(197aa) Machine Learning Anisotropic Coarse-Grained Potentials

Molecular dynamics (MD) simulations enable prediction of equilibrium structure and properties of complex molecular systems, but slow relaxation times and high computational cost of highly-resolved (all-atom and united atom) representations prohibits the interrogation of many multi-million-atom morphologies. This is where coarse-grained molecular dynamics (CGMD) proves useful, as it reduces the computational complexity by grouping atoms into larger simulation elements. Here we focus on the use of non-spherical beads to more accurately capture the structure of materials with complex shapes and directional dependencies that are difficult to represent with spherical simulation elements.

We combine machine learning (ML) with CGMD to learn the forces and torques between nonspherical beads that will result in equivalent trajectories to all-atom simulations. A proof-of-concept neural network model is developed using PyTorch for the organic semiconductor PPS, which takes in the relative positions and orientations of atomistic simulations to learn a latent representation of the local pair interactions by incorporating neighbor data. The trained neural network is used to infer the forces and torques acting on coarse-grained simulation elements. We compare the accuracy of models trained against data sets of varying size, and those biased to include configurations where the intermolecular potential gradients are largest. We implement the resulting ML potentials in the HOOMD-Blue simulation engine and compare against all-atom rigid models of PPS quantitatively and qualitatively. We conclude with a discussion of future directions and network architectures that may offer further performance and accuracy enhancements. In sum we develop tools for performing efficient large-scale structural predictions of complex molecules including organic semiconductors, semiconducting polymers, and enable scientific interrogations related to these materials.

This material is based upon work supported by the National Science Foundation under Grant No. 1653954. This research was partially supported by the National Aeronautics and Space Administration (NASA) under the University Leadership Initiative program; grant number 80NSSC20M0165.