Data-Efficient Machine Learning-Enabled Aerodynamic Optimization of Blades

Abstract

Rotor blade design requires accurate aerodynamic evaluation across a high-dimensionaldesign space, where high-fidelity Computational Fluid Dynamics (CFD) simulations are too expensive for iterative optimization and simplified low-fidelity methods (like potential flow-based solvers) lack sufficient accuracy. This dissertation presents a comprehensive machine learning framework for aerodynamic surrogate modeling and optimization, spanning airfoil-level prediction to full rotor blade design. A Proper Orthogonal Decomposition-based Artificial Neural Network (POD-ANN) surrogate is developed to predict the complete aerodynamic performance polar of airfoil geometries across a wide range of operating conditions, achieving a minimum drag error of 0.44 drag counts while evaluating a full polar, seven orders of magnitude faster than the CFD. A systematic study on the airfoil parameterization technique is also performed. The airfoil inverse design problem is addressed through a Tandem Neural Network (TNN) architecture that overcomes the ill-posedness of the inverse mapping by modifying the loss function in terms of target performance rather than geometry. This allows the TNN to predict the airfoil geometry that matches the target polar with high accuracy (within 1 drag counts). An auxiliary network is used to enforce hard geometric constraints to get practical airfoil geometries. A novel predictor-corrector type multi-fidelity artificial neural network (MFANN) architecture is developed to predict the performance polar of airfoil geometry with much less high-fidelity CFD data. MFANN uses a combination of low-fidelity data and a much smaller number of high-fidelity data (one-tenth) to predict the performance polar of an airfoil without compromising on the accuracy of the model (within 1 drag counts). Such architecture allows the extension of the neural network model to multiple Mach and Reynolds numbers without a significant increase in the cost of data generation. An uncertainty quantification framework is also presented, wherein aleatoric uncertainty arising from geometric variability is quantified using Monte Carlo sampling and epistemic uncertainty is estimated using neural network ensembles. The results demonstrate that the geometric variability and modeling assumptions can influence aerodynamic predictions, especially in the near- and post-stall regions. The impact of boundary-layer transition on airfoil tables is also investigated, and the resulting variations are propagated to wind turbine performance predictions. The framework is extended to the full rotor blade, where a single ANN trained on a 73-dimensional blade geometry parameter predicts thrust and power with R2 > 0.97. A surrogate-assisted Genetic Algorithm achieves a 12.77 percent reduction in hover power at constant thrust, simultaneously optimizing blade planform and spanwise airfoil geometry. This dissertation establishes a scalable and computationally efficient machine learning pipeline for surrogate-based aerodynamic design and optimization of rotorcraft blades.