Machine learning techniques for blind searches of new gamma-ray pulsars with Fermi

The searches for gamma-ray pulsars whose spin frequency f0 and its evolution f1 are unknown, as for radio-quiets, are called blind searches and are computationally expensive: it is necessary a scan on the possible values of f0 and f1 of the signal through a Fourier analysis. This search for the periodicity of the signal can be boosted by studying the differences in the arrival times of photons within a certain time window, rather than the arrival times themselves. For a given source, this analysis returns a set of candidate pulsar frequencies (f0, f1) (10^6 ), that need to be further filtered. With them a set of phase vs time graphs can be constructed and be examined by image classification models using machine learning. In this thesis we have developed a deep learning algorithm based on 2D convolutional neural networks to improve searches for new gamma-ray pulsars based on data collected by the Fermi Large Area Telescope. We initially simulated pulsar signals and non-pulsating signals with which we built a dataset of phasograms vs time. We divided these images into datasets for training, validation and testing, then we used them to train and evaluate different image classification models to obtain the best performing architecture. The approach has shown that an accuracy of 99% can be achieved. This machine-learning method can be expanded and further implemented, with the final goal of train the model and apply it to real Fermi data