Harnessing the Power of Deep Learning for Astrophysical Discoveries

Abstract

The rapid growth of astrophysical data is ushering us into a new era of astrophysics and cosmology, one where many current statistical methods will struggle to keep pace with the volume of available information. These datasets will offer us unprecedented opportunities to gain deeper insights into our Universe, but making optimal use of them requires either development of new techniques or optimization of existing methodologies. This thesis explores how advances in deep learning can be leveraged to tackle the challenge of analyzing the next generation of large astronomical datasets. We delve into a range of astrophysical observations, including strong gravitational lensing images, time-domain data of supernovae, and galaxy surveys, and demonstrate how deep learning methods have the potential to accelerate scientific discoveries in these domains. The first part of this thesis focuses on neural likelihood-ratio estimation applied to strong gravitational lensing observations. Substructures in strong gravitational lenses are powerful probes of the nature of dark matter, but conventional methods for analyzing them are computationally costly and cannot scale up to the volume of images made available by upcoming surveys. To address this, we introduce neural likelihood-ratio estimation as a promising tool to accelerate analysis and demonstrate its ability to measure the density profile of substructures, first from simulated images and then from telescope observations. The second part of this thesis introduces Maven, a foundation model for supernova science. Maven is pre-trained with a large synthetic dataset and then fine-tuned on observations using a contrastive objective. We demonstrate that this training scheme is valuable for integrating large supernova datasets from multiple modalities into a joint embedding space which can then be used for downstream inference. In the last part of this thesis, we work towards a more scalable framework to analyze the large-scale structures in our Universe. We build a hybrid simulation-based inference framework for cosmological inference on galaxy clustering, combining analytical perturbative methods and neural density estimation to bypass the need for full-volume N-body simulations. This work lays the foundation for scaling up simulation-based inference for increasingly large galaxy surveys in the years to come.