HAL-CentraleSupelec
Not a member yet
77624 research outputs found
Sort by
Optimal transport for statistical matching
No full textMotivation: Statistical matching integrates multiple datasets on the same population, sharing some variables but each containing unique ones. The goal is to build a unified synthetic dataset where all variables are jointly available, improving the reliability of subsequent analyses.Results: We introduce two optimal transport-based methods that align the distributions of shared and distinct variables, either within each data source or between sources, while simultaneously estimating a predictive function to impute missing values. The methods are supported by theoretical guaranties, including error of generalisation bounds. A Monte Carlo study compares their performance against other OT approaches across scenarios. Finally, the methods are applied to real-world cohort data, demonstrating their practical interest in addressing common challenges in public health research
Linking and Combining Digital Music Sources in Digital Archives: The CollabScore approach
No full textInternational audienceLibraries and archives that propose access to music resources can expose a wide variety of heterogeneous digital representations: texts and textbooks, recordings, score sheets, video and audio. Each representation, although related to the same intellectual content, addresses a specific communication purpose, serves quite distinct motivation (listening, analyzing, performing, to name a few) and proposes to the public a quite specific interaction. The downside of this diversity is obviously the lack of correspondence between these documents at a fine level of granularity. It would highly be desirable for instance to combine, in a same presentation, a performance, the corresponding music notation, a scan of a manuscript and possibly textual annotations. Such artifacts can greatly enhance the user experience and valorize the digital contents.Producing such combined documents is currently hindered by several obstacles. First, we need to manually annotate images, audio and textual fragments. Second, these annotations must be associated in a complex structure. Finally, a player apt at exposing this rich information is required. This results in a costly and tedious process.We will present a solution designed by the French research project CollabScore. It consists in a workflow that (i) recognizes score images, (ii) produces music notation, (iii) aligns recorded performances on this notation, and finally (iv) creates a compliant IIIIF structure that glues together these resources. A standard IIIF player can then be used to display the interlinked resources. The design and methodology of each step will be detailed, and the result will be demonstrated on digital libraries managed by the project partners: BnF and Fondation Royaumon
Influence of Ageing on Mechanical, Thermal, and Physical Properties of Concrete with Recycled Polystyrene Granules
No full textInternational audienc
Energy-Efficient Right-Sizing of Kafka-like Message Brokers for IoT Workloads
No full textInternational audienceIoT data pipelines rely on message brokers, such as Apache Kafka and Redpanda for continuous telemetry ingestion. When it comes to capacity planning of these systems, the absence of clear sizing guidance often leads to conservative over-provisioning and unnecessary energy use. We present a calibration-based methodology for energy-efficient right-sizing of Kafka-compatible clusters for IoT ingest. Using a small set of initial experiments on 3-4 nodes, we fit a performance model that predicts maximum sustainable throughput and per-node power, enabling operators to choose the smallest cluster that satisfies a target ingest rate with headroom while minimizing energy consumption.We substantiate the approach with an experimental study of Kafka and Redpanda across three hardware generations (HDD, SATA SSD, NVMe), varying partition counts, node counts, and resource limits. We find that storage technology is the primary determinant of throughput, horizontal scaling is near-linear, and vertical CPU scaling yields diminishing returns; the two brokers exhibit distinct energy proportionality properties. On previously unseen hardware, the model predicts throughput and power with median errors under 10% and 7%, respectively.Our results provide a practical, reproducible capacity-planning workflow that maps IoT workload requirements (message size and rate) to concrete, energy-aware deployment decisions.</p
Transition 1point5, the platform for French academic research laboratories engaged in their environmental transition
No full textThis article presents an analysis, quantitative and qualitative, of the actions shared by French laboratories on the Transition 1point5 platform. This is an open-source web application, developed by the French research group Labos 1point5, that allows volunteer French laboratories to share the measures -the actions-they have taken to reduce their environmental footprint. We present the analysis of the 274 actions shared by 69 laboratories on the platform on 11 June 2025, addressing the following questions. What measures are research laboratories implementing to adopt sustainable practices? Is there a single transition path that applies to all laboratories? What role do awareness-raising initiatives play in this process? Can participatory workshops add value to the transition process? Have laboratories undergoing transition effectively reduced their GHG (greenhouse gases) emissions?The main result of this study is that, rather than a single transition pathway, the platform's laboratories employ various approaches depending on their specific characteristics. These approaches share common features, such as the use of awareness-raising activities, the creation of working groups dedicated to environmental issues, and increasingly participatory workshops to discuss and implement transition actions. Another key finding is that it is too early to measure the effectiveness of the implemented measures in terms laboratories' GHG (greenhouse gas) reductions, since most of the laboratories only implemented their actions in 2024 or early 2025.Nevertheless, the platform provides valuable insights into how research laboratories are engaging in the transition, revealing organizational processes and demonstrating that another way to do research is explored.Cet article présente une analyse quantitative et qualitative des actions partagées par les laboratoires français sur la plateforme Transition 1point5. Il s'agit d'une application web open source, développée par le GDR Labos 1point5, qui permet aux laboratoires français volontaires de partager les mesures - les actions- qu'ils ont prises pour réduire leur empreinte environnementale. Nous présentons l'analyse des 274 actions partagées par 69 laboratoires sur la plateforme au 11 juin 2025, en abordant les questions suivantes. Quelles mesures les laboratoires de recherche mettent-ils en œuvre pour adopter des pratiques durables ? Existe-t-il une trajectoire de transition unique qui s'applique à tous les laboratoires ? Quel rôle jouent les initiatives de sensibilisation dans ce processus ? Les ateliers participatifs peuvent-ils apporter une valeur ajoutée au processus de transition ? Les laboratoires en transition ont-ils effectivement réduit leurs émissions de GES (gaz à effet de serre) ?Le principal résultat de cette étude est que, plutôt qu'une trajectoire de transition unique, les laboratoires de la plateforme emploient diverses approches en fonction de leurs caractéristiques spécifiques. Ces approches partagent des caractéristiques communes, telles que le recours à des activités de sensibilisation, la création de groupes de travail dédiés aux questions environnementales et des ateliers participatifs pour discuter et mettre en œuvre des actions de transition.Une autre conclusion importante est qu'il est trop tôt pour mesurer l'efficacité des mesures mises en œuvre en termes de réduction des émissions de gaz à effet de serre (GES) des laboratoires, car la plupart d'entre eux n'ont mis en œuvre leurs actions qu'en 2024 ou au début de 2025.Néanmoins, la plateforme fournit des informations importantes sur la manière dont les laboratoires de recherche s'engagent dans la transition, en révélant les processus organisationnels et en démontrant qu'une autre façon de mener des recherches est explorée
Output Feedback Regulation of Heat Equations
No full textInternational audienceThis paper presents an output feedback control strategy for the setpoint regulation control of heat equations. Inspired by classical finite-dimensional methods, the adopted approach leverages an integral component of the tracking error. The reported approach relies on spectral reduction methods, directly applied to the heat equation once augmented with the integral component. It is shown that the adopted boundary control design strategy can always be successfully implemented for the setpoint regulation control of a either Dirichlet or Neumann trace despite the presence of an unknown boundary disturbance
Embedding de Graphes de Connaissances Biomédicaux à base de GNN et de modèles LLM biomoléculaires pour la découverte de médicaments
No full textInternational audienceWe are interested in learning vectorial embeddings for entities of a biomedical knowledge graph (KG) to predict new links between drugs, diseases, or therapeutic targets. We chose toleverage the PrimeKG graph, constructed to enhance drug discovery. Our first objective is to measure the contribution of a GNN-type embedding model (RGCN) compared to geometric models. The next question is whether it is possible to improve the quality of the learned representation and the predicted links by using a GNN model but limiting ourselves to unbiased characteristics through type hot-encoding or representations derived from pre-trained models on large volumes of biological sequences called protein LLMs and chemical LLMs. We will present the work carried out and the initial results obtained on PrimeKG.ous nous intéressons à l'apprentissage de représentations (embeddings) pour les entités d'un graphe de connaissances (GC) biomédical afin de prédire des nouveaux liens entre médicaments et maladies ou cibles thérapeutiques (protéines responsables de maladies). Nous avons choisi d'exploiter le graphe PrimeKG construit pour favoriser la recherche de médicaments notamment pour les maladies rares. Notre premier objectif est de mesurer l'apport d'un modèle d'embedding de type GNN (RGCN) par rapport aux modèles géométriques. Le second, dans un souci de frugalité, est la possibilité d'améliorer la qualité de la représentation apprise et des liens prédits en utilisant ce modèle GNN mais en nous limitant à des caractéristiques peu biaisées à travers un encodage des types ou de représentations issues de modèles pré-entraînés sur des grands volumes de séquences biologiques appelés protein LLMs et chemical LLMs. Nous présentons les travaux réalisés et les premiers résultats obtenus sur PrimeKG
Bipyrimidine powder for potential applications in integrated photonics on polymer materials for the creation of a broad spectrum
No full textInternational audienceOrganic materials for nonlinear optics have created a wide range of applications in the field of photonics, optics and in the area of biological science. A series of bipyrimidine basedchromophores are prepared with alkoxy styryl groups and connected with electron accepting 2,2’-bipyrimidine cores. Their linear and nonlinear optical properties are investigated throughvarious techniques such as light transmission measurements, optical measurements and as well as the size properties are characterized through scanning electron microscopy. So, from all thecharacteristics thus bipyrimidine for nonlinear optics confirms the size, behavior and nano size range between 10-100 nm. Bipyrimidine seems too attractive because the electron affinity of thepyrimidine rings is much higher than that of the pyridine ring and this property has been rooted for designing the pyrimidine contains conjugated systems with an interesting way of motion ofcomplementary colors for luminescence and two photon absorption properties. Orange crystals are suitable for the X-ray diffraction analysis for the crystalline nature and detect various phasesand helps to know about the good planarity of conjugated systems, so it is applicable for to mix with polymers and also help in the optoelectronic applications such as the sensors and microresonators as of in the area of integrated photonics and for creation of broad band spectrum emission for the non-linear optics
Revisiting SIF abstraction rules with SPARQL for querying BioPAX
No full textInternational audienceBioPAX (for Biological PAthway eXchange) is a standard Semantic Web format based on OWL(Ontology Web Language) designed for the representation of metabolic, signaling pathways, molecularand genetic interactions and gene regulation networks [1]. The BioPAX ontology provides classes andproperties allowing to finely describe biological information using a controlled vocabulary. However,the counterpart of its high expressivity is an important complexity of the resulting BioPAX graphs,that hinders downstream analyses and reasoning tasks [2]. Overall, there is a lack of easy-to-use toolsto work with BioPAX models, making its accessibility limited to a range of specialist users. There isthus an important need to exploit BioPAX’ richness while accommodating its complexity. Abstractionmethods on BioPAX models partially address this need by extracting contextual information from theknowledge graph while getting rid of part of its complexity.Here, we focus on the abstraction rules for the conversion of BioPAX to SIF (Simple InteractionFormat), a binary interaction file format allowing to represent interactions involving proteins andchemicals. The SIF format was introduced in the early 2000s for describing interactions in a biologicalnetwork using the Cytoscape network visualization software [3]. SIF interactions are presented intabular format with three columns: a first column for the source entity, a middle column for theinteraction type, and a third column for the destination entity. This simplicity of representationmakes the format easy-to-use, understandable by end-users and easily importable into a wide range ofnetwork visualization and analysis softwares. As for the tools available to convert BioPAX to SIF, thereare few easy-to-use tools for working with BioPAX, and even fewer providing abstraction methods. ThePaxtools Java library [4] is the only complete BioPAX API developed to date, capable of handlingBioPAX’ high intrinsic complexity and performing multiple tasks on BioPAX files (manipulation,validation and conversion). Paxtools, and its derivative visualization tool ChiBE [5], are the onlytools able to abstract BioPAX models to SIF. The abstraction is done using inference rules that wewill call SIF abstraction rules. These rules were defined by the SIF development team and are availableon the PathwayCommons [6] website 5. SIF abstraction rules describe 14 biological configurations, andsummarize them in protein-protein, protein-chemical and chemical-chemical interactions covered bythe SIF format. Technically, SIF abstraction rules are modeled in the Paxtools code with Java graphpatterns. A pattern is a fixed number of BioPAX elements that satisfy a list of constraints [7]. However,the 14 SIF abstraction rules and their corresponding Paxtools patterns are ambiguously documented,potentially leading to errors or misunderstandings of the abstraction results.In this work we investigated how SIF rules are described on PathwayCommons and in the Pax-tools code. We highlighted that SIF abstraction rules are described at different levels, revealing threepossible meanings of description for each one of the 14 rules: a diagram on the PathwayCommonswebsite associated to a short textual description (meaning 1), a textual description in the Paxtoolscode (meaning 2) and the code of the Paxtools pattern modeling individual SIF rules (meaning 3).We explored the formalization of SIF abstraction rules through a detailed comparison of the threeidentified meanings of description of the rules, and highlighted the different constraints found in eachmeaning. The result highlights an important gap between the description of SIF rules presented to theuser, and what SIF interactions are truly extracted by the Paxtools patterns in the BioPAX graph.We also investigated the potential errors generated by insufficient description of abstraction rules. Weillustrated and quantified them with a biological example, by generating the three distinct SIF abstrac-tions of the BioPAX Kidney function pathway from PathBank [8] obtained with the three meaningsof description previously identified. Overall, SIF abstraction rules descriptions and documentation lack transparency, as evidenced by the three identified meanings of descriptions that give rise to threedifferent types of SIF abstractions with distinct constraints not being explicited to the user, and thatyield different results.Our contribution is to propose a transparent and FAIR BioPAX to SIF abstraction methodby introducing SPARQL queries recapitulating SIF abstraction rules at their three meanings of de-scription. We developed a SPARQL query for each one of the 14 SIF abstraction rules, and for eachone of the three meanings of description, highlighted the different SIF interactions that can be ob-tained with the three meanings. Each SPARQL query is transparent and documented with a cleartextual description accompanied by a diagram of the BioPAX pattern extracted, and the SIF configu-ration obtained. As a case study, we generated using our SPARQL queries the SIF abstractions of theBioPAX exports of two large pathway databases: PathBank (version 2019, from PathwayCommons)and Reactome BioPAX Homo Sapiens (version 94, September 2025). The SPARQL queries scale wellon such large BioPAX models. SPARQL queries and associated Python codes are available as Jupyternotebooks on GitHub at: https://github.com/CecileBeust/BioPAX-To-SIF-SPARQL.gi
Opérateurs potentiels de couches pour problèmes de transmission sur domaines d'extension
No full textInternational audienceWe use the well-posedness of transmission problems on classes of two-sided Sobolev extension domains to give variational definitions for (boundary) layer potential operators and Neumann-Poincaré operators. These classes of domains contain Lipschitz domains, and also domains with fractal boundaries. Although our variational formulation does not involve any measures on the boundary, we recover the classical results in smooth domains by considering the surface measure on the boundary. We discuss properties of these operators and generalize basic results in imaging beyond the Lipschitz case.On utilise le caractère bien posé de problèmes de transmission sur des classes de domaines d'extension de Sobolev bilatéraux pour définir de façon variationnelle les opérateurs potentiels de couches (au bord) et Neumann-Poincaré. Ces classes contiennent des domaines lipschitziens, ainsi que des domaines à bord fractal. Bien que cette formulation variationnelle ne fasse pas intervenir de mesure sur le bord, on retrouve les résultats classiques sur les domaines classiques en considérant la mesure de surface au bord. On discute des propriétés de ces opérateurs et l'on généralise des résultats d'imagerie standards au-delà du cas lipschitzien