Fast prediction of car driving direction velocity field based on convolutional neural network with data of flow simulation nodes after feature enhancement

• Published in: Advanced engineering informatics Vol. 64; p. 103045
• Main Authors: Shen, Shengrong, Han, Tian, Pang, Jiachen
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2025
• Subjects: CNN; Computer-aided design; Drag coefficient; Flow velocity fields predict; CNN; Flow velocity fields predict; Computer-aided design; Drag coefficient
• ISSN: 1474-0346
• DOI: 10.1016/j.aei.2024.103045

•A deep learning method has been introduced into the flow velocity field analysis of a car profile.•A simulation node data enhancement method is proposed based on the principle of matrix mapping and completion as a deep learning model input.•One method for extracting fluent field value features is carried out from the side profile of cars. The analysis of the drag coefficient of the car is a crucial step during the entire car design process, which is typically conducted through wind tunnel testing or fluid simulation with fine modelling and parameters setting, and drag coefficients are obtained through extensive calculations. A convolutional neural network-based method for predicting the flow velocity field of the car profile is proposed. The method can acquire the corresponding car side profile flow data with significantly less time than traditional simulations. The core of this method is to take the node coordinates obtained from the simulation mesh division as the basic data, which are mapped into a matrix and disassembled as the input data to the convolutional neural network to obtain an accurate distribution of fluid velocity values in the car driving direction. The method develops the prediction model by mapping the car profile simulation data and using convolution and anti-convolution structures. For drag coefficient ranging from 0.28 to 0.30, the average percentage error between the model predicted and simulated values around the car profile for real car profile flow fields is maximum 9.69%, and minimum 2.11%.