In this work, we implement molecular dynamics (MD) simulations with deep neural network (DNN) potential trained with the datasets from ab initio calculations to determine the dielectric spectra of crystal. The fluctuations of the total dipole moment of crystal, which are obtained from MD, can be directly related to the frequency-dependent permittivity according to the work of Neumann and Steinhauser [Chem. Phys. Lett. 102, 508–513 (1983)]. We generalize their theoretical work to express the permittivity in the form of a tensor and perform MD simulations for cubic silicon carbide (3C-SiC) with 8000 atoms to assess the accuracy. The infrared resonance frequency and the phonon linewidth obtained by the DNN potential are compared with those obtained by the empirical Vashishta potential and experiments. The results of the DNN potential are in good agreement with the experimental measurements. It shows that we can carry out MD simulations for large systems with the accuracy of ab initio calculations to obtain dielectric properties.

No abstract

Using the reference hypernetted chain (RHNC) integral equation theory and a rigorous stability analysis method, we investigate the phase behavior of a mixture of hard-sphere dipoles and neutral hard spheres based on the correlations of the homogeneous isotropic phase. Lowering the temperature down to the points where the RHNC equations fail to have a solution, several fluctuations strongly increase. At low densities our results indicate the onset of chain formation, which is similar with the pure DHS system. At high densities, the results indicate the appearance of isotropic-to-ferroelectric transitions (small neutral hard spheres concentrations) and demixing transitions (large neutral hard spheres concentrations).