Machine learning en astrofísica y cosmología

Resumen

Nearly 380.000 years after the Big Bang, photons decoupled from baryons and freely traveled along the Universe. Today, they still can be observed in the microwave regime. This effect is called the cosmic microwave background, and it is a key probe for cosmologists to understand the nature and evolution of the Universe. However, at the frequencies where the cosmic microwave background can be observed, there are several emissions from our Galaxy and extragalactic sources called foregrounds, which contaminate both temperature and polarization maps. The characterization of these emissions and therefore the recovery of the cosmic microwave background depend mainly on the quality of the chosen methodology. Due to the increasing in the quality of the instruments used in Astrophysics and Cosmology, the quantity of available data in future cosmic microwave background experiments will also increase, requiring more sophisticated and automatic methods. Due to the increasing in the computational capability, machine learning models, which have the ability of learning from data a particular task but require high amounts of data and memory, have been increased their impact in many areas of human live. Furthermore, artificial neural networks, which are machine learning models inspired in neuroscience, are perfect for cosmic microwave background recovery and foreground characterization since they are designed to deal with non-linear behaviors from data, which are precisely the ones that characterize these emissions. This PhD thesis presents new methodologies based on artificial neural networks for several cosmic microwave background analyses. More precisely, by cutting squared patches of the microwave sky as seen by the Planck satellite, several convolutional neural networks have been trained with realistic simulations for radio galaxies detection, for the constraining of their polarization properties and for the recovery of the cosmic microwave background in both temperature and polarization. Lastly, future uses and developments of these neural networks will be described.