Thermodynamic Transferability\nin Coarse-Grained Force\nFields Using Graph Neural Networks

Coarse-graining is a molecular modeling technique in\nwhich an atomistic\nsystem is represented in a simplified fashion that retains the most\nsignificant system features that contribute to a target output while\nremoving the degrees of freedom that are less relevant. This reduction\nin model complexity allows coarse-grained molecular simulations to\nreach increased spatial and temporal scales compared with corresponding\nall-atom models. A core challenge in coarse-graining is to construct\na force field that represents the interactions in the new representation\nin a way that preserves the atomistic-level properties. Many approaches\nto building coarse-grained force fields have limited transferability\nbetween different thermodynamic conditions as a result of averaging\nover internal fluctuations at a specific thermodynamic state point.\nHere, we use a graph-convolutional neural network architecture, the\nHierarchically Interacting Particle Neural Network with Tensor Sensitivity\n(HIP-NN-TS), to develop a highly automated training pipeline for coarse-grained\nforce fields, which allows for studying the transferability of coarse-grained\nmodels based on the force-matching approach. We show that this approach\nyields not only highly accurate force fields but also that these force\nfields are more transferable through a variety of thermodynamic conditions.\nThese results illustrate the potential of machine learning techniques,\nsuch as graph neural networks, to improve the construction of transferable\ncoarse-grained force fields.