Data-Driven Analysis of Rotors' Performance Using Machine Learning

This thesis focuses on the use of machine learning for the design, optimization, and performance analysis of rotors. The ML models were trained using a public dataset provided by the University of Illinois, containing both data related to rotor geometry and performance data under static conditions, in which the advance ratio J is zero, and dynamic conditions, in which J > 0. The training was carried out using MATLAB’s Regression Learner, selecting the most suitable models based on prediction errors. During the course of the work, two codes were developed, designed to be flexible and scalable. The first allows for the design of rotors starting from performance or geometric specifications provided as input; the user can choose whether to generate a Pareto front using a multi-objective genetic algorithm, or to directly obtain a single configuration through a standard genetic algorithm. The second code is dedicated to the optimization of existing rotors: starting from an initial configuration, the model modifies the geometry within user-defined tolerances in order to improve a single selected performance metric, namely the thrust coefficient (CT) or the power coefficient (CP). In addition to design and optimization, the developed models are used for the prediction of the performance of new rotors, allowing for a significant time saving compared to traditional numerical simulations. The results obtained are compared with those derived from BEM (Blade Element Momentum) and CFD (Computational Fluid Dynamics) analyses in order to verify their consistency. Overall, the work constitutes a basis for future extensions, such as the integration of neural networks and the use of larger datasets to improve the generalization capability of the models.