A multi-dimensional machine learning framework for superior propeller design choices

Abstract

In this study, PropAI is presented, which is a scalable, multi-dimensional surrogate modelling framework for propeller performance prediction that compiles a large 5D baseline database and couples it to a KD-tree Gaussian radial-basis-function (RBF) interpolator. A full-factorial sweep over five design parameters (rotational speed, freestream velocity, blade pitch, diameter, and number of blades) yields 88,540 operating points evaluated via low-fidelity BEMT in QBlade. The trained surrogate reproduces these data with near-machine precision (global RMSE = 6.2e-4, MAE = 1.0e-5, R2 = 1.000), and parity plots of predicted vs true thrust and power lie essentially on the 45 deg line (1:1 line). Cross-validation of leave-one-out (LOO) errors confirms excellent generalisation, with both thrust and power below 0.6% of full-scale output. Integrated with an UNSGA‑III optimiser, the model performs multi-objective optimisation of thrust, power, and thrust-to-power ratio, producing Pareto fronts that reveal the trade-offs among these metrics. This lightweight, gradient-capable surrogate enables rapid design-space exploration (e.g. via 1D/2D response slices, pair-plots, parallel coordinates) and provides insight into parametric interactions.