Connecting Scales: Learning Dynamics for Efficient Ionic Conductivity Predictions with Graphs

Abstract

Multiscale approaches are crucial for advancing our understanding of material properties, particularly in the search for novel solid electrolytes essential for solid-state batteries. Estimating ionic conductivity through traditional molecular dynamics (MD) simulations is computationally intensive, requiring significant time to capture macro-scale behavior from micro-scale interatomic interactions. This work addresses the challenge of connecting micro-scale interatomic potentials with macro-scale conductivity measurements. We propose using equivariant graph neural networks to develop a faster mapping between these scales, significantly enhancing the efficiency of ionic diffusion predictions. This proof-of-concept demonstrates the potential to accelerate material discovery for solid electrolytes, addressing a critical need in energy storage technology.