Towards Design of Airfoil Pressure Tap Locations for Real-Time Predictions Under Uncertainty Using Bayesian Neural Networks

Flow separation can be highly dangerous for aircraft operations especially during takeoffs and landings, and active flow control is an important research area tasked with developing technologies to delay the effects of flow separation. Designing active controllers requires reliable real-time in-flight detection of aerodynamic characteristics from sensor measurements. In this work, we focus on (a) constructing a mapping from sparse and noisy pressure tap readings to aerodynamic quantities of interest using Bayesian neural networks (BNNs) trained from SU2 flow simulations; and (b) designing pressure tap location configurations by assessing the BNN performance on testing data with an uncertainty-incorporated loss criterion. BNNs are a type of machine learning model that can provide rapid online predictions as well as uncertainty information surrounding the predicted values, and thus highly useful for supporting decision-making that may impact aircraft and passenger safety. This paper presents a proof-of-concept study of a real-time flow separation detection tool in a simplified two-dimensional environment.