Exploring Liquid Silica with Deep Machine Learning Potentials

Silica, the most abundant component of Earth's crust, plays a fundamental role in understanding the origin and formation of life. Liquid silica exhibits complex structural and thermodynamic behaviors, and its abundant anomalies similar to water have aroused considerable interests in fundamental glass physics and materials science. However, experimental investigations are hindered by the extreme temperatures required, while density functional theory (DFT) calculations are computationally prohibitive, and classical potentials such as BKS (Beest–Kramer–van Santen) and SHIK (Sundararaman–Huang–Ispas–Kob) introduce significant deviations from real systems. To address these challenges, we employ deep potential molecular dynamics (DeePMD) to construct an interatomic potential for silica that achieves DFT-level accuracy at the computational cost of classical models. Using this framework, we systematically investigate the thermodynamic and structural properties of liquid silica, as well as its surface behavior, and compare the results with classical potentials. Specifically, we analyze the radial and angular distribution functions, density profiles, energy, and heat capacity. Our results demonstrate that DeePMD provides a more accurate description of geometric structures, capturing a weak symmtry breaking of local tetrahedrality at relatively high temperatures without an evident liquid–liquid phase transition as observed in water. Furthermore, the model predicts orientational ordering at the silica surface, offering new insights into surface properties with broad implications for studying silicate materials.