HAL Portal IOGS (nstitut d'Optique Graduate School)
Not a member yet
12589 research outputs found
Sort by
Propriétés optiques de structures auto-organisées induites par laser dans les films plasmoniques nanocomposites
No full textNational audienc
Quasinormal mode as a foundational framework for all electromagnetic Fano resonances
No full textFano profiles are observed across various fields of wave physics. They emerge from interference phenomena and are quantified by the asymmetry parameter q. In optics, q is usually considered as a phenomenological coefficient obtained by fitting experimental or numerical data. In this work, we introduce an ab initio Maxwellian approach using quasinormal modes to analytically describe line shapes in light scattering problems. We show that the response of each individual quasinormal mode inherently exhibits a Fano profile and derive an explicit analytical formula for the Fano parameter. Experimental and numerical validations confirm the formula's accuracy across a broad spectrum of electromagnetic systems. The general expression for q opens new possibilities for fine-tuning and optimizing spectral line shapes in electromagnetism
Atomistic investigation of the oxidation and laser synthesis of nickel nanoparticles
No full textNational audienceNickel nanoparticles (NiNPs) exhibit unique magnetic and chemical properties that make them attractive for applications in catalysis, sensing, and nanotechnology. In this study, reactive molecular dynamics simulations were used to investigate the oxidation behavior of NiNPs in oxygen-rich environments, focusing on the roles of particle size, temperature, and pulsed thermal excitation. The results reveal that oxidation proceeds via a surface-limited mechanism, with reaction rates increasing systematically with nanoparticle diameter and temperature. Simulated pulsed heating, designed to mimic femtosecond laser excitation, significantly enhances oxidation by inducing transient high-temperature conditions that promote irreversible surface reactions. These findings provide fundamental insight into the size-and temperature-dependent oxidation dynamics of NiNPs and underscore the importance of laser-induced thermal histories in controlling their reactivity during laser-based processing or synthesis
zIncubascope: Long-term quantitative imaging of multi-cellular assemblies inside an incubator
No full textInternational audienceRecent advances in bioengineering have made it possible to develop increasingly complex biological systems to recapitulate organ functions as closely as possible in vitro . Monitoring the assembly and growth of multi-cellular aggregates, micro-tissues or organoids and extracting quantitative information is a crucial but challenging task required to decipher the underlying morphogenetic mechanisms. We present here an imaging platform designed to be accommodated inside an incubator which provides high-throughput monitoring of cell assemblies over days and weeks. We exemplify the capabilities of our system by investigating human induced pluripotent stem cells (hiPSCs) enclosed in spherical capsules, hiPSCs in tubular capsules and yeast cells in spherical capsules. Combined with a customized pipeline of image analysis, our solution provides insight into the impact of confinement on the morphogenesis of these self-organized systems
Engineering Deterministic, Tunable, and Reversible Folds in Graphene with the Use of Ultrafast Laser Micro-Patterned Stretchable Polymer Substrate
No full textInternational audienceThe unique atomic monolayer structure of graphene gives rise to a broad range of remarkable mechanical folding properties. However, significant challenges remain in effectively harnessing them in a controllable and scalable manner. In this study, we introduce an innovative approach that employs micron-scale cavities, fabricated through ultrafast laser patterning, in a stretchable polymer substrate to locally modulate adhesion and strain transfer to a graphene monolayer. This technique enables the deterministic induction of single folds in graphene with fold dimensions, width and height in the hundreds of nanometers, tunable through the geometry of the polymer cavities and the applied strain. Importantly, these folds are reversible, returning to a flat morphology with minimal structural damage, as confirmed by Raman spectroscopy. Additionally, our method allows for the creation of fields of folds with reproducible periodicity, defining clear potential for practical applications. These findings pave the way for the development of advanced devices that would leverage the strain and morphology-sensitive properties of graphene
Addressing the Correlation of Stokes-Shifted Photons Emitted from Two Quantum Emitters
No full textInternational audienc
Saturation des réseaux neuronaux récurrents : Expressivité, Apprenabilité et Généralisation
No full textIn the vast class of algorithms that make up Neural Networks, Recurrent Neural Networks (RNNs) have a singular place. On the one hand, their main function is to process sequential data such as sequences, text, audio, video and more. But also because of the ambivalence of their internal operation, resembling both feed-forward Neural Networks in terms of their structure, but exhibiting incomparable behaviour. Therefor the ambition of the ANR project TAUDoS is to establish theoretical facts and algorithms that can give more understanding about the functioning of RNNs.In this thesis we present our theoretical contributions on RNNs. Our contributions span across several fields such as: expressivity, learnibility and generalisation. Expressivity is a global property, in the sense that we talk about the expressivity of a set of functions. Determining the expressivity of a set of RNNs F is the act of finding the most complex class of functions that the RNNs in F can simulate. Learnability is, in a way, expressivity but in a more constrained way, where the aim is to determine what is possible (or not possible) to learn for a given RNN architecture. Guarantees of generalisation have been a topical issue since the introduction of learning algorithms. In very simple terms these guarantees are used to ensure that the model, in our case a RNN, has learned what it is supposed to learn.In the field of expressivity, we started from a recently introduced framework and extended it in order to study a larger portion of RNN architectures. We propose a reformulation of this framework allowing us to import some of the handy properties to a real world scenario. On the one hand, this reformulation enabled us to fill a theoretical gap, but it also paved the way for the other contributions. In the field of learnability our contribution is a theoretical demonstration of the limits of what can be learned by gradient descent for a class of RNNs the Simple Recurent Networks (SRNs). We build on the phenomenon of the vanishing gradient and combine this fact with the finite precision of computers, in order to delimit areas of the parameter space inaccessible for SRNs with current gradient descent algorithms. This demonstration is accompanied by experiments and a proof that the inaccessible areas of parameter space contain functions of interest such as finite-state machines. Finally in the field of generalisation guaranties, we derive a PAC-Bayes bound for SRN that is emph{independent of data length}. We emphasis the fact that the generalisation guarantee is length independent because due to the recursive nature of RNNs, a vast proposition of theoretical results are heavily dependent on the data length. The singular features of our work that set us apart are: 1) all our main results are independent of data length, 2) we have always considered the case where RNNs are defined with non-linear activation functions. Contrary to results where RNNs are assumed to be defined with linear or ReLU functions to simplify the analysis, we have instead incorporated the particularities of bounded functions such as the hyperbolic tangent function and the sigmoidal function. 3) we have incorporated the computer finite precision into our theoretical work.Dans la vaste classe d'algorithmes que composent les Réseaux Neuronaux, les Réseaux Neuronaux Récurrents (RNN acconyme anglais) occupent une place singulière. D'une part, leur fonction principale est de traiter des données séquentielles telles que du texte, de l'audio, de la vidéo, etc. Mais aussi en raison de l'ambivalence de leur fonctionnement interne, ressemblant à la fois aux réseaux neuronaux non récurrent, de par leur structure, mais à la fois présentant des comportements incomparables avec les réseaux non récurrents. Cela motive le projet ANR TAUDoS qui a pour ambition d'établir des faits théoriques et des algorithmes permettant de mieux comprendre le fonctionnement des RNN. Dans cette thèse, nous présentons nos contributions théoriques sur les RNN. Nos contributions couvrent plusieurs domaines tels que : expressivité, apprenabilité et généralisation. L'expressivité est une propriété globale, dans le sens où nous parlons de l'expressivité d'un ensemble de fonctions. Déterminer l'expressivité d'un ensemble de RNN F consiste à trouver la classe de fonctions la plus complexe que les RNN de F peuvent simuler. La capacité d'apprentissage est, en quelque sorte, l'expressivité, mais d'une manière plus contraignante, l'objectif étant de déterminer ce qu'il est possible (ou pas possible) d'apprendre pour une architecture de RNN. Les garanties de généralisation est une question d'actualité depuis l'introduction des algorithmes d'apprentissage. D'une manière très simplifiée ces garanties sont utilisées pour s'assurer que le modèle, dans notre cas un RNN, a appris ce qu'il est supposé apprendre.Dans le domaine de l'expressivité, nous sommes partis d'un cadre récemment introduit et l'avons étendu ce qui nous a permis d'étudier l'expressivité d'architectures de RNN ne pouvant pas être étudiées dans le cadre initiale. Nous proposons une reformulation de ce cadre qui nous permet d'importer certaines des propriétés pratiques dans un scénario du monde réel. D'une part, cette reformulation nous a permis de combler une lacune théorique, mais elle a également ouvert la voie aux autres contributions de cette thèse. Dans le domaine de l'apprenabilité, notre contribution est une démonstration théorique des limites de ce qui peut être appris par descente de gradient pour une classe de RNN, les Réseaux Récurrents Simples (SRN accronyme anglais). Nous nous appuyons sur le phénomène de disparition du gradient et combinons ce fait avec la précision finie des ordinateurs, afin de délimiter des zones de l'espace des paramètres inaccessibles pour les SRNs avec les algorithmes usuels de descente de gradient. Cette démonstration est accompagnée d'expériences et d'une preuve que les zones inaccessibles de l'espace des paramètres contiennent des fonctions d'intérêt telles que les machines à états finis. Enfin, dans le domaine des garanties de généralisation, nous dérivons un PAC-Bayes pour les SRN qui sont emph{indépendant de la longueur des données}. Nous soulignons le fait que la garantie de généralisation sont indépendante de la longueur des données car, en raison de la nature récursive des RNN, une vaste proposition de résultats théoriques dépend fortement de la longueur des données. Les caractéristiques singulières de notre travail qui nous distinguent sont les suivantes : 1) tous nos résultats principaux sont indépendants de la longueur des données, 2) nous avons toujours considéré le cas où les RNN sont définis avec des fonctions d'activation non linéaires. Contrairement a une grande proportion des résultats où les RNN sont supposés être définis avec des fonctions linéaires ou ReLU pour simplifier l'analyse, nous avons au contraire incorporé les particularités des fonctions d'activation bornées telles que la fonction tangente hyperbolique et la fonction sigmoïdale dans nos résultats. 3) nous avons intégré la précision finie de l'ordinateur dans notre travail théorique
Relevance of Human Body Pose Estimation Methods for Complex Dance Movements Analysis
No full textInternational audienceThe relevance and robustness of pose estimation methods is of primary importance in dance movements analysis. The main objective of this keynote paper is to qualitatively compare the most recent pose estimation models for dance movements analysis and discuss their strengths and weaknesses. For this purpose, we developed a specific methodology and tools. The second objective of this paper is to discuss the interest of human body pose estimation for dance movements analyse and to show that the accuracy of body pose estimation is, in the context of dance movements analyse, important but less important than the accuracy of the dance movements modelling. Beyond these objectives, we also discuss the lack of efficient models to describe the kinematics of dance movements
Multi-objective optimization of nanogrids for remote telecom base stations in Canada
No full textInternational audienceThe telecommunications sector targets net-zero emissions by 2050, yet many remote Canadian base stations rely on diesel generators, incurring high costs and emissions. Most hybrid renewable energy system (HRES) studies overlook snow accumulation, limiting relevance in northern climates. This work proposes a snow-aware hybrid nanogrid for a telecom base station in Dorval Lodge, Quebec, using bifacial PV modules, lithium iron phosphate (LFP) batteries, and a diesel generator. A preliminary HOMER Pro study showed 99% renewable penetration is technically possible but at high cost and without snow, bifacial, or aging effects. We developed a high-fidelity model including hourly snow coverage, seasonal albedo, battery aging, and diesel fuel emission behavior. A joint multi-objective optimization minimizing life cycle cost (LCC) and annual CO 2 under LP SP < 0.0001% was solved using a Controlled Elitist NSGA-II algorithm. Three stages were tested: baseline, fixed controls, and monthly adaptive controls. The adaptive strategy achieved the largest gains, cutting CO 2 by 18.59% and LCC by 5.26% versus baseline, with the most sustainable setup using 856 L/year (2.93 t CO 2 ). Sensitivity analysis showed snow-aware designs avoid up to 40.9% higher LCC and 139.7% more CO 2 seen in snowunaware cases. Integrating climate-specific snow modeling with adaptive controls enhances economic and environmental performance, offering a robust, transferable solution for remote telecom power in harsh climates.</div