Learning world models for advanced autonomous agent

World models are internal representations of the external environment that an agent can use to interact with the real world. They are essential for understanding the physics that govern real-world dynamics, making predictions, and planning long-horizon actions. World models can be used to simulate real-world interactions and enhance the interpretability and explainability of an agent's behavior within this environment, making them key components for advanced autonomous agent models.
Nevertheless, building an accurate world model remains challenging. The goal of this PhD is to develop methodology to learn world models and study their use in the context of autonomous driving, particularly for motion forecasting and developing autonomous agents for navigation.

Scalable thermodynamic computing architectures

Large-scale optimisation problems are increasingly prevalent in industries such as finance, materials development, logistics and artificial intelligence. These algorithms are typically realised on hardware solutions comprising clusters of CPUs and GPUs. However, at scale, this can quickly translate into latencies, energies and financial costs that are not sustainable. Thermodynamic computing is a new computing paradigm in which analogue components are coupled together in a physical network. It promises extremely efficient implementations of algorithms such as simulated annealing, stochastic gradient descent and Markov chain Monte Carlo using the intrinsic physics of the system. However, no clear vision of how a realistic programmable and scalable thermodynamic computer exists. It is this ambitious challenge that will be addressed in this PhD topic. Aspects ranging from the development computing macroblocks, their partitioning and interfacing to a digital system to the adaptation and compilation of algorithms to thermodynamic hardware may be considered. Particular emphasis will be put on understanding the trade-offs required to maximise the scalability and programmability of thermodynamic computers on large-scale optimisation benchmarks and their comparison to implementations on conventional digital hardware.

Multipath-based Cooperative Simultaneous Localization & Mapping through Machine Learning

The goal of this PhD is to explore the potential of machine learning (ML) tools for simultaneous localization and mapping (SLAM) applications, while leveraging multipath radio signals between cooperative wireless devices. The idea is to identify characteristic features of the propagation channels observed over multiple radio links, so as to jointly determine the relative positions of the mobile radio devices, as well as those of scattering objects present in their vicinity. Such radio features typically rely on the arrival times of multipath echos of the transmitted signals. The envisaged approach is expected to benefit from multipath correlation as the radio devices are moving, as well as from spatial diversity and information redundancy through multi-device cooperation. The developed solution will be evaluated on both real measurements collected with integrated Ultra Wideband devices in a reference indoor environment, and synthetic data generated with a Ray-Tracing simulator. Possible applications of this research concern group navigation in complex and/or unknown environments (incl. fleets of drones or robots, firefighters…).

New machine learning methods applied to side-channel attacks

Products secured by embedded cryptographic mechanisms may be vulnerable to side-channel attacks. Such attacks are based on the observation of some physique quantities measured during the device activity, whose variation may provoke information leakage and lead to a security flaw.
Today, such attacks are improved, even in presence of specific countermeasures, by deep learning based methods.
The goal of this thesis is go get familiarity with semi-supervised and self-supervised Learning state-of-the-art and adapt promising methods to the context of the side-channel attacks, in order to improve performances of the attacks in very complex scenarios. A particular attention will be given to attacks against secure implementations of post-quantum cryptographic algorithms.

Towards a better understanding of membrane proteins through AI

Despite the remarkable advances in artificial intelligence (AI), particularly with tools like AlphaFold, the prediction of membrane protein structures remains a major challenge in structural biology. These proteins, which represent 30% of the proteome and 60% of therapeutic targets, are still significantly underrepresented in the Protein Data Bank (PDB), with only 3% of their structures resolved. This rarity is due to the difficulty in maintaining their native state in an amphiphilic environment, which complicates their study, especially with classical structural techniques.

This PhD project aims to overcome these challenges by combining the predictive capabilities of AlphaFold with experimental small-angle scattering (SAXS/SANS) data obtained under physiological conditions. The study will focus on the translocator protein TSPO, a key marker in neuroimaging of several serious pathologies (cancers, neurodegenerative diseases) due to its strong affinity for various pharmacological ligands.

The work will involve predicting the structure of TSPO, both in the presence and absence of ligands, acquiring SAXS/SANS data of the TSPO/amphiphile complex, and refining the models using advanced modeling tools (MolPlay, Chai-1) and molecular dynamics simulations. By deepening the understanding of TSPO’s structure and function, this project could contribute to the design of new ligands for diagnostic and therapeutic purposes.

Online analysis of actinides surrogates in solution by LIBS and AI for nuclear fuel reprocessing processes

The construction of new nuclear reactors in the coming years will require an increase in fuel reprocessing capacity. This evolution requires scientific and technological developments to update process monitoring equipment. One of the parameters to be continuously monitored is the actinide content in solution, which is essential for process control and is currently measured using obsolete technologies. We therefore propose to develop LIBS (laser-induced breakdown spectroscopy) for this application, a technique well suited for quantitative online elemental analysis. As actinide spectra are particularly complex, we shall use multivariate data processing approaches, such as several artificial intelligence (AI) techniques, to extract quantitative information from LIBS data and characterize measurement uncertainty.
The aim of this thesis is therefore to evaluate the performance of online analysis of actinides in solution using LIBS and AI. In particular, we aim to improve the characterisation of uncertainties using machine learning techniques, in order to strongly reduce them and to meet the monitoring needs of the future reprocessing plant.
Experimental work will be carried out on non-radioactive actinide simulants, using a commercial LIBS equipment. The spectroscopic data will drive the data processing part of the thesis, and the determination of the uncertainty obtained by different quantification models.
The results obtained will enable publishing at least 2-3 articles in peer-reviewed journals, and even to file patents. The prospects of the thesis are to increase the maturity level of the method and instrumentation, and gradually move towards implementation on a pilot line representative of a reprocessing process.

Generative AI for Robust Uncertainty Quantification in Astrophysical Inverse Problems

Context
Inverse problems, i.e. estimating underlying signals from corrupted observations, are ubiquitous in astrophysics, and our ability to solve them accurately is critical to the scientific interpretation of the data. Examples of such problems include inferring the distribution of dark matter in the Universe from gravitational lensing effects [1], or component separation in radio interferometric imaging [2].

Thanks to recent deep learning advances, and in particular deep generative modeling techniques (e.g. diffusion models), it now becomes not only possible to get an estimate of the solution of these inverse problems, but to perform Uncertainty Quantification by estimating the full Bayesian posterior of the problem, i.e. having access to all possible solutions that would be allowed by the data, but also plausible under prior knowledge.

Our team has in particular been pioneering such Bayesian methods to combine our knowledge of the physics of the problem, in the form of an explicit likelihood term, with data-driven priors implemented as generative models. This physics-constrained approach ensures that solutions remain compatible with the data and prevents “hallucinations” that typically plague most generative AI applications.

However, despite remarkable progress over the last years, several challenges still remain in the aforementioned framework, and most notably:

[Imperfect or distributionally shifted prior data] Building data-driven priors typically requires having access to examples of non corrupted data, which in many cases do not exist (e.g. all astronomical images are observed with noise and some amount of blurring), or might exist but may have distribution shifts compared to the problems we would like to apply this prior to.
This mismatch can bias estimations and lead to incorrect scientific conclusions. Therefore, the adaptation, or calibration, of data-driven priors from incomplete and noisy observations becomes crucial for working with real data in astrophysical applications.

[Efficient sampling of high dimensional posteriors] Even if the likelihood and the data-driven prior are available, correctly sampling from non-convex multimodal probability distributions in such high-dimensions in an efficient way remains a challenging problem. The most effective methods to date rely on diffusion models, but rely on approximations and can be expensive at inference time to reach accurate estimates of the desired posteriors.

The stringent requirements of scientific applications are a powerful driver for improved methodologies, but beyond the astrophysical scientific context motivating this research, these tools also find broad applicability in many other domains, including medical images [3].

PhD project
The candidate will aim to address these limitations of current methodologies, with the overall aim to make uncertainty quantification for large scale inverse problems faster and more accurate.
As a first direction of research, we will extend recent methodology concurrently developed by our team and our Ciela collaborators [4,5], based on Expectation-Maximization, to iteratively learn (or adapt) diffusion-based priors to data observed under some amount of corruption. This strategy has been shown to be effective at correcting for distribution shifts in the prior (and therefore leading to well calibrated posteriors). However, this approach is still expensive as it requires iteratively solving inverse problems and retraining the diffusion models, and is critically dependent on the quality of the inverse problem solver. We will explore several strategies including variational inference and improved inverse problem sampling strategies to address these issues.
As a second (but connected) direction we will focus on the development of general methodologies for sampling complex posteriors (multimodal/complex geometries) of non-linear inverse problems. Specifically we will investigate strategies based on posterior annealing, inspired from diffusion model sampling, applicable in situations with explicit likelihoods and priors.
Finally, we will apply these methodologies to some challenging and high impact inverse problems in astrophysics, in particular in collaboration with our colleagues from the Ciela institute, we will aim to improve source and lens reconstruction of strong gravitational lensing systems.
Publications in top machine learning conferences are expected (NeurIPS, ICML), as well as publications of the applications of these methodologies in astrophysical journals.

References
[1] Benjamin Remy, Francois Lanusse, Niall Jeffrey, Jia Liu, Jean-Luc Starck, Ken Osato, Tim Schrabback, Probabilistic Mass Mapping with Neural Score Estimation, https://www.aanda.org/articles/aa/abs/2023/04/aa43054-22/aa43054-22.html

[2] Tobías I Liaudat, Matthijs Mars, Matthew A Price, Marcelo Pereyra, Marta M Betcke, Jason D McEwen, Scalable Bayesian uncertainty quantification with data-driven priors for radio interferometric imaging, RAS Techniques and Instruments, Volume 3, Issue 1, January 2024, Pages 505–534, https://doi.org/10.1093/rasti/rzae030

[3] Zaccharie Ramzi, Benjamin Remy, Francois Lanusse, Jean-Luc Starck, Philippe Ciuciu, Denoising Score-Matching for Uncertainty Quantification in Inverse Problems, https://arxiv.org/abs/2011.08698

[4] François Rozet, Gérôme Andry, François Lanusse, Gilles Louppe, Learning Diffusion Priors from Observations by Expectation Maximization, NeurIPS 2024, https://arxiv.org/abs/2405.13712

[5] Gabriel Missael Barco, Alexandre Adam, Connor Stone, Yashar Hezaveh, Laurence Perreault-Levasseur, Tackling the Problem of Distributional Shifts: Correcting Misspecified, High-Dimensional Data-Driven Priors for Inverse Problems, https://arxiv.org/abs/2407.17667

Prediction of Soiling on PV modules/systems through Real-World Environment Modeling and Data Fusion

Photovoltaic (PV) systems, particularly those installed in regions prone to soiling such as arid areas, coastal sites, and agricultural zones, can experience energy losses of up to 20–30% annually. These losses translate to financial impacts exceeding €10 billion in 2023.
This thesis aims to develop a robust and comprehensive method to predict soiling accumulation on PV modules and systems by combining real-world environmental modeling with operational PV data (electrical, thermal, optical). The research will follow a bottom-up approach in three stages:

1. Component/Module Level: Reproduction and modeling of soiling accumulation in laboratory conditions, followed by experimental validation. This stage will leverage the CEA’s expertise in degradation modeling, including accelerated testing.

2. Module/System Level: Implementation of monitoring campaigns to collect meteorological, operational, and imaging data, combined with field soiling tests on a pilot site. The data will validate and enhance CEA diagnostic tools by introducing innovative features such as AI-driven soiling propagation prediction.

3. System/Operational Level: Validation of the proposed method on commercial PV modules in utility-scale PV plants, aiming to demonstrate scalability and real-world applicability.

The outcomes of this thesis will contribute to the development of an innovative tool/method for comprehensive soiling diagnostics and prognostics in PV installations, enabling the minimization of energy losses while anticipating and optimizing cleaning strategies for PV plants.

Learning Interpretable Models for Stress Corrosion of Stainless Steels Exposed in the Primary Environment of PWRs

Stress corrosion cracking (SCC) of austenitic alloys in water-cooled nuclear reactors is one of the most significant component degradation phenomena. SCC occurs due to the synergistic effects of tensile stresses, environment and material susceptibility. For reactor life extension, understanding this mechanism is essential. The methodology most frequently employed to investigate SCC cracking is an experimental one, requiring lengthy and costly tests of several thousand hours. Furthermore, the considerable number of critical parameters that influence susceptibility to SCC cracking and coupling effects have resulted in test grids increasing in length and complexity. This thesis proposes a novel approach based on the use of interpretable models that are driven by the artificial intelligence of fuzzy logic. The aim is to reduce the length and cost of research activities by focusing on relevant tests and parameters that can improve environmental performance. The key issues here will be to add the performance of artificial intelligence to the experimental approach, with the aim of defining susceptibility domains for the initiation of SCC cracks as a function of the critical parameters identified in the model, and providing data for the development of new materials by additive manufacturing. The thesis will develop a numerical model that can be used as guidance in decision-making regarding the stress corrosion mechanism. The future PhD student will also carry out experimental work to validate this new numerical approach.

Scalable NoC-based Programmable Cluster Architecture for future AI applications

Context
Artificial Intelligence (AI) has emerged as a major field impacting various sectors, including healthcare, automotive, robotics, and more. Hardware architectures must now meet increasingly demanding requirements in terms of computational power, low latency, and flexibility. Network-on-Chip (NoC) technology is a key enabler in addressing these challenges, providing efficient and scalable interconnections within multiprocessor systems. However, despite its benefits, designing NoCs poses significant challenges, particularly in optimizing latency, energy consumption, and scalability.
Programmable cluster architectures hold great promise for AI as they enable resource adaptation to meet the specific needs of deep learning algorithms and other compute-intensive AI applications. By combining the modularity of clusters with the advantages of NoCs, it becomes possible to design systems capable of handling ever-increasing AI workloads while ensuring maximum energy efficiency and flexibility.
Summary of the Thesis Topic
This PhD project aims to design a scalable, programmable cluster architecture based on a Network-on-Chip tailored for future AI applications. The primary objective will be to design and optimize a NoC architecture capable of meeting the high demands of AI applications in terms of intensive computing and efficient data transfer between processing clusters.
The research will focus on the following key areas:
1. NoC Architecture Design: Developing a scalable and programmable NoC to effectively connect various AI processing clusters.
2. Performance and Energy Efficiency Optimization: Defining mechanisms to optimize system latency and energy consumption based on the nature of AI workloads.
3. Cluster Flexibility and Programmability: Proposing a modular and programmable architecture that dynamically allocates resources based on the specific needs of each AI application.
4. Experimental Evaluation: Implementing and testing prototypes of the proposed architecture to validate its performance on real-world use cases, such as image classification, object detection, and real-time data processing.
The outcomes of this research may contribute to the development of cutting-edge embedded systems and AI solutions optimized for the next generation of AI applications and algorithms.

The work performed during this thesis will be presented at international conferences and scientific journals. Certain results may be patented.