BEMT-driven neural network for aerodynamic performance prediction and optimisation of high-altitude UAV propellers

This study proposes an efficient aerodynamic prediction and optimisation framework integrating blade element momentum theory (BEMT), a Bayesian-optimised aerodynamic residual network (BO-ARN), and a genetic algorithm (GA) for high-altitude long-endurance solar UAV propellers under stratospheric low-Reynolds-number conditions. An aerodynamic database was generated using Latin hypercube sampling and XFoil. A fully connected residual network was developed to learn nonlinear mappings from Reynolds number, Mach number and angle-of-attack to lift and drag coefficients, while tree-structured Parzen estimator Bayesian optimisation selected network depth, width, learning rate, weight decay and batch size. The trained BO-ARN surrogate was embedded in BEMT to replace repeated XFoil evaluations, and a GA was then used to optimise the spanwise chord and twist distributions under motor power and torque constraints. Compared with a Bayesian-optimised multilayer perceptron, BO-ARN reduced the RMSEs of lift and drag coefficient prediction by 25.65% and 37.80% on the training set and by 4.06% and 7.71% on the test set, respectively. For one propeller operating condition, BO-ARN-supported BEMT reduced the calculation time from 43.322 s to 0.411 s. After optimisation, cruise thrust increased from 17.89 N to 19.49 N, corresponding to an 8.94% improvement, while system propulsion efficiency increased from 68.75% to 70.59%, with the motor constraints satisfied. These results demonstrate that the proposed framework provides an accurate and computationally efficient approach for propeller-motor matching and aerodynamic shape optimisation of high-altitude UAV propulsion systems.