Machine learning potentials for amorphous oxides modeling and simulations

Machine learning (ML) offers a powerful approach to generate inter-atomic potentials that are superior to the conventional classical functions used in simulations. With an accuracy comparable to first-principles calculations and a much lower calculational cost, potentials generated by ML enable simulations of large number of atoms and better predictions of structures and energy landscape properties for amorphous materials, especially for the doping systems where current classical pair potentials are not available or fail to reproduce the atomic structure features from experiments. In this work, we demonstrate a ML potential for zirconia-doped amorphous tantala based on spectral neighbor analysis potential (SNAP), which allows us to model the amorphous LIGO mirror coatings. Compared to first-principles calculations, the SNAP ML potential has very low energy mean absolute error (MAE) (~10^-4 eV/atom) and force MAE (~10^-1 eV/A), and is over 1000 times faster. Structure models from the SNAP potential also show better agreement with the experimental metal-metal pair distribution functions compared to existing well-built Morse-BKS potentials.