Thèse Simulation en Temps Réel des Procédés de Soudage par Friction pour le Développement de leurs Jumeaux Numériques Application au Soudage par Friction Linéaire Lfw H/F Doctorat.Gouv.Fr

Établissement : Institut National Polytechnique de Toulouse École doctorale : MEGEP - Mécanique, Energétique, Génie civil, Procédés Laboratoire de recherche : LGP - Laboratoire Génie de Production Direction de la thèse : Amèvi TONGNE ORCID 0000000253605121 Début de la thèse : 2027-09-01 Date limite de candidature : 2026-11-23T23:59:59 L'allègement des structures constitue un enjeu majeur pour réduire l'empreinte carbone du secteur des transports, notamment dans le domaine aérospatial. Cet objectif favorise le développement de nouveaux alliages, dont l'intégration soulève toutefois d'importants défis en matière d'assemblage. Les procédés de soudage en phase solide se distinguent par leur capacité à préserver les propriétés spécifiques de ces alliages, dont la composition est optimisée pour garantir des performances mécaniques et thermiques élevées.

Des techniques telles que le soudage par friction-malaxage (FSW) et ses variantes, notamment le soudage par friction-malaxage à épaulement fixe (SSFSW), ainsi que le soudage par friction linéaire (LFW), présentent un fort potentiel. Cependant, les écoulements complexes de matière et les mécanismes de cohésion mis en jeu au cours de ces procédés sont particulièrement sensibles aux paramètres opératoires. Une maîtrise insuffisante de ces paramètres peut entraîner l'apparition de défauts. Chaque nouvelle application nécessite alors des analyses approfondies et de nombreux essais, ce qui ralentit l'adoption industrielle de ces technologies.

L'intelligence artificielle (IA) et les jumeaux numériques offrent des perspectives prometteuses pour relever ces défis. En constituant des répliques virtuelles évolutives des systèmes physiques, les jumeaux numériques permettent de simuler, de surveiller et d'optimiser les procédés industriels en temps réel. L'IA facilite l'intégration des connaissances métiers, l'exploitation des données expérimentales et l'ajustement dynamique des paramètres opératoires. Elle contribue ainsi à l'optimisation du procédé et à la détection précoce des défauts.

Ce projet de thèse se concentre sur la simulation en temps réel du soudage par friction linéaire. Il vise à développer des modèles de substitution fondés sur des réseaux de neurones informés par la physique et enrichis par des données expérimentales multi-échelles, allant de la microstructure jusqu'à l'échelle de la machine. Ces modèles constitueront le coeur d'un jumeau numérique destiné à mieux comprendre, surveiller et optimiser le procédé. Linear Friction Welding (LFW) is a solid-state joining process with significant potential for assembling high-performance alloys that are difficult to weld using conventional fusion-based techniques. Since the materials do not reach their melting point, LFW limits solidification-related defects and helps preserve the intrinsic properties of advanced alloys. These advantages make the process particularly attractive to the aerospace industry, where structural lightweighting, mechanical performance, and joint reliability are critical requirements.

Nevertheless, the physical phenomena governing LFW remain particularly complex. Heat is generated through friction and severe plastic deformation, while the resulting temperature increase strongly affects the material behaviour, contact conditions, interfacial friction, material flow, and bonding mechanisms. These phenomena are highly coupled and evolve rapidly during the different stages of the process. Moreover, the final weld quality depends not only on macroscopic variables, such as temperature, stress, strain, flash formation, and applied force, but also on microstructural transformations occurring at and around the interface. An inadequate combination of operating parameters may therefore lead to insufficient bonding, voids, cracks, or undesirable microstructural changes.

Experimental investigations provide essential information for understanding these mechanisms. However, the high cost of testing, the short duration of the process, and the difficulty of accessing the welding interface during operation limit the amount of information that can be obtained experimentally. Furthermore, extensive experimental campaigns are generally required for each new material and welding configuration. High-fidelity numerical models can complement these investigations by predicting the thermomechanical history, material flow, contact behaviour, and microstructure evolution. However, their computational cost remains incompatible with real-time applications.

Real-time simulation is an essential requirement for the future monitoring, optimisation, and control of LFW. It would make it possible to estimate physical quantities that cannot be measured directly, detect deviations from nominal process conditions, and predict potential defects before the end of the welding operation. Achieving this objective requires reduced-order or surrogate models that preserve the relevant physical behaviour while considerably reducing computation time.

Artificial neural networks provide rapid predictions but generally require large, representative datasets and may produce physically inconsistent results outside their training domain. Physics-Informed Neural Networks (PINNs) offer a promising alternative by incorporating governing equations, boundary conditions, and physical constraints into the learning process. They can also combine heterogeneous information obtained from numerical simulations, experimental measurements, and expert knowledge. Nevertheless, their application to LFW raises several scientific challenges, including the representation of strongly coupled and highly nonlinear thermomechanical phenomena, moving and evolving interfaces, complex contact conditions, multiscale behaviour, and rapid transient events. Their convergence, accuracy, robustness, and computational efficiency must also be demonstrated under process-relevant conditions.

In this context, the present PhD project will focus on developing real-time surrogate models for LFW based on physics-informed machine-learning approaches and multiscale experimental data. The models will aim to reproduce the main thermomechanical and microstructural responses of the process, from the welding interface to the machine scale, and to establish relationships between operating parameters, physical phenomena, and defect formation. Particular attention will be paid to contact and bonding conditions, material flow, temperature and stress fields, and microstructure evolution.

The project does not aim to develop a complete digital twin or its closed-loop decision-making and control architecture. Instead, it will address one of its most critical scientific components: a fast, physically consistent, and experimentally validated computational model capable of providing real-time predictions. This framework will ultimately constitute a computational building block for future digital-twin applications dedicated to the monitoring, optimisation, and control of Linear Friction Welding.
The main objective of this PhD project is to develop a real-time computational framework for the simulation of Linear Friction Welding (LFW). The proposed approach will rely on surrogate models, particularly Physics-Informed Neural Networks (PINNs), combining physical knowledge with multiscale experimental data ranging from the microstructural scale to the machine scale. The models will aim to predict the process behaviour and the formation of defects and anomalies, such as voids and cracks, according to the operating parameters. Particular attention will be paid to material flow, contact and bonding conditions, and their effects on microstructure evolution. Ultimately, this real-time simulation framework will provide a key computational building block for the future development of a digital twin dedicated to the monitoring and optimisation of the LFW process. The methodology combines experimental investigations, high-fidelity numerical simulations, and physics-informed machine learning to develop a real-time computational model of Linear Friction Welding.

First, an experimental campaign based on Design of Experiments will investigate representative combinations of oscillation parameters, axial force, and welding time. Force, displacement, and temperature data will be recorded, while post-weld observations will assess flash formation, interface evolution, bonding quality, and defects.

An Arbitrary Lagrangian-Eulerian finite element model will then be developed, calibrated, and validated using these experimental data. The model will account for thermomechanical coupling, heat generation, contact conditions, and material flow. Once validated, it will generate a numerical database containing temperature, stress, strain, contact pressure, material-flow fields, and defect-related indicators.

Neural-network-based surrogate models will be trained on this database to provide rapid predictions. The best-performing model will subsequently be enriched with physical equations, boundary conditions, and experimental measurements using a physics-informed learning strategy.

Finally, the model will be validated on unseen experimental configurations. Its accuracy, robustness, generalisation capability, and computation time will be evaluated to demonstrate its suitability for real-time simulation. The resulting model will constitute a computational building block for future monitoring, optimisation, and digital-twin applications.