Developing machine learning interaction potentials for polyolefins

In the field of molecular modeling and simulation, machine learning (ML) methods are increasingly employed to represent the potential energy surface (PES) of molecular systems, as they allow us to bridge the long-standing speed and accuracy gap between quantum and classical modeling approaches. In this regard, polymeric materials are particularly challenging, due to the range of length and time scales inherent to the physics of polymer chains. Providing a faithful representation of detailed atomic-level interactions (including possible chemical reactions) across these scales is, therefore, computationally expensive, thus opening the door for ML-based approaches to potentially solve this challenge. In this study, we aim to use the deep learning-based Deep Potential Molecular Dynamics (DeePMD) scheme [NeurIPS pp. 4436 - 4446 (2018)] to build ML-potentials for polyolefins. However, best practices for assembling the ML-potential training data in polymeric systems are essentially unknown. As a first attempt, we used the TraPPE united atom forcefield to generate MD simulation trajectories of isolated polyethylene chains and amorphous polyethylene melts and used them to train DeePMD models. We evaluate the ML-potentials’ ability to predict polyethylene's chain-level structure and physical properties, to develop rules-of-thumb for polymeric ML model development. We then discuss our ongoing efforts to build models targeted for studying polyolefin thermal degradation.