Coarse-Graining of Molecular Dynamics Using Neural Ordinary Differential Equations

Coarse-graining (CG) improves computational feasibility of simulating complex molecular systems (e.g. proteins, or polymer chains) by reducing the number of degrees of freedom considered. Here, we collect many particles making up a composite body to a single centre of mass and orientation. Defining the CG interaction potential between the bodies that minimizes loss of information is non-trivial with no clear analytical solution. We use neural ordinary differential equations (ODE) to learn such CG potentials in a data-driven manner. We show a proof-of-concept application on a toy problem and outline the next steps towards an automated CG software pipeline.