Abstract
Computational Fluid Dynamics (CFD) is crucial for automotive design,requiring the analysis of large 3D point clouds to study how vehicle geometryaffects pressure fields and drag forces. However, existing deep learningapproaches for CFD struggle with the computational complexity of processinghigh-resolution 3D data. We propose Factorized Implicit Global Convolution(FIGConv), a novel architecture that efficiently solves CFD problems for verylarge 3D meshes with arbitrary input and output geometries. FIGConv achievesquadratic complexity $O(N^2)$, a significant improvement over existing 3Dneural CFD models that require cubic complexity $O(N^3)$. Our approach combinesFactorized Implicit Grids to approximate high-resolution domains, efficientglobal convolutions through 2D reparameterization, and a U-shaped architecturefor effective information gathering and integration. We validate our approachon the industry-standard Ahmed body dataset and the large-scale DrivAerNetdataset. In DrivAerNet, our model achieves an $R^2$ value of 0.95 for dragprediction, outperforming the previous state-of-the-art by a significantmargin. This represents a 40% improvement in relative mean squared error and a70% improvement in absolute mean squared error over previous methods.