Large-scale Molecular-dynamics Simulations of SiO<sub>2</sub> Melt and Glass under High Pressure with Robust Machine-learning Interatomic Potentials

We have carried out molecular-dynamics simulations of SiO2 melt and glass at 60 GPa in a system of about 30,000 atoms, with robust machine-learning interatomic potentials based on ab initio calculations. The large-scale simulations not only reproduced the ab initio calculations, but were also found to significantly improve the precision of the intermediate-range structure and reproduce the emergence of a nanometer-scale heterogeneous intermediate state during the pressure-induced phase transformation. Interatomic potentials determined using machine learning easily destabilize simulations of highly covalent melts under high pressure in an extended system with over one order of magnitude more atoms than those in the initial calculations used for data generation. The combination of potential averaging and active learning can efficiently suppress computational instability.