Deep learning insights into cosmological structure formation

Dark matter halos are the fundamental building blocks of cosmic large-scale structure. Improving our theoretical understanding of their evolution and formation is an essential step towards understanding how galaxies form. I will present a deep learning model, based on 3D convolutional neural networks (CNNs), trained to learn the relationship between the initial conditions density field and the final mass of dark matter halos in N-body simulations. Our goal is to utilize deep learning for knowledge extraction: we aim to gain new physical understanding of halo formation by interpreting the findings of the deep learning model. I will present a simple and effective technique to reveal which information the deep learning model extracts from the inputs to make the final predictions. This then allows us to interpret the features learnt by the deep learning model in relation to physical properties of the initial conditions of the Universe.