What if a control surface could find its own angle?
Formula 1 active aero balances the downforce needed for cornering against the drag reduction needed on the straights. We modeled a simplified two element wing, trained a neural network to predict its aerodynamics from flap angle alone, then inverted that model to solve for the angle a given downforce target actually needs.
A sweep, a surrogate, and a search run in reverse
A fixed mainplane and a rotating flap, modeled as a simplified 2D two element wing and run through OpenFOAM at 44 flap angles from fully closed to fully open. Flap closed produced Cd 0.61 and Cl -0.89, heavy downforce and heavy drag. Flap open produced Cd 0.10 and Cl +0.13, an 84% drop in drag, and the wing stopped making downforce entirely.
Those 44 data points trained a small neural network, a two layer MLP, to map flap angle to Cd and Cl in under a millisecond. Freezing its weights turns it into a stand in for the solver, and running gradient descent on the angle itself instead of the usual forward direction lets it work in reverse: given a downforce target instead of an angle, find the angle that produces it.
What the search finds
Flap closed against flap open: the same wing, two different aero states.
Given a downforce target, the frozen model converges on the angle that hits it.
Fed a general objective, minimize drag while keeping meaningful downforce, the optimizer converges on an interior angle of about 10.5 degrees, Cd 0.05, Cl -0.41. Fed a specific target, Cl -0.70 for a given corner, it converges on 15.8 degrees and lands within half a percent of the target.
Where this scales
The model at the center of this is a small MLP, and MLPs are universal function approximators: given enough capacity, they can represent arbitrarily complex mappings between inputs and outputs. Here that is one input and two outputs. The same property holds as you add dimensions. Grid sweeping a design space by brute force gets combinatorially expensive as you add variables. A trained function approximator does not need every combination sampled directly, it generalizes across the space from the data it has seen.
That is what scales: the same architecture that maps one flap angle to two aerodynamic coefficients can map dozens of inputs to whatever a real control surface needs to solve for, without changing the underlying approach. A racecar wing, a drone control surface, or any other aerodynamic problem that needs to react rather than run one fixed setting.