Benchmarking Convolutional Neural Network and Graph Neural Network based Surrogate Models on a Real-World Car External Aerodynamics Dataset

Aerodynamic optimization is a critical aspect in the development of contemporary vehicles, aiming to achieve enhanced eco-friendliness, improved aerodynamic performance, and appealing aesthetics. This process necessitates close collaboration between aerodynamicists and stylists, which is often hindered by the substantial computational time required for traditional aerodynamic simulations. Surrogate models present a viable methodology to mitigate this computational overhead, offering accelerated predictions. This research undertakes a comparative evaluation of two distinct surrogate modeling paradigms for the prediction of aerodynamic drag, utilizing a real-world automotive dataset. The methodologies investigated include a Convolutional Neural Network (CNN) model, which processes signed distance fields as input, and a Graph Neural Network (GNN) based commercial tool, which

operates directly on surface meshes. Unlike prior investigations that typically rely on datasets derived from parameterized geometries, the dataset employed in this study comprises 343 unique geometries originating from 32 baseline vehicle designs across five distinct car projects. This diverse dataset accurately reflects the complex, free-form modifications commonly encountered throughout the vehicle development lifecycle. Experimental results indicate that the CNN-based approach achieves a mean absolute error of 2.3 drag counts, while the GNN-based method yields a mean absolute error of 3.8 drag counts. Both methodologies demonstrate an approximate 77% accuracy in predicting the directional change of drag relative to the baseline geometry.

Message: Hi FE, I need help with “Benchmarking Convolutional Neural Network and Graph Neural Network based Surrogate Models on a Real-World Car External Aerodynamics Dataset” in “Aerospace Engineering”