Machine Learning-Driven Multiscale Simulations: Unveiling Atomic Diffusion in Superionic Materials

Machine learning-driven multiscale simulations can significantly accelerate the design and optimization of superionic thermoelectric and cathode materials. Using systems such as Ag₈SiTe₆, Zn₃.₆₊ₓSb₃, and the Li-rich layered iron sulfide cathode (Li₂FeS₂), we illustrate the efficacy of these simulations. Machine-learning interatomic potentials (MLIPs) combine DFT-level accuracy with the scalability of classical molecular dynamics. By integrating optimized simulation methods across multiple scales, we can explore diffusion mechanisms in systems containing thousands of atoms, with simulation times extending beyond 500 ps. This approach enables more precise modeling and insights into atomic behavior, which are critical for developing high-performance superionic materials.