On-the-fly machine learning force fields for alkali silicate glasses

Molecular dynamics simulations can provide structural information of glass materials with atomic resolution and for very short time scales. In turn, this requires accurate and effective force fields. Here, focusing on alkali silicate glasses, we develop transferable force fields that are applicable across a wide range of temperatures and compositions. Using on-the-fly machine learning routines, these force fields produce energy and force predictions with accuracy comparable to that of density functional theory. The potential energy landscape is described using structural descriptors that incorporate both radial and angular information within a reduced-dimensional descriptor space. We then employ these accurate machine learning force fields (MLFFs) to prepare silica and binary alkali $({\mathrm{Li}}_{2}\mathrm{O},$ ${\mathrm{Na}}_{2}\mathrm{O},$ ${\mathrm{K}}_{2}\mathrm{O})$ silicate glass samples, allowing us to study their structures. We assess the composition dependence of the structural characteristics of the glasses by computing several key parameters, including neutron structure factor, pair distribution function, angular distribution function, coordination number, and ${Q}^{n}$ speciation. The trained MLFFs accurately reproduce the splitting of the first sharp diffraction peak in potassium silicate glasses, consistent with experimental findings, and reveal that its structural origin is tied to Si-K interactions. Overall, our research demonstrates the power of combining advanced computational techniques with classical structural analysis methods to enhance our structural understanding of complex glass systems.