This paper obtains downstream flow field data under the positioning grid through experiments and simulations. The mixing factor evaluation system is used to calculate the mixing factor parameters, and a database of mixing characteristics is built. Additionally, we use two neural network architectures, Multilayer Perceptron (MLP) and Kolmogorov–Arnold Networks (KAN) to predict the mixing characteristic parameters. For the MLP model, by cross-validating activation functions, optimization functions, and optimizing parameters such as regularization coefficients and learning rates, the vortex factor determination coefficient reached 0.97, and the turbulence intensity determination coefficient reached 0.93, showing excellent performance. For the KAN model, through pruning and optimizing the model structure, the vortex factor determination coefficient reached 0.89, and the turbulence intensity determination coefficient reached 0.96.
For the MLP model, the SHapley Additive exPlanations(SHAP) method was used for model interpretability analysis, providing an intuitive understanding of how each feature affects different mixing characteristic parameters. For the KAN model, by analyzing the pruned network structure, the relationship between input features and mixing characteristic parameters can be clearly analyzed. Moreover, symbolic representation can output accurate mixing characteristic parameter equations, allowing for a quantifiable analysis of hidden data relationships.
 

Spacer grid; mixing vane; KAN; MLP； SHAP