AI Surrogates Promise to Slash CFD Simulation Time from Hours to Seconds in Motorsports

Since the introduction of wings to racing cars halfway through the 1960s, airflow has been everything in racing. Until that point, the focus was on making a car as slippery as possible; less drag meant more top speed on the straights. Then designers like Jim Hall at Chaparral and Colin Chapman at Lotus realized they could use the air to push the car onto the track, increasing grip and allowing it to go faster through the corners. Things haven't been the same since.

Finding aerodynamic downforce started as something of a dark art. The use of wind tunnels to simulate its effect on scale models of cars was in its infancy, so teams were mostly limited to expensive and sometimes dangerous track testing. But wind tunnels can run day and night, rain or shine, and you can't crash a car or injure a driver (or worse) in the process. Wind tunnel work became even more important when F1 began restricting on-track testing to help teams cut budgets. Consequently, teams would do as much work with models as possible before validating the results during the limited test sessions they were allowed.

Computational fluid dynamics (CFD) simulation came next. In racing, everyone is looking for an advantage over their competitors, and it was finally possible to model, with some fidelity, the effect of airflow on a virtual model of a car. Not only were CFD sims cheaper than wind tunnel time, but they were also much faster at iterating. Early design work is now done in silico before being validated with scale models in a wind tunnel, as most series—including Formula 1, the World Endurance Championship, Formula E, and NASCAR—have tightly restricted on-track testing.

But as CFD has gotten more capable, it has also become more expensive. It can take thousands of hours of processor time to model a car, and tens of thousands more once you start to explore the effect of things like pitch and yaw. This, too, has become a new bottleneck for motorsport teams, which is why they're increasingly looking to AI to serve as the next helper.

Unlike in an office environment, there's nowhere to hide when it comes to race car design: Either the car is competitive or it isn't. If a tool doesn't help that happen, no one's boss will tell them to use it anyway.

Today, IBM and Dallara published new research showing it's possible to train AI surrogates to run simulations in seconds that would take hours conventionally, and with comparable error margins. You might not have heard of Dallara, but you've seen its cars—it designs and builds the entire Indy Car and Super Formula grids, as well as numerous single-seater feeder series. It has also made plenty of sports prototypes throughout the years, building carbon-fiber chassis for Le Mans programs like Audi and Ferrari, as well as the backbone or spine that BMW, Cadillac, and soon McLaren will use for their prototypes.

In this study, IBM fed its new Gauge-Invariant Spectral Transformer neural operator a gigantic bucket of CFD data on a simulated LMP2 sports prototype (think the fastest-but-one class at Le Mans). Until now, publicly available CFD models have focused on smooth road car shapes, but Dallara's dataset allowed the company to model things like how wakes from rotating wheels interact with the car's shaped underfloor, both in steady-state and cornering conditions.

After finding that GIST works better than other public tools—thanks to retaining "the flexibility of a point-cloud representation while rigorously treating the surface as a manifold mesh,"—the authors show it's as accurate as a conventional CFD simulation in modeling drag and downforce coefficients across a range of rear diffuser angles. But GIST was able to do it in seconds on a single CPU, "compared to the tens of thousands of core-hours an equivalent CFD campaign would require," the researchers wrote.

In the rarefied world of F1, the use of AI as a way to boost CFD work has been underway for a few seasons now. Not content with limiting real-world testing, F1 now also strictly limits the number of hours a team can use a wind tunnel—which can be only 60 percent scale—as well as the number of hours of CFD simulations. This is partly to reduce costs, but it's also used to level performance—the higher you finish in the championship one year, the less wind tunnel and CFD time you're allowed.

So teams like Red Bull have turned to Neural Concept, a startup helping at least four F1 teams use machine learning to model aerodynamics and challenges like how to cool the cells in the hybrid power unit's battery pack.

"It's more a way to extract all the value you can get from your CFD credits and track testing and internal time that you have. But yeah, it's really a way to go from 100 or 1,000 CFD runs to be able to have 1 million data points at the end of the day," said Pierre Baqué, CEO and founder of Neural Concept.

"It sounds magical, but the reality is that the accuracy of the model is only guaranteed within a specific range of situations that are not too far from what you have already explored," Baqué told me. "So all the trick and the gap from the idea to the value is to find what are the right workflows, what kind of data do I need to generate to be able to explore what kind of configurations afterward in which type of setting, and how often do I need to retrain my model, all the data hygiene around the design workflow."