Portail HAL edf
Not a member yet
11029 research outputs found
Sort by
Off-the-grid learning of mixtures from a continuous dictionary
No full textWe consider a general non-linear model where the signal is a finite mixture of an unknown, possibly increasing, number of features issued from a continuous dictionary parameterized by a real non-linear parameter. The signal is observed with Gaussian (possibly correlated) noise in either a continuous or a discrete setup. We propose an off-the-grid optimization method, that is, a method which does not use any discretization scheme on the parameter space, to estimate both the non-linear parameters of the features and the linear parameters of the mixture.We use recent results on the geometry of off-the-grid methods to give minimal separation on the true underlying non-linear parameters such that interpolating certificate functions can be constructed. Using also tail bounds for suprema of Gaussian processes we bound the prediction error with high probability. Assuming that the certificate functions can be constructed, our prediction error bound is up to -factors similar to the rates attained by the Lasso predictor in the linear regression model. We also establish convergence rates that quantify with high probability the quality of estimation for both the linear and the non-linear parameters.We develop in full details our main results for two applications: the Gaussian spike deconvolution and the scaled exponential model
Improving hydroacoustic methods for monitoring suspended‐sand flux and grain size in sediment‐laden rivers
No full textInternational audienceAbstract Suspended‐sand concentration and grain‐size data in rivers provide valuable information on the catchment's dynamics for scientists and river managers. Producing continuous measurements of suspended‐sand concentrations remains a scientific challenge due to their high spatial and temporal variability. Traditional methods such as sediment‐rating curves may be highly uncertain, and optical turbidity is insensitive to coarse particles when there are many fine particles. Surrogate hydroacoustic methods aim to improve sand concentration measurements. These single‐ or dual‐frequency acoustic methods use acoustic attenuation and/or backscatter to estimate fine‐sediment (i.e., silt and clay) and/or sand concentration and possibly grain size. New methods have recently been developed and applied in rivers exhibiting a wide range of sediment conditions in North America but not independently tested elsewhere by other researchers. In this article, we apply, adapt and evaluate hydroacoustic methods to continuously estimate suspended‐sand concentration and grain size in an Alpine river with high suspended‐sediment concentrations. From the example of the River Isère at Grenoble Campus, France, we show that the hydroacoustic methods adapted to local conditions may yield valuable sand concentration estimates consistent with traditional measurements. Compared with prior knowledge, limited additional information on the grain size can be obtained due to high uncertainties. Hydroacoustic concentration estimates are more sensitive to real changes in concentration at the event scale than traditional rating‐curve methods that relate concentration to discharge only. These findings open the perspective for facilitated sand concentration monitoring at a higher temporal resolution with decreased field work
Radionuclide sorption dynamics in the Rhone River: Experimental and modelling approach
No full textInternational audienceThe transfer of radionuclides discharged into rivers by nuclear facilities are conditioned by their solid/liquid fractionation, commonly represented by an equilibrium approach using the distribution coefficient K_d. This coefficient, largely used in modeling, assumes an instantaneous and completely reversible reaction. However, such assumptions are rarely verified. Compared to instantaneous adsorption of radionuclides onto particles, slower reactions may lead to an underestimation of the dissolved fraction, and modifications of environmental conditions (e.g. at confluences, dams…) may induce a change in solid/liquid partition. Considering this background, this study aims to assess whether models incorporating one or more kinetics to describe this fractionation allow more accurate estimations than the equilibrium approach. A large dataset has been obtained experimentally to compare and test fractionation models. The exchanges of four radionuclides (137Cs, 60Co, 54Mn, and 110mAg) between solution and riverine suspended particulate matter (from the Rhone River, France) were followed in laboratory. Adsorption kinetics were monitored starting from 30 minutes up to 2 months. They stabilized only after several days and up to 2 weeks. Dilutions of the contaminated suspension were carried out after 1 hour, 3 days, 10, 21 and 31 days of prior adsorption, to simulate a change in environmental conditions, such as the input of uncontaminated water from a tributary. The dilution induced a quick release from solid to liquid at first, followed by a non-expected re-adsorption for all four radionuclides. 31-days sorption data were used to fit the parameters of a Kd model and derived models involving one (EK) or two fractionation kinetics (KK). Predictions were then carried out to evaluate the model capacities to react to a dilution of the contamination. Unlike the Kd model, which is a constant, the kinetic models predicted a variation in the solid/liquid activity ratio over time and when environmental conditions changed. However, models developed on one-step sorption showed limitations to reproduce the adsorption kinetics observed after a dilution, especially when a previous steady-state was reache
Approximation results on resource leveling problems
No full textInternational audienceThis work deals with resource leveling problems. A set of jobs is given as well as a resource level representing a capacity that may be exceeded at some cost. Jobs have integer processing times, must be scheduled non-preemptively and consume one unit of resource while processed. More precisely, the objective to be maximized is the resource use below the resource level, i.e., the complementary of the total overload cost. Two main families of problems are investigated: either with or without precedence constraints. The case with no precedence constraints is shown to admit an EPTAS; a quasi-linear time approximation algorithm with constant ratio 7/8 is also provided. The case with precedence constraints is shown to be significantly harder to solve as it does not admit a PTAS under some classical complexity assumption. Approximation algorithms with constant ratios are provided for special cases with in-tree precedence graph or with fixed resource level
Optimizing computation time in 3D air quality models by using aerosol superbins within a sectional size distribution approach: Application to the CHIMERE model
No full textInternational audienceOne limitation of the operational application of air-quality models at high resolution for forecasting or for the evaluation of emission mitigation scenario is the computational cost. It may also be an important limitation to the use of more complex (but more realistic) secondary organic aerosol (SOA) schemes. While the size distribution may be accurately described with a sectional approach to resolve processes involved in aerosol dynamics, it also leads to large CPU time due to the number of size bins that need to be used.In this study, we developed a “superbin” approach consisting in lumping for a given species several size bins into a single size superbin and to use a specified size distribution to distribute the superbin concentration into the different bins of CHIMERE when needed.Together with the revision of the numerical resolution algorithm, the ”superbin” approach was implemented into a new version of CHIMERE (based on v2020r1) in order to optimize the CPU time performance. The computation time was reduced by 60% with induced errors on PM10 concentrations around 3% to 7% over most of Europe. The use of the “superbin” approach proved to be much more efficient in terms of computational time and errors compared to simply reducing the number of bins
A Combined Slump-Sliding Test for Dredged Fine Sediments
No full text(IF 1.2;Q4)International audienceThis study presents a new laboratory apparatus for measuring and estimating the ability of shoveling, loading, and transport of fine soils and sediments. This device permits the investigation of the slump and sliding behavior of fine sediments based on their consistency and water content. Initially, a manual version of the apparatus was designed. The combined slump-sliding test (SST) and its use through a simple procedure were described in detail. Measurements done during tests on three different sediments and smooth steel support allow the presentation of the typical results obtained from the SST test. Because consistency governs the slump and sliding behavior of sediments, grain size distribution and Atterberg limits are the most useful characteristics of sediments. Typical relationships among slumps, cone footprints, and sliding rotation angle versus water content during moistening of sediments were established and explained. Based on these relationships, criteria for the ability to shovel sediments were proposed
Faster than real-time, phase-resolving, data-driven model of wave propagation and wave–structure interaction
No full textInternational audienceA machine learning time-series prediction approach is proposed for wave propagation and wave load prediction. Under unidirectional wave conditions and variable bathymetry, given a wave gauge upstream, a model is shown to reproduce wave elevation or wave forces downstream under irregular steep and either non-breaking or breaking conditions. Attempts to perform the opposite calculation, predicting upstream conditions from downstream measurements, results in higher error, likely due to information loss under breaking conditions. For choice of machine learning approach, comparisons show that the Time-series Dense Encoder (TiDE) approach results in a good balance between model complexity, stability, computational time, and error. Over a flat bottom, time-series of wave elevation can be predicted up to 10 wavelengths away, though with a degraded accuracy compared to shorter distances. Similar results are shown for time-series of forces on a vertical cylinder, showing better results than a simple Morison approach, as used in engineering tools such as OpenFAST, but with a similarly fast computational time. Generalizations show that training on irregular wave data permit extrapolations to periodic wave cases. Finally, the same method also is also demonstrated at field-scale, comparing results between two offshore buoys
Discrete Weber inequalities and related Maxwell compactness for hybrid spaces over polyhedral partitions of domains with general topology
No full textInternational audienceWe prove discrete versions of the first and second Weber inequalities on -like hybrid spaces spanned by polynomials attached to the faces and to the cells of a polyhedral mesh. The proven hybrid Weber inequalities are optimal in the sense that (i) they are formulated in terms of - and -like hybrid semi-norms designed so as to embed optimally (polynomially) consistent face penalty terms, and (ii) they are valid for face polynomials in the smallest possible stability-compatible spaces. Our results are valid on domains with general, possibly non-trivial topology. In a second part we also prove, within a general topological setting, related discrete Maxwell compactness properties
Co‐Doping Approach for Enhanced Electron Extraction to TiO 2 for Stable Inorganic Perovskite Solar Cells
No full textInternational audienceInorganic perovskite CsPbI 3 solar cells hold great potential for improving the operational stability of perovskite photovoltaics. However, electron extraction is limited by the low conductivity of TiO 2 , representing a bottleneck for achieving stable performance. In this study, a co‐doping strategy for TiO 2 using Nb(V) and Sn(IV), which reduces the material's work function by 80 meV compared to Nb(V) mono‐doped TiO 2 , is introduced. To gain fundamental understanding of the processes at the interfaces between the perovskite and charge‐selective layer, transient surface photovoltage measurements are applied, revealing the beneficial effect of the energetic and structural modification on electron extraction across the CsPbI 3 /TiO 2 interface. Using 2D drift‐diffusion simulations, it is found that co‐doping reduces the interface hole recombination velocity by two orders of magnitude, increasing the concentration of extracted electrons by 20%. When integrated into n–i–p solar cells, co‐doped TiO 2 enhances the projected T S80 lifetimes under continuous AM1.5G illumination by a factor of 25 compared to mono‐doped TiO 2 . This study provides fundamental insights into interfacial charge extraction and its correlation with operational stability of perovskite solar cells, offering potential applications for other charge‐selective contacts
Provably safeguarding a classifier from OOD and adversarial samples
No full textInternational audienceThis paper aims to transform a trained classifier into an abstaining classifier, such that the latter is provably protected from out-of-distribution and adversarial samples. The proposed Sample-efficient Probabilistic Detection using Extreme Value Theory (SPADE) approach relies on a Generalized Extreme Value (GEV) model of the training distribution in the latent space of the classifier. Under mild assumptions, this GEV model allows for formally characterizing out-of-distribution and adversarial samples and rejecting them. Empirical validation of the approach is conducted on various neural architectures (ResNet, VGG, and Vision Transformer) and considers medium and large-sized datasets (CIFAR-10, CIFAR-100, and ImageNet). The results show the stability and frugality of the GEV model and demonstrate SPADE's efficiency compared to the state-of-the-art methods