ED Mathématiques et Informatique
Scheduling recursive task graphs
by Thomas MORIN (LaBRI - Laboratoire Bordelais de Recherche en Informatique)
The defense will take place at 14h00 - amphi du LaBRI Amphitéatre du LaBRI Domaine universitaire, 351 Cr de la Libération, 33405 Talence
in front of the jury composed of
- Samuel THIBAULT - Professeur des universités - Université de Bordeaux - Directeur de these
- Vicenç BELTRAN - Directrice de recherche - Barcelona Supercomputing Center - Rapporteur
- Hatem LTAIEF - Directeur de recherche - KAUST - Rapporteur
- Abdou GUERMOUCHE - Maître de conférences - Université de Bordeaux - CoDirecteur de these
- Cédric AUGONNET - Ingénieur de recherche - Nvidia Research - Examinateur
- Thomas HÉRAULT - Directeur de recherche - Inria Bordeaux Sud-Ouest - Examinateur
- Loris MARCHAL - Directeur de recherche - Laboratoire de l'Informatique du Parallélisme - Examinateur
- Florina CIORBA - Associate Professor - University of basel - Examinateur
Over the past decades, supercomputers have become increasingly com- plex, with a growing number of nodes, GPUs, CPUs, and levels of memory hierar- chy, leading to ever more powerful platforms that, in 2026, exceed two quintillion floating-point operations per second. Fully exploiting these platforms is challeng- ing. To avoid having to rewrite an application whenever the underlying hardware changes, runtime systems have been developed. They provide an abstraction of the machine, allowing users to focus primarily on application development, while low-level details such as memory management, driver initialization, and commu- nications are handled automatically. Among existing runtime systems, many now rely on task-based programming, in which an application is divided into small units of work, called tasks. These tasks may depend on one another, thereby forming a Directed Acyclic Graph (DAG) that the runtime system can exploit. In particular, the Sequential Task Flow (STF) programming model makes it easy to express such DAGs by automatically constructing dependencies based on data accesses. However, STF, and more generally task-based programming models, face a major challenge: selecting the size of tasks, noted as granularity. Indeed, these models require the granularity to be selected at task submission time in order to ensure consistent execution. Moreover, CPUs and GPUs require different task granularities to be used efficiently, meaning that a statically selected granularity can sometimes limit performance. To address this issue, modern runtime systems such as OpenMP, PaRSEC, OmpSs, StarPU, and IRIS extend the task model by introducing recursive tasks. At runtime, recursive tasks can either be executed as regular tasks or submit a subDAG of tasks. This enables the final execution to use a finer granularity than the one selected at initial submission, thereby improving the dynamism of runtime systems. However, most runtime systems supporting re- cursive tasks leave the selection of the final granularity to the user, which reduces application portability. Indeed, each change in hardware generation may require the task granularity to be adjusted. This thesis addresses this issue by introducing several strategies to dynamically determine task granularity using the recursive task model. We propose different strategies to improve the utilization of both homogeneous and heterogeneous plat- forms. We also identify the adaptations required within runtime systems to enable dynamic and automatic task granularity selection using recursive tasks. Further- more, we investigate how the recursive task model can help address issues arising from the use of the STF model in a distributed-memory context. Throughout this thesis, we use the recursive task model provided by the StarPU runtime sys- tem, and evaluate the performance of our strategies using the dense linear algebra library Chameleon on different platforms.
ED Sciences Chimiques
Magnesium-rich ternary phases (LPSO and defined compounds) for hydrogen storage
by Mohamed Najib EL RHAZFOURI (ICMCB - Institut de Chimie de la Matière Condensée de Bordeaux)
The defense will take place at 14h00 - AMPHITHEATRE-ICMCB Institut de Chimie de la Matière Condensée de Bordeaux (ICMCB), UMR 5026, 87 avenue du Dr. Albert Schweitzer 33608 Pessac Cedex, France
in front of the jury composed of
- Jean-Louis BOBET - Professeur des universités - Université de Bordeaux - Directeur de these
- Étienne GAUDIN - Maître de conférences - Université de Bordeaux - CoDirecteur de these
- Claudia ZLOTEA - Directrice de recherche - CNRS / ICMPE - Rapporteur
- Laetitia LAVERSENNE - Directrice de recherche - CNRS / Institut NEEL / Université de Grenoble Alpes - Rapporteur
- Mathieu PASTUREL - Directeur de recherche - CNRS / Institut des Sciences Chimiques de Rennes / Université de Rennes - Examinateur
- Cyril AYMONIER - Directeur de recherche - CNRS / Institut de Chimie de la Matière Condensée de Bordeaux / Université de Bordeaux - Examinateur
Abstract: In the context of the energy transition and the reduction of greenhouse gas emissions, hydrogen is emerging as a promising energy carrier, capable of contributing to the development of more sustainable energy systems. However, its large-scale use is notably limited by the challenges associated with its storage. Among the various technologies, solid-state storage represents a particularly interesting avenue due to the high volumetric capacities it can offer. This work focuses on the study of magnesium-rich ternary phases formed in RE–TM–Mg systems (RE = Rare Earths and TM = Transition Metals), with a view to assessing their potential for solid hydrogen storage. The first part of this work involves studying the Gd–Ni–Mg system along the Gd/Ni = 4/3 composition line, corresponding to the stoichiometry of the Gd₈Ni₆ clusters characteristic of a family of magnesium-rich ternary phases known as Long-Period Stacking Ordered phases (LPSO O). The samples were synthesised with magnesium contents ranging from 70 to 95 at%. Their characterisation revealed a direct relationship between the magnesium content and the nature of the LPSO phase formed. LPSO phases of the 12R, 10H, 18R and 14H types were thus observed as the magnesium content increased. Mixtures of phases were frequently observed; however, in the case of the initial composition containing 80 at.% Mg, with the formula Gd11.4Ni8.6Mg80, it was possible to obtain a single 10H-type LPSO phase, which had not previously been reported in this system. Upon the first hydrogen absorption, this phase irreversibly decomposes into GdH2, Mg2NiH4 and MgH2. Subsequent hydrogenation cycles rely mainly on the reversibility of the Mg/MgH2 and Mg2Ni/Mg2NiH4 pairs. In the case of the other samples, an increase in the Mg content leads to an increase in reversible storage capacity, but is accompanied by a gradual slowing of the hydrogen sorption kinetics. The second, more exploratory part of this work is devoted to investigating the influence of the nature of the rare earth element and the transition metal on the stabilisation of magnesium-rich ternary phases. Replacing Gd with heavier rare earth elements (Tb, Dy, Ho and Er) consistently leads to the formation of LPSO phases. For the compositions studied in these systems, structural characterisation generally reveals the coexistence of LPSO phases, mainly of the 10H and 12R types, with a new magnesium-rich ternary phase of general composition TRNiMg2. However, a particularly notable case was observed for the composition Ho5.7Ni4.3Mg90, for which a single LPSO 18R-type phase was obtained. Replacing Gd with Pr, a lighter rare earth element, does not lead to the formation of LPSO phases. The study of the Pr–Ni–Mg system has enabled the identification of several ternary compounds, including Pr4Ni8Mg80, whose hydrogen storage properties have been investigated. Finally, the replacement of Ni with Co revealed a new magnesium-rich ternary phase in the Gd–Co–Mg system, with the composition Gd7Co3Mg90, which appears to correspond to an LPSO phase of the 18R type; this hypothesis remains to be confirmed. Keywords: Solid hydrogen storage; Magnesium; LPSO phases; Metal hydrides.
ED Sciences de la Vie et de la Santé
predictive imaging and clinical monitoring of post-stroke emotional disability
by Suhrit DUTTAGUPTA (Institut de neurosciences cognitives et intégratives d'Aquitaine)
The defense will take place at 14h00 - CARF Salle de conférence 146 Rue Léo Saignat, Centre d'Appui à la Recherche et de Formation, 33000, Bordeaux
in front of the jury composed of
- Igor SIBON - Professeur des universités - praticien hospitalier - Université de Bordeaux - Directeur de these
- Solène MOULIN - Professeure des universités - praticienne hospitalière - Hopital Maison Blanche - CHU Reims - Examinateur
- Nicolas FARRUGIA - Enseignant-Chercheur - Institut Mines-Télécom Atlantique - Rapporteur
- Charles LAIDI - Professeur des universités - praticien hospitalier - Université Paris-Est-Créteil-Val-de-Marne - Rapporteur
- Sandra CHANRAUD - Maîtresse de conférences - École Pratique des Hautes Études - Examinateur
Stroke is a leading cause of death and disability. Although patient rehabilitation mainly targets motor and functional impairments, post-stroke mood disruptions are prevalent and equally debilitating. Conditions such as post-stroke depression (PSD), anxiety (PSA), and fatigue (PSF) are difficult to predict. Conventional neuroimaging markers such as lesion location and white matter disruptions have provided limited predictive value, while advanced imaging techniques are difficult to implement in clinical practice. The main purpose of this thesis was to improve the management of post-stroke mood impairments. We aimed to improve the predictability of mood outcomes using neuroimaging markers derived from routine clinical MRI data and evaluate clinical measures that capture mood symptoms longitudinally. Data were acquired from 1085 patients across four French hospital-based cohorts: Brain Before Stroke (BBS; Bordeaux), Groupe de Réflexion pour l'Evaluation COGnitive Vasculaire (GRECOGVasc; Amiens), memoQUEST (Bordeaux), and MObile Technologies In the preVention of POSt-stroke DEPression (MOTIV-POSDEP; Bordeaux). Four studies were conducted. Study I examined PSF at 3 months post-stroke, where traditional voxel-based analyses provided limited results. Using a data-driven approach with principal component analysis (PCA), whole-brain networks were derived from lesion distributions. We identified distinct white- and gray-matter networks associated with specific fatigue domains, highlighting the multifaceted nature of post-stroke neuropsychological impairments. Building on this framework of data compression, Study II developed a neuroimaging pipeline using diffusion-weighted MRI to generate global diffusivity components through PCA. These components were entered into nonparametric regression models to predict 6-month PSA and PSD scores, comparing performances with clinical and disconnectome-based models. Overall, external predictive performance was low and clinical models performed best largely due to including baseline mood status. Nevertheless, the imaging pipeline added modest predictive value across evaluations, supporting the use of whole-brain markers to complement established predictors. Noting the reliance on baseline mood in the predictive models, we explored whether remote evaluations following hospital discharge could provide more utility. Study III evaluated the use of ecological momentary assessments (EMA), collecting daily measures of patient mood and activity through smartphones for 12 weeks. Based on symptoms matching the clinical diagnostic criteria for mood disorders, anxiety and depression mood indices were created. The association between mood indices and follow-up PSA and PSD scores improved with closer proximity. Compared to baseline clinical scores, averaged EMA mood index values from the second month offered better early prediction of follow-up mood disorders with 90.6% adherence. Additionally, responses for specific core symptoms, including anhedonia, sadness, and worrying had stronger associations with follow-up mood than the overall index scores. These observations led to Study IV evaluating a lower-burden approach based on SMS responses shortly before follow-up. Brief symptom ratings collected within two weeks prior to follow-up sessions showed strong associations with clinical evaluations for PSA, PSD, and PSF, demonstrating the potential of simplified remote monitoring to prioritize patients at risk of developing mood disorders. In summary, post-stroke neuropsychological outcomes are multifaceted and interacting syndromes with limited predictability. Whole-brain neuroimaging components derived through a novel framework showed modest predictive value. Our findings are consistent with the growing recognition of the need for holistic approaches, suggesting that incorporating multidomain models and longitudinal remote assessments may improve the early characterization of individual mood symptoms.
ED Sociétés, Politique, Santé Publique
Prediction and decision under partial information
by Pierre CATOIRE (Bordeaux Population Health Research Center)
The defense will take place at 14h00 - Module 2.6 bât CROUS 2ème étage Université de Bordeaux, Campus Carreire 146 Rue Léo Saignat, 33000 Bordeaux
in front of the jury composed of
- Cécile PROUST-LIMA - Directrice de recherche - Inserm / Université de Bordeaux - Directeur de these
- Ruth KEOGH - Professor - London School of Hygiene & Tropical Medicine - Rapporteur
- Ewout STEYERBERG - Professor - UMC Utrecht - Rapporteur
- Robin GENUER - Maître de conférences - Université de Bordeaux - CoDirecteur de these
- Hélène JACQMIN-GADDA - Directrice de recherche - Inserm / Université de Bordeaux - Examinateur
- Marine LE MORVAN - Chargée de recherche - INRIA - Examinateur
- Laure WYNANTS - Professeure assistante - Maastricht University - Examinateur
Modern predictive and decision algorithms are developed, evaluated, and deployed in environments where predictors may be partially observed. In such settings, the dominant theoretical framework is inherited from the statistical inference literature, where missingness is primarily treated as a nuisance that prevents consistent estimation of model parameters. This thesis argues that this perspective is incomplete when the objective is prediction rather than parameter estimation. In predictive settings, the missingness pattern itself may carry information about the outcome of interest. Consequently, a fundamental distinction arises between two target probability distributions: the distribution of the outcome conditional on the observed predictors alone, and the distribution conditional on both the observed predictors and the observation pattern. We develop a unified theoretical framework for prediction and decision under partial information, distinguishing between prediction strategies that condition on the missingness pattern and those that do not. We characterise the conditions under which these strategies coincide or differ, and assumptions under which consistent prediction is guaranteed. For both targets, those conditions are weaker than the "Missing at Random" assumption required in inference. This distinction clarifies several apparent inconsistencies in the literature and provides a principled interpretation of existing methods. Building on this framework, we derive conditions of validity for commonly used approaches at the stages of model development, evaluation, and deployment. In particular, we show that evaluating a prediction model intended for prediction under complete observation of predictors generally does not yield consistent estimation of the predictive performance to be expected at deployment. Finally, we extend the paradigm to decision analysis, where predictive models inform action choices. We characterise optimal decisions under restricted information and study the value of additional data acquisition. Application of the theoretical results are illustrated using clinical examples in emergency medicine context.