Predicting dark matter halo formation in N-body simulations with deep regression networks

Dark matter haloes play a fundamental role in cosmological structure formation. The most common approach to model their assembly mechanisms is through N-body simulations. In this work, we present an innovative pathway to predict dark matter halo formation from the initial density field using a Deep Learning algorithm. We implement and train a Deep Convolutional Neural Network to solve the task of retrieving Lagrangian patches from which dark matter haloes will condense. The volumetric multilabel classification task is turned into a regression problem by means of the Euclidean distance transformation. The network is complemented by an adaptive version of the watershed algorithm to form the entire protohalo identification pipeline. We show that splitting the segmentation problem into two distinct subtasks allows for training smaller and faster networks, while the predictive power of the pipeline remains the same. The model is trained on synthetic data derived from a single full N-body simulation and achieves deviations of similar to 10 per cent when reconstructing the dark matter halo mass function at z = 0. This approach represents a promising framework for learning highly non-linear relations in the primordial density field. As a practical application, our method can be used to produce mock dark matter halo catalogues directly from the initial conditions of N-body simulations.