Artificial Neural Networks Applied to Airplane Design

Abstract

Multi-disciplinary frameworks for airplane optimal design require a lot of computational power, which grows enormeously if higher fidelity tools are used to model aeronautical disciplines like aerodynamics, loads, flight dynamics, performance, and structural analysis. In order to address properly and elegantly this issue, surrogate models are employed. In this highlight, the main goal of the present work is the design and application of an artificial neural network to predict aerodynamic coefficients of wing-body configurations of transport airplanes in order to replace a computational fluid dynamic code for future optimizations. The artificial neural network system that was developed is able to predict lift and drag coefficients for wing-fuselage configurations of transport airplanes. The input parameters for the neural network are the wing planform, airfoil geometry, and flight condition. An aerodynamic database consisting of 100,000 cases evaluated with the BLWF V2.81 full-potential code is used for the neural network training. The neural network training is carried out with the back-propagation algorithm, the scaled gradient algorithm, and the Nguyen-Wridow weight initialization. Networks with different numbers of neurons are evaluated in order to minimize the regression error. The optimum networks reduce the computation time of the aerodynamic coefficients in 4000 times when compared with BLWF V2.81, and with an average error of only five counts for the drag coefficient. We present an adaptation of the back-propagation algorithm that allows the computation of the gradients of the neural network outputs in a scalable manner, and then we use it for an airplane optimization task. The results are then compared with similar tasks that were performed by calling the full potential code. An adjoint-method is employed to use ANN in constraint functions as well. The resulting geometry obtained with the ANN methodology is compared to one designed directly with BLWF. The optimal geometry is practically the same, and the drag coefficients predicted by each method differ in four drag counts.