Accurate Ion Transport in Polymer Electrolytes with UF3 Machine-Learned Potentials

Polymer electrolytes offer advantages over conventional liquid electrolytes, such as enhanced safety, selective ion conductivity, high energy density, and electrochemical stability. However, their adoption is limited by low ionic conductivity. Understanding the microscopic mechanisms of ion diffusion is crucial to overcome this limitation, and molecular dynamics (MD) simulations are vital for providing these insights. In this work, we develop and benchmark an Ultra-Fast Forcefield (UF3) machine-learned interatomic potential (MLIP) for the polyethylene oxide-LiTFSI system, enabling efficient MD simulations. UF3 models the system's energy as a linear combination of two-body and three-body interactions, blending the interpretability of classical force fields with the flexibility of cubic B-spline basis sets and the statistical rigor of machine learning. To optimize the UF3 model, we implemented an active learning scheme that generates new training data, minimizing uncertainties in the fitted coefficients for molecular motifs encountered in simulations. We validate UF3 by comparing computed ionic diffusivity to experimental values, demonstrating the model's accuracy. Furthermore, we reveal that achieving low energy and force errors in the UF3 model does not always ensure accurate diffusivity predictions in MD simulations, underscoring the importance of incorporating additional physical metrics during model development to improve performance in real-world applications.