Abstract
With the goal of producing fast mock observations of the Sunyaev-Zeldovich (SZ) effect, I trained two conceptually different machine learning models to read images of dark matter density as input and produce maps of the thermal SZ effect as output. The maps were computed from the particle data in the Romulus simulations, and I wrote a parallelized program in C to integrate along multiple lines of sight to generate training data to be used by the models. The models utilize transfer learning to converge quickly. The first model contains a pre-trained encoder based on the ResNet model, followed by a decoder to generate SZ maps. A similar structure works as the generator in the second model which is a Generative Adversarial Network (GAN), also including a discriminator which competes with the generator to distinguish between generated images and images from the simulation. The first model performed best on the training data, but I argue that it is mostly due to overfitting and without much real value. Both models performed about as well on validation data, but the GAN model is better in general as the training and validation error is more consistent, indicating a more robust model overall
Supervisors:
Associate Professor Sijing Shen, Institute of Theoretical Astrophysics, UiO (main-supervisor)
Researcher Signe Reimer-Sørensen, Sintef Digital (co-supervisor)
Intern. assessor: Professor Øystein Elgarøy, Institute of Theoretical Astrophysics, UiO
Extern. assessor: Astrophysicist and science communicator Peter Laursen, Niels Bohr Institute, University of Copenhagen