We propose a method for evaluating ionic conductivity by integrating machine learning models with nonequilibrium molecular dynamics simulations under a constant electric field. The method computes the forces exerted on atoms by the external electric field within the linear response regime, using the Born effective charge tensor predicted by an equivariant graph convolutional neural network. These field-induced forces are then combined with nonperturbed forces predicted by an equivariant neural network potential. We applied this approach to ${\mathrm{Li}}_{10}{\mathrm{GeP}}_2{\mathrm{S}}_{12}$, a representative solid lithium-ion electrolyte. The method enables accurate evaluation of conductivity at a level comparable to first-principles molecular dynamics simulations, but at a fraction of their prohibitive computational cost, and further allows simulations of realistic ion conduction dynamics under an applied electric field. The predicted Born effective charges indicate that Li ions located near ${\mathrm{GeS}}_4$ and ${\mathrm{PS}}_4$ tetrahedra exhibit higher ionization. These ions are localized around ${\mathrm{GeS}}_4$ and ${\mathrm{PS}}_4$ tetrahedra, which serve as structural barriers that interrupt conduction pathways. In contrast, other Li ions are less ionized and primarily contribute to ionic conduction. The proposed method offers a nonarbitrary and physically interpretable framework for the quantitative evaluation of ionic dynamics, including charge fluctuations that are essential for complex systems such as inhomogeneous crystals and interfaces.

Close

PrevNext

View figure in article

• Density functional theory
• Electrical properties of membranes
• First-principles calculations
• Ionic conductivity
• Machine learning
• Neural network simulations