Whether it’s an IndyCar or a Le Mans Hypercar, driving a race car to victory takes grit and skill, but above all, raw speed. This puts smart engineering at the heart of any winning strategy.

In the search for a competitive edge, engineers iterate endlessly on new race car designs, tweaking the geometry of the body, front wing, rear diffuser, and tiny add-on aerodynamic components, to see if any of these new configurations can translate to meaningful performance gains on the track.

Computer-aided design and simulation have augmented the creative process, making it faster and easier to test new ideas in a virtual environment. Engineers today use computational fluid dynamics (CFD) to predict how subtle design changes will affect aerodynamic forces like downforce and drag, as well as indirectly influencing vehicle handling, balance, and efficiency which determine how a vehicle behaves at speed and under different track conditions.

CFD is a powerful tool, but solving the complex equations that describe air flows and other physical phenomena takes time. Aerodynamic engineers can spend several hours evaluating a narrow surface adjustment and weeks to months designing an entire car.

IBM Research and the Dallara Group recently teamed up to see if they could speed up these evaluations while preserving the underlying physics at a high level of accuracy.

Dallara designs and supplies high-performance cars for some of the world’s top racing series, including IndyCar, Formula 2, and the IMSA WeatherTech SportsCar Championship. IBM is a leader in classical and quantum computing, with a portfolio of ground-breaking AI models designed to predict flood and wildfire extents, weather and climate, solar dynamics, and even fusion plasma dynamics.

Together, the two companies want to advance vehicle design and optimization using AI surrogate models trained on physics-based simulation data and, eventually, explore how they might incorporate quantum algorithms into the workflow to improve simulation fidelity.

In one experiment, Dallara took a Le Mans Prototype 2 (LMP2)-like car and designed several versions of a rear diffuser, the part under the body that increases downforce and the car’s grip on the track. Engineers then evaluated the designs using traditional CFD simulation as well as an AI surrogate model built by IBM and trained on Dallara’s proprietary simulation data.

Both methods correctly identified the optimal design, but the CFD analysis took several hours while the AI evaluation was finished in about 10 seconds. Dallara estimates that applying the AI surrogate to a typical set of several hundred geometry configurations could cut simulation time from days to minutes.

These preliminary results are reported in a new white paper aptly titled Faster by Design, that IBM researcher Cristiano Malossi recently presented at the AI & PDE Workshop at ICLR 2026.The model itself is a graph-based neural operator designed by IBM researchers. Called the Gauge-Invariant Spectral Transformer (GIST), the model addresses the key challenge of how to efficiently process graph-structured simulation data based on partial differential equations.

Scaling the graph transformer

From eyeglasses to airplanes, many of today’s mass-produced products are designed on the computer using a 3D grid, or mesh, that defines the product’s shape, volume, and surface contours through a dense network of points, links, and planar surfaces.

Previous graph models for predicting aerodynamic forces treated a car's mesh as a simple point cloud. While a collection of points may be fine to describe the geometry of the typical passenger car, race cars have finely detailed aerodynamic components, from the front flick to the gurney on the rear wing, that can have outsize effects on vehicle performance.

Points on opposing sides of a mesh fabric may be close physically, but subject to opposite forces aerodynamically. For an AI model to treat these surfaces properly, it needs to be taught how points on the mesh are connected. IBM’s GIST architecture encodes the coordinates of both the mesh’s points and its connecting links to more accurately capture its topology. The result is sharper, more physically accurate predictions, particularly for intricate components like the ones that matter most for advanced race car performance.

Capturing these complex relationships, however, is computationally intensive. To reduce the complexity that limited previous graph transformers from scaling, the researchers generated graph embeddings based on random projections, a compressed sensing technique known for its power and simplicity.

They also designed a “gauge-invariant” transformer architecture to ensure the model could seamlessly generalize across different embedding projections and mesh densities. In physics, gauge-invariance means that the fundamental properties of a system remain constant regardless of the arbitrary mathematical choices used to describe it.

“The coordinates of Boston and New York may vary depending on whether you use GPS or a 2D map, but the actual physical distance between the two cities does not," said Mattia Rigotti, one of the main IBM researchers behind the work. “Coordinate choices are like gauge choices. Physical relationships, whether it’s distances between cities, or the structure of points and links in a graph, are gauge-invariant. We carried that insight into our model design."

Researchers trained and tested the model on Dallara’s LMP2 dataset, which includes simulations under different track conditions and maneuvering, from high-speed cornering to heavy braking. Each configuration in the dataset was reviewed and verified by Dallara engineers against expected aerodynamic behavior. Evaluated on an unseen subset of the data, the GIST model predicted aerodynamic forces like pressure and shear stress more accurately than other leading AI surrogates.

“We’re honored that IBM chose Dallara for this innovative collaboration,” said Elisa Serioli, who leads CFD methodology at Dallara. “IBM’s AI-based model predictions show highly promising results compared to traditional CFD, and we see encouraging levels of generalization, which is particularly compelling within Dallara’s multi product environment.”

The road ahead

IBM and Dallara plan to refine their AI surrogate model by bringing in more of Dallara’s proprietary high-performance car data. But they are also looking ahead to the next era in computing and starting to explore how quantum and quantum–classical approaches could further improve and accelerate the design of high-performance vehicles.

By combining Dallara’s expertise in vehicle engineering and CFD-driven design with IBM’s leadership in AI and quantum computing, researchers will explore ways to augment traditional simulation workflows and eventually put the innovative designs they discover to practical use.

The implications go beyond the racetrack. Aerodynamic design is important for passenger cars and commercial planes, among other vehicles, and for any industry where aerodynamics figures into the design process. Reducing drag in everyday vehicles by a few percentage points can add up to meaningful improvements in fuel efficiency, creating savings at the pump as well as reduced carbon emissions.