arxiv: 2605.00070 · v1 · submitted 2026-04-30 · 💻 cs.LG

Recognition: unknown

CRADIPOR: Crash Dispersion Predictor

Edgar Chaillou , Sebastian Rodriguez , Yves Tourbier , Francisco Chinesta

Authors on Pith no claims yet

Pith reviewed 2026-05-09 20:33 UTC · model grok-4.3

classification 💻 cs.LG

keywords numerical dispersionautomotive crash simulationRank Reduction Autoencoderfinite element modelssupervised classificationsignal representationslope featureswavelet analysis

0 comments

The pith

An RRAE framework identifies numerical-dispersion-sensitive regions in single-run automotive crash simulations and outperforms a Random Forest baseline.

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

The paper presents CRADIPOR as a post-processing tool that predicts where finite-element crash models are most affected by numerical variability. It extracts latent features from single-run signals using a Rank Reduction Autoencoder, then applies supervised classification to label sensitive regions. On the studied dataset this approach beats a Random Forest baseline, with slope-based signal features delivering the strongest performance and wavelet representations also effective. The goal is to let engineers evaluate dispersion effects during ordinary post-processing instead of running costly repeat simulations. If the method works, routine crash analyses could incorporate dispersion awareness without extra compute.

Core claim

The RRAE-based framework processes post-processing signals from single crash simulations to classify regions sensitive to numerical dispersion. It achieves higher effectiveness than a Random Forest baseline on the dataset examined, and among tested representations the slope-variation inputs yield the best classification results while wavelet-based inputs also perform well.

What carries the argument

The Rank Reduction Autoencoder (RRAE) that learns structured latent representations from simulation signals for subsequent supervised classification of dispersion-sensitive regions.

Load-bearing premise

Dispersion patterns learned from the training simulations will generalize to new crash models when only single-run signals are available.

What would settle it

Measure classification accuracy on a fresh collection of crash models that were never seen during training and check whether performance drops below the reported levels.

Figures

Figures reproduced from arXiv: 2605.00070 by Edgar Chaillou, Francisco Chinesta, Sebastian Rodriguez, Yves Tourbier.

**Figure 2.** Figure 2: Illustration of the effect of numerical dispersion on design ranking based on a single [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Finite element model of the vehicle at the initial time step (pre-impact configuration) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Vehicle-fixed coordinate sys [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Evolution of the number of intersections over time for a dispersed node and a non [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Schematic representation of the RRAE–MLP architecture. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

**Figure 7.** Figure 7: Pseudocode of the RRAE–MLP training procedure. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

**Figure 8.** Figure 8: Superposition of the steering rack housing obtained from two identical simulation runs, [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗

**Figure 9.** Figure 9: Confusion matrix obtained with the supervised learning approach with raw data [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗

**Figure 10.** Figure 10: Classification performance obtained from features extracted using the Fourier transform. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗

**Figure 11.** Figure 11: Classification performance obtained from features extracted using the wavelet transform. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗

**Figure 12.** Figure 12: Classification performance obtained from temporal slope variations. [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗

**Figure 13.** Figure 13: Confusion matrix obtained using position trajectories. [PITH_FULL_IMAGE:figures/full_fig_p013_13.png] view at source ↗

**Figure 14.** Figure 14: Confusion matrix obtained using wavelet-based input (DWT, db4). [PITH_FULL_IMAGE:figures/full_fig_p014_14.png] view at source ↗

**Figure 15.** Figure 15: Confusion matrix obtained using slope variations. [PITH_FULL_IMAGE:figures/full_fig_p014_15.png] view at source ↗

read the original abstract

We present CRADIPOR, a numerical dispersion prediction tool for automotive crash simulations. Finite Element (FE) crash models are widely used throughout vehicle development, but their predictions are not strictly repeatable because of parallel computation and model complexity. As a result, performance criteria evaluated during post-processing may exhibit significant numerical dispersion, which complicates engineering decision-making. Although dispersion can be estimated by repeating the same simulation, this approach is generally impractical because of its high computational cost. This work therefore investigates a prediction tool that can be applied during routine crash-simulation post-processing without repeating the computation. The proposed approach relies on a Rank Reduction Autoencoder (RRAE) combined with supervised classification in order to identify regions sensitive to numerical dispersion. The comparative analysis suggests that the RRAE-based framework is more effective than the Random Forest baseline on the studied dataset. Among the tested signal representations, wavelet-based and slope-based inputs appear to be the most promising, with slope variations providing the best classification performance. These results support the use of structured latent representations for improving numerical-dispersion detection in automotive crash post-processing.

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

CRADIPOR applies RRAE to flag dispersion-sensitive zones from single-run crash signals and beats random forest on slope inputs in their tests, but the lack of numbers and cross-model checks leaves the practical gain unclear.

read the letter

The main point is that this paper builds a classifier on rank-reduction autoencoders to identify regions in automotive crash simulations that are prone to numerical dispersion from parallel solves and model complexity. It processes post-processing signals from one run, encodes them with RRAE, and classifies sensitivity, reporting that slope-based inputs work better than other representations and that the whole setup beats a random forest baseline on the data they used. The motivation is clear: repeating simulations to measure variability is too expensive for routine vehicle programs, so a cheap predictor at post-processing time would help. The signal-representation experiments and the choice of RRAE for structured latent features are the concrete steps forward here. The work stays grounded in the engineering constraint that labels come from repeated runs on the training models, so any compute saving happens only at inference. That framing is honest. The soft spots are straightforward. The abstract gives no accuracy figures, no dataset size, no cross-validation scheme, and no statistical tests, which makes it impossible to judge whether the reported edge over random forest is reliable or just dataset-specific. More critically, there is no evidence that patterns learned on one set of crash models transfer to new geometries, materials, or solvers; dispersion sources are model-dependent, yet the paper does not test held-out model types. This keeps the claim tied to the studied cases. The methods themselves are standard supervised learning, so nothing looks circular or invented. This is aimed at crash analysts who already run large FE suites and want a filter for sensitive zones without extra compute. A reader in that niche could extract the idea of slope or wavelet inputs for similar tasks, but anyone wanting to deploy it would need the missing numbers and validation first. I would bring it to a reading group focused on ML for simulation workflows, but I would not cite it in its current form. It deserves peer review because the application is relevant and the pipeline is testable; referees could ask for the quantitative results and cross-model experiments that are currently absent.

Referee Report

3 major / 3 minor

Summary. The paper presents CRADIPOR, a tool that uses a Rank Reduction Autoencoder (RRAE) to extract latent representations from single-run post-processing signals in finite-element automotive crash simulations, followed by supervised classification to identify regions sensitive to numerical dispersion. It reports that this RRAE-based approach outperforms a Random Forest baseline on the studied dataset, with slope-based and wavelet-based input representations (particularly slope variations) yielding the best classification performance, enabling dispersion prediction without repeated simulations.

Significance. If the comparative results and generalization hold with proper validation, the work addresses a practical issue in vehicle development where numerical dispersion from parallel computing and model complexity affects engineering decisions. By avoiding repeated high-cost simulations at inference time, it could reduce computational overhead in routine post-processing workflows, provided the learned patterns transfer across crash models.

major comments (3)

[Abstract and Results] Abstract and Results section: The central claim that the RRAE framework outperforms Random Forest and that slope inputs are best lacks any quantitative metrics (accuracy, F1, AUC, etc.), dataset size, number of simulations or regions, cross-validation procedure, or statistical significance tests. Without these, the comparative analysis cannot be evaluated and the performance advantage remains unverified.
[Methods and Experiments] Methods and Experiments: Supervised labels for dispersion-sensitive regions are generated via repeated simulations on the training crash models, yet no cross-model validation, ablation on held-out geometries/materials/solvers, or tests on new crash types are reported. Since the introduction notes that dispersion arises from model-specific factors (parallelism and complexity), this leaves the generalization claim unsupported and ties utility to the single studied dataset.
[Signal representations and classification] § on signal representations and classification: The assertion that single-run post-processing signals suffice for reliable identification is not tested against the possibility that dispersion patterns require multiple realizations to capture; the paper provides no ablation comparing single-run vs. multi-run label quality or performance drop on unseen models.

minor comments (3)

[Methods] Clarify the exact definition and training procedure for the RRAE latent dimension and any hyperparameters in the methods section, as these are listed as free parameters.
[Figures and Tables] Figure captions and tables should explicitly state the number of samples, classes, and any preprocessing steps applied to the signals.
[Discussion] Add a limitations or future work subsection discussing the scope of generalization and computational overhead of the RRAE training phase.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment point by point below, providing honest responses based on the current work. We will revise the manuscript to incorporate additional details, clarifications, and a limitations discussion where appropriate.

read point-by-point responses

Referee: [Abstract and Results] Abstract and Results section: The central claim that the RRAE framework outperforms Random Forest and that slope inputs are best lacks any quantitative metrics (accuracy, F1, AUC, etc.), dataset size, number of simulations or regions, cross-validation procedure, or statistical significance tests. Without these, the comparative analysis cannot be evaluated and the performance advantage remains unverified.

Authors: We agree that the abstract and results section would benefit from explicit quantitative support for the claims. The manuscript reports comparative performance on the studied dataset but presents it without listing specific metrics in the abstract. In the revised version, we will expand the results section and abstract to include accuracy, F1, AUC (or other relevant metrics), dataset size details (number of simulations and regions), the cross-validation procedure, and any statistical tests. This will enable proper evaluation of the RRAE advantage over Random Forest and the preference for slope-based inputs. revision: yes
Referee: [Methods and Experiments] Methods and Experiments: Supervised labels for dispersion-sensitive regions are generated via repeated simulations on the training crash models, yet no cross-model validation, ablation on held-out geometries/materials/solvers, or tests on new crash types are reported. Since the introduction notes that dispersion arises from model-specific factors (parallelism and complexity), this leaves the generalization claim unsupported and ties utility to the single studied dataset.

Authors: This is a valid concern. Our experiments and label generation are confined to the single studied dataset of crash models, with no cross-model validation or ablations on held-out geometries, materials, solvers, or new crash types performed. We will revise the manuscript to add an explicit limitations section that acknowledges the model-specific nature of numerical dispersion (as already noted in the introduction) and clarifies that generalization claims are not supported beyond the current dataset. We will also discuss the implications for practical utility. revision: partial
Referee: [Signal representations and classification] § on signal representations and classification: The assertion that single-run post-processing signals suffice for reliable identification is not tested against the possibility that dispersion patterns require multiple realizations to capture; the paper provides no ablation comparing single-run vs. multi-run label quality or performance drop on unseen models.

Authors: We acknowledge the absence of such an ablation study. The proposed method uses single-run signals at inference time while generating supervised labels from repeated simulations during training. In the revision, we will clarify this design choice in the methods and signal representations section, add discussion of the untested aspects (including potential differences in label quality and performance on unseen models), and note this as a limitation to better support the claims about single-run post-processing sufficiency. revision: yes

Circularity Check

0 steps flagged

No circularity: standard supervised ML pipeline on simulation data

full rationale

The paper proposes an RRAE-based classifier trained on post-processing signals from repeated crash simulations to label dispersion-sensitive regions, then applies it at inference without repeats. No equations, derivations, or self-citations reduce any claimed prediction or result to its inputs by construction. The comparative performance claims are empirical evaluations on the given dataset against a Random Forest baseline, with no self-definitional loops, fitted inputs renamed as predictions, or uniqueness theorems imported from overlapping-author prior work. The method is a conventional supervised learning setup whose outputs are not forced by the training labels or architecture definitions.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract-only review; full methods, training details, and data characteristics unavailable.

free parameters (1)

RRAE latent dimension and training hyperparameters
Required for the autoencoder but not specified; typical in such models and would need fitting to the crash data.

axioms (1)

domain assumption Numerical dispersion in finite-element crash models can be adequately captured by wavelet or slope representations of single-run post-processing signals
Central modeling choice that enables the classification step without repeated simulations.

pith-pipeline@v0.9.0 · 5489 in / 1217 out tokens · 43948 ms · 2026-05-09T20:33:15.023708+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

33 extracted references · 17 canonical work pages

[1]

Accessed: 2026

Euroncapassessmentprotocol–frontalimpact.https://www.euroncap.com, 2020. Accessed: 2026

2020
[2]

Accessed: 2026

Iihs crashworthiness evaluation – frontal crash tests.https://www.iihs.org, 2020. Accessed: 2026

2020
[3]

Dynamique temporelle multivariéeinvarianted’échelle: del’analysespectrale(fourier)àl’analysefractale(ondelette)

Patrice Abry, Herwig Wendt, Stéphane Jaffard, and Gustavo Didier. Dynamique temporelle multivariéeinvarianted’échelle: del’analysespectrale(fourier)àl’analysefractale(ondelette). Comptes Rendus Physique, 2019

2019
[4]

Klaus-Jurgen Bathe, 2006

Klaus-Jürgen Bathe.Finite Element Procedures. Klaus-Jurgen Bathe, 2006. URL https://soaneemrana.org/onewebmedia/Finite%20Element%20Procedures%20in% 20Engineering%20Analysis%20Bathe%20K.J.pdf

2006
[5]

Wiley, 2013

Ted Belytschko, Wing Kam Liu, Brian Moran, and Khalil Elkhodary.Nonlinear Finite Ele- ments for Continua and Structures. Wiley, 2013

2013
[6]

Rank reduction autoencoders for mechanical design: Advancing novel and efficient data-driven topology optimization.arXiv preprint arXiv:2601.23269, 2026

Ismael Ben-Yelun, Mohammed El Fallaki Idrissi, Jad Mounayer, Sebastian Rodriguez, and Francisco Chinesta. Rank reduction autoencoders for mechanical design: Advancing novel and efficient data-driven topology optimization.arXiv preprint arXiv:2601.23269, 2026. doi: 10.48550/arXiv.2601.23269

work page doi:10.48550/arxiv.2601.23269 2026
[7]

Representation learning: a review and new perspectives

Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(8): 1798–1828, 2013. doi: 10.1109/TPAMI.2013.50

work page doi:10.1109/tpami.2013.50 2013
[8]

A random forest guided tour.arXiv:1511.05741, 2015

Gerard Biau and Erwan Scornet. A random forest guided tour.arXiv:1511.05741, 2015

work page arXiv 2015
[9]

Jour- nal of Automated Reasoning26(2), 107–137 (2001) https://doi.org/10.1023/A: 1026518331905

Leo Breiman. Random forests.Machine Learning, 45(1):5–32, 2001. doi: 10.1023/A: 1010933404324

work page doi:10.1023/a: 2001
[10]

Breitung

K. Breitung. Asymptotic approximations for multinormal integrals.Journal of Engineering Mechanics, 110(3):357–366, 1984. doi: 10.1061/(ASCE)0733-9399(1984)110:3(357)

work page doi:10.1061/(asce)0733-9399(1984)110:3(357 1984
[11]

Numericalreproducibil- ity for the parallel reduction on multi- and many-core architectures.Parallel Computing, 49: 83–97, 2015

SylvainCollange, DavidDefour, StéfGraillat, andRomanIakymchuk. Numericalreproducibil- ity for the parallel reduction on multi- and many-core architectures.Parallel Computing, 49: 83–97, 2015. doi: 10.1016/j.parco.2015.08.001

work page doi:10.1016/j.parco.2015.08.001 2015
[12]

A new framework for generative design, real-time predic- tion, and inverse design optimization: Application to microstructure.Research Square, 2025

Mohammed El Fallaki Idrissi, Ismael Ben-Yelun, Jad Mounayer, Sebastian Rodriguez, Chady Ghnatios, and Francisco Chinesta. A new framework for generative design, real-time predic- tion, and inverse design optimization: Application to microstructure.Research Square, 2025. doi: 10.21203/rs.3.rs-7478922/v1

work page doi:10.21203/rs.3.rs-7478922/v1 2025
[13]

Generative parametric design (GPD): A framework for real-time geometry generation and on-the-fly multiparametric approximation.arXiv preprint arXiv:2512.11748,

Mohammed El Fallaki Idrissi, Jad Mounayer, Sebastian Rodriguez, Fodil Meraghni, and Fran- cisco Chinesta. Generative parametric design (GPD): A framework for real-time geometry generation and on-the-fly multiparametric approximation.arXiv preprint arXiv:2512.11748,

work page arXiv
[14]

doi: 10.48550/arXiv.2512.11748

work page doi:10.48550/arxiv.2512.11748
[15]

What every computer scientist should know about floating-point arithmetic,

David Goldberg. What every computer scientist should know about floating-point arithmetic. ACM Computing Surveys, 23(1):5–48, 1991. doi: 10.1145/103162.103163

work page doi:10.1145/103162.103163 1991
[16]

Golub and Charles F

Gene H. Golub and Charles F. Van Loan.Matrix Computations. Johns Hopkins University Press, 2013

2013
[17]

MIT Press, 2016

Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning. MIT Press, 2016

2016
[18]

Springer, 2009

Trevor Hastie, Robert Tibshirani, and Jerome Friedman.The Elements of Statistical Learning. Springer, 2009

2009
[19]

Higham.Accuracy and Stability of Numerical Algorithms

Nicholas J. Higham.Accuracy and Stability of Numerical Algorithms. SIAM, 2002

2002
[20]

Reducing the dimensionality of data with neural networks

Geoffrey E. Hinton and Ruslan R. Salakhutdinov. Reducing the dimensionality of data with neural networks.Science, 313(5786):504–507, 2006. doi: 10.1126/science.1127647

work page doi:10.1126/science.1127647 2006
[21]

Jolliffe and Jorge Cadima.Principal Component Analysis

Ian T. Jolliffe and Jorge Cadima.Principal Component Analysis. Springer, 2016. 17

2016
[22]

Cambridge University Press, 2010

Norman Jones.Structural Impact. Cambridge University Press, 2010

2010
[23]

Academic Press, 1999

Stéphane Mallat.A Wavelet Tour of Signal Processing. Academic Press, 1999

1999
[24]

Rank reduction autoencoders.arXiv, abs/2405.13980, 2024

Jad Mounayer, Sebastian Rodriguez, Chady Ghnatios, Charbel Farhat, and Francisco Chinesta. Rank reduction autoencoders.arXiv, abs/2405.13980, 2024. URLhttps: //arxiv.org/abs/2405.13980

work page arXiv 2024
[25]

RRAEDy: Adaptive latent linearization of nonlinear dynamical systems.Scientific Reports, 2026

Jad Mounayer, Sebastian Rodriguez, Jerome Tomezyk, Chady Ghnatios, and Francisco Chinesta. RRAEDy: Adaptive latent linearization of nonlinear dynamical systems.Scientific Reports, 2026. doi: 10.1038/s41598-026-47609-0

work page doi:10.1038/s41598-026-47609-0 2026
[26]

Oppenheim and Alan S

Alan V. Oppenheim and Alan S. Willsky.Signals and Systems. Prentice Hall, 1999

1999
[27]

Rackwitz and B

R. Rackwitz and B. Fiessler. Structural reliability under combined random load sequences. Computers & Structures, 9(5):489–494, 1978. doi: 10.1016/0045-7949(78)90046-9

work page doi:10.1016/0045-7949(78)90046-9 1978
[28]

Sebastian Rodriguez, Marc Rébillat, Nazih Mechbal, Amine Ammar, and Francisco Chinesta. Damage detection algorithm based on an innovative nonlinear model-order reduction tech- nique: The rank reduction autoencoder (RRAE) conditioned to learn damage features.e- Journal of Nondestructive Testing, 31(2), 2026. doi: 10.58286/32462

work page doi:10.58286/32462 2026
[29]

Data-driven discovery of roughness descriptors for surface character- ization and intimate contact modeling of unidirectional composite tapes.arXiv preprint arXiv:2603.20418, 2026

Sebastian Rodriguez, Mikhael Tannous, Jad Mounayer, Camilo Cruz, Anais Barasinski, and Francisco Chinesta. Data-driven discovery of roughness descriptors for surface character- ization and intimate contact modeling of unidirectional composite tapes.arXiv preprint arXiv:2603.20418, 2026. doi: 10.48550/arXiv.2603.20418

work page doi:10.48550/arxiv.2603.20418 2026
[30]

Learning representations by back-propagating errors.Nature1986,323, 533–536

David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors.Nature, 323:533–536, 1986. doi: 10.1038/323533a0

work page doi:10.1038/323533a0 1986
[31]

Strogatz.Nonlinear Dynamics and Chaos

Steven H. Strogatz.Nonlinear Dynamics and Chaos. CRC Press, 2018

2018
[32]

Asian Productivity Organization, 1986

Genichi Taguchi.Introduction to Quality Engineering: Designing Quality into Products and Processes. Asian Productivity Organization, 1986

1986
[33]

Timoshenko and James M

Stephen P. Timoshenko and James M. Gere.Theory of Elastic Stability. McGraw-Hill, 2 edition, 1961. 18

1961