Portail HAL edf
Not a member yet
11029 research outputs found
Sort by
Concevoir une IA explicable, appropriable et de confiance : approche située par et pour les métiers
No full textInternational audienceLes récents progrès de l'intelligence artificielle (IA) ont permis des innovations dans de nombreux domaines, notamment grâce aux techniques de machine learning et de deep learning. Cependant, ces techniques, souvent perçues comme opaques, soulèvent d'importants défis en matière de transparence et d'explicabilité, en particulier dans les contextes de prise de décision. Pour y répondre, un courant appelé intelligence artificielle explicable (Explainable AI ou XAI) a émergé, visant à améliorer la compréhension de ces technologies. Toutefois, les recherches actuelles en XAI restent majoritairement technocentrées, en négligeant le contexte d'usage et l'activité des utilisateurs. Ce projet de thèse propose donc une approche anthropocentrée, cherchant à comprendre et à transformer l'activité humaine à travers la co-conception d'un dispositif d'aide au diagnostic
How to introduce an initial crack in phase field simulations to accurately predict the linear elastic fracture propagation threshold?
No full textInternational audienceVariational phase field fracture models are now widely used to simulate crack propagation in structures. A critical aspect of these simulations is the correct determination of the propagation threshold of pre-existing cracks, as it highly relies on how the initial cracks are implemented. While prior studies briefly discuss initial crack implementation techniques, we present here a systematic investigation. Various techniques to introduce initial cracks in phase field fracture simulations are tested, from the crack explicit meshing to the replacement by a fully damaged phase field, including different variants for the boundary conditions. Our focus here is on phase field models aiming to approximate, in the -convergence limit, Griffith quasi-static propagation in the framework of Linear Elastic Fracture Mechanics. Therefore, a sharp crack model from classic linear elastic fracture mechanics based on Griffith criterion is the reference in this work. To assess the different techniques to introduce initial cracks, we rely on path-following methods to compute the sharp crack and the phase field smeared crack solutions. The underlying idea is that path-following ensures staying at equilibrium at each instant so that any difference between phase field and sharp crack models can be attributed to numerical artifacts. Thus, by comparing the results from both models, we can provide practical recommendations for reliably incorporating initial cracks in phase field fracture simulations. The comparison shows that an improper initial crack implementation often requires the smeared crack to transition to a one-element-wide phase band to adequately represent a displacement jump along a crack. This transition increases the energy required to propagate the crack, leading to a significant overshoot in the force-displacement response. The take-home message is that to predict the propagation threshold accurately and avoid artificial toughening; the crack must be initialized either setting the phase field to its damage state over a one-element-wide band or meshing the crack explicitly as a one-element-wide slit and imposing the fully cracked state on the crack surface
Au-delà des discours, l’IA générative à l’épreuve des usages réels en entreprise
No full textInternational audienc
Online Episodic Convex Reinforcement Learning
No full textInternational audienceWe study online learning in episodic finite-horizon Markov decision processes (MDPs) with convex objective functions, known as the concave utility reinforcement learning (CURL) problem. This setting generalizes RL from linear to convex losses on the state-action distribution induced by the agent’s policy. The non-linearity of CURL invalidates classical Bellman equations and requires new algorithmic approaches. We introduce the first algorithm achieving near-optimal regret bounds for online CURL without any prior knowledge on the transition function. To achieve this, we use an online mirror descent algorithm with varying constraint sets and a carefully designed exploration bonus. We then address for the first time a bandit version of CURL, where the only feedback is the value of the objective function on the state-action distribution induced by the agent's policy. We achieve a sub-linear regret bound for this more challenging problem by adapting techniques from bandit convex optimization to the MDP setting
Apprentissage Automatique pour la sélection de modèles additifs généralisés en ligne : application à la prévision de consommation électrique
No full textInternational audienceElectricity demand forecasting is key to ensuring that supply meets demand lest the grid would blackout. Reliable short-term forecasts may be obtained by combining a Generalized Additive Models (GAM) with a State-Space model (Obst et al., 2021), leading to an adaptive (or online) model. A GAM is an over-parameterized linear model defined by a formula and a state-space model involves hyperparameters. Both the formula and adaptation parameters have to be fixed before model training and have a huge impact on the model's predictive performance. We propose optimizing them using the DRAGON package of Keisler (2025), originally designed for neural architecture search. This work generalizes it for automated online generalized additive model selection by defining an efficient modeling of the search space (namely, the space of the GAM formulae and adaptation parameters). Its application to short-term French electricity demand forecasting demonstrates the relevance of the approach
Hot electrons and cold holes: operation, efficiency and design of a two-temperature hot-carrier solar cell
No full textInternational audienceHot-carrier solar cells (HCSCs) offer potential for enhancing the energy-conversion efficiency of photovoltaic devices up to 86%. However, most HCSC models to date assume that electrons and holes have the same temperature, while many reports in III-V materials indicate that electrons can be much hotter than their counterparts. We present here a detailed balance HCSC model that includes different temperatures for electrons and holes. We focus on the impact of the temperature imbalance on the voltage of such a HCSC, and on its power-conversion efficiency. Surprisingly, a temperature imbalance at a fixed effective temperature leads to a slight power-conversion efficiency increase, up to 1-2 percentage points, primarily due to an increase in fill factor and possibly of opencircuit voltage. Yet, we show that the knowledge of the effective temperature alone is sufficient to design a satisfying HCSC
Dynamics of COVID-19 crisis management in hospitals and its long-term effects: An analysis using organizational resilience
No full textInternational audienceThis empirical and qualitative study focuses on COVID-19 crisis management in a French hospital, analyzing it from the perspective of organizational resilience to understand its evolution over time. The study identifies adaptation factors during pandemic management, supported by success factors, and it also identifies difficulty factors associated with resilience. These factors are analyzed at the different waves of the pandemic to understand the evolution of organizational resilience across various crisis management temporalities. The results highlight how certain factors initially considered as resilience capabilities, evolve to become vulnerability factors of the sociotechnical system, in particular due to their impacts on healthcare personnel
Local-scale experimental investigation of a two-phase cross-flow in a tube bundle and flow-induced vibration; Bubbly flow regime
No full textInternational audienceTwo-phase cross-flows can induce vibrations in several industrial situations. This is especially the case in nuclear power plant U-tube Steam Generator (SG) tube bundles. With the purpose of gathering high quality data for the validation of multiphase CFD simulation tools, a new experimental apparatus was designed and put in operation. The facility is instrumented for the two-phase air–water flow and tube vibration characterization. The two-phase flow was investigated starting from the bubble generation at the gas injection, to inside the tube bundle. Bubble sizes, shapes and velocities were measured by means of high speed camera image post-processing in the region comprised from the gas injection to the inlet of the tube bundle: the bubbles are generated with unstable size and shape, this results in promoting their breakup before reaching the tube bundle. The local behavior of the two-phase flow within the tube bundle was studied through an optical dual-tip probe, placed in different positions in order to obtain profiles of void fraction, bubble diameter and gas velocity. To study the flow-induced vibrations, the central tube of the bundle was designed to be flexible, and its vibrational response was studied by accelerometers. The root mean square displacements could be derived and the variation of the added mass with the void fraction could be observed. This paper represents the first part of two papers and focuses on bubbly flow experimental tests; the second part focuses on experimental tests performed at the so-called intermittent flow-regime conditions
Experimental investigation of Dean-vortices oscillation downstream of a 90° Bend
No full textInternational audienceThe present work experimentally investigates the dynamics of a pipe flow downstream of a 90 degrees bend in turbulent regime. The study is carried out on two different test benches, enabling to cover a decade in Reynolds numbers (Re is an element of [1.2 x 104, 5.4 x 105]). Laser-based metrology techniques are employed to capture the three velocity components in several flow sections, namely, cross-sections orthogonal to the main flow and vertical diametral planes (parallel to the main flow direction). Time-resolved and long-time decorrelated measurements allow the characterisation of both the dynamics and the statistics of the flow. These measurements highlight the behaviour of the fully three-dimensional flow generated downstream of the bend. In particular, an oscillation of the dipolar structure generated by the bend, known as the Dean vortices, is measured and analysed using a Lamb-Chaplygin analytical model. The dependency of the flow behaviour as a function of both the Reynolds number and the distance downstream of the bend and the return to axisymmetric flow are evaluate
Comment surveiller et prévoir la qualité microbiologique des sites de baignade dans les cours d'eau urbains ? Le site d'étude du bassin de la Villette (Paris)
No full textInternational audienceSwimming in urban rivers is becoming increasingly popular. In order to limit the health risks to bathers, a warning system can be used to anticipate episodes of microbiological contamination and thus better manage any bathing closures. This article presents a short-term (24 hours) forecasting system developed at La Villette bathing site in Paris, where a public bathing area was opened in 2017. It is based on the measurement of physico-chemical and hydraulic variables and on the in situ measurement of faecal indicator bacteria (FIB) upstream of the bathing area. The functioning of this early warning system is illustrated during 2021 summer. The data collected from the 1st June to the 31st August 2021 showed a major contamination event following heavy rainfall in mid-July. This event led to the closure of the bathing facility. By integrating these measurements with a hydrodynamic model, TELEMAC-3D, the transport of bacteria and their spatial distribution were simulated. The observations coupled with the simulations also highlighted the thermal stratification and its impact on the velocity and flow direction, which can significantly alter the transfer time of FIB. The results demonstrate the ability of this approach to anticipate contamination peaks, providing useful information to manage closure and reopening of urban bathing areas.La baignade dans les cours d’eau urbains bénéficie d’un attrait grandissant. Afin de limiter les risques sanitaires pour les baigneurs, un système d’alerte permet d’anticiper des épisodes de contamination microbiologique et ainsi de mieux gérer l’éventuelle fermeture à la baignade. Cet article présente un système prédictif à court terme (48 heures) développé sur le site d’étude du bassin de La Villette à Paris, où une baignade publique est ouverte en été depuis 2017. Il s’appuie sur la mesure de variables physico-chimiques, hydrauliques et sur la mesure in situ de bactéries indicatrices fécales (BIF) en amont de l’espace de baignade. Le fonctionnement de ce système de suivi et d’alerte est illustré sur l’été 2021. Les données collectées du 1er juin au 31 août 2021 ont révélé un événement de forte contamination microbiologique, ayant entrainé une fermeture du site de baignade, après de fortes précipitations survenues à la mi-juillet. L'intégration de ces mesures à un modèle hydrodynamique, TELEMAC-3D, a permis de simuler le transport des bactéries et leur distribution spatiale. Les mesures, ainsi que la modélisation, ont aussi permis de montrer les effets de la stratification thermique sur la vitesse et la direction de l’écoulement, modifiant alors fortement le temps de transfert des BIF. Les résultats mettent en évidence la capacité de cette approche à anticiper les pics de contamination, fournissant ainsi des informations utiles pour décider de la fermeture et la réouverture des zones de baignade