Natural language interactions for anomaly detection in mono and multi-variate time series using fondation models and retrieval augmented generation
Anomaly detection in mono and multi-variate time series highly depends on the context of the task. State-of-the-art approaches rely usually on two main approaches: first extensive data acquisition is sought to train artificial intelligence models such as auto-encoders, able to learn useful latent reprensations able to isolate abnormality from expected system behaviors; a second approach consists in careful features construction based on a combination of expert knowledge and artificial intelligence expert to isolate anomalies from normal behaviors using limited examples. An extensive analysis of the literature shows that anomaly detection refer to an ambiguous definition, because a given pattern in time series could appear as normal or abnormal depending on the application domain and the immediate context within the successive observed data points. Fondation models and retrieval-augmented generation has the potential to substantially modify anomaly detection approaches. The rationale is that domain expert, through natural language interactions, could be able to specify system behavior normality and/or abnormality, and a joint indexing of state-of-the-art literature and time series embedding could guide this domain expert to define a carefully crafted algorithm.
Tool supported model completion with support for design pattern application
Generative AI and large language models (LLMs), such Copilot and ChatGPT can complete code based on initial fragments written by a developer. They are integrated in software development environments such as VS code. Many papers analyse the advantages and limitations of these approaches for code generation, see for instance Besides some deficiencies, the produced code is often correct and the results that are getting increasingly better.
However, a surprisingly small amount of work has been done in the context of completion software models (for instance based on UML). The paper concludes that while the performance of the current LLMs for software modeling is still limited (in contrast to code generation), there is a need that (in contrast to code generation) we should adapt our model-based engineering practices to these new assistants and integrate these into MBSE methods and tools.
The integration of design-patterns is a complementary part of this work. Originally coming from building architecture, the term design patterns has been adopted in the software domain to capture a proven solution for a given problem along with its advantages and disadvantages. A bit later, the term anti-pattern has been proposed to identify patterns that are known not to work or having severe disadvantages. Thus, when proposing a completion, then assistant could explicitly reference an existing design pattern with its implications. The completion proposal can be based either on identified model fragments (including modeled requirements) or an explicit pattern selection. This thesis will explore the state-of-the-art of model completion with AI and design patterns and associated tool support. Up to now, little work is available on pattern formalization and the use in model tools. It will propose to identify the modelers intention, based on partial models. The task could be rule-based but should also explore machine-learning approaches. Implement a completion proposal in the context of a design tool, notably Papyrus SW designer. The solution will be evaluated.
AI-assisted generation of Instruction Set Simulators
The simulation tools for digital architectures rely on various types of models with different levels of abstraction to meet the requirements of hardware/software co-design and co-validation. Among these models, higher-level ones enable rapid functional validation of software on target architectures.
Developing these functional models often involves a manual process, which is both tedious and error-prone. When low-level RTL (Register Transfer Level) descriptions are available, they serve as a foundation for deriving higher-level models, such as functional ones. Preliminary work at CEA has resulted in an initial prototype based on MLIR (Multi-Level Intermediate Representation), demonstrating promising results in generating instruction execution functions from RTL descriptions.
The goal of this thesis is to further explore these initial efforts and subsequently automate the extraction of architectural states, leveraging the latest advancements in machine learning for EDA. The expected result is a comprehensive workflow for the automatic generation of functional simulators (a.k.a Instruction Set Simulators) from RTL, ensuring by construction the semantic consistency between the two abstraction levels.
Software and hardware acceleration of Neural Fields in autonomous robotics
Since 2020, Neural Radiance Fields, or NeRFs, have been the focus of intense interest in the scientific community for their ability to implicitly reconstruct 3D and synthesize new points of view of a scene from a limited set of images. Recent scientific advances have drastically improved initial performance (reduction in data requirements, memory needs and processing speed), paving the way for new uses of these networks, particularly in embedded applications, or for new purposes.
This thesis therefore focuses on the use of these networks for autonomous robotic navigation, with the embedded constraints involved: power consumption, limited computing and memorization hardware resources, etc. The navigation context will involve extending work already underway on incremental versions of these neural networks.
The student will be in charge of proposing and designing innovative algorithmic, software and hardware mechanisms enabling the execution of NeRFs in real time for autonomous robotic navigation.
Combination study of high throughput screening techniques and artificial intelligence (AI) to identify innovative materials for next generation of battery
In recent years, the CEA has set up an experimental high-throughput screening (HTS) activity for lithium battery materials, based on combinatorial synthesis by sputtering and various high-throughput characterisation techniques on large substrates (typically 4 inches). Optimisation of material compositions is traditionally carried out by analysing experimental designs. In the framework of this thesis, we propose to compare the results of this conventional method with the Artificial Intelligence tools developed at CEA-LIST (symbolic AI) and CEA-CTREG (connectionist AI). The objetive is to demonstrate that AI can advantageously replace standard experimental design in order to offer an innovative, high-performance high-throughput screening tool.
Benefits of Physics-Informed Neural Networks methods for complex problems in physics applications
Strategies for approximating the solutions of partial differential equations by means of Neural Networks have recently gained in popularity. These so-called Physics-Informed Neural Networks (PINNs) are issued from recent advances in the field of Artificial Intelligence and bring a new paradigm compared to conventional numerical methods such as Finite Volume or Finite Element Methods. The core of the method consists in enforcing the physical model thanks to the loss function by minimizing the residual of the operators. Although these methods show promising results on academic problems, they also bring many specific questions regarding their benefits for complex applications and their mathematical properties. The present Ph.D proposal aims at studying both aspects.
The candidate will first perform a state of the art study in order to understand the PINNs approach and their potential as a industrial grade simulation method. We propose then to focus on several problems involving different types of complexity issued from physical processes applications like the two-phase flows or the coupling of neutronics and thermal-hydraulics.
Electronic structure calculation with deep learning models
Ab initio simulations with Density Functional Theory (DFT) are now routinely employed across scientific disciplines to unravel the intricate electronic characteristics and properties of materials at the atomic level. Over the past decade, deep learning has revolutionized multiple areas such as computer vision, natural language processing, healthcare diagnostics, and autonomous systems. The combination of these two fields presents a promising avenue to enhance the accuracy and efficiency of complex materials properties predictions, bridging the gap between quantum-level understanding and data-driven insights for accelerated scientific discovery and innovation. Many efforts have been devoted to build deep learning interatomic potentials that learn the potential energy surface (PES) from DFT simulations and can be employed in large-scale molecular dynamics (MD) simulations. Generalizing such deep learning approaches to predict the electronic structure instead of just the energy, forces and stress tensor of a system is an appealing idea as it would open up new frontiers in materials research, enabling the simulation of electron-related physical properties in large systems that are important for microelectronic applications. The goal of this PhD is to develop new methodologies relying on equivariant neural networks to predict the DFT Hamiltonian (i.e. the most fundamental property) of complex materials (including disorder, defects, interfaces, etc.) or heterostructures.
Generative artificial intelligence algorithms for understanding and countering online polarization
Digital platforms enable the widespread dissemination of information, but their engagement-centric business models often promote the spread of ideologically homogeneous or controversial political content. These models can lead to the polarization of political opinions and impede the healthy functioning of democratic systems. The PhD will investigate innovative generative AI models devised for a deep understanding of political polarization and for countering its effects. It will mobilize several areas of AI: generative learning, frugal AI, continual learning, and multimedia learning. Advances will be associated with the following challenges:
-the modeling of political polarization, and the translation of the obtained domain model into actionable implementation requirements that will be used as inputs of AI algorithms;
-the curation of massive and diversified multimodal political data to ensure topical and temporal coverage, and to map these data to a common semantic representation space;
-the training of politics-oriented generative models to encode relevant knowledge effectively and efficiently and to generate labeled training data for downstream tasks;
-the specialization of the models for the specific tasks needed for a fine-grained understanding of polarization (topic detection, entity recognition, sentiment analysis);
-the continual update of the politics-oriented generative models and polarization-specific tasks to keep pace with the evolution of political events and news.
Modeling and ALARA optimization of maintenance operations in fusion nuclear power plants with Artificial Intelligence and Virtual Reality techniques
In view to the development of future fusion reactors, the maintenance operations in these nuclear facilities will be a diffculty, as part of them will have to be carried out hands-on. Safety rules govern interventions in a radioactive environment. They take into account the level of effective dose received by the operator, a factor that characterizes the risk to which the operator is exposed (dose depending on ambient dose rate and time).
In the aim of optimizing this dose in line with the ALARA principle and the safety constraints associated with these installations, the prior simulation of operations in Virtual Reality is an asset in terms of design optimization and worker training. Calculating dose during these simulations would be an important contribution to discriminating between different options. The simulation methods currently used to calculate dose rates are in some cases imprecise and in others very costly in terms of simulation time.
The aim of this work is to propose a new method for dynamic dose rate estimation in reduced time (or even real time) as a function of the movements of both the activation sources of a fusion installation, the maintenance operator and the shield protecting the latter. These dynamic configurations are representative of real intervention conditions. This method will implement Artificial Intelligence techniques coupled with Neutronics methods, and should be able to be integrated into a Virtual Reality tool based on existing development environments such as Unity3D.
Anomaly Detection Machine learning Methods for Experimental Plasma Fusion Data Quality - Application to WEST Data
Fusion plasmas in tokamaks have complex non-linear dynamics. In the WEST Tokamak, of the same family as the ITER project, a large amount of heterogeneous experimental fusion data is collected. Ensuring the integrity and quality of this data in real time is essential for the stable and safe operation of the Tokamak. Continuous monitoring and validation are essential, as any disturbance or anomaly can significantly affect our ability to ensure plasma stability, control performance and even lifetime. The detection of unusual patterns or events within the collected data can provide valuable insights and help identify potentially abnormal behavior in plasma operations.
This Ph.D. research aims to study and develop anomaly detection system for WEST -- prefiguring what could be installed on ITER -- by integrate machine learning algorithms, statistical methodologies and signal processing techniques to validate various diagnostic signals in Tokamak operations, including density, interferometry, radiative power and magnetic data.
The expected outcomes are:
– The development of dedicated machine-learning algorithms capable of detecting anomalies in selected time series data from WEST Tokamak.
– The fine-tuning of an operational autonomous system able to ensure data quality in Tokamak reactors, integrated into the WEST AI platform.
– The constitution of a comprehensive database.
– The validation of a data quality framework built for the specific needs of plasma fusion research.