Revisiting collective-variable guided sampling with normalizing flows

Deep generative models parametrize very flexible families of distributions able to fit complicated datasets of images or text. These models provide independent samples from complex high-distributions at negligible costs. On the other hand, sampling exactly a target distribution, such as the Boltzmann distribution of a physical system, is typically challenging: either because of dimensionality, multi-modality, ill-conditioning or a combination of the previous. A recent line of work using generative models to accelerate sampling has shown promises but still struggles as the system size gets large. In this talk, I will discuss an approach tackling this challenge by using a normalizing flow to explore the configuration space of a collective-variable (CV) and a non-equilibrium candidate Monte Carlo to recover unbiased all-atoms configurations. The approach revisits CV-guided sampling with two main advantages. Firstly, the collection of CVs need not be restricted to a few variables and can include tens or hundreds of degrees of freedom. Secondly, updates in the CV space are non-local thanks to the generative model, leading to a fast exploration regardless of free energy barriers.