Slowly Quenched, High Pressure Glassy B$_2$O$_3$ at DFT Accuracy

TL;DR

This paper develops a machine learning potential for B$_2$O$_3$ glass that accurately models its structure and properties under slow quenching and high pressure, aligning well with experimental data.

Contribution

The authors created a DFT-trained machine learning potential enabling realistic simulations of B$_2$O$_3$ glass at slow quenching rates and high pressures, surpassing previous methods.

Findings

MLP accurately predicts glass densification and structure.

Simulations match experimental structure factors.

High-pressure simulations reveal varied coordination geometries.

Abstract

Modeling inorganic glasses requires an accurate representation of interatomic interactions, large system sizes to allow for intermediate-range structural order, and slow quenching rates to eliminate kinetically trapped structural motifs. Neither first principles- nor force field-based molecular dynamics (MD) simulations satisfy these three criteria unequivocally. Herein, we report the development of a machine learning potential (MLP) for a classic glass, B$_2$O$_3$, which meets these goals well. The MLP is trained on condensed phase configurations whose energies and forces on the atoms are obtained using periodic quantum density functional theory. Deep potential MD (DPMD) simulations based on this MLP accurately predict the equation of state and the densification of the glass with slower quenching from the melt. At ambient conditions, quenching rates larger than 10$^{11}$ K/s are shown to lead to artifacts in the structure. Pressure-dependent X-ray and neutron structure factors from the simulations compare excellently with experimental data. High-pressure simulations of the glass show varied coordination geometries of boron and oxygen, which concur with experimental observations.