IBM Research and Dallara: AI surrogate model GIST evaluates racing car aerodynamics in 10 seconds instead of hours of classical CFD simulation

IBM Research and Dallara, the Italian chassis manufacturer for racing series including LMP2 and IndyCar, announced on April 30, 2026, a joint collaboration in which an AI surrogate model takes over part of the aerodynamic design process. The model, named GIST, replaces expensive classical CFD (computational fluid dynamics) simulation in the early iterative phase and reduces individual design evaluation from several hours to approximately 10 seconds.

How does GIST replace computational fluid dynamics?

GIST (Gauge-Invariant Spectral Transformer) is a graph-based neural operator that represents the output velocity and pressure field around a geometry in the spectral domain, trained to be invariant to the choice of local coordinate system. In practice, the model learns to map a 3D geometry mesh to a predicted flow field that is close enough to the CFD reference simulation for ranking design candidates. The specific use case is the optimization of the rear diffuser of an LMP2-type vehicle, where Dallara typically generates hundreds of geometric variants and ranks them by aerodynamic efficiency. CFD analysis of each variant takes several hours, and Dallara estimates that with GIST, iteration time could drop from days to minutes. The paper, titled “Faster by Design,” was presented at the AI & PDE Workshop within the ICLR 2026 conference.

What’s next: quantum methods and broader application

The team on IBM’s side is led by Mattia Rigotti, while CFD methodology on Dallara’s side is led by Elisa Serioli. The next planned research phase is integrating quantum and hybrid quantum-classical computing into the optimization pipeline, which IBM believes could further accelerate the search through design space. The authors emphasize that the application is not limited to racing: the technology is relevant for consumer vehicles where the goal is reducing aerodynamic drag for fuel efficiency, and for commercial aviation aerodynamics where similar requirements for rapid geometry iteration apply.