Fast flow simulation method based on POD and BPNN

基于POD和BPNN的流场快速计算方法

Received: 31 March 2021

Abstract

It is computationally expensive to perform a large number of flow analysis by using high-fidelity CFD simulation. The paper proposes a fast flow-analysis method based on the proper orthogonal decomposition(POD) and back propagation based neural network(BPNN). First samples are generated in the geometric parameter space. Then a POD model is built to map the high-dimensional flow-field data to low-dimensional base modal coefficients, and further a BPNN model is fitted from geometric parameters to base modal coefficients to achieve fast flow prediction. During constructing the POD and BPNN model, the partitioning strategy and K-means clustering are implemented to improve modeling efficiency, as well as to reduce model training time. The results of predicting the steady flow of variable geometries show that: at subsonic, the trained model possesses good accuracy, especially in predicting the pressure isolines of the flow field and the pressure coefficient distribution of the airfoil. The average prediction errors of the lift and drag coefficient are smaller than 0.4%, while they are smaller than 1.4% at transonic. The shock location can be well predicted as well.

摘要

采用高精度CFD仿真进行大量流场分析存在计算成本高、耗时长的问题。提出了一种基于本征正交分解（proper orthogonal decomposition，POD）和反向传播神经网络（back propagation based neural network，BPNN）的流场快速计算方法。在几何参数化设计空间中抽样，然后利用POD将高维流场数据映射到低维基模态空间，并用BPNN建立几何参数到基模态系数的多层神经网络模型，实现流场快速预测。在POD和BPNN模型构建中分别引入分区和聚类取样策略，以提高建模效率，降低模型训练耗时。变几何翼型的定常流场案例结果表明：在亚声速情况下，训练所得的模型可以保证流场中等压线、翼面压力系数等信息的预测精度，其升阻力系数平均预测误差在0.4%之内；在跨声速情况下，训练所得的模型升阻力系数平均预测误差在1.4%之内，并且激波位置也可以得到较准确的预测。