Generalizable Deep Learning Module for Rotorcraft Inverse Design Applications

Airfoil design optimization involves iterative processes, often using less accurate low-fidelity tools due to limitations on computational costs. Surrogate modeling bridges the gap between low computational costs and high-accuracy analysis tools (such as computational fluid dynamics). This study explores the benefits of a vectorized implementation of artificial neural networks (ANNs) as a surrogate modeling tool to predict the performance polar (profiles of lift, drag, and pitching moment coefficients) of airfoils at various angles of attack with high accuracy and a more uniform error distribution. The ANNs are additionally trained on the boundary-layer thickness at the airfoil’s trailing edge. For inverse design challenges, we introduce the tandem neural network (T-NN) architecture, a novel approach utilizing the entire performance polar instead of a single point. Comprising two ANNs working in series, T-NNs feature a modified and flexible loss function, thereby enhancing accuracy in inverse-design problems. Practical design constraints are seamlessly integrated into the design loop, making this architecture a valuable tool for designers. To illustrate real-world applicability, we include the implementation of this deep learning framework for rotor blade design, showcasing the versatility and effectiveness of the proposed methodology.