all six displacement–stress test metrics across the three panels. Experiments on BumperBeam and Plasticity further evaluate its applicability to coupled-field collision and deformation-sequence prediction.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Impact response prediction, Vehicle-derived structures, Geometric routing, Graph neural networks, Mechanics-guided learning, Finite element surrogate modeling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_kikaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a geometry-grounded transformer architecture to predict the physical response of thin shells when they encounter moderate point impacts. The method is augmented with a new benchmark for this task using automotive data, and the paper shows that it outperforms or is competitive with the state of the art on all tests.

Critical strengths

The paper is mostly well written, and does a great job of intuitively justifying the need for the specific geometric representation used. It tackles an important problem in a way that outperforms the state of the art, and it does so while providing theoretical insights that will likely inspire future work. Given the usual replication crisis that plagues work on neural PDE surrogates, I also think that the benchmark is a great contribution for the community.

Critical weaknesses

In my opinion, the main weakness of the paper is in its justification and comparison to classical alternatives. The paper uses finite element analysis to generate training data: how does inference on this model compare to direct, classical finite element analysis? Is inference faster? The sentence in the introduction "parametric studies become expensive when these simulations..." seems to hint at this, but it is unclear if it's a general statement or it applies to this paper's specific problem setting. Any finite element method will have a tradeoff curve between accuracy and runtime as one decreases the grid/mesh resolution. Even discounting training time, where does this method's inference fall in relation to this curve? Some of this information is in Appendix J.2 but it is quite incomplete: what do the error of the two methods look like if the FE mesh / tolerances are reduced to match equal runtime? The same effort put into hyper parameter tuning of this method and baselines should be put into finding the optimal setting for FE, and then compared at equal time and equal error with the method proposed.

To be clear, I do not think it is necessary for the method to outperform classical FE analysis, but I really do think that "how well does this research line compare to traditional alternatives" is a critical question any paper on this topic should answer, comprehensively (and in the main text).

For such an application-driven paper, I also missed a result in which the neural surrogate is used for some real engineering task. For example, how good is the surrogate for tasks like shape optimization? Or to predict fracture with von Mises stress? After all, these are the motivations stated for the work. I don't think this is a massive problem precluding publication, but it would have made the paper stronger.

Finally, I believe the clarity of Sections 3.4-3.X should be improved. Currently, the text relies on the reader having knowledge of a prior work or reading sections of the Appendix to even understand most of the notation and get even just an intuitive understanding of how the architecture works. While I really appreciate the intuition in 3.3; if necessary, it can be compressed (after all, it seems somewhat intuitive that deformation depends on relative positioning to both the impact and the rest of the ship) in favor of a clearer and more complete explanation of the method.

Questions for authors

The main thing that would affect my assessment is an answer on the comparison against classical methods (see above). Where does the method compare to traditional FE on an error vs runtime tradeoff curve? Do the authors plan to include such a result in a revised version of the paper?

I would also appreciate if the related work section could be improved. For example, the difference between "graph-based physical modeling" and "neural operators and latent global interaction" is unclear to me. Don't graph neural network also "learn mappings between input and response fields"? Perhaps an introductory paragraph, a clarification of how these methods are related to the present paper, and clearer subheadings would be an improvement.

5 comments
AuthorsOpenReview ↗

Response to Reviewer kika

Thank you for your constructive comments on the finite element comparison, engineering relevance, and clarity of the paper. We address these points below and describe the corresponding additions and planned revisions.

1. Accuracy–runtime comparison with finite elements

We thank the reviewer for suggesting an accuracy–runtime comparison with finite element analysis. We have added Section 4.6 and Figure 5, comparing GeoSlice-MGN with an FE mesh-resolution sweep on ten floorfrontdriver impact cases using a common numerical reference.

GeoSlice-MGN achieves a mean joint relative difference of 20.31% with a GPU forward-pass time of 15.33 ms per case, compared with 22.17% in 50.65 s and 17.49% in 61.77 s for the two coarser FE meshes. These results demonstrate comparable differences relative to the numerical reference at substantially lower execution cost, while finer FE meshes achieve smaller differences at higher cost.

2. Engineering relevance and application scope

Our work focuses on rapidly predicting displacement and von Mises stress fields in automotive floor panels under varying impact conditions. Excessive deformation may cause interference with surrounding components, while regions of elevated stress warrant further structural assessment. These responses are therefore relevant to engineering evaluation, but obtaining them across numerous impact conditions through nonlinear FE simulations is computationally expensive.

We evaluate the prediction accuracy on three vehicle-derived floor panels and further compare accuracy and execution cost against FE simulations at different mesh resolutions in Section 4.6. These results establish the accuracy and computational efficiency of the surrogate for the targeted prediction task. We agree that downstream applications would provide additional engineering validation. Shape optimization requires evaluation across changing design geometries, while fracture prediction requires additional failure modeling beyond von Mises stress. Both are valuable directions for future work.

3. A self-contained method description

Thank you for pointing this out. We agree that the method description should be self-contained and easier to follow without referring to prior work or the appendix. We will streamline the intuition in Section 3.3 and revise the subsequent sections to present the complete forward process, from local mesh encoding and geometry-conditioned slice assignment to occupancy-aware aggregation, global interaction, deslicing, and local refinement. We will also define the necessary notation and tensor dimensions alongside the corresponding operations to make the connections between modules clear.

4. Organization of related work

We agree that the previous organization did not clearly distinguish the perspectives of graph-based physical modeling and neural operators. Both can learn mappings from physical inputs to response fields, and graph-based computation can also be used within neural operators. Rather than treating them as separate categories, we have revised Section 2 to clarify their relationship and organize the discussion around local mesh propagation, global interaction, and geometry-aware modeling. The revised introduction and subheadings also make the connections between these approaches and GeoSlice-MGN more explicit.

Area_Chair_YFY2OpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer kika,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

Thanks, AC. It would be very helpful if the authors' revision had highlighted changes!

Reviewer_bfY8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
3
Read review
Paper summary

This paper proposes GeoSlice-MGN, which uses impact-relative geometry and bounding box-relative position to modulate content-driven node-to-slice assignment. The model uses a geometric routing bias and an occupancy encoding. Experiments show that GeoSlice-MGN achieves the lowest average error and the contributions of the method lead to lower error.

Critical strengths

S1: The paper is clearly written, and the prediction task of GeoSlice-MGN is clearly defined. It also introduces a useful simulation dataset based on realistic automotive structures.

S2: The proposed method performs well in the experiments. It achieves the lowest mean displacement and stress errors in the main benchmark.

S3: The paper includes extensive experiments and detailed ablation studies. The ablations examine key components and help assess their contributions to the overall performance.

Critical weaknesses

W1: The backbone mainly combines MGN with Transolver-style slice attention. I think the main additions are a geometric routing bias and an occupancy encoding. However, the Introduction and Methodology sections do not clearly explain the specific contributions beyond existing methods. More detailed motivations are needed for these design choices. The authors should explain which limitations each change addresses and why it is needed.

W2: The paper discusses GeoTransolver but does not include it in the experiments. This is a closely related method that adds geometric information to slice attention. A direct comparison would help show the value of the proposed routing design. If such a comparison is not feasible, the authors should explain the practical limitations.

W3: The main benchmark contains only three fixed panel geometries. This limits the evidence for generalization to more diverse structures. The manuscript does not directly compare the learned slice patterns across geometries. The authors should evaluate more diverse structures and compare slice assignments across panels. This would help explain how the geometric routing adapts to structural differences.

W4: Transolver recomputes the slice assignments in each attention block. GeoSlice-MGN computes them only once per forward pass and then applies four attention blocks to the slice features. Could the authors explain why the assignments are not updated between blocks? A comparison with repeated slicing would help clarify the effect of this design.

W5: The model uses 48 slices per head on two panels and 96 on the third. The paper does not explain how these values are chosen or test sensitivity to the slice count. This choice may affect both prediction accuracy and computational cost. Results with different slice counts would help clarify this trade-off.

W6: Please clarify how the floorfrontdriver and floorfrontR differ and why both are included in the benchmark. It would also be useful to test a case and its mirrored version. The geometry, loading, and support conditions should be reflected consistently. Do the predicted displacement vectors follow the same reflection? Do the stress values remain unchanged at corresponding elements?

Questions for authors

Please see the questions in the weaknesses section, and I will consider raising my score if your responses resolve my concerns.

7 comments
AuthorsOpenReview ↗

Response to Reviewer bfY8 (Part 1/2)

Thank you for your detailed comments on the contribution, experimental comparisons, and geometric scope of GeoSlice-MGN. We respond to your six concerns below.

1. Contributions and design motivation

Finite-depth MGN captures local mesh interactions but has limited long-range reach, while Transolver-style slicing enables global exchange. However, for moving impacts on a fixed structure, content-only assignments must infer the changing load–structure geometry indirectly.

Motivated by finite-plate response, our gated geometric residual directly conditions node-to-slice logits on impact-relative and structure-relative descriptors. This allows the current impact–structure configuration to guide global grouping, rather than relying solely on geometry encoded in node features.

Furthermore, normalized slice states retain weighted-average features but do not separately preserve total assignment weight. Our relative occupancy encoding supplies this complementary grouping statistic to global slice interaction. The ablations support the combined design and favor geometric injection at the routing stage.

We will clarify these motivations in the Introduction and Sections 3.4–3.5.

2. Direct comparison with GeoTransolver

We agree that GeoTransolver is a closely related baseline and that a direct comparison is important. We have therefore evaluated GeoTransolver on all three panels for both nodal displacement and element stress prediction. We will add these results to Table 1 and provide the adaptation details in the appendix.

3. Geometry diversity and the scope of generalization

Our benchmark is designed to evaluate prediction under varying impact conditions on fixed panel geometries, rather than generalization to unseen structures. We train and evaluate a separate model for each of the three panels, considering variations in impact location, velocity, and material parameters. Cross-geometry generalization is therefore beyond the scope of this study.

The geometric routing is designed to capture changes in the spatial relationship between the impact and the structure, even when the underlying mesh remains fixed. Figure 4 illustrates how the learned slice assignments vary with impact location on the same panel. Since the models are trained independently, their slice indices are not aligned across panels, and direct slice-by-slice comparisons would not meaningfully characterize cross-geometry adaptation. Our results therefore demonstrate impact-conditioned routing within fixed geometries, rather than generalization across different structures.

AuthorsOpenReview ↗

Response to Reviewer bfY8 (Part 2/2)

4. Single versus repeated slicing

GeoSlice-MGN uses a single slice assignment to separate geometry-conditioned grouping from global information exchange. After local MGN encoding, the routing module establishes node–slice assignments based on learned features and the spatial relationships between the impact and the structure. Since these geometric relationships remain fixed for a given impact case, the four Transformer blocks progressively model long-range interactions within the same geometry-conditioned grouping. The resulting global information is then mapped back to the FE mesh for local refinement. In contrast, repeated slicing requires intermediate node updates and reconstruction of the slice representations, rather than maintaining a continuous global interaction stage.

We also experimented with repeated slicing, where assignments are recomputed and node features are updated after each attention block. This variant yielded higher mean displacement and stress errors across all three panels. We will present the ablation results in Appendix F.4.

5. Slice-count selection and sensitivity

We selected the slice counts during model development, using S = 48 per head for driver and R, and S = 96 for trunk to accommodate its larger and more geometrically complex mesh. To examine the sensitivity to this choice, we evaluated S ∈ {24, 48, 96, 128} on all three panels, with three training seeds per configuration and all other settings fixed.

The results show that prediction errors vary non-monotonically with the number of slices, with different trends for displacement and stress. Compared with the original three-seed means in Table 1, none of the alternative configurations improves both metrics simultaneously on any panel. We therefore retain the original settings and provide the complete sensitivity analysis in Appendix F.5.

6. Front-panel differences and reflection consistency

The floorfrontdriver and floorfrontR panels originate from opposite sides of the same full-vehicle finite element model, but they are not exact mirror counterparts. They differ in their holes, stiffening features, local boundaries, and mesh connectivity. The driver panel contains 7,408 nodes and 7,374 shell elements, whereas R contains 12,011 nodes and 12,055 elements. We will add detailed mesh visualizations in the appendix to illustrate their origins and structural differences.

These differences are also important for interpreting the cross-panel transfer experiments in Appendix I. We evaluate direct transfer and limited-data fine-tuning from floorfrontR to floorfrontdriver. To align the two sides of the vehicle, we reflect the target panel's geometry, impact conditions, and displacement fields across the vehicle center plane. This transformation aligns their spatial frames without eliminating the structural differences between the panels. Thus, the experiment examines transfer between two distinct geometries, rather than the reflection equivariance suggested by the reviewer. The latter would require comparing a structure with its fully reflected counterpart under correspondingly reflected loading and boundary conditions, whereas reflection equivariance is not explicitly enforced by our architecture.

Thank you for your reply. I suggest you highlight the revised sections using \textcolor{blue}{} and this will help me see which parts have been updated.

Reviewer_7e79OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes GeoSlice-MGN to predict peak displacement and stress fields on automotive floor panels under localized impact by combining local mesh message passing with Transolver-style global slicing. The key idea is to let impact-relative geometry directly bias the node-to-slice assignment, and the design is sensibly motivated by plate mechanics. What I find most convincing is the same routing module also improves a vanilla Transolver, so the idea is not tied to the authors' proposed own backbone. The released benchmark on three real vehicle panels is also a useful contribution.

My reservations are twofold: the comparisons leave out several strong recent methods the paper itself cites (like GeoTransolver and Transolver++), and the reported gains are uneven and rest on small test sets with few seeds. Overall, I see this paper a sound and well-executed paper, but the evidence is thinner than the claims suggest.

Critical strengths
  1. The design is mechanically motivated and easy to follow. The finite-plate analysis gives a plausible reason for why both impact-relative and structure-relative positions should matter, and the architecture follows naturally from that reasoning rather than feeling like a pile of tricks.

  2. In my view, the geometric routing idea seems to be the real deal. Grafting the same GeoRoute10 module onto a vanilla Transolver still improves 5 of the 6 metrics, which tells me the contribution is portable and not just an artifact of the authors' own backbone.

  3. I appreciate the released AutoPanelImpact benchmark: 1,500 nonlinear impact cases on three panels taken from an actual production vehicle model. This kind of controlled, realistic testbed is useful to the community.

Critical weaknesses

My first issue is that the paper doesn't compare against stronger baselines, missing the most recent works as well. The main table stops at Transolver, GINO and Geo-FNO, all 2023/2024 era methods. The authors know the newer literature, since GeoTransolver is right in their related work, yet it never shows up in an experiment. Same for Transolver++ and MeshGraphNet-Transformer. And GAOT (Geometry Aware Operator Transformer as an efficient and accurate neural surrogate for PDEs on arbitrary domains, NeurIPS 2025), which is arguably a closely related geometry-aware operator architecture and already shows up as a standard baseline in Transolver-3, isn't even cited. So when the abstract says lowest error on all six metrics, what it really means is lowest among the older methods they chose to run.

Second, I'm not sure the numbers carry the weight the paper puts on them. Compared with vanilla Transolver, the gains vary substantially: displacement error on floorfrontdriver decreases by about 12%, but stress error on the same panel decreases by less than 3%. The roughly 20% stress error reduction on trunkfloor is substantially larger than that on the other two panels. Therefore, I can see these results establish the lowest reported mean errors among the evaluated methods, but three seeds and 50 test cases per panel provide limited evidence for the statistical robustness of the smaller margins,

Questions for authors
  1. Could you compare with GeoTransolver and Transolver++ or MeshGraphNet-Transformer, and clarify the relationship to GAOT (or even making comparison) ?

  2. Could you add seeds or confidence intervals to verify the smaller gains, particularly the stress improvement on floorfrontdriver?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 7e79

Thank you for your careful review and for highlighting the need for stronger recent baselines and more evidence for the smaller improvements. We have completed additional baseline comparisons and training runs, and address these points below.

1. Recent baselines and the relationship to GAOT

Thank you for highlighting these recent methods. We have added comparisons with GeoTransolver, Transolver++, and GAOT on all three panels, using three training seeds per method and panel. Each backbone was adapted to the same joint displacement–stress prediction task. The results are now included in Table 1, with implementation details provided in Appendix D.

We will also clarify how these methods relate to GeoSlice-MGN in Section 2. While they employ different mechanisms for geometric modeling and global interaction, GeoSlice-MGN specifically incorporates impact-relative and structure-relative geometry into node-to-slice assignments, enriches slice representations with relative occupancy, and retains local FE-mesh propagation. In particular, unlike GAOT's combination of geometry-aware graph operators and a latent Transformer, our approach directly conditions the slice routing on the impact–structure configuration.

2. Additional seeds and statistical robustness

Thank you for this suggestion. To further assess robustness, we evaluated both GeoSlice-MGN and Transolver using two additional random seeds on each panel.

The results below report the mean ± standard deviation over the two additional seeds:

PanelMethodDisplacement Rel. L2Stress Rel. L2
floorfrontdriverGeoSlice-MGN0.1504 ± 0.00940.1824 ± 0.0003
floorfrontdriverTransolver0.1730 ± 0.00720.2018 ± 0.0051
floorfrontRGeoSlice-MGN0.1597 ± 0.00070.1834 ± 0.0007
floorfrontRTransolver0.1676 ± 0.00130.2038 ± 0.0007
trunkfloorGeoSlice-MGN0.3031 ± 0.01860.2996 ± 0.0095
trunkfloorTransolver0.3554 ± 0.00980.3710 ± 0.0150

GeoSlice-MGN achieves lower mean errors on all six metrics in these additional runs. In particular, floorfrontdriver stress error decreases by approximately 9.6% relative to Transolver. These results provide additional evidence that the observed improvements are not confined to the original random seeds.

Area_Chair_YFY2OpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer 7e79,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

Reviewer_7e79OpenReview ↗

Largely satisfied with the response and revisions

The authors' response arrived a little late, but I have now tried my best to read the updated manuscript. The new baseline comparisons address my first concern, and the extra runs give me more confidence in the results. Still, I would suggest reporting the mean and standard deviation over all five seeds together to my second concern. Overall, most of my concerns have been addressed, and I am willing to raise my final score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.