HAL Arts et Métiers
Not a member yet
14127 research outputs found
Sort by
Cryogenic effects on the mechanical behavior of bulk metallic glasses
No full textInternational audienceBulk metallic glasses (BMGs), unlike crystalline alloys, exhibit significantly enhanced plastic deformation when tested at cryogenic temperatures. This enhanced plasticity is primarily characterized by the slowed propagation of shear bands and the formation of multiple shear bands, which play a crucial role in the material's behavior at low temperatures. Due to their amorphous nature, BMGs are prone to catastrophic fractures once shear band nucleation and propagation occur, a behavior distinct from that of crystalline materials. However, the underlying mechanisms of BMG failure and the effect of strain rate remain controversial. This study investigates the mechanical behavior of a Zr-based BMG under cryogenic conditions. Compression tests were conducted at room temperature and -180°C, using liquid nitrogen, across a range of strain rates. The results show that, at cryogenic temperatures, ductility increases, though it remains relatively low, leaving uncertain its impact on machinability. Notably, larger stress drops were observed at ambient temperature, likely linked to shear band formation. Additionally, the study identified two distinct fracture modes during dynamic tests, warranting further investigation. This research provides valuable insights into the behavior of BMGs under cryogenic conditions and their machinability
Elaboration of 316L/CU composite alloy using a hybrid pvd/sps process
No full textInternational audiencePowder metallurgy is an ideal field for developing advanced materials with complex geometries and compositions. Iron-based alloys produced through sintering often exhibit significant porosity
Trunk Acceleration Asymmetries in Transfemoral and Transtibial Amputee Runners
No full textInternational audienc
Observation and large‐eddy simulation of an offshore atmospheric undular bore during sea‐breeze initiation
No full textInternational audienceA sea‐breeze (SB) initiation under land synoptic wind near a peninsula is analyzed by means of LiDAR measurements and large‐eddy simulations (LES) using the Weather Research and Forecasting (WRF) framework. In the simulation results, local SBs initiated over several coast segments converge in the morning and form a front near the peninsula. As the marine atmospheric boundary layer is stably stratified, the front generates an undular bore featuring gravity waves (GW). Despite the absence of cloud signatures, the GW are detected in the LiDAR horizontal scans, providing a direct observation. The GW have a low propagation speed, a small wavelength, and their amplitude decreases with increasing distance from the coastline. The GW amplitude increases with the strengthening of the local convergence and then decreases when the local SBs merge into a regional SB. The turbulent kinetic energy (TKE) profile in the SB without GW forms a peak in the center of the SB cell. Within the GW, a significant part of the TKE calculated from simulation results is related to the GW horizontal motion. A method is proposed to extract only the part originating from the turbulent field. A peak in the TKE profile is also observed, and its intensity is 50%–100% higher near the crests than at the troughs. Analysis of the TKE budget demonstrates that, at the peak height, the shear production is positive only in the ascending phase of the GW, and that the TKE maximum at the crests mainly results from the advection process
Comparison of parametric model order reduction methods to solve magneto-quasistatic and electro-quasistatic problems
No full textInternational audienceIn this paper, we compare two parametric model order reduction methods, the multi-moment matching method and the interpolation of projection subspaces method for the magneto-quasistatic (MQS) and electro-quasistatic (EQS) problems derived from Maxwell’s equations and discretized with the Finite Element (FE) method. The two problems considered are both governed by the differential–algebraic equations. The material characteristic parameters as well as the geometry parameters have been considered. The applications are two realistic test cases: an EQS model of a transformer bushing under insulation defect uncertainty and a MQS model of a planar inductor with geometric and material variations. The result shows that both methods approximate well global quantities, such as the current or the voltage, as well as the local quantities like field distributions. The multi-moment matching method remains always faster in the online stage, since the reduced basis is not parameter dependent, requiring no reduced basis calculation. The multi-moment matching method requires an affine decomposition of the FE model, which is not easy to obtain when considering geometry parameters. A hybrid method is proposed and tested leading to more accurate results than the interpolation of projection subspaces method but much easier to implement than the multi-moment matching method
Apprendre avec un Jumeau Numérique
No full textInternational audienceApprendre avec un Jumeau Numériqu
Effects of the Tool Microgeometry on Thermo-Mechanical Loads for Ti-6Al-4V Finishing Cutting Operations
No full textInternational audienceAchieving a high-quality surface while minimizing the adverse effects associated with machining operations, particularly in finishing processes, is a major challenge in aerospace component manufacturing. To overcome the limitations of conventional numerical simulations of cutting operations, which involve chip formation and complex contact management, this paper proposes a simplified approach to model the thermal loading generated during cutting using a chipless method. After analytically determining the effort-based thermal sources and validating the chipless model, the influence of tool microgeometry on thermo-mechanical loading is investigated emphasizing this parameter as an important factor for finishing cutting operations
Optimization and Multi-Model Approaches for Maximizing Ship Decarbonization
No full textInternational audienceAligned with the International Maritime Organization's (IMO) aim to lower greenhouse gas emissions, the use of Wind-Assisted Ship Propulsion (WASP) in maritime transport is gaining increased attention. The study and optimization of WASP remain challenging due to complex physics and the high number of parameters involved. This includes interactions between aerodynamics, hydrodynamics, and structural dynamics, which require advanced multi-physics modeling. Additionally, numerous factors must be optimized, from design and control parameters to route selection, all within the variable maritime environment. Addressing these challenges demands multi-model approaches and multi-criteria optimization methods to maximize energy efficiency effectively. In this context, the laboratories of Ecole Navale, IFREMER, ENSM, and ENSTA Bretagne are collaborating on several research projects, with a selection of two thematic studies presented here.The first study is carried out as part of the SHIVA and SAWASP projects, jointly led by Ecole Navale, IFREMER and ENSTA Bretagne. The goal of these projects is to optimize the hydrodynamic performance of innovative, fully electric, Vertical-Axis Propellers (VAPs) and their optimal use in conjunction with a wind-assisted ships. To achieve this, the SHIVA project implements a multi-criteria optimization of the propeller blade-pitching laws, utilizing multi-fidelity numerical and experimental surrogate models. These optimizations enable the determination of a set of optimal pitch laws for different operating points of the propeller. In the SAWASP project, a 6-meter wind-assisted ship equipped with VAPs is developed to study the optimal aerodynamic-hydrodynamic coupling. In particular, the energy gains from using VAPs as the main propulsion system, generating a lateral anti-drift force, are studied. The use of Reinforcement Learning (RL) methods to maintain optimal ship operation performance at sea, as an uncertain environment, is also part of this project.The second study is conducted within the framework of the SOMOS project, jointly managed by ENSM and ENSTA Bretagne. The project's goal is to create and validate a set of numerical tools, that allow for rapid and precise assessment of the energy efficiency of wind-assisted ships. For the purposes of this study, a modular and comprehensive ship motion solver is formulated as an optimal control optimization problem, to evaluate, compare, and optimize energy performance. Such an approach is very complex to implement and, depending on the fidelity-level used, may require very high modeling costs. This is why most research efforts focus on specific aspects of the broader problem, often overlooking the coupling of maritime routing, ship motion analysis, and the optimization of control parameters along the planned sea route. The present work provides an innovative approach for the calculation of optimized trajectories for wind-assisted ship, by both considering the ship's maneuvering capabilities and the optimization of ship control and/or design parameters. In contrast to conventional routing methods, the proposed approach achieves high computational efficiency and relies on direct multiple shooting methods to determine optimal ship control parameters (RPM, rudder angle, etc.) and design variables (rotor Flettner or sail positioning, rudder area, etc.) along the sea route, satisfying the constraints of complying with the ship's equations of motion
Techniques d’apprentissage pour l’optimisation en Ingénierie : outils dirigés par les données pour applications mécaniques et biomédicales
No full textThe present doctoral thesis encompasses a series of case studies in various fields of Engineering and Medicine, whose common points are their conceptual and computational complexity and thus compromise the availability and reliability of experimental data supporting their corresponding constituent models, if existent at all. Often times, such models based on analytic expressions suffer from computationally unreachable complexity ("curse of dimensionality") and/or are burdened by biases and vagueness induced in their definition (e.g. non-observable internal variables).There are various proven and widely used methodological solutions allowing the avoidance or reduction of the effect of such problems by building models based mostly on observed data, as opposed to preconceived expressions, yielding the so-called surrogate models. These techniques include Model Order Reduction (MOR) and Artificial Intelligence (AI), or more specifically Machine Learning (ML). The latter consists in redefining the complete model as a version with somewhat lower resolution but reduced computational cost, for which methods such as Proper Orthognal/Generalized Decomposition (POD/PGD) are used. Machine Learning, on the other hand, is based on statistical regressions built upon large databases labeled as "inputs" and "outputs", whose underlying behavior is adjusted in a phase of "training" and evaluated in a subsequent "essay" with new data - i.e. not previously processed. This methodology is commonly enforced by means of neural networks, based (with limitations) on the biological ones in the brain and offering different architectures according to the type of data to be processed: multi-layer perceptron (MLP) for vectors and matrices - locally condensed by convolutions if the database is very large - convolutional networks for images and text (CNN), recurrent (RNN) for time sequences - with long short-term memory (LSTM) if there is historical dependency, or based on graphs (GNN).Graphs as mathematical objects offer many possibilities for the compact representation of data with states as vertices and their relationships as segments. This analogy is directly applicable to the two main case studies tackled in this work, namely mechanical structures (focusing on metamaterials) and biological neuronal networks, being possible in both cases to assimilate nodes/neurons as vertices and bars/axons as segments of a graph. Bearing this in mind, the various needs and challenges of these research fields (complexity, non-linearity, reduced observability, multi-scale and multi-objective optimization, etc.) will be studied using the tools described above, and making new contributions should the current state of the art be limited. To this end, a number of case studies have been devised for various industrial and/or medical applications, and the results of the proposed methodologies upon which have been successfully compared with similar work. Finally, the limitations of the proposed surrogate models have been analysed and some improvements and extensions have been suggested for future developments, with a view to create accurate and efficient replicas for the modelled phenomena (Digital Twins).La présente thèse de doctorat englobe une série de casuistiques dans divers domaines de l’ingénierie et de la médecine dont les points communs sont leur complexité conceptuelle et numérique. Cela compromet à son tour la disponibilité et la fiabilité des données expérimentales qui avalisent leurs modèles constitutifs correspondants, si existants. Souvent, ces modèles basés sur des expressions analytiques souffrent d’une complexité inabordable pour l'ordinateur ("malédiction de la dimensionnalité") et/ou sont alourdis par des biais induits dans leur définition (e.g. variables internes non observables).Il y a plusieurs solutions méthodologiques éprouvées et largement utilisées qui permettent de éviter ou d’atténuer l’effet de ces problèmes en construisant des modèles basés principalement sur des données observées plutôt que sur des expressions préconçues, ce qui donne naissance à des modèles dits substitutifs. Parmi ces techniques, on peut citer la réduction de l’ordre des modèles (MOR) et l’intelligence artificielle (IA), ou plus précisément l'apprentissage automatique (ML). La première consiste à redéfinir le modèle complet comme une version de résolution légèrement inférieure mais coût de calcul réduit, pour lequel on utilise des méthodes telles que la Décomposition Orthogonale ou Générale en valeurs propres (POD/PGD). En revanche, l’apprentissage automatique est basé sur des régressions statistiques provenants de grandes bases de données étiquetées comme "entrées" et "sorties", dont le comportement sous-jacent est ajusté dans une phase de "formation" et évalué dans un "essai" subséquent avec de nouvelles données non traitées précédemment. Cette méthodologie se manifeste principalement par des réseaux neuronaux inspirés (avec limitations) par les réseaux cérébraux, offrant différentes architectures selon le type de données traitées: perceptron multi-couches (MLP) pour vecteurs et matrices - condensées localement par des convolutions si la base de données est très grande - réseaux convolutifs pour les images et le texte (CNN), récurrents (RNN) pour les séquences temporelles - avec mémoire à court terme étendue (LSTM) s’il y a dépendance historique, ou basés sur des graphes (GNN).En tant qu’objet mathématique, les graphes offrent de multiples possibilités pour la représentation compacte de données avec des états comme vertex et leurs interrelations comme segments. Cette analogie s’applique directement aux deux principaux cas d’étude qui concernent ce travail, à savoir les structures mécaniques (centré sur les métamatériaux) et les réseaux neuronaux biologiques, pouvant dans les deux cas assimiler des nœuds/neurones comme vertex et des poutres/axons comme segments d’un graphe. Dans cet esprit, les divers besoins et défis de ces domaines de recherche (complexité, non-linéarité, observabilité réduite, optimisation multi-échelle et multi-objectif, etc.) seront étudiés au moyen des outils décrits ci-dessus, en formulant de nouvelles contributions lorsque l’état actuel de la technique est limité. A cet effet, plusieurs études pratiques ont été conçues pour diverses applications industrielles et/ou médicales et les résultats des méthodologies proposées ont été comparés avec succès à des travaux similaires. Enfin, les limites des modèles substitutifs proposés ont été analysées et quelques points d’amélioration et d’extension suggérés pour de futurs développements, en vue de la création de répliques fidèles et efficaces des phénomènes modélisés (Jumeaux Numériques)
Imagerie ultrasonore 3D par rayonnement de plaque mince appliquée à l'interaction homme-machine sans contact
No full textThe wide range of digital technology applications and the diverse contexts in which they are used drive the development of integrated multimodal interaction devices, enabling both tactile and contactless interactions with screens. This thesis contributes to this field through the development of a sensor for contactless gestural interaction based on the acoustic radiation of thin plates such as display screens. This radiation, controlled by a limited number of piezoelectric transducers coupled to the plate, is reflected by objects in the volume above and captured by microphones set in the plane of the plate. Using a 3D imaging technique based on unfocused ultrasound emission combined with tensor decomposition using HOSVD, this work achieved real-time imaging (33 Hz) of the volume (160 x 110 x 80 mm3) above the plate, with a spatial resolution of 1 mm. These advancements enabled first applications of detection and tracking of finger movements in air. Additionally, the hand movements generate a Doppler shift, which was utilized to identify 11 distinct 3D gestures performed by 19 participants. A low-complexity recurrent neural network was used to detect and classify these contactless gestures with no noticeable latency. In addition to contactless interaction, the methods developed in this thesis for fast imaging with a small number of transducers and Doppler pattern recognition open up new applications in medical imaging and non-destructive testing.La multiplicité des usages du numérique et la variété des contextes d'utilisation poussent au développement de dispositifs d'interactions multimodaux intégrés permettant une interaction tactile tangible ou sans-contact avec des écrans. Dans ce contexte, ces travaux de thèse ont porté sur le développement d'un capteur pour l'interaction gestuelle sans contact reposant sur le rayonnement acoustique de plaques minces, telles que des écrans. Ce rayonnement, contrôlé par un faible nombre de transducteurs piézoélectriques couplés à la plaque, est réfléchi par les objets du volume situés au-dessus et capté par des microphones placés dans le plan de la plaque. En s'appuyant sur une méthode d'imagerie 3D par émission ultrasonore non focalisée combinée à une décomposition tensorielle par HOSVD, ces travaux ont permis de produire une image en temps réel (33 Hz) du volume (160 x 110 x 80 mm3) au-dessus de la plaque avec une résolution spatiale de 1 mm. Ces résultats permettent de premières applications de détection et de suivi de déplacement de doigts dans l'air. Le mouvement des mains produit également un décalage Doppler qui a été exploité pour reconnaître 11 gestes 3D réalisés au-dessus de la plaque par 19 participants. Un réseau de neurones récurrent de faible complexité permet leur détection et classification sans latence perceptible. Au-delà de l'interaction sans contact, les méthodes développées dans cette thèse d'imagerie rapide à faible nombre de transducteurs et de reconnaissance de motifs Doppler ouvrent des perspectives d'applications en imagerie médicale ou contrôle non destructif