✨ TL;DR
This paper introduces GIST, a neural network surrogate that predicts race-car aerodynamics 10,000× faster than traditional CFD simulations while maintaining accuracy suitable for early-stage design. The work includes a new high-fidelity dataset of LMP2 race-car aerodynamics validated by industry experts at Dallara, enabling interactive design exploration in motorsport.
Computational Fluid Dynamics (CFD) is essential for developing race-car aerodynamics, but each high-fidelity simulation requires tens of thousands of core-hours, making it prohibitively expensive to explore large design spaces within realistic budgets. While AI-based surrogate models could accelerate this process, existing public datasets focus on simple passenger-car shapes with smooth surfaces. These datasets fail to capture the complexity of motorsport components—thin wings, complex multi-element assemblies, and highly loaded surfaces—that are critical for race-car performance, limiting the development and validation of surrogates for industrial motorsport applications. The challenge is compounded by the need for surrogates that can handle complex mesh topologies with tightly packed geometric features while maintaining physical accuracy and discretization invariance. Previous neural operators struggle with the intricate geometries and mesh connectivity patterns found in race-car aerodynamics, and there has been no validation of whether AI surrogates can achieve accuracy levels suitable for actual industrial motorsport design workflows.
The authors develop a three-part solution. First, they create a high-fidelity RANS CFD dataset based on a parametric LMP2-class race-car CAD model, covering six operating conditions including straight-line and cornering regimes. This dataset is generated and validated by aerodynamics experts at Dallara to ensure it preserves features relevant to industrial motorsport practice. Second, they introduce the Gauge-Invariant Spectral Transformer (GIST), a graph-based neural operator architecture designed specifically for complex aerodynamic geometries. GIST uses spectral embeddings that encode mesh connectivity information, enabling better predictions on tightly packed, intricate components. The architecture guarantees discretization invariance—meaning predictions are independent of how the geometry is meshed—and scales linearly with mesh size, making it computationally efficient. The model processes surface meshes as graphs and learns to map geometric parameters directly to aerodynamic flow fields and integrated forces.