Abstract: Efficient sampling of the Boltzmann distribution of molecularsystems is a long-standing challenge. Recently, instead of generating longmolecular dynamics simulations, generative machine learning methods such asnormalizing flows have been used to learn the Boltzmann distribution directly,without samples. However, this approach is susceptible to mode collapse andthus often does not explore the full configurational space. In this work, weaddress this challenge by separating the problem into two levels, thefine-grained and coarse-grained degrees of freedom. A normalizing flowconditioned on the coarse-grained space yields a probabilistic connectionbetween the two levels. To explore the configurational space, we employcoarse-grained simulations with active learning which allows us to update theflow and make all-atom potential energy evaluations only when necessary. Usingalanine dipeptide as an example, we show that our methods obtain a speedup to molecular dynamics simulations of approximately $15.9$ to $216.2$ compared to the speedup of $4.5$ of the current state-of-the-art machine learning approach.