Use of artificial neural networks to correct computer simulations of small-scale propellers

Abstract

© 2021, American Institute of Aeronautics and Astronautics Inc.. All rights reserved.Propellers are one of the most efficient ways to generate propulsion for low-speed flights. About 84% of the energy generated by the engines is utilized, being therefore widely used in several different aircraft. However, studies show that propellers with a diameter less than 16 inches have efficiency reduced by up to 15% when compared to larger ones. This deficiency is not always captured by the mathematical models, since they are not as accurate for that scale. The present study aims to increase the accuracy of simulations performed by a blade element/vortex software to predict the performance of different motor-propeller assemblies. For this purpose, neural networks are trained to correct thrust and torque values given by the software in relation to wind tunnel tests. For this, 28 propellers from different manufacturers and geometries are tested in wind tunnel and simulated in the software under the same conditions to generate the training database. Geometric data of propellers and operational conditions were used as inputs for the neural networks. The outputs are the difference between the results of the test in a wind tunnel and the software simulation. The use of neural networks to correct the simulation results reduced the mean squared error of the estimates at least in 80% in the case of thrust and 70% in the case of torque.