Solving electromagnetic integral equations: from high-order discretization to H-matrix compression
The simulation of electromagnetic (EM) wave problems plays a key role in many fields, ranging from object characterization and inspection to radar stealth. A common approach consists in transforming the initial 3D volumetric problem into a 3D surface integral equation defined at material interfaces and discretizing it using a Boundary Element Method (BEM). High-order (HO) discretization schemes accelerate approximation convergence and thus provide a significant gain in accuracy for a given number of mesh elements, but complicates the assembly of the linear system. In particular, the calculation of singular and quasi-singular integrals becomes challenging, while the integration of HO methods into fast compression algorithms based on the hierarchical matrix formalism (HMAT) raises questions about the overall efficiency of the BEM. This postdoctoral position aims to address these in order to obtain a solution that is both efficient and robust for the targeted EM applications.
Instrumentation and Physics-Informed 3D Reconstruction for Quantum Magnetic Diagnosis of Li-ion Batteries
The OPTIMAQ project of the CEA aims to reduce the reject rate in Li-ion cell formation — currently up to 50%, with a target of 5% — through a non-invasive quantum magnetic measurement system. This post-doctoral position covers both the software and experimental strands of the project: setting up the measurement bench (helium-4 optically pumped magnetometer, shielding, cell translation system) and developing a physics-informed 3D current density reconstruction algorithm (RBF-FD/PINN) to detect defects as early as the formation stage.
Hosted at CEA-LETI in Grenoble within the LSSC laboratory (sensor signal processing, reconstruction), in close collaboration with the L2EP/LAIC laboratories (battery system instrumentation, quantum magnetometry) at CEA-LETI and the LAPS/LMPS laboratories (Li-ion electrochemistry: experimental implementation and testing, modeling and simulation) at CEA-LITEN, this position is part of a 24-month project targeting TRL 3, with prospects for publications and patents.
Artificial Intelligence for Rapid Source Term Estimation (AIFORSTE)
The AIFORSTE project of the CEA aims to improve emergency management for atmospheric releases resulting from industrial accidents or malicious acts. We will develop a source estimation method operating in near real-time, enabling a fast and effective response to these critical events where time is always a fundamental constraint.
Given the impossibility of using detailed CFD simulations for immediate intervention, and the inaccuracy of simplified analytical models, this postdoctoral position will be devoted to a hybrid approach based on Physics-Informed Artificial Intelligence.
The core of the work will consist in designing and optimizing deep neural network architectures combining PINNs (Physics-Informed Neural Networks) and Neural Operators (FNO, DeepONet). The objective will be to quasi-instantaneously evaluate the position, intensity, and kinetics of one or several unknown release sources from a limited number of noisy measurements, while respecting the physical conservation laws.
Hosted at CEA-Leti (Grenoble) and co-supervised by CEA-DAM, the project will apply to realistic case studies (urban districts, industrial sites) using data generated through high-fidelity simulation. This 18-month project offers a unique opportunity to publish in leading journals and to help build strategic technological autonomy in a field where security and environmental monitoring are of major importance.
Development and validation of a numerical solver for the neutron noise equations
There is currently no operational technique capable of continuously measuring the reactivity of nuclear systems—that is, their distance from criticality—during low-power operation. This issue is of major importance, for example, for power reactors during the fuel loading phase. In these subcritical systems driven by weak sources, the average detector signal provides little information; by contrast, the temporal fluctuations in the neutron signal — stochastic neutron noise, arising from the fundamental randomness in the number of neutrons emitted by fission — provide direct and quantitative information on the multiplication process. This work, carried out as part of the SENTINELLE project proposed by the CEA, aims to develop a deterministic GPU-based solver for the moment equations (mean and correlations) of the neutron population, enabling the rapid calculation of noise observables (Feynman-a, Rossi-a, local multiplicities). The solver will then be validated using experimental data from the SENTINELLE project.
Robust path-following solvers for the finite element simulation of cracking in complex heterogeneous media: application to reinforced concrete structures
Path-following (or continuation) procedures are used to describe the unstable responses of structures exhibiting snap-back or snap-through phenomena. These methods consist in adapting the external load during the deformation process in order to satisfy a control constraint, by introducing an additional unknown, the load multiplier. Several variants exist depending on the controlled quantity: degrees of freedom, strain measures, or variables related to energy dissipation.
In addition to enabling the tracing of unstable responses, a major advantage of these approaches lies in improving the convergence of incremental Newton-type solvers by reducing the number of iterations required. This gain often compensates for the additional computational cost associated with the continuation algorithm. Some formulations have proven both efficient and simple to implement.
However, no objective criterion yet allows one to determine which formulation is best suited for the simulation of reinforced concrete structures, where multiple dissipation mechanisms coexist along with a strong spatial variability of the material properties.
The proposed postdoctoral work aims to develop robust path-following algorithms for such structures, building upon previous research carried out at CEA. It will include a critical analysis of existing formulations, an evaluation of their performance (monolithic or partitioned solvers), followed by their implementation. Finally, representative test cases of industrial structures will be simulated to assess the gain in robustness and computational cost compared to standard solvers.
Study of the Velocity-Vorticity-Pression formulation for discretising the Navier-Stokes equations.
The incompressible Navier-Stokes equations are among the most widely used models to describe the flow of a Newtonian fluid (i.e. a fluid whose viscosity is independent of the external forces applied to the fluid). These equations model the fluid's velocity field and pressure field. The first of the two equations is none other than Newton's law, while the second derives from the conservation of mass in the case of an incompressible fluid (the divergence of velocity vanishes). The numerical approximation of these equations is a real challenge because of their three-dimensional and unsteady nature, the vanishing divergence constraint and the non-linearity of the convection term. Various discretisation methods exist, but for most of them, the mass conservation equation is not satisfied exactly. An alternative is to introduce the vorticity of the fluid as an additional unknown, equal to the curl of the velocity. The Navier-Stokes equations are then rewritten with three equations. The post-doc involves studying this formulation from a theoretical and numerical point of view and proposing an efficient algorithm for solving it, in the TrioCFD code.
VALERIAN: caracterizing electron transport for the ITkPix modules of ATLAS
A precise description of the transport of electrons and photons in matter is crucial in several of the CEA's flagship fields, notably radiation protection and nuclear
instrumentation. Their validation requires dedicated parametric studies and measurements.Given the scarcity of public experimental data, comparisons between calculation codes are also used. The challenge for the coming years is to qualify these codes in a broad energy domain, as certain discrepancies between their results have been identified during preliminary SERMA studies involving the coupled transport of neutrons, photons and electrons. The VALERIAN project involves seizing the opportunity created by a unique data collection Campaign planned for 2025-2026 at the IRFU (DRF) to better characterise these discrepancies. The IRFU has undertaken to check at least 750 pixel modules for the new trajectograph of the ATLAS experiment, as part of the rejuvenation of the large detectors at CERN. Numerous measurements with beta sources will be carried out in 2025-2026 for the qualification of these modules.
Advanced reconstruction methods for cryo-electron tomography of biological samples
Cryo-electron tomography (CET) is a powerful technique for the 3D structural analysis of biological samples in their near-native state. CET has seen remarkable advances in instrumentation in the last decade but the classical weighted back-projection (WBP) remains by far the standard CET reconstruction method. Due to radiation damage and the limited tilt range within the microscope, WBP reconstructions suffer from low contrast and elongation artifacts, known as ‘missing wedge’ (MW) artifacts. Recently, there has been a revival of interest in iterative approaches to improve the quality and hence the interpretability of the CET data.
In this project, we propose to go beyond the state-of-the-art in CET by (1) applying curvelet- and shearlet-based compressed sensing (CS) algorithms, and (2) exploring deep learning (DL) strategies with the aim to denoise et correct for the MW artifacts. These approaches have the potential to improve the resolution of the CET reconstructions and facilitate the segmentation and sub-tomogram averaging tasks.
The candidate will conduct a comparative study of iterative algorithms used in life science, and CS and DL approaches optimized in this project for thin curved structures.
Modeling of charge noise in spin qubits
Thanks to strong partnerships between several research institutes, Grenoble is a pioneer in the development of future technologies based on spin qubits using manufacturing processes identical to those used in the silicon microelectronics industry. The spin of a qubit is often manipulated with alternating electrical (AC) signals through various spin-orbit coupling (SOC) mechanisms that couple it to electric fields. This also makes it sensitive to fluctuations in the qubit's electrical environment, which can lead to large qubit-to-qubit variability and charge noise. The charge noise in the spin qubit devices potentially comes from charging/discharging events within amorphous and defective materials (SiO2, Si3N4, etc.) and device interfaces. The objective of this postdoc is to improve the understanding of charge noise in spin qubit devices through simulations at different scales. This research work will be carried out using an ab initio type method and also through the use of the TB_Sim code, developed within the CEA-IRIG institute. This last one is able of describing very realistic qubit structures using strong atomic and multi-band k.p binding models.
detection of multiplets and application to turkey-Syria seismic crisis of february 2023
The correlation technique, or template matching, applied to the detection and analysis of seismic events has demonstrated its performance and usefulness in the processing chain of the CEA/DAM National Data Center. Unfortunately, this method suffers from limitations which limit its effectiveness and its use in the operational environment, linked on the one hand to the computational cost of massive data processing, and on the other hand to the rate of false detections that could generate low-level processing. The use of denoising methods upstream of processing (example: deepDenoiser, by Zhu et al., 2020), could also increase the number of erroneous detections. The first part of the research project consists of providing a methodology aimed at improving the processing time performance of the multiplets detector, in particular by using information indexing techniques developed in collaboration with LIPADE (L-MESSI method , Botao Peng, Panagiota Fatourou, Themis Palpanas. Fast Data Series Indexing for In-Memory Data. International Journal on Very Large Data Bases (VLDBJ) 2021). The second part of the project concerns the development of an auto-encoder type “filtering” tool for false detections built using machine learning. The Syria-Turkey seismic crisis of February 2023, dominated by two earthquakes of magnitude greater than 7.0, will serve as a learning database for this study.