Metodi di deep learning per la stima di proprietà fisiche delle galassie

Euclid is a mission of the European Space Agency (ESA) that will be launched in July 2023 with the objective of exploring the evolution of the dark universe. Euclid will study billions of galaxies up to a distance of 10 billion light years, covering more than a third of the entire sky. This will allow astronomers to measure the expansion history of the Universe and the growth rate of cosmic structures. Machine learning methods, and especially deep architectures, are some of the most efficient state-of-the-art tools for managing huge amounts of data. For this reason, they are suitable for handling the data that Euclid will collect and for estimating galaxy physical properties more carefully with respect to traditional methods used in astronomy. In this work we applied Deep Learning methods to estimate three well known and important galaxy physical properties, i.e. redshift, stellar mass and star formation rate (SFR). The dataset used consists of artificially generated images that mimic the real images that Euclid will capture. Each image is associated with tabular data corresponding to the magnitude measurement performed by Euclid in 9 spectral bands. We developed an architecture called Fusion Network, which consists mainly of two parts, ResNet50 for image processing and a Multilayer perceptron (MLP) for tabular data processing. These two networks allow to extract the meaningful features from the respective input data: the features are then concatenated and used as input to a final MLP for obtaining the estimate of redshift, stellar mass and star formation rate. The main result obtained by our proposed deep learning approach is the simultaneous estimation of these properties, that in previous works were estimated through ad-hoc models built for each one. Besides, the exploitation of the correlations among the three physical properties allows our model to beat state-of-the-art methods.