(376ap) Systematic Coarse-Grained Mapping and Potential Learning

The choice of coarse-grained (CG) mapping operator and the derivation of the corresponding coarse-grained (CG) potentials are critical in determining how effective a CG model for a molecule is. Here we present a technique for extracting a CG mapping operator from a hierarchical graph which encodes multiple mapping operators. We demonstrated the used of relevant thermodynamic properties, like entropy, to automate CG mapping selection from the hierarchical graph. We have also added the capability of learning CG potentials on-the-fly via force-matching to a recently developed plugin, HOOMD-TF. The plugin integrates TensorFlow (TF), a Machine Learning platform, with the GPU-accelerated simulation engine, HOOMD-blue. We have compared the CG potentials learned by a neural network to those modeled using a basis-set. These systematic strategies of mapping operator selection and CG potential learning were implemented for methanol, alkanes and peptides.