HAL-MINES ParisTech
Not a member yet
27348 research outputs found
Sort by
Kriging Alluvial Thicknesses in Valley Bottoms Using Nonstationary Geometric Anisotropies
No full textInternational audienceModeling the geometry and volumes of alluvial infill at the bottom of river valleys is essential for a better understanding of river dynamics and alluvial storage, which buffers sedimentary signal transmission from source to sink. This study introduces kriging methods for predicting alluvial thickness geometries in the valley bottom using random fields with nonstationary geometric anisotropies. These anisotropies are constructed based on the contours of the alluvial plain. The two methods (one based on covariance and the other on Stochastic Partial Differential Equations (SPDE)) are compared to kriging with stationary anisotropies. The main parameters are estimated by maximum likelihood and the different methods are tested using cross-validation. The methods incorporating nonstationary anisotropies exhibit lower uncertainties than the stationary anisotropy approach and produce patterns that align more realistically with river geomorphic and sedimentologic knowledge. Evidence of braided patterns, rifflepool geometry related to meandering, and scouring at confluences is observed at the scale of the valley width. Finally, it is estimated that about 2.30 km 3 of alluvium is currently stored in the studied Oise valley bottom. Although the method was applied to an alluvial setting, it could be easily adapted to other geological contexts where anisotropy fields can be estimated, such as faulted and folded structures, supergene mineralization or environmental context to track pollutants dissemination in rivers
Sparse multitask group lasso for genome-wide association studies
No full textInternational audienceA critical hurdle in Genome-Wide Association Studies (GWAS) involves population stratification, wherein differences in allele frequencies among subpopulations within samples are influenced by distinct ancestry. This stratification implies that risk variants may be distinct across populations with different allele frequencies. This study introduces Sparse Multitask Group Lasso (SMuGLasso) to tackle this challenge. SMuGLasso is based on MuGLasso, which formulates this problem using a multitask group lasso framework in which tasks are subpopulations, and groups are population-specific Linkage-Disequilibrium (LD)-groups of strongly correlated Single Nucleotide Polymorphisms (SNPs). The novelty in SMuGLasso is the incorporation of an additional ℓ 1 -norm regularization for the selection of population-specific genetic variants. As MuGLasso, SMuGLasso uses a stability selection procedure to improve robustness and gap-safe screening rules for computational efficiency. We evaluate MuGLasso and SMuGLasso on simulated data sets as well as on a case-control breast cancer data set and a quantitative GWAS in Arabidopsis thaliana . We show that SMuGLasso is well suited to addressing linkage disequilibrium and population stratification in GWAS data, and show the superiority of SMuGLasso over MuGLasso in identifying population-specific SNPs. On real data, we confirm the relevance of the identified loci through pathway and network analysis, and observe that the findings of SMuGLasso are more consistent with the literature than those of MuGLasso. All in all, SMuGLasso is a promising tool for analyzing GWAS data and furthering our understanding of population-specific biological mechanisms
Plug-and-play learned proximal trajectory for 3D sparse-view X-ray computed tomography
No full textPlug-and-Play algorithms (PnP) have recently emerged as a powerful framework for solving inverse problems in imaging. They leverage the power of Gaussian denoising algorithms to solve complex optimization problems. This work focuses on the challenging task of 3D sparse-view X-ray computed tomography (CT). We propose to replace the Gaussian denoising network in Plug-and-Play with a restoration network, i.e. a network trained to remove arbitrary artifacts. We show that using a restoration prior tailored to the specific inverse problem improves the performances of Plug-and-Play algorithms. Besides, we show that plugging a basic restoration network into a PnP scheme is not sufficient to obtain good results. Thus, we propose a procedure to train the restoration network to be a robust approximation of a proximal operator along a pre-defined optimization trajectory. We demonstrate the effectiveness and scalability of our approach on two 3D Cone-Beam CT datasets and outperform state-of-the-art methods in terms of PSNR. Code is available at https://github.com/romainvo/ pnp-learned-proximal-trajectory</p
Decision-Focused Learning for Power System Decision-Making Under Uncertainty
No full textInternational audienceMore accurate forecasts may not necessarily lead to better decision-making. To address this challenge, decision-focused learning (DFL) has been proposed as a new branch of machine learning that replaces traditional statistical loss with a decision loss to form an end-to-end paradigm. Applications of DFL in power systems have been developed in recent years. However, existing applications remain fragmented without systematic analysis of methodologies or comparative benchmarks. This review addresses this gap by performing a set of scenario analysis, taxonomy analysis, application analysis and comparative analysis. It first illustrates the inherent mismatch between statistical accuracy and operational decisions through power system example scenarios. It then establishes a structured taxonomy of DFL techniques, categorizing methods by model structure (direct/indirect) and gradient handling (gradient-based/free). An application-based analysis reviews existing DFL applications by forecasting targets and decision contexts. Furthermore, an open-source comparative benchmark is developed to assess different DFL models through power system-specific metrics like cost reduction, forecasting accuracy, decision speed, providing a baseline for future research. Finally, this paper identifies the challenges to adopting DFL in power systems and presents future research directions, offering researchers a roadmap to advance DFL beyond theoretical analysis into power grid-tailored models
Imagerie des structures internes d'un glissement de terrain par une approche couplée gravimétrie - tomographie de résistivité électrique.
No full textInternational audienceL'imagerie des structures internes des glissements de terrain reste un défi important pour caractériser la dynamique des instabilités de versant. Cette étude porte sur l'imagerie interne d'un glissement de terrain situé sur la bordure sud des Causses du Larzac (France, Lodévois). Ici, la déformation des versants est associée à des épisodes de fortes pluies qui surviennent principalement après des périodes estivales chaudes et arides. Les failles et réseaux de fractures permettent l'infiltration des eaux de pluies jusqu'aux couches profondes du Trias, favorisant la déformation du versant. Comme ces glissements de terrain endommagent des infrastructures majeures dont l'autoroute A75, le risque gravitaire est très présent dans cette vallée. Des études sont donc essentielles pour documenter la géométrie complexe de ces glissements et mieux comprendre leur fonctionnement.Le glissement de terrain étudié est un glissement profond (50 m) avec un déplacement relativement lent (3 à 4 mm/an). Nous proposons ici une approche originale combinant des mesures de résistivité électrique et de la micro-gravimétrie. L'étude s'appuie également sur des données de forage et des observations géomorphologiques. Les résultats obtenus mettent en évidence les dépôts de pente caractérisés par une forte résistivité électrique et une anomalie négative de densité. Ces dépôts ont une épaisseur moyenne de 30 m, avec des variations latérales allant de 15 à 50 m dues à la géométrie complexe de l'interface entre les colluvions et les formations sous-jacentes ainsi qu'à la présence de blocs calcaires en profondeur. Les valeurs de porosité apparente estimées pour ces colluvions sont comprises entre 0,05 et 0,23, suggérant une capacité de stockage élevée et de possible variations importantes de la pression interstitielle lors de fortes précipitations pouvant induire des mouvements de pente. Cette étude souligne la complémentarité de la micro-gravimétrie avec les méthodes géophysiques conventionnelles, telles que la tomographie de résistivité électrique, sismique ou les diagraphies pour imager les structures internes. La gravimétrie offre une vision plus directe de la densité de la roche et de la porosité apparente associée, qui sont essentielles pour comprendre la dynamique interne des glissements de terrain
Estimation des matériaux, géométrie et sémantique à partir d'images pour la compréhension et édition de scènes
No full textLn this thesis, we explore multiple approaches for understanding materials and learning dense multi-task predictions from images. The thesis is divided into two parts, each focusing on either material understanding or dense multi-task learn- ing. Part l emphasizes material understanding, addressing the extraction, transformation, and editing of physical surface properties from images. We present methods for learning material mappings in 3D reflectance fields, extracting material palettes from single photographs through unsupervised domain adaptation, and performing lighting-aware material swaps in the image domain. Part ll shifts the focus to dense multi-task learning, where we predict multiple pixel-wise outputs such as depth, surface normals, semantic segmentation, and edges from a unified model. We introduce correlation-guided attention mechanisms that leverage task relationships to improve predictions and propose a zero-shot framework that re- purposes latent diffusion models to enable training on partially annotated synthetic datasets while generalizing to real images. Throughout both parts, we demonstrate how insights from computer vision, graphics, and deep learning can becombined to develop practical methods that bridge the gap between perception and rendering.Dans cette thèse, nous explorons plusieurs approches pour la compréhension des matériaux et l'apprentissage de pré- dictions multi-tâches denses à partir d'images. La thèse est divisée en deux parties, chacune se concentrant soit sur la compréhension des matériaux, soit sur l'apprentissage multi-tâches dense. La Part I met l'accent sur la compréhen- sion des matériaux, en abordant l'extraction, la transformation et l'édition des propriétés physiques des surfaces à partir d'images. Nous présentons des méthodes pour apprendre des correspondances de matériaux dans des champs de réflectance 3D, extraire des palettes de matériaux à partir de photographies uniques par adaptation de domaine non supervisée, et effectuer des échanges de matériaux tenant compte de l'éclairage dans le domaine image. La Part II déplace l'attention vers l'apprentissage multi-tâches dense, où nous prédisons plusieurs sorties pixel-à-pixel telles que la profondeur, les normales de surface, la segmentation sémantique et les contours à partir d'un modèle unifié. Nous introduisons des mécanismes d'attention guidés par corrélation qui exploitent les relations entre tâches pour améliorer les prédictions et proposons un cadre zero-shot qui réutilise les modèles de diffusion latents pour permettre l'entraînement sur des ensembles de données synthétiques partiellement annotés tout en généralisant aux images réelles. À travers ces deux parties, nous démontrons comment les connaissances issues de la vision par ordinateur, de l'infographie et de l'apprentissage profond peuvent être combinées pour développer des méthodes pratiques qui comblent l'écart entre perception et rendu
CAMS radiation service for solar energy: Exploring the error space with data-driven and spatiallyresolved methods and service evolution
No full textInternational audienceThe Copernicus Atmosphere Monitoring Service (CAMS) provides open data access to solar irradiances through its CAMS Radiation Service (CRS). Data is accessible via both the CAMS Atmospheric Data Store and user-specific Python libraries such as PV-LIB. The services meet the needs of European and national policy development and the requirements of partly commercial downstream services in the solar energy sector for e.g. planning, monitoring, efficiency improvements, and integration of renewable energies into the energy supply system. Observations from the meteorological satellites are combined with modelled aerosols, water vapor, and ozone from the CAMS Integrated Forecasting System to retrieve Surface Solar Irradiance (SSI). SSI time series in 1 min, 15 min, hourly and daily temporal resolution are produced 'on-the-fly' on user request at the desired location inside the domain of the MSG and the HIMAWARI satellites field of view. In addition to the standard access as time series, a gridded dataset in 0.1° and 15 min for the years 2004-2023 is available for land surfaces of Europe and Africa. Validation of the CRS is done regularly on a quarter-annual basis with a 6 months delay between data acquisition and assessment due to the delay in ground observation data provision from ground networks. The focus of this validation is on direct, diffuse and global irradiation components against ground observations and it provides classical, station-wise metrics such as bias, RMSE, and correlations. To widen the coverage of the quality assessment of the CRS, the ground observations database is continuously being extended with additional networks. Nevertheless, special care must be taken as these observations may have uncertainties of the order of the observed uncertainty in the CRS. Typically, in an operational setup, from the physical point of view, the retrieval of SSI is limited in the treatment of heterogeneous or multi-layer clouds and not well-defined aerosol conditions. Moreover, in a general setting the errors at the pixel-level are hard to estimate. Europe’s CAMEO project supports the CRS service evolution through a detailed assessment of the error patterns. In this project, data-driven methods e.g., SHAP analysis, are being exploited using all input parameters to bring light into the error space of the CRS. In addition, sensitivity studies and ML-based surrogate models provide additional insights on the errors. Mainly, we use retrieval inherent information to provide a single retrieval uncertainty estimate or a physics-informed bias correction. We aim at methods without the need for any auxiliary information which is not available in a stand-alone real-time satellite-based retrieval process. Furthermore, the CRS team is currently evaluating the 0.5 km visible channel of HIMAWARI to prepare for the upcoming MTG capabilities. Impact of this high-resolution information for solar radiation accuracy will be addressed, either as input for an ML-based uncertainty quantification or alternatively for a direct high-resolution retrieval product
A scalable activity-based and physics-informed bottom-up model for residential electricity demand using open data
No full textIn the context of the ongoing energy transition, accurately modeling the spatial and temporal heterogeneity of residential electricity load curves has become increasingly critical. Such models are essential for prospective studies ranging from emerging technology adoption to multi-scale distribution system analyses. This paper introduces an activity-based, physics-informed bottom-up model of residential electricity demand, developed exclusively from open data for its construction, calibration, and validation. The activity-based structure enables a realistic representation of occupant behavior, while the physics-based analytical expressions allow for the explicit modeling of transition pathways as well as the impact of climate change. To ensure scalability over large geographical domains, a hierarchical clustering methodology is employed to identify a minimal yet representative subset of districts. This approach enables the reconstruction of aggregated load profiles across extensive regions while preserving computational efficiency and accuracy. The resulting framework constitutes a scalable, parametrizable and robust tool for electricity consumption modeling that can be leveraged in a wide range of energy transition prospective studies.Dans le contexte de la transition énergétique en cours, la modélisation précise de l’hétérogénéité spatiale et temporelle des profils de charge résidentiels est devenue de plus en plus cruciale. De tels modèles sont essentiels pour des études prospectives allant de l’adoption de nouvelles technologies à des analyses multi-échelles des systèmes de distribution. Cet article présente un modèle de demande résidentielle en électricité basé sur les activités et informé par la physique, construit de manière descendante (« bottom-up ») et développé exclusivement à partir de données ouvertes pour sa construction, sa calibration et sa validation. La structure basée sur les activités permet une représentation réaliste du comportement des occupants, tandis que les expressions analytiques fondées sur la physique autorisent une modélisation explicite des trajectoires de transition ainsi que de l’impact du changement climatique. Afin d’assurer la scalabilité sur de vastes domaines géographiques, une méthodologie de regroupement hiérarchique est employée pour identifier un sous-ensemble minimal mais représentatif de districts. Cette approche permet la reconstruction de profils de charge agrégés sur de grandes régions tout en préservant l’efficacité computationnelle et la précision. Le cadre ainsi obtenu constitue un outil évolutif, paramétrable et robuste pour la modélisation de la consommation électrique, pouvant être utilisé dans une large gamme d’études prospectives sur la transition énergétique
Average Predictor-Feedback Control Design for Switched Linear Systems
No full textInternational audienc