Prediction of swirling flow field in combustor based on deep learning

Abstract

This paper considers the application of machine learning to the modeling of the swirling flow field in a combustor. A deep learning prediction model based on experimental data from particle image velocimetry is established, with the state parameters and spatial parameters as input and the axial and radial velocities as output. An optimal network framework is determined by comparative analysis of goodness of fit, prediction error, and training cost. The swirling flow field is reconstructed using 1.5, 2.0, 2.5, and 3.0 kPa data sets, and an extrapolation prediction is made using a 4.0 kPa data set. The results indicate that the data-driven deep learning model is able to capture the nonlinear spatial characteristics of the swirling flow field and learn the complex relationship between input and output parameters. The results for the velocity contour distribution, streamline shape, and vortex center position of the swirling flow field are in good agreement with the experimental results. The constructed deep learning prediction model has good prediction accuracy based on previously seen data sets and is able to reconstruct the swirling flow field. In addition, it is able to perform extrapolation prediction based on a previously unseen data set and has some generalization capability. Finally, several potential engineering applications of the deep learning prediction model constructed here are pointed out.