Development and Validation of Neural Network Potentials for Multicomponent Oxide Glasses

A neural network potential (NNP) for molecular dynamics simulation (MD) of multicomponent oxide glasses was developed with a particular focus on structural reproducibility. The NNP was constructed through pretraining using a density functional theory (DFT) data set provided by the Open Catalysis project and fine-tuning based on a data set of nine-component glasses. Thorough validation was performed by comparing a glass structure derived using neural network molecular dynamics simulation (NNMD) and the NNP developed here with previous experimental and DFT-MD data. The accuracy of the NNMD was investigated in terms of the local structure of the glass, in comparison with a glass derived from MD with conventional potentials. The composition dependence of the local structure in Na2O–SiO2 and Na2O–B2O3 glass systems was well-reproduced for the NNMD-derived glass. The ability to reproduce the glass structure was demonstrated in the four-coordinated boron population, by formation of superstructures in alkali borate glasses, and by the Al local structure in a novel Al-rich binary aluminoborosilicate glass. The importance of pretraining was investigated by comparing NNMD results obtained using the NNP developed with and without pretraining. Although better metric scores were obtained for the NNP without pretraining, the resulting structure was not realistic. This is an important lesson, showing that the metric score alone is inadequate to determine the accuracy of an NNP for glasses. Finally, the developed NNMD was used to model a reference nuclear waste glass (60.1SiO2–3.84Al2O3–15.97B2O3–12.65Na2O–2.87CaO–2.86MgO–1.72ZrO2), and the charge compensation mechanism of the cations was investigated.