Speaker
Description
Donggyu Lee, Takuji Oda
Seoul National University
Lithium titanate (Li2TiO3) is a promising ceramic-type tritium breeder material for nuclear fusion reactors. While extensive experimental studies have been conducted to investigate its properties, the understanding of its behavior under radiation environments remains limited. Molecular dynamics (MD) simulations provide crucial insights into defect formation and evolution and their impact on material properties. Previous MD studies on Li2TiO3 have primarily relied on a Buckingham-type empirical two-body potential combined with the universal ZBL potential.
In recent years, machine learning (ML) potentials have emerged as a game changer in atomistic simulations, achieving first-principles accuracy while significantly reducing computational cost. We have recently developed an ML potential for Li2TiO3 based on the moment tensor potential framework. Our evaluation shows that the ML potential outperforms the conventional Buckingham-type potential in reproducing bulk material properties, such as heat capacity and thermal conductivity, and is in better agreement with DFT calculations and experimental data.
In this presentation, we will demonstrate the applicability of our ML potential in analyzing defect systems. Our ML potential can assist in geometry optimization tasks by providing approximate defect structures and energies at a lower computational cost than first-principles calculations, and allows us to thoroughly perform geometry optimization, which is a formidable task for first-principles calculations because the flat potential landscape of Li2TiO3 causes slow convergence. Furthermore, the accuracy and efficiency of ML potentials allow the exploration of various defect compositions and configurations. These results illustrate how ML potentials facilitate a comprehensive analysis of defect structure and stability.
