In computational fluid dynamics (CFD), accelerating calculations is of great significance. During the numerical computation process, for non-uniform initialization, initialization is performed based on boundary and initial conditions and the known internal field distributions, assigning the values of initial pressure, velocity, and other field variables to each grid throughout the computational domain. Machine learning methods can model and characterize more complex field distribution properties based on a priori knowledge of historical data compared to interpolation algorithms. In this study, a neural network is applied to the generation of initial fields data in CFD to explore the applicability of accelerated convergence of the CFD computational process. A complex flow scenario with separated flow was used to validate and evaluate the effectiveness of the proposed method in accelerating computational convergence, using a case of blockage flow within a narrow rectangular channel. The results indicate that the method proposed in this study significantly improves computational efficiency when handling high-dimensional, nonlinear, complex flow calculations.

Computational fluid dynamics (CFD),Neural network modeling,Numerical computation,Non-Uniform initialization,Machine learning,Computational acceleration