HAL-MINES ParisTech
Not a member yet
27348 research outputs found
Sort by
The Endless Repetition of Graffiti Writing and Removing: Gestures, Sites, Traces
No full textInternational audienceAlthough graffiti writing has been the object of numerous analyses for fifty years, the entangled practices of graffiti writers and buffers’ interventions remain understudied. By considering their actual gestures in front of urban surfaces, this text emphasizes the similarities of practices usually seen as antagonistic. Drawing on Deleuze’s masterful work about repetition and difference, it notably points out the sensitive, attentional, and technical common features of graffiti writers and buffers, the choreographic dynamics at the core of their performances, and the layered material rhythm resulting from their intertwined actions. Whether they take the form of a soothed dialogue moved by a shared experience, a trench warfare in performances, or an intense unintended collaboration, the relationships between graffiti writers and buffers appear to be driven by the transformative principle of repetition
sCellST predicts single-cell gene expression from H& E images
No full textInternational audienceUnderstanding the spatial organization of individual cell types within tissue and how this organization is disrupted in disease, is a central question in biology and medicine. Hematoxylin and eosin-stained slides are widely available and provide detailed morphological context, while spatial gene expression profiling offers complementary molecular insights, though it remains costly and limited in accessibility. Predicting gene expression directly from histological images is therefore an attractive goal. However, existing approaches typically rely on small image patches, limiting resolution and the ability to capture fine-grained morphological variation. Here, we introduce a deep learning approach that predicts single-cell gene expression from morphology, matching patch-based methods on spot level prediction tasks. The model recovers biologically meaningful expression patterns across two cancer datasets and distinguishes fine cell populations. This approach enables molecular-level interpretation of standard histological slides at scale, offering new opportunities to study tissue organization and cellular diversity in health and disease
De la socialisation professionnelle à la socialisation organisationnelle : le rôle des agents socialisateurs externes
No full textInternational audienceL’intégration des décrocheurs scolaires dans les organisations soulève des défis liés à leur adaptation aux codes professionnels, mais représente aussi une opportunité, notamment dans des secteurs en tension comme le bâtiment. En mobilisant le cadre de la socialisation organisationnelle, cette recherche explore le rôle des agents socialisateurs externes, encore peu étudié. Elle questionne leur intervention dans l’articulation entre socialisation professionnelle et organisationnelle des jeunes décrocheurs. À travers une étude qualitative menée auprès des Compagnons du Devoir, nous montrons que ces agents réduisent l’écart entre les attentes des entreprises et les besoins des jeunes via huit comportements clés regroupés en trois finalités : expliquer les règles, adapter les parcours et permettre l’expérimentation du métier. Cette recherche enrichit la compréhension de la perspective interactionniste relationnelle de la socialisation organisationnelle et propose aux managers des pistes pour gérer l’intégration de ces publics
Dossier "La Société à Mission : Réalités Du Déploiement et Nouvelles Perspectives". Présentation
No full textInternational audienc
Analyse spatio-temporelle expérimentale et numérique de la dynamique de l'effet Portevin-Le Chatelier dans un superalliage à base de nickel
No full textInternational audienceThis study examines the Portevin-Le Chatelier (PLC) effect in the nickel-based superalloy Inconel 718 through a combined experimental-computational approach using statistical indicators from nonlinear dynamical systems theory. We develop a finite element model incorporating the Kubin-Estrin-McCormick constitutive law to capture Dynamic Strain Ageing effects, accounting for machine stiffness influence, and reproduce various dynamics under both hard and soft loading conditions. Statistical analysis reveals that machine stiffness significantly affects serration morphology and dynamics, influencing mean amplitudes, stress drop periods, and dynamical indicators such as the correlation dimension and Lyapunov exponents. Comparison of simulated and experimental stress time series demonstrates chaotic behaviour for type B and C serrations across all strain rates, with no evidence of self-organised criticality in experimental data. However, simulations predict self-organised criticality at high strain rates corresponding to type A bands, consistent with literature references. Statistical indicators reveal power-law behaviour for stress drop amplitudes as a function of strain rate, with critical exponents dependent on band type. Multifractal analysis shows that simulations overestimate complexity relative to experimental observations, suggesting the need for additional internal variables and finer-scale dynamics in modelling. Additionally, multifractal analysis of spatio-temporal diagrams reveals power-law distributions of plastic strain rates with consistent critical exponents across all strain rates, demonstrating its potential for characterising PLC spatio-temporal dynamics. Statistical, dynamical, and multifractal indicators show consistent correlations, collectively capturing transitions between serration regimes and serving as reliable quantitative metrics for characterising PLC dynamics. The analysis is finally applied to the spatio-temporal strain fields measured by digital image correlation. The results demonstrate the value of multi-indicator analysis for assessing the agreement between experiment and simulation and subsequently improving constitutive model parameter identification
Integrated optimization and machine learning through hyperparameter selection: an application to predictive maintenance of wind turbines
No full textInternational audienceNumerous real-life problems involve two complex challenges: prediction, because of unknown or uncertain parameters or variables, and decision, because of the need to make a good one or the best one. These two challenges are not solved with the same tools: the firstone requires learning and the second one constrained optimization. The literature proposes various ways to link decision and prediction problems. The most common and obvious way is to perform the prediction first and then to optimize the decision based on this prediction. This sequential paradigm is called predict-then-optimize (PtO). Even though this paradigm is widespread, it often leads to sub-optimal decisions becausethe learning model used for the prediction aims at minimizing a loss function on the object to predict and not on the decision that will be taken based on this prediction (2).More recently, authors proposed a new paradigm, called Decision-Focused Learning (DFL), including the decision problem into the prediction problem by changing the loss function of the learning task to minimize the loss on the final decision. It supposes several conditions to be satisfied, because of the need to be able to compute the derivative, or a substitute, of the loss function. This is not immediate when considering a constrained optimization problem, in particular with integer variables (3).The model we propose optimizes the hyperparameter search of the learning model not with a traditional score, such as mean squared error or f1-score, but with a customized score being the objective function of the constrained optimization model of the decision problem. The decision is then integrated into the learning phase. As the scoring function of the hyperparameter search is not subject to the assumptions of derivability, which is the case for the loss function used in DFL, it requires much less effort to integrate the decision and the prediction phases.Predictive maintenance (PdM) is a case involving a decision problem, using constrained optimization, and a prediction problem, using learning. This paper examines how optimization and machine learning can be combined to improve PdM in wind farms. PdM is a maintenance strategy that aims at performing maintenance tasks little before a failure is likely to occur. Being able to schedule a task a little before a failure occurs assumes having information on the future operational state of the equipment. This information can be under the form of remaining useful life, ie. the time remaining before a failure, or of a binary information on the likeliness of a failure occurring the next period of time (4). Having these kinds of forecasts is a prediction problem. Wind turbines are equipped with multiple sensors monitoring and storing various quantities, such as temperatures, angles, rotation speeds, etc. The availability of data makes it suitable to use learning for failures prediction (1).Operation and Maintenance decision-makers wish to elaborate a good maintenance schedule, or even the best one, in the sense that it should minimize the maintenance costs and both the planned and unplanned downtime. The goal is to achieve a compromise between curative maintenance, that systematically causes unplanned downtime and high maintenance costs but guarantees a maximal use of the components, and preventive maintenance, that avoids some of the failures by planning operation at regular time interval with lower maintenance costs but still cause planned downtime, may lead to over-maintaining and does not erase the risk of failures. Predictive tasks should then replace preventive and curativeones.Maintenance planning is thus a decision problem. The maintenance planning is constrained by field rules such as potentiality, security, etc. (5). The tool used to find the best planning while satisfying the filed rules is constrained optimization.In the case of PdM for wind turbines, the constrained optimization program seeks to minimize a cost function composed of maintenance costs and of downtime costs, both planned and unplanned. The decision variables are the beginning of each maintenance task and is a Mixed Integer Linear Program (MILP). The tasks to schedule can be curative, preventive or curative tasks. A predictive task can replace a preventive or curative task if it is scheduled closely enough before this task’s due date. The data we work with to predict the failures is labeled with historical failures, thus thelearning task is not anomaly detection but supervised machine learning. Due to the fact that the failures do not happen as often as normal functioning, the dataset is highly skewed. We use a gradient boosting classification model for the prediction, as it is robust to skewed datasets. This still leads to an excessive amount of false positive, even using the f1-score as scoring function to find the best hyper-parameters, which combines precision and recall to ensure the quality of the prediction of a classifier.This paper proposes a model integrating the failure predictions, using a classification gradient boosting learning algorithm, and the planning optimization, using a scheduling MILP. The integration is done through the hyperparameters selection of the learning model as thescoring function of the grid search is a loss function based on the objective function of the scheduling MILP. Experimentations were carried out on industrial data shared by a partner. We compared the integrated predictive approach to a maintenance strategy purely preventive and curative and to a second maintenance strategy including predictive maintenance according to a PtO model, ie. without the loop on the learning-decision phases. Results show that the integrated model reduces the amount of false negative predictions, as expected. This has the effect of scheduling a number of predictive tasks that is closer to what is in fact needed, replacing preventive tasks and not performing too many predictive tasks. Results are convincing: the integrated predictive approach allows to gain up to 30% compared to the purely preventive and curative strategy and from 10% to 20% compared to the PtO predictive strategy, in maintenance costs
Direct multi-model dark-matter search with gravitational-wave interferometers using data from the first part of the fourth LIGO-Virgo-KAGRA observing run
No full textInternational audienceGravitational-wave detectors can probe the existence of dark matter with exquisite sensitivity. Here, we perform a search for three kinds of dark matter -- dilatons (spin-0), dark photons (spin-1) and tensor bosons (spin-2) -- using three independent methods on the first part of the most recent data from the fourth observing run of LIGO--Virgo--KAGRA. Each form of dark matter could have interacted with different standard-model particles in the instruments, causing unique differential strains on the interferometers. While we do not find any evidence for a signal, we place the most stringent upper limits to-date on each of these models. For scalars with masses between eV that couple to photons or electrons, our constraints improve upon those from the third observing run by one order of magnitude, with the tightest limit of at a mass of . For vectors with masses between eV that couple to baryons, our constraints supersede those from MICROSCOPE and Eöt-Wash by one to two orders of magnitude, reaching a minimum of at a mass of eV. For tensors with masses of eV (the full mass range analyzed) that couple via a Yukawa interaction, our constraints surpass those from fifth-force experiments by four to five orders of magnitude, achieving a limit as low as at eV. Our results show that gravitational-wave interferometers have become frontiers for new physics and laboratories for direct multi-model dark-matter detection
Two-way Coupling of Fluid-Structure Interaction for Elastic Magneto-Swimmers: A Finite Element ALE Approach
No full textArtificial micro-swimmers actuated by external magnetic fields hold significant promise for targeted biomedical applications, including drug delivery and micro-robot-assisted therapy. However, their dynamics remain challenging to control due to the complex nonlinear coupling between magnetic actuation, elastic deformations, and fluid interactions in confined biological environments. Numerical modeling is therefore essential to better understand, predict, and optimize their behavior for practical applications. In this work, we present a comprehensive finite element framework based on the Arbitrary Lagrangian-Eulerian formulation to simulate deformable elastic micro-swimmers in confined fluid domains. The method employs a fullorder model that resolves the complete fluid dynamics while simultaneously tracking swimmer deformation and global displacement on conforming meshes. Numerical experiments are performed with the open-source finite element library Feel++, demonstrating excellent agreement with experimental data from the literature. The validation benchmarks in both two and three dimensions confirm the accuracy, robustness, and computational efficiency of the proposed framework, representing a foundational step toward developing digital twins of magneto-swimmers for biomedical applications
Solid-fluid equilibria of mixtures of interest in LNG production: Measurement and modelling of methane + carbon dioxide + neo-pentane systems
No full textInternational audienc
Why companies use academic partnerships to invest in basic research
No full textLSE Business ReviewConventional wisdom assumes that companies invest in academic partnerships to produce short-term, commercially useful outcomes, and that academics cannot “do” business, making it difficult to align research agendas with industrial needs. Benjamin Cabanes and Quentin Plantec find that both those ideas are outdated and that industry is willing to initiate and fund basic research, even when immediate applications are unclear