Unify thermodynamically consistent and kinetically faithful coarse-grained models

Determine whether thermodynamically consistent coarse-grained potentials and kinetically faithful coarse-grained propagators can be unified in a single model, so that reduced-resolution simulations reproduce both equilibrium distributions and the relevant dynamical behavior of the underlying atomistic system.

Background

Machine-learned coarse-grained models reduce the number of degrees of freedom by replacing atomistic coordinates with a smaller set of retained variables. Bottom-up approaches can reproduce equilibrium distributions by learning a coarse-grained free-energy surface or potential of mean force, but integrating out fast degrees of freedom generally removes friction and memory effects that influence the dynamics of the retained coordinates.

Consequently, a coarse-grained potential that is thermodynamically accurate will not generally reproduce atomistic diffusion constants or transition rates. The paper discusses alternative approaches based on generalized Langevin or Mori–Zwanzig formulations with learned memory kernels, and approaches that learn transition densities or propagators instead of explicit energy functions. The unresolved problem is whether these thermodynamic and kinetic objectives can be achieved simultaneously within one model.