ED Mathématiques et Informatique
Development of models and numerical methods for the degradation of a pyrolysable material
by Alexis CAS (IMB - Institut de Mathématiques de Bordeaux)
The defense will take place at 14h00 - Amphithéâtre du LaBRI 351, cours de la Libération F-33405 Talence cedex Bâtiment A30: LaBRI
in front of the jury composed of
- Heloïse BEAUGENDRE - Professeure des universités - Bordeaux INP - Directeur de these
- Céline BARANGER - Ingénieure de recherche - CEA-CESTA - CoDirecteur de these
- Jean LACHAUD - Professeur des universités - Université de Bordeaux - Examinateur
- Nicolas DELLINGER - Ingénieur de recherche - ONERA - Examinateur
- Pierre SCHROOYEN - Ingénieur de recherche - VKI - Examinateur
- Alexandre MARTIN - Professor - University of Kentucky - Rapporteur
- Pauline LAFITTE - Professeure des universités - CentraleSupélec - Rapporteur
During atmospheric reentry, a spacecraft is subjected to significant heat fluxes, which degrade the material of its heat shield. Heat propagation is limited by pyrolysis -- a deep chemical degradation process -- and by ablation -- a surface physicochemical degradation process. The aim of this work is to develop a macroscopic model of the material's thermomechanical behavior and to investigate numerical methods for simulating arc-jet tests. First, a macroscopic pyrolysis model is developed to describe the material's anisotropic thermal response. This is coupled with an ablation model based on a B' formulation and an empirical deformation model. Numerical schemes and resolutions are then implemented to solve this system of nonlinear partial differential equations while studying mass and energy conservation at discrete and continuous levels. Finally, an ALE formulation is employed to account for mesh movements related to ablation, deformation and mesh adaptation while preserving the number of mesh nodes. The numerical models and methods were first developed in the SPA2RC code -- which was implemented from scratch during this work -- and then validated against reference cases before certain methods were integrated into the in-house CEA-CESTA aerothermodynamic code. Finally, arc-jet experiments were reproduced in order to simulate the wall displacements caused by ablation and deformation, as well as the temperature profile. The model accurately reproduces wall displacements while improving the thermal response. Including deformations in the model influences both the ablation rate and the material's heating rate, which appears essential for accurately reproducing the material's thermomechanical behavior.
ED Sciences de la Vie et de la Santé
Effects of chronic administration of 5-HT agonists on autonomous functions in spinal cord injured rats
by Aurélie BONILLA (Institut de neurosciences cognitives et intégratives d'Aquitaine)
The defense will take place at 14h00 - Amphithéâtre Bâtiment Biologie Santé (BBS) 2 rue du Dr Hoffmann Martinot 33076 BORDEAUX
in front of the jury composed of
- Sandrine BERTRAND - Directrice de recherche - CNRS - Directeur de these
- Pascal FOSSAT - Professeur des universités - Université de Bordeaux - Examinateur
- Florence PERRIN - Professeure des universités - MMDN, INSERM, Université de Montpellier - Examinateur
- Flavien CHARPENTIER - Directeur de recherche - CNRS - Rapporteur
- Christian GESTREAU - Professeur des universités - Université d'Aix-Marseille - Rapporteur
Spinal cord injuries (SCI) disrupt communication between descending pathways from the brain and their spinal targets, resulting in a complete or partial loss of motor, sensory, and autonomic functions. When these injuries involve the upper thoracic spinal segments, they lead to profound chronic cardiovascular impairments, characterized in particular by persistent hypotension and the development of autonomic dysreflexia. Cardiovascular disease is now one of the leading causes of morbidity and mortality among individuals with spinal cord injury. In parallel, serotonin plays a major role in the functioning of spinal neuronal networks. Serotonergic agonists, particularly quipazine, promote locomotor recovery after spinal cord injury by activating sublesional circuits. However, their cardiovascular effects remain largely unknown, despite the widespread distribution of the serotonergic system, both centrally and peripherally, within structures involved in cardiovascular regulation. The aim of this thesis was to evaluate the acute and chronic cardiovascular effects of quipazine in a rat model of complete T4 thoracic spinal cord transection and to identify the underlying mechanisms of action. Following SCI, certain serotonergic agonists, such as quipazine, promote functional locomotor recovery. Nevertheless, their effects on cardiovascular regulation remain poorly understood, despite the abundant presence of the serotonergic system, both in the central nervous system and in peripheral tissues involved in cardiovascular control. The objective of this thesis was therefore to evaluate the acute and chronic effects of serotonergic agents on cardiovascular function in a rat model of complete T4 thoracic SCI and to identify their mechanisms of action. Using implanted telemetry transmitters in freely behaving rats, we demonstrated that chronic quipazine administration improves cardiovascular recovery by attenuating the hypotension observed during the subacute post-injury phase. It also modulates autonomic balance by reducing the post-lesional cardiac autonomic imbalance. However, quipazine does not prevent the development of autonomic dysreflexia and even enhances the pressor responses associated with these episodes. We are currently investigating the anatomical remodeling of the cardiovascular system and its autonomic innervation following SCI, as well as their evolution during chronic quipazine treatment, using histological and imaging approaches. Through acute administration of serotonergic agonists and antagonists, we sought to characterize the pharmacological targets of quipazine before and after spinal network remodeling induced by SCI. In intact animals, quipazine produced a dose-dependent increase in arterial blood pressure that was primarily mediated by 5-HT2A receptors and was associated with a biphasic chronotropic response also involving 5-HT3 receptors. In SCI animals, the same effects were observed but with markedly greater amplitudes, suggesting a post-lesional hypersensitivity of autonomic circuits. In addition to providing a detailed characterization of the impact of high thoracic SCI on the cardiovascular system, this work demonstrates, for the first time, the potential of a serotonergic agonist to improve cardiovascular function following spinal cord injury. These findings open new therapeutic perspectives aimed at simultaneously enhancing locomotor and cardiovascular recovery after SCI.
ED Sciences Physiques et de l'Ingénieur
Prediction of mechanical fields for multi-scale simulation using physics-informed neural networks
by Manuel GUEVARA GARBAN (I2M - Institut de Mécanique et d'Ingénierie de Bordeaux)
The defense will take place at 13h30 - Amphithéâtre La Rochefoucauld-Liancourt École Nationale Supérieure des Arts et Métiers, Esplanade des Arts et Métiers, 33400 Talence
in front of the jury composed of
- Yves CHEMISKY - Professeur des universités - Université Grenoble Alpes - Directeur de these
- Olivier LEZORAY - Professeur des universités - Université de Caen Normandie - Rapporteur
- David RYCKELYNCK - Professeur des universités - MINES Paris – PSL - Rapporteur
- Pascal DESBARATS - Professeur des universités - Université de Bordeaux - Examinateur
- Stéphanie CHAILLAT-LOSEILLE - Directrice de recherche - CNRS - Examinateur
Advances in additive manufacturing have made it possible to print architectured materials such as lattice structures, triply periodic minimal surfaces, and reinforced composites, at the scale of real components, for applications ranging from lightweight aerospace structures to bone implants. The macroscopic behavior of these materials emerges from their microstructure, so predicting their mechanical response means simulating two scales at once: the component, and the unit cell it is built from. The reference tool for this task, the FE² method, solves a full boundary value problem on a representative volume element at every macroscopic integration point and every load increment; its cost is prohibitive for design practice. Faster alternatives exist, such as reduced order models and neural surrogates of the homogenized response, but they share one sacrifice: the local mechanical fields inside the microstructure are lost after the simulation, and with them every pointwise criterion that decides the integrity of a component. This thesis develops deep learning models that recover the local fields at the cost of a surrogate, fast enough for interactive use, while enforcing the mechanical equilibrium that training data alone does not guarantee. Three contributions are presented. The first contribution introduces a physics-informed graph neural network (GNN), namely P-DivGNN, for reconstructing local stress fields from the homogenized macroscopic stress in periodic microstructures. Training such a network on simulated finite element data alone does not guarantee physically admissible predictions. To reinforce the underlying physics, a regularization term enforcing discrete local equilibrium (the stress divergence-free condition), expressed at FEM mesh nodes, is incorporated into the loss function. The model is validated on a 2D plate-with-hole under both linear elasticity and nonlinear hyperelasticity, achieving approximately 500-fold speedup over finite element computations in the nonlinear case. The second contribution couples a long short-term memory (LSTM) recurrent network with a GNN to handle history-dependent elasto-plastic materials. The LSTM encodes the macroscopic strain-stress trajectory and passes its hidden state to the GNN for spatial localization. This coupling achieves up to 3500-fold speedup and generalizes to mesh topologies and refinement levels unseen during training. The third contribution extends the LSTM–GNN framework to three-dimensional elasto-plastic microstructures. The GNN reconstructs 3D stress fields on a TPMS Gyroid microstructure, at inference speeds compatible with interactive visualization. As in the 2D case, the model is tied to its training geometry and does not generalize to other representative volume elements. An exploratory study on a 3D convolutional neural network for TPMS classification and density estimation from voxel grids is presented as a candidate geometric feature descriptor for future conditioning of the constitutive model. Taken together, these contributions demonstrate the potential of deep learning architectures to reduce the computational cost of multi-scale simulation while respecting the physical constraints of mechanical equilibrium.