Exploring non-equilibrium systems with normalizing flows

Normalizing flows are generative invertible neural-network models that gradually map a complicated probability distribution to a simple one, e.g. a normal multi-dimensional Gaussian. They can learn to sample from an empirically observed distribution and at the same time provide an estimate for this distribution. This allows for the use of information-theoretical concepts like the Kullback-Leibler divergence to explore phase diagrams, classify trajectories in non-equilibrium systems in an unsupervised fashion, as well as efficiently obtain effective model descriptions. We apply normalizing flows to examples of equilibrium and non-equilibrium physical systems.