Efficient Prediction of Vehicle Drag Coefficient Using Physics-Enhanced DeepONet

Under the "Dual Carbon" strategy, this study presents an automotive aerodynamic optimization method based on Scientific Machine Learning (SciML), which combines scientific computing with machine learning to address the high cost and low efficiency of traditional Computational Fluid Dynamics (CFD) simulations. We developed a hybrid neural operator model integrating physical information and data-driven approaches——yellow a deep learning model capable of learning and simulating complex physical processes—to achieve efficient prediction from vehicle geometry to pressure fields and drag coefficient (Cd). The study used the 3D ShapeNet Car dataset and designed an innovative lightweight DeepONet architecture, incorporating Navier-Stokes (N-S) equations into the loss function as prior information. Experimental results indicate that this method significantly improves simulation efficiency while maintaining computational accuracy, with only 0.09M parameters, 5% of Unet3D, and reducing errors by more than 15.8% compared to models like Unet3D.