Boron coordination and three‐membered ring formation in sodium borate glasses: a machine‐learning molecular dynamics study

Abstract

Classical molecular dynamics (CMD) simulations that employ analytical force fields have been commonly utilized to investigate mechanical, chemical, and thermal properties of oxide glasses owing to their superior computational efficiency. Conversely, simple functional forms limit the accuracy in modeling complicated glass structures, specifically, in alkaline borate glasses, which exhibit boron coordination numbers that vary nonlinearly with changes in glass composition and temperature. Machine‐learning potentials (MLPs), which are trained using datasets on energy and force evaluated via the density functional theory (DFT), are garnering significant attention as a novel simulation technology for enhancing the accuracy in modeling materials. Therefore, this study applied a universal MLP, PreFerred Potential (PFP) (trade‐name: Matlantis), to model sodium borate glasses, and its accuracy was verified in reproducing boron coordination and ring structures by comparing its results to available experimental data. We found that PFP can quantitatively reproduce the boron coordination change with glass composition without any empirical correction, while the boron coordination in the melts at high temperatures is overestimated, even though the qualitative variation was better estimated than CMD simulations. Furthermore, the MLP could generate many 3‐rings, unlike the analytical force‐field. Accordingly, we demonstrated superior accuracy of the MLP in modeling alkaline borate glasses, while discussing the challenges faced in reproducing the elaborated microstructures in borate glasses.