Early Prediction of Ion Transport Properties in Molecular Dynamics Simulations of Solid Polymer Electrolytes Using Machine Learning

Solid polymer electrolytes have received much interest for developing a new generation of safe, high-performance Li-ion batteries. To this end, Molecular Dynamics (MD) simulations have been widely used to quickly screen polymer candidates for desirable properties, such as high ionic conductivity and mechanical robustness. Unfortunately, these simulations can be time-consuming, and accurate predictions of these properties can require MD simulation lengths of 20ns or more which correspond to wall clock times on the order of hundreds of hours using conventional computational resources. Therefore, accelerating MD simulations is critical to expedite the screening process, and subsequently, the design of new polymers. In this study, we show that with the correct choice of descriptors, we can make predictions of equilibrated transport properties at 10% of the total simulation time. The new set of descriptors used in the current study combines the configuration of ion clusters with the early time evolution of transport properties. Specifically, we find that descriptors that include information about anion-cation interactions and dynamics of ion transport in the polymer environment outperform features extracted from the molecular structure of the polymers. We show that these descriptors have several advantages over polymer structure-only descriptors, as features can be extracted and predictions made at any time point, increasing the applicability of this method to a wide range of polymer systems, simulation times, and conditions such as different temperatures and concentrations.