Themes: Mathematics, Informatic
Thesis location: Explosion and Fire Laboratory (LIE) - Cadarache (13)
Start: October 2021
Master's degree in computer science or statistics or data science
Age limit: 26 years old unless otherwise stated.
The research carried out at IRSN aims to better describe the physical phenomena that can lead to an accident and to assess the consequences. One of the results of this research activity is the development of softwares to simulate the evolution of accident scenarios. These softwares are partly based on a "mechanistic" approach that is to say on a precise mathematical formalization of the physical processes involved. The operational use of modeling tools in the case of a safety assessment relies in large part on the ability to be able to feed the software with data, the time required to perform the calculations and the performance and quality of the predictions. IRSN recently engaged in the design of new artificial intelligence software tools, the purpose of which is to make the most of the knowledge conveyed through business software, to help with the diagnosis of the scenario, the prognosis of its evolution or the important elements to control. The general idea is to build large databases resulting from millions of numerical simulations making it possible to cover a wide range of initial conditions and limits that may be encountered during an accident. The artificial intelligence software then uses this data dynamically to identify the configurations and changes compatible with the known data. IRSN has thus developed two expert systems using the Bayesian network technique as part of fire safety studies. The theoretical objective of this PhD thesis will be to conduct research on the methods used to generate databases, as well as on the most suitable probabilistic causal models and algorithms to exploit these databases. The work already carried out within the institute has made it possible to define a methodology for the creation and dynamic operation of large databases and to confirm the feasibility in terms of computer engineering. However, it reveals theoretical difficulties of Bayesian networks to model complex systems. Indeed, Bayesian networks impose a precise and complete description of the studied system, which makes difficult a generic design for which only part of the relational information between the components of the system can be established. Furthermore, since these components interact with each other over time, they cannot be modeled completely independently. The operational objective of this PhD thesis will be to integrate the results of this research in order to complete an existing expert system with "satellite" expert systems and to allow the propagation and aggregation of information between several expert systems. These "satellite" expert systems would, for example, integrate the geometry and specific features of the installations that one wishes to study.