Development of a neural network for fractographic analysis of hydrogenated zirconium alloys

M.I. Petrov ORCID logo , M.G. Isaenkova, N.S. Saburov, V.A. Markelov, A.A. Plyasov show affiliations and emails
Received 28 December 2024; Accepted 17 February 2025;
Citation: M.I. Petrov, M.G. Isaenkova, N.S. Saburov, V.A. Markelov, A.A. Plyasov. Development of a neural network for fractographic analysis of hydrogenated zirconium alloys. Lett. Mater., 2025, 15(1) 49-54
BibTex   https://doi.org/10.48612/letters/2025-1-49-54

Abstract

SEM fractographic image of a hydrogenated zirconium alloy, processed using VGG-16 and U-Net neural networks to determine the brittle fraction and segment brittle/ductile regions, with corresponding outputs showing the dependence of brittle fraction on hydride orientation and segmented fracture surfaces.A methodology for the quantitative assessment of fracture characteristics in hydrogenated zirconium alloys using electron microscopy images has been developed. The study utilized a model Zr-Nb-Sn-Fe zirconium alloy with hydrides of varying orientations to analyze fracture surfaces from mechanical tests on ring-shaped specimens. Two convolutional neural networks were employed: a modified VGG-16 for determining the fraction of brittle components and a U-Net for segmenting images into brittle and ductile regions. These networks achieved high accuracy, with the VGG-16 and U-Net performing at 98.2 % and 94.6 %, respectively, relative to manual annotation. The results demonstrated that the brittle fraction extracted from fractographic images provides a more precise and reliable method for determining fracture characteristics compared to conventional macroscopic parameters, such as the relative change in cross-sectional area. Furthermore, it was shown that the brittle fraction can classify the fracture mechanism into one of two categories — brittle fracture (0 – 0.3) or ductile fracture (0.7 –1) — with high reliability. This study highlights the potential of microparameter-based analysis in evaluating structural fracture characteristics and standardizing fractographic methods. The automated fracture image processing methodology developed is not only applicable to hydrogenated zirconium alloy tubes but also extends to other materials and components. These findings emphasize the advantages of machine learning in automating and enhancing the accuracy of quantitative fractography, with significant potential for industrial applications in material property analysis.

References (23)

1. M. P. Puls, The Effect of Hydrogen and Hydrides on the Integrity of Zirconium Alloy Components, Springer-Verlag London, London, 2012, 451 p.
14. Z. Sun, J. Wu, Q. Li, MCNN-TE: a novel method for fractographic images classification based on multi-channel CNN and texture enhancement, UPB Sci. Bull., Ser. C 81 (2019) 121-132.
16. N. S. Saburov, M. G. Isaenkova, M. I. Petrov, P. S. Dzhumaev, A. A. Plyasov, V. A. Markelov, S. A. Bekrenev, Fracture parameters of the Zr-Nb-Sn-Fe tubes samples with different hydrides orientations during the mechanical testing and plasticity of the material, Phys. At. Nucl. (2025). (in press).
22. K. Simonyan, A. Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition, ArXiv Prepr. ArXiv1409.1556 (2014). http://arxiv.org/abs/1409.1556.

Funding

1. Ministry of Science and Higher Education of the Russian Federation - 075-15-2021-1352