Determining the Limit of Extrapolation for Macromolecular Machine Learning Potentials

Machine learning potentials have great promise for increasing what is computationally possible for molecular simulation, enabling atomistic simulation at spatiotemporal scales that would otherwise be inaccessible from traditional molecular dynamics and electronic structure calculations. One such promise is the ability to train a machine-learned potential on smaller, computationally cheap systems and extrapolate the behavior of complex molecular interactions across longer-length scales. Here, we analyze the consequences of this approach in polymeric systems, specifically the limitations of extrapolating long-chain and branched polyalkane potentials from ab-initio energetics of smaller hydrocarbons. Specifically, we address the question: what is the smallest polymer chain we can build a machine-learned potential on to learn complex macromolecular interactions? In doing so, we also derive fundamental insights into the origins of locally determined force field construction as it relates to hydrocarbon systems and we hope to further conversations of transfer learning and machine-learning architecture as it relates to the polymers community.