2D and 3D analysis improvements with machine learning for muography applications

Transmission muography is a non-invasive and non-destructive imaging method which allows to estimate the integrated density of a volume in a given direction (also referred as opacity). It is used in multiple societal applications like archaeology, nuclear safety or geoscience. It relies on the reconstruction of muon tracks that crossed the studied volume compared to the corresponding open sky expectation. The portable experimental setups developed by CEA/Irfu group operates four Micromegas gaseous detectors, HV and DAQ modules, and an embedded computer allowing remote control. Used Micromegas detectors have a multiplexed readout to optimize the DAQ system, while keeping good spatial resolution. However, the natural muon flux is relatively low, so muography images could have high levels of statistical noise, which would be propagated to the reconstructed 3D images. For this reason, we propose three new methods, using machine learning, which increase significantly the quality of 2D images and 3D tomographies. We developed a new demultiplexing method for the Micromegas. It showed its efficiency both for 1D and 2D multiplexed detectors, in a hodoscopic tracker and a Time Projection Chamber (TPC). In the TPC, this method can embed a particle identification between muon-hits and electron-hits. We demonstrated how neural networks could denoise muography images. In the context of the scanning of a nuclear reactor, we used data augmentation, modeling few hundreds fake nuclear reactors. Muographies from different points of view were simulated to train a denoising neural network. These new methods significantly improved the muography images. Nonetheless, the used 3D tomography algorithm still has some limitations. We proved that it was possible to build a 3D post-process neural network, which was trained to compensate the reconstruction algorithm’s limitations.