Original Article

Deep learning approach for predicting hoop-direction mechanical behavior of tubular materials in hydroforming

¹ Laboratoire de Génie Mécanique, Ecole Nationale d'Ingénieurs de Monastir, Av. Ibn Eljazzar Monastir 5019, Université de Monastir, Tunisia
² CEMMPRE, Department of Mechanical Engineering, University of Coimbra, Rua Luís Reis Santos, Coimbra, 3030-788, Portugal
³ Institut Supérieur des Sciences Appliquées et de Technologie de Sousse, Cité Ibn Khaldoun, 4003 Sousse, Université de Sousse, Tunisia

^* e-mail: Cette adresse e-mail est protégée contre les robots spammeurs. Vous devez activer le JavaScript pour la visualiser.

Received: 22 August 2025
Accepted: 12 January 2026

Abstract

Characterizing the mechanical behavior of anisotropic tubular materials remains challenging due to their curved geometry and manufacturing-induced heterogeneity. Unlike sheet metals, tubes cannot be easily sampled along multiple directions, and conventional approaches such as flattening introduce pre-strains that compromise accuracy. This difficulty is particularly critical in the hoop direction, which governs the performance of tubular components in hydroforming processes and biomedical applications. The ring hoop tensile test (RHTT) is an effective method for probing hoop-direction behavior; however, the measured force–displacement response is strongly affected by friction between the specimen and the loading device, leading to overestimation of the true material response. This paper presents a novel approach, based on deep learning techniques, to predict the actual mechanical behavior of an AA6063 aluminum alloy tube along the hoop direction using the ring hoop tensile test. We explore the potential of convolutional neural networks, a class of deep learning models widely recognized for their accuracy in inverse identification tasks. A one-dimensional convolutional neural network (1D-CNN) is trained using numerically generated RHTT datasets incorporating a Swift hardening law and varying friction conditions. Data augmentation and hyperparameter tuning are employed to enhance robustness under limited data availability. The performance of the proposed model is quantitatively compared with artificial neural networks (ANNs) and physics-informed neural networks (PINNs). The optimized model achieves a coefficient of determination of approximately R² = 0.945 for the identified hoop stress–strain curves and R² = 0.97 for the friction coefficient outperforming the ANN and PINN models. The identified hoop stress–strain curve for the aluminum alloy AA6063-O tube remains within ± 3% of the Hill48-based reference over the investigated deformation range, thereby validating the numerical model and the inverse identification strategy. These results demonstrate the ability of the proposed approach to provide accurate and efficient material and friction identification from a single mechanical test, offering clear advantages for the design and analysis of tubular components in hydroforming and biomedical applications.