Abstract

Precision cosmology is entering a new golden era, with current and upcoming surveys mapping the distribution of galaxies to an unprecedented level of detail. However, uncertainties in the theoretical modelling of the Universe on small, non-linear scales remain a major roadblock to interpreting these cosmological measurements. While machine learning has greatly enhanced our ability to analyse large datasets, its “black box” nature often limits physical interpretability and trust in their results. I will discuss recent advances in modelling cosmological observables in the non-linear regime using artificial intelligence (AI), and on the impact of baryonic feedback on cosmological observables. I will introduce deep learning frameworks that are explicitly designed to be interpretable and explainable in terms of the underlying physics of interest, and demonstrate their application to properties of cosmic structures. I will then present applications to the cosmic microwave background, revealing to which parameters the temperature power spectrum is sensitive in the context of early dark energy models.