[3C01-05] Data-Driven Surface Science 

10:45 AM - 11:15 AM JST(1:45 AM - 2:15 AM UTC)

[3C04【依頼講演】] Development and applications of charge-aware machine-learning models to investigate ionic behaviors under electric fields

Anh Khoa Augustin Lu^2,1, *Satoshi Watanabe¹ (1. Department of Materials Engineering, The University of Tokyo (Japan), 2. Research Center for Materials Nanoarchitectonics (MANA), National Institute for Materials Science (NIMS) (Japan))

Conventional machine-learning interatomic potentials cannot directly describe ionic dynamics under electric fields because they do not predict field-responsive quantities. To overcome this limitation, we developed neural-network models for Born effective charge (BEC) tensors and, more recently, SevenNet-Polar, an equivariant graph-neural-network framework for accurate prediction of energies, forces, stresses, BECs, and dielectric tensors. Molecular dynamics simulations under electric field revealed enhanced Li-ion migration in amorphous Li₃PO₄, even without explicitly introduced defects, and enhanced oxygen-ion diffusion in defect-laden ZrO₂. We are also investigating distinct ionic behavior near ZrO₂ grain boundaries. SevenNet-Polar achieves high prediction accuracy and enables large-scale simulations, including systems containing up to 1.5 million atoms. The presentation will discuss the models, their scaling performance, and recent simulation results.

Password required to view