Coarse-Grained Modeling of Fluids and Polymer-Gas Systems Using Differentiable Molecular Dynamics

Developing accurate coarse-grained (CG) models requires systematic optimization of interaction parameters or tuning of pairwise potential functions. This is traditionally achieved either through iterative trial-and-error, or expensive derivative-free methods due to the difficulty of computing gradient functions through molecular dynamics simulations. Automatic differentiation (AD) overcomes this limitation and enables efficient gradient calculation. In this work, we integrate AD with molecular dynamics simulations to construct CG models for simple fluids and gas-polymer systems. We implement AD to obtain gradients of the potential energy functions and optimize CG model parameters. First, we demonstrate coarse-graining of liquid water from the four-site TIP4P/2005 model at 1 bar and 298 K to a single-site CG model. We then extend the methodology to optimize coarse grain force field for polymer membranes used in gas separation application. For the polymer systems, we apply multi-objective optimization to match different structural and thermodynamic distributions from full-atomistic simulations. After building the model, we validate the performance of the CG system with respect to gas solubility and transport properties against atomistic simulations. We further extend this method to coarse graining gas molecules alongside the polymer matrix. Finally, we also examine the transferability of these CG models to different thermodynamic conditions.