Surrogate Assisted Pedestrian Protection Design via a Foundation Model Orchestrated Workflow

Akihiko Katagiri; Jun Shiraishi; Masato Sasaki; Osamu Ito; Shin Saeki; Yoshikazu Nakagawa

arxiv: 2606.17577 · v1 · pith:CCBMLXIFnew · submitted 2026-06-16 · 💻 cs.AI

Surrogate Assisted Pedestrian Protection Design via a Foundation Model Orchestrated Workflow

Osamu Ito , Akihiko Katagiri , Yoshikazu Nakagawa , Shin Saeki , Jun Shiraishi , Masato Sasaki This is my paper

Pith reviewed 2026-06-27 01:22 UTC · model grok-4.3

classification 💻 cs.AI

keywords pedestrian protectionsurrogate modelingfoundation modelscrash safety designNSGA-IImultiobjective optimizationautomotive engineering

0 comments

The pith

A foundation model orchestrates surrogates and evolutionary search to generate dozens of pedestrian-safe bumper designs from one run instead of weeks of manual iteration.

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

The paper presents a workflow that trains a surrogate on crash simulations to predict leg injury metrics, then uses NSGA-II to search for feasible designs under constraints while an LLM coordinates the steps and a vision-language model compares outputs. This setup replaces hours-long CAE runs with second-scale evaluations and yields 35 distinct compliant front-bumper alternatives in an automotive case study. A sympathetic reader cares because crash safety design has resisted data-driven methods due to nonlinear contact and material behavior, so a working integration layer could make rapid iteration practical in a regulated domain.

Core claim

The foundation model-orchestrated workflow combines a surrogate achieving average R²=0.87 with conformal prediction intervals, NSGA-II multiobjective search, a morphing geometry generator, and natural-language orchestration to produce 35 distinct safety-compliant front-bumper designs in a single automated exploration that would otherwise require weeks of conventional CAE iteration.

What carries the argument

The foundation model-orchestrated workflow that links the trained surrogate predictor, NSGA-II search, morphing geometry generator, and LLM-driven interface to enable surrogate-assisted design exploration.

If this is right

The workflow produces 35 distinct safety-compliant design alternatives from a single exploration run.
Evaluation time drops from hours per CAE simulation to seconds per surrogate query.
Foundation models can act as integration layers between ML surrogates and physics-based simulation in safety-critical domains.
An LLM natural-language interface plus vision-language comparison supports semantic review of generated geometries.

Where Pith is reading between the lines

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

The same orchestration pattern could shorten iteration cycles in other engineering problems that combine nonlinear simulation with multiobjective constraints.
Conformal intervals may allow engineers to set risk thresholds when accepting surrogate-guided designs without exhaustive re-simulation.
Extending the geometry generator to additional vehicle subsystems would test whether the workflow scales beyond the front-bumper case.

Load-bearing premise

The surrogate's average R squared of 0.87 together with its conformal prediction intervals remains accurate enough that the parameter sets returned by the search still satisfy injury constraints when re-checked with full nonlinear CAE contact simulations.

What would settle it

Re-evaluate the 35 generated designs with the original high-fidelity CAE simulations and check whether their actual injury metrics fall inside the conformal prediction intervals and meet the user-specified constraints.

Figures

Figures reproduced from arXiv: 2606.17577 by Akihiko Katagiri, Jun Shiraishi, Masato Sasaki, Osamu Ito, Shin Saeki, Yoshikazu Nakagawa.

**Figure 1.** Figure 1: shows the system architecture. We use GPT-4o for its instruction-following and structured output capabilities; however, the workflow is not tied to any specific model. Any LLM with comparable abilities for task classification, parameter extraction, and multi-turn dialogue can serve as the orchestrator [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Workflow execution. Five steps from reference geometry to VLM-assisted comparison, with the LLM coordinating surrogate evaluation, evolutionary search, and morphing. 4 SURROGATE-GUIDED INJURY PREDICTION 4.1 PROBLEM FORMULATION Pedestrian leg protection is assessed by impacting standardized leg-form impactors (FLEX-PLI and aPLI) against the vehicle front and measuring injury metrics. We consider thirteen in… view at source ↗

**Figure 3.** Figure 3: Training data generation. Left: 10 design parameters (7 geometric, 3 load). Hood reaction force is fixed because hood leading edge geometry is strongly constrained by styling requirements. Right: CAE simulation using a spring–mass vehicle model with detailed FE impactors (FLEX-PLI, aPLI). 1,800 parameter combinations were simulated. 4.3 ROLE OF THE SURROGATE IN THE DESIGN LOOP Within our workflow, the sur… view at source ↗

**Figure 4.** Figure 4: , control points are placed at 100 mm intervals and are tied to the same geometric styling parameters used by the surrogate (hood, grille, and upper/lower bumper positions). This shared parameterization ensures that designs deemed feasible by the surrogate correspond to realizable geometries. We implement morphing using an industry-standard CAE pre-processing tool [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: illustrates the VLM evaluation interface. This capability is intended to complement rather than replace human judgment: the VLM provides structured observations that help designers efficiently narrow down candidates, while final selection remains with human experts who can weigh factors the VLM cannot assess [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

**Figure 6.** Figure 6: Three representative designs from 35 feasible candidates. Each satisfies all pedestrian leg injury constraints while exhibiting distinct styling: (1) minimal deviation from reference; (2) elevated design lines with prominent grille; (3) sharper headlight impression from differential upper/lower adjustments. protruding grille, resulting in a more aggressive appearance. Design 3 raises the upper bumper whil… view at source ↗

read the original abstract

AI-driven engineering workflows face particular challenges in crash safety design: unlike aerodynamics, crash events involve highly nonlinear contact dynamics, material nonlinearity, and discrete state transitions that are difficult to capture with data-driven surrogate models. To the best of our knowledge, we present the first foundation model--orchestrated workflow for crash safety design that enables surrogate-assisted exploration for pedestrian protection, reducing evaluation time from hours per CAE simulation to seconds. The workflow integrates four components: (1) a surrogate trained on CAE crash simulations to predict pedestrian leg injury metrics from design parameters, achieving an average $R^2=0.87$ and providing distribution-free conformal prediction intervals; (2) multiobjective evolutionary search (NSGA-II) to discover diverse feasible parameter sets under user-specified constraints; (3) a morphing-based geometry generator that maps parameters to topology-preserving 3D shapes; and (4) a natural-language interface in which an LLM orchestrates the workflow and a vision--language model supports semantic comparison of generated designs. In an automotive front-bumper case study, the workflow produces 35 distinct safety-compliant alternatives from a single exploration, a process that would require weeks with conventional CAE iteration. These results suggest that foundation models can serve as integration layers between ML surrogates and physics-based simulation, helping bring AI capabilities to safety-critical engineering domains.

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.

Desk Editor's Note private letter to a colleague

The paper integrates existing tools into a crash-safety workflow but leaves the 35 claimed designs unverified against full CAE.

read the letter

The core takeaway is that this is an applied integration paper: surrogate model plus NSGA-II plus morphing plus LLM orchestration, demonstrated on a front-bumper pedestrian-protection case. It reports an R² of 0.87 with conformal intervals and says the process yields 35 feasible designs in far less time than manual CAE loops.

The integration itself is the main piece of work. Using conformal prediction for the surrogate outputs is a reasonable engineering choice, and routing the pieces through an LLM interface is a practical way to make the loop usable by designers who are not ML specialists. The morphing generator keeps the geometry changes topology-preserving, which matters for real CAD handoff.

The soft spot is exactly where the stress test points: an average R² of 0.87 on training data does not automatically mean the optimizer stays inside the true feasible region once the designs are re-run in the original nonlinear contact solver. Crash problems have sharp transitions at contact and material failure, so surrogate error near constraint boundaries can produce designs that look good only on the model. The abstract gives no training-set size, no cross-validation breakdown, and no post-optimization CAE confirmation on the 35 outputs. Without those checks the headline claim stays provisional.

This is aimed at CAE teams in automotive safety who already run large simulation campaigns and are looking for workflow speed-ups. A reader who wants concrete numbers on surrogate accuracy or wants to see the validation campaign will find the current version thin. It is coherent on its own terms and shows clear thinking about how the pieces fit, so it is worth sending to referees once the missing verification data are added.

Referee Report

2 major / 1 minor

Summary. The paper presents a foundation model-orchestrated workflow for surrogate-assisted multi-objective optimization in pedestrian protection crash safety design. It combines a surrogate model (average R²=0.87 with conformal prediction intervals) trained on CAE simulations, NSGA-II search under constraints, a morphing-based geometry generator, and LLM/VLM components for orchestration and semantic comparison. The central result is the generation of 35 distinct safety-compliant front-bumper designs from a single exploration, claimed to reduce evaluation time from hours per CAE run to seconds.

Significance. If the surrogate-guided designs are shown to remain feasible under full nonlinear CAE re-evaluation, the work would illustrate a viable integration layer for foundation models between data-driven surrogates and physics-based simulation in safety-critical domains, addressing the challenges of nonlinear contact dynamics that standard surrogates struggle with.

major comments (2)

[Abstract] Abstract and workflow description: the headline claim that the workflow produces 35 safety-compliant alternatives rests on the surrogate (R²=0.87 plus conformal intervals) guiding NSGA-II to feasible points, yet the manuscript reports no post-optimization CAE verification campaign on these 35 designs using the original nonlinear contact solver. Crash problems involve sharp state transitions; without this check the practical correctness of the discovered set remains unverified.
[Methods (surrogate)] Surrogate model section: no training-set size, cross-validation procedure, or hold-out performance details are supplied beyond the average R²=0.87, leaving unclear whether the model generalizes sufficiently near constraint boundaries for the evolutionary search to be reliable.

minor comments (1)

[Abstract] The 'first' claim in the abstract would benefit from a concise literature positioning paragraph to substantiate novelty relative to prior surrogate-assisted crash design studies.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments on our manuscript. We address each major comment below and will revise the paper accordingly to improve clarity and rigor.

read point-by-point responses

Referee: [Abstract] Abstract and workflow description: the headline claim that the workflow produces 35 safety-compliant alternatives rests on the surrogate (R²=0.87 plus conformal intervals) guiding NSGA-II to feasible points, yet the manuscript reports no post-optimization CAE verification campaign on these 35 designs using the original nonlinear contact solver. Crash problems involve sharp state transitions; without this check the practical correctness of the discovered set remains unverified.

Authors: We agree that the absence of post-optimization verification with the full nonlinear CAE solver is a limitation, particularly given the nonlinear contact dynamics and potential sharp state transitions in crash problems. The current claims rely on the surrogate predictions (including conformal intervals) to identify feasible designs. In the revised manuscript, we will add a dedicated verification section reporting results from re-evaluating the 35 designs (or a representative subset) using the original CAE solver to confirm compliance and quantify any surrogate prediction errors. revision: yes
Referee: [Methods (surrogate)] Surrogate model section: no training-set size, cross-validation procedure, or hold-out performance details are supplied beyond the average R²=0.87, leaving unclear whether the model generalizes sufficiently near constraint boundaries for the evolutionary search to be reliable.

Authors: We acknowledge that these details were omitted from the surrogate model description. The revised manuscript will explicitly report the training-set size (number of CAE simulations), the cross-validation procedure (e.g., k-fold details), and additional hold-out test performance metrics. This will include analysis of performance near constraint boundaries to better substantiate the reliability of the NSGA-II optimization. revision: yes

Circularity Check

0 steps flagged

No circularity: workflow integrates external standard components without self-referential reduction

full rationale

The manuscript presents an engineering workflow that chains a trained surrogate (R²=0.87), NSGA-II, a morphing generator, and LLM orchestration to produce design candidates. No equations define any quantity in terms of itself, no fitted parameter is relabeled as an independent prediction, and no load-bearing premise rests on a self-citation chain. The headline result (35 compliant designs) is an empirical output of the described pipeline rather than a quantity forced by construction from its inputs. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the surrogate training process, conformal calibration, and morphing mapping are treated as black boxes whose internal assumptions cannot be audited.

pith-pipeline@v0.9.1-grok · 5793 in / 1256 out tokens · 30444 ms · 2026-06-27T01:22:58.828606+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 6 canonical work pages · 1 internal anchor

[2]

Christopher Baker, Karen Rafferty, and Mark Price

URL https://arxiv.org/abs/2107.07511. Christopher Baker, Karen Rafferty, and Mark Price. Large language models in mechanical engi- neering: A scoping review of applications, challenges, and future directions.Big Data and Cog- nitive Computing, 9(12):305,

Pith/arXiv arXiv
[3]

URLhttps://doi.org/ 10.3390/bdcc9120305

doi: 10.3390/bdcc9120305. URLhttps://doi.org/ 10.3390/bdcc9120305. Wei Chen and Arun Ramamurthy. Deep generative model for efficient 3d airfoil parameterization and generation,

work page doi:10.3390/bdcc9120305
[4]

Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T

URLhttps://arxiv.org/abs/2101.02744. Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic algorithm: NSGA-II.IEEE Transactions on Evolutionary Computation, 6(2):182–197, April

arXiv
[5]

Meyarivan

doi: 10.1109/4235.996017. URLhttps://doi.org/10.1109/4235.996017. Osamu Ito, Jun Shiraishi, Kazuo Imura, Takeshi Yamatoda, and Yasuhiro Kumatani. Prediction of pedestrian protection performance using machine learning. InProceedings of the 26th Interna- tional Technical Conference on the Enhanced Safety of V ehicles (ESV), pp. 1–10,

work page doi:10.1109/4235.996017
[6]

URLhttps://doi.org/10.1007/ s00158-024-03771-5

doi: 10.1007/s00158-024-03771-5. URLhttps://doi.org/10.1007/ s00158-024-03771-5. Published: 26 February

work page doi:10.1007/s00158-024-03771-5
[7]

Jones, Matthias Schonlau, and William J

doi: 10.1023/A:1008306431147. URLhttps://doi.org/10.1023/A:1008306431147. Bharat Kaushik, P. Daphal, P. Khare, and S. Koralla. Pedestrian safety performance prediction using machine learning techniques. SAE Technical Paper 2021-26-0026, SAE International,

work page doi:10.1023/a:1008306431147 2021
[8]

Jonas Kneifl, J ¨org Fehr, Steven L

URL https://doi.org/10.4271/2021-26-0026. Jonas Kneifl, J ¨org Fehr, Steven L. Brunton, and J. Nathan Kutz. Multi-hierarchical surrogate learning for explicit structural dynamical systems using graph convolutional neural networks. Computational Mechanics, 75:1115–1135,

work page doi:10.4271/2021-26-0026 2021
[9]

A graph neural network surrogate model for mesh-based crashworthiness prediction of vehicle panel components

doi: 10.1007/s00466-024-02553-6. URL https://doi.org/10.1007/s00466-024-02553-6. Haoran Li, Yingxue Zhao, Haosu Zhou, Tobias Pfaff, and Nan Li. A new graph-based surrogate model for rapid prediction of crashworthiness performance of vehicle panel components, 2025a. URLhttps://arxiv.org/abs/2503.17386. Jiahui Li, Haodong Wei, Chenhui Yang, Yihang Liu, Ruiy...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/s00466-024-02553-6
[10]

Alexandre Picard, Simon Lafleur, Arnaud Gros, Ayelet Bessissow, and Florence Petit

URLhttps://arxiv.org/abs/2510.15201. Alexandre Picard, Simon Lafleur, Arnaud Gros, Ayelet Bessissow, and Florence Petit. From con- cept to manufacturing: Evaluating vision-language models for engineering design,

arXiv
[11]

Jamshid A

URL https://arxiv.org/abs/2311.12668. Jamshid A. Samareh. Survey of shape parameterization techniques for high-fidelity multidisci- plinary shape optimization. Technical report, NASA Langley Research Center,

arXiv
[12]

11 Published as a conference paper at ICLR 2026 Thomas W

URL https://ntrs.nasa.gov/citations/19990050940. 11 Published as a conference paper at ICLR 2026 Thomas W. Sederberg and Scott R. Parry. Free-form deformation of solid geometric models. In Proceedings of SIGGRAPH, pp. 151–160,

arXiv 2026
[13]

URLhttps: //arxiv.org/abs/2509.12224. AUTHORCONTRIBUTIONS Osamu Ito (corresponding author) conceived the project, designed the workflow, implemented the surrogate modeling and optimization pipeline, and wrote the manuscript. Akihiko Katagiri, Yoshikazu Nakagawa, Shin Saeki, Jun Shiraishi, and Masato Sasaki contributed to discussions, pro- vided domain exp...

arXiv 2026

[1] [2]

Christopher Baker, Karen Rafferty, and Mark Price

URL https://arxiv.org/abs/2107.07511. Christopher Baker, Karen Rafferty, and Mark Price. Large language models in mechanical engi- neering: A scoping review of applications, challenges, and future directions.Big Data and Cog- nitive Computing, 9(12):305,

Pith/arXiv arXiv

[2] [3]

URLhttps://doi.org/ 10.3390/bdcc9120305

doi: 10.3390/bdcc9120305. URLhttps://doi.org/ 10.3390/bdcc9120305. Wei Chen and Arun Ramamurthy. Deep generative model for efficient 3d airfoil parameterization and generation,

work page doi:10.3390/bdcc9120305

[3] [4]

Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T

URLhttps://arxiv.org/abs/2101.02744. Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and T. Meyarivan. A fast and elitist multiobjective genetic algorithm: NSGA-II.IEEE Transactions on Evolutionary Computation, 6(2):182–197, April

arXiv

[4] [5]

Meyarivan

doi: 10.1109/4235.996017. URLhttps://doi.org/10.1109/4235.996017. Osamu Ito, Jun Shiraishi, Kazuo Imura, Takeshi Yamatoda, and Yasuhiro Kumatani. Prediction of pedestrian protection performance using machine learning. InProceedings of the 26th Interna- tional Technical Conference on the Enhanced Safety of V ehicles (ESV), pp. 1–10,

work page doi:10.1109/4235.996017

[5] [6]

URLhttps://doi.org/10.1007/ s00158-024-03771-5

doi: 10.1007/s00158-024-03771-5. URLhttps://doi.org/10.1007/ s00158-024-03771-5. Published: 26 February

work page doi:10.1007/s00158-024-03771-5

[6] [7]

Jones, Matthias Schonlau, and William J

doi: 10.1023/A:1008306431147. URLhttps://doi.org/10.1023/A:1008306431147. Bharat Kaushik, P. Daphal, P. Khare, and S. Koralla. Pedestrian safety performance prediction using machine learning techniques. SAE Technical Paper 2021-26-0026, SAE International,

work page doi:10.1023/a:1008306431147 2021

[7] [8]

Jonas Kneifl, J ¨org Fehr, Steven L

URL https://doi.org/10.4271/2021-26-0026. Jonas Kneifl, J ¨org Fehr, Steven L. Brunton, and J. Nathan Kutz. Multi-hierarchical surrogate learning for explicit structural dynamical systems using graph convolutional neural networks. Computational Mechanics, 75:1115–1135,

work page doi:10.4271/2021-26-0026 2021

[8] [9]

A graph neural network surrogate model for mesh-based crashworthiness prediction of vehicle panel components

doi: 10.1007/s00466-024-02553-6. URL https://doi.org/10.1007/s00466-024-02553-6. Haoran Li, Yingxue Zhao, Haosu Zhou, Tobias Pfaff, and Nan Li. A new graph-based surrogate model for rapid prediction of crashworthiness performance of vehicle panel components, 2025a. URLhttps://arxiv.org/abs/2503.17386. Jiahui Li, Haodong Wei, Chenhui Yang, Yihang Liu, Ruiy...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/s00466-024-02553-6

[9] [10]

Alexandre Picard, Simon Lafleur, Arnaud Gros, Ayelet Bessissow, and Florence Petit

URLhttps://arxiv.org/abs/2510.15201. Alexandre Picard, Simon Lafleur, Arnaud Gros, Ayelet Bessissow, and Florence Petit. From con- cept to manufacturing: Evaluating vision-language models for engineering design,

arXiv

[10] [11]

Jamshid A

URL https://arxiv.org/abs/2311.12668. Jamshid A. Samareh. Survey of shape parameterization techniques for high-fidelity multidisci- plinary shape optimization. Technical report, NASA Langley Research Center,

arXiv

[11] [12]

11 Published as a conference paper at ICLR 2026 Thomas W

URL https://ntrs.nasa.gov/citations/19990050940. 11 Published as a conference paper at ICLR 2026 Thomas W. Sederberg and Scott R. Parry. Free-form deformation of solid geometric models. In Proceedings of SIGGRAPH, pp. 151–160,

arXiv 2026

[12] [13]

URLhttps: //arxiv.org/abs/2509.12224. AUTHORCONTRIBUTIONS Osamu Ito (corresponding author) conceived the project, designed the workflow, implemented the surrogate modeling and optimization pipeline, and wrote the manuscript. Akihiko Katagiri, Yoshikazu Nakagawa, Shin Saeki, Jun Shiraishi, and Masato Sasaki contributed to discussions, pro- vided domain exp...

arXiv 2026