Frugal and Explainable Open-World Object Detection through Visual Concept Discovery and Transfer
Unlike classical detectors trained on a closed set of categories, and VLMs (vision-language models) that are heavy, non-explainable, and limited to a fixed vocabulary, this thesis proposes an explainable and frugal approach to open-world object detection. Indeed, object detectors in an open world are capable not only of recognizing known objects, but also of signaling unknown or unexpected objects.
The thesis project relies on using VLMs solely as offline teachers during training in order to distill their knowledge into a lightweight, real-time detector. By leveraging learning and reasoning from a hierarchy of shared visual concepts (covering appearance, attributes, and semantic families), the proposed model will be able not only to recognize known objects, but also to signal and characterize unknown objects without prior prompts. This approach will thus provide visual explanations directly grounded in the image, meeting the demands for robustness, transparency, and frugality essential in applications across numerous safety-critical domains (mobility, defense and security, healthcare, and industrial inspection).
Development of mid-infrared sensors and Fourier-transform spectrometers in silicon integrated photonics
Fourier-transform infrared (FTIR) spectrometers have become a reference tool for chemical analysis. One of the next challenges is to integrate them on photonic integrated circuits (PICs), using silicon platforms optimized for the mid-infrared (MIR), in order to improve compactness and scalability. You will simulate key MIR PIC components, with a focus on photonic and thermo-optic modeling, design photolithography masks, follow up on fabrication, and characterize the FTIR spectrometers and sensors. Sensing applications will primarily target the detection and identification of greenhouse gases, but may also extend to plastic pollution and biosensing. Although dedicated engineers will be in charge of silicon wafer fabrication in our cleanrooms, you will be involved in defining specifications and in monitoring in-line quality control to ensure that the fabricated devices perform as intended. You will be responsible for characterizing the FTIR and sensor devices at wafer and chip level on custom mid-infrared optoelectronic test benches. A first generation of non-optimized devices will be available upon your arrival. You will be part of the team upgrading the characterization tools to enable faster, more accurate automated measurements. You will analyze and interpret experimental data to deepen our understanding of the advantages and limitations of integrated FTIR spectrometers compared with their bulkier free-space counterparts. You will be expected to publish research papers and present at international conferences.
Integrated superconducting nanowire single photon detecors on SiN waveguides with photon number resolution capability for photonic quantum computing
Among the various qubit technologies, photonic qubits represent a highly promising approach to the development of quantum computers capable of solving problems beyond the reach of the most powerful supercomputers. The core of photonic quantum computing processors relies on the generation, manipulation, and detection of single photons. In this context, this PhD project at CEA-LETI aims to develop a new generation of high-performance single-photon detectors, offering not only high detection efficiency and low noise, but also photon-number-resolving capabilities, i.e., the ability to discriminate between 0, 1, 2, or more incident photons. This functionality is essential for implementing quantum error correction codes and ultimately enabling the transition from physical qubits to logical qubits. The research will focus on the design of innovative detector circuit architectures, the fabrication of integrated prototypes in CEA-LETI's 200 mm cleanroom, and their experimental characterization at cryogenic temperatures. This work will contribute to the development of a strategic technology for next-generation photonic quantum computers.
Multi-level modeling and accelerated simulation of entropy sources for pre-silicon analysis of TRNGs
This topic focuses on the modeling and accelerated simulation of entropy sources used in True Random Number Generators (TRNGs). The candidate will study how to preserve the precision of analog, netlist, and post-layout simulations while greatly accelerating their execution. The goal is to develop intermediate models that incorporate key physical phenomena such as noise, jitter, variability, coupling effects, and layout dependencies. These models will be based on theoretical foundations from the state of the art and validated against full analog simulations.
An experimental component will also allow the candidate to validate these models using real-world measurements of entropy sources, particularly with the VASCO component. The candidate will aim to link the physical parameters of the circuit to the statistical properties and entropy of the generated sequences.
The work contributes to improving pre-silicon analysis, design, and evaluation of TRNGs. Additionally, the developed approaches may be extended to Ring Oscillator-based Physical Unclonable Functions (PUFs).
Certifying Event Non-Occurrence in Video Large Language Models
The reliability of artificial intelligence models capable of analyzing videos (Video Large Language Models) is a major challenge, particularly in application domains where misinterpreting visual information can lead to inappropriate decisions or generate misinformation.
Currently, these models tend to confuse what is actually visible in a video with what is merely plausible or expected. As a result, they may claim that an action has been completed even when the video does not provide visual evidence that it actually occurred.
The main objective of this thesis is to equip these models with the ability to distinguish between three situations: (1) an event has occurred, (2) an event did not occur (supported by sufficient visual evidence), or (3) the available information is insufficient to reach a conclusion.
Three complementary research directions will be explored:
· The first aims to develop methods that teach models to identify the visual cues that are truly decisive in confirming or refuting an action.
· The second focuses on developing reasoning mechanisms capable of identifying the observations that should necessarily be visible if an event had occurred. This approach will enable the generation of "non-occurrence certificates", i.e., structured justifications based on visual evidence supporting the conclusion that an event did not take place.
· The third investigates whether using these certificates as additional supervision can improve model performance while reducing the amount of training data required.
The proposed approaches will be evaluated in terms of accuracy, robustness to ambiguous situations, and data efficiency. Ultimately, this research aims to develop more reliable video models capable of justifying their conclusions rather than simply predicting what seems likely.
Generative AI for the Design of Next-Generation Power Converters, End-of-Life Guaranteed: Accelerated Topology Discovery and Multi-Criteria Optimization Integrating Ageing and Eco-Design
Next-generation wide-bandgap (SiC/GaN) power converters are at the heart of the electrification of our economy, yet their ageing and environmental impact are only assessed a posteriori, on a frozen design: this leads either to costly oversizing margins or, conversely, to nominally optimal designs that no longer meet their specifications at end of life. This thesis aims to remove this obstacle by establishing a closed-loop, generative-AI-based design loop in which end-of-life specification compliance and environmental impact per kWh converted become first-rank objectives, on a par with efficiency, volume, mass, EMC and consumption. The method couples a topology generation constrained by a versioned electrotechnical ontology, frontier LLM models, a propagation of multi-component drift (R_DS(on) and V_th of GaNs, ESR and capacitance of capacitors) up to system-level performance, recalibrated on accelerated tests, a multi-fidelity evaluation (physics-informed surrogates, simulation tools), and a multi-objective Bayesian optimization under probabilistic constraints, integrating LCA impact through a functional unit that depends on the predicted lifetime. The work will be validated experimentally on a converter prototype measured both new and aged, with explainability of the trade-offs. It capitalizes on the CEA's local generative-AI infrastructure (AITIANA Studio), aiming for measurable gains in reliability and environmental footprint, a reusable method, and open data.
Human–humanoid robot object handover based on shared tactile perception: contact modelling, learning from human data and trust dynamics
Humanoid robots are starting to be tested on factory floors, where they can take over workstations designed for people. To work alongside an operator, they must be able to hand over a part or a tool, and take it back. The gesture looks simple, but it lasts about half a second and relies on touch: the giver loosens their grip at the pace the receiver tightens theirs. Current robot hands measure contact force well, but not what matters here: the gradual shift of the load from one hand to the other and the onset of slip, especially with oily or unbalanced parts.
The thesis proposes to give the robot a finer sense of touch, able to follow these phenomena during the handover. The same device will be fitted to the robot’s hand and to a glove worn by the operator. The PhD student will thus be able to record handovers between people, then use these data to teach the robot when to let go. Experiments with volunteers will then measure whether this robot inspires justified trust, neither excessive nor insufficient.
The expected results are a tactile sensor prototype, a release-decision method integrated into a humanoid robot, an open dataset of human handovers and a method for measuring trust. The work takes place at CEA List and combines sensors, robotics, machine learning and human factors.
Acoustic Interrogation of Passive Sensors Through Metal Walls
This PhD topic addresses critical challenges in the nuclear sector, such as monitoring reactors, fuels, and waste, by eliminating risks associated with drilling metal walls (leaks, contamination) while ensuring resistance to extreme environments (radiation, high temperature, pressure). It also offers applications in the medical field, particularly for implantable devices and non-invasive diagnostics. The primary objective is to develop passive acoustic telemetry systems capable of measuring physical parameters behind a metal wall without requiring electronics on the inaccessible side. Inspired by radio frequency (RF) techniques, this approach aims to overcome the limitations of current sensors, which often cannot penetrate metal barriers or require embedded electronics. The work will build on the expertise of the Systems Department in acoustic modeling, piezoelectric materials, and frequency transposition techniques, while fostering cross-disciplinary collaboration across various CEA entities, particularly in RF and MEMS domains. The expected results include novel passive acoustic telemetry techniques and associated design methods, providing a robust and innovative alternative to existing sensors. Structured in three phases (literature review, development, and experimentation), this thesis is part of a collaborative effort within the CEA, with concrete application potential in nuclear, medical, and industrial fields.
Evaluation of new low-damage GaN etching strategies for the realization of high performance MicroLED
For optical communication and display applications (display, AR/VR), the need for micro-displays with high resolution µLEDs GaN is becoming increasingly important. The formation of the pixel constitutes the critical brick because it generates defects on the sides of the µLED, which limit the emission performance of the device. These degradations have even more impacts on electro-optical performance for aggressive LED pitches.
The subject proposes a comprehensive approach, ranging from the development of new GaN etching strategies limiting damage to the pixel edges, to its characterization through dedicated direct means, and validated on a complete device by an electro-optical test vehicle.
Backpropagation-Free Langevin Sampling for Bayesian Learning on ReRAM Architectures
Bayesian neural networks allow artificial intelligence systems to account for uncertainty in their predictions, but training them requires substantial computational resources. Resistive memories (ReRAM), which can both store data and perform computations directly within electronic circuits, offer a promising solution to reduce this cost.
This PhD project aims to develop a new Bayesian learning method on ReRAM hardware, based on Langevin sampling. This approach efficiently explores neural network parameters by combining random perturbations with gradient estimates. The objective is to exploit the intrinsic physical variability of ReRAM devices to estimate these gradients without relying on conventional backpropagation. The possibility of reusing memory programming attempts will also be investigated to reduce energy consumption and device wear.
The work will combine algorithm development, simulations, and experiments on existing ReRAM circuits at CEA-Leti. Expected outcomes include a demonstration of Bayesian learning on real hardware and an assessment of its performance, accuracy, and energy efficiency compared with existing approaches.