Investigation on compressor flow field reconstruction using deep neural networks with a fine-tuned strategy

Abstract

In numerical simulations, finer grids generally yield higher accuracy compared to coarse grids, but significantly more computational resources and time are demanded. In scientific researches, it is often necessary to compute many cases with similar boundary conditions. Traditional Computational Fluid Dynamics simulations solve each case individually, thereby ignoring the shared characteristics among similar cases, which leads to inefficient use of computational resources. To address the issue, this paper proposes a flow field reconstruction model for compressors based on deep neural networks combined with a fine-tuned strategy (FTDNN). This model captures common flow patterns among similar cases to enable fast and high-precision predictions for unseen cases. On a variable aerodynamic dataset, a deep neural network with two prediction modes is constructed: one is a pre-trained mode, and the other uses a fine-tuned strategy for high-accuracy predictions. In deep learning models, generalization under multiple operating conditions and precision under specific conditions often present a trade-off. This paper quantitatively analyzes the relationship between the number of frozen hidden layers and prediction accuracy. It is ultimately found that when the number of frozen and trainable hidden layers is equal, a balance between generalization and accuracy is achieved, resulting in optimal prediction performance. The FTDNN model achieves a maximum average relative error of only 1.29% on extreme flow fields near the upper limit of positive attack angles in the test set. The model also demonstrates excellent accuracy in predicting flow fields at the compressor blade leading edge, within boundary layers, and in wake distributions.