Portail HAL edf
Not a member yet
11029 research outputs found
Sort by
Rethinking Synthetic Twins: Average Treatment Effect Estimation with Latent Representations Learning
No full textInternational audienceLa demande croissante de flexibilité énergétique a conduit au développement de programmes de réduction de la consommation en période de pointe. EDF R&D a lancé plusieurs initiatives utilisant les données des compteurs Linky. Cependant, évaluer leur impact reste difficile en raison des modèles de consommation individualisés, du biais de sélection dû à la participation volontaire et des limites des méthodes classiques d'inférence causale. Nous simulons un environnement contrôlé pour examiner ces défis et comparons les approches traditionnelles avec des méthodes d'apprentissage automatique comme SyncTwin, qui crée des jumeaux synthétiques pour gérer les variables cachées. Nos résultats montrent que les séries temporelles complexes défient les méthodes classiques, surtout avec des variables non observées influençant le traitement. Nous expliquons théoriquement pourquoi ces approches échouent, en raison de covariables inaccessibles influençant le score de propension. Nos travaux ouvrent la voie à de nouvelles techniques pour améliorer l'évaluation des programmes d'économie d'énergie
Méthodes à noyaux rapides: Sobolev, régression informée parla physique, et modèles additifs
No full textKernel methods are powerful tools in statistical learning, but their cubic complexity in the sample size n limits their use on large-scale datasets. In this work, we introduce a scalable framework for kernel regression with O(n log n) complexity, fully leveraging GPU acceleration. The approach is based on a Fourier representation of kernels combined with non-uniform fast Fourier transforms (NUFFT), enabling exact, fast, and memory-efficient computations. We instantiate our framework in three settings: Sobolev kernel regression, physics-informed regression, and additive models. When known, the proposed estimators are shown to achieve minimax convergence rates, consistent with classical kernel theory. Empirical results demonstrate that our methods can process up to tens of billions of samples within minutes, providing both statistical accuracy and computational scalability. These contributions establish a flexible approach, paving the way for the routine application of kernel methods in large-scale learning tasks.Les méthodes à noyaux sont des outils puissants en apprentissage statistique, mais leur complexité cubique en fonction de la taille de l’échantillon n limite leur utilisation sur des ensembles de données de grande dimension. Dans ce travail, nous introduisons un cadre général pour ramener la régression à noyaux à une complexité en O(n log n), tout en exploitant pleinement l’accélération par carte graphique (GPU). L’approche repose sur une représentation de Fourier des noyaux combinée avec des transformées de Fourier rapides non uniformes (NUFFT), permettant des calculs exacts, rapides et peu coûteux en mémoire. Nous instancions notre cadre dans trois contextes : la régression à noyau de Sobolev, la régression informée par la physique et les modèles additifs. Quand ils sont connus, les estimateurs proposés atteignent des vitesses de convergence minimax, en accord avec la théorie classique des noyaux. Les résultats empiriques montrent que nos méthodes peuvent traiter jusqu’à plusieurs dizaines de milliards d’échantillons en quelques minutes, tout en offrant à la fois une précision statistique et une scalabilité computationnelle. Ces contributions ouvrent la voie à une application courante des méthodes à noyaux dans des tâches d’apprentissage à grande échelle
Research and development to bring the PUMAS (Plutonium Uranium Mono-Amide Separation) process to indus-trial maturity
No full textInternational audienceAs an integral part of the multi-recycling of materials, future processing plants will be able to handle a wider variety of spent fuels, not only those based on uranium (UOX and ERU) but also those based on a mixture of uranium and plutonium (MOX LWR then MOX RNR). Capacitive processing of spent MOX requires the separation and purification of in-creased flows of plutonium to ensure multi-recycling and thus progress in the circular economy of uranium resources.In addition to the industrially established PUREX process, CEA has been carrying out more intensive research over the past 15 years, in conjunction with industry, to simplify the separation/purification process with a view to a new plant capable of processing fuels with increasingly high plutonium concentrations. This would require more extensive use of the current PUREX process, based on a Pu redox principle, leading to more complex man-agement of safety, reagent consumption, effluent treatment and nitrate waste.The PUMAS process, the result of this research, is based on the remarkable properties of an extractant molecule from the monoamide family, selected according to multiple rigor-ously tested criteria. The Pu/U selectivity is no longer based on variations in redox for pluto-nium, but on controlled variations in the acidity of the aqueous phase in contact with the solvent, constituting a major overall simplification.The current development program is aimed at industrial development along three lines, with deadlines set jointly by CEA, Orano and EDF:- Area 1: Basic data (physical chemistry of the process);- Area 2: Multi-criteria performance validation;- Area 3: suitability for industrialization (endurance tests, pilots on various scales
Assessing Agentic AI & AI Agents Concepts and Platforms through the lens of Traditional Multi-Agent Systems Principles
No full textInternational audienceMulti-Agent Systems are a long-standing subfield of artificial intelligence, with foundational concepts emerging as early as the 1980s. Multi-Agent Systems involves collections of autonomous agents interacting within an environment and organized within a system. Although early research showed strong potential, computational limitations at the time restricted real-world adoption. The recent rise of AI Agents and Agentic AI -driven by advances in Large Language Models -has renewed interest in Multi-Agent concepts. These developments have given rise to a new wave of Agent-based platforms and applications. However, many of these modern implementations only partially reflect the theoretical foundations of traditional Multi-Agent Systems, often resulting in limited explicability, robustness, coordination, or scalability. As a first contribution in this paper, we clarify the evolving terminology surrounding "AI Agents" and "Agentic AI", through the lens of Traditional Multi-Agent Systems formalisms. We propose a synthesis combining theoretical rigor of Traditional Multi-Agent Systems with current technological capabilities of Agentic AI systems, which is paving the way to what we call "Modern MAS". As a second contribution, we use the formalisms of Traditional MAS to compare contemporary Agentic AI platforms
Convergence analysis of overlapping domain decomposition preconditioners for nonlinear problems
No full textInternational audienceNumerical simulations of nonlinear partial differential equations often involve solving large nonlinear systems, for which Newton's method is widely employed due to its fast convergence near the solution. However, its performance can deteriorate in the presence of strong nonlinearities or poor initial guesses. Nonlinear overlapping domain decomposition methods, such as RASPEN and Substructured RASPEN (SRASPEN), have proven effective in addressing these challenges. Because SRASPEN reduces the problem size by restricting computations to a substructure, it does not update the solution outside the substructure, so that no natural initial guesses for the nonlinear local solution exists that might lead to additional inner subdomain nonlinear iterations or even prevent the local solvers to converge. In this study, we analyze the convergence of RASPEN. We show how domain decomposition improves the convergence rate of the Newton's method by highlighting the key role of the substructure on the global error contraction. Moreover, our analysis provides insight into an inexpensive modification to SRASPEN that mitigates the lack of iterations outside the substructure. The proposed variant significantly reduces computational cost while improving overall efficiency compared to existing techniques in the literature. Numerical experiments confirm the computational performance and robustness of the improved SRASPEN, establishing it as a reliable approach for solving large-scale nonlinear systems
Vers une modélisation semi-distribuée pour la simulation des étiages : comparaison de quatre modèles hydrologiques sur le bassin de la Meuse
No full textNational audienc
Accelerating Solar Cell Research Through Physics Modeling, Bayesian Machine Learning, and Automation
No full textInternational audienceThe race to develop more efficient and stable solar technologies demands advanced tools to unravel complex performance factors and degradation mechanisms. In this work, we present an integrated approach that combines physics-based modeling with Bayesian machine learning to gain deeper insights into photovoltaic performance and degradation. We show how this tool has been used to investigate the impact of various passivation layers in perovskite solar cells, revealing key efficiency factors that were experimentally validated through our collaboration with NTU via the SINERGIE French-Singaporean research network. Building on the findings of this study and other related work, we outline the development of our automated platform for the fabrication and characterization of next-generation solar cells. This platform will leverage the physics-informed Bayesian optimization tools we have developed to explore the parameter space more intelligently and efficiently, ultimately contributing to the creation of an international network of automated research platforms
Game Theory and Multi-Agent Reinforcement Learning for Zonal Ancillary Markets
No full textWe characterize zonal ancillary market coupling relying on noncooperative game theory. To that purpose, we formulate the ancillary market as a multi-leader single follower bilevel problem, that we subsequently cast as a generalized Nash game with side constraints and nonconvex feasibility sets. We determine conditions for equilibrium existence and show that the game has a generalized potential game structure. To compute market equilibrium, we rely on two exact approaches: an integrated optimization approach and Gauss-Seidel best-response, that we compare against multi-agent deep reinforcement learning. On real data from Germany and Austria, simulations indicate that multi-agent deep reinforcement learning achieves the smallest convergence rate but requires pretraining, while bestresponse is the slowest. On the economics side, multi-agent deep reinforcement learning results in smaller market costs compared to the exact methods, but at the cost of higher variability in the profit allocation among stakeholders. Further, stronger coupling between zones tends to reduce costs for larger zones
An alternative elastoplastic model for ductile fracture with graded plasticity
No full textInternational audienceThe prediction of failure modes in metallic structures is a crucial step in the safety analysis of industrial components subjected to significant mechanical loads (e.g., nuclear power plant components, pipelines, etc.). To perform such analyses, it is essential to accurately simulate the propagation of a defect in the ductile regime, characterized by large plastic deformations before and during propagation. Predictive numerical simulation of ductile fracture remains an open scientific and technicalchallenge, despite significant progress in recent years. The so-called local approach to fracture is widely used to model ductile fracture; however, like all softening models, it exhibits dependency on spatial discretization. The objective of this work is to propose a robust approach for modeling and simulating ductile fracture at the structural scale. It introduces an elastoplastic model with a bounded gradient of accumulated plastic strain and an associated set of internal constraints. This method aims to eliminate mesh dependency, without increasing the number of degrees of freedom of the numerical problem as it can be the case for most of non-local models. So, the proposed model strives for fast and efficient numerical computations at the structural scale. The feasibility of this approach is demonstrated through the resolution of problems on simple geometries, such as cylinders or spheres made of composite elastoplastic materials. Tests on 316L steel specimens are also conducted to further validate the robustness of this method
Réponse structurelle et fonctionnelle des communautés benthiques et piscicoles aux modifications long-terme des paramètres abiotiques du Rhône.
No full textInternational audienceA search for similarities or differencies in the temporal trends of functioning and evolution of long-term series in large rivers was carried out from the confrontation of different hydroclimatic and physicochemical variables put in relation to each other. the temporal evolution and spatial variability of two biological descriptors, fish and benthic macroinvertebrates of the Rhône over the period 2000-2019 over a stretch of nearly 250 km of river, in contrasted sectors both from of geographical point of view than hydromorphological.The analyzes showed the existence of spatial co-structures and temporal co-dynamics between the abiotic environment of the Rhône and the composition of its populations, making it possible to bring out a set of concomitant evolutions between the latter and the environmental variables characterizing the current functioning of the river. They gradually highlighted marked trends linked to global changes. However, the speed and magnitude of the changes observed require more precise determination of cause and effect relationships in order to identify potential levers for action. These different observations also highlight the complementarity of the biological models studied, such a holistic approach allowing an efficient understanding of the structure and functioning of contemporary large regulated rivers.Une recherche de similarités ou de discordances des tendances temporelles de fonctionnement et d’évolution de séries long-terme en grand cours d’eau a été menée à partir de la confrontation de différentes variables hydroclimatiques et physico-chimiques mises en vis-à-vis de l’évolution temporelle et de la variabilité spatiale de deux descripteurs biologiques, les poissons et les macroinvertébrés benthiques du Rhône sur la période 2000-2019 sur un linéaire de près de 250 km de fleuve, dans des secteurs fluviaux contrastés tant du point de vue géographique qu’hydromorphologique.Les analyses ont montré l’existence de co-structures spatiales et de co-dynamiques temporelles entre l’environnement abiotique du Rhône et la composition de ses peuplements, permettant de faire ressortir un ensemble d’évolutions concomitantes entre ces derniers et les variables environnementales caractérisant le fonctionnement actuel du fleuve. Elles ont progressivement mis en exergue des tendances marquées liées aux changements globaux. Cependant, la rapidité et l’ampleur des changements observés impliquent de déterminer plus précisément les relations de cause à effet dans le but d’identifier de potentiels leviers d’action. Ces différents constats mettent également en avant la complémentarité des modèles biologiques étudiés, une telle approche holistique permettant une compréhension efficiente de la structure et du fonctionnement des hydrosystèmes contemporains anthropisés