Abstract
In this paper, a new and unique surrogate-based forward propagation algorithm for aerodynamic geometric uncertainty quantification (UQ) is proposed. The proposed algorithm extends the recent efficient global optimization with neural network (NN)-based prediction and uncertainty (EGONN) algorithm which was created for unconstrained optimization problems. The proposed extended EGONN algorithm for UQ (uqEGONN) constructs a global surrogate model of the aerodynamics characteristics using a NN based on data sampled from physics-based computational fluid dynamics (CFD) simulations. The NN model is adaptively enhanced using a NN model of the prediction uncertainty. In each sampling cycle, the prediction NN is used to compute the summary statistics with Monte Carlo simulations (MCS). The algorithm terminates when the absolute relative change in the summary statistics reach a specified tolerance, or when a specified maximum number of samples is reached. The algorithm is demonstrated on the UQ of the RAE 2822 airfoil at Mach 0.734, angle of attack 2.89 deg, and Reynolds number of 6.5 million. The geometric uncertainty is represented using twelve normally distributed parameters. The results show that the proposed algorithm yields comparable results as those obtained from direct CFD-based MCS with less than 100 CFD samples. The mean and standard deviation of the airfoil drag coefficient are within 0.45 drag counts, for the lift coefficient it is within 0.04 lift counts, and for the pitching moment coefficient it is within 0.11 (× 10−2).
