• Framework that predicts unseen 2-bladed propellers performance at varying conditions. • Mixed dataset improves robustness and aligns predictions with experiments. • MLP model outperforms BEMT predictions across wide operating conditions. Accurate and fast prediction of aerodynamic forces is an essential step in the design of any Unmanned Aerial Vehicle (UAV), in particular the prediction of the thrust and torque produced by a propeller. Existing approaches to this span from low-fidelity Blade Element Momentum Theory (BEMT) to high fidelity Computational Fluid Dynamics (CFD) or even experimental testing and present different trade-offs between accuracy, cost, and practicality. This work introduces a machine learning approach to predict thrust and torque for 2-bladed propellers by APC manufacturer with diameters between 7 and 20 inches. To train the Multi-Layer Perceptron (MLP) used in this approach, a mixed dataset was built by combining numerical data generated through an adapted BEMT implementation with experimental data gathered from wind tunnel tests on 19 propellers. The results show that the model accurately predicts thrust and torque under various operational conditions, reaching values below 10% for thrust and demonstrating reliability comparable to traditional methods while significantly reducing computational or experimental demands. This study confirms the viability of machine learning as a predictive tool for propeller performance of small UAV, especially in early-stages of the design process where fast, reliable estimates are needed.
Claro et al. (Wed,) studied this question.