Example of the prediction problem considered in this article. Top panel: Slice of the three-dimensional initial density field of an N-body simulation. Each voxel (represented here as a pixel) corresponds to a particle that can become part of a halo at later times. Bottom panel: Regions in the initial condition space (same slice as the top panel) that are part of different DM haloes at redshift z = 0. Pixels coloured in white do not belong to any halo. Pixels with different colours belong to different haloes. In this work, we present a machine-learning approach to predict the formation of haloes (as in the bottom panel) from the initial condition field (top panel).