Machine learning modeling based on informer for predicting complex unsteady flow fields to reduce consumption of computational fluid dynamics simulation

Accurately predicting the dynamic behaviour of complex flow fields has always been a major challenge in Computational Fluid Dynamics (CFD) research. This paper proposes an innovative approach based on the Informer model for efficient prediction of unsteady flow fields. This study focuses on the two-dimensional National Advisory Committee for Aeronautics (NACA) 0009 hydrofoil wake vortices and introduces the machine learning model Informer to predict unsteady wake vortices. In the prediction area, a grid of 10 x10 monitoring points is established, and various error evaluation criteria are employed to assess the prediction results. Simultaneously, the predicted cloud maps at 24-time steps are compared with the CFD-calculated cloud maps to validate the feasibility of using machine learning models for predicting unsteady flow fields. The results indicate that the Informer model achieves favorable predictions for unsteady flow fields, with average Root Mean Square Error (RMSE) values along three paths being 0.0061, 0.017, and 0.0103, respectively. As the prediction length increases, the R-square (R2) increases from 0.9917 to 0.9984. The Informer model demonstrates commendable performance in predicting vorticity positions, sizes, and shapes, affirming its suitability for forecasting unsteady flow fields and consequently mitigating CFD computational resource consumption.