Development and\nValidation of Neural Network Potentials\nfor Multicomponent Oxide Glasses

A neural\nnetwork potential (NNP) for molecular dynamics\nsimulation\n(MD) of multicomponent oxide glasses was developed with a particular\nfocus on structural reproducibility. The NNP was constructed through\npretraining using a density functional theory (DFT) data set provided\nby the Open Catalysis project and fine-tuning based on a data set\nof nine-component glasses. Thorough validation was performed by comparing\na glass structure derived using neural network molecular dynamics\nsimulation (NNMD) and the NNP developed here with previous experimental\nand DFT-MD data. The accuracy of the NNMD was investigated in terms\nof the local structure of the glass, in comparison with a glass derived\nfrom MD with conventional potentials. The composition dependence of\nthe local structure in Na<sub>2</sub>O–SiO<sub>2</sub> and\nNa<sub>2</sub>O–B<sub>2</sub>O<sub>3</sub> glass systems was\nwell-reproduced for the NNMD-derived glass. The ability to reproduce\nthe glass structure was demonstrated in the four-coordinated boron\npopulation, by formation of superstructures in alkali borate glasses,\nand by the Al local structure in a novel Al-rich binary aluminoborosilicate\nglass. The importance of pretraining was investigated by comparing\nNNMD results obtained using the NNP developed with and without pretraining.\nAlthough better metric scores were obtained for the NNP without pretraining,\nthe resulting structure was not realistic. This is an important lesson,\nshowing that the metric score alone is inadequate to determine the\naccuracy of an NNP for glasses. Finally, the developed NNMD was used\nto model a reference nuclear waste glass (60.1SiO<sub>2</sub>–3.84Al<sub>2</sub>O<sub>3</sub>–15.97B<sub>2</sub>O<sub>3</sub>–12.65Na<sub>2</sub>O–2.87CaO–2.86MgO–1.72ZrO<sub>2</sub>), and the charge compensation mechanism of the cations was investigated.