CoreWeave says AI halved Aston Martin F1’s aerodynamic prediction error
The model has guided pre-race setup choices throughout 2026. The newly disclosed result measures prediction accuracy, not racing performance.
Loading page…
The model has guided pre-race setup choices throughout 2026. The newly disclosed result measures prediction accuracy, not racing performance.
Listen to this story
CoreWeave and Aston Martin’s Vehicle Science team built and deployed a model that estimates how wind-tunnel aerodynamic predictions will diverge from track performance, informing component and setup choices before each 2026 race. Developed in six weeks and validated against aerodynamic sensors on the car, it reportedly cut prediction error by 50%; CoreWeave did not disclose the error metric or absolute error levels. The result concerns prediction accuracy—not lap-time gains—and Aston Martin can retrain the workflow as it gathers more race data.
The teams delivered the model before the 2025 season finale in Abu Dhabi, after six weeks of joint work.
Telemetry arrives at 100 readings per second, but noise such as curb strikes leaves engineers with limited useful data from each race weekend.
The teams jointly chose and prepared model inputs, then compared statistical and classical approaches with a newer modeling approach.
Aston Martin’s Formula One engineers are using machine learning to anticipate where wind-tunnel predictions will miss reality before a car reaches the circuit. In a November 16 disclosure, CoreWeave said a model developed with the team cut that prediction error by 50%—a result from a system already used throughout the 2026 season.
The starting problem was a gap between controlled testing and racing conditions. Aston Martin uses its wind tunnel to measure downforce and drag and help choose components and configurations. But on track, aerodynamic behavior changes under braking, through corners and over curbs. Track temperature, surface conditions and disturbed air behind another car add further differences.
Engineers call the difference between the prediction and the actual performance the delta. Before this project, the team learned the car’s true aerodynamic behavior during Friday practice and adjusted from there. That learning informed setup changes, the team’s understanding of the car’s aerodynamics and the calibration of its driver-in-the-loop simulator, which incorporates a human driver.
Anticipating that difference changes the component decision. If a floor or front wing will deliver less downforce on track than in the tunnel, engineers can account for the shortfall before the car leaves the factory. The model’s assignment is to predict that correction for the circuit ahead.
CoreWeave’s Physical AI Field Engineering group worked alongside Aston Martin’s Vehicle Science team for six weeks, delivering the model before the 2025 season finale in Abu Dhabi. The collaboration began with how engineers actually make setup decisions, rather than with a predetermined modeling technique.
The car supplies readings including wind speed, vehicle speed, steering angle and ride height. Telemetry arrives at 100 readings per second, but the volume is misleading: a curb strike can produce a sudden jump in a sensor reading. Engineers reduce the sampling rate to find useful patterns, leaving a small dataset from each race weekend.
The two teams jointly selected and prepared the signals the model would use. CoreWeave credits that work with making the limited data usable. They began with statistical methods, then tried classical modeling techniques before finding that what CoreWeave describes as a modern modeling approach performed better than both.
Experiments were recorded in Weights & Biases Models so the approaches could be compared and reproduced before the team chose one. CoreWeave says the final model beat Aston Martin’s existing solution, with the reduction in prediction error checked against aerodynamic sensors on the car.
The wind tunnel will always be our reference, but the track is where performance is decided.
Enrico Cardile, chief technical officer, Aston Martin Aramco Formula One Team
During the 2026 season, Aston Martin has used the model in the week before each race to guide component choices and setup. Aston Martin’s own November 16 announcement also describes combining wind-tunnel and real-track data to help engineers focus their optimization work where it will matter most on the car.
The published result has a narrow meaning: less error in predicting aerodynamic performance from tunnel data. CoreWeave does not specify the error metric or give absolute before-and-after values, so the percentage does not show how large the remaining mismatch is. It is not a measurement of faster lap times.
Aston Martin retained the pipeline, giving its Vehicle Science team a workflow it can retrain and extend as race weekends add data. The next iteration is therefore in the team’s hands, rather than requiring it to start the modeling process again.
Loading discussion...
Join the conversation
Explain when you think quality can—or cannot—compensate for quantity.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.