Portail HAL ONERA
Not a member yet
12842 research outputs found
Sort by
Simulation hybride CFD-DSMC pour les jets de moteur-fusée à haute altitude
No full textAt high altitudes, the expansion of rocket plumes into rarefied atmospheres generates complex flowfields characterized by strong non-equilibrium effects. Accurately predicting this environment is essential to assess the signature of launch vehicles and the risk of communication blackouts. Conventional CFD methods based on the Navier-Stokes equations fail to capture rarefaction phenomena, while rarefied gas solvers such as DSMC are computationally prohibitive in dense regions. As a result, hybrid strategies that couple CFD and DSMC have been proposed, yet their application to full-scale propulsion systems remains limited.This work evaluates and extends hybrid coupling methodologies for rocket plume simulations at high altitude. First, the one-way coupling method, in which CFD supplies boundary conditions to DSMC, is validated using cold free-jet configurations. It is shown that the Maxwell-Smoluchowski wall law improves CFD accuracy near the nozzle wall and extends the coupling interface upstream. When conservative rarefaction criteria are used, hybrid solutions match full DSMC simulations for velocity and density. However, thermal fields require careful placement of the coupling interface.The two-way coupling method, allowing feedback from DSMC to CFD, is then assessed through three test cases: free jet, impinging jet, and toss back. While one-way coupling remains sufficient for expansion-dominated flows, only two-way coupling correctly captures upstream interactions in toss back configurations.The one-way method is applied to the third stage of JAXA's M-V launcher. Simulations highlight two key phenomena: backflow of jet species upstream of the vehicle, and radial segregation of light species such as hydrogen. A DSMC-DSMC coupling technique is developed to simulate the far plume, based on approximating velocity distribution functions as sums of maxwellians. This enables simulation up to 500 km downstream, revealing slow re-equilibration and persistent species stratification.Finally, a semi-analytical model is proposed to predict backflow intensity based on dimensionless numbers and validated against DSMC results. These findings support the use of one-way and two-way hybrid methods for multi-scale plume simulations, and open the way for further improvements including chemical reactivity, radiative effects, and dispersed phase dynamics.À haute altitude, l'expansion des panaches de fusée dans des atmosphères raréfiées génère des écoulements complexes, caractérisés par de forts effets de non-équilibre. Une prédiction précise de cet environnement est essentielle pour évaluer la signature des lanceurs ainsi que les risques d'interruptions de communication. Les méthodes classiques de CFD, basées sur les équations de Navier-Stokes, ne permettent pas de capturer les phénomènes liés à la raréfaction, tandis que les solveurs de gaz raréfié, tels que le DSMC, deviennent prohibitifs en termes de coût de calcul dans les régions denses. En conséquence, des approches hybrides couplant CFD et DSMC ont été proposées, bien que leur application à des systèmes de propulsion à l'échelle réelle reste encore limitée.Ce travail évalue et étend les méthodologies de couplage hybride pour la simulation des panaches de fusée en haute altitude. Dans un premier temps, la méthode de chaînage one-way, où la CFD fournit les conditions aux limites au DSMC, est validée à l'aide de configurations de jets libres froids. Il est montré que la loi de paroi de Maxwell-Smoluchowski améliore la précision de la CFD à proximité de la paroi de la tuyère et permet de remonter l'interface de couplage vers l'amont. Lorsque des critères de raréfaction conservatifs sont utilisés, les solutions hybrides reproduisent fidèlement les résultats DSMC de référence en termes de vitesse et de densité. En revanche, les champs thermiques requièrent un positionnement soigneux de l'interface de couplage.La méthode de couplage two-way, autorisant une rétroaction de la DSMC vers la CFD, est ensuite évaluée au travers de trois cas tests : jet libre, jet impactant, et jet à contre courant. Si le chaînage one-way demeure suffisant dans les écoulements dominés par l'expansion, seul le couplage two-way permet de capturer correctement les interactions amont dans les configurations de jet à contre courant.La méthode one-way est ensuite appliquée au troisième étage du lanceur M-V de la JAXA. Les simulations mettent en évidence deux phénomènes clés : le retour en amont des espèces du jet vers le véhicule (backflow), et la ségrégation radiale des espèces légères telles que l'hydrogène. Une technique de couplage DSMC-DSMC est développée pour simuler le panache lointain, en approximant les fonctions de distribution de vitesses par des sommes de lois de Maxwell. Cela permet de simuler jusqu'à 500 km en aval, révélant une rééquilibration lente ainsi qu'une stratification persistante des espèces.Enfin, un modèle semi-analytique est proposé pour prédire l'intensité du retour (backflow) en amont à partir de nombres sans dimension, et validé par comparaison avec les résultats DSMC. Ces travaux confirment la pertinence des méthodes hybrides one-way et two-way pour les simulations multi-échelles de panaches, et ouvrent la voie à de futurs développements incluant la réactivité chimique, les effets radiatifs, et la dynamique de la phase dispersée
Apprentissage incrémental pour le passage à l'échelle des algorithmes d'interprétation des images de télédétection
No full textThe volume of satellite and aerial imagery is rapidly increasing. Collected from diverse sensors over large territories, these data enable land-use mapping via semantic segmentation (per-pixel labeling). Segmentation models are trained from annotations and often assume i.i.d. data available all at once. Scaling. This thesis aims to develop a single model that progressively extends its geographic coverage without degrading performance on previously seen areas. However, generalization beyond the training distribution is limited, and retraining from cumulative data at every new site is virtually impossible due to compute, storage, and data availability. We therefore adopt an incremental framework with local updates, leveraging foundation models pretrained at scale. Domain-Incremental Learning (DIL). In DIL, classes remain fixed while the data distribution shifts with acquisition context. The main challenge is catastrophic forgetting: adapting to new domains can degrade performance on earlier ones. Controlling this plasticity-stability trade-off is the core objective of this work. Contributions. (1) In an initial study with U-Net, we show that updating only the decoder, even with replay, is insufficient when the frozen upstream representation lacks robustness to successive distribution shifts. (2) We then propose a framework with a frozen foundation encoder (ViT) and the trade-off confined to a parsimonious generative decoder based on Probabilistic PCA (PPCA); in classification, this achieves a satisfactory balance, strongly dependent on the encoder's representation quality. (3) For incremental segmentation, we introduce decision conditioning on a latent domain, inferred automatically from representations, to reduce inter-domain interference and stabilize updates. The trio — frozen foundation encoder, lightweight generative decoder, and latent-domain-conditioned decisions — provides a practical route to geographic scaling in semantic segmentation.Le volume d'images satellitaires et aériennes augmente rapidement. Issues de capteurs variés et couvrant de vastes territoires, ces données permettent la cartographie de l'occupation et des usages des sols par segmentation sémantique (étiquette par pixel). Les modèles de segmentation sémantique sont appris à partir d'annotations et supposent souvent des données d'entraînement disponibles d'un seul tenant et i.i.d. Le passage à l'échelle. Cette thèse vise à développer un modèle unique capable d'étendre progressivement sa couverture géographique sans dégrader les performances sur les zones déjà apprises. Or, la généralisation au-delà de la distribution d'entraînement est limitée, et entraîner à nouveau le modèle à partir des données cumulées à chaque nouveau site d'acquisition est virtuellement impossible (calcul, stockage, disponibilité). Nous adoptons donc un cadre incrémental avec mises à jour locales, en tirant parti des modèles de fondation pré-entraînés à grande échelle. Apprentissage incrémental de domaine (AID). En AID, les classes restent stables tandis que la distribution des données évolue selon le contexte d'acquisition (distribution shift). Le défi principal est l'oubli catastrophique : l'adaptation à de nouveaux domaines peut dégrader les performances sur les domaines antérieurs. Maîtriser ce compromis plasticité — stabilité est l'objectif central de ce travail. Contributions. (1) Dans une première étude avec U-Net, nous montrons que mettre à jour uniquement le décodeur, même avec replay, ne suffit pas lorsque la représentation amont figée manque de robustesse face à des distributions shifts successifs. (2) Nous proposons ensuite un schéma où l'encodeur de fondation (ViT) est figé et où le compromis plasticité — stabilité est confiné à un décodeur génératif parcimonieux fondé sur l'Analyse en Composantes Principales Probabiliste (ACPP) ; en classification, ce choix atteint un équilibre satisfaisant, fortement dépendant de la qualité de représentation de l'encodeur. (3) Enfin, pour la segmentation incrémentale, nous introduisons un conditionnement de la décision par un domaine latent, inféré automatiquement à partir des représentations, afin de réduire les interférences entre distributions et stabiliser les mises à jour. Ce cadre — encodeur de fondation figé, décodeur génératif léger, décision conditionnée par domaine latent — offre une voie pragmatique vers le passage à l'échelle géographique en segmentation sémantique
Urban Air Temperature Modeling: Combining Physical Simulation and Data-Driven Fine-Tuning
No full textInternational audienceAccurate urban climate modeling is crucial for addressing the growing impacts of urban heat islands (UHI) and climate change. Physics-based tools such as the Urban Weather Generator (UWG) are widely used but often limited by high parameterization needs and a lack of specialized data. In this study, we develop a hybrid framework combining UWG simulations with deep learning, introducing two models: NUWG-Sim (Neural Urban Weather Generator on Simulations), trained solely on simulated data, and NUWG-city, which is fine-tuned with ground weather station data. To systematically evaluate model performance across heterogeneous urban contexts, we structure our experiments around Local Climate Zones (LCZs) in Toulouse, France. Our methodology involves generating over 3400 UWG initialization files, simulating urban air temperatures time series for diverse surface parameters, and training a neural model on these series. We then fine-tune the model with observed data from selected weather stations, analyzing how the number and diversity of stations environments impact performance on unseen stations from different LCZs. Results show that even limited fine-tuning significantly improves performance, particularly when training includes stations from LCZs similar to the test set. The approach highlights the potential of physics-informed neural models for city-specific urban climate monitoring
Vortex Particle Velocity Diffusion Method Using Dynamic Turbulence Control Based on Enstrophy
No full textInternational audienceThe Vortex Particle Method (VPM) represents continuous, incompressible, unsteady, viscous flows by using singular, independent particles represented by small concentrations of vorticity. The VPM is particularly convenient for representing viscous flow fields, which is useful for simulating vortex reconnection in an unbounded domain and flow mixing. This study aims to evaluate the accuracy of the VPM predictions using, among others, the diffusion velocity method to represent the molecular diffusion of vortex structures. The VPM can also represent the viscous diffusion of turbulent flow fields using a turbulent viscosity model. However, such models are defined up to a given constant. The dynamic computation of this constant is required for the turbulence model to encompass the diffusion of all the flow scales. A new fast dynamic technique based on the local loss of enstrophy in the flow is proposed. The coupling of the Diffusion Velocity Method with this dynamic procedure is presented. Finally, the method is validated on several known analytical flow solutions, such as the Lamb–Oseen vortex and isolated and leapfrogging vortex rings
Multipactor in the Presence of Dielectrics: An Overview of Charging Effects Specificities and Physical Consequences
No full textInternational audienceThe multipactor effect is a resonant electron avalanche phenomenon that can occur in microwave components operating under vacuum conditions, when electrons are accelerated by radio-frequency (RF) fields. This phenomenon arises from the synchronization between the RF electromagnetic wave and the electron emission process, potentially leading to an exponential increase in electron density. Multipactor discharges are considered detrimental in various high-performance applications such as particle accelerators, controlled nuclear fusion systems, and, notably, space telecommunication payloads.The presence of multipactor can degrade system performance by increasing noise levels, and inducing local heating and pressure rises. Such local pressure increases can trigger secondary discharges, including corona effects and electrical breakdowns, which may damage RF components. While the multipactor effect is now relatively well understood and accurately modelled, particularly for configurations involving only conductive materials, its prediction remains significantly more complex when dielectric materials are involved.Benchmark studies (ESA EVEREST project) comparing simulations and experimental results have shown that multipactor thresholds can be reliably predicted in metal-only structures, provided that the Total Electron Emission Yield (TEEY) models are well calibrated against experimental data. TEEY, defined as the ratio of emitted secondary (SE ) and backscattered(BSE) to incident electrons, plays a crucial role in these models.However, when dielectrics are introduced within the critical gap, additional phenomena come into play. Surface charge accumulation on dielectric materials, driven by effects such as triboelectrification, exposure to space radiation, or even the multipactor process itself can generate quasi-static electric fields. These fields influence both the secondary emission processes at the dielectric surface and the trajectories of electrons in the vacuum. Despite their benefits in reducing the size and mass of microwave components, dielectrics introduce considerable uncertainty into multipactor predictions due to poorly understood charge effects.As a result, conservative RF power margins, up to 12 dB, are routinely applied in the design of satellite-borne components that include dielectrics. These large margins reflect the current limitations in measuring or modelling the TEEY of dielectric materials and in predicting the dynamics of surface charge: its magnitude, spatial distribution, and evolution over both short (during discharge events) and long timescales (across the mission duration).This article aims to explore these challenges and assess their impact on the prediction of multipactor thresholds in RF systems incorporating dielectric materials. We begin by reviewing the main mechanisms leading to excess surface charging in dielectrics within satellite payloads. We then examine how these charges affect electron emission, with particular emphasis on the TEEY. The trend toward miniaturization in space hardware also prompts us to consider the emergence of electrostatic discharge phenomena. Finally, we conclude with an outline of the key research challenges and suggest directions for future investigation
Straightforward Method to Orient Black Phosphorus from Bulk to Thin Layers using a Standard Green Laser
No full textInternational audienceThe crystallographic orientation of anisotropic 2D materials plays a crucial role in their physical properties and device performance. However, standard orientation techniques such as transmission electron microscopy (TEM) or X‐ray diffraction can be complex and less accessible for routine characterization. Herein, the orientation of black phosphorus (BP) from bulk crystals to thin layers is investigated using angle‐resolved polarized Raman spectroscopy with a single‐wavelength (514 nm) Raman setup. By incorporating thickness‐dependent interference effects and anisotropic optical indices, this approach provides a reliable framework for orientation determination across different BP thicknesses. The method is validated through direct orientation measurements using TEM and electron backscattering diffraction, confirming its applicability to both thick and ultrathin samples. Given its simplicity and compatibility with widely available Raman setups, this approach offers a practical solution for characterizing BP orientation without requiring advanced structural characterization techniques
Time Schemes for Solving Maxwell’s Equations in a Mesh Hybridization Strategy
No full textInternational audienceThe efficient resolution of Maxwell’s equations remains a priority in the field of electromagnetic simulation due to the increasing complexity of the structures and systems to be studied or qualified. In the time domain, several schemes have been and are still being studied due to their advantages. In our case, for several years, we have believed that efficient simulation of a computational scene requires dividing it into multiple zones, each treated with the appropriate scheme. In this strategy, we focus on a global meshing of the scene into a Cartesian grid with polyhedral cell inclusions. This approach has the advantage of maximizing the number of Cartesian cells, which is particularly efficient for our methods, while allowing for zones adapted to curved geometries and refinement areas that account for the multi-scale aspects of the computational scene. In our presentation, we propose to showcase the multi-domain/multi-method strategy implemented by providing examples and studies conducted on the hybridization of certain schemes. We will then focus more specifically on the latest research carried out within the framework of theses, using FVTD (Finite Volume Time Domain) and CDO (Compatible Discrete Operator) schemes
Extracting aircraft conflict-resolution situations from historical ADS-B data
No full textInternational audienceExisting conflict resolution models are often based on theoretical frameworks that, while providing optimal solutions under specific criteria, may not fully align with real-world controller decision-making practices. This gap between model predictions and actual behaviour can lead to low acceptance of automated tools. Understanding how controllers resolve conflicts in daily operations could help design assistance tools that generate advisories more likely to be accepted and integrated into their workflow. This study introduces a data-driven methodology for identifying and cataloguing air traffic deconfliction instances using historical ADS-B data. By analysing trajectory deviations and their impact on predicted aircraft separations, we extract instances of deconfliction and encode them into a structured dataset. This dataset captures key elements of each event, including sector information, deviated aircraft details, predicted non-deviated trajectories, and surrounding traffic conditions. Our approach facilitates large-scale analysis of air traffic control decision-making, providing a foundation for developing conflict resolution models that better reflect operational practices. This paper details the methodology and process used to construct this dataset
Physique d’un revêtement furtif acoustique de type fractal applicable dans le domaine naval
No full textAcoustique sous-marine et navale, bioacoustique en milieu marin; GAPSUS - Acoustique Physique, Sous-Marine et Ultra-Sonore: GVB - Vibro acoustique et Contrôle du Bruit: GBIO - BioacoustiqueNational audienceLes revêtements furtifs acoustiques actuels applicables dans le domaine naval sont principalement basés sur le principe d’absorption : ainsi on utilise aujourd’hui très majoritairement des "tuiles anéchoïques", c’est à dire des matériaux qui dissipent l’énergie d’une onde incidente sur une certaine bande de fréquences. On peut distinguer différents types de matériaux potentiellement utilisables en milieu marin, soit les polymères pour leur amortissement, les matériaux fibreux à base de carbone ou verre pour améliorer la raideur à faible masse ajoutée ou les métaux sous forme de tiges ou de partitions pour leur faible compressibilité et leur densité. Ils sont associés à un nombre de concepts relativement limité destinés à réduire la réflexion des ondes : premièrement, pour protéger et isoler une antenne sonar vis-à-vis de bruits parasites externes et internes ou, deuxièmement, dans un but de furtivité ou de discrétion, si appliqué sur une coque d’engin sous-marin. Il s’agit, dans le cas présent, d’évaluer par simulation l’interaction entre des éléments en acier de forme évolutive suivant la profondeur et un matériau anéchoique en élastomère. En l’occurrence, sont proposées des configurations avec des éléments de type fractal « Cantor » d’ordre variable. L’effet de l’inclusion d’éléments pleins en acier ou des coques de forme cylindrique, hexagonale et auxétique, disposés entre les éléments en acier de type fractale, est également étudié. Il apparait que les hétérogénéités fractales peuvent piloter la position des maxima d’amplitude d’accélération au sein du revêtement (modes locaux). Par ailleurs, la modification d’absorption procurée par la présence des inclusions est d’autant plus importante lorsque les éléments sont situés sur les ventres initiaux. Dans tous les cas, l’accélération effective transversale est modifiée localement. Il y a donc un transfert d’énergie entre déplacements longitudinal et transversal
StrikeNet: A Deep Neural Network to predict pixel-sized lightning location
No full textInternational audienceForecasting the location of electrical activity at a very short time range remains one of the most challenging predictions to make, primarily attributable to the chaotic nature of thunderstorms. Additionally, the punctual nature of lightning further complicates the establishment of reliable forecasts. This article introduces StrikeNet, a specialized Convolutional Neural Network (CNN) model designed for very short-term forecasts of electrical activity locations, utilizing sequences of temporal images as input and only two data types. Employing soft Non-Maximum Suppression (NMS) techniques, incorporating morphological features within residual blocks, and implementing dropout regularization, StrikeNet is specifically designed for detecting and predicting pixel-sized objects in images. This design seamlessly aligns with the task of forecasting imminent electrical activity giving great scores of 0.42 for precision, 0.78 for detection, and an F1-Score about 0.54