基于分层卷积自编码器的钝体湍流流场降阶分析

Reduced-order modelling of a bluff body turbulent wake flow field using hierarchical convolutional neural network autoencoder

本文采用了一种非线性的分层卷积自编码器,类比于本征正交分解的方法,可以对提取到的低维特征进行能量排序,同时又能在一定范围内达到更高的降阶性能。文中以雷诺数为20000的三维圆柱钝体湍流尾迹流动为例,分析分层卷积自编码器对该流场的降阶能力,并与本征正交分解的结果作对比。同时,在此基础上延伸出组合模态的概念,并增加每一组低维潜在向量数,观察重构流场与原始流场的均方误差变化。结果表明,在分层层数及每层潜在空间向量数较少时,非线性的分层卷积自编码器相对于本征正交分解的方法对流场有更强的还原能力,但是其优势会随着层数和潜在向量数的增加而减弱。

In this study,a nonlinear hierarchical convolutional autoencoder(H-CNN-AE)is employed to sort the energy content of the latent vectors of AE,which is analogous to the method of proper orthogonal decomposition(POD),and at the same time result in better performance in terms of reduced-order modelling within limits.This method is applied to a turbulent wake behind a three-dimensional circular cylinder bluff body at Re=20000.We assess the ability of H-CNN-AE with L2 error and compares it with the results of POD.Furthermore,the concept of grouping AE-modes is extended.We observe the change of mean square error between the reconstructed and the original flow when adding the number of latent vectors of each group.It is demonstrated that when the number of subnetworks and low-dimensional vectors in latent space of each subnetwork is small,H-CNN-AE has better capability to restore the flow field than POD.However,it is also found that the strength of H-CNN-AE will weaken with the increase of the number of subnetworks and latent AE modes and will even be inferior to POD under certain conditions.