HAL-MINES ParisTech
Not a member yet
27348 research outputs found
Sort by
Bridging the gap: Industrial PhD supervisors as hidden architects of collaborative value
No full textInternational audienceThis study investigates the micro-foundations of university-industry collaborations (UIC) through the lens of industrial PhD supervisors in France, focusing on the CIFRE program. Using a mixed-methods approach, we analyze a unique dataset of 24,365 CIFRE agreements and 6,209 supervisor profiles complemented by 23 semi-structured interviews. Our findings reveal that the supervisors’ prior exposure to academia significantly influences their strategies for engaging with industrial PhDs and the types of value they derive. PhD graduates are more likely to specialize in repeat supervision, leveraging their academic networks to explore innovative research avenues and integrate knowledge into R&D departments. Conversely, non-PhDs often engage in supervision as a means to address specific organizational challenges or gain internal legitimacy. Communities of practice and organizational routines play a pivotal role in sustaining the impact of these collaborations beyond individual projects, facilitating knowledge transfer and organizational learning. This work extends the literature on absorptive capacity, knowledge transfer, and boundary spanners by highlighting the diverse pathways and outcomes of industrial supervisors in UICs. It also provides actionable insights for firms seeking to optimize the integration of academic research into their innovation ecosystems
Cumulative Energy Demand and Global Warming Potential of metals and minerals production: Assessment, projections and mitigation options
No full textInternational audienceThis study estimates the cumulative energy demand and global warming potential for 50 metals and 22 non-metallic commodities, incorporating uncertainties through three distinct assessments: low, high, and median, based on the Ecoinvent 3.8 database. The results show that the cumulative energy demand of these commodities represents between 11.3% and 18.2% of global primary energy consumption in 2019, with a median estimate of 13.8%. Similarly, the global warming potential accounts for 13.4%–19.1% of global greenhouse gas (GHG) emissions in the same year, with a median estimate of 17.7%. Iron and steel, aluminum, and cement production emerge as the dominant contributors, responsible for approximately 80% of cumulative energy demand and 90% of GHG emissions across all scenarios. The study also quantifies the potential environmental benefits of enhanced recycling and the accelerated adoption of electric arc furnace technology for steel production in China. These measures can potentially reduce global cumulative energy demand by up to 2.1% and 1.6% respectively and GHG emissions by up to 1.8% and 0.9%. The analysis further examines the energy and climate impacts of mining requirements for a clean energy transition, using IEA's Sustainable Development Scenario. Considering current extraction technologies, the production surplus associated with this scenario could increase global cumulative energy demand by 1.1% and global warming potential by 0.6%. However, these increases would be outweighed by the corresponding reductions in fossil fuel consumption. The paper concludes by discussing the sectors and regions most associated with the environmental burdens of metal and mineral production and highlight key strategies for mitigating these impacts
Dissipative Protection of a GKP Qubit in a High-Impedance Superconducting Circuit Driven by a Microwave Frequency Comb
No full textInternational audienceWe propose a novel approach to generate, protect, and control Gottesman-Kitaev-Preskill (GKP) qubits. It employs a microwave frequency comb parametrically modulating a Josephson circuit to enforce a dissipative dynamics of a high-impedance circuit mode, autonomously stabilizing the finite-energy GKP code. The encoded GKP qubit is robustly protected against all dominant decoherence channels plaguing superconducting circuits but quasiparticle poisoning. In particular, noise from ancillary modes leveraged for dissipation engineering does not propagate at the logical level. In a state-of-the-art experimental setup, we estimate that the encoded qubit lifetime could extend 2 orders of magnitude beyond the break-even point, with substantial margin for improvement through progress in fabrication and control electronics. Qubit initialization, readout, and control via Clifford gates can be performed while maintaining the code stabilization, paving the way toward the assembly of GKP qubits in a fault-tolerant quantum computing architecture. Published by the American Physical Society 202
Interpretable Embeddings for Segmentation-Free Single-Cell Analysis in Multiplex Imaging
No full textMultiplex Imaging (MI) enables the simultaneous visualization of multiple biological markers in separate imaging channels at subcellular resolution, providing valuable insights into cell-type heterogeneity and spatial organization. However, current computational pipelines rely on cell segmentation algorithms, which require laborious fine-tuning and can introduce downstream errors due to inaccurate single-cell representations. We propose a segmentation-free deep learning approach that leverages grouped convolutions to learn interpretable embedded features from each imaging channel, enabling robust cell-type identification without manual feature selection. Validated on an Imaging Mass Cytometry dataset of 1.8 million cells from neuroblastoma patients, our method enables the accurate identification of known cell types, showcasing its scalability and suitability for high-dimensional MI data
Modélisation bi-niveaux pour l'optimisation de la planification de maintenance préventive d'un parc photovoltaïque
No full textInternational audienc
intelligence artificielle haute-performance efficace en énergie : de la mesure et la modélisation à l'ordonnancement multi-objectifs
No full textThe convergence of High-Performance Computing (HPC) and Artificial Intelligence (AI) has led to an era of unprecedented computational demand, driven by the explosive growth of Deep Learning (DL) models. However, the demise of traditional scaling paradigms like Moore's Law and Dennard scaling has extended the challenge in designing next-generation systems from raw performance to energy efficiency. The massive power consumption of modern heterogeneous infrastructures now represents a first-order constraint, thus necessitating a fundamental shift towards sustainable computing.State-of-the-art approaches often lack necessary tools for fine-grained energy measurement across heterogeneous components. To tackle this issue, this dissertation presents an integrated end-to-end methodological framework for the analysis, modeling, and optimization of energy efficiency of AI workloads. Our approach follows a virtuous cycle of "Measure, Understand, Model, and Optimize", for which we have developed an operational suite of novel tools and strategies.Firstly, to enable robust empirical analysis, we introduce a flexible and multi-platform tool for fine-grained energy profiling. Secondly, to move from observation to prediction, we developed a comprehensive analytical framework to model the performance and energy of DL training on GPUs. Finally, to translate energy prediction into active optimization, we present a systematic methodology for the co-design of energy-aware job schedulers.In summary, this thesis provides a set of tools and a structured engineering methodology that transform energy optimization from an ad-hoc practice into a systematic process. We hope through our contributions to have made a significant step towards the design and operation of more intelligent, efficient, and sustainable infrastructures for the future of AI.La convergence du calcul haute performance (HPC) et de l'intelligence artificielle (IA) est marquée par une demande en puissance de calcul sans précédent, portée par la croissance fulgurante des modèles d'apprentissage profond (Deep Learning, DL). Toutefois, le déclin des paradigmes classiques de passage à l'échelle, tels que la loi de Moore ou ou celle de Dennard, a déplacé le défi de la conception des systèmes de nouvelle génération : il ne s'agit plus seulement d'atteindre des performances maximales, mais également de garantir une efficacité énergétique accrue. En effet, la consommation électrique massive des infrastructures hétérogènes modernes constitue désormais une contrainte majeure, imposant une transition vers des approches de calcul plus durables.Les approches actuelles présentent encore des limites, notamment le manque d'outils permettant une mesure fine et détaillée de l'énergie des systèmes. Pour faire face à celà, cette thèse propose un cadre méthodologique intégré, couvrant l'ensemble de la chaîne — analyse, modélisation et optimisation — afin d'améliorer l'efficacité énergétique des applications en IA. Notre démarche repose sur un cycle vertueux articulé autour de quatre étapes : « Mesurer, Comprendre, Modéliser et Optimiser », pour lesquelles nous avons conçu un ensemble cohérent d'outils et de stratégies innovantes.Afin de permettre une analyse empirique robuste, nous introduisons d'abord un outil flexible et multiplateforme permettant un profilage énergétique fin. Ensuite, pour passer de la simple observation à la prédiction, nous développons un cadre analytique complet pour modéliser la performance et la consommation énergétique de l'entraînement de modèles de DL. Enfin, afin de transformer les prédictions en leviers concrets d'optimisation, nous proposons une méthodologie systématique de co-conception de planificateurs de tâches conscients des contraintes énergétiques.En somme, cette thèse met à disposition un ensemble d'outils ainsi qu'une méthodologie structurée qui permettent de faire passer l'optimisation énergétique, d'une démarche ponctuelle et empirique à un processus systématique. Nous espérons ainsi contribuer à franchir une étape significative vers la conception et l'exploitation d'infrastructures d'IA plus intelligentes, plus efficaces et plus durables
Avancées algorithmiques pour l'optimisation contrainte non lisse et non convexe
No full textThis thesis addresses the development of scalable and theoretically grounded algorithms for solving nonsmooth and nonconvex constrained optimization problems, with a focus on models arising in DC programming and stochastic programming. These problems are central to a wide range of applications, including structural mechanics, machine learning, and optimal transport, yet remain challenging due to the lack of smoothness, convexity, and the presence of uncertainty.In the first part, we propose a proximal-type algorithm based on improvement functions and epigraphical nesting. This method is capable of handling general nonsmooth and nonconvex constraints and is shown to converge to what we defined as model-critical points under mild assumptions. The framework unifies and extends several existing approaches, offering a flexible tool for constrained optimization.The second part of the thesis focuses on the computation of constrained Wasserstein barycenters, a problem that arises in distributional clustering and robust optimization. We extend the Method of Averaged Marginals to incorporate convex constraints and propose a progressive decoupling algorithm for the nonconvex case. Theoretical convergence results and numerical experiments demonstrate the effectiveness of the proposed methods.Overall, this work contributes new algorithmic tools and theoretical insights for structure-exploiting optimization in nonsmooth and uncertain environments, and opens several avenues for future research in both theory and applications.Cette thèse porte sur le développement d'algorithmes évolutifs et théoriquement fondés pour résoudre des problèmes d'optimisation contraints, non lisses et non convexes, en particulier ceux issus de la programmation DC et de la programmation stochastique. Ces problèmes sont au cœur de nombreuses applications, telles que la mécanique des structures, l'apprentissage automatique et le transport optimal, mais restent difficiles à traiter en raison de l'absence de régularité, de convexité et de la présence d'incertitude.Dans la première partie, nous proposons un algorithme de type proximal basé sur la fonction d'amélioration et l'imbrication épigraphique. Cette méthode permet de traiter des contraintes générales non lisses et non convexes, et converge vers des points critiques du modèle sous des hypothèses modérées. Le cadre proposé unifie et étend plusieurs approches existantes, offrant un outil flexible pour l'optimisation sous contraintes.La seconde partie de la thèse est consacrée au calcul de barycentres de Wasserstein contraints, un problème rencontré dans la classification de distributions et l'optimisation robuste. Nous étendons la méthode des marges moyennes pour intégrer des contraintes convexes, et proposons un algorithme de découplage progressif pour le cas non convexe. Les résultats théoriques de convergence et les expériences numériques confirment l'efficacité des méthodes proposées.Dans l'ensemble, ce travail apporte de nouveaux outils algorithmiques et des perspectives théoriques pour l'optimisation exploitant la structure dans des environnements non lisses et incertains, et ouvre plusieurs pistes de recherche futures, tant sur le plan théorique que pratique
Absorption of Acid Gases: Evaluation of Thermodynamic and Kinetic Aspects of an Aqueous Solution of 2-(2-Diethylaminoethoxy) Ethanol and 1,3-Dimethyl-2-imidazolidinone
No full textInternational audienceThe removal of CO2 and H2S from a gas by absorption with an aqueous alkanolamine solution is an energy-intensive process. The addition of a physical cosolvent might lower the regeneration energy but also affects the thermodynamics and kinetics of acid gas absorption. In this work, we study the impact on the thermodynamics and kinetics of CO2 and H2S absorption of the addition of 1,3-dimethyl-2-imidazolidinone (DMI) to the aqueous 2-(2-diethylaminoethoxy)ethanol (DEAE-EO) solvent. A comparison is made with the impact of DMI on an aqueous methyldiethanolamine (MDEA) solvent. In all cases, DMI reduced the solubility and the absorption rate of the acid gases. DMI also reduced the initial kinetic H2S/CO2 selectivity, although this effect seems much more pronounced in aqueous MDEA than in aqueous DEAE-EO. The equilibrium thermodynamic H2S/CO2 selectivity is higher in the presence of DMI
Feasibility, conditions, and opportunities for achieving net-negative emissions in the global cement industry
No full textInternational audienceThe cement industry possesses multiple options to decarbonize its operations, including material efficiency, energy efficiency, clinker content reduction, hydrogen utilization, bioenergy, and carbon capture and storage (CCS). By integrating bioenergy and CCS (BECCS), the industry could produce net-negative cement, surpassing the 2050 carbon neutrality pledge of the Global Cement and Concrete Association. In TIAM-FR, a bottom-up optimization model of the global energy system, we developed an explicit model of the global cement industry to analyze the potential contribution of BECCS to producing cleaner cement. We investigated the technical and policy conditions favorable to BECCS deployment and sustainability, considering different future biomass potentials, yields, rotation periods, and management costs. Our findings demonstrate that BECCS can significantly contribute to cement decarbonization, making it easier, quicker, and more cost-effective to achieve. However, the current bioenergy use and policy landscape falls short of meeting the 2050 target. Scaling bioenergy use from 3% of the global energy mix to more than 40%, along with strengthening global climate policies, is essential. By leveraging the potential of bioenergy substitution up to 80% and extensively invest in CCS processes, carbon neutrality in cement production could be advanced by 10 to 18 years, enabling the production of net-negative cement. Finally, we propose a technical roadmap for the decarbonization of the global cement industry
Highly Cited Researchers : anatomy of a list
No full textInternational audienceThe list of Highly Cited Researchers (HCR) published each year by Clarivate occupies a special place in the academic landscape, due to its use in the Shanghai rankings. This article looks at the evolution of this list, based on material communicated between 2001 and 2023 by its various producers (the Institute for Scientific Information, Thomson Reuters and Clarivate) on their respective websites. Three main phases in its trajectory have then been identified. The first is characterised by the creation of a database (2001-2011), the second by the affirmation of an indicator (2012-2018) and the third by the weakening of a strategy (2019-2023). An analysis of this trajectory provides a better understanding of the importance of this list and the challenges it faces today, in a context where some of the key issues of research evaluation and scientific integrity are being called into question