Université de technologie de Troyes open archive
Not a member yet
10722 research outputs found
Sort by
Directional light scattering in Mie-resonant Si particles with ultra-thin Au shells
No full textInternational audienceMetamaterial research has sought to create nanostructures with strong directional optical scattering to control light propagation at the nanoscale. Core–shell architectures comprised of both resonant cores and resonant shells are suggested as candidate particles in which the spectral overlap of the electric and magnetic dipoles is controlled to create strong directional scattering. In this study, Au-decorated Si core–shell (Si@Au) particles are presented, studying the role of the architecture (particulate, discontinuous shells vs continuous) and dimensions of the shell. The core–shell particles are synthesized by first creating Si particles, through the thermal disproportionation of hydrogen silsesquioxane (HSQ), which are then decorated with ≈4 nm diameter Au nanoparticles. The resonant behavior of the core–shell particles is characterized using electron energy-loss spectroscopy mapping and optical single-particle scatter spectroscopy. These observations are supported by T-matrix simulations and Mie-theory calculations of the scattering spectra, which show that, compared to Si, Si@Au particles demonstrate a dampened magnetic dipole resonance for smaller Si core diameters (100–130 nm) and an enhanced magnetic dipole resonance for larger Si core sizes (150–200 nm). The study indicates that the previously reported hybridized modes do not exist in particulate Au shells around a Si core and can only exist in continuous plasmonic shells. Thus, it is shown here how important it is to be as precise as possible regarding the nanomaterial architecture used in simulations. No configuration of Si@Au core–shell particles with a particulate shell could be found that strongly enhanced directional scattering, and a continuous shell may do so only modestly. However, the simulations show that the synthesis of thin, continuous Ag shells might represent an alternative route towards achieving good directional scattering properties.</p
Variational Quantum Brushes
No full textQuantum brushes are computational arts software introduced by Ferreira et al (2025) that leverage quantum behavior to generate novel artistic effects. In this outreach paper, we introduce the mathematical framework and describe the implementation of two quantum brushes based on variational quantum algorithms, Steerable and Chemical. While Steerable uses quantum geometric control theory to merge two works of art, Chemical mimics variational eigensolvers for estimating molecular ground energies to evolve colors on an underlying canvas. The implementation of both brushes is available open-source at https://github.com/moth-quantum/QuantumBrush and is fully compatible with the original quantum brushes
Statistical modeling and likelihood ratio testing for resampling detection in TIFF images
No full textInternational audienc
Quantum Approaches to the Minimum Edge Multiway Cut Problem
No full textInternational audienceWe investigate the minimum edge multiway cut problem, a fundamental task in evaluating the resilience of telecommunication networks. This study benchmarks the problem across three quantum computing paradigms: quantum annealing on a D-Wave quantum processing unit, photonic variational quantum circuits simulated on Quandela s Perceval platform, and IBM s gate-based Quantum Approximate Optimization Algorithm (QAOA). We assess the comparative feasibility of these approaches for early-stage quantum optimization, highlighting trade-offs in circuit constraints, encoding overhead, and scalability. Our findings suggest that quantum annealing currently offers the most scalable performance for this class of problems, while photonic and gate-based approaches remain limited by hardware and simulation depth. These results provide actionable insights for designing quantum workflows targeting combinatorial optimization in telecom security and resilience analysis
Targeted Pooled Latent-Space Steganalysis Applied to Generative Steganography, with a Fix
No full textInternational audienceSteganographic schemes dedicated to generated images modify the seed vector in the latent space to embed a message. Whereas most steganalysis methods attempt to detect the embedding in the image space, this paper proposes to perform steganalysis in the latent space by modeling the statistical distribution of the norm of the latent vector. Specifically, we analyze the practical security of a scheme proposed by Hu et al. for latent diffusion models, which is both robust and practically undetectable when steganalysis is performed on generated images. We show that after embedding, the Stego (latent) vector is distributed on a hypersphere while the Cover vector is i.i.d. Gaussian. By going from the image space to the latent space, we show that it is possible to model the norm of the vector in the latent space under the Cover or Stego hypothesis as Gaussian distributions with different variances. A Likelihood Ratio Test is then derived to perform pooled steganalysis. The impact of the potential knowledge of the prompt and the number of diffusion steps is also studied. Additionally, we show how, by randomly sampling the norm of the latent vector before generation, the initial Stego scheme becomes undetectable in the latent space
Achieving stable LaVO₃ without a reducing atmosphere: Zn/Sb doping toward non-halide perovskite photoconductive devices
No full textInternational audienc
Prévision des fournitures médicales par stratification des types : gestion de la demande régulière et irrégulière à fréquence hebdomadaire et mensuelle.
No full textBackground and Objective: Forecasting the demand for medical supplies is essential for patient safety while also being a major challenge in inventory management. Each product exhibits its own stochastic demand pattern, making manual forecasting infeasible at scale. Forecasting models can assist, but they often struggle with sparse, irregular, or bursty demand, especially in hospital settings. The objective of this study is to develop a type-stratified forecasting framework that specifically addresses lumpy and intermittent medical supply demands, improving their forecasts without contaminating models trained on smoother series.Methods: Medical supply series were first classified using the ADI–CV2 taxonomy into smooth, erratic, intermittent, and lumpy categories. Smooth and erratic series were handled by a multioutput regression pipeline, while intermittent and lumpy series were processed through a twostage Classification + Regression pipeline. The first stage predicts demand occurrence using a CatBoostClassifier, the second estimates magnitude using a RandomForestRegressor, and a recurrent-demand mapping layer adjusts predictions toward historically recurrent magnitudes. The approach was evaluated at both weekly and monthly frequencies, with performance measured via nRMSEI, nRMSEII, ROC–AUC, and the cost-sensitive SPEC metric.Results: The proposed pipeline achieved competitive accuracy across all datasets. On intermittent and lumpy series, it reduced event-conditioned magnitude error (nRMSEI) from 0.72 to 0.67 at weekly frequency and from 1.99 to 1.92 at monthly frequency after applying the recurrent-demand mapping. SPEC analysis showed a consistent reduction in shortage-related costs compared to classical models such as ARIMA, Prophet, or Croston, even when nRMSE values were similar. Weekly Updated evaluation provided slightly better performance than Recursive inference, while monthly data also favored the Updated strategy due to error accumulation over longer horizons.Conclusions: The study confirms that segmenting time series by demand type and explicitly modeling occurrence and magnitude substantially improves operational forecasting for medical supplies. The proposed lightweight machine learning pipeline, though less expressive than deep learning architectures, achieves robust performance with low computational cost, making it suitable for on premise hospital deployment without GPU infrastructure. This trade-off between interpretability, energy efficiency, and predictive accuracy represents a pragmatic step toward sustainable and secure AI applications in healthcare logistics
Quantification de l’incertitude dans le cas des signaux médicaux avec bruit d’annotation
No full textNational audienceAvec la généralisation de l’utilisation de l’apprentissage automatique pour l’analyse enimagerie médicale, les jeux de données et leur annotation par des experts deviennent l’en-jeu central des performances des modèles. Cependant, ces annotations peuvent contenirdes erreurs. Plus spécifiquement dans le cadre médical, la vérité terrain peut être ambiguëdu fait de la complexité et de l’insuffisance de l’image, et est soumise à la subjectivité etl’expérience de l’annotateur. D’autre part, la nécessité de donner au clinicien les cléspour la décision du diagnostic final motivent l’ajout d’une estimation d’incertitude auxrésultats de classification, voire d’une distinction de l’incertitude épistémique (ignorancedu modèle) et aléatoire (difficulté de la donnée, conflit entre classes). Si les im-pacts du bruit d’annotation ont été étudiés en terme de performance, leur effet surl’incertitude des modèles reste inexploré.En particulier, nous montrons l’évolution de l’incertitude des modèles entraînés avec unbruit d’annotation symétrique et proposons une première stratégie pour filtrer les don-nées difficiles dans ce contexte. Nous nous restreignons à la quantification des incertitudespar Monte Carlo dropout, la méthode de quantification d’incertitude majoritaire enimagerie médicale, appliquée à des modèles entraînés sur des jeux de données cor-rompus par du bruit d’annotation symétrique. Les effets sont observés avec deux famillesd’architectures sur trois jeux de données dont deux médicaux. Un réseau convolutionnel(CNN) est utilisé pour les images naturelles MNIST et sur les spectrogrammes de signauxDoppler transcrânien pour la classification des emboles du jeu de données privé HITS.Une architecture de CNN-Transformer permet la classification de l’épilepsie à partir d’unjeu d’électro-encéphalogrammes pour la base Epileptic Seizure Recognition (ESR). Lesmodèles sont entraînés sur des jeux de données soumis à différents taux de bruit d’anno-tation symétrique et évalués sur un jeu de test non bruité.Le bruit symétrique a deux effets sur le modèle appris : ses performances de classificationsont dégradées et les distributions des prédictions en sortie de la softmax tendent à s’uni-formiser. Ce deuxième phénomène est traduit à la fois par l’augmentation del’incertitude aléatoire et la diminution de l’incertitude épistémique. Cela coïncide avec le fait qu’il y a plus de conflits entre classes dans le jeu d’entraî-nement à cause du bruit, et qu’il y a moins de variabilité du modèle par MC dropoutpuisque le modèle tend vers un classifieur complètement incertain. On observe les mêmestendances pour les trois jeux de données, les distributions de ces deux incertitudes donnentdonc une première information sur le taux de bruit. Concernant la diminution des per-formances, les prédictions erronées ont une distribution d’incertitude aléatoire distincteet plus élevée en moyenne des prédictions justes. L’incertitude aléatoire est donc robustepour filtrer les données trop incertaines dans le contexte des jeux de données bruités. Eneffet, la précision augmente significativement lorsque l’on filtre les données par incertitudealéatoire décroissante. Celle-ci est donc un outil fiable pour filtrer les faussesprédictions. Finalement, la combinaison des incertitudes aléatoire et épistémique permetplutôt un diagnostic non-supervisé du modèle, notamment en termes de bruit, et l’incer-titude aléatoire semble la plus adaptée pour l’aide à la prise de décision sur des donnéespotentiellement bruitées
Partage de tournées entre producteurs en circuits courts : une approche bi-objectif conciliant coût et équité
No full textInternational audienc
A Hybrid Machine Learning and NGO Algorithm Approach for Fault Classification and Localization in Electrical Distribution Lines
No full textInternational audienceToday's distribution networks are becoming increasingly complex, necessitating highly accurate and robust fault diagnosis methods. Traditional methods based on impedance or traveling waves often lack flexibility and precision in these dynamic environments. This study proposes a hybrid approach based on the synergy between machine learning (ML) techniques and a recent metaheuristic, the Northern Goshawk Optimizer (NGO). Fault location is performed using a cubic spline interpolation model. Classification is handled by a decision tree, while fault resistance-a key parameter that significantly influences diagnostic performance-is optimized using the NGO algorithm. The effectiveness of the proposed method is evaluated through a series of experiments conducted on the IEEE 34-bus test network. These experiments encompass various fault scenarios (single lineto-ground, line-to-line, double line-to-ground, and three-phase faults) as well as voltage and load variation conditions. Fault resistance values considered in the study are 0, 10, 50 and 100 ohms. The results highlight the robustness and efficiency of the hybrid approach, achieving an accuracy rate of up to 99.999% in fault location. This level of performance enables reliable identification of both the fault location and the affected line.</div