Aerodynamic simulations are crucial in vehicle design and performance evaluation. Traditionally, these simulations utilize Computational Fluid Dynamics (CFD) techniques to compute flow quantities such as velocity, pressure, and wall-shear stresses. Accurate prediction of these quantities is vital for estimating drag and lift forces, which directly impact fuel efficiency, stability, and acoustics. This study focuses on developing an AI surrogate for aerodynamic design of production mideo-size SUVs using NVIDIA’s PhysicsNeMo framework.
Firstly, high-fidelity 3D CFD data are generated using first-principles solvers on 102 different geometry variants at a uniform inlet velocity of 38.89 m/s and a fixed set of boundary conditions. The DoMINO (Decomposable Multiscale Iterative Neural Operator) AI model, part of the PhysicsNeMo framework, is then used to train on this dataset, accurately predicting surface pressure and flow fields around vehicles for rapid estimation of critical aerodynamic metrics such as drag and lift. DoMINO is a neural operator that learns local geometry representations from point cloud data and predicts PDE solutions on discrete points using dynamically constructed computational stencils in local regions. By leveraging both short- and long-range geometric features, the DoMINO model predicts solution fields on the vehicle’s surface and in the surrounding flow domain—capabilities essential for informed design and engineering decisions in industrial applications.
In this study, the DoMINO model is evaluated on a realistic production mid-size SUV vehicle designed by General Motors. Comprehensive hyperparameter tuning is conducted to optimize model performance, along with an analysis of input grid sensitivity. The findings highlight DoMINO's effectiveness as a robust and accurate tool for aerodynamic analysis in the automotive sector, enabling accelerated design cycles and enhanced vehicle performance.