Lightweight CNN and Causal GNN for scene understanding
Scene understanding is a major challenge in computer vision, with recent approaches dominated by transformers (ViT, LLM, MLLM), which offer high performance but at a significant computational cost. This thesis proposes an innovative alternative combining lightweight convolutional neural networks (Lightweight CNN) and causal graph neural networks (Causal GNN) for efficient spatio-temporal analysis while optimizing computational resources. Lightweight CNNs enable high-performance extraction of visual features, while causal GNNs model dynamic relationships between objects in a scene graph, addressing challenges in object detection and relationship prediction in complex environments. Unlike current transformer-based models, this approach aims to reduce computational complexity while maintaining competitive accuracy, with potential applications in embedded vision and real-time systems.
Development of automatic gamma spectrum analysis using a hybrid machine learning algorithm for the radiological characterization of nuclear facilities decommissioning.
The application of gamma spectrometry to radiological characterization in nuclear facility decommissioning, requires the development of specific algorithms for automatic gamma spectrum analysis. In particular, the classification of concrete waste according to its level of contamination, is a crucial issue for controlling decommissioning costs.
Within CEA/List, LNHB, in collaboration with CEA/DEDIP, has been involved for several years in the development of tools for the automatic analysis of low-statistics gamma spectra, which can be applied to scintillator detectors (NaI(Tl), plastics). In this context, an original approach based on a hybrid machine learning/statistics spectral unmixing algorithm has been developed for the identification and quantification of radionuclides in the presence of significant deformations in the measured spectrum, due in particular to interactions between the gamma emission from the radioactive source and its environment.
The proposed subject follows on from thesis work that led to the development of the hybrid algorithm with the aim of extending this approach to the radiological characterization of concrete surfaces. The candidate will be involved in the evolution of the hybrid machine learning/statistical algorithm for the characterization of concrete for classification as conventional waste. The work will include a feasibility study of modeling the deviations of the learned model to optimize the robustness of decision-making.
Optimized control of a modular energy hub with minimal EMC signature
The integration of renewable energy sources (RES) has become an important issue for power converters. The increasing number of these converters and their average utilization rate allows for a rethink of energy exchange management at the system level. This leads us to the concept of an energy hub, which can interface, for example, a photovoltaic (PV) system, an electric vehicle, a grid, and stationary storage with loads.
The main objective of this thesis is to improve the efficiency, compactness, and modularity of the energy hub through control. Several ideas emerge to achieve this, such as advanced control to minimize losses, the use of AC input opposition to reduce electromagnetic compatibility (EMC) filtering, series/parallel DC output configurations to address 400Vdc/800Vdc batteries, and increasing the switching frequency to reduce volume, etc.
Thus, this thesis will, in the medium term, lead to the development of an optimal converter in terms of both energy efficiency and environmental impact.
Learning Mechanisms for Detecting Abnormal Behaviors in Embedded Systems
Embedded systems are increasingly used in critical infrastructures (e.g., energy production networks) and are therefore prime targets for malicious actors. The use of intrusion detection systems (IDS) that dynamically analyze the system's state is becoming necessary to detect an attack before its impacts become harmful.
The IDS that interest us are based on machine learning anomaly detection methods and allow learning the normal behavior of a system and raising an alert at the slightest deviation. However, the learning of normal behavior by the model is done only once beforehand on a static dataset, even though the embedded systems considered can evolve over time with updates affecting their nominal behavior or the addition of new behaviors deemed legitimate.
The subject of this thesis therefore focuses on studying re-learning mechanisms for anomaly detection models to update the model's knowledge of normal behavior without losing information about its prior knowledge. Other learning paradigms, such as reinforcement learning or federated learning, may also be studied to improve the performance of IDS and enable learning from the behavior of multiple systems.
Side-Channel based Reverse-Engineering
The characterization of the security of embedded systems in "black box" or "gray box" against Side-Channel attacks often requires a preparatory phase of Reverse-Engineering, which can be particularly time-consuming, especially on a complex System-on-Chip that can be found in smartphones or in the automotive industry. This phase can, for example, consist of detecting a cryptographic primitive within Side-Channel measurements for a future observation attack, or a target routine for a fault injection attack. The objective of this thesis is to develop a methodology and non-profiled tools that allow the automation of this detection phase, while enabling the exploitation of prior knowledge of a potential attacker.
Gyrokinetic modelling of the nonlinear interaction between energetic particle-driven instabilities and microturbulence in tokamak plasmas
Tokamak plasmas are strongly nonlinear systems far from thermodynamic equilibrium, in which instabilities of very different spatial scales coexist, ranging from large-scale macroscopic oscillations to microturbulence. The presence of energetic ions produced by fusion reactions or by auxiliary heating further enhances these instabilities through wave–particle resonances. Microturbulence is responsible for heat and particle transport in the thermal plasma, while instabilities driven by energetic particles can induce their radial transport and, consequently, their losses. Both phenomena degrade the performance of present tokamak plasmas, and possibly also those of burning plasmas such as ITER.
Recent results, however, show that these instabilities, which have long been studied separately, can interact nonlinearly, and that this interaction may lead to an unexpected improvement of plasma confinement.
The objective of this project is to investigate these multiscale interactions using the gyrokinetic code GTC, which is able to simultaneously simulate turbulence and energetic-particle-driven instabilities. This work aims to improve the understanding of the nonlinear mechanisms governing plasma confinement and to identify optimal regimes for future fusion plasmas.
Electromagnetic Signature Modeling and AI for Radar Object Recognition
This PhD thesis offers a unique opportunity to work at the crossroads of electromagnetics, numerical simulations, and artificial intelligence, contributing to the development of next-generation intelligent sensing and recognition systems. The intern will join the Antenna & Propagation Laboratory at CEA-LETI, Grenoble (France), a world-class research environment equipped with state-of-the-art tools for propagation channel characterization and modelling. A collaboration with the University of Bologna (Italy) is planned during the PhD.
This PhD thesis aims to develop advanced electromagnetic models of near-field radar backscattering, tailored to radar and Joint Communication and Sensing (JCAS) systems operating at mmWave and THz frequencies. The research will focus on the physics-based modeling of the radar signatures of extended objects, accounting for near-field effects, multistatic and multi-antenna configurations, as well as the influence of target materials and orientations. These models will be validated through electromagnetic simulations and dedicated measurement campaigns, and subsequently integrated into scene-level and multipath propagation simulation tools based on ray tracing. The resulting radar signatures will be exploited to train artificial intelligence algorithms for object recognition, material property inference, and radar imaging. In parallel, physics-assisted AI approaches will be investigated to accelerate electromagnetic simulations and reduce their computational complexity. The final objective of the thesis is to integrate radar backscattering-based information into a 3D Semantic Radio SLAM framework, in order to improve localization, mapping, and environmental understanding in complex or partially obstructed scenarios.
We are seeking a student at engineering school or Master’s level (MSc/M2), with a strong background in signal processing, electromagnetics, radar, or telecommunications. An interest in artificial intelligence, physics-based modeling, and numerical simulation is expected. Programming skills in Matlab and/or Python are appreciated, as well as the ability to work at the interface between theoretical models, simulations, and experimental validation. Scientific curiosity, autonomy, and strong motivation for research are essential.The application must include a CV, academic transcripts, and a motivation letter.
Study of Failure Modes and Mechanisms in RF Switches Based on Phase-Change Materials
Switches based on phase change materials (PCM) demonstrate excellent RF performance (FOM <10fs) and can be co-integrated into the BEOL of CMOS processes. However, their reliability is still very little studied today. Failure modes such as heater breakage, segregation, or the appearance of cavities in the material are shown during endurance tests, but the mechanisms of these failures are not discussed. The objective of this thesis will therefore be to study the failure modes and mechanisms for different operating conditions (endurance, hold, power). The analysis will be carried out through electrical and physical characterizations and accelerated aging methods will be implemented.
Numerical simulation of turbulence models on distorted meshes
Turbulence plays an important role in many industrial applications (flow, heat transfer, chemical reactions). Since Direct Simulation (DNS) is often an excessive cost in computing time, Reynolds Models (RANS) are then used in CFD (computational fluid dynamics) codes. The best known, which was published in the 70s, is the k - epsilon model.
It results in two additional non-linear equations coupled to the Navier-Stokes equations, describing the transport, for one, of turbulent kinetic energy (k) and, for the other, of its dissipation rate (epsilon). ). A very important property to check is the positivity of the parameters k and epsilon which is necessary for the system of equations modeling the turbulence to remain stable. It is therefore crucial that the discretization of these models preserves the monotony. The equations being of convection-diffusion type, it is well known that with classical linear schemes (finite elements, finite volumes, etc ...), the numerical solutions are likely to oscillate on distorted meshes. The negative values of the parameters k and epsilon are then at the origin of the stop of the simulation.
We are interested in nonlinear methods allowing to obtain compact stencils. For diffusion operators, they rely on nonlinear combinations of fluxes on either side of each edge. These approaches have proved their efficiency, especially for the suppression of oscillations on very distorted meshes. We can also take the ideas proposed in the literature where it is for example described nonlinear corrections applying on classical linear schemes. The idea would be to apply this type of method on the diffusive operators appearing in the k-epsilon models. In this context it will also be interesting to transform classical schemes of literature approaching gradients into nonlinear two-point fluxes. Fundamental questions need to be considered in the case of general meshes about the consistency and coercivity of the schemes studied.
During this thesis, we will take the time to solve the basic problems of these methods (first and second year), both on the theoretical aspects and on the computer implementation. This can be done in Castem, TrioCFD or Trust development environments. We will then focus on regular analytical solutions and application cases representative of the community.
Staggered schemes for the Navier-Stokes equations with general meshes
The simulation of the Navier-Stokes equations requires accurate and robust numerical methods that
take into account diffusion operators, gradient and convection terms. Operational approaches have
shown their effectiveness on simplexes. However, in some models or codes
(TrioCF, Flica5), it may be useful to improve the accuracy of solutions locally using an
error estimator or to take into account general meshes. We are here interested in staggered schemes.
This means that the pressure is calculated at the centre of the mesh and the velocities on the edges
(or faces) of the mesh. This results in methods that are naturally accurate at low Mach numbers .
New schemes have recently been presented in this context and have shown their
robustness and accuracy. However, these discretisations can be very costly in terms of memory and
computation time compared with MAC schemes on regular meshes
We are interested in the "gradient" type methods. Some of them are based on a
variational formulation with pressure unknowns at the mesh centres and velocity vector unknowns on
the edges (or faces) of the cells. This approach has been shown to be effective, particularly in terms of
robustness. It should also be noted that an algorithm with the same degrees of freedom as the
MAC methods has been proposed and gives promising results.
The idea would therefore be to combine these two approaches, namely the "gradient" method with the same degrees of freedom as MAC methods. Initially, the focus will be on recovering MAC schemes on regular meshes. Fundamental
questions need to be examined in the case of general meshes: stability, consistency, conditioning of
the system to be inverted, numerical locking. An attempt may also be made to recover the gains in
accuracy using the methods presented in for discretising pressure gradients.
During the course of the thesis, time will be taken to settle the basic problems of this method (first and
second years), both on the theoretical aspects and on the computer implementation. It may be carried
out in the Castem, TrioCFD, Trust or POLYMAC development environments. The focus will be on
application cases that are representative of the community.