Aerodynamic Uncertainty Analysis of the Common Research Model Wing Using Autoencoders

Uncertainties in the operations and design of aircraft wings are unavoidable, making it necessary to analyze uncertainties in both surface pressure distributions and aerodynamic coefficients to ensure robustness and reliability in wing design. However, uncertainty analysis through Monte Carlo simulations is prohibitive due to the computational expense of numerical simulations required to evaluate a wing design. Surrogate models can be used to alleviate the expense, but current models typically do not offer predictions of both surface pressure distributions and aerodynamic coefficients for highly nonlinear flow regimes to analyze uncertainties in both simultaneously. As such, this paper proposes UQAeroNet, a neural network architecture that can simultaneously provide a prediction of the surface pressure distribution and the aerodynamic coefficients for the uncertainty analysis of both types of quantities. The proposed approach is characterized by using the NASA Common Research Model wing in transonic viscous flow at a Mach number of 0.80 and Reynolds number of 5.18 million. Quasi-Monte Carlo simulations are utilized to propagate uncertainties in twist angles and operating conditions to analyze the uncertainties in the outputs. The output of the proposed approach is compared with a direct quasi-Monte Carlo simulation of a high-fidelity model. The results demonstrate that the UQAeroNet architecture can accurately model the surface pressure distribution and aerodynamic coefficients, achieving a normalized root mean squared error of 2.5% and below 1%, respectively, with 250 training samples. The results furthermore show that the percentage difference between most summary statistics obtained via UQAeroNet and those obtained directly from the high-fidelity solver is less than or close to 1%. The summary statistics of the pressure distribution are also well captured by the UQAeroNet model and show a large resemblance with those obtained from the high-fidelity solver. This demonstrates that the UQAeroNet architecture is an effective model for aerodynamic uncertainty analysis of wing geometries in transonic flow.