AgI exhibits superionic conductivity of Ag ions only above 147 °C. However, AgI-As₂Se₃ mixed glass shows high ionic conductivity (≈ 10^-3 S cm^-1) even at room temperature [1]. AgI-As₂Se₃ glass can be formed by melt-quench method over a wide AgI composition range (20~70 mol% AgI), and the ionic conductivity changes exponentially with changes in the AgI fraction [2]. In order to elucidate the structural origin of these behavior, we first performed a first-principles molecular dynamics (FPMD) study, but due to the time and space limitation of PFMD, it was not possible to accurately evaluate the diffusivity of silver ions and its relationship with the medium-range structure of the glass. In this study, we overcome these limitations by constructing a machine learning interatomic potential (ML-IP) that closely mimics first-principles calculations using Allegro [3], a type of E(3)-equivariant graph neural network, and running large-scale and long-time neural network potential molecular dynamics (NNP-MD) simulations. MD simulations were performed for 40 mol% AgI and 60 mol % AgI compositions, and the glass structure was obtained by quenching the atomic structure in the molten state at 2000 K over 10 ns to 300 K. The diffusion of silver ions was measured for 100 ns for the obtained glass structure, and long-range diffusion over 20 Å, which was not captured by the FPMD calculation, was observed. The calculated silver ion conductivity from the NNP-MD simulation trajectory reproduces the experimental values well for the 60 mol % AgI composition, indicating that the simulation was performed reliably. The conductivity at 40 mol% AgI composition was lower than that at 60 mol% AgI composition, and the change in conductivity with composition was reproduced. On the other hand, the conductivity was overestimated from the experimental value, and there are some remaining issues such as insufficient calculation time and the possibility of insufficient reproduction of glass structure due to too fast quenching rate. Analysis of the number density distribution of silver in the resulting glass structures revealed the presence of local sites of silver heterogeneously embedded in the glass matrix and diffusion pathways connecting them. Analysis of the free energy surface of silver calculated from the number density distribution reveals that the local sites with a higher coordination number of iodine tend to have lower activation energies between the local sites.