Predicting highly correlated hydride-ion diffusion in SrTiO$_3$ crystals based on the fragment kinetic Monte Carlo method with machine-learning potential

TL;DR

This paper introduces a machine-learning-enhanced kinetic Monte Carlo method to accurately simulate hydride ion migration in SrTiO$_3$ crystals, revealing the interplay between hydride and oxygen diffusion and matching experimental activation barriers.

Contribution

The study develops a novel approach combining neural networks with kinetic Monte Carlo to predict complex ionic migration barriers dynamically in oxyhydrides.

Findings

Simulation results agree with experimental activation barriers.

Hydride migration is hindered by slow oxygen diffusion.

Oxygen diffusion accelerates due to barrier changes.

Abstract

Oxyhydrides have drawn attention because of their fast ion conductivity and strong reducing properties. Recently, hydride ion migration in SrTiO$_{3-x}$H$_{x}$ oxyhydride crystals has been investigated, showing that hydride ion migration is blocked by slow oxygen diffusion. In this study, we investigate the hydride-ion migration mechanism using a kinetic Monte Carlo approach to understanding the relationship between the hydride and oxygen ions. The difficulties in applying the method to hydride and oxygen ion migration involve complex changes in the ionic migration barrier, which shifts dynamically depending on the characteristics of the surrounding hydride and oxygen ions. We can predict these complex changes using a machine-learning neural network model. The simulation can then be performed using this model to predict the temperature-dependent ionic-migration behavior. We found that our simulation results with respect to the activation barrier for hydride ion diffusion accorded well with those obtained by experiment. We also found that hydride ion migration is affected by slow oxygen diffusion and that oxygen diffusion is accelerated by changes in the ionic migration barriers. The parallel-processing efficiency of our proposed method was 84.92 \% for our 1,000-CPU implementation, suggesting that the approach should be widely applicable to simulations of ionic migration in crystals at a reasonable computational cost.