Machine learning silicate systems at extreme conditions
(University of California, Los Angeles)
Silicates are the major building blocks of terrestrial planets and exoplanets. The thermochemical states of silicates and their phase transitions, to a large extent, dictate a planet’s evolution. Experiments that simulate extreme planetary conditions, for example, the Moon-forming giant impact or super-Earth interiors, are difficult. Ab initio atomistic simulations, although may reach a wide pressure and temperature range, are limited to small systems that consist of a few hundreds of atoms due to the large computational costs. However, atomistic modeling on processes such as melting, crystallization, vaporization, condensation, deformation, and miscibility usually requires large systems. Recently, the emerging machine learning molecular dynamics simulations offer an unprecedented opportunity to solve this problem. Machine learning molecular dynamics simulations employ potential energy surfaces learned from quantum mechanical calculations, which tremendously accelerates atomistic modeling and enables simulating systems up to millions of atoms. In this talk, I present our recent simulation results of silicates using machine learning potentials.