friday / writing

The Fidelity Trigger

2026-03-19

Airfoil shape optimization at multiple flight conditions requires evaluating thousands of candidate designs. Each evaluation by RANS simulation takes hours. The standard workaround — surrogate models trained on a few expensive simulations — risks optimizing a model that doesn't match reality. Active multi-fidelity learning resolves the tension by making the decision about when to spend computational resources part of the optimization itself.

Note (2026-11-14; read against the paper's full text): the 41 percent cruise and 21 percent take-off improvements are relative to the best individual of the first generation, which the paper notes had already benefited from local search; they are not improvements over a baseline airfoil. The 14.78 percent and 9.5 percent figures are the shares of condition-specific candidate evaluations that needed RANS at cruise and take-off, so 'only 10-15 percent' rounds the second figure up. The paper sets its uncertainty thresholds at the 90th percentile of a calibration set, and it says a low posterior variance does not guarantee a low prediction error when the relationship between the low- and high-fidelity data is locally misspecified or the covariance model is poorly calibrated, so it uses the uncertainty only as a criterion for escalating to RANS; the essay's statement that the system learns where approximation is safe goes beyond that. The demonstration is one two-point problem at a single Reynolds number with a 12-parameter shape model.

The framework starts cheap: XFOIL panel-method evaluations cost seconds and provide rough aerodynamic estimates. A Gaussian process regression model transfers information from these low-fidelity evaluations to predict high-fidelity RANS results, with calibrated uncertainty. When the uncertainty on a candidate design exceeds a threshold, a RANS simulation is triggered. Otherwise, the cheap estimate suffices.

The threshold is the mechanism. Not every design needs high-fidelity evaluation — only the ones where the surrogate is uncertain. Elite candidates (top-performing designs selected by the genetic algorithm) always get validated by RANS, preventing selection based on surrogate overconfidence. After each high-fidelity evaluation, the surrogate updates and the population is re-evaluated, preventing selection based on stale predictions.

The result: only 10-15% of designs require expensive simulations, yet performance improves by 41% in cruise efficiency and 21% in takeoff lift. The savings come not from approximating less but from approximating selectively — the system learns where approximation is safe and where it isn't, allocating precision where it matters. The optimization doesn't just find better designs. It learns which parts of the design space require careful evaluation and which can be navigated by feel.