超声速湍流密度脉动预测的神经网络方法

Neural network method for predicting density fluctuations in supersonic turbulence

针对气动光学效应对湍流密度脉动预测的需求,发展了超声速湍流密度脉动预测的神经网络方法,从直接数值模拟(DNS)的超声速湍流边界层的流场数据中挖掘规律,建立了包含5个隐含层的密度脉动模型。实验结果表明,所发展的神经网络方法可以很好地预测密度脉动均方值,它不仅能很好地预测训练样本,对测试样本预测的精度和稳定性也显著高于传统模型,且具有一定的泛化能力。通过特征选择和加入先验信息,确定了密度脉动模型的7个输入参数特征量,进一步提高了模型的泛化能力和实用性。

Considering the demand for predicting density fluctuation in turbulence of in prediction of the aero-optical effect,this paper develops a neural network method to predict the density fluctuation in supersonic turbulence.A model for the density fluctuation with 5 hidden layers is built by mining the regularity from the flow field data of supersonic turbulent boundary layers,which are simulated by the Direct Numerical Simulation（DNS）method.The experimental results show that the neural network method proposed is able to predict the mean square values of density fluctuation accurately.The method can predict the training samples well,and can get the prediction of test samples with better accuracy and stability than traditional models,and has certain capability of generalization.By selecting features and adding prior information,the 7 features of input parameters of the density fluctuation model are determined to further improve the capability of generalization and practicability of the model.