Portail HAL edf
Not a member yet
11029 research outputs found
Sort by
Cutting the Black Box: Conceptual Interpretation of a Deep Neural Net with Multi-Modal Embeddings and Multi-Criteria Decision Aid
No full textInternational audienceThis paper tackles the concept-based explanation of neural models in computer vision, building upon the state of the art in Multi-Criteria Decision Aid (MCDA). The novelty of the approach is to leverage multi-modal embeddings from CLIP to bridge the gap between pixel-based and concept-based representations. The proposed Cut the Black Box (CB2) approach disentangles the latent representation of a trained pixel-based neural net, referred to as teacher model, along a 3-step process. Firstly, the pixel-based representation of the samples is mapped onto a conceptual representation using multi-modal embeddings. Secondly, an interpretable-by-design MCDA student model is trained by distillation from the teacher model, using the conceptual sample representation. Thirdly, the alignment of the teacher and student latent representations spells out the concepts relevant to explaining the teacher model. The empirical validation of the approach on ResNet, VGG, and VisionTransformer on Cifar-10, Cifar-100, Tiny ImageNet, and Fashion-MNIST showcases the effectiveness of the interpretations provided for the teacher models. The analysis reveals that decision-making predominantly relies on few concepts, thereby exposing potential bias in the teacher's decisions
Effect of shoaling length on rogue wave occurrence
No full textInternational audienceThe impact of shoaling on linear water waves is well known, but it has only been recently found to significantly amplify both the intensity and frequency of rogue waves in nonlinear irregular wave trains atop coastal shoals. At least qualitatively, this effect has been partially attributed to the ‘rapid’ nature of the shoaling process, i.e. shoaling occurs over a distance far shorter than that required for waves to modulate themselves and adapt to the reduced water depth. Through a theoretical model and highly accurate nonlinear simulations, we disentangle the respective effects of the length and angle of a shoal's slope. We investigate the effects of the shoaling process rapidness on the evolution of key statistical and spectral sea-state parameters. We let the wave field evolve over a slope with constant angle in all cases while we vary the slope length. Our results indicate that the non-equilibrium dynamics is not affected by the slope length, because further extending the slope length does not influence the magnitude of the statistical and spectral measures as long as the non-equilibrium dynamics dominates the wave evolution. Thus, the shoaling effect on rogue waves is deduced to be mainly driven by the slope magnitude rather than the slope length
Failure modes of silicon heterojunction photovoltaic modules in damp heat environment : Sodium and moisture effects
No full textInternational audienceSilicon heterojunction (SHJ) solar cells are expected to gain significant market share in the coming years. In the field, among identified degradation modes, moisture-induced degradation can be a significant concern for this solar cell technology and should be monitored. This work investigates the moisture-induced degradation mechanisms in SHJ cells encapsulated in different module configurations. Damp heat (DH) testing was performed under IEC 61215 standard conditions (85 °C and 85% relative humidity) for up to 2000 h. Different degradation mechanisms are identified after DH aging, due to moisture alone or in combination with sodium ions originating from photovoltaic glass leaching. Under the influence of moisture, these ions can migrate into the cell and degrade the cell passivation, resulting in massive power losses up to 57.6% of the initial value after 1500 h of DH aging. By using other types of glass, glass-glass module configurations show less than 3% of power losses after 2000 h of DH aging. The front side of the cell is much more sensitive than the rear side where the emitter of the cell is. After highlighting the impact of sodium, moisture alone was studied with a module configuration without glass. In that case, the degradation is characterized by increased series resistance without passivation losses
Austenite-martensite interfacial patterns and energy dissipation of phase transformation in Ni-Mn-Ga single crystal
No full textInternational audienceThe martensitic phase transformation occurs in Shape Memory Alloys (SMA) via the nucleation/propagation of the Austenite-Martensite interface (A-M interface), which is a transition region (domain) between the two coexisting phases. For providing compatibility, the transition region consists of various martensitic twin structures (laminates), forming different interfacial patterns, such as parallel laminates and branching laminates. Due to the energy accumulation in the interfacial structures (e.g., elastic mismatch and twin-boundary surface energy), the interfacial patterns should be relevant to the driving force (energy dissipation) and the associated kinetics of the phase transformation. In this paper, we adopt a special thermal loading, small-temperature-gradient "heatingcooling-reheating", to control the A-M interface's forward and reverse propagation to generate different interfacial patterns in a Ni-Mn-Ga single crystal SMA with the observation on the twin structures (by optical microscope and SEM) and the InfraRed measurement on the temperature hysteresis of the interface propagation (for characterizing the thermal driving force). Simple energetic analysis indicates that the thermal driving force (energy dissipation) is directly related to the stored energy in the interfacial structure, particularly the large mismatch elastic energy near the habit plane. This study not only provides the details of the various interfacial patterns and their dependence on the 2 / 34 loading path, but also indicates the driving force and the associated mechanism about the pattern evolution to understand the phase transformation process
Représentation des structures en béton armé dans les simulations basées sur la FDTD pour calculer les champs magnétiques induits par la foudre
No full textInternational audienceReinforced concrete structures are generally well protected against the adverse effects of a direct lightning strike. Nevertheless, when it occurs, the electromagnetic field generated by the lighting current flowing through their reinforcement can cause a malfunction of sensitive electronic devices. The electromagnetic field can be estimated using full-wave methods, although the accuracy of the estimation depends on the representatives of the model, which is why experimental validation is essential. This paper presents the numerical model of a reinforced concrete structure where a current was injected to emulate a direct lightning strike. Full-wave simulations are conducted using TEMSI-FD and the results are compared to the measurements made at the testing facility of EDF TEGG. The simulation results were in good agreement with the measurements, especially when the soil was used as the return path for the current.Les structures en béton armé sont généralement bien protégées contre les effets néfastes d'un coup de foudre direct. Néanmoins, lorsque cela se produit, le champ électromagnétique généré par le courant d'éclairage circulant à travers leur armature peut provoquer un dysfonctionnement des appareils électroniques sensibles. Le champ électromagnétique peut être estimé à l'aide de méthodes à ondes pleines, bien que la précision de l'estimation dépende des représentants du modèle, c'est pourquoi la validation expérimentale est essentielle. Cet article présente le modèle numérique d'une structure en béton armé dans laquelle un courant a été injecté pour simuler un coup de foudre direct. Des simulations pleine onde sont réalisées à l'aide de TEMSI-FD et les résultats sont comparés aux mesures effectuées à l'installation d'essai d'EDF TEGG. Les résultats de la simulation sont en bon accord avec les mesures, en particulier lorsque le sol est utilisé comme voie de retour pour le courant
Méthodes d'inversion de type one-shot et décomposition de domaine
No full textOur main goal is to analyze the convergence of a gradient-based optimization method, to solve inverse problems for parameter identification, in which the corresponding forward and adjoint problems are solved by an iterative solver. Coupling the iterations for the three unknowns (the inverse problem parameter, the forward problem solution and the adjoint problem solution) yields the so-called one-shot inversion methods. Many numerical experiments showed that using very few inner iterations for the forward and adjoint problems may still lead to a good convergence for the inverse problem. This motivates us to develop a rigorous convergence theory for one-shot methods using a fixed small number of inner iterations, with a semi-implicit scheme for the parameter update and a regularized cost functional. Our theory covers a general class of linear inverse problems in the finite-dimensional discrete setting, for which the forward and adjoint problems are solved by generic fixed point iteration methods. By studying the spectral radius of the block iteration matrix of the coupled iterations, we prove that for sufficiently small descent steps the (semi-implicit) one-shot methods converge. In particular, in the scalar case, where the unknowns belong to one-dimensional spaces, we establish not only sufficient but even necessary convergence conditions on the descent step. Next, we apply one-shot methods to (linearized and then non-linear) inverse conductivity problems, and solve the forward and adjoint problems by domain decomposition methods, more specifically nonoverlapping optimized Schwarz methods. We analyze a domain decomposition algorithm that simultaneously calculates the forward and adjoint solutions for a given conductivity. By combining this algorithm with the gradient descent parameter update, we obtain a domain decomposition one-shot method that solves the inverse problem. We propose two discretized versions of the coupled algorithm, the second of which (in the case of the linearized inverse conductivity problem) falls into the abstract framework of our convergence theory. Finally, several numerical experiments are provided to illustrate the performance of the one-shot methods, in comparison with the classical gradient descent in which the forward and adjoint problems are solved using direct solvers. In particular, we observe that, even in the case of noisy data, very few inner iterations may still guarantee good convergence of the one-shot methods.Notre objectif principal est d’analyser la convergence d’une méthode d’optimisation basée sur le gradient, pour résoudre des problèmes inverses d’identification de paramètres, dans laquelle les problèmes directs et adjoints correspondants sont résolus par un solveur itératif. Le couplage des itérations pour les trois inconnues (le paramètre du problème inverse, la solution du problème direct et la solution du problème adjoint) donne ce que l’on appelle les méthodes d’inversion de type one-shot. De nombreux tests numériques ont montré que l’utilisation de très peu d’itérations internes pour les problèmes directs et adjoints peut néanmoins conduire à une bonne convergence pour le problème inverse. Cela nous motive à développer une théorie de convergence rigoureuse pour les méthodes de type one-shot en utilisant un petit nombre fixe d’itérations internes, avec un schéma semi-implicite pour la mise à jour du paramètre et une fonction de coût régularisée. Notre théorie couvre une classe générale de problèmes inverses linéaires dans le cadre discret de dimension finie, pour lesquels les problèmes directs et adjoints sont résolus par des méthodes génériques d’itération de point fixe. En étudiant le rayon spectral de la matrice par blocs des itérations couplées, nous prouvons que pour des pas de descente suffisamment petits, les méthodes de type one-shot (semi-implicites) convergent. En particulier, dans le cas scalaire, où les inconnues appartiennent à des espaces à une dimension, nous établissons des conditions de convergence suffisantes et même nécessaires sur le pas de descente. Ensuite, nous appliquons des méthodes de type one-shot aux problèmes inverses de conductivité (linéarisés et puis non linéaires), et résolvons les problèmes directs et adjoints par des méthodes de décomposition de domaines, plus spécifiquement des méthodes de Schwarz optimisées sans recouvrement. Nous analysons un algorithme de décomposition de domaine qui calcule simultanément les solutions directe et adjointe pour une conductivité donnée. En combinant cet algorithme avec la mise à jour du paramètre par descente de gradient, nous obtenons une méthode one-shot de décomposition de domaine qui résout le problème inverse. Nous proposons deux versions discrétisées de l’algorithme couplé, dont la seconde (dans le cas du problème inverse de conductivité linéarisé) s’inscrit dans le cadre abstrait de notre théorie de convergence. Enfin, plusieurs expériences numériques sont fournies pour illustrer les performances des méthodes de type one-shot, en comparaison avec la méthode de descente de gradient classique dans laquelle les problèmes directs et adjoints sont résolus par des solveurs directs. En particulier, nous observons que, même dans le cas de données bruitées, très peu d’itérations internes peuvent toujours garantir une bonne convergence des méthodes de type one-shot
Spatial variability in the seasonal precipitation lapse rates in complex topographical regions – application in France
No full textInternational audienceSeasonal precipitation estimation in ungauged mountainous areas is essential for understanding and modeling a physical variable of interest in many environmental applications (hydrology, ecology, and cryospheric studies). Precipitation lapse rates (PLRs), defined as the increasing or decreasing rate of precipitation amounts with the elevation, play a decisive role in high-altitude precipitation estimation. However, the documentation of PLR in mountainous regions remains weak even though their utilization in environmental applications is frequent. This article intends to assess the spatial variability and the spatial-scale dependence of seasonal PLRs in a varied and complex topographical region. At the regional scale (10 000 km2), seven different precipitation products are compared in their ability to reproduce the altitude dependence of the annual/seasonal precipitation of 1836 stations located in France. The convection-permitting regional climate model (CP-RCM) AROME is the best in this regard, despite severe precipitation overestimation in high altitudes. The fine resolution of AROME allows for a precise assessment of the influence of altitude on winter and summer precipitation on 23 massifs at the sub-regional scale (∼ 1000 km2) and 2748 small catchments (∼ 100 km2) through linear regressions. With AROME, PLRs are often higher in winter at the catchment scale. The variability in the PLR is higher in high-altitude regions such as the French Alps, with higher PLRs at the border than inside the massifs. This study emphasizes the interest of conducting a PLR investigation at a fine scale to reduce spatial heterogeneity in the seasonal precipitation–altitude relationships
Improving the accuracy of the Jin-Xin relaxation scheme
No full textIn the present work, a proposition is made in order to improve the accuracy of the Jin-Xin relaxation scheme. This scheme is composed of two steps: a first step in which scalar quantities corresponding to relaxed variables are advected, and a second step associated with a relaxation procedure that allows retrieving the original variables. Both steps introduce some numerical diffusion, but we focus here on the first one. The idea is to test the GRU projection step which has recently been proposed. This algorithm is very simple and has interesting anti-diffusive properties. It is based on a random choice and is directly inspired from Glimm’s ideas. This first-order scheme is very well-suited for the advection of scalar quantities with discontinuities and can be applied to unstructured meshes
Optimization of functions defined over sets of points in polygons with evolutionary algorithms based on Wasserstein barycenters
No full textInternational audienc
A hybrid approach of semantic modeling and co-simulation for a better consideration of the physics of physical phenomena in a smart building
No full textInternational audienceCombination of simulation and semantic knowledge of the physical phenomena will improve decision making process of IoT systems deployed in smart buildings