Thesis of Harizo Rajaona

Adaptive Bayesian inference for source reconstruction in atmospheric dispersion

In atmospheric physics, reconstructing a pollution source is a challenging but important question : it provides better input parameters to dispersion models, and gives useful information to first-responder teams in case of an accidental toxic release. Various methods already exist, but using them requires an important amount of computational resources, especially as the accuracy of the dispersion model increases. A minimal degree of precision for these models remains necessary, particularly in urban scenarios where the presence of obstacles and the unstationary meteorology have to be taken into account. One has also to account for all factors of uncertainty, from the observations and for the estimation. The topic of this thesis is the construction of a source term estimation method based on adaptive Bayesian inference and Monte Carlo methods. First, we describe the context of the problem and the existing methods. Next, we go into more details on the Bayesian formulation, focusing on adaptive importance sampling methods, especially on the AMIS algorithm. The third chapter presents an application of the AMIS to an experimental case study, and illustrates the mechanisms behind the estimation process that provides the source parameters’ posterior density. Finally, the fourth chapter underlines an improvement of how the dispersion computations can be processed, thus allowing a considerable gain in computation time, and giving room for using a more complex dispersion model on both rural and urban use cases.


- Directeur(s) de thèse : - François SEPTIER (Télécom Lille / CRIStAL) - Rapporteurs : - Thierry CHONAVEL (Télécom Bretagne) - Hichem SNOUSSI (Université de Technologie de Troyes) - Examinateurs : - Lionel SOULHAC (Ecole Centrale de Lyon) - Patrick ARMAND (CEA) - Patrick BAS (Ecole Centrale de Lille / CRIStAL)

Thesis of the team SIGMA defended on 21/11/2016