Método de aplicação de Redes Neurais Convolucionárias (CNN) para identificação de fusões de galáxias em dados simulados pelo Projeto Illustris

Abstract

Since its origin, the universe has been in constant evolution, which includes the formation and destruction of black holes, galaxies, stars, and other cosmological structures. Observing this evolution allows us to understand, discover, and test models and theories related to the foundations of physics, such as the Big Bang, relativity, the existence of dark matter and dark energy, etc. In addition to experimental observation, theoretical models can be tested in computer simulations, in which theoretical variables can be directly controlled and time accelerated or reversed, allowing analyses that are impossible to perform on real scales. Recent models, however, produce a vast amount of data, which makes their manual analysis difficult. The same is true for the vast amount of images acquired by the Hubble Space Telescope (HST) and James Webb Space Telescope (JWST). To facilitate this analysis, we demonstrate here the use of convolutional neural networks (CNNs) to identify galaxy mergers in data simulated by the Illustris project. We successfully reproduced some results from the literature, improving the hyperparameters of the neural network to allow faster and more reliable convergence. Data collection was performed on open-access platforms, GitHub, and the Space Telescope Science Institute (STSI). The source code developed is also made available to the community so that it can be used in this and other problems and improved. We hope to contribute to the use of neural networks and artificial intelligence not only in astrophysics but also in other problems that require image classification and that have a large database.