Archive ouverte de Centrale Lyon
Not a member yet
32420 research outputs found
Sort by
What pins edge dislocations in random alloys? Comparing size and elastic heterogeneities
No full textInternational audienceThis work compares the impact of size mismatch and elastic heterogeneities on solute pinning of edge dislocations in random body-centered cubic alloys -solid solutions in which different atomic species occupy lattice sites at random. Using atomic-scale modeling of both parametric alloys and more realistic binary systems (W-Ta and W-Nb) across the full composition range, we show that the mismatch of elastic constants between the alloy constituents can have an effect comparable to, or even greater than, the size mismatch. This contrasts with some solid solution strengthening theories, which assume a homogeneous elastic medium. Our results further demonstrate that the combined effects of size and modulus mismatches are well captured by a quadratic form, consistent with historical phenomenological approaches
Explainable Analysis of Patient Profiles with Sleep Disorders: Clustering Approach and Identification of Key Factors
No full textSleep disorders have a major impact on patients, but their diagnosis remains complex due to the diversity of symptoms. Today, technological advances and the analysis of medical data offer new perspectives for better understanding them. In this work, we use real data from the KANOPEE application to analyze these disorders. Our objective is to better understand sleep disorders using artificial intelligence. In particular, eXplainable Artificial Intelligence (XAI), which aims to make AI decisions understandable and interpretable by users. We propose two methods: (1) a regression method to predict the average sleep duration, and (2) a method based on clustering to classify patients according to different disorder profiles. In both cases, an XAI approach is used to identify the key features influencing the target variable. A concrete application on anonymized real-world data demonstrates the relevance of our methods
Carbon dioxide sorption in earthen plasters and its impact on Indoor Air Quality
No full textInternational audienceRaw earth-based materials play a significant role in creating comfortable indoor environments, particularly due to the presence of clay minerals, which can interact with indoor air constituents such as moisture and provide passive regulation. This study explores the interactions between raw earth plasters and carbon dioxide (CO) in indoor air, with a focus on their potential role in passive CO regulation, particularly in densely occupied or poorly ventilated spaces. Experiments were conducted using an emission chamber with continuous CO monitoring to assess: (i) the influence of humidity on sorption processes, (ii) sorption differences between two types of earth materials and a plasterboard, and (iii) the effect of plaster thickness. Results show that raw earth plasters effectively adsorb CO, with higher adsorption under high humidity. Furthermore, CO diffusion occurs within the plaster, leading to greater adsorption with increased plaster thickness. Key coefficients, adsorption and desorption rate constants (, ) and the material/air partition coefficient (), were determined. Using these coefficients, simulations were performed to predict the impact of earthen plasters on CO concentrations in a typical bedroom. Simulations predict that earth plasters can reduce CO levels, with a modest, yet measurable, 8% reduction overnight compared to a room without plasters
Characterization of β-Ga2O3 and device fabrication
No full textInternational audienceMonoclinic β-Ga2O3 is an ultra-wide bandgap (UWBG) semiconductor (Eg ~ 4,5 eV) with great physical properties, including a high breakdown field (up to 8 MV/cm) [1], a short absorption edge, and excellent chemical/thermal stability. These characteristics, as well as the recent progress in single-crystal substrate growth of this material, have raised interest in gallium oxide for high-voltage and low-loss power electronics, as a complement or potential alternative to wide-bandgap semiconductors such as silicon carbide (SiC) and gallium nitride (GaN)
GrootWatch à EXIST 2025 : Détection Automatique du Sexisme sur les Réseaux Sociaux - Classification de Tweets et de Memes
No full textInternational audienceThis paper presents our participation in the EXIST (sEXism Identification in Social neTworks) challenge at CLEF 2025, focusing on the classification of tweets and memes. We participated in all the tasks for tweets and memes, including both hard and soft classifications for tweets and hard classification for memes. For tweet classification, we propose a multi-task headed BERT model enriched with relevant information surrounding the tweet, helping the model achieve a full understanding of the tweet and its context. For memes, the paper explores the use of a Vision-Language Model (VLM)-based application to detect and categorise sexism in different scenarios, leveraging the ability of such models to understand the relationship between images and text in situations where sexist ideas are often expressed subtly. Our solutions achieved excellent performance, ranking first in all soft-soft tweet classification tasks and second in all hard-hard meme classification tasks.Content Warning: This paper includes examples of hateful, explicit and sexist language presented for illustrative purposes.Ce document présente notre participation au défi EXIST (sEXism Identification in Social neTworks) lors de CLEF 2025, en nous concentrant sur la classification des tweets et des memes. Nous avons participé à toutes les tâches pour les tweets et les memes, incluant à la fois des classifications hard et soft pour les tweets et des classifications hard pour les memes. Pour la classification des tweets, nous proposons un modèle BERT multi-tâche enrichi d'informations pertinentes entourant le tweet, aidant le modèle à comprendre pleinement le tweet et son contexte. Pour les memes, ce document explore l'utilisation d'une application basée sur un modèle de langage-visuelle (VLM) pour détecter et catégoriser le sexisme dans différents scénarios, en exploitant la capacité de tels modèles à comprendre la relation entre les images et le texte dans des situations où les idées sexistes sont souvent exprimées subtilement. Nos solutions ont obtenu d'excellents résultats, se classant premiers dans toutes les tâches de classification de tweets soft-soft et seconds dans toutes les tâches de classification de memes hard-hard.Avertissement de contenu : Ce document comprend des exemples de langage haineux, explicite et sexiste présentés à titre illustratif
Enhancing Knowledge Tracing with Large Language Models (LLMs)
No full textInternational audienceIn intelligent tutoring systems (ITS), knowledge tracing (KT) is a fundamental requirement for effective education and data mining. The main objective of KT is to model and predict the evolving understanding level of a student on different educational tasks. Traditional KT methods, such as the factor analysis method (FAM), Bayesian KT (BKT), and deep KT (DKT) approaches, have achieved high-performance effectiveness but often fail to identify reasoning processes, diverse learning trajectories, and complex interdependence relationships between skills associated with educational questions. The existing challenges highlight the requirement for personalization and contextual adaptability for effective student modeling. Addressing these challenges is important for building an ITS that can provide personalized educational experiences and support lifelong learning for diverse students.</div
Processeur sécurisé et reconfigurable incluant des technologies émergentes
No full textThe Internet of Things (IoT) is a rapidly expanding ecosystem in which intelligent objects interact via communicating networks. This rapid growth leads to a massive increase in the amount of data collected, posing challenges in terms of energy efficiency, particularly at sensor nodes. To improve this efficiency, it is essential to process data as close as possible to the sensors, thus reducing the load of communication and operations on the main computing unit. The Near Sensor Computing approach and the use of non-volatile memories (NV), capable of maintaining the state of sensors in standby mode, make it possible to meet these challenges while significantly increasing the energy efficiency of IoT-related systems.Alongside these energy challenges, data security in the IoT has become a major concern, particularly for applications where security is a crucial aspect, such as connected vehicles and smart healthcare systems. Attacks on secure communications have demonstrated the vulnerability of today's systems to sophisticated threats. Protecting data as soon as it is collected, even before it is transmitted, is therefore crucial. The SECRET project aims to meet this need by applying memory-centric computing paradigms and integrating reconfigurable NV operators as close as possible to the sensors. This provides a first line of defense, while optimizing energy consumption by reducing memory accesses. The emergence of new technologies, such as ferroelectric field-effect transistors (FeFETs) and resistive memories (RRAMs), opens up promising prospects for the design of circuits based on memory-centric architectures.These approaches, which break away from Von Neumann or Harvard architectures, make it possible to limit the bottlenecks associated with data transfers between computing and storage units. However, their integration into complex circuits raises several major technical challenges. On the one hand, the choice of models must be judicious in order to ensure the accuracy of simulations and the reliability of results throughout the design flow. On the other hand, current design tools, mainly developed for conventional architectures, are still poorly adapted to the specificities of circuits exploiting an intrinsic memory effect. In particular, taking into account the NV properties of components and their impact on logic synthesis requires adapted methodologies. This thesis shows that it is possible to design logic gates based on FeFETs, and then proposes an application within NV operators.L'Internet des Objets, raccourcie en IoT pour Internet of Things, est un écosystème en pleine expansion, où des objets intelligents interagissent via des réseaux communicants. Cette croissance rapide entraîne une augmentation massive des données collectées, posant des défis en termes d'efficacité énergétique, notamment au niveau des nœuds de capteurs. Pour améliorer cette efficacité, il est essentiel de traiter les données au plus près des capteurs, ce qui réduit la charge de communication et d'opérations sur l'unité de calcul principale. L'approche du Near Sensor Computing et l'utilisation de mémoires non volatiles (NV), capables de maintenir l'état des capteurs en veille, permettent de répondre à ces défis tout en augmentant significativement l'efficacité énergétique des systèmes liés à l'IoT.Parallèlement à ces défis énergétiques, la sécurité des données dans l'IoT est devenue une préoccupation majeure, notamment pour des applications où la sécurité est un aspect primordiale comme les véhicules connectés et les systèmes de santé intelligents. Des attaques sur les communications sécurisées ont démontré la vulnérabilité des systèmes actuels face à des menaces sophistiquées. Ainsi, la protection des données dès leur collecte, avant même qu'elles ne soient transmises, est cruciale. Le projet SECRET tente de répondre à ce besoin en appliquant des paradigmes de calcul centrés sur la mémoire et en intégrant des opérateurs reconfigurables et NV au plus près des capteurs. Offrant ainsi une première ligne de défense tout en optimisant la consommation énergétique grâce à la réduction des accès mémoire. L'émergence de technologies émergentes, telles que les transistors à effet de champ ferroélectriques (FeFET) et les mémoires résistives (RRAM), ouvre des perspectives prometteuses pour la conception de circuits basés sur des architectures centrées sur la mémoire.Ces approches, en rupture avec les architectures de Von Neumann ou Harvard, permettent de limiter les goulots d'étranglement liés aux transferts de données entre unités de calcul et de stockage. Cependant, leur intégration dans des circuits complexes soulève plusieurs défis techniques majeurs. D'une part, le choix des modèles doit être judicieux afin d'assurer la précision des simulations et la fiabilité des résultats tout au long du flot de conception. D'autre part, les outils de conception actuels, principalement développés pour des architectures conventionnelles, sont encore peu adaptés aux spécificités des circuits exploitant un effet mémoire intrinsèque. En particulier, la prise en compte des propriétés NV des composants et leur impact sur la synthèse logique nécessitent des méthodologies adaptées. Cette thèse montre qu'il est possible de concevoir des portes logiques basées sur des FeFET puis propose une application au sein d'opérateurs NV
Le projet 4BLife : intégrer le vieillissement des batteries dans le dimensionnement et la gestion des micro-réseaux
No full textInternational audienceThis paper presents the results of the 4BLife project, which aims to integrate aging laws into the sizing and management of storage systems using lithium-ion batteries. This involves defining one or more aging models as well as tools for performance characterization and State of Health (SoH) monitoring. Laboratory and field data are used to identify aging models and validate SoH tracking tools. Two battery technologies, LFP and NMC, suitable for both embedded and stationary applications, are studied. The integration of aging laws into the design and management of stationary storage systems is explored through simulation. The validation of aging impacts under real-world operating conditions is currently under investigation.Ce papier présente les résultats du projet 4BLife dont l'objectif est d'intégrer des lois de vieillissement dans le dimensionnement et la gestion de dispositifs de stockage utilisant des batteries lithium-ion. Cela nécessite de définir une ou plusieurs lois de vieillissement ainsi que des outils de caractérisation de performances et de suivi de l'état de santé (SoH). Des données en laboratoire et de terrain sont exploitées pour identifier les modèles de vieillissement et valider les outils de suivi de SoH. Deux technologies de batteries LFP et NMC aptes à fonctionner en condition embarquée et stationnaire sont étudiées. L'intégration de lois de vieillissement dans le dimensionnement et la gestion de systèmes de stockage stationnaire est abordée par la simulation. La validation des impacts sur le vieillissement des batteries en condition réelle est en cours d'étude
Enhancing Structural Reliability Analysis with Machine Learning-Based Surrogate Models: Theoretical and Experimental Insights
No full textThis paper presents a comprehensive framework for structural dynamics and reliability analysis by integrating theoretical modeling of continuous systems and experimental investigations of multi-storey steel frames. First, a beamlike structure with infinite degrees of freedom is analyzed using the Rayleigh-Ritz method, highlighting the importance of partial di!erential equations in capturing distributed mass, sti!ness, and damping. Tuned Mass Dampers (TMDs) are incorporated to mitigate resonant vibrations, and advanced sampling techniques (Latin Hypercube Sampling, Monte Carlo simulation) are employed to quantify structural reliability under uncertain loading conditions. Machine learning models, including Random Forest and Neural Networks, are then developed as surrogate models to predict failure probabilities, revealing critical nonlinear relationships between system parameters. To validate and extend these insights, the framework is applied to a twostorey steel frame tested under controlled laboratory conditions. A mechanical shaker supplies dynamic excitations with varying statistical characteristics-kurtosis, root mean square (RMS), skewness, and crest factor-while force and acceleration measurements capture the structure's real-time responses. By training machine learning algorithms (Random Forest, Gradient Boosting, XGBoost, and Neural Networks) on time-domain features, the study demonstrates the capability of data-driven methods to capture complex vibratory behaviors beyond what standard mechanical models predict.</div
Comment prendre en compte le sentiment de compétence des apprenant.e.s au travers des EIAH
No full textNational audienceLe sentiment de compétence, impactant les performances et l'orientation des apprenant.e.s, est un sujet complexe qu'il serait intéressant de prendre en compte dans les EIAH, puisque ceux-ci se sont déjà avérés utiles sur des questions de performance et motivation. Nous proposons ici notre approche, basée sur les traces d'interactions des apprenant.e.s afin de le détecter. Cette approche combine l'utilisation du Raisonnement à Partir de l'Expérience Tracée pour rechercher dans les traces des patterns définis par un expert humain et l'utilisation du Machine Learning pour faire émerger des patterns. Ces deux approches, confrontées l'une à l'autre, sont également confrontées à des questionnaires permettant une auto-évaluation du sentiment de compétence des apprenant.e.s.</div