Dual-convolutional neural network based aerodynamic prediction and multi-objective optimization of a compact turbine rotor

Abstract

Abstract With the development of neural network technology, surrogate models and dimensionality reduction strategies based on machine learning have become the research hotspots of aerodynamic shape optimization recently. In order to further improve the accuracy and interpretability of the traditional surrogate models, this research establishes a deep learning model, named Dual Convolutional Neural Network (Dual-CNN) for the aero-engine turbines. The aerodynamic performances are predicted and the pressure, temperature fields are reconstructed for multiple rotor profile conditions. The prediction of efficiency is compared with the accuracy of Gaussian Process Regression (GPR) and Artificial Neural Network (ANN) models. The results show that the proposed Dual-CNN model can accurately reconstruct the fields, thus interpreting the mechanism for the change of aerodynamic performance. Dual-CNN is more accurate than GPR and ANN in predicting efficiency and torque, whose error is within an acceptable range of optimization. Then, efficiency and torque are selected as the objective functions to perform a gradient-based multi-objective optimization by the automatic differentiation method and a Pareto solution is obtained. The trained Dual-CNN provides rapid and accurate prediction of performance without CFD calculation in the optimization. Finally, the sensitivity to train size is analyzed for the Dual-CNN model, which indicates that the sampling of 1500 cases for eight design variables in this dataset enables Dual-CNN to achieve favorable effect of field reconstruction and performance prediction.