This thesis describes improvements in muon imaging analysis in the scope of a characterization of the G3 nuclear reactor located at CEA-Marcoule (France). Due to the size and composition of the reactor, muography is the only non invasive method capable of observing inside G3's airtight concrete chamber. Other techniques, as X-ray tomography, are not penetrating enough and present other deployment constraints, as radiological ones.
Muography images were taken under the G3 reactor in order to observe its internal parts from 46 different points of view. The data-taking lasted 1 year and used 4 muon telescopes built at CEA/Irfu.
To analyze this data, new developments were made at multiple steps of the analysis pipeline. A new demultiplexing was developed for the Micromegas muon detectors, based on neural networks trained with detector simulations. The muography images obtained were analyzed in order to estimate the opacity of the reactor. The opacity images were then denoised, using a reverse diffusion neural network, which make possible to decrease the acquisition duration.
A 3D tomography was computed using the opacity images and allowed reconstructing the inner parts of the reactor. The tomography underwent a post-process by a neural network to improve its shape. Finally, the opacities and the tomography highlighted differences between the reactor model and its actual status, mainly in the loading block and the graphite core.
