A novel automotive side window flow field excitation prediction process based on frequency segmented training method

Abstract

The burden of time and computation of full-vehicle computational fluid dynamics (CFD) simulations hinders research and development (R&D) in automotive aerodynamics. This article introduces a rapid prediction process that is based on neural network modeling to tackle this difficulty. The process establishes a mapping relationship between acoustic excitations and styling geometric parameters to enable quick evaluation of candidate schemes. Furthermore, to overcome the challenges of high complexity in neural network modeling and substantial sample requirements encountered in engineering applications for the process, this research innovatively developed a frequency segmented training strategy based on the varying contributions of geometric parameters to acoustic responses at different frequencies. The strategy involves firstly quantifying the frequency-domain contribution weights of geometric parameters to acoustic excitations using the ridge regression. Subsequently, the entire frequency domain is divided into multiple sub-bands based on the weight distribution, and neural network prediction models are constructed for each sub-band. These models are then integrated to achieve rapid excitation prediction across the complete frequency spectrum. The validation results from two separate samples indicate that the relative errors for acoustic pressure excitation predictions are 1.06 %/0.99 % (with average errors of 0.64 dB/0.56 dB), while those for hydrodynamic pressure predictions are 0.17 %/0.20 % (with average errors of 0.14 dB/0.17 dB). The prediction process facilitates the rapid and precise evaluation of the acoustic excitations of schemes during the initial stage and could direct the screening of candidate schemes, thereby reducing the burden of CFD validation and enhancing the R&D efficiency of automotive aerodynamic design.