Various numerical models based on mathematical physics equations in geosciences, including numerical modeling of subduction dynamics, necessarily involve computational grids consisting of discrete nodes, which control the accuracy of numerical calculations and thus the credibility of numerical models in solving practical geoscientific problems. The numerical modeling of subduction dynamics has made great progress in recent years, however, the numerical accuracy due to the arbitrary use of computational grids is still unclear. In this paper, we construct numerical models based on three sets of computational grids with different resolutions for a classical scientific problem, and evaluate the possible impact of low- resolution grids in practical research. It is argued that the computational grid with a resolution of 2 km×2 km for the encrypted zone, which has been more commonly used in the last decade, is likely to obtain computational results containing significant numerical errors, which in turn affects the application of numerical modeling in subduction dynamics. Therefore, it may be necessary to revisit models with low resolution grids and their corresponding geologic conclusions in recent years. As numerical models of subduction dynamics become more and more highly nonlinear, the choice of computational grids with the highest possible resolution may be inevitable. For the case of using low- resolution grids for highly nonlinear problems, definitive evidence of grid reliability is needed. We propose a new set of grid pattern suitable for subduction dynamics: locally encrypted structured quadrilateral grids containing hanging nodes. This grid may accomplish high- precision numerical calculations with a small total number of grids and is relatively simple to implement.