We investigated the training dependency of neural network interatomic potentials for molecular dynamics simulation of a Ru–Si–O mixed system. Our neural network interatomic potential was improved using a data augmentation technique for the training dataset, including data points of reference energies and forces related to reference structures. We demonstrated that the data augmentation technique, focusing on the lattice expansion coefficient of bulk structures in the training dataset, requires moderation to ensure optimal training of the neural network interatomic potential. We found that Ru/SiO2 interfaces were accurately represented using the neural network interatomic potential trained with Ru and SiO2 surfaces in addition to Ru/SiO2 interfaces. In the case of modeling Ru/SiO2 interfaces including unbonded atoms, training the surfaces with unbonded atoms is effective in generalizing the neural network interatomic potential. Our demonstration and finding shed light on the pivotal role of the training dataset on the development of the neural network interatomic potential for the Ru–Si–O mixed system.
Read full abstract