Efficient Sampling of Equilibrium Distributions with Scalable Autoregressive Flow Models

Deep generative models are a powerful tool for sampling equilibrium states of physical systems and estimating observable averages. However, training these models to sample complex molecular systems can be challenging since the underlying distributions are high-dimensional in nature and likely contain nonlocal interactions. We evaluate the quality of autoregressive flow models in the context of free energy estimation in solid lattice systems, using results from the Frenkel-Ladd technique as a benchmark. Moreover, the autoregressive structure enables partial reconstruction of the system of interest, which can be exploited to improve the scalability of the method. The autoregressive model can be trained on only parts of the system and autoregressively generate configurations of the entire system.