pith. sign in

arxiv: 2605.15231 · v1 · pith:DHASQDOCnew · submitted 2026-05-13 · 💻 cs.LG · cs.CV

Mask-Morph Graph U-Net: A Generalisable Mesh-Based Surrogate for Crashworthiness Field Prediction under Large Geometric Variation

Haoran Li , Tobias Lehrer , Yingxue Zhao , Haosu Zhou , Philipp Stocker , Tobias Pfaff , Nan Li This is my paper

Pith reviewed 2026-05-19 16:34 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords Graph U-NetMesh surrogate modelingCrashworthiness simulationGeometric variationBarycentric parameterisationMasked pretrainingTransfer learningGraph neural networks
0
0 comments X p. Extension
pith:DHASQDOC Add to your LaTeX paper What is a Pith Number? 
\usepackage{pith}
\pithnumber{DHASQDOC}

Prints a linked pith:DHASQDOC badge after your title and writes the identifier into PDF metadata. Compiles on arXiv with no extra files. Learn more

The pith

Mask-Morph Graph U-Net uses coarse-graph morphing and masked pretraining to generalise hierarchical GNNs to new mesh geometries in crash simulations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a surrogate model that predicts crashworthiness fields on meshes with large geometric changes, avoiding the need to retrain for each new design. It keeps the accurate edge-specific layers of a Graph U-Net by first morphing a fixed coarse graph onto each input mesh through feature-aligned barycentric parameterisation. Node masking during supervised pretraining, followed by freezing the high-parameter layers during fine-tuning, then reduces train-test gaps and cuts the data required for transfer to new components. If the approach holds, it turns expensive nonlinear finite-element runs into fast, reusable predictions that support iterative vehicle design without sacrificing accuracy across varying shapes.

Core claim

By morphing the coarsened graph hierarchy to each input mesh using feature-aligned barycentric parameterisation before constructing cross-graph edges, and applying node masking during supervised pretraining with parameter-efficient fine-tuning where high-parameter edge-specific layers are frozen, the Mask-Morph Graph U-Net retains the benefits of edge-specific layers while improving spatial correspondence and generalisability across varying graph structures for crash field prediction under large geometric variation.

What carries the argument

Coarse-graph morphing via feature-aligned barycentric parameterisation together with masked supervised pretraining inside a hierarchical Graph U-Net.

If this is right

  • Coarse-graph morphing improves test accuracy relative to a fixed-coarse-graph baseline.
  • Masked supervised pretraining reduces train-test discrepancy.
  • Masked pretraining improves data efficiency in cross-component transfer settings.
  • The model achieves lower prediction error than external baselines.
  • The method supplies a practical route to reusable, data-efficient mesh-based surrogates for crashworthiness design exploration.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same morphing step could be tested on other mesh-based physics problems that require transfer across shape families.
  • Freezing the expensive layers after masked pretraining may lower the cost of maintaining a library of component-specific surrogates.
  • If the correspondence errors remain small, the technique could be combined with gradient-based shape optimisation loops that query the surrogate thousands of times.
  • Extending the masking strategy to include geometric features might further improve robustness when component topologies differ more radically.

Load-bearing premise

Feature-aligned barycentric parameterisation creates accurate enough spatial correspondence between the morphed coarse graph and each new input mesh without systematic interpolation errors that would propagate through the U-Net layers.

What would settle it

A controlled test in which the morphed-coarse-graph version produces higher mean Euclidean distance or maximum intrusion errors than the fixed-coarse-graph baseline on out-of-distribution meshes would falsify the accuracy claim.

Figures

Figures reproduced from arXiv: 2605.15231 by Haoran Li, Haosu Zhou, Nan Li, Philipp Stocker, Tobias Lehrer, Tobias Pfaff, Yingxue Zhao.

Figure 1
Figure 1. Figure 1: Illustration of spatial-proximity-based cross-graph edge construction. When the [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed MMGUNet architecture. Shared-weight MLP-based [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Case-study geometries and B-pillar morphing directions. (a) Base component [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: B-pillar side-impact simulation setup. (a) Fixed boundary regions and impactor [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Feature-aligned UV parameterisation of a representative B-pillar mesh. (a) [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Four-step visualisation of template-to-target mesh morphing in a shared UV [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of positional encodings: (a) the first five Laplacian eigenvectors (LE); [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Mask-ratio ablation: (a) MED (mm) versus mask ratio and (b) MIPE (%) versus [PITH_FULL_IMAGE:figures/full_fig_p027_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Baseline comparison on the B-pillar A3 case study. (a) Distribution of MED [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: However, similar ranking trend remains. The external baselines [PITH_FULL_IMAGE:figures/full_fig_p029_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: Baseline comparison on the U-channel case study. (a) Distribution of MED [PITH_FULL_IMAGE:figures/full_fig_p030_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Trial-wise training-strategy comparison: (a) MED and (b) MIPE. Each trial [PITH_FULL_IMAGE:figures/full_fig_p031_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Trial B (shape transfer learning) data-requirement analysis: (a) MED and (b) [PITH_FULL_IMAGE:figures/full_fig_p033_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Trial B qualitative comparison of ground-truth and predicted [PITH_FULL_IMAGE:figures/full_fig_p034_13.png] view at source ↗
read the original abstract

Nonlinear finite element crash simulations are accurate but computationally expensive, limiting their use in iterative design optimisation. Machine-learning surrogate models based on graph neural networks (GNNs) offer a faster alternative. Message-passing GNNs are widely used for mesh simulation, and their shared node and edge update functions are relatively generalisable across varying graph structures. By contrast, non-shareable edge-specific aggregation layers can capture nonlinear relationships more accurately but usually require fixed graph connectivity, which limits generalisability. This paper presents Mask-Morph Graph U-Net (MMGUNet), a practical approach to addressing the limitation of hierarchical Graph U-Net architectures that use edge-specific downsampling and upsampling layers. Fixed coarse graph connectivity is required for edge-specific layers. To retain this while improving spatial correspondence, the proposed method morphs the coarsened graph hierarchy to each input mesh using feature-aligned barycentric parameterisation before constructing cross-graph edges. It further applies node masking during supervised pretraining, followed by parameter-efficient fine-tuning in which high-parameter edge-specific layers are frozen. The proposed approach is evaluated in in-distribution, out-of-distribution, and cross-component transfer settings using mean Euclidean distance and maximum intrusion percentage error. Results show that coarse-graph morphing improves test accuracy relative to a fixed-coarse-graph baseline, while masked supervised pretraining reduces the train-test discrepancy and improves data efficiency during transfer. The proposed model also achieves lower prediction error compared with external baselines. These results demonstrate a practical route toward reusable, data-efficient mesh-based surrogate modelling for crashworthiness design exploration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Mask-Morph Graph U-Net (MMGUNet), a hierarchical GNN surrogate for predicting crashworthiness fields (e.g., displacement and intrusion) on meshes with large geometric variations. It addresses the fixed-connectivity limitation of edge-specific down/upsampling layers in Graph U-Nets by morphing a fixed coarse graph hierarchy to each input mesh via feature-aligned barycentric parameterisation, then constructing cross-graph edges. Supervised pretraining with node masking is followed by parameter-efficient fine-tuning that freezes high-parameter edge-specific layers. Evaluations across in-distribution, out-of-distribution, and cross-component transfer settings report lower mean Euclidean distance and maximum intrusion percentage errors than fixed-coarse-graph baselines and external methods, attributing gains to improved spatial correspondence from morphing and reduced train-test discrepancy from masking.

Significance. If the central claims hold after addressing the noted gaps, the work offers a concrete, practical route to reusable mesh surrogates for nonlinear crash simulations. It combines explicit morphing to retain edge-specific layers with masking-based pretraining for data efficiency, which could meaningfully reduce reliance on per-geometry retraining in iterative design loops. The multi-setting evaluation (in-dist/OOD/transfer) and comparison to baselines are strengths that, once supported by error quantification and ablations, would strengthen the case for generalisable GNN surrogates in engineering.

major comments (2)
  1. [§3.2] §3.2 (Morphing and Cross-Graph Construction): The accuracy of feature-aligned barycentric parameterisation is not quantified (no mean/max node displacement, Hausdorff distance, or per-region error relative to input mesh resolution). This is load-bearing for the strongest claim that morphing improves test accuracy via better correspondence; without these metrics it remains unclear whether observed gains over the fixed-coarse-graph baseline arise from reduced interpolation error or from other factors, and whether residual offsets propagate through the hierarchical message-passing layers.
  2. [§5] §5 (Experiments and Ablations): No ablation isolates the morphing step from the masking schedule, nor reports quantitative error bars or details on train-test splits and data exclusion criteria. Because the central claims rest on morphing improving accuracy and masking improving data efficiency in transfer, the absence of these controls makes it difficult to attribute performance differences specifically to the proposed components rather than to training schedule or dataset partitioning choices.
minor comments (2)
  1. [§3] Notation for the morphed coarse-graph nodes and the cross-graph edge construction could be made more explicit in the methods to avoid ambiguity when readers compare against the fixed-graph baseline.
  2. [Figures in §5] Figure captions and axis labels in the result plots would benefit from stating the exact error metric (mean Euclidean vs. max intrusion) and whether shaded regions represent standard deviation across runs.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and indicate where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Morphing and Cross-Graph Construction): The accuracy of feature-aligned barycentric parameterisation is not quantified (no mean/max node displacement, Hausdorff distance, or per-region error relative to input mesh resolution). This is load-bearing for the strongest claim that morphing improves test accuracy via better correspondence; without these metrics it remains unclear whether observed gains over the fixed-coarse-graph baseline arise from reduced interpolation error or from other factors, and whether residual offsets propagate through the hierarchical message-passing layers.

    Authors: We agree that quantifying the morphing accuracy is necessary to substantiate the claim that improved spatial correspondence drives the observed gains. In the revised manuscript we will add explicit metrics for the feature-aligned barycentric parameterisation, including mean and maximum node displacements, Hausdorff distance, and per-region error relative to input mesh resolution. These will be reported in an expanded §3.2 and/or a new appendix, allowing readers to assess residual offsets and their potential propagation through the hierarchy. revision: yes

  2. Referee: [§5] §5 (Experiments and Ablations): No ablation isolates the morphing step from the masking schedule, nor reports quantitative error bars or details on train-test splits and data exclusion criteria. Because the central claims rest on morphing improving accuracy and masking improving data efficiency in transfer, the absence of these controls makes it difficult to attribute performance differences specifically to the proposed components rather than to training schedule or dataset partitioning choices.

    Authors: We concur that isolating the contributions of morphing and masking, together with statistical rigour, is required for clear attribution. The revised §5 will include (i) an ablation that decouples the morphing step from the masking pretraining schedule, (ii) quantitative error bars (standard deviation across multiple random seeds or runs), and (iii) expanded details on train-test split construction and any data exclusion criteria. These additions will directly address the concern about confounding factors. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper introduces explicit methodological components (feature-aligned barycentric morphing of coarse graphs and node masking in pretraining) whose definitions and implementations are independent of the final evaluation metrics (mean Euclidean distance and maximum intrusion percentage error). These steps are described as practical solutions to fixed-graph limitations in hierarchical GNNs, with empirical validation on held-out meshes in multiple settings. No equations, self-citations, or fitted parameters are presented as reducing the central claims to their own inputs by construction. The provided text contains no load-bearing self-references, ansatzes smuggled via citation, or renamings of known results as novel derivations.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the premise that barycentric morphing preserves enough geometric fidelity for the subsequent GNN layers to learn accurate nonlinear mappings; no free parameters are explicitly named in the abstract, but the method implicitly depends on choices of coarse-graph resolution and masking ratio.

axioms (1)
  • domain assumption Barycentric parameterisation produces a spatial correspondence that is sufficiently accurate for hierarchical message passing across large geometric variations.
    Invoked when the paper states that morphing the coarsened graph improves spatial correspondence before constructing cross-graph edges.

pith-pipeline@v0.9.0 · 5837 in / 1381 out tokens · 39555 ms · 2026-05-19T16:34:58.057667+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 55 canonical work pages · 2 internal anchors

  1. [1]

    S. R. Wu, Convergence study on explicit finite element for crashwor- thiness analysis, Tech. Rep. 2006-01-0672, SAE International (2006). doi:10.4271/2006-01-0672

    work page doi:10.4271/2006-01-0672 2006

  2. [2]

    Chang, T

    J. Chang, T. Tyan, M. El-Bkaily, J. Cheng, A. Marpu, Q. Zeng, J. Santini, Implicit and explicit finite element methods for crash safety analysis, Tech. Rep. 2007-01-0982, SAE International (2007). doi:10.4271/2007-01-0982

    work page doi:10.4271/2007-01-0982 2007

  3. [3]

    E. Albak, Optimization design for circular multi-cell thin-walled tubes with discrete and continuous design variables, Mechanics of Advanced Materials and Structures 30 (24) (2023) 5091–5105

    work page 2023

  4. [4]

    Xiong, D

    F. Xiong, D. Wang, Z.-D. Ma, L. Yang, Z. Li, B. Song, Multi-objective lightweight and crashworthiness optimization for the side structure of an automobile body, Structural and Multidisciplinary Optimization 58 (4) (2018) 1823–1843

    work page 2018

  5. [5]

    Ahmadi Dastjerdi, M

    A. Ahmadi Dastjerdi, M. Moshref-Javadi, H. Ahmadian, M. Gholam- pour, Crushing analysis and multi-objective optimization of different length bi-thin walled cylindrical structures under axial impact loading, Engineering Optimization 51 (2019) 1–18

    work page 2019

  6. [6]

    Zende, H

    P. Zende, H. Dalir, Multi-objective optimization of composite square tube for minimizing peak crushing force and maximizing specific en- ergy absorption using artificial neural network and genetic algorithm, in: ASME International Mechanical Engineering Congress and Exposi- tion, Proceedings (IMECE), 2022

    work page 2022

  7. [7]

    Rogala, J

    M. Rogala, J. Gajewski, M. Ferdynus, The effect of geometrical non- linearity on the crashworthiness of thin-walled conical energy-absorbers, Materials 13 (21) (2020) 4857

    work page 2020

  8. [8]

    Sakaridis, N

    E. Sakaridis, N. Karathanasopoulos, D. Mohr, Machine-learning based prediction of crash response of tubular structures, International Journal of Impact Engineering 166 (2022). 37

    work page 2022

  9. [9]

    C. P. Kohar, T. K. Eller, D. S. Connolly, K. Inal, Using artificial intel- ligence to aid vehicle lightweighting in crashworthiness with aluminum, MATEC Web of Conferences 326 (2020) 01006

    work page 2020

  10. [10]

    W. Guo, P. Xu, C. Yang, J. Guo, L. Yang, S. Yao, Machine learning- basedcrashworthinessoptimizationforthesquareconeenergy-absorbing structure of the subway vehicle, Structural and Multidisciplinary Opti- mization 66 (8) (2023) 182

    work page 2023

  11. [11]

    C. P. Kohar, L. Greve, T. K. Eller, D. S. Connolly, K. Inal, A machine learning framework for accelerating the design process using cae simula- tions: An application to finite element analysis in structural crashwor- thiness, Computer Methods in Applied Mechanics and Engineering 385 (2021) 114008. doi:10.1016/j.cma.2021.114008

    work page doi:10.1016/j.cma.2021.114008 2021

  12. [12]

    H. Li, H. Zhou, N. Li, An integrated convolutional neural network- based surrogate model for crashworthiness performance prediction of hot-stamped vehicle panel components, in: MATEC Web of Confer- ences, Vol. 401, 2024, p. 03013. doi:10.1051/matecconf/202440103013

    work page doi:10.1051/matecconf/202440103013 2024

  13. [13]

    H. Li, Y. Zhao, H. Zhou, T. Pfaff, N. Li, A graph neural network surrogate model for mesh-based crashworthiness predic- tion of vehicle panel components, Results in Engineering (2026). doi:10.1016/j.rineng.2026.110925

    work page doi:10.1016/j.rineng.2026.110925 2026

  14. [14]

    Z. Wen, Y. Li, H. Wang, Y. Peng, Data-driven spatiotemporal modeling for structural dynamics on irregular domains by stochastic dependency neural estimation, Computer Methods in Applied Mechanics and Engi- neering 404 (2023) 115831. doi:10.1016/j.cma.2022.115831

    work page doi:10.1016/j.cma.2022.115831 2023

  15. [15]

    Le Guennec, T

    Y. Le Guennec, T. Defoort, J. V. Aguado, D. Borzacchiello, Compar- ing traditional surrogate modelling and neural fields for vehicle crash simulation data, in: SIA Simulation numérique 2025, 2025

    work page 2025

  16. [16]

    André, M

    V. André, M. Costas, M. Langseth, D. Morin, Neural network modelling of mechanical joints for the application in large-scale crash analyses, International Journal of Impact Engineering 177 (2023) 104490

    work page 2023

  17. [17]

    S. Thel, L. Greve, B. van de Weg, P. van der Smagt, Introducing fi- nite element method integrated networks (femin), Computer Methods in Applied Mechanics and Engineering 427 (2024) 117073. 38

    work page 2024

  18. [18]

    S. Thel, L. Greve, M. Karl, P. van der Smagt, Accelerating crash sim- ulations with finite element method integrated networks (femin): Com- paring two approaches to replace large portions of a fem simulation, Computer Methods in Applied Mechanics and Engineering 443 (2025) 118046

    work page 2025

  19. [19]

    Zhang, Y

    G. Zhang, Y. Liu, Y. Quan, J. Yan, A mesh-based geometric deep learning framework for rapid response prediction of large-scale and multi-component mechanical structures in engineering, Computer Methods in Applied Mechanics and Engineering 448 (2026) 118435. doi:10.1016/j.cma.2025.118435

    work page doi:10.1016/j.cma.2025.118435 2026

  20. [20]

    M. A. Nabian, S. Chavare, D. Akhare, R. Ranade, R. Cherukuri, S. Tadepalli, Automotive crash dynamics modeling accelerated with machine learning, in: SAE Technical Paper, 2025, pp. 2026–01–0568. doi:10.4271/2026-01-0568

    work page doi:10.4271/2026-01-0568 2025

  21. [21]

    Pfaff, M

    T. Pfaff, M. Fortunato, A. Sanchez-Gonzalez, P. Battaglia, Learning mesh-based simulation with graph networks, in: International Confer- ence on Learning Representations (ICLR), 2021, arXiv:2010.03409

    work page arXiv 2021

  22. [22]

    H. Wu, H. Luo, H. Wang, J. Wang, M. Long, Transolver: A fast transformer solver for pdes on general geometries, arXiv preprint arXiv:2402.02366 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [23]

    H. Luo, H. Wu, H. Zhou, L. Xing, Y. Di, J. Wang, M. Long, Tran- solver++: An accurate neural solver for pdes on million-scale geome- tries, arXiv preprint arXiv:2502.02414 (2025)

    work page arXiv 2025

  24. [24]

    GitHub (2023)

    PhysicsNeMoContributors, Nvidiaphysicsnemo: Anopen-sourceframe- work for physics-based deep learning in science and engineering, released February 24, 2023. GitHub (2023)

    work page 2023

  25. [25]

    Deshpande, J

    S. Deshpande, J. Lengiewicz, S. Bordas, MAgNET: A graph U-Net ar- chitecture for mesh-based simulations, arXiv preprint arXiv:2211.00713 (2022)

    work page arXiv 2022

  26. [26]

    J. He, Y. Jiao, S. P. A. Bordas, On the use of graph neural networks and shape-function-based gradient computation in the deep energy method, International Journal for Numerical Methods in Engineering 124 (4) (2023) 864–879. 39

    work page 2023

  27. [27]

    X. Fu, J. Hu, X. Zhang, H. Feng, An finite element analysis surrogate model with boundary oriented graph embedding approach for rapid de- sign, Journal of Computational Design and Engineering 10 (3) (2023) 1026–1046

    work page 2023

  28. [28]

    Q. Chen, X. Wen, Y. Zhang, Predicting dynamic responses of contin- uous deformable bodies: A graph-based learning approach, Computer Methods in Applied Mechanics and Engineering 420 (2024) 116669

    work page 2024

  29. [29]

    H. Zhou, Y. Zhao, H. Li, T. Pfaff, N. Li, A multi-level graph- based surrogate model for real-time high-fidelity sheet forming simulations, Advanced Engineering Informatics 66 (2025) 103458. doi:10.1016/j.aei.2025.103458

    work page doi:10.1016/j.aei.2025.103458 2025

  30. [30]

    Sanchez-Gonzalez, J

    A. Sanchez-Gonzalez, J. Godwin, T. Pfaff, R. Ying, J. Leskovec, P. Battaglia, Learning to simulate complex physics with graph networks, in: International Conference on Machine Learning (ICML), 2020

    work page 2020

  31. [31]

    Fortunato, T

    M. Fortunato, T. Pfaff, K. Stachenfeld, P. Battaglia, Multiscale mesh- graphnets, in: 2nd AI4Science Workshop at the 39th International Con- ference on Machine Learning (ICML), 2022

    work page 2022

  32. [32]

    Y. Cao, S. Bhat, A. Subhash, J. Li, S. Ovie, K. Dvijotham, L. N. Chel- liah, K. Ramakrishnan, A. Kalinowska, M. J. Glueck, A. G. Schwing, A. Anandkumar, G. Cecchi, Efficient learning of mesh-based physical simulation with bsms-gnn, arXiv preprint (2022). arXiv:2210.02573

    work page arXiv 2022

  33. [33]

    Y. Zhao, Q. Chen, H. Li, H. Zhou, H. R. Attar, T. Pfaff, T. Wu, N. Li, Recurrent U-Net-based graph neural network (RUGNN) for accurate de- formation predictions in sheet material forming, Advanced Engineering Informatics 69 (2026) 104021. doi:10.1016/j.aei.2025.104021

    work page doi:10.1016/j.aei.2025.104021 2026

  34. [34]

    de Boer, M

    A. de Boer, M. S. van der Schoot, H. Bijl, Mesh deformation based on radial basis function interpolation, Computers & Structures 85 (11–14) (2007) 784–795. doi:10.1016/j.compstruc.2007.01.013

    work page doi:10.1016/j.compstruc.2007.01.013 2007

  35. [35]

    T. W. Sederberg, S. R. Parry, Free-form deformation of solid geometric models, in: Proceedings of the 13th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 1986, pp. 151–160. doi:10.1145/15922.15903. 40

    work page doi:10.1145/15922.15903 1986

  36. [36]

    Kraevoy, A

    V. Kraevoy, A. Sheffer, Cross-parameterization and compatible remesh- ing of 3d models, ACM Transactions on Graphics 23 (3) (2004) 861–869. doi:10.1145/1015706.1015811

    work page doi:10.1145/1015706.1015811 2004

  37. [37]

    Kraevoy, A

    J. Schreiner, A. Asirvatham, E. Praun, H. Hoppe, Inter-surface mapping, ACM Transactions on Graphics 23 (3) (2004) 870–877. doi:10.1145/1015706.1015812

    work page doi:10.1145/1015706.1015812 2004

  38. [38]

    Alexa, D

    M. Alexa, D. Cohen-Or, D. Levin, As-rigid-as-possible shape interpola- tion, in: Proceedings of EUROGRAPHICS 2000, Blackwell Publishers Ltd, 2000, pp. 157–164. doi:10.1111/1467-8659.00339

    work page doi:10.1111/1467-8659.00339 2000

  39. [39]

    W. T. Tutte, How to draw a graph, Proceedings of the London Mathe- matical Society s3-13 (1) (1963) 743–767. doi:10.1112/plms/s3-13.1.743

    work page doi:10.1112/plms/s3-13.1.743 1963

  40. [40]

    M. S. Floater, Parametrization and smooth approximation of surface triangulations, Computer Aided Geometric Design 14 (3) (1997) 231–

    work page 1997

  41. [41]

    doi:10.1016/S0167-8396(96)00031-3

    work page doi:10.1016/s0167-8396(96)00031-3

  42. [42]

    M. S. Floater, Mean value coordinates, Computer Aided Geometric De- sign 20 (1) (2003) 19–27. doi:10.1016/S0167-8396(03)00002-5

    work page doi:10.1016/s0167-8396(03)00002-5 2003

  43. [43]

    B. Lévy, S. Petitjean, N. Ray, J. Maillot, Least squares conformal maps for automatic texture atlas generation, in: Proceedings of the 29th Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), 2002, pp. 362–371. doi:10.1145/566654.566590

    work page doi:10.1145/566654.566590 2002

  44. [44]

    Rabinovich, R

    M. Rabinovich, R. Poranne, D. Panozzo, O. Sorkine-Hornung, Scalable locally injective mappings, ACM Transactions on Graphics 36 (2) (2017) 16:1–16:16. doi:10.1145/2977606

    work page doi:10.1145/2977606 2017

  45. [45]

    K. He, X. Chen, S. Xie, Y. Li, P. Dollár, R. Girshick, Masked autoen- coders are scalable vision learners, in: Proceedings of the IEEE/CVF ConferenceonComputerVisionandPatternRecognition(CVPR),2022, pp. 16000–16009

    work page 2022

  46. [46]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint (2019). arXiv:1810.04805. 41

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  47. [47]

    Z.Hou, X.Liu, Y.Cen, Y.Dong, H.Yang, C.Wang, J.Tang, Graphmae: Self-supervised masked graph autoencoders, in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2022, pp. 594–604

    work page 2022

  48. [48]

    Z. Hou, X. Liu, Y. Cen, Y. Zhao, Y. Dong, H. Yang, C. Wang, J. Tang, Graphmae2: A decoding-enhanced masked self-supervised graph learner, in: Proceedings of the ACM Web Conference (WWW), 2023, pp. 737–746

    work page 2023

  49. [49]

    Garnier, V

    P. Garnier, V. Lannelongue, J. Viquerat, E. Hachem, Meshmask: Pre- training mesh simulators with masked autoencoding, arXiv preprint (2025). arXiv:2502.10841

    work page arXiv 2025

  50. [50]

    Whalen, C

    E. Whalen, C. Mueller, Toward reusable surrogate models: Graph- basedtransferlearningontrussstructures, JournalofMechanicalDesign 144 (2) (2022) 021704, aSME Digital Collection

    work page 2022

  51. [51]

    Zhang, J

    X. Zhang, J. Zhou, Z. Feng, B-pillar collision test method, Tech. rep., Aiways Automobile Shanghai Co Ltd, China (2020)

    work page 2020

  52. [52]

    M. D. McKay, R. J. Beckman, W. J. Conover, A comparison of three methods for selecting values of input variables in the analysis of output from a computer code, Technometrics 21 (2) (1979) 239–245

    work page 1979

  53. [53]

    Lehrer, P

    T. Lehrer, P. Stocker, F. Duddeck, M. Wagner, Ucsm: Dataset of u-shaped parametric cad geometries and real-world sheet metal meshes for deep drawing, Computer-Aided Design 188 (2025) 103924. doi:10.1016/j.cad.2025.103924

    work page doi:10.1016/j.cad.2025.103924 2025

  54. [54]

    V. P. Dwivedi, X. Bresson, A generalization of transformer networks to graphs, arXiv preprint arXiv:2012.09699 (2021). URLhttps://arxiv.org/abs/2012.09699

    work page arXiv 2012

  55. [55]

    Garnier, J

    P. Garnier, J. Viquerat, E. Hachem, Multi-grid graph neural networks with self-attention for computational mechanics (2024). arXiv:2409.11899. URLhttps://arxiv.org/abs/2409.11899 42 Appendix A. Dataset parameters In this appendix, we provide the detailed parameter definitions used to generate the B-pillar and U-channel datasets. For the B-pillar case, as ...

    work page arXiv 2024