Content area
Abstract
Solid electrolyte materials are a key component for next-generation solid-state batteries. Ab initio simulations, particularly density functional theory (DFT) calculations, have offered profound insights into critical considerations, such as electrochemical stability, electrode compatibility, mechanical properties, doping effects, and diffusion mechanisms, of solid electrolytes. However, the high computational cost of DFT has limited its application in modeling the behavior of solid electrolytes under practical conditions, for example, diffusivity at room temperature and in complex microstructures with disordered atomic arrangements. This thesis demonstrates how machine learning interatomic potentials (MLIPs) can extend DFT accuracy to larger simulation scales to directly probe solid electrolyte under these practical conditions. Furthermore, as a rapidly advancing technology, MLIPs have been continuously improved in terms of accuracy and reliability. This thesis showcases the development of a DImensionality-Reduced Encoded Clusters with sTratified (DIRECT) sampling algorithm, leveraging universal graph neural network models for efficient and robust training set generation for MLIPs. This thesis exemplifies the transformative journey from pure ab initio methods to machine learning paradigms in the field of computational materials science, promising to unlock new horizons in solid electrolytes and other materials.
