The three-dimensional (3D) nanocrystalline models of lithium silicates with the log-normal grain size distribution are constructed by constrained Voronoi tessellation. During evolution process, the algorithm is improved. We proposed a new algorithm idea by combining Genetic Algorithm (GA) with Least Square (LS) method to make up for the disadvantages of traditional genetic algorithm which may be easily trapped in local optimal solution. In the process of modeling, it is the first time, to the best of our knowledge, that we keep the whole sample showing the charge neutrality by deleting the excess atoms on the polyhedron boundary during the modeling. By using the molecular-dynamics method, the relaxation procedure of nanostructured Li4SiO4 is carried out. The results show that the average mass density of the sample is slightly lower than the experimental data of the perfect crystal after relaxation process. In addition, boundary component proportion (BCP) and density reduction proportion (DRP) of the sample is obtained, respectively. The present results display a significantly reduced BCP but an increased DRP when increasing the mean grain size of the sample.