In computational fluid dynamics, mesh quality greatly affects the accuracy and computational efficiency of numerical simulation results. The Bubble does not require the consideration of intersection judgments and has a relatively simple data structure, which has significant advantages in mesh generation efficiency and quality. The process of improving the mesh quality by moving nodes based on the traditional Bubble is optimized in this article, and we define it as the Bubble-Opt method. In this method, a bubble radius selection method combined with neural networks is used to generate the initial bubbles, and an improved bubble dynamic movement technique is used to adjust the bubbles to the appropriate position. The Delaunay method is used to connect the center of bubbles to form the final optimized mesh. Then, the optimization effects of different bubble radius selection methods and Bubble-Opt methods are compared under different process parameters. Taking the flow around a 2D cylinder as an example, the geometric quality and transition ratio of the mesh before and after optimization are tested. For this example, there is a set of optimal parameters and a radius selection method that achieve the best mesh quality optimization effect. The average transition ratio can be improved by about 17.37%, the average mesh quality can be improved by about 13.60%, and the minimum transition ratio and minimum mesh quality can be significantly improved. Finally, under the radius selection method and process parameters, taking two-dimensional cylindrical flow and NACA0012 airfoil flow as examples, the numerical simulation results are compared with experimental data from both qualitative and quantitative perspectives, indicating a significant improvement in the overall grid quality.