Development and Validation of Neural Network Potentials for Multicomponent Oxide Glasses

A neural network potential (MLP) for molecular dynamics simulation (MD) of multicomponent oxide glasses was developed with special consideration of structural reproducibility. The MLP was constructed through pre-training using the dataset of density functional theory (DFT) data provided by the Open Catalysis project and fine-tuning for the dataset of nine-component glasses. A thorough validation based on previous experimental and DFT-MD data was performed for the glass structure derived from neural network molecular dynamics simulation ({\MLMD}) with the developed {\MLP}. The accuracy of {\MLMD} was investigated in terms of the local structure of the glass, comparing it to the glass derived from MD with conventional potentials. The composition dependence of the local structure in \ce{Na2O}--\ce{SiO2} and \ce{Na2O}--\ce{B2O3} glass systems was well reproduced for the {\MLMD}-derived glass. The ability to reproduce the glass structure was demonstrated in the population of four-coordinated boron population and formation of superstructures in alkali borate glasses, and the Al local structure in the novel Al-rich binary aluminoborosilicate glass. The importance of pre-training was revealed by comparing {\MLMD} results using {\MLP} developed with and without pre-training. Although better metric scores for {\MLP} without pre-training can be achieved, the resultant structure was not realistic. This is an important lesson that the metric score alone is inadequate to construct accurate {\MLP} for glasses. Finally, the developed {\MLMD} was applied to the modeling of the reference nuclear waste glass (60.1{\Si}--3.84{\Al}--15.97{\B}--12.65{\Na}--2.87{\Ca}--2.86{\Mg}--1.72{\Zr}), and the charge compensation mechanism of the cations was discussed.