New Gaussian Process Framework Boosts Aerodynamic Prediction Accuracy by 94-96%

Predicting aerodynamic coefficients (lift, drag, moment) for aircraft is notoriously difficult. High-fidelity simulations are computationally prohibitive, while low-fidelity models are biased. A common workaround is to calibrate cheap simulations using experimental data—but that data is sparse, and both control inputs and measurements carry uncertainty.

Davis and Renganathan tackle this with a Bayesian latent Gaussian process (GP) framework built on the Kennedy-O'Hagan calibration approach. They first train a GP surrogate on abundant low-fidelity data, then calibrate it against a limited set of experimental measurements. Critically, they marginalize the surrogate over input uncertainty while matching the marginal mean and variance of the measured output uncertainty. This produces a calibrated model that predicts aerodynamic coefficient uncertainty with remarkable fidelity.

In validation tests against true uncertainty intervals, the model placed 94.2–95.8% of its predictive samples inside the nominal 95% confidence bands, and endpoint cumulative probabilities closely matched the expected 0.025 and 0.975 levels. The method works even at extrapolative input settings, making it practical for flight conditions not covered in training data.

The framework is published on arXiv (2606.28871) and spans 29 pages with 9 figures. It combines tools from machine learning (stat.ML, cs.LG) and computational physics (physics.comp-ph), offering a rigorous, statistically sound path to aerodynamic uncertainty quantification without expensive direct simulations.