Artificial intelligence is now seen as the next major computing revolution, with computational capabilities far beyond traditional algorithms. This technology has applications across numerous sectors, including Formula 1, where it is already being used in various areas. One of the most interesting is aerodynamic analysis, with the potential to reduce the reliance on wind tunnels and traditional tools in the future.
Existing technologies
CFD programs for virtual simulations start with equations describing physical phenomena to predict the behavior of an aerodynamic component, calculating how speed, pressure, and forces generated by the air interact with the surface. However, CFD models are always simplified compared to reality, as a complete physical analysis would be too time-consuming and resource-intensive. Since not every aspect can be considered, design teams must choose only the most relevant ones. This involves deciding which phenomena to analyze, such as whether to focus solely on aerodynamics or also on thermal aspects, whether to study stationary or transient behavior, and, most importantly, at what scale of turbulence, with vortices ranging from microscopic sizes to several meters in diameter.
Wind tunnels also have limitations, as they cannot accurately replicate how a car behaves in the real world. For instance, it is possible to analyze a car at various heights from the ground, but its movements on the suspension cannot be simulated, preventing data collection in more dynamic conditions. Additionally, the ability to study a physical phenomenon depends on the availability of sensors to measure key parameters, not to mention the difficulty in controlling air temperature, which modern cars are highly sensitive to. As a result, CFD and wind tunnels are useful tools for developing a rough idea of a particular solution, but only the track can provide the final verdict.
A reverse approach
Contrary to popular belief, artificial intelligence is not a recent development. Theoretical foundations were laid decades ago, but only in recent years have processors reached the necessary computing power to apply them. For example, in the 1990s, Benetton experimented with neural networks to improve the drivability and vehicle dynamics of their cars.
While CFD starts from physical laws to study the behavior of an aerodynamic solution, artificial intelligence algorithms work in the opposite way. They start with existing examples, identifying trends in available data to predict the behavior of new solutions or the same object under different conditions. AI’s “reasoning” is therefore deductive, based on known information. A prerequisite is building a database, a large sample of data that, in the case of aerodynamics, can come from the track, wind tunnels, or CFD analysis, which is necessary to train the model to recognize trends and predict different scenarios.
Examples
Formula 1 is no stranger to this technology, having already used it in developing the 2022 regulations. The focus then was on the aerodynamic disturbance caused by the car ahead, requiring a study of the turbulent wake, which was extremely complex to model and analyze in CFD. The FIA and Formula 1, however, also exploited the predictive power of machine learning, with algorithms maturing to the point where they could predict the behavior of the turbulent wake at a distance of one meter and up to 40 meters behind the car.
An earlier case was with Porsche, when between 2017 and 2018 they developed the EVO version of their 919 Hybrid LMP1 prototype. In designing the rear wing, the German manufacturer started with an existing database of 1,600 wing profiles, training a computational model to identify key parameters like chord, geometry, and curvature. The goal was to map the “genes” that described the DNA of each profile, identifying five key parameters. However, with five values per parameter, two overlapping flaps on the same wing, and other variables such as incidence and relative distance, the possible combinations exceeded 30 billion.
At that point, Porsche used another machine learning algorithm to solve the optimization problem. The model returned hundreds to a few thousand of the most promising configurations, which were then virtually simulated using traditional CFD analysis with a cluster. The sample size was progressively reduced until reaching the final wing configuration, which was then verified in a wind tunnel.
Possible coexistence
The Porsche example highlights how, in its various forms, artificial intelligence can cooperate with traditional simulation tools without necessarily replacing them. As seen, these models use deductive logic but are still limited in replicating human intuition. To explore completely new shapes, concepts, or situations, AI requires an existing database, whereas for entirely novel explorations, CFD or wind tunnels must still be used, starting from the fundamental physics. Furthermore, traditional tools can be employed to verify AI’s predictions, as was the case with the Porsche 919 Hybrid EVO.
This does not diminish AI’s immense potential, going beyond the limits of CFD and wind tunnels. In recent years, there has been much talk about teams investing in infrastructure, modernizing production departments, and building new wind tunnels. However, research and development in artificial intelligence could make an even bigger difference in the medium-term future, much like CFD did at the end of the 1990s and early 2000s.
Leave a Reply