Using the relative entropy to sample free energy landscapes with transferable coarse-grained models

Coarse-grained molecular models permit quantitative simulations of complex systems with large length and time scales, often enabling free energy calculations that would be difficult or impossible with all-atom models. However, their utility tends to be highly constrained by accuracy over a limited set of state conditions. We discuss a fundamental approach to the identification of accurate coarse models and more generally of emergent physical behavior from many-body systems. The relative entropy measures the information lost upon coarse graining and we hypothesize that its minimization provides a universal variational principle for coarse-graining. This broad statistical-mechanical framework can improve and even detect both analytical and simulation models of complex systems. We discuss conceptual and numerical aspects of this approach. In particular we show recent efforts to create improved coarse grained models with high degrees of transferability suitable for free energy and phase equilibrium calculations, using unconventional interaction potentials that capture multibody effects.