Proceedings/Recueil Des Communications Année : 2025

Machine learning methods for analysing damage in energetical granular materials

Steve Belon
Élodie Kaeshammer

Résumé

An energetic material suffering an unintentional impact can undergo a possible shock-to-detonation transition. Indeed, the impact generates a shock wave propagating through the material, interacting with microstructural het- erogeneities and increasing pressure and temperature locally. This phenomenon is known as hot spots. Thus, understanding the consequences of a low-velocity impact on energetic materials is essential for safety. While certain microstructural features like porosity are known to play an important role (Bowden and Yoffe, 1952), the effect of damage parameters such as microcracking and decohesion has not been elucidated to the same extent (Gillibert and Jeulin, 2011). To address this problem, one must first quantify the amount of defects undergone by the material in both its initial state and after an impact. The aim of this work is to present an approach with a neural network to identify and analyze statistically the damage in a shocked microstructure. To this end, butalite samples (polybutadiene and ammonium perchlorate) are subjected to an impact test at a controlled stress level to cause damage without triggering a decomposition reaction. Samples are imaged before and after the impact test using micro-computed tomography. Then, fragments are match spatially to the sound grains from which they originated using image registration. Then, sound microstructure grains are segmented individually in three dimensions using watershed, structural and intensity analysis (Chabardès, 2018). On the other hand, damage such as cracking, fragments and decohesion are segmented from damaged microstructures using thresholding and morphological tools (top-hat, watershed). Then, segmented images of both sound and damaged microstuctures are used to train a neural network whose ultimate goal will be to analyze statistically the damage of grains solely from damaged microstructure images. Finally, this work will attempt to estimate the degree of post-impact damage of a material based on the morphological characteristics of sound material.

Fichier principal
Vignette du fichier
soumis.pdf (16.87 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-05423264 , version 1 (18-12-2025)

Licence

Identifiants

  • HAL Id : hal-05423264 , version 1

Citer

Camille Robin, François Willot, Steve Belon, Lionel Borne, Petr Dokládal, et al.. Machine learning methods for analysing damage in energetical granular materials. ICT 2025 - 54th International Annual Conference : Artificial Intelligence, Machine Learning and Data Science in Energetic Materials Research, Jun 2025, Karlsrhue, Germany. 2025. ⟨hal-05423264⟩
94 Consultations
108 Téléchargements

Partager

  • More