HAL Portal IOGS (nstitut d'Optique Graduate School)
Not a member yet
12589 research outputs found
Sort by
Coherent beam combining strategies for highenergy and high-repetition rate lasers dedicated to inertial nuclear fusion applications
No full textInternational audienceSpatial Coherent Beam Combining (CBC) strategies are particularly promising for lasers dedicated to nuclear fusion applications in the purpose of developing high-energy and high-repetition-rate lasers. Let us examine their potential application to high-energy, largesize beams with repetition rates that preclude phase correlation between subsequent pulses. Both passive and active CBC approaches are investigated in this study to identify the most appropriate and effective strategy for large-aperture, high-energy laser amplifiers. Specific experiments have been conducted to validate promising architectures for lasers dedicated to inertial confinement fusion (ICF) applications. Based on these experimental results, examples of suitable and well-adapted setups are propose
Nanoscopic Mapping of the Extracellular Space in Amyloid Plaque‐rich Cortex
No full textInternational audienceA hallmark of Alzheimer's disease (AD) is the accumulation of amyloid plaques, primarily composed of misfolded amyloid β (Aβ) peptides. Complementary high-resolution imaging techniques are employed to investigate the plaque penetrability and the extracellular space (ECS) rheology in a mouse model of AD. Two-photon shadow imaging in vivo confirms that a dense ring of cells surrounds cortical amyloid plaques but highlights the diffusional penetrability of the amyloid core. Quantum dot tracking unveils that ECS diffusional parameters are heterogeneous in and around plaques, with an elevated diffusivity within and around plaques compared to wild-type-tissue. The amyloid core shows low nanoparticle density, varying by plaque phenotype. Carbon nanotube tracking confirms these altered local rheological properties at the level of the whole cortex of AD mice. Finally, the extracellular matrix is found to be dysregulated within the amyloid plaque, which may account for the observed alterations in diffusivity. This study provides fresh insights for understanding Aβ plaque penetration, a prerequisite for therapeutic developmen
Distillation fédérée en une seule étape à l'aide d'enseignants monoclasses : étude sur la fragmentation des connaissances et la supervision hors distribution
No full textInternational audienceThe performance of machine learning models critically depends on the quality and diversity of training data. However, privacy, legal, and proprietary concerns often limit direct data sharing. Many organizations possess high-quality data for specific classes and may wish to share the knowledge derived from it without revealing the data or engaging in collaborative training. While federated learning (FL) enables distributed model training, it typically assumes mutual benefit, requires repeated communication, and produces a shared global model. Another paradigm, knowledge distillation (KD), allows a student model to learn from teacher predictions. We propose a one-shot federated distillation method in which a single client learns from monoclass teacher models trained independently by multiple providers. Each provider shares its model once, and the client combines these with unlabeled data to distill a multiclass student model-aggregating knowledge from disjoint, class-specific sources. This unidirectional, asymmetric setup poses a key challenge: out-of-distribution (OOD) supervision, where monoclass teachers often mispredict unseen inputs, leading to noisy signals for the student. The main contribution of this work is a systematic study of knowledge fragmentation in one-shot federated distillation with monoclass teachers. We evaluate five configurations with varying class coverage per provider and show that increasing fragmentation intensifies OOD supervision, degrading student performance. Experiments on MNIST, FashionMNIST, and CIFAR-10 confirm that fragmentation consistently reduces student accuracy. To mitigate this, we discuss three strategies: (1) exposing teachers to diverse off-class examples, (2) penalizing overconfidence, and (3) using contrastive learning to sharpen feature boundaries.Les performances des modèles d'apprentissage automatique dépendent fortement de la qualité et de la diversité des données d'entraînement. Cependant, les préoccupations liées à la confidentialité, aux aspects juridiques et à la propriété limitent souvent le partage direct des données. De nombreuses organisations possèdent des données de haute qualité pour des classes spécifiques et peuvent souhaiter partager les connaissances qui en découlent sans révéler les données ni s'engager dans une formation collaborative. Si l'apprentissage fédéré (FL) permet l'entraînement distribué des modèles, il suppose généralement un avantage mutuel, nécessite des communications répétées et produit un modèle global partagé. Un autre paradigme, la distillation des connaissances (KD), permet à un modèle élève d'apprendre à partir des prédictions du modèle enseignant. Nous proposons une méthode de distillation fédérée en une seule étape dans laquelle un seul client apprend à partir de modèles enseignants monoclasses formés indépendamment par plusieurs fournisseurs. Chaque fournisseur partage son modèle une seule fois, et le client les combine avec des données non étiquetées pour distiller un modèle élève multiclasses, en agrégeant les connaissances provenant de sources disjointes et spécifiques à chaque classe. Cette configuration unidirectionnelle et asymétrique pose un défi majeur : la supervision hors distribution (OOD), où les enseignants monoclasses font souvent des prédictions erronées sur des entrées invisibles, ce qui entraîne des signaux bruités pour l'élève. La principale contribution de ce travail est une étude systématique de la fragmentation des connaissances dans la distillation fédérée en une seule fois avec des enseignants monoclasses. Nous évaluons cinq configurations avec une couverture de classe variable par fournisseur et montrons que l'augmentation de la fragmentation intensifie la supervision OOD, dégradant ainsi l'élève
Self‐Assembled Small Interfering RNA‐Gold Supraclusters Detected at the Single‐Molecule Level in the Near‐Infrared‐II Window
No full textInternational audienceGold nanoclusters (AuNCs) possess unique photophysical properties that make them excellent candidates for advanced bioimaging and single‐particle detection. In this article, the self‐assembly of highly emissive, positively charged near‐infrared‐II AuNCs stabilized by cysteamine, directed by small interfering RNA (siRNA), is reported, which serves as both a structural and electrostatic modulator. The resulting supramolecular assemblies exhibit quasi‐spherical morphologies around 100 nm in diameter, with outstanding colloidal stability, photostability, and enzymatic resistance. Their strong photoluminescence, extending up to 1400 nm, enables robust single‐particle detection in solution. Spectroscopic and structural analyses—including fluorescence spectroscopy, small‐angle X‐ray scattering, and single‐particle tracking—highlight the pivotal role of siRNA in tuning the assembly process via charge balance and concentration‐dependent interactions. Beyond providing insights into the structural and photophysical behavior of nucleic acid‐guided AuNC assemblies, these results underscore their promise as multifunctional nanoplatforms for integrated imaging and gene‐silencing therapies in biophotonic and theranostic applications
Miroirs interferentiels multicouches pour les impulsions attosecondes
No full textInternational audienceLes impulsions attosecondes, découvertes il y a une vingtaine d’années et célébrées par les prix Nobel de physique 2023, ont suscité le développement de composants optiques spécifiques. Le spectre de ces impulsions ultrabrèves se situe dans l’extrême ultraviolet (EUV), domaine spectral où les revêtements interférentiels multicouches sont indispensables pour réfléchir efficacement la lumière. En optimisant la structure de ces empilements, il est possible de transporter et même de compresser temporellement ces impulsions ultrabrèves
CryoRhodopsins: A comprehensive characterization of a group of microbial rhodopsins from cold environments
No full textInternational audienceMicrobial rhodopsins are omnipresent on Earth; however, the vast majority of them remain uncharacterized. Here, we describe a rhodopsin group found in microorganisms from cold environments, such as glaciers, denoted as CryoRhodopsins (CryoRs). A distinguishing feature of the group is the presence of a buried arginine residue close to the cytoplasmic face. Combining single-particle cryo–electron microscopy and x-ray crystallography with rhodopsin activation by light, we demonstrate that the arginine stabilizes an ultraviolet (UV)–absorbing intermediate of an extremely slow CryoRhodopsin photocycle. Together with extensive spectroscopic characterization, our investigations on CryoR1 and CryoR2 proteins reveal mechanisms of photoswitching in the identified group. Our data suggest that CryoRs are sensors for UV irradiation and are also capable of inward proton translocation modulated by UV light
Side-Channel Extraction of Dataflow AI Accelerator Hardware Parameters
No full textInternational audienceDataflow neural network accelerators efficiently process AI tasks on FPGAs, with deployment simplified by readyto-use frameworks and pre-trained models. However, this convenience makes them vulnerable to malicious actors seeking to reverse engineer valuable Intellectual Property (IP) through Side-Channel Attacks (SCA). This paper proposes a methodology to recover the hardware configuration of dataflow accelerators generated with the FINN framework. Through unsupervised dimensionality reduction, we reduce the computational overhead compared to the state-of-the-art, enabling lightweight classifiers to recover both folding and quantization parameters. We demonstrate an attack phase requiring only 337 ms to recover the hardware parameters with an accuracy of more than 95% and 421 ms to fully recover these parameters with an averaging of 4 traces for a FINN-based accelerator running a CNN, both using a random forest classifier on side-channel traces, even with the accelerator dataflow fully loaded. This approach offers a more realistic attack scenario than existing methods, and compared to SoA attacks based on tsfresh, our method requires 940× and 110× less time for preparation and attack phases, respectively, and gives better results even without averaging traces.</div
Étude des transitions de phase non-équilibres dans le silicium sous excitation laser femtoseconde
No full textInternational audienceNon-equilibrium phase transitions in silicon induced by ultrashort laser pulses are essential for advanced applications in microelectronics and optoelectronics. This work presents a numerical modeling of laser-induced phase transitions in silicon using a hybrid atomistic-continuum model. By combining molecular dynamics (MD) simulations and a modified continuum model (nTTM), we study melting, ablations, and amorphization phenomena under extreme fluence and pressure conditions. Our results show that free-carrier temperature and thermal energy play a crucial role in determining the melting depth and phase transition mechanisms, highlighting significant differences between heterogeneous and homogeneous melting mechanisms. In particular, we observe a marked dependence of the melting behavior on the silicon crystal orientation, with <111> orientations showing faster melting than <001> orientations. Furthermore, the hybrid MD-nTTM model allows for better prediction of melting depth and amorphization dynamics compared to conventional continuum models, such as the nTTM model, and provides valuable information for the control of surface structuring processes and information storage applications in silicon-based materials. These results pave the way for the optimization of laser-based manufacturing processes for high-precision applications in the field of microelectronics.Les transitions de phase non-équilibres dans le silicium induites par des impulsions laser ultracourtes sont essentielles pour des applications avancées en microélectronique et optoélectronique. Ce travail présente une modélisation numérique des transitions de phase induites par laser dans le silicium à l'aide d'un modèle hybride atomistique-continuum. En combinant des simulations de dynamique moléculaire (MD) et un modèle continu modifié (nTTM), nous étudions les phénomènes de fusion, d’ablations et d’amorphisation dans des conditions extrêmes de fluence et de pression. Nos résultats montrent que la température des porteurs libres et l’énergie thermique jouent un rôle crucial dans la détermination de la profondeur de fusion et des mécanismes de transition de phase, mettant en évidence des différences significatives entre les mécanismes de fusion hétérogène et homogène. En particulier, nous observons une dépendance marquée du comportement de fusion à l’orientation cristalline du silicium, avec des orientations <111> montrant une fusion plus rapide que les orientations <001>. De plus, le modèle hybride MD-nTTM permet une meilleure prédiction de la profondeur de fusion et des dynamiques d'amorphisation par rapport aux modèles continus classiques, comme le modèle nTTM, et fournit des informations précieuses pour le contrôle des processus de structuration de surface et des applications de stockage d’informations dans les matériaux à base de silicium. Ces résultats ouvrent la voie à l’optimisation des procédés de fabrication basés sur les lasers pour des applications de haute précision dans le domaine de la microélectronique
