Bayesian Machine Learning for Predicting Wing Pressure Distributions at Transonic Flow Conditions

In aircraft design, computational fluid dynamics plays an important role by providing invaluable insights into aerodynamic characteristics, facilitating enhancements in aircraft performance and efficiency. High-fidelity simulations based on the Reynolds-averaged Navier–Stokes equations are essential to capture phenomena like shock waves and flow separation, particularly in the transonic regime. Due to the significant computational cost, conducting the numerous evaluations required to cover the complete flight envelope becomes impractical. Alternatively, leveraging machine learning to construct surrogate models provides approaches to approximate quantities of interest within the design space. However, since these models are not exact, incorporating a measure of model epistemic uncertainty is advantageous. The Bayesian paradigm offers a rigorous framework to train and analyze uncertainty-aware models. This study focuses on evaluating the effectiveness of Bayesian surrogate models, specifically Bayesian neural networks and Gaussian processes, in predicting aerodynamic pressure distribution. A comprehensive comparison is conducted, encompassing both model accuracy and predicted uncertainty, for an airfoil and a representative commercial aircraft model. Variations in surrogate model quality are identified when predicting pressure distribution under transonic flow conditions while exhibiting similar behavior otherwise.