Neural Network for Propeller Performance Prediction and CFD Validation of its Optimal

The article presents a two-layer feed-forward neural network trained with a large set of experimental data to predict the thrust and torque coefficients for small-sized propellers. The training method is the Levenberg-Marquardt, with a K-fold data division into training, testing, and validation. The input data consist of a set of 12 variables that describes the chord and pitch distribution, the rotational speed, the advance ratio J, and the propeller diameter. The neural network shows a fitting higher than 0.75. Furthermore, a particle swarm optimization to explore the neural network design space is carried out for a study case, where the optimal propeller is simulated through Computational Fluid Dynamics, showing adequate results. The problem addressed by this proposal is to generate a fast tool to predict the blade performance with enough accuracy and test the feasibility of the artificial neural networks to be another methodology for it besides the well-known blade element theory. The results show an error similar to blade element theory methods, which proves the suitability of the artificial neural networks for propeller prediction and suggests further improvements in the neural network to reach similar accuracy of 3D computational fluid dynamics results.