Simulation of structure and dynamics of phosphate glasses with fluoride additives: A hybrid approach combining universal and specialized neural network potentials

IA Balyakin and KP Arslanov and MI Vlasov, COMPUTATIONAL MATERIALS SCIENCE, 258, 114065 (2025).

DOI: 10.1016/j.commatsci.2025.114065

The universal neural network potential (CHGNet model) was employed to perform molecular dynamics simulations of thermophysical and structural properties in the complex oxide glass system Na2O Al2O3 Fe2O3 P2O5 B2O3 NaF. A DeePMD model was subsequently trained on CHGNet data to enhance computational efficiency while maintaining accuracy. Through a series of molecular dynamics cooling simulations at various rates, both with and without NaF addition, we identified B, Al, and P as network-forming cations, while Fe and Na were found to be network modifiers. The study revealed significant structural transformations during cooling. Due to the multicomponent nature of the system, different atomic species exhibited distinct dynamic slowing behaviors, making the determination of glass transition temperature from MD simulations particularly challenging. The introduction of small amounts of sodium fluoride was shown to disrupt the oxygen network by creating terminal sites, thereby reducing the fraction of bridging oxygen atoms and compromising the overall network connectivity. CHGNet demonstrated robust performance in modeling this complex oxide glass system. Future research directions include expansion of modeling parameters and refinement of the models using advanced first-principles methods.

Return to Publications page