Construction of databases for radionuclide identification based on neural networks (NANTISTA project)
The project NANTISTA (Neuromorphic Architecture for Nuclear Threat Identification for SecuriTy Applications) deals with the prevention of illegal traffic of nuclear materials at international borders. The project aims at the development of a detection platform using plastic scintillators for fast radionuclide identification (such as fissile materials) based on neural networks. The post-doctoral subject consists in the development of the detection system and the construction of databases dedicated to the learning process and the optimization of the neural networks. The databases will be built with experimental measurements given by radioactive sources. Radiation-matter simulations (Monte-Carlo codes Geant4 and Penelope) will also be implemented for the construction of the databases.
Distributed optimal planning of energy resources. Application to district heating
Heating district networks in France fed more than one million homes and deliver a quantity of heat equal to about 5% of the heat consumed by the residential and tertiary sector. Therefore, they represent a significant potential for the massive introduction of renewable and recovery energy. However, heating networks are complex systems that must manage large numbers of consumers and producers of energy, and that are distributed in extended and highly branched geographical zones. The aim of the STRATEGE project, realized in collaboration among the CEA-LIST and the CEA-LITEN, is to implement an optimal and dynamic management of heating networks. We propose a multidisciplinary approach, by integrating the advanced network management using Multi-Agent Systems (MAS) and by considering simplified physical models of transport and recovery of heat developed on Modelica.
The post-doc’s goal is to design mechanisms of planning and optimization for allocation of heat resources that consider the geographical information from a GIS and the predictions of consumption, production and losses calculated with the physical models. In this way, several characteristics of the network will be considered: the continuous and dynamic aspect of the resource; sources with different behaviors, capabilities and production costs; the dependence of consumption/production to external aspects (weather, energy price); the internal characteristics of the network (losses, storage capacity). The developed algorithms will be implemented in a existing MAS management plateform and will constitute the main brick of a decision-support engine for the management of heating systems. It will initially operate in a simulated environment and in a second time online on a real system.
Compressed Sensing for ultrasonic imaging: disruptive method development and prototyping
In non-destructive ultrasonic testing, multi-element sensors are used for the inspection of structures to ensure the safety of people and infrastructures. Currently, one of the driving factor of an ultrasonic method is the number of elements of the sensor, influencing the speed and efficiency of the inspection but also the cost and the volume of the equipment. This project aims at developing a prototype of a multi-element sensor with a limited number of elements compared to current state of the art equipment, without losing imaging resolution. To achieve this goal, Compressed Sensing (CS), a recent technique of signal processing allowing to go beyond the traditional sampling theorems and to reconstruct data from severely undersampled measurements, will be used. The ultrasonic inspection procedure will need to be entirely rethought to meet the CS requirements, specifically the sparsity of the measured data and the incoherence of the measurement process. The expected results is a significant reduction (of the order of 5) of the number of elements to conduct imaging, which would be a true revolution in NDT with direct applications in various industrials sectors.
The following laboratories, all located in Saclay (France) of the CEA (the French atomic commission), will participate to the project: the NDT department for its expertise in multi-element ultrasonic testing and Neurospin and Cosmostat for their expertises in the field of CS, mainly applied to medical RMI imaging and astrophysics, respectively. The collaboration between these three labs, each among the worldwide leading institutes in their respective fields, will ensure the creation of a new and disruptive family of sensors.
Distributed multiagent resources allocation. Application to district heating
Heating district networks in France fed more than one million homes and deliver a quantity of heat equal to about 5% of the heat consumed by the residential and tertiary sector. Therefore, they represent a significant potential for the massive introduction of renewable and recovery energy. However, heating networks are complex systems that must manage large numbers of consumers and producers of energy, and that are distributed in extended and highly branched geographical zones. The aim of the SIGMA project, realized in collaboration among the CEA-LIST and the CEA-LITEN, is to implement an optimal and dynamic management of heating networks. We propose a multidisciplinary approach, by integrating the advanced network management using Multi-Agent Systems (MAS), by taking into account spatial constraints using Geographic Information Systems (GIS) and by considering simplified physical models of transport and recovery of heat.
The post-doc’s goal is to design mechanisms for dynamically allocating resources that consider the geographical information from the GIS and the predictions of consumption, production and losses calculated with the physical models. In this way, several characteristics of the network will be considered: the continuous and dynamic aspect of the resource; sources with different behaviors, capabilities and production costs; the dependence of consumption / production to external aspects (weather, energy price); the internal characteristics of the network (losses, storage capacity). The coupling with a GIS should allow implementing self-configuration mechanisms for the management of different networks and different levels of granularity obtained by reduction of the original GIS. The MAS should dynamically establish the link between the suitable simplified models and the desired level of granularity and then it will create the agents needed to represent the system.
Real time low cost algorithms for brain computer interface with multiple degrees of freedom
The topic of the postdoctoral project is the optimization of BCI methods and algorithms for medical application in humans (quadriplegic subjects).
Namely the particular goal of the postdoctoral fellow will be optimization and the acceleration of calculation to allow multiple degrees of freedom (up to 26) in real time. Selecting the appropriate features subset will improve the computational efficiency and the quality of control. To this purpose the algorithms of sparse modeling will be applied.
To map ECoG recordings to the spatial-temporal-frequency space, continuous wavelet transform (CWT) is applied. Optimization will include the implementation of low cost CWT and C++ coding.
The project will include the test and the adaptation of BCI algorithms to wireless signal transmission with the implant WIMAGINE.
Finally the adaptation of algorithms to medical environment of quadriplegic subjects (the use of imaginary tasks, presence of stimuli in the signal, the restricted duration of experiments) will be under responsibility of postdoctoral scientist.
Development of a Metal Supported Cell for Hydrogen production by High Temperature Steam Electrolysis
The development of Metal Supported Cells (MSC) for High Temperature Steam Electrolysis (HTSE) constitutes an interesting innovation able to reduce the degradation of this component under operation. An increase in the cell life time would be a valuable contribution to cost reduction and is able at positioning HTSE as an alternative process to other hydrogen production technologies. However, some progresses in the elaboration of MSCs are still required. Within the current process, functional ceramic layers of the MSC are joined to the metallic substrate at elevated temperature (> 1000 °C). Mismatch of the mechanical properties of the materials as well as the reducing conditions fixed by the metal substrate during sintering lead to MSCs having insufficient electrochemical performances. The post-doctorate aims, on the one hand, at obtaining a better understanding of the mechanisms that occur in the multilayer structure during sintering and, on the other hand, at proposing and testing technological solutions able to improve to reliability of MSC elaboration.
Development of numerical tools for the simulation of ultrasonic nondestructive inspection.
The CEA LIST develops the CIVA software platform (http://www-civa.cea.fr), in position of global leader for the simulation and expertise of non-destructive testing (ultrasonic and electromagnetic methods, radiography, X tomography). In the context of extending and improving capabilities of the CIVA platform, the post-doc fellow will contribute to the development of numerical methods for the ultrasonic testing (UT) simulation module.
The semi-analytical models of CIVA are based on physical simplifying hypotheses for the propagation of the ultrasonic beam and its interaction with defects. These methods provide efficient computational performances with accurate results in a wide range of realistic applications. However, configurations for which these models are not valid. The so called “numerical models” including finite element method (FEM), finite difference method (FDM) or boundary element method (BEM) enable to deal with these configurations without any simplifying hypotheses. However, the computation time is prohibitive in an industrial context, especially for 3D applications. The approach adopted in the CIVA platform consists in hybridizing semi analytical models with numerical ones in order to benefit of both the numerical efficiency of the former and the capability of simulating complex phenomena of the later.
The tasks associated to the post include the development of numerical methods themselves, in collaboration with academic partners, the integration in CIVA of a meshing solution enabling the data formatting for the numerical solving and specific developments for the coupling with semi analytical methods. Multi-boxes solutions, enabling for example the optimization of configurations with several defects or a coupling around the source and the defect, will be studied.
Compensation methods (for magnetic perturbation) for shape capture via orientation sensors
Our laboratory works for several years on shape capture (curves, surfaces) in static and moving positions, via inertial sensors - e.g. accelerometers ans magnetometers - able to provide information about their own orientation. In fact, in real conditions, sensors do not exactly provide their orientation, the measure is disturbed with external contributions (own motion acceleration, vibrations, magnetic perturbations). This work consists in analysing these disturbances, proposing preprocessing to clean data to obtain "denoised" tangential information to allow the reconstruction of these curves and surfaces.
First, we study the case of the reconstruction of a metallic pipe: we want to reconstruct a curve with magnetic sensors disturbed (the surface reconstruction will be explored afterwards). This work consists in finding the best methods allowing to extract the needed information from these "noised" signals (data fusion, source separation, model of perturbations, adding a new sensor modality,... are domains to explore). In this goal, a bibliographic study will be done firstly by the Post Doc student, then he will have to implement the different methods found, and test the performances with real signals acquired with our system of shape capture in a disturbed environment.
Model reduction in dynamics : application to earthquake engineering problems
The complexity and refinement of the numerical models used to predict the behavior of structures under seismic loading often impose computation times of several days for solving the partial differential equations of the reference problem.
Furthermore, in the context of optimization , model identification, or parametric and stochastic analyses, the aim is not to predict the response of a unique model but of a family of models.
To reduce the computation time, model reduction techniques (Proper Orthogonal/Generalized Decomposition) may be considered. This post-doctoral study proposes to define and implement (especially in the FE code CAST3M) a technique suitable for the reduction of reinforced concrete type models subjected to seismic loading.
Unsupervised Few-Shot Detection of Signal Anomalies
Our laboratory, located at Digiteo in CEA Saclay, is looking for a postdoc candidate working on the subject of anomaly detection in manufacturing processes, for a duration of 18 months starting from Feburary 2022. This postdoc is part of HIASCI (Hybridation des IA et de la Simulation pour le Contrôle Industriel), a CEA LIST project in an internal collaboration which aims at building a platform of AI methods and tools for manufacturing applications, ranging from quality control to process monitoring. Our laboratory contributes to HIASCI by developping efficient methods of anomaly detection in acoustic or vibrational signals, operating with small amounts of training data. In this context, the detection of signal anomalies (DSA) consists of extracting from data the information about the physical process of manufacturing, which is in general too complex to be fully understood. Moreover, real data of abnormal states are relatively scarce and often expensive to collect. For these reasons we privilege a data-driven approach under the framework of Few-Shot Learning (FSL).