Statistical and Generative Models for Inference and Reasoning in Astrophysics and Beyond

Abstract

Despite considerable leaps in our understanding of the Universe, murmurs of discordance in the measurements of cosmological parameters and unanswered questions about the nature of dark matter and dark energy remain. Addressing these gaps will likely require advances in three complementary thrusts: observation, theory, and analysis methods to bridge the two. This thesis draws from tools across the evolving machine learning landscape to address problems in each of these domains. On the observational front, we used simple interpretable filters to perform a more pristine separation of stars from extragalactic interlopers to synthesize maps of Galactic extinction that possess lower correlations with the large scale structure of the Universe. We then turn our attention to two interrelated tasks in the domain of cosmological parameter estimation. Cosmological simulations play a key role in generating predictions from theory; however, they are often expensive. We demonstrate the utility of diffusion generative models as emulators of dark matter density fields conditional on cosmological parameters, and show that the same model can solve the inverse problem of constraining the cosmological parameters of an input field. Next, we present two investigations that examine the potential of large language models (LLMs) in accelerating the scientific research pipeline, in particular, scientific code generation. We first evaluate the ability of LLMs to solve general numerical analysis problems. This thesis concludes by returning to cosmology, specifically, the theoretical frontier, where we introduce a framework to enable an LLM-driven agent to autonomously implement and propose alternate cosmologies.