Most stars form in clumpy and sub-structured clusters. These properties also emerge in hydro-dynamical simulations of star-forming clouds, which provide a way to generate realistic initial conditions for N-body runs of young stellar clusters. However, producing large sets of initial conditions by hydro-dynamical simulations is prohibitively expensive in terms of computational time. We introduce a novel technique for generating new initial conditions from a given sample of hydro-dynamical simulations, at a tiny computational cost. In particular, we apply a hierarchical clustering algorithm to learn a tree representation of the spatial and kinematic relations between stars, where the leaves represent the single stars and the nodes describe the structure of the cluster at larger and larger scales. This procedure can be used as a basis for the random generation of new sets of stars, by simply modifying the global structure of the stellar cluster, while leaving the small-scale properties unaltered.