Electron Identification with Neural Networks and Measurement of the TeV Cosmic Electrons Flux with the DAMPE Experiment

When one thinks about the cosmos, one imagines the vast structures that populate the night sky: constellations, galaxies, clusters and many others. Yet, other curious, much smaller objects inhabit this dark canvas: cosmic rays. Sort of a space radioactivity, they are highly energetic subatomic particles that reach us from every corner of the sky. Despite a century of studies and experiments, we still do not know much about them. The nature of their sources and of the mechanisms that accelerate them is, for example, still largely unknown. In the past few decades, highly precise experiments have provided a wealth of new observations and solid hints towards the answers we're seeking. Yet some of these observations caused a lot of turmoil. It indeed turns out that there are more high energy electrons than expected, and the same holds true for their antimatter partner, the positron. Is this caused by new cosmic accelerators? By the medium they traverse? Or by the still-elusive dark matter, a highly sought substance outweighing familiar matter in the universe by a factor five? Seeking answers to these questions among others, the Dark Matter Particle Explorer (DAMPE) rose to space in December 2015. This satellite-based system of four detectors working as one was designed to push the boundaries of cosmic ray physics up to energies that had never been studied directly. After a mere one and a half year of operations DAMPE published its first results, the direct detection of a break in the cosmic electron+positron spectrum, triggering a lot of discussions within the astroparticle physics community. Yet this observation was but a mere preview of the raw potential of DAMPE, designed to push electron measurements beyond energies of 10 TeV. Doing so, however, requires both a larger amount of data and a more powerful analysis tool. Which are the goals of the present thesis. We propose a tool coming from the field of data sciences, and more precisely of deep learning. Using billions of Monte Carlo simulated events, created over years of continuous usage of a thousand of CPUs, we train a four-layers deep neural network for the task of electron identification and proton rejection. After a careful and extensive optimisation campaign, the trained network reaches excellent performances, notably achieving at highest energies up to eight times better background rejection than conventional methods for the same signal efficiency. When turning to real data, the network shows consistent outputs and distributions with simulations, thus demonstrating its reliability. The combination of five years of data with the deep learning classifier allows to extend the cosmic electron measurement beyond 10 TeV. For that purpose, we carry an analysis with first and foremost a set of selection criteria to clean up DAMPE data and remove detected events that either cannot be exploited or are obviously not electrons. Each of these filters is carefully validated thanks in no small part to the highly precise Monte Carlo. It is only after said cleaning that the neural network takes its final decision. The entire analysis chain results in a effiency over 90\% in the multi-TeV area, crucially identifying most of the rare electrons at such energies, while inducing only few systematic uncertainties as they stay well below the statistical error at high energies. The resulting cosmic electron+positron flux, pushed up to 13.8 TeV, first confirms the break observed in the previous measurement but also observes a potential new feature in the form of a spectral hardening at 3.6 TeV. Such a feature has never been observed by any other experiment, and if confirmed may yield valuable informations on cosmic ray sources and accceleration.