Mikaël Jacquemont – Phd’s defense – nov. 26, 2021

Gamma-ray astronomy is the astronomical observation of the most energetic photons produced by violent astrophysical phenomena. Their study allows understanding better the physics ruling the birth of stars or the evolution of the galaxies. It also allows exploring a new physics. Ground-based Cherenkov telescopes detect indirectly gamma rays via the particle shower they generate when entering the atmosphere. The purpose of the image analysis is to estimate the energy and direction of the primary particle and to separate the gamma rays from the cosmic ray background. This step is complex because cosmic rays can generate very similar images and the signal-to-noise ratio is typically lower than 1/1000.The Cherenkov Telescope Array (CTA) is the next generation of very high-energy universe observatories. Composed of more than 100 telescopes distributed on 2 sites, CTA will enhance the sensitivity by a factor of 10 compared to the current observatories, while improving the accuracy of the reconstruction. As a counterpart, once built, CTA will generate a tremendous amount of data (210 PB / year) to be analyzed. 

Moreover, all the data will be reprocessed yearly to benefit from the improvement of analysis models. Due to these properties of CTA, standard analysis methods either are too slow or lack sensitivity at low energies. We need then to explore other methods. Following the recent advances of artificial neural networks, this thesis proposes a new deep learning approach to analyze CTA data, especially the data from the Large-Sized Telescope 1 (LST1), the first on-site prototype built.The first contribution of this thesis is an original method to apply the deep learning techniques to any kind of pixel organization. It provides convolution and pooling operators, that were crucial for neural networks successes. This method respects the real neighborhood of the pixels, and avoids preprocessing and image size increase. This work demonstrates its interest over standard resampling methods for the hexagonal pixel images of the LST1. 

The second and main contribution of this thesis is a physically inspired deep multi-task architecture, γ-PhysNet, to perform full event reconstruction. It benefits from an attention mechanism enhancing its robustness to initialization conditions. Its evaluation on the simulated data used to prepare analysis algorithms shows that γ-PhysNet has a significantly better sensitivity and spatial resolution than a standard method relying on the Hillas parameter extraction and a multivariate method, such as the Random Forest. In particular, it achieves very interesting sensitivity below 200,GeV, and could enhance the study of transient phenomena. γ-PhysNet is also 800 times faster than the state-of-the-art method. Besides, using a visual explanation method, this thesis presents a preliminary analysis of the model proposed to understand better its behavior. Finally, the first observation data from the “Crab” produced by the LST1, still in commissioning phase, are analyzed with γ-PhysNet. This very preliminary analysis highlights the need for a better adaptation to real data. Finally, the last contribution is a framework to ease the deep learning procedure with CTA data, and to ensure the traceability and reproducibility of the experiments.