A Machine Learning Model for Rotor Blade Efficiency Prediction

A machine learning model is presented that is capable of predicting rotor hover performance metrics of thrust coefficient (Ct), power coefficient (Cp), and figure of merit (FM) based on blade geometry and rotor collective operating at a fixed tip Mach number. An original set of sixteen real-world blade geometries is used to construct a large dataset of synthesized blade geometries through blade function interpolation. Several blade geometry parameterization techniques are considered, and a stacked input Principal Component Analysis approach is chosen to reduce input dimensionality to the machine learning models. Model hyperparameters are tuned and a simple linear regressor is compared with the final multi-layer perceptron (MLP) model to quantify the performance gains associated with a more complex model architecture. MLP predictions for blades within the design space are accurate to within ∼ 0.2% normalized root mean squared error (NRMSE), with FM predictions within ∼ 2.3% NRMSE for blade geometries outside the original design space used for training, showing good performance both in interpolation and extrapolation tasks. With this model, rotor hover performance can be evaluated for arbitrary blade geometries at much reduced computational cost than those of traditional simulation approaches.