Abstract: The formation of cosmic structures in the Universe is driven by the gravitational collapse of small matter density fluctuations present in the early Universe. The non-linear nature of gravitational collapse makes it difficult to develop a physical understanding of how complex late-time cosmic structures emerge from these linear initial conditions. In this talk, I will show how explainable machine learning models can be used to extract new knowledge about the underlying physics of cosmological structure formation. I will focus on two applications to the density profiles of dark matter halos. I will address the question of what determines the final profile of a halo, in particular focussing on the role of the initial conditions and the halo’s assembly history. I will then show how neural networks can be used to discover the independent degrees of freedom that characterise the halo density profiles. I will show that the model is able to reproduce the known variations encapsulated by previous empirical approaches, and discovers an additional factor of variation in the outer profile related to the splashback radius.